content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
The properties and the evolution of massive clumps hosting the precursors of the highest mass stars currently forming in our Galaxy are poorly known.
Massive clumps at an early evolutionary phase,
{ thus, prior to the emergence of luminous massive young stellar objects and {UC-H\,{\scriptsize II}} regions},
are excellent candidates to host high-mass protostars in their earliest stages (e.g.\, \citealp{Zhang2009,Bontemps2010,Csengeri2011a,Csengeri2011b,Palau2013,SM2013}).
Large samples have only recently been identified based on large area surveys (e.g.\,\citealp{BT2012,Tackenberg2012,Traficante2015,Svoboda2016,Csengeri2017}), which
show that the early evolutionary stages are short lived (e.g.\,\citealp{M07,Csengeri2014}), as star formation proceeds rapidly.
Using the Atacama Large Millimeter/submillimeter Array (ALMA),
here we present the first results of a statistical study of early stage fragmentation
to shed light on the physical processes
at the origin of high-mass collapsing entities,
and to search for the youngest
precursors of O-type stars.
\section{The sample of infrared quiet massive clumps}
Based on a flux limited sample
of the 870 $\mu$m APEX Telescope LArge Survey of the GAlaxy (ATLASGAL, \citealp{Schuller2009, Csengeri2014}), \citet{Csengeri2017}
identified the complete sample of massive infrared quiet clumps
with { the highest peak surface density ($\Sigma_{\rm cl}\geq0.5$\,g\,cm$^{-2}$)\footnote{In the ATLASGAL beam of {19{\rlap{\arcsec}{.}2}.}}
and low bolometric luminosity, $L_{\rm bol}$$<$$10^4$\,L$_{\odot}$,
corresponding to the ZAMS luminosity of a late O type star.
Their large mass reservoir and low luminosity suggest
that infrared quiet massive clumps correspond to the early evolutionary phase,
some already exhibiting signs of ongoing (high-mass)
star formation such as EGOs and Class~II methanol masers.
Here we present the sample of
35 infrared quiet massive clumps
located within $d\leq4.5$\,kpc, which could be conveniently grouped on the sky as targets for ALMA.
They cover 70\% of all the most massive and nearby infrared quiet clumps from
\citet{Csengeri2017}, and
are thus a representative selection of a
homogenous sample of
early phase massive clumps in the inner Galaxy.
\section{Observations and data reduction}
We present observations carried out in Cycle~2 with the ALMA 7m array using
$9$ to $11$ of the 7m antennas with baselines ranging between 8.2\,m (9.5k$\lambda$)
to 48.9\,m (53.4k$\lambda$).
We used a low-resolution wide-band setup
in Band 7, yielding $4\times1.75$\,GHz
effective bandwidth with a spectral resolution of 976.562 kHz.
The four basebands were centred on
347.331, 345.796, 337.061, 335.900\,GHz, respectively.
The primary beam at this frequency is 28.9\arcsec.
Each source was observed for $\sim$5.4~min in total.
The system temperature, $T_{\rm sys}$ varies between $100-150$~K.
{ The targets have been
split according to Galactic longitude in five observing groups (Table\,\ref{tab:table-obs}).}
{
The data was calibrated using standard procedures in CASA~4.2.1.}
To obtain line-free continuum images, we first identified the channels with
spectral lines towards each source, and excluding these averaged the remaining channels.
We used a robust weight of 0.5 for imaging, and the CLEAN algorithm
for the deconvolution, and corrected for the primary beam attenuation.
The synthesized beam varies between $3.5$\arcsec\ to $4.6$\arcsec\
taking the geometric mean of the major and minor axes.
{ The noise has been measured in an emission free area close to the center of the maps { including the side-lobes}.}
The achieved median $rms$ noise level is $54$\,mJy/beam and
varies among the targets due to a combination of restricted bandwidth available for continuum, dynamic range or mediocre observing conditions. In particular for
groups 4 and 5, the observations have been carried out at low elevation
resulting in an elongated beam and poor $uv$-sampling.
{ The observing parameters per group are summarized in Table\,\ref{tab:table-obs},
and for each source in Table\,\ref{table:long}.}
\begin{table*}[!ht]
\centering
\caption{Summary of observations.}\label{tab:table-obs}
\begin{tabular}{cccclrrrrrrrrrrr}
\hline\hline
\multicolumn{2}{c}{Observing group} & Date & Bandpass & Phase & Flux & \multicolumn{3}{c}{Synthesized beam\tablefootmark{a}} & $\sigma_{\rm rms}$ \tablefootmark{b} \\
& & & calibrator & calibrator & calibrator & [\arcsec$\times$\arcsec] & [$^{\circ}$]& [\arcsec] & [mJy] \\
\hline
1 & $320<\ell<330^{\circ}$& 8, 16 July 2014& J1427-4206 & J16170-5848 & Titan, Ceres & $5.0\times2.9$& -78.6 & $3.8$ & $19.3-83.5$\\
2 & $330<\ell<340^{\circ}$ & 18, 21 July 2014 &J1427-4206 & J1617-5848 & Titan, Ceres &
$4.6\times 2.8$& 14.9& 3.6& $20.7-119.2$\\
3 & $340<\ell<350^{\circ}$ & 19, 21 July 2014 & J1517-2422 & J1636-4102 & Titan, Ceres &
$4.7\times 2.6$ & -83.4 & 3.5
& $22.9-105.3$ \\
4 & $350<\ell<360^{\circ}$& 14, 15 June 2014 & J1733-1304 & J1717-3342 & Neptune &
$9.2\times 2.4$ & -76.2 & 4.6 & $28.7-175.8$ \\
5 & $30<\ell<40^{\circ}$& 8 June 2014 & J1751+0939 &J1851+0035 & Neptune &
$5.8\times2.4$& -68.2 & 3.7 & $16.4-45.8$ \\
\hline
\end{tabular}
\tablefoot{
\tablefoottext{a}{Averaged properties.}
\tablefoottext{b}{The minimum and maximum $\sigma_{\rm rms}$ noise is averaged over the line-free channels in the total 7.5\,GHz bandwidth.}
}\end{table*}
\begin{figure}[]
\centering
\includegraphics[width=0.74\linewidth]{FIG/G320p2325-0p2844.eps}
\caption{\small
{\it Left:} Clump-scale view by ATLASGAL of an example source.
{\it Right:} Line-free continuum emission at 345\,GHz by the ALMA 7m array. Contours start at 7$\sigma_{\rm rms}$ noise and increase in a logarithmic scale. White crosses mark the extracted sources (see Table\,\ref{table:long}). The synthesized beam is shown in the lower left corner.
}
\label{fig:overview}%
\end{figure}
\input{mass_input_T25_ag2_short}
\section{Results and analysis}\label{sec:res}
Compact continuum emission is detected towards all clumps
(see Fig.\,\ref{fig:overview} for an example, and
Fig.\,A\,\ref{fig:all} for all targets).
We find sources that stay single ($\sim$14\%)
at our resolution and sensitivity.
Fragmentation is, in fact, limited towards the majority of the sample;
$45$\% of the clumps hosts up to two, while $77$\% host up to three compact sources.
Only a few clumps host more fragments.
We identify and measure the parameters of the compact sources
using the Gaussclumps task in GILDAS\footnote{Continuum and Line Analysis Single-Dish Software http://www.iram.fr/IRAMFR/GILDAS}, which performs a 2D
Gaussian fitting. A total number of 124 fragments down to a $\sim$7\,$\sigma_{\rm rms}$ noise
level are systematically identified within the primary beam, where
the noise is measured towards each field.
This gives on average, $\bar{N}_{\rm fr}$=3 sources per clump
corresponding to a population of cores
at the typically achieved physical resolution of $\sim$0.06\,pc.
We can directly compare
the integrated flux in compact sources seen by the ALMA 7m array with
the ATLASGAL flux densities measured over the primary beam of the array
as both datasets have similar centre frequencies\footnote{
The centre frequency for the ALMA dataset is at 341.4\,GHz, while for the
LABOCA filter, it is around 345\,GHz.
A spectral index of $-3.5$
gives 10\% change in the flux
up to a difference of 10\,GHz in the centre frequencies. This is below
our absolute flux uncertainty.
}.
We recover between 16-47\%
of the flux, the rest of the emission is filtered above the typically 19\arcsec\
largest angular scale sensitivity
of the ALMA 7m array observations.
To estimate the mass, we assume optically thin dust emission and use
the same formula as in \citet{Csengeri2017}; $M = S_{\rm 870\mu m} d^2\,\kappa_{\rm 870\mu m}\,^{-1}\,B_{{\rm 870\mu m}}(T_{\rm d})^{-1}$, where $S_{\rm 870\mu m}$ is the integrated flux density, $d$ is the distance, $\kappa_{\rm 870\mu m}=0.0185$\,g\,cm$^{-2}
$ from \citet{OH1994} accounting for a gas-to-dust ratio of 100, and $B_{\nu}(T_{\rm d})$ is the Planck function. While on the $\sim$0.3\,pc scales of clumps \citet{Csengeri2017} adopt $T_{\rm d}$=18\,K, on the smaller scales of cores heating due to the embedded protostar may result in elevated dust temperatures that are poorly constrained. Following the model of \citet{GoldreichKwan1974}, we estimate
$T_{\rm d}$=15-38\,K
for the luminosity range of $10^2-10^4$\,L$_{\odot}$\, at a typical radius of half the deconvolved $FWHM$ size of 0.025\,pc.
We adopt thus $T_{\rm d}$=25\,K
which results up to a factor of two uncertainty in the mass estimate.
\begin{figure}[]
\centering
\includegraphics[width=3.7cm,angle=90]{FIG_NEW2/histo_mass_new_final.eps}
\caption{\small
Mass distribution of MDCs within $d$$\leq$4.5\,kpc. The Poisson error of each bin is shown as a grey line above the 10$\sigma_{\rm rms}$ completeness limit of 50\,M$_{\odot}$, the power-law fit is shown in a solid black line.
Hashed area shows the distribution of the brightest cores ($M^{\rm max}_{\rm MDC}$) per clump. Dashed lines show the slope of the CMF/IMF \citep{Andre2014},
and CO clumps \citep{Kramer1998}. }
\label{fig:flux_distr}
\end{figure}
The extracted cores have a mean mass of $\sim$63\,M$_{\odot}$\
corresponding to massive dense cores (MDCs as in \citealp{M07}),
and about 40\% of the sample hosts cores more massive than 150\,M$_{\odot}$.
They are, in terms of physical properties, similar to SDC335-MM1 \citep{Peretto2013},
which is here the most massive core with
$\sim$400\,M$_{\odot}$\ within a deconvolved FWHM size of $0.054$\,pc\footnote{
Our mass estimates for SDC335-MM1
can be reconciled with \citet{Peretto2013} using a dust emissivity index of $\beta$$\sim$1.2 between 93\,GHz and 345\,GHz. A similarly low value of $\beta$ is also suggested by \citet{Avison2015}.}.
In these clumps the second brightest sources are also typically massive,
on average 78\,M$_{\odot}$\ suggesting a preference to form more massive cores.
Except for one clump, no core is detected below 35\,M$_{\odot}$\
which is well above the typical detection threshold considering
the
mean 7$\sigma_{\rm rms}$ mass sensitivity of 11.2\,M$_{\odot}$\ at the mean distance of 2.6\,kpc, and may indicate a
lack of intermediate mass (between 10--40\,M$_{\odot}$) cores.
Similar findings have been reported towards a handful of other young massive sources by \citet{Bontemps2010} and \citet{Zhang2015}.
Clumps with single sources host strictly massive cores with $M_{\rm MDC}$$>$40\,M$_{\odot}$,
and about half of them reaches the highest mass range of $M_{\rm MDC}$$>$150\,M$_{\odot}$.
We show the mass distribution of cores as $\Delta N/\Delta {\rm log}\,M\sim M^{\alpha}$ in Fig.\,\ref{fig:flux_distr}, and indicate the 10$\sigma_{\rm rms}$ completeness limit of 50\,M$_{\odot}$,
set by the highest noise in the poorest sensitivity data.
The distribution tends to be
flat up to the completeness limit, and then shows a decrease at
the highest masses.
The distribution of $M^{\rm max}_{\rm MDC}$ (hatched histogram) shows that the majority of the clumps host at least one massive core,
while a few host only at most intermediate mass fragments.
The least square power-law fit
to the highest mass bins above the completeness limit gives $\alpha=-1.01\pm0.20$, which is steeper than the distribution of CO clumps ($\alpha$=$-0.6$ to $-0.8$, \citealp{Kramer1998}), and tends to be shallower than the low-mass prestellar CMF and the stellar initial mass function (IMF) ($\alpha$=-1.35-- -1.5, \citealp{Andre2010}), although
at the high-mass end the scatter of the measured slopes is more significant \citep{Bastian2010}.
Using Monte Carlo methods we test the uncertainty of $\alpha$ due to the unknown
dust temperature, and simulated a range of $T_{\rm d}$ between 10$-$50\,K using a normal distribution with
a mean of 25\,K, and a power-law distribution.
We fitted to the slope the same way, as above, and repeated the tests
until the standard deviation of the measured slope reached convergence.
In good agreement with the observational results, { the
normal} temperature distribution gives $\alpha_{\rm MC}$=-1.01$\pm$0.11, and thus
constrains the error of the fit suggesting an intrinsically shallower slope than the IMF.
A power-law temperature distribution in the same mass range with an exponent of $-0.5$,
could reproduce, however, the slope of the IMF, assuming that
the brightest sources are intrinsically warmer.
Alternatively, a larger level of fragmentation of the brightest cores
on smaller scales
could also reconcile our result with the IMF.
\section{Discussion}
\begin{figure}[]
\centering
\includegraphics[width=5.0cm,angle=90]{FIG/sigma_mass_paper.eps}
\caption{\small
Surface-density versus mass diagram, coloured dotted lines in different shades show constant radius (green) and $n_{\rm H}$ number density (red)
(c.f.\,\citealp{Tan2014}).
Colored large circles show clumps (ATLASGAL), while smaller circles the cores (ALMA 7m array), colors scaling from blue to red with increasing $M^{\rm max}_{\rm MDC}$. We mark two massive cores with $M_{\rm MDC}=60$\,M$_{\odot}$\ (C1-S, \citealt{Tan2013}) and 55\,M$_{\odot}$\ (CygX-N63, \citealp{Bontemps2010}). For comparison IRDC clumps \citep{Kainulainen2013} and cores are shown \citep{BT2012}. Gray arrows show two models: 1) a uniform clump density, and 2) a single central object with an $r^{-2}$ density profile.
}
\label{fig:surfdens}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=5.0cm,angle=90]{FIG/cfe.eps}
\caption{\small
CFE versus average clump density ($\bar{n}_{\rm cl}$).
Green triangles show cores of $\rho$~Oph \citep{M98} and red diamonds of Cygnus-X \citep{Bontemps2010}.
}
\label{fig:cfe}
\end{figure}
\subsection{Limited fragmentation from clump to core scale}\label{sec:monolothic_collapse}
The thermal Jeans mass in massive clumps is low ($M_{\rm J}\sim$1\,M$_{\odot}$\
at $\bar{n}_{\rm cl}$=4.6$\times10^5$\,cm$^{-3}$, $T$=18\,K), which is expected to lead to
a high degree of fragmentation.
In contrast, the observed
infrared quiet massive clumps exhibit here limited fragmentation with $\bar{N}_{\rm fr}=3$,
from clump to core scales.
We even find single clumps/MDCs at our resolution.
This is intriguing also because these most massive clumps of the Galaxy are expected to form rich clusters.
The selected highest peak surface density clumps could therefore correspond to a phase of compactness
where the large level of fragmentation to form a cluster has not yet developed.
We find that the mass surface density ($\Sigma$) increases towards
small scales (Fig.\,\ref{fig:surfdens}, c.f.\,\citealp{Tan2014})
corresponding to a high concentration of mass.
80\% of the clumps host MDCs above 40\,M$_{\odot}$, and the
most massive fragments scale with the mass of their clump.
Two models are shown with arrows in Fig.\,\ref{fig:surfdens}:
1) clumps with a uniform mass distribution
forming low mass stars correspond to a roughly constant mass surface density;
2) clumps with all the mass concentrated in a single object corresponding to $n(r)\sim r^{-2}$ density profile. The majority of the sources fit better the steeper than uniform
density profile.
The early fragmentation of massive clumps thus
does not seem to follow thermal processes,
and shows fragment masses largely exceeding the local Jeans-mass
(see also \citealp{Zhang2009, Bontemps2010,Wang2014,Beuther2015, BT2012}).
The significant concentration of mass on small scales also manifests in a high core formation efficiency (CFE), which { is the ratio of the total mass in fragments and the total clump mass from \citet{Csengeri2017} adopting the same physical parameters} (Fig.\,\ref{fig:cfe}). The CFE suggests an increasing concentration of mass in cores with the average clump volume density ($\bar{n}_{\rm cl}$), a trend which has been seen, although inferred from smaller scales, towards high-mass infrared quiet MDCs in Cygnus-X (\citealp{Bontemps2010}), and low-mass cores in $\rho$\,Oph (\citealp{M98}), and a sample of infrared bright MDCs (\citealp{Palau2013}). Although the CFE shows variations at high densities with $\bar{n}_{\rm cl}>10^5$\,cm$^{-3}$, exceptionally high CFE of over 50\%, can only be reached towards the highest average clump densities.
\subsection{Which physical processes influence fragmentation?}
\label{sec:stability}
What can explain that the thermal Jeans mass does not represent well the observed fragmentation properties in the early stages?
A combination of turbulence, magnetic field, and radiative feedback
could increase the necessary mass scale for fragmentation.
Using the Turbulent Core model \citep{MT03}
for cores with $M_{\rm MDC}$$>$$150$\,M$_{\odot}$\
at the average radius of 0.025\,pc,
we estimate
from their Eq.\,18 a turbulent line-width of $\Delta v_{\rm obs}$$\gtrsim$$6$\,km~s$^{-1}$\ at the surface of cores,
which is a factor of two higher than the average $\Delta v_{\rm obs}$
at the clump scale \citep{Wienen2015}. The magnetic critical mass at the
average clump density corresponds to $M_{\rm mag}$$<$400\,M$_{\odot}$\ at
the typically observed magnetic field values of 1\,mG towards massive clumps
(e.g.\,\citealp{Falgarone2008,Girart2009,Cortes2016,Pillai2016})
following Eq.2.17 of \citet{BertoldiMcKee1992}. This suggests that
moderately strong magnetic fields could explain the large core masses, however,
at the high core densities of $\bar{n}_{\rm core}$=4$\times$10$^7$\,cm$^{-3}$
considerably stronger fields, at the order of B$>$10\,mG, would be required
to keep the most massive cores subcritical.
Although radiative feedback could also
limit fragmentation (e.g.\,\citealp{Krumholz2007, Longmore2011}),
infrared quiet massive clumps are at the onset of star formation activity
and we lack evidence for a potential deeply embedded population of low-mass protostars
needed to heat up the collapsing gas.
\subsection{Can global collapse explain the mass of MDCs?}
The rather monolithic fashion of collapse
suggests that fragmentation is at least partly determined already at the clump scale,
which would be in
agreement with observational signatures of global collapse of massive
filaments
(e.g.\,\citealp{Schneider2010, Peretto2013}).
If entire cloud fragments undergo collapse, and equilibrium may not be reached on small scales
leading to the observed limited fragmentation and a high core formation efficiency at early stages.
Mass replenishment beyond the clump scale could fuel the formation of the lower mass population of stars leading to an
increase in the number of fragments with time, and
allowing a Jeans-like fragmentation to develop at more evolved stages (e.g.\,\citealp{Palau2015}).
At the scale of cloud fragments, if collapse sets in at a lower density range of $\bar{n}_{\rm cloud}=10^2$\,cm$^{-3}$,
the initial thermal Jeans mass could reach $M_{\rm J}$$\sim$50\,M$_{\odot}$\ assuming T=18\,K,
at a characteristic $\lambda_{\rm Jeans}$ of about 2.3\,pc.
This
is consistent with the extent of globally collapsing clouds,
the involved mass range is, however, not sufficient to
explain the mass reservoir of the most massive cores.
Considering the turbulent nature of molecular clouds
in the form of large-scale flows, their shocks
could compress larger extents of gas at higher
densities depending on the turbulent mach number (c.f.\,\citealp{Chabrier2011}),
and lead to an increase in the initial mass reservoir.
Fragmentation inhibition and the observed high CFE
are thus consistent with a collapse setting in at parsec scales.
The origin of their initial mass reservoir, however, still poses a challenge
to current star formation models.
\subsection{Towards the highest mass stars}
The mass distribution of MDCs could be reconciled with the IMF either if multiplicity prevailed on smaller than 0.06\,pc scales, or if the temperature distribution scales with the brightest fragments. Similar results have been found towards
MDCs in Cygnus-X by \citet{Bontemps2010}, but also towards Galactic infrared-quiet clumps, such as G28.34+0.06 P1 \citep{Zhang2015}, and G11.11-0.12 P6 \citep{Wang2014}.
Alternatively, the high core formation efficiency and a shallow core mass distribution could suggest
an intrinsically top-heavy distribution of high-mass protostars at the early phases.
Considering the twelve highest mass cores with $M_{\rm MDC}$=150$-$400\,M$_{\odot}$\ and an efficiency ($\epsilon$) of $10-30$\% (e.g.\,\citealp{Tanaka2016}), we could expect a population of stars with a final stellar mass of $M_{\star}\sim \epsilon\times M_{\rm MDC}=15-120$\,M$_{\odot}$, reaching the highest mass O-type stars.
\section{Conclusions}
We study the fragmentation of a representative selection of a homogenous sample of massive
infrared-quiet clumps, and reveal a population of MDCs reaching up to $\sim$400\,M$_{\odot}$.
A large fraction (77\%) of clumps exhibit limited fragmentation,
and host MDCs.
The fragmentation of massive clumps
suggests a large concentration of mass at small scales and a high CFE.
We lack observational support for strong enough turbulence and magnetic field to keep
the most massive cores virialized.
Our results are consistent with entire cloud fragments in global collapse,
while the origin of their pre-collapse mass reservoir still challenges current star formation models.
\begin{acknowledgements}
We thank the referee for constructive comments on the manuscript. This paper makes use of the ALMA data: ADS/JAO.ALMA 2013.1.00960.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. T.Cs. acknowledges support from the \emph{Deut\-sche For\-schungs\-ge\-mein\-schaft, DFG\/} via the SPP (priority programme) 1573 'Physics of the ISM'.
HB acknowledges support from the European Research Council under the Horizon 2020 framework program via the ERC Consolidator Grant CSF-648505. LB acknowledges support from CONICYT PFB-06 project.
A.P. acknowledges financial support from UNAM-DGAPA-PAPIIT IA102815 grant, M\'exico.
\end{acknowledgements}
\vspace{-1cm}
\bibliographystyle{aa}
|
\section{\textbf{Introduction:}}
The notion of convergence of real double sequences was first introduced by Pringsheim \cite{Pr}. A double
sequence $x = \{x_{jk}\}_{ j, k \in \mathbb {N}} $ of real numbers is said to converge to a real
number $l$, if for any $\varepsilon > 0 $, there exists $ m \in \mathbb{N} $ such that for all $ j, k \geq m $
\begin{center}
$| x_{jk} - l | < \varepsilon $ .
\end{center}
In this case we write $ \lim\limits_{\stackrel{\stackrel{j\rightarrow
\infty}{k\rightarrow \infty}} ~} x_{jk} = l $. This notion of convergence
of real double sequences has been extended to statistical convergence by Mursaleen et. al. \cite{Mu}
( also by Moricz \cite{Mo} who introduced it for multiple sequences) using double natural density of $\mathbb{N} \times \mathbb{N}$.
A subset $K \subset \mathbb{N} \times \mathbb{N} $ is said to have natural density $d(K)$ if
\begin{center}
$d(K) = \lim\limits_{\stackrel{\stackrel{m\rightarrow
\infty}{n\rightarrow \infty}} ~} \frac{| K(m,n)|}{m.n} $,
\end{center}
where $K(m,n) = \{(j,k) \in \mathbb{N} \times \mathbb{N}: j \leq m, k \leq n;(j,k)\in K \} $ and $\left|K(m,n)\right|$ denotes number of elements of the set $K(m,n)$.
A double sequence $x = \{x_{jk}\}_{ j, k \in \mathbb{N}}$ of real numbers is said to
be statistically convergent to $\xi \in \mathbb{R}$, if
for any $\epsilon > 0$, we have $d(A(\epsilon)) = 0$, where
$A(\epsilon) = \{ (j,k) \in \mathbb{N}\times \mathbb{N} : \parallel x_{jk} - \xi \parallel \geq \varepsilon \}$.
More investigation and applications of statistical convergence of double sequences can be found in \cite{Da1,Tr2} and many others.
The notion of statistical convergence of double sequences has been further generalized to $I$-convergence of double sequences by Das et. al. \cite{Da2} using ideals in $\mathbb{N}\times\mathbb N$. For more details one can see \cite{Da1, Da3, Tr1} etc.
Recently Belen et. al. \cite{Be} introduced the notion of ideal statistical convergence of double sequences, which is a new generalization of the notions of statistical convergence and usual convergence. More investigation and applications on this notion can be found in \cite{Be, Ya}.
The concept of rough convergence of double sequence was first introduced by Malik et. al.\cite{Ma1}. If $x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}$ be a double sequence in some normed linear space $(X,\left\|.\right\|)$ and $r$ be a non negative real number, then $x$ is said to be $r$-convergent to $\xi\in X$ if for any $\varepsilon>0$, there exists $ m \in \mathbb{N} $ such that for all $ j, k \geq m $
\begin{center}
$| x_{jk} - \xi | < r+\epsilon $.
\end{center}
Further this notion of rough convergence of double sequence has been extended to rough statistical convergence of double sequence by Malik et. al.\cite{Ma2} using double natural density of the subsets of $\mathbb{N}\times\mathbb N$ in a similar way as the notion of convergence of double sequence in Pringsheim sense was generalized to statistical convergence of double sequences. Further the notion of rough statistical convergence of double sequences was generalized to rough $I$-convergence of double sequences by Dunder et. al. \cite{Du}. So it is quite natural to think , if the new notion of $I$-statistical convergence of double sequences can be introduced in the theory of rough convergence.
In this paper we introduced and study the notion of rough $I$-statistical convergence of double sequences in a normed linear space $(X,\left\|.\right\|)$ which naturally extends both the notions of rough convergence as well as rough statistical convergence of double sequences in a new way. We also define the set of all rough $I$-statistical limits of a double sequence and investigate some topological properties of this set.
\section{\textbf{Basic Definitions and Notations}}
\begin{defn}\cite{Ma2}
Let $x = \{x_{jk}\}_{j,k \in \mathbb{N}}$ be a double sequence in a
normed linear space $(X, \parallel.\parallel)$ and $r$ be a non negative real number.
$x$ is said to be $r$- statistically convergent to $\xi$, denoted by $x \overset{r-st}\longrightarrow \xi$,
if for any $\varepsilon > 0$ we have $d(A(\varepsilon)) = 0$, where
$A(\varepsilon) = \{ (j,k) \in \mathbb{N}\times \mathbb{N} :~~ \parallel x_{jk} - \xi \parallel \geq r + \varepsilon \}$.
In this case $\xi$ is called the $r$-statistical limit of $x$.
\end{defn}
\begin{defn}
A class $I$ of subsets
of a nonempty set $X$ is said to be an ideal in $X$ provided
(i) $\phi\in I$.
(ii) $ A,B\in I$ $~$implies $A\bigcup B\in I$.
(iii) $ A\in I,B\subset A $$~$ implies $~~$ $B\in I$.
$I$ is called a nontrivial ideal if $X\notin I$.
\end{defn}
\begin{defn}
A non empty class $F$ of
subsets of a nonempty set $X$ is said to be a filter in $X$ provided
(i) $\phi\notin F$.
(ii) $A,B\in F$ $~$ implies $~~$ $A\bigcap B\in F$.
(iii) $A\in F,A\subset B$ $~$ implies $~~$ $B\in F$.
If $I$ is a nontrivial ideal in $X$, $X\neq\phi$, then the class
\begin{center}
$F(I)=\{ M \subset X : M = X \setminus A$ for some $A \in I \}$
\end{center}
is a filter on $X$, called the filter associated with $I$.
\end{defn}
\begin{defn}
A nontrivial ideal $I$ in $X$ is called admissible if $\{x\} \in I$ for each $x \in X $.
\end{defn}
\begin{defn}
A nontrivial ideal $I$ on $\mathbb{N} \times \mathbb{N} $ is called strongly admissible if $\left\{i\right\}\times \mathbb{N}$ and $\mathbb{N}\times\left\{i\right\}$ belong to $I$ for each $i\in \mathbb{N}$.
\end{defn}
Clearly every strongly admissible ideal is admissible.
Throughout the paper we take $I$ as a strongly admissible ideal in $\mathbb{N} \times \mathbb{N} $.
\begin{defn}\cite{Ma3}
Let $ x = \{x_{jk}\}_{j,k \in \mathbb{N}} $ be a double sequence in a normed linear space $ (X, \parallel . \parallel) $ and $ r $ be a non negative real number. Then $ x $ is said to be rough $I$-convergent or $ r-I$-convergent to $ \xi $,
denoted by $ x \overset{r-I}\longrightarrow \xi $, if for any $ \varepsilon > 0 $ we have $ \{(j,k) \in \mathbb{N} \times \mathbb{N} :\parallel x_{jk} - \xi \parallel \geq r + \varepsilon \} \in I $. In this case $ \xi $ is called rough $ I $-limit of $ x $ and $x$ is called rough $I$-convergent to $\xi$ with $r$ as roughness degree.
\end{defn}
Now we give the definition of $I$-asymptotic density of a subset of $\mathbb{N}\times\mathbb N$.
\begin{defn}
A subset $K \subset \mathbb{N} \times \mathbb{N} $ is said to have $I$-asymptotic density $d_I(K)$ if
\begin{center}
$d_I(K) = I-\lim\limits_{\stackrel{\stackrel{m\rightarrow
\infty}{n\rightarrow \infty}} ~} \frac{| K(m,n)|}{m.n} $,
\end{center}
where $K(m,n) = \{(j,k) \in \mathbb{N} \times \mathbb{N}: j \leq m, k \leq n;(j,k)\in K \} $ and $\left|K(m,n)\right|$ denotes number of elements of the set $K(m,n)$.
\end{defn}
\begin{defn}$[1]$
A double sequence $x = \{x_{jk}\}_{j,k\in \mathbb{N}}$ of real numbers is $I$-statistically convergent to $L$, and we write $x\stackrel{I-st}{\rightarrow}L$, provided that for any $\epsilon> 0$ and $\delta> 0$
\begin{center}
$\left\{\left(m,n\right)\in\mathbb{N}\times\mathbb{N} :\frac{1}{mn}\left|\left\{\left(j,k\right):\left|x_{jk}-L\right|\geq\epsilon,j\leq m,k\leq n\right\}\right|\geq\delta\right\}\in I$.
\end{center}
\end{defn}
\begin{defn}
Let $x=\{x_{jk}\}_{j,k \in \mathbb{N}}$ be a double sequence in a normed
linear space $\left(X,\left\|.\right\|\right)$ and $r$ be a non
negative real number. Then $x$ is said to be rough
$I$-statistically convergent to $ \xi $ or
$r\mbox{-}I$-statistically convergent to $\xi$ if for any
$\varepsilon > 0$ and $ \delta > 0$
\begin{center}
$\{ (m,n) \in {\mathbb{N}\times\mathbb N}: \frac{1}{mn}|\{(j,k):j\leq m, k \leq n; \parallel x_{jk} - \xi \parallel \geq r+ \varepsilon\}|
\geq \delta \} \in I $.
\end{center}
\end{defn}
In this case $\xi$ is called the rough $I$-statistical limit of
$x=\{x_{jk}\}_{j,k \in \mathbb{N}}$ and we denote it by $ x
\overset{r\mbox{-}I \mbox{-}st}\longrightarrow \xi $.
\begin{rem}
Note that if $I$ is the ideal $I_0 = \{ A \subset \mathbb{N} \times \mathbb{N}: \exists
m(A) \in \mathbb{N} ~such ~ that ~ i,j \geq m(A) \Rightarrow (i,j) \notin A \}$, then rough $I$-statistical
convergence coincide with rough statistical convergence.
\end{rem}
Here $r$ in the above definition is called the roughness degree of
the rough $I$-statistical convergence. If $r=0$, we obtain the
notion of $I \mbox{-}$statistical convergence. But our main
interest is when $ r > 0 $. It may happen that a double sequence $x=\{x_{jk}\}_{j,k \in \mathbb{N}}$ is not $ I $-statistically convergent
in the usual sense, but there exists a double sequence $y=\{y_{jk}\}_{j,k \in \mathbb{N}}$, which is $ I $-statistically convergent and
satisfying the condition $\| x_{jk} -y_{jk} \|\leq r $ for all $j,k$. Then $ x $ is
rough $I$-statistically convergent to the same limit.
From the above definition it is clear that the rough
$I$-statistical limit of a double sequence is not unique. So we consider
the set of rough $I$-statistical limits of a double sequence $x$ and we
use the notation $I\mbox{-}st\mbox{-}\mbox{LIM}_x^r$ to denote the
set of all rough $I$-statistical limits of a double sequence $x$. We say
that a double sequence $x$ is rough $I$-statistically convergent if
$I\mbox{-}st\mbox{-}\mbox{LIM}_x^r\neq \phi$.
Throughout the paper $X$ denotes a normed linear space $(X, \|. \|)$ and $x$ denotes the double sequence $\{x_{jk}\}_{j,k \in \mathbb{N}}$ in $X$ .
We now provide an example to show that there exists a double sequence which is neither
rough statistically convergent nor $I$-statistically convergent but is rough $I$-statistically convergent.
\begin{Example}
Let $ I $ be a nontrivial strongly admissible ideal in $\mathbb{N}\times\mathbb N$ which contains at least one infinite subset of $\mathbb{N}\times \mathbb N$. Choose an
infinite set $ A \subset {\mathbb{N}\times\mathbb N} $, whose $I$-asymptotic density is zero but
double natural density does not exists. We define a double sequence $x=\{x_{jk}\}_{j,k \in \mathbb{N}}$ in the following way
\[ x_{jk} = \left\{
\begin{array}{l l}
(-1)^{j+k}, & \quad \text{if $ (j,k)\notin A$ }\\
jk, & \quad \text{if $(j,k)\in A$}.
\end{array} \right.\]\\
Then $x$ neither rough statistically convergent nor
$I$-statistically convergent but
\[ I\mbox{-}st\mbox{-}LIM^r_x = \left\{
\begin{array}{l l}
\emptyset, & \quad \text{if $ r < 1 $ }\\
\left[1-r, r-1\right], & \quad \text{otherwise.}
\end{array} \right.\]\\
\end{Example}
\section{\textbf{Main Results}}
In this section we discuss some basic properties of rough $I$-statistical convergence of double sequences.
\begin{thm}
Let $x = \{x_{jk} \}_{j, k \in \mathbb{N}} $ be a double sequence in $X$ and
$r \geq 0 $. Then diam ($I\mbox{-}st\mbox{-}\mbox{LIM}_x^r) \leq 2r $.
In particular if $x$ is $I$-statistically convergent to $\xi$,
then $I\mbox{-}st\mbox{-}\mbox{LIM}_x^r
={\overline{{B_r}(\xi)}}(=\left\{y\in
X:\left\|y-\xi\right\|\leq r\right\})$ and so diam
($I\mbox{-}st\mbox{-}\mbox{LIM}_x^r) = 2r $.
\end{thm}
\begin{proof}
If possible let, diam ($I\mbox{-}st\mbox{-}\mbox{LIM}_x^r) > 2r $. Then there exist $ y,z \in I\mbox{-}st\mbox{-}\mbox{LIM}_x^r $ such that $\| y - z \| > 2r $. Now choose $ \varepsilon > 0$
so that $ \varepsilon < \frac{\| y - z \|}{2} - r $.
Let
$ A = \{(j,k) \in \mathbb{N\times N} : \|x_{jk} - y \| \geq r +\varepsilon \}$
and
$ B = \{(j,k) \in \mathbb{N\times N} : \|x_{jk} - z \| \geq r + \varepsilon \}$.
Then
\begin{center}
$ \frac{1}{mn} |\{ (j,k):j\leq m;k \leq n,(j,k) \in A \cup B \}|
\leq \frac{1}{mn}|\{(j,k):j\leq m; k \leq n,(j,k) \in A\}| + \frac{1}{mn}|\{(j,k):j\leq m; k \leq n ,(j,k)\in B \}|$,
\end{center}
and so by the property of $I$-convergence \\
$ I \mbox{-}\lim\limits_{\stackrel{\stackrel{m\rightarrow
\infty}{n\rightarrow \infty}} ~} \frac{1}{mn} |\{ (j,k):j\leq m;k \leq n,(j,k) \in A \cup B \}| \leq
I \mbox{-}\lim\limits_{\stackrel{\stackrel{m\rightarrow
\infty}{n\rightarrow \infty}} ~}\frac{1}{mn}|\{(j,k):j\leq m; k \leq n,(j,k) \in A\}| + \frac{1}{mn}|\{(j,k):j\leq m; k \leq n ,(j,k)\in B \}|= 0 $.
Thus $\{(m,n )\in \mathbb{N\times N} : \frac{1}{mn} |\{(j,k):j\leq m; k \leq n, (j,k)\in A \cup B \}| \geq \delta\} \in I $ for all $\delta>0$.
Let $ K = \{(m, n) \in \mathbb{N\times N} : \frac{1}{mn} |\{(j,k):j\leq m;k \leq n,(j,k )\in A \cup B \}| \geq \frac{1}{2}\} $. Clearly $K\in I$,
so choose $ (m_0, n_0) \in \mathbb{N\times N} \setminus K $. Then $ \frac{1}{m_0n_0}|\{(j,k):j\leq m_0;k \leq n_0,(j, k) \in A \cup B \}| < \frac{1}{2}$.
So $\frac{1}{m_0 n_0} | \{(j,k):j\leq m_0;k \leq n_0,(j, k) \notin A \cup B \}| \geq 1- \frac{1}{2} = \frac{1}{2} $
i.e., $ \{(j, k) :(j, k) \notin A \cup B \}$ is a nonempty set. Take $(j_0, k_0 )\in \mathbb{N\times N}$ such that $(j_0, k_0) \notin A \cup B$. Then $(j_0, k_0) \in A^c \cap B^c$ and hence $ \|x_{j_0 k_0} - y \| < r + \varepsilon $ and $ \| x_{j_0 k_0} - z \| < r + \varepsilon $. So
$\|y - z \| \leq \|x_{j_0 k_0} -y \| + \|x_{j_0k_0} - z \| \leq 2(r + \varepsilon) < \| y - z \|$,
which is absurd. Therefore diam ($I\mbox{-}st\mbox{-}\mbox{LIM}_x^r) \leq 2r $.\\
If $ I\mbox{-}st\mbox{-}\lim x = \xi $, then we proceed as follows.
Let $ \varepsilon > 0 $ and $ \delta > 0 $ be given. Then
$ A = \{(m,n) \in \mathbb{N\times N}: \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \|\geq \varepsilon \}|\geq \delta \} \in I $.
Then for $ (m,n) \notin A $ we have
$\frac{1}{mn}|\{(j,k):j\leq m;k \leq n, \|x_{jk} - \xi \| \geq \varepsilon \} < \delta $,
i.e.,
\begin{eqnarray}
\frac{1}{mn} |\{(j,k):j\leq m;k \leq n, \|x_{jk} - \xi \| < \varepsilon \} \geq 1 - \delta.
\end{eqnarray}
Now for each $y\in {\overline{{B_r}(\xi)}}(=\left\{y\in
X:\left\|y-\xi\right\|\leq r\right\})$ we have
\begin{eqnarray}
\|x_{jk} - y \| \leq \| x_{jk} - \xi \| + \|\xi - y \| \leq \|x_{jk} - \xi \| + r.
\end{eqnarray}
Let $B_{mn }= \left\{ (j,k):j\leq m;k \leq n ,\|x_{jk} - \xi \| < \varepsilon \right\}$. Then for $ (j,k) \in B_{mn}$ we have
$\|x_{jk} -y \| < r + \varepsilon$.
Hence $ B_{mn} \subset \{ (j,k):j\leq m;k \leq n, \|x_{jk} - y \|< r + \varepsilon\}$. This implies,
$ \frac{|B_{mn}|}{mn} \leq \frac{1}{mn}|\{(j,k):j\leq m;k \leq n , \|x_{jk} -y \| < r + \varepsilon\}|$
i.e., $\frac{1}{mn}|\{(j,k):j\leq m;k \leq n,\|x_{jk} - y\|< r + \varepsilon \}| \geq 1 -\delta $.
Thus for all $ (m,n) \notin A $, $ \frac{1}{mn} |(j,k):j\leq m;k \leq n, \|x_{jk} - y \|\geq r + \varepsilon \}|<1-(1 - \delta) $.
Hence we have
$\{(m,n)\in\mathbb{N\times N}: \frac{1}{mn} |\{(j,k):j\leq m;k \leq n,\|x_{jk} - y \| \geq r +\varepsilon \}| \geq \delta\} \subset A $.
Since $ A \in I $, so
$\{(m,n)\in\mathbb{N\times N}: \frac{1}{mn} |\{(j,k):j\leq m;k \leq n,\|x_{jk} - y \| \geq r + \varepsilon \}| \geq \delta\} \in I $.
This shows that $ y \in I\mbox{-}st\mbox{-}LIM_x^r$. Therefore $I\mbox{-}st\mbox{-}\mbox{LIM}_x^r
\supset{\overline{{B_r}(\xi)}} $.
Conversely, let $ y \in I\mbox{-}st\mbox{-}LIM_x^r$. If possible , let $\left\|y-\xi\right\|>r$. Take $\varepsilon=\frac{\left\|y-\xi\right\|-r}{2}$. Let
$ K_1 = \{(j,k) \in \mathbb{N\times N} : \|x_{jk} - y \| \geq r +\varepsilon \}$
and
$ K_2= \{(j,k) \in \mathbb{N\times N} : \|x_{jk} - \xi \| \geq \varepsilon \}$.
Then
\begin{center}
$ \frac{1}{mn} |\{ (j,k):j\leq m;k \leq n,(j,k) \in K_1 \cup K_2\}|
\leq \frac{1}{mn}|\{(j,k):j\leq m; k \leq n,(j,k) \in K_1\}| + \frac{1}{mn}|\{(j,k):j\leq m; k \leq n ,(j,k)\in K_2 \}|$,
\end{center}
and so by the property of $I$-convergence \\
$ I \mbox{-}\lim\limits_{\stackrel{\stackrel{m\rightarrow
\infty}{n\rightarrow \infty}} ~} \frac{1}{mn} |\{ (j,k):j\leq m;k \leq n,(j,k) \in K_1 \cup K_2 \}| \leq
I \mbox{-}\lim\limits_{\stackrel{\stackrel{m\rightarrow
\infty}{n\rightarrow \infty}} ~}\frac{1}{mn}|\{(j,k):j\leq m; k \leq n,(j,k) \in K_1\}| + \frac{1}{mn}|\{(j,k):j\leq m; k \leq n ,(j,k)\in K_2 \}|= 0 $. Let $ K = \{(m, n) \in \mathbb{N\times N} : \frac{1}{mn} |\{(j,k):j\leq m;k \leq n,(j,k )\in K_1 \cup K_2 \}| \geq \frac{1}{2}\} $. Clearly $K\in I$ and we choose $ (m_0, n_0) \in \mathbb{N\times N} \setminus K $. Then $ \frac{1}{m_0n_0}|\{(j,k):j\leq m_0;k \leq n_0,(j, k) \in K_1 \cup K_2\}| < \frac{1}{2}$.
So $\frac{1}{m_0 n_0} | \{(j,k):j\leq m_0;k \leq n_0,(j, k) \notin K_1 \cup K_2 \}| \geq 1- \frac{1}{2} = \frac{1}{2} $
i.e., $ \{(j, k) :(j, k) \notin K_1 \cup K_2 \}$ is a nonempty set. We choose $(j_0, k_0 )\in \mathbb{N\times N}$ such that $(j_0, k_0) \notin K_1 \cup K_2$. Then $(j_0, k_0) \in {K_1}^c \cap {K_2}^c$ and hence $ \|x_{j_0 k_0} - y \| < r + \varepsilon $ and $ \| x_{j_0 k_0} - \xi \| < \varepsilon $. So
\begin{center}
$\|y - \xi \| \leq \|x_{j_0 k_0} -y \| + \|x_{j_0k_0} - \xi \| \leq r +2 \varepsilon < \| y - z \|$,
\end{center}
which is absurd. Therefore $\left\|y-\xi\right\|\leq r$ and so $y\in {\overline{{B_r}(\xi)}}$. Consequently we have $I\mbox{-}st\mbox{-}\mbox{LIM}_x^r
={\overline{{B_r}(\xi)}}$ and this completes the proof.
\end{proof}
\begin{defn}
A double sequence $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}$ is said to be $ I $- statistically bounded if there exists a positive number $ T $ such that for any $ \delta > 0 $ the set $ A = \{(m,n)\in\mathbb{N\times N}: \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n,\|x_{jk}\| \geq T \}| \geq \delta\}\in I $.
\end{defn}
The next result provides a relationship between boundedness and rough $I$-statistical convergence of double sequences.
\begin{thm}
If a double sequence $x = \{ x_{jk}\}_{ j,k \in \mathbb{N}}$ is
bounded then there exists $r \geq 0$ such that $I\mbox{-}st\mbox{-}LIM_x^r \neq
\emptyset$.
\end{thm}
\begin{proof}
Let $x = \{ x_{jk}\}_{ j,k \in \mathbb{N}}$ be a bounded double sequence. Then there exists a positive
real number $T$ such that $\parallel x_{jk} \parallel < T$ for all
$(j,k) \in \mathbb{N} \times \mathbb{N}$. Let $\varepsilon>0$ be given. Then
\begin{center}
$\left\{(j,k):\left\|x_{jk}-0\right\|\geq T+\varepsilon\right\}=\emptyset$.
\end{center}
Therefore $0 \in I\mbox{-}st\mbox{-}LIM_x^T$ and so $I\mbox{-}st\mbox{-}LIM_x^T
\neq \emptyset$.
\end{proof}
\begin{note}
The converse of the above theorem is not true. For example,
let us consider the double sequence $x = \{ x_{jk}\}_{ j,k \in
\mathbb{N}}$ in $\mathbb{R}$ defined by
\begin{eqnarray*}
x_{jk} &=& jk ~~~~, \mbox{if}~ j ~and~ k~ are~ squares \\
&=& 3 ~~~~, ~ otherwise.
\end{eqnarray*}
Then $I\mbox{-}st\mbox{-}LIM_x^0= \{3\} \neq \emptyset$ but the double sequence
$x$ is unbounded.
\end{note}
We now show that the converse of Theorem 3.2 is true if the double sequence is $I$-statistically bounded.
\begin{thm}
A double sequence $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N} $ in $X$ is
$I$-statistically bounded if and only if there exists a non
negative real number $ r $ such that $
I\mbox{-}st\mbox{-}LIM_x^r \neq \emptyset $.
\end{thm}
\begin{proof}
Let $ x=\left\{x_{jk}\right\}_{k\in\mathbb N} $ be an
$I$-statistically bounded double sequence in $X$. Then there exists a
positive real number $T$ such that for $\delta > 0 $ we have $\{ (m,n)
: \frac{1}{mn}|\{(j,k):j\leq m; k \leq n, \|x _{jk} \| \geq T \}| \geq \delta \} \in
I $. Let $ A = \{ (j,k): \|x_{jk} \| \geq T \} $. Then $
I\mbox{-}\underset{ m,n \rightarrow \infty}{\lim}\frac{1}{mn}|\{(j,k) :j\leq m;k
\leq n,(j, k) \in A \}| = 0 $. Let $ r' = sup\{\| x_{jk}\| : (j,k) \in
A^c\}$. Then $ 0\in I\mbox{-}st\mbox{-}LIM^{r'}_x $. Hence $ I\mbox{-}st\mbox{-}LIM^r_x \neq \emptyset
$ for $ r = r'$.
Conversely, let $ I\mbox{-}st\mbox{-}LIM_x^r \neq \emptyset $ for
some $ r \geq 0 $. Let $ \xi \in I\mbox{-}st\mbox{-}LIM_x^r $. Choose
$\varepsilon = \| \xi \| $. Then for each $ \delta
> 0 $, $ \{ (m,n)\in \mathbb{N\times N}: \frac{1}{mn}|\{(j,k):j\leq m;k \leq n,
\|x_{jk} - \xi \| \geq r + \varepsilon \}| \geq \delta \} \in I $.
Now taking $ T = r + 2 \| \xi \|$, we have $ \{ (m,n) \in \mathbb{N\times N}:
\frac{1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk}\| \geq T \}| \geq \delta \} \in
I $. Therefore $ x $ is $I$-statistically bounded.
\end{proof}
Now let $ \{j_p\}_{p \in \mathbb{N}} $ and $ \{k_q\}_{q \in \mathbb{N}}$ be two strictly increasing sequences
of natural numbers. If $ x = \{ x_{jk}\}_{j,k \in \mathbb{N}}$ is a double sequence in $(X, \parallel.\parallel)$,
then we define $\{x_{j_{p}k_{q}}\}_{p,q \in \mathbb{N}}$ as a subsequence of $x$.
\begin{defn}
A subsequence $x^{\prime} = \{x_{j_{p}k_{q}}\}_{p,q \in \mathbb{N}}$ of a double sequence $x = \{x_{jk}\}_{j,k \in \mathbb{N}}$
is called a $I$-dense subsequence if $d_I(\{(j_{p}, k_{q}); p, q \in \mathbb{N} \}) = 1 $.
\end{defn}
In (\cite{Ma1}, Theorem 3.4), Malik et. al. have already shown that if $x^{'} = \{x_{{j_p}{k_q}}\}_{p,q \in \mathbb{N}}$ is a subsequence
of $ x = \{ x_{jk}\}_{j,k \in \mathbb{N}}$, then $LIM_x^r \subset LIM_{x^{'}}^r$. But this result is not true for rough $I$-statistical convergence.
To show this we consider the following example.
\begin{Example}
Consider the double sequence $ x = \{ x_{jk}\}_{j,k \in \mathbb{N}}$ in $\mathbb{R}$ defined by
\begin{eqnarray*}
x_{jk} &=& jk ~~~~, \mbox{if j and k both are cubes} \\
&=& 0 ~~~~, \mbox{otherwise}.
\end{eqnarray*}
Then $x^{\prime} = \{x_{j_{p}k_{q}}\}_{p,q \in \mathbb{N}}$ is a subsequence of $x$, where
$j_{p} = p^{3}, p \in \mathbb{N} $ and $k_{q} = q^{3}, q \in \mathbb{N} $ and for all $ r \geq 0 $,
$I\mbox{-}st\mbox{-}LIM_x^r = [-r,r]$ but $I\mbox{-}st\mbox{-}LIM_{x'}^r = \emptyset $.
\end{Example}
We now present $I$-statistical analogue of Theorem 3.4 \cite{Ma1} in the following form.
\begin{thm}
If $x'=\{x_{{j_p}{k_q}}\}_{p,q\in \mathbb{N}}$ is a $I$-dense subsequence of
$x = \{ x_{jk}\}_{ j,k \in \mathbb{N}}$, then $ I\mbox{-}st\mbox{-}LIM_x^r \subset I\mbox{-}st\mbox{-}LIM_{x'}^r $.
\end{thm}
\begin{proof}
Let $\xi \in I\mbox{-}st\mbox{-}LIM_x^r$. Then for any $\varepsilon > 0$, $d_I(A(\varepsilon)) = 0$, where
$A(\varepsilon) = \{(j,k) \in \mathbb{N} \times \mathbb{N}; \parallel x_{jk} - \xi \parallel \geq r + \varepsilon\}$.
Then $d_I(A^c(\varepsilon)) = 1$.\\
Since $x'=\{x_{{j_p}{k_q}}\}_{p,q\in \mathbb{N}} $ is a $I$-dense subsequence of $x$, so $d_I(K) = 1 $, where $K = \{(j_{p}, k_{q} ); p, q \in \mathbb{N} \}$.
Then $d_I(A^{c} (\varepsilon) \cap K ) = 1 $. \\
Let $A^{'}(\varepsilon) = \{(p,q) \in \mathbb{N} \times \mathbb{N}; \parallel x_{{j_p}{k_q}} - \xi \parallel \geq r + \varepsilon\}$. Now $\{(p,q); \parallel x_{{j_p}{k_q}} - \xi \parallel < r + \varepsilon \} \supset A^{c} (\varepsilon) \cap K $. Therefor $d_I(A^{\prime}(\varepsilon))^{c} = 1 $ and so $d_I(A^{\prime}(\varepsilon)) =0 $-which
implies $ \xi \in I\mbox{-}st\mbox{-}LIM_{x'}^r $.
\end{proof}
\begin{thm}
Let $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}$ be double sequence and $ r \geq 0 $ be a real number. Then the rough $I$-statistical limit set of the double sequence $ x $ i.e., the set
$ I\mbox{-}st\mbox{-}LIM_x^r $ is closed.
\end{thm}
\begin{proof}
If $ I\mbox{-}st\mbox{-}LIM_x^r = \emptyset $, then nothing to prove.\\
Let us assume that $ I\mbox{-}st\mbox{-}LIM_x^r \neq \emptyset $. Now consider a double sequence $\{y_{jk}\}_{j,k \in \mathbb{N}}$
in $ I\mbox{-}st\mbox{-}LIM_x^r $ with $ \lim\limits_{\stackrel{\stackrel{j\rightarrow
\infty}{k\rightarrow \infty}} ~}y_{jk} = y $.
Choose $ \varepsilon > 0 $ and $ \delta > 0 $. Then there exists $ i_{\frac{\varepsilon}{2}} \in \mathbb{N} $ such that $ \|y_{jk} - y \| < \frac{\varepsilon}{2} $ for all $ j > i_{\frac{\varepsilon}{2}}$ and $ k > i_{\frac{\varepsilon}{2}}$.
Let $ j_0 > i_{\frac{\varepsilon}{2}}$ and $ k_0 > i_{\frac{\varepsilon}{2}}$. Then $ y_{j_0k_0} \in I\mbox{-}st\mbox{-}LIM_x^r $.
Consequently, we have
$ A = \{ (m,n)\in\mathbb{N\times N} : \frac {1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk} - y_{j_0k_0}\| \geq r + \frac{\varepsilon}{2}\}| \geq \delta \} \in I $. Clearly $ M = \mathbb{N\times N} \setminus A $ is nonempty, choose $ (m,n) \in M $.
We have
\begin{eqnarray*}
\frac{1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk} - y_{j_0k_0} \| \geq r + \frac{\varepsilon}{2} \}| < \delta.
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk} - y_{j_0k_0} \| < r + \frac{\varepsilon}{2} \}| \geq 1 - \delta.
\end{eqnarray*}
Put $ B_{mn} = \{ (j,k):j\leq m;k \leq n , \|x_{jk} - y_{j_0k_0} \| < r + \frac{\varepsilon}{2} \} $. Choose $(j, k) \in B_{mn} $.
Then
\begin{eqnarray*}
\|x_{jk} - y \| \leq \| x_{jk} - y_{j_0k_0}\| + \|y_{j_0k_0} - y \| < r + \frac{\varepsilon}{2} + \frac{\varepsilon}{2}
= r + \varepsilon.
\end{eqnarray*}
Hence $ B_{mn} \subset \{ (j,k):j\leq m;k \leq n , \|x_{jk} - y \| < r + \varepsilon \}$, which implies
\begin{center}
$ 1-\delta \leq\frac{ |B_{mn}|}{mn} \leq \frac{1}{mn}\left|\{ (j,k):j\leq m;k \leq n , \|x_{jk} - y \| < r + \varepsilon\}\right| $.
\end{center}
Therefore $\frac{1}{mn}|\{ (j,k):j\leq m;k \leq n , \|x_{jk} - y \| \geq r + \varepsilon\}| < 1-(1-\delta)=\delta$.\\
Thus we have
\begin{eqnarray*}
\{(m, n): \frac{1}{mn} | \{ (j,k):j\leq m;k \leq n, \|x_{jk} - y \| \geq r + \varepsilon \}| \geq \delta \} \subset A \in I.
\end{eqnarray*}
This shows that $ y \in I\mbox{-}st\mbox{-}LIM_x^r $. Hence $ I\mbox{-}st\mbox{-}LIM_x^r $ is a closed set.
\end{proof}
\begin{thm}
Let $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}$ be double sequence and $ r \geq 0 $ be a real number. Then the rough $I$-statistical limit set
$ I\mbox{-}st\mbox{-}LIM_x^r $ of the double sequence $ x $ is a convex set.
\end{thm}
\begin{proof}
Let $ y_0, y_1 \in I\mbox{-}st\mbox{-}LIM_x^r $. Let $ \varepsilon > 0 $ be given. Let\\
\begin{eqnarray*}
A_0 = \{ (j,k) \in \mathbb{N\times N} : \| x_{jk} - y_0 \| \geq r + \varepsilon \}~ \\
A_1 = \{ (j,k) \in \mathbb{N\times N} : \| x_{jk} - y_1 \| \geq r + \varepsilon \}.
\end{eqnarray*}
Then by theorem 3.1 for $ \delta > 0 $ we have
$ \{ (m,n) : \frac{1}{mn} |\{(j,k):j\leq m;k \leq n ,(j, k) \in A_0 \cup A_1 \}| \geq \delta \} \in I $.
Choose $ 0 < {\delta}_1 < 1 $ such that $ 0 < 1 - {\delta}_1 < \delta $.
Let $ A = \{ (m,n) : \frac{1}{mn} |\{(j,k):j\leq m;k \leq n ,(j, k) \in A_0 \cup A_1 \}| \geq {1-\delta}_1 \}$.
Then $ A \in I $. Now for all $ (m,n) \notin A $ we have
\begin{eqnarray*}
\frac{1}{mn} |\{(j,k):j\leq m;k \leq n ,(j, k )\in A_0 \cup A_1 \}| < 1 - {\delta}_1
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{1}{mn} |\{(j,k):j\leq m;k \leq n ,(j, k) \notin A_0 \cup A_1 \}| \geq \left\{1-(1 - {\delta}_1)\right\}=\delta_1.
\end{eqnarray*}
Therefore $ \{ (j,k):(j, k) \notin A_0 \cup A_1 \}$ is a nonempty set. Let us take $ (j_0,k_0) \in {A_0}^c \cap {A_1}^c $ and
$ 0 \leq \lambda \leq 1 $. Then
\begin{eqnarray*}
\|x_{j_0k_0} - [(1 - \lambda)y_0 + {\lambda}y_1] \|
& = & \|(1 - \lambda)x_{j_0k_0} + {\lambda}x_{j_0k_0} - [(1 - \lambda)y_0 + {\lambda}y_1] \|\\
& \leq & (1 - \lambda)\|x_{j_0k_0} - y_0 \| + \lambda \|x_{j_0k_0} - y_1\|\\
& < & (1 - \lambda)(r + \varepsilon) + \lambda(r + \varepsilon) = r + \varepsilon.
\end{eqnarray*}
Let $ B = \{ (j,k) \in \mathbb{N\times N}: \| x_{jk} - [(1-\lambda)y_0 + {\lambda}y_1]\|\geq r + \varepsilon\}$.
Then clearly, $ {A_0}^c \cap {A_1}^c \subset B^c $. So for $ (m,n) \notin A $,
\begin{eqnarray*}
{\delta}_1 \leq \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n,(j, k) \notin A_0 \cup A_1 \} \leq \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n ,(j, k) \notin B \}.
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{1}{mn}|\{ (j,k):j\leq m;k \leq n,(j, k) \in B\}| < 1 - {\delta}_1 < \delta.
\end{eqnarray*}
Thus $ A^c \subset \{ (m,n) : \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n,(j, k )\in B\}| < \delta \}$.
Since $ A^c \in F(I) $, so $\{(m, n) : \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n,(j, k) \in B\}| < \delta \} \in F(I)$ and so $\{ (m,n) : \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n,(j, k) \in B\}| \geq \delta \} \in I $.
This completes the proof.
\end{proof}
\begin{thm}
Let $ r > 0 $. Then a double sequence $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}$ is rough $I$-statistically convergent to $ \xi $ if and only if there exists a double sequence $ y = \{y_{jk}\}_{j,k \in \mathbb{N}} $ such that $ {I\mbox{-}st\mbox{-}\lim}~ y = \xi $ and $ \parallel x_{jk} - y_{jk} \parallel \leq r $ for all $(j, k) \in \mathbb{N\times N} $.
\end{thm}
\begin{proof}
Let $ y = \{y_{jk}\}_{j,k \in \mathbb{N}}$ be a double sequence in $X$, which is $I$-statistically convergent to $\xi$ and $\|x_{jk} - y_{jk} \| \leq r $ for all $(j, k) \in \mathbb{N\times N} $. Then for any $ \varepsilon > 0 $ and $ \delta > 0 $ the set
$ A = \{ (m,n)\in\mathbb{N\times N}: \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n , \|y_{jk} - \xi \|\geq \varepsilon \} \geq \delta \} \in I $
. Let $ (m,n) \notin A $. Then we have
\begin{eqnarray*}
\frac{1}{mn} |\{(j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| \geq \varepsilon \}| < \delta
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{1}{mn}|\{ (j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| < \varepsilon \}| \geq 1 - \delta.
\end{eqnarray*}
Let $ B_{mn} = \{ (j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| < \varepsilon \}$. Then for $ (j,k) \in B_{mn} $, we have
\begin{eqnarray*}
\|x_{jk} - \xi \| \leq \| x_{jk} - y_{jk} \| + \|y_{jk} - \xi \| < r + \varepsilon.
\end{eqnarray*}
Therefore,
\begin{eqnarray*}
B_{mn} \subset \{ (j,k):j\leq m;k \leq n, \|x_{jk} - \xi \| < r + \varepsilon \}.
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{|B_{mn}|}{mn} \leq \frac{1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| < r + \varepsilon \}|.
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{1}{mn} |\{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| < r + \varepsilon \}| \geq 1 - \delta.
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{1}{mn} |\{ (j,k):j\leq m;k \leq n , \| x_{jk} - \xi \| \geq r + \varepsilon \}| < 1 - (1 - \delta) = \delta.
\end{eqnarray*}
Thus, $\{ (m,n)\in\mathbb{N\times N}: \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| \geq r + \varepsilon \}| \geq \delta \} \subset A $ and since $ A \in I $, we have $ \{ (m,n)\in\mathbb{N\times N} : \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| \geq r + \varepsilon \}|\geq\delta\} \in I $. Hence $ {I\mbox{-}st\mbox{-}\lim}~ x = \xi $.\\
Conversely, suppose that $ {I\mbox{-}st\mbox{-}\lim}~ x = \xi $. Then for $ \varepsilon > 0 ~~\mbox{and}~~ \delta >0 $,
\begin{eqnarray*}
A = \{ (m,n)\in\mathbb{N\times N} : \frac{1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| \geq r + \varepsilon \}| \geq \delta \} \in I.
\end{eqnarray*}
Let $ (m,n )\notin A $. Then
\begin{eqnarray*}
\frac{1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| \geq r + \varepsilon \}| < \delta.
\end{eqnarray*}
\begin{eqnarray*}
\Rightarrow\frac{1}{mn} | \{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| < r + \varepsilon \}| \geq 1 - \delta .
\end{eqnarray*}
Let $ B_{mn}= \{ (j,k):j\leq m;k \leq n , \|x_{jk} - \xi \| < r + \varepsilon \}$. Now we define a double sequence $y=\{y_{jk}\}_{j,k\in\mathbb N}$ as follows,
\[ y_{jk} = \left\{
\begin{array}{l l}
\xi, & \quad \text{if $\|x_{jk} - \xi \| \leq r $ }\\
x_{jk} + r \frac{\xi - x_{jk}}{\|x_{jk} - \xi \|}, & \quad \text{otherwise.}
\end{array} \right.\]\\
Then
\begin{eqnarray*}
\|y_{jk} -\xi \| & = & \left\{
\begin{array}{l l}
0, & \quad \text{if $\|x_{jk} - \xi \| \leq r $ }\\
\|x_{jk} - \xi + r \frac{\xi - x_{jk}}{\|x_{jk} - \xi \|}\|, & \quad \text{otherwise.}
\end{array} \right.\\
\\
& = & \left\{
\begin{array}{l l}
0, & \quad \text{if $\|x_{jk} - \xi \| \leq r $ }\\
\| x_{jk} - \xi \| - r , & \quad \text{otherwise.}
\end{array} \right.\\
\end{eqnarray*}
Let $ (j,k) \in B_{mn} $.
Then
\begin{eqnarray*}
\|y_{jk} - \xi \| = 0 , ~~\mbox{if}~~\|x_{jk} -\xi \| \leq r
\end{eqnarray*}
and
\begin{eqnarray*}
\|y_{jk} - \xi \| < \varepsilon, ~~\mbox{if}~~ r < \|x_{jk} - \xi \| < r + \varepsilon.
\end{eqnarray*}
Then $ B_{mn} \subset \{ (j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| < \varepsilon \} $. This implies
\begin{center}
$ \frac{|B_{mn}|}{mn} \leq \frac{1}{mn} |\{ (j,k):j\leq m;k \leq n , \|y_{jk} - \xi\| < \varepsilon\} |$.
\end{center}
Hence,
\begin{center}
$ \frac{1}{mn}|\{(j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| < \varepsilon \}| \geq 1 - \delta $
\end{center}
\begin{center}
$ \Rightarrow \frac{1}{mn}|\{(j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| \geq \varepsilon \}| < 1 -(1 - \delta) = \delta $.
\end{center}
Thus $\{(m,n)\in\mathbb{N\times N}: \frac{1}{mn}|\{(j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| \geq \varepsilon\} | \geq \delta \} \subset A $.
Since $ A \in I $, we have
\begin{center}
$ \{ (m,n)\in\mathbb{N\times N}: \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n , \|y_{jk} - \xi \| \geq \varepsilon \}| \geq \delta \} \in I $.
\end{center}
So $ {I\mbox{-}st\mbox{-}\lim}~ y = \xi $.
\end{proof}
\begin{defn}
A point $\lambda \in X $ is said to be an $I$-statistical cluster point of a double sequence $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}$ in $X$ if for any $\varepsilon > 0 $
\begin{eqnarray*}
d_I(\{(j,k):\|x_{jk} - \lambda \| < \varepsilon \}) \neq 0
\end{eqnarray*}
where
$ d_I(A) = I - \lim\limits_{\stackrel{\stackrel{m\rightarrow
\infty}{n\rightarrow \infty}} ~} \frac{1}{mn}|\{ (j,k):j\leq m;k \leq n, (j,k) \in A \}$ if exists.
\end{defn}
The set of $I$-statistical cluster point of $x$ is denoted by ${\Lambda}_x^S(I) $.
\begin{lem}\cite{Ma4}
Let $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}\in\mathbb {R}^n$ be $I$-statistically bounded double sequence. Then for every $\varepsilon>0$ the set
\begin{center}
$\left\{(j,k):d({\Lambda}_x^S(I),x_{jk})\geq\varepsilon\right\}$
\end{center}
has $I$-asymptotic density zero, where $d({\Lambda}_x^S(I),x_{jk})=inf_{y\in{\Lambda}_x^S(I)}\left\|y-x_{jk}\right\|$, the distance from $x_{jk}$ to the set ${\Lambda}_x^S(I)$.
\end{lem}
\begin{thm}
For any arbitrary $ c \in {\Lambda}_x^S(I) $ of a double sequence $ x=\left\{x_{jk}\right\}_{j,k\in\mathbb N}$ we have $ \| \xi - c \| \leq r $ for all $ \xi \in I\mbox{-}st\mbox{-}LIM_x^r $.
\end{thm}
\begin{proof}
Assume that there exists a point $ c \in {\Lambda}_x^S(I) $ and $ \xi \in I\mbox{-}st\mbox{-}LIM^r_x $ such that
$\| \xi - c \| > r $. Let $ \varepsilon = \frac{\| \xi - c \| - r}{3} $. Then
\begin{eqnarray}
\{ (j,k) \in \mathbb{N\times N}: \|x_{jk} - \xi \| \geq r + \varepsilon \} \supset \{ (j,k) \in \mathbb{N\times N}: \|x_{jk} - c \| < \varepsilon \}.
\end{eqnarray}
Since $ c \in {\Lambda}_x^S(I) $ we have $ d_I(\{(j,k):\|x_{jk} - c \| < \varepsilon \}) \neq 0 $.
Hence by (3) we have $ d_I(\{(j,k):\|x_{jk} - c \| \geq r + \varepsilon \}) \neq 0 $, which contradicts that
$ \xi \in I\mbox{-}st\mbox{-}LIM_x^r $. Hence $ \|\xi - c \| \leq r $.
\end{proof}
\begin{thm}
Let $(\mathbb{R}^{n}, \parallel. \parallel)$ be a strictly convex space and $ x = \{x_{jk}\}_{j,k \in \mathbb{N}} $ be a double sequence in this space.
For any $ r > 0 $, let $ y_1, y_2 \in I\mbox{-}st\mbox{-}LIM^r_x $ with $\parallel y_1 - y_2 \parallel = 2r $. Then $x$ is $I$-statistically convergent to $\frac{1}{2}(y_1 + y_2)$.
\end{thm}
\begin{proof}
Let $y_3$ be an arbitrary $I$-statistical cluster point of $x$. Now since $y_1, y_2 \in I\mbox{-}st\mbox{-}LIM^r_x $, so by Theorem 3.9,
\begin{center}
$\parallel y_1 - y_3 \parallel \leq r ~ \mbox {and} ~ \parallel y_2 - y_3 \parallel \leq r $.
\end{center}
Then $ 2r = \parallel y_1 - y_2 \parallel \leq \parallel y_1 - y_3 \parallel + \parallel y_3 - y_2 \parallel \leq 2r $. Therefore $\parallel y_1 - y_3 \parallel = \parallel y_2 - y_3 \parallel = r $. Now
\begin{equation}
\frac{1}{2}(y_1 -y_2) = \frac{1}{2}[(y_3 - y_1) + (y_2 - y_3)]
\end{equation}
Since $\parallel y_1 - y_2 \parallel = 2r $, so $\parallel \frac{1}{2}(y_2 - y_1) \parallel = r $. Again since the space is strictly convex so by (4) we get, $\frac{1}{2}(y_2 - y_1) = y_3 -y_1 = y_2 - y_3 $. Thus $ y_3 = \frac{1}{2}(y_1 + y_2)$ is the unique $I$-statistical cluster point of the double sequence $x$. Again by the given condition $I\mbox{-}st\mbox{-}LIM_x^r \neq \emptyset $ and so by Theorem 3.3, $x$ is $I$-statistically bounded. Since $y_3$ is the unique $I$-statistical cluster point of the $I$-statistically bounded double sequence $x$, so by Lemma 3.8, $x$ is $I$-statistically convergent to $ y_3 = \frac{1}{2}(y_1 + y_2)$.
\end{proof}
\begin{thm}
Let $ x =\{x_{jk} \}_{j,k \in \mathbb{N}}$ be a double sequence in $X$.
(i) If $ c \in {{\Lambda}_x^S}(I)$, then $ I\mbox{-}st\mbox{-}LIM_x^r \subset \overline{{B_r}(c)} $.
(ii) $ I\mbox{-}st\mbox{-}LIM_x^r = \underset{ c \in
{\Lambda}_x^S(I) }{\bigcap}{\overline{{B_r}(c)}} = \{\xi \in X:
{\Lambda}_x^S(I) \subset \overline{B_r(\xi)}\}$.
\end{thm}
\begin{proof}
(i)Let $ c \in {\Lambda}_x^S(I) $. Then by Theorem 3.9, for all $
\xi \in I\mbox{-}st\mbox{-}LIM_x^r $, $ \|\xi - c \| \leq r $ and
hence the result follows.
(ii) By (i) it is clear that $ I\mbox{-}st\mbox{-}LIM_r^x \subset
\underset{ c \in {\Lambda}_x^S(I) }{\bigcap}{\overline{{B_r}(c)}}
$. Now for all $ c \in {\Lambda}_x^S(I) $ and $ y \in \underset{ c
\in {\Lambda}_x^S(I) }{\bigcap}{\overline{{B_r}(c)}}$ we have $\|y
- c \| \leq r $. Then clearly $ \underset{ c \in {\Lambda}_x^S(I)
} {\bigcap}{\overline{{B_r}(c)}} \subset \{ \xi \in X:
{\Lambda}_x^S(I) \subset \overline{{B_r}(\xi)}\} $.
Now, let $ y \notin I\mbox{-}st\mbox{-}LIM_x^r $. Then there
exists an $ \varepsilon > 0 $ such that $ {d_I}(A) \neq 0 $, where
$ A = \{ (m,n) \in \mathbb{N\times N}: \|x_{mn} - y \| \geq r + \varepsilon \} $
. This implies the existence of an $I$-statistical cluster point
$c$ of the sequence $ x $ with $ \|y - c \| \geq r + \varepsilon
$. This gives ${\Lambda}_x^S(I) \nsubseteq \overline{{B_r}(y)}$and
so $ y \notin \{ \xi \in X: {\Lambda}_x^S(I) \subset
\overline{{B_r}(\xi)} \} $. Hence $ \{ \xi \in X: {\Lambda}_x^S(I)
\subset \overline{{B_r}(\xi)} \} \subset I
\mbox{-}st\mbox{-}LIM_x^r $. This completes the proof.
\end{proof}
\noindent\textbf{Acknowledgement:}
The second author is grateful to Government of India for his fellowship funding under UGC-JRF scheme during the preparation of this paper.
\\
|
\section{Introduction}
The ability to achieve controlled, deterministic interactions between photons and atomic media constitutes an important resource in applications ranging from quantum information processing to metrology. As single photons and atoms typically do not interact efficiently, a common approach has been to employ atomic ensembles, where the interaction probability with a given optical mode is enhanced via a large number of atoms~\cite{HSP10}. Atomic ensembles have enabled a number of spectacular proof-of-principle demonstrations of quantum protocols, such as coherent photon storage and quantum memories for light \cite{HHD99,FIM05,HSP10}, entanglement generation between light and atomic spins \cite{JKP01}, nonlinear interactions between photons at the level of individual quanta \cite{PWA13,PFL12,GTS14,TBS14}, and quantum-enhanced metrology \cite{WJK10,LSV10,SKN12}. It has also been proposed that such systems could lead to exotic many-body physics, such as strongly correlated photon ``gases'' \cite{NA16}.
A fundamental limitation in nearly all such possibilities arises from spontaneous emission, wherein photons in a desired optical mode~(\textit{e.g.}, a Gaussian input beam) that facilitate the process are absorbed by the atoms and then re-scattered into other inaccessible modes or channels. Within the context of quantum light-matter interfaces based upon atomic ensembles, it is typically assumed that spontaneous emission occurs independently, and at the same rate given by a single, isolated atom. In that case, the infidelity or error arising from spontaneous emission for a desired process typically decreases with the ``optical depth'' $D$ of the medium as $1/D$ or slower. The optical depth is given by $D\sim (\lambda_0^2/A_\text{eff})N$, where $N$ is the atom number, and $\lambda_0^2/A_\text{eff}$ represents the interaction probability between a single atom and a single photon in the preferred optical mode~($\lambda_0$ being the wavelength associated with the atomic transition and $A_\text{eff}$ the beam area). Intuitively, the $1/D$~(or $1/N$) scaling directly reflects the fact that a given atom is assumed to succeed or fail independently, and that the success is enhanced by the number of atoms involved.
Technically, however, the assumption of independent emission cannot strictly be correct. In particular, as scattering is a wave phenomenon, the emission into other directions may exhibit collective interference. In fact, the possibility that an atomic ensemble can experience a significantly enhanced radiation rate via interference (``superradiance'') was already pointed out in the seminal work of Dicke \cite{D1954}, and has been thoroughly studied for decades \cite{GH1982}. The complementary phenomenon of subradiance, in which photon emission becomes highly suppressed, has also been theoretically studied \cite{PG00,SCS10,JR12,POR15,BGA15,BGA16,FJR16,SR16,ZR10,ZR11,KSP16}, and even observed in recent experiments \cite{DB96,GAK16,SBF17}. Clearly the possibility to enhance atom-light interfaces by suppressing unwanted emission is a tantalizing one, and has started to gain theoretical interest \cite{POR15}. However, finding protocols where subradiance clearly improves the scaling of errors remains an elusive task, in part because techniques to efficiently address subradiant states remain poorly developed.
In this paper, we provide a comprehensive description of subradiance in the case where atoms form ordered arrays. We also present an explicit construction of a protocol exploiting suppressed emission into undesired channels, which enables an exponential improvement in infidelity as a function of atom number over previously known bounds. Our main results are summarized as follows:
\begin{itemize}
\item We first consider infinite 1D or 2D arrays of atoms, which consist of an electronic ground state $\ket{g}$ and excited state $\ket{e}$ that couple to light through a dipole transition. Examining the case of a single collective excitation, we find that a set of perfectly subradiant states with zero decay rate emerge, which can be interpreted as optical ``guided modes.'' Specifically, in exact analogy to guided modes of conventional optical fibers or photonic structures, the spin-wave excitations that constitute these subradiant states have associated wave vectors that are mismatched from free-space radiation fields, which consequently prevents the decay of energy from these states.
\item In the case of a finite array, a set of single-excitation collective atomic modes can exhibit decay rates which are polynomially suppressed with atom number $N$. The finite decay rate can be understood as emerging from scattering of guided excitations into free space through the boundaries of the array.
\item We go beyond the most frequently studied case of subradiance within a single-excitation manifold, and investigate the nature of multi-excitation subradiant modes. Specifically, we show that subradiance is largely destroyed when excitations spatially overlap, as the scattering of two excitations generates many wave vectors that couple to free-space radiation due to the ``hard core'' nature of spins. In 1D arrays, we find that a ``fermionic'' ansatz works well to describe multi-excitation subradiant modes, where these multi-excitation states are constructed from anti-symmetric combinations of single-excitation subradiant modes in order to enforce a spatial repulsion~(\textit{i.e.}, ``Pauli exclusion'') of excitations. These states preserve the same polynomial suppression of decay rate with atom number for any low density of excitations.
\item Having elucidated the salient properties of subradiant states in free-space atomic arrays, we introduce the new concept of ``selectively radiant'' states. In particular, while subradiant states couple weakly to all electromagnetic modes, to realize an efficient atom-light interface it is instead desirable to construct states that are simultaneously superradiant to a preferred photonic mode and subradiant to all the others. We show that one natural way to achieve such a scenario is by coupling an atomic array to the guided modes of a nanophotonic structure, such as an optical nanofiber \cite{VRS10,GCA12,GSO15,polzik,CGC16}. As the wave vectors of the guided modes of the structure itself are mismatched from free-space radiation, it becomes possible for a set of atomic spin waves to efficiently couple to these guided modes, while retaining a suppressed coupling to free-space modes. We analyze the specific protocol of photon storage \cite{L03,FIM05} using an atomic array coupled to a nanofiber \cite{SCA15,GMN15}, and find numerically a storage infidelity that is exponentially small in the atom number or optical depth, $\sim \exp(-D)$. This scaling represents an exponential improvement over the best previously established error bound of $\sim 1/D$~\cite{GAF07,GAL07}, derived assuming that emission into undesired modes is independent.\end{itemize}
This article is structured as follows. In Sec. II we begin by introducing a theoretical framework for atom-light interactions that does not invoke the typical assumption of independent atomic emission, and instead formally and exactly describes collective emission and interactions of atoms via photon fields. In Sec. III we apply this formalism to investigate single- and multi-excitation subradiant states in atomic arrays, with the main results having already been summarized above. In Sec. IV we present the idea of selectively radiant states, and analyze the efficiency of a quantum memory consisting of a chain of atoms close to a nanofiber. Finally, in Sec. V we discuss possible implementations and other photonic platforms for observing subradiant physics. An outlook is provided in Sec. VI.
\section{Spin model}
Here we introduce a theoretical formalism to describe the fully quantum interaction between atoms and radiation fields, which is valid in the presence of any linear, isotropic, dielectric media. This rather general formalism will enable us to equally treat the case of atomic arrays in free space (Sec. III), or interacting via the guided modes of an optical fiber (Sec. IV). In particular, we present a model in which the field is integrated out and the dynamics of the atomic internal (``spin") degrees of freedom follow a master equation that only depends on atomic operators. Moreover, once the time evolution of the atoms is solved for, one can recover the field at any point in space by means of a generalized input-output equation.
The first step to describe how atoms couple to radiation is to quantize the electromagnetic field. The traditional approach involves explicitly finding a normal mode decomposition of the fields, and associating bosonic annihilation and creation operators to each mode. This is well-suited to cases where a limited number of modes are assumed to be relevant (such as a high-Q cavity). In our case, though, as we want to exactly capture collective effects in spontaneous emission involving all modes, such an approach becomes unwieldy (as in free space) \cite{GH1982,BS06} or impossible, such as for complex dielectric structures. We require a more general technique that allows us to treat these situations. Such a framework was developed by Welsch and coworkers \cite{GW96,DKW02,BW07,BI12}, and is based on the classical electromagnetic Green's function (or Green's tensor).
The Green's function ${\bf G}(\rb,\rb',\omega)$ is the fundamental solution of the electromagnetic wave equation, and obeys \cite{NH06}:
\begin{align}\label{gf}
\bm{\nabla}\times\bm{\nabla}\times {\bf G}(\rb,\rb',\omega)-\frac{\omega^2}{c^2}\epsilon(\rb,\omega)\, {\bf G}(\rb,\rb',\omega)=\delta(\rb-\rb')\mathbb{1},
\end{align}
where $\epsilon(\rb,\omega)$ is the position- and possibly frequency-dependent relative permittivity of the medium. The Green's function physically describes the field at point $\rb$ due to a normalized, oscillating dipole at $\rb'$. ${\bf G}_{\alpha\beta}$ is a tensor quantity ($\{\alpha,\beta\}=\{x,y,z\}$), as $\alpha$ and $\beta$ refer to the possible orientations of the field and dipole, respectively. Here, we will deal with cases where the Green's function can be solved analytically, but for more complex structures it is also possible to obtain it numerically ~\cite{ORI10,RH07,HGA16}. In the following, we introduce a prescription of how to write down an equation that relates the field and the atomic coherence operators, built upon the intuition provided by classical physics. For a more formal derivation of the field quantization, we refer the reader to Refs.~\cite{GW96,DKW02,BW07,BI12,AHC17}.
In the frequency domain, the analogous classical problem that one would like to solve is to find the total field $\Eb(\rb,\omega)$ at point $\rb$, given a known input field $\Eb_{\rm p}(\rb,\omega)$ and a collection of $N$ polarizable dipoles $\pb_j(\omega)$ located at $\rb_j$, which are excited by the fields and re-scatter light themselves. The values of $\pb_j(\omega)$ are not known a priori, since they depend on the polarizability and the total field at $\rb_j$ [solving for $\pb_j(\omega)$ will be discussed in following steps]. As the field at any given point in space is just the sum of the external or driving field and the field re-scattered by the dipoles, we find $\Eb(\rb,\omega)=\Eb_{\rm p}(\rb,\omega)+\mu_0\omega^2\,\sum_{j=1}^N{\bf G}(\rb,\rb_j,\omega)\cdot\pb_j(\omega)$, where $\mu_0$ is the vacuum permeability.
The question is how to translate this classical equation into an equation for quantum operators. In fact, the quantum nature of the field is inherited from the quantum properties (\textit{e.g.}, correlations and fluctuations) of the sources, while the field propagation remains the same as both the quantum and classical fields obey Maxwell's equations. Therefore, the above equation is valid for quantum fields, but replacing $\pb_j(\omega)$ by the dipole moment operator $\hat{\pb}_j(\omega)$, and $\Eb(\rb,\omega)$ by the field operator $\hat{\Eb}^+(\rb,\omega)$, where the superscript refers to the positive-frequency component. In the case that the quantum dipoles are atoms, one can make a further approximation, taking advantage of the fact that an atom only has a significant optical response in a narrow bandwidth around its resonance frequency $\omega_0$. Thus, one is able to approximate ${\bf G}(\rb,\rb_j,\omega)$ by ${\bf G}(\rb,\rb_j,\omega_0)$, which allows the Fourier transform of the equation to become local in time. Then, one arrives to the generalized input-output equation in the time domain, which reads \cite{DKW02,CMS15,XF15}
\begin{align}\label{fielddef}
\hat{\Eb}^+(\rb)&=\hat{\Eb}_{\rm p}^+(\rb)+\mu_0 \omega_0^2\sum_{j=1}^N {\bf G}(\rb,\rb_j,\omega_0)\cdot\db\,\hge^j.
\end{align}
To obtain the above expression, we have made use of the fact that $\hat{\pb}_j = \db^* \, \hat{\sigma}^j_{eg} + \db \, \hat{\sigma}^j_{ge}$, where $\hat{\sigma}_{eg}^j=\ket{e_j}\bra{g_j}$ is the atomic coherence operator between the ground and excited states of atom $j$, and $\db$ is the dipole matrix element associated with that transition. This equation is valid in the Markovian regime, where the dispersion in the Green's function can be neglected and the replacement of ${\bf G}(\rb,\rb_j,\omega)$ by ${\bf G}(\rb,\rb_j,\omega_0)$ is well-founded. For this to be true, two conditions have to be fulfilled. First, the retardation arising from the physical distance between atoms can be ignored \cite{CJG12,SCC15}. For atoms in free space, this means that they should sit much closer than the length of a spontaneously emitted photon ($\lesssim 1$ meter). Second, the electromagnetic environment itself should not have very narrow-bandwidth features (\textit{e.g}., one must avoid the strong coupling regime of cavity QED \cite{TRK92}).
What remains now is to solve for the dipoles (in this case, $\hge$) themselves. We do so by writing down Heisenberg-Langevin equations for the atomic internal degrees of freedom, starting from the full atom-field Hamiltonian. Intuitively, the atomic spin $\hat{\sigma}_{eg}^{i}$ will be driven by the quantum field at position $\rb_i$. However, as the field itself only depends on other atoms via the input-output equation, the atomic dynamics can be fully derived from an equivalent master equation of the form $\dot{\hat{\rho}}_{\rm A}=-(\ii/\hbar)\,[\mathcal{H},\hat{\rho}_{\rm A}]+\mathcal{L}[\hat{\rho}_{\rm A}]$ \cite{MS1990}, where $\hat{\rho}_{\rm A}$ is the atomic density matrix, and the Hamiltonian and Lindblad operators read
\begin{subequations}
\begin{equation}\label{ham}
\mathcal{H}=\hbar\omega_0\sum_{i=1}^N\hat{\sigma}_{ee}^i+\hbar\sum_{i,j=1}^N J^{ij}\hat{\sigma}_{eg}^i\hat{\sigma}_{ge}^j,
\end{equation}
\begin{equation}\label{lind}
\mathcal{L}[\hat{\rho}_{\rm A}]=\sum_{i,j=1}^N\frac{\Gamma_{ij}}{2}\,\left(2\hat{\sigma}_{ge}^j\hat{\rho}_{\rm A}\hat{\sigma}_{eg}^i-\hat{\sigma}_{eg}^i\hat{\sigma}_{ge}^j\hat{\rho}_{\rm A}-\hat{\rho}_{\rm A}\hat{\sigma}_{eg}^i\hat{\sigma}_{ge}^j\right).
\end{equation}
\end{subequations}
In the above expressions, the rates for coherent and dissipative interactions between atoms $i$ and $j$ are respectively given by
\begin{subequations}\label{rates}
\begin{align}\label{shiftrate}
J^{ij}&=-\frac{\mu_0\omega_0^2}{\hbar}\,\db^*\cdot\text{Re}\,\mathbf{G}(\rb_i,\rb_j,\omega_0)\cdot\db,
\end{align}
\begin{equation}\label{dissiprate}
\Gamma^{ij} =\frac{2\mu_0\,\omega_0^2}{\hbar}\,\db^*\cdot\text{Im}\,\mathbf{G}(\rb_i,\rb_j,\omega_0)\cdot\db,
\end{equation}
\end{subequations}
where the sign of $J^{ij}$ is taken to be opposite to that of Refs.~\cite{GW96,DKW02,BW07,BI12,HGA16,AHC17}. In the above Hamiltonian, we have neglected Casimir interactions between ground-state atoms (of the form $\hat{\sigma}_{gg}^i\hat{\sigma}_{gg}^j$), as their spatial decay is very fast ($\sim 1/d^6$ in free space, $d$ being the inter-atomic distance) \cite{BI12}.
The dynamics under the master equation can analogously be described in the quantum jump formalism of open systems \cite{MS07}. In this formalism, the atomic wave function evolves deterministically under an effective non-Hermitian Hamiltonian that reads $\mathcal{H}=\hbar\omega_0\sum_{i=1}^N\hat{\sigma}_{ee}^i+\mathcal{H}_{\rm eff}$, with
\begin{equation}\label{heff}
\mathcal{H}_{\rm eff}=-\mu_0\omega_0^2\sum_{i,j=1}^N \,\db^\ast\cdot\mathbf{G}(\rb_i,\rb_j,\omega_0)\cdot\db\,\,\hat{\sigma}_{eg}^i\hat{\sigma}_{ge}^j,
\end{equation}
along with stochastically applied ``quantum jump" operators to account for the population recycling term ($\hge^{j}\hat{\rho}_\text{A}\hat{\sigma}_{eg}^{i}$) of Eq.~\eqref{lind}. While $\mathcal{H}_{\rm eff}$ just describes the interaction of atoms through emission and re-absorption of photons, one can directly add other terms to the Hamiltonian to account for external driving fields.
To conclude, we point out that although the full formalism above has only been rigorously and generally developed in recent years, many aspects have long been used within atomic physics and quantum optics. For example, for a single atom or other quantum emitter, the spin model becomes trivial and just yields the total spontaneous emission rate. Thus, the calculation of enhancement of spontaneous emission near dielectric structures is standardly reduced to the calculation of the Green's function \cite{CPS1978,LVN04,EFW05}. Alternatively, such equations are often used to model the optical response of dense three-dimensional atomic gases \cite{SSH16,JBS16,JSG16,BZB16,GK17}.
\section{Free space: Subradiant states}
\label{SecIII}
We now apply the spin model we describe in the previous section to investigate the properties of subradiant states associated with ordered atomic arrays in free space. Recently, the peculiar linear optical properties of periodic atomic arrays have started to attract interest \cite{BGA15,BGA16,BGA16b,SWL17,KSP16,SR16,ZR11,POR15,HFS16,JR12,FJR16}. This includes the identification of guided modes supported by infinite arrays \cite{ZR10,SWL17,SR16}, and states with very long lifetimes in finite arrays \cite{BGA15,BGA16,BGA16b,KSP16,SR16,ZR11,POR15,JR12,HFS16,FJR16}. Here, we provide a clear and intuitive connection between the existence of guided modes in infinite arrays and subradiant states in a finite system. We provide conditions for the lattice constants in 1D and 2D that enable single-excitation guided Bloch modes with zero decay rate to emerge, which are decoupled from free-space radiation due to wave vector mismatch. We then analyze a single excitation in a finite lattice, and show how the guided modes acquire a non-zero decay rate due to scattering into electromagnetic radiation at the system boundaries. We also analyze the scaling of the decay rates with system size and elucidate the spatial structure of subradiant states. Finally, we go beyond previous studies of single-excitation subradiance (where the atoms can equivalently be treated as classical dipoles) to the rich physics of the multi-excitation case. In particular, in one dimension, we show that multi-excitation subradiant states exist for any low density of excitations, and that their wave functions have fermionic character.
The atoms are assumed to be tightly trapped, so that we can treat the positions of the particles as classical points rather than dynamical variables. In this situation, we substitute in Eqs.(\ref{fielddef}-\ref{heff}) the free-space Green's tensor ${\bf G}(\rb_i,\rb_j,\omega_0)={\bf G}_{0}(\rb_{ij},\omega_0)$, with $\rb_{ij}=\rb_i-\rb_j$. Here ${\bf G}_{0}(\rb,\omega_0)$ is the solution to Eq.(\ref{gf}) when setting $\epsilon(\rb,\omega)=1$, and can be written as:
\begin{align}
{\bf G}_0 (\rb,\omega_0) = \frac{e^{\ii k_0 r}}
{4\pi k_0^2 r^3} \left[(k_0^2 r^2+\ii k_0 r -1) \mathbb{1} \right.+ \notag \\
\qquad{}\left. + (-k_0^2 r^2 -3\ii k_0 r + 3) \frac{\rb \otimes \rb}{r^2} \right], \label{Greens_def}
\end{align}
where $r=|\rb|$ and $k_0=2\pi/\lambda_0=\omega_0/c$ is the wave number corresponding to the atomic transition energy. For a single atom, evaluating Eq.~\eqref{dissiprate} simply reproduces the well-known vacuum emission rate $\Gamma^{ii}=\Gamma_0$, where $\Gamma_0=\omega_0^3|\db|^2/3\pi\hbar\epsilon_0 c^3$. The single-atom energy shift $J^{ii}$ in Eq.\eqref{shiftrate} arising from ${\bf G}_0$ formally yields a divergence and will be set to zero in what follows, as it should be incorporated into a re-normalized resonance frequency $\omega_0$. In Sec. IV, the Green's function of a nanofiber will be decomposed into a free-space and a scattered component, ${\bf G}={\bf G}_0+{\bf G}_\text{sc}$, where ${\bf G}_\text{sc}$ does produce a finite, observable contribution to $J^{ii}$.
For concreteness, we will restrict ourselves to the following lattice geometries: one-dimensional (1D) linear chains and closed circular rings, two-dimensional (2D) square and three-dimensional (3D) cubic lattices. However, it should become clear that the underlying principles should be general to other lattice structures as well. In the following, the number of atoms and lattice constant are denoted by $N$ and $d$, respectively.
\subsection{Infinite Lattice (Single excitation)}
\label{SecIIIA}
\begin{figure}[h!]
\centerline{\includegraphics[width=\linewidth]{fsFig1}}
\caption{{\bf (a)} Generic dispersion relation of frequency $\omega(k_z)$ versus Bloch wave vector $k_z$ for single-excitation modes of an infinite, one-dimensional chain. The Bloch vector $k_z$ is only uniquely defined within the first Brillouin zone $|k_z|\leq \pi/d$. The dashed black line is the light line, and corresponds to the dispersion relation of light in vacuum propagating along the $\hat{z}$ direction, \textit{i.e.}, $\omega = c |k_z|$. Atomic modes in the region enclosed within the light line (shaded) are generally unguided and radiate into free space. Outside the light line ($|k_z| > \omega /c$) the modes are guided and subradiant, as the electromagnetic field is evanescent in the directions transverse to the chain. The dispersion relation is generally expected to be rather flat and centered around the bare atomic resonance frequency $\omega_0$. {\bf (b)} Collective frequency shifts and {\bf (c)} decay rates for an atomic chain along $\hat{z}$ with lattice constant $d/\lambda_0 = 0.2$, for parallel (blue) and transverse (red) atomic polarization. Circles correspond to the results for a finite system with $N=50$ atoms. The analytical expressions for the infinite chain are denoted by solid lines and approximate well the finite chain results, except for a small region close to the light line. In the infinite lattice case the modes with $|k_z|>\omega_0/c$ are perfectly guided and the decay rate $\Gamma_{k_z}$ is exactly zero. The light line (black dashed) appears vertical over the very narrow frequency ranges plotted here.}
\label{fsFig1}
\end{figure}
Let us consider first a perfectly ordered infinite array of atoms. Despite the infinite lattice being unrealistic, it provides insight into the problem thanks to its mathematical simplicity. In this case, the system is perfectly translationally invariant by any lattice vector displacement, and thus, both atomic and electromagnetic eigenmodes must obey Bloch's theorem.
For a single excitation stored in the system, the eigenstates of the effective atomic Hamiltonian of Eq.(\ref{heff}) are spin waves, with well-defined quasi-momentum $\textbf{k}$, which can always be chosen to be within the first Brillouin zone. For such states, whose creation operators can be written as $S^\dagger_{\textbf{k}}= N^{-1/2} \sum_j e^{\ii \textbf{k} \cdot \rb_j} \hat{\sigma}_{eg}^j$, the single excitation is delocalized and shared in a coherent way among all the atoms. Classically, these states are analogous to oscillating dipoles where the phase of dipole $j$ is given by $e^{\ii \textbf{k} \cdot \rb_j}$.
As the Bloch modes are eigenstates of the effective Hamiltonian, they satisfy $\mathcal{H}_{\rm eff} \,S^\dagger_\textbf{k} \ket{g}^{\otimes N}=\hbar(J_\textbf{k} - \ii \Gamma_\textbf{k}/2) \,S_\textbf{k}^\dagger \ket{g}^{\otimes N}$. Here $J_\textbf{k}$ and $\Gamma_\textbf{k}$ are real quantities and can be identified as the frequency shift of mode $\textbf{k}$ (relative to the bare atomic frequency $\omega_0$) and the decay rate, respectively. One can readily show that, in terms of the single-atom spontaneous emission rate $\Gamma_0$, they are given by:
\begin{subequations}
\begin{align}
\frac{J_\textbf{k}}{\Gamma_0} &=- \frac{3\pi}{k_0} \dbu^* \cdot \text{Re} \,\tilde{{\bf G}}_{0} (\textbf{k}) \cdot \dbu,
\end{align}
\begin{align}
\frac{\Gamma_\textbf{k}}{\Gamma_0} &= \frac{6\pi}{k_0} \dbu^*\cdot \text{Im} \,\tilde{{\bf G}}_{0} (\textbf{k}) \cdot \dbu, \label{Eq:Couplings}
\end{align}
\end{subequations}
where $\tilde{{\bf G}}_{0} (\textbf{k}) = \sum_{j} e^{-\ii \textbf{k} \cdot \rb_j} {\bf G}_{0} (\rb_j)$ is the discrete Fourier transform of the free-space Green's tensor.
We will now show that, when the atoms are placed at close enough distances, dipole-dipole interactions can dramatically modify the decay rates of collective states. As the simplest case, let us consider an infinite one-dimensional chain of atoms first, oriented along the $\hat{z}$ direction. In that case, the wave vector $k_z$ constitutes an index for the modes, and one can consider the dispersion relation of frequency $\omega(k_z)=\omega_0+J_{k_z}$ versus $k_z$. For a periodic structure, regardless of the system details, one expects the dispersion relation to exhibit general characteristics [see Fig.~\ref{fsFig1}(a)]. First, and as mentioned before, $k_z$ is only uniquely defined within the first Brillouin zone ($|k_z|\leq \pi/d$) and thus, it suffices to plot the dispersion relation in that region. Second, it is helpful to draw the ``light line'', \textit{i.e.}, the dispersion relation $\omega=c|k_z|$ corresponding to light propagating in free space along the $\hat{z}$ direction [dashed line of Fig.~\ref{fsFig1}(a)]. Physically, the light line is significant because it separates states of very different character, as we now describe.
To see this, let's consider the field generated by a spin-wave excitation, which is given by Eq.(\ref{fielddef}) under the replacement $\hge^{j}\rightarrow e^{\ii k_z z_j}$ (it is sufficient to consider the limit of classical dipoles for this argument). One can always expand the field $\Eb(\rb)$ in terms of plane wave components, $\Eb(\rb)=\sum_{q_z,{\bf q_\perp}} \Eb_{q_z,{\bf q_\perp}} e^{\ii q_z z + \ii{\bf q_\perp} \cdot{\bf r_\perp}}$. The state is clearly of Bloch's form, and thus, only a discrete set of wave vectors $q_z = k_z+g_z$ ($g_z$ being any reciprocal lattice vector) will contribute. At the same time, the wave equation requires that the axial and perpendicular components of the wave vector satisfy $(q_z)^2+{\bf q}_\perp \cdot {\bf q}_\perp = (\omega/c)^2$. Thus, one can readily verify that a spin wave outside the light line ($|k_z|>\omega/c$) has an associated electromagnetic field composed of axial wave vectors $|q_z|>\omega/c$. This in turn implies that ${\bf q_\perp}$ is imaginary, and the field is guided and decays evanescently away from the structure. Therefore, these guided modes are decoupled from all optical modes propagating in free space, and their inability to radiate away energy leads to perfect subradiance (exactly zero decay rate). Conversely, modes within the light line are generally unguided and can radiate energy out to infinity.
The concepts outlined above regarding the separation of the dispersion relation into guided and radiative regions are actually quite general, and well-known in the context of periodically modulated dielectric waveguides (``photonic crystals'' \cite{JMW95}). An atomic chain might appear quite different physically, but mathematically the same set of principles apply. Furthermore, while it is difficult to prove independent of lattice geometry and atomic level configuration, one would generally expect that for atoms any guided modes would occur within a narrow bandwidth (on the order of the atomic transition linewidth $\Gamma_0/2\pi \lesssim 10 {\rm~MHz}$) around the resonance frequency ($\omega_0/2\pi \sim 300 {\rm ~THz}$), where the atoms have a significant optical response. Thus, in Fig.~\ref{fsFig1}(a) the band structure will appear rather flat. Then, a sufficient condition for guided modes to exist in an atomic chain is essentially that the light line intersects the edge of the Brillouin zone $k_z = \pi/d$ at a frequency $\omega(k_z)$ greater than the atomic resonance. This condition can be rewritten as a condition on the lattice constant $d<\lambda_0/2$ required to support guided modes.\\
Equipped with this general intuition, we now quantitatively investigate the dispersion relation for the 1D infinite chain of two-level atoms with polarization parallel or transverse to the array. The collective frequency shifts are derived in greater detail in Appendix~\ref{AppA}, and read:
\begin{subequations}
\begin{align}
\frac{J_{k_z}^{||}}{\Gamma_0} &=-\frac{3}{2 k_0^3 d^3}\textrm{Re} \left[ \textrm{Li}_{3}(e^{\ii (k_0+k_z)d})+ \textrm{Li}_{3}(e^{\ii (k_0-k_z)d}) \right. \notag\\
&\quad{}\left. -\ii k_0 d \textrm{Li}_{2}(e^{\ii (k_0+k_z)d})-\ii k_0 d \textrm{Li}_{2}(e^{\ii (k_0-k_z)d}) \right], \label{Eq:ShiftAnal1}
\end{align}
\begin{align}
\frac{J_{k_z}^{\perp}}{\Gamma_0} &=\frac{3}{4 k_0^3 d^3} \textrm{Re} \left[\textrm{Li}_{3}(e^{\ii (k_0+k_z)d})+ \textrm{Li}_{3}(e^{\ii (k_0-k_z)d}) \right.\notag\\
&\quad{}-\ii k_0 d \textrm{Li}_{2}(e^{\ii (k_0+k_z)d})- \ii k_0 d \textrm{Li}_{2}(e^{\ii (k_0-k_z)d}) \notag\\
&\quad{} \left. +k_0^2 d^2 \textrm{Ln}(1-e^{\ii(k_0+k_z)d}) +k_0^2 d^2 \textrm{Ln}(1-e^{\ii(k_0-k_z)d}) \right], \label{Eq:ShiftAnal2}
\end{align}
\end{subequations}
where $\textrm{Li}_{n}(x)$ is the PolyLogarithm of order $n$. These expressions are plotted in Fig.~\ref{fsFig1}(b), for the particular value of $d/\lambda_0=0.2$. Here, the light line is indicated as before by a dashed line, but since $\Gamma_0/ \omega_0 \sim 10^{-8}$, it appears essentially as a vertical line.
As anticipated, we can see in this figure that the bands occupy only a narrow bandwidth around the resonance frequency, except close to the light line for transverse polarization. The exact shape of the bands depends on the value of $d/\lambda_0$ and the polarization direction, and for instance, the effective mass at the zone edge ($|k_z|=\pi/d$) is negative (positive) for parallel (transverse) polarization. Exactly at the light line the expression for $J_{k_z}^{\perp}$ ($J_{k_z}^{||}$) becomes non-analytic and diverges (has a derivative that diverges).
\begin{figure}
\centerline{\includegraphics[width=\linewidth]{fsFig2}}
\caption{{\bf (a)} Illustration of a square lattice of atoms in the $\hat{y}$-$\hat{z}$ plane, with lattice constant $d$. {\bf (b)} Corresponding reciprocal lattice in 2D, with lattice constant $2\pi/d$. The collective modes have well defined quasi-momentum ${\textbf{k}}=(k_y,k_z)$ within the first Brillouin zone, which is indicated by the blue square. A circle of radius $| \textbf{k} |=k_0$ defines the set of propagating electromagnetic modes in vacuum in the $\hat{y}$-$\hat{z}$ plane at the atomic frequency. Collective spin waves outside of this circle will be guided, with a decay rate $\Gamma_{\textbf{k}}=0$. For $k_0>k^\textrm{max} = \sqrt{2} \pi / d$ (or equivalently, $d/\lambda_0> 1/\sqrt{2}$) all collective eigenstates lie inside of the circle. {\bf (c)} and {\bf (d)} Collective decay rates for the infinite square lattice for parallel and transverse atomic linear polarization, respectively. The lattice constant is set to $d/\lambda_0=0.2$. For transverse polarization, $\Gamma_{\textbf{k}}$ also vanishes for $\textbf{k} \sim (0,0)$, provided that $d/\lambda_0<1$.} \label{fsFig2}
\end{figure}
The collective decay rates can also be analytically derived (see Appendix~\ref{AppA}):
\begin{subequations}
\begin{align}
\frac{\Gamma_{k_z}^{||}}{\Gamma_0} &=\frac{3\pi}{2 k_0 d} \sum_{\substack{g_z \\ |k_z+g_z|\leq k_0}}\left(1-\frac{(k_z+g_z)^2}{k_0^2} \right),
\end{align}
\begin{align}
\frac{\Gamma_{k_z}^{\perp}}{\Gamma_0} &=\frac{3\pi}{4 k_0 d} \sum_{\substack{g_z \\ |k_z+g_z|\leq k_0}} \left(1+\frac{(k_z+g_z)^2}{k_0^2} \right).
\end{align}
\end{subequations}
These summations run over reciprocal lattice vectors that satisfy $|g_z +k_z| \leq k_0$. That is, only the diffracted waves enclosed within the light line will contribute to the decay rate. When $|k_z|>k_0$, there are no values of $g_z$ satisfying the above condition. Thus the decay rates are zero and we mathematically recover the result previously anticipated -- modes beyond the light line are perfectly guided without radiative losses. The decay rates are plotted in Fig.~\ref{fsFig1}(c). As we can see from the expressions above, at the light line the state can be subradiant or radiant depending on the polarization direction. This results in a discontinuity at the light line for transverse polarization.\\
A similar set of results can be obtained for a 2D array. Considering a square lattice in the $\hat{y}$-$\hat{z}$ plane [Fig.~\ref{fsFig2}(a)], the corresponding first Brillouin zone for Bloch wave vectors extends over the region $|k_y|,|k_z| \leq \pi/d$, as shown in Fig.~\ref{fsFig2}(b). The set of electromagnetic fields propagating in the plane at the atomic frequency $\omega_0$ have a wave vector of magnitude $k_0$ which defines a circle centered around the origin in $\textbf{k}$-space, as illustrated in Fig.~\ref{fsFig2}(b). Similar to 1D, a sufficient condition for spin-wave excitations to be guided is that the wave vector lies outside of this circle. It should be noted that the longest ``distance'' in the first Brillouin zone from the origin extends along the diagonal, and has magnitude $k^\text{max} = \sqrt{2}\pi/ d $. Thus, in 2D, guided modes exist as long as $k_0<k^\text{max}$, which translates into a maximum allowed lattice constant $d/\lambda_0 = 1/\sqrt{2}$.
Analogous to the 1D case, we can obtain closed mathematical expressions for the decay rates in the 2D lattice. They are given by:
\begin{subequations}\label{rate2D}
\begin{align}
\frac{\Gamma_{\textbf{k}}^{||}}{\Gamma_0} &=\frac{3\pi}{ k_0^3 d^2} \sum_{\substack{\gb \\ |\textbf{k}+\gb| \leq k_0}} \frac{k_0^2 - |(\textbf{k}+ \gb)\cdot \dbu|^2}{ \sqrt{k_0^2 -|\textbf{k} + \gb|^2}}, \label{Eq:Decay2D_parallel}
\end{align}
\begin{align}
\frac{\Gamma_{\textbf{k}}^{\perp}}{\Gamma_0} &=\frac{3\pi}{k_0^3 d^2} \sum_{\substack{\gb \\ |\textbf{k}+\gb| \leq k_0}} \frac{|\textbf{k} + \gb|^2}{ \sqrt{k_0^2 -|\textbf{k} + \gb|^2}}, \label{Eq:Decay2D_trans}
\end{align}
\end{subequations}
from which we recover again the important result that Bloch states with $|\textbf{k}|>k_0$ do not radiate out to infinity. For these states, the electromagnetic field is now confined within the plane, and evanescently decays away from the lattice in the transverse direction. The decay rates are plotted in Fig.~\ref{fsFig2}(c)-(d) for atomic polarizations along $\hat{z}$ and $\hat{x}$ directions, for the particular value of $d/\lambda_0=0.2$, which defines the light line as the circle $k_0=0.4\pi/d$, beyond which the decay rate is exactly zero.
We would like to remark that the previous considerations are valid regardless of the specific atomic structure, provided that the atom in question only contains a single ground state (see Sec.~V for a discussion of the subtleties associated with a ground-state manifold).
The previous analysis for the 2D lattice provides another interesting result. As Fig.~\ref{fsFig2}(d) shows, for transverse atomic polarization in a 2D square lattice, subradiance can emerge not only outside the light line, but also at the center of the Brillouin zone, that is, for Bloch states with quasi-momentum $\textbf{k} \sim (0,0)$. Physically, the origin of this effect can be understood as follows. On one hand, and as previously discussed, for $d/\lambda_0 < 1$ the field created by such a state is generally evanescent at all diffraction orders except for the component $\gb = 0$ [c.f. Eq.(\ref{Eq:Decay2D_trans})], which corresponds to a plane wave propagating perpendicularly to the atomic plane. On the other hand, the state $\textbf{k} = (0,0)$ corresponds to an array of dipoles that are in phase. However, dipoles oscillating in phase and perpendicularly to the atomic plane are forbidden to radiate energy in the perpendicular direction, and thus, the state must be subradiant. In contrast, as soon as $d/\lambda_0 > 1$, there will be other $\gb$ components that are not evanescent, yielding a radiative state. We note that, although only the state $\textbf{k} = (0,0)$ has exactly zero decay rate, other modes around this point will also show a strong suppression in the emission rate relative to $\Gamma_0$.\\
\begin{figure*}
\centerline{\includegraphics[width=\linewidth]{fsFig3.pdf}}
\caption{Single-excitation collective modes in a finite 1D chain of two-level atoms with polarization along the chain. {\bf (a)} Decay rates of the $N$ modes at different lattice constants $d/\lambda_0$ for a finite chain of $N=50$ atoms. Subradiant modes only arise if $d/\lambda_0\leq 1/2$. A vertical cut at the fixed value of $d/\lambda_0 = 0.2$ corresponds to the blue circles depicted in Fig. \ref{fsFig1}(c). {\bf (b)} Field intensity (arbitrary units) in the $\hat{y}$-$\hat{z}$ plane ($x=5d)$ created by the most subradiant mode in a chain of $N=50$ atoms. The field is largely evanescent transverse to the bulk of the chain, while most of the energy is radiated out through scattering at the ends of the chain. White circles denote atomic positions. {\bf (c)} Scaling with atom number of decay rates for the three most subradiant modes. A fit for large $N$ yields $\Gamma \sim N^{-3}$. {\bf (d)} Scaling with mode index $\xi$ of decay rates at fixed $N=50$. Here $\xi$ is used to label the magnitude of the decay rates in increasing order ($\xi=1$ is the most subradiant state, while $\xi=N$ has the largest decay rate). A fit for small $\xi$ yields $\Gamma \sim \xi^2$. Open and solid symbols denote the results obtained by exact diagonalization and from the ansatz of Eq.(\ref{Eq:SingleParticleStates}), respectively. The black dashed line corresponds to the eigenstate whose dominant wave vector $k$ crosses the light line ($k=k_0$). (b), (c) and (d) are for $d/\lambda_0=0.3$.} \label{fsFig3}
\end{figure*}
\subsection{Finite Lattice (Single excitation)}
\label{SecIIIB}
In this section, we analyze the decay rates and spatial properties of single-excitation eigenstates, for a lattice of finite size. We show that all eigenstates now acquire a non-zero decay rate, and subradiant states can be identified as those for which the rate is suppressed with increasing system size. The small value of the decay can be interpreted as arising from the finite system boundaries, which scatter a mode propagating in the bulk into free space.\\
{\bf{1D Linear Chain.}}\\
In the following, we consider a finite chain of atoms along $\hat{z}$, with a linear polarization along the chain (unless otherwise stated). However, a similar set of conclusions is obtained for the transverse polarization case.
{\bf \textit{Scaling of the most subradiant decay rates with system size.--}} The effective atomic Hamiltonian of Eq.(\ref{heff}) conserves the excitation number in the system, and thus, it can be diagonalized in blocks with fixed excitation number. Before proceeding futher, we discuss a technical but important point. Since the effective Hamiltonian is non-Hermitian, in general the eigenstates will not be orthogonal in the standard quantum mechanical sense (\textit{i.e.}, two eigenstates $\ket{\psi_i}$ and $\ket{\psi_j}$ will not satisfy $\Braket{\psi_i}{\psi_j}=\delta_{ij}$) \cite{AHC17}. The infinite lattice case presented an exception, as Bloch's theorem is still enforced. While this implies that general quantum mechanical rules, such as for eigenstate decompositions of states and observables, do not apply, we will nonetheless investigate the properties of the eigenstates further. This is physically motivated as they still represent non-evolving states under the Hamiltonian (aside from an overall phase and amplitude); thus, for example, they might be expected to shed light on how a general state behaves at long times.
We consider the case of a 1D chain of $N$ atoms with lattice constant $d$, for which numerical diagonalization of $\mathcal{H}_\textrm{eff}$ in the one-excitation manifold produces $N$ eigenstates (denoted by $\ket{\psi_\xi}$, $1\leq \xi \leq N$) and complex eigenvalues. As in the infinite case, the eigenvalues can be written in the form $J_{\xi}-i\Gamma_\xi /2$, with $J_{\xi}$ and $\Gamma_{\xi}$ representing the frequency shift relative to $\omega_0$ and decay rate, respectively. For concreteness, the eigenstates will be ordered in increasing decay rate, such that $\xi=1$ represents the most subradiant state and $\xi=N$ the most radiant one. To start understanding the properties of this system we fix the atomic number $N=50$ and change the lattice constant $d$. For each value of $d$, we diagonalize $\mathcal{H}_\textrm{eff}$, and obtain the $N$ different values for the decay rates and frequency shifts associated with each of the collective modes. Figure \ref{fsFig3}(a) shows the resulting single-excitation decay rates $\Gamma_\xi$ for each collective mode, normalized by the free-space single-atom emission rate $\Gamma_0$, in the case of atomic polarization parallel to the chain. In this plot, a vertical cut at a fixed value of $d/\lambda_0$ contains the $N$ different values of $\Gamma_\xi$. As expected, for large interparticle distances, the collective decay rates tend to the spontaneous emission rate of a single atom. As the distance decreases, they are periodically modulated, showing for $d/\lambda_0 < 1/2$ a qualitatively distinct behavior. In this region, the decay rates of some of the modes are dramatically suppressed ($\Gamma_\xi/\Gamma_0 \ll 1$), in accordance with the condition for the emergence of modes with zero decay rate derived in the infinite lattice case.
The subradiant modes in the finite chain are closely related to those derived in the infinite chain. First, having established that subradiant states in the infinite chain correspond to guided modes, the nonzero decay rates in the finite chain can be interpreted as emerging from scattering of these guided modes from the ends of the system. This can be seen in Fig.~\ref{fsFig3}(b), where we have plotted the field intensity in the plane $x=5d$ generated by the most subradiant state when the atomic polarization is parallel to the chain (we choose a distance $x$ offset from the $x=0$ plane containing the atomic chain in order to avoid seeing the divergent near-fields associated with each atom). Clearly, this figure shows that the field vanishes when moving away from the chain transversally, while it is very intense at the tips of the chain, where the spin wave scatters into an outgoing photon. The field intensity is computed from Eq.(\ref{fielddef}), by taking $\bra{\psi_1} \hat{\Eb}^- (\rb) \hat{\Eb}^+ (\rb) \ket{\psi_1}$. As the input field is vacuum, the intensity only involves calculating two-body correlations $\hat{\sigma}_{eg}^i \hat{\sigma}_{ge}^j$ of the eigenstate.
While the wave vector $k_z$ is strictly a good index for the modes only in the case of the infinite chain, in practice one can also unambiguously associate a distinct, dominant wave vector $k$ with each of the modes $\xi$ in the finite case. Specifically, the discrete Fourier transform of the coefficients that define each mode is peaked around a different value $k_\xi$, which can be used to label the state. In particular, let us consider a general single-excitation state, which can be written as $\ket{\psi_\xi} =\sum_j c_{\xi}^j \ket{e_j}$, where $\ket{e_j}\equiv \hat{\sigma}_{eg}^j \ket{g}^{\otimes{N}}$ is defined as the state where atom $j$ is excited while all others are in their ground states. Then, we define the discrete Fourier transform of the associated coefficients $\tilde{c}_{\xi}^k \equiv N^{-1/2} \sum_j e^{\ii k j d} c_{\xi}^j$, for discrete values of $k =2\pi m /N d$ ($1 \leq m \leq N$). For each value of $\xi$, the function $\tilde{c}_{\xi}^{k}$ shows a well defined peak at a distinct value of $k=k_{\xi}$. In Figs.~\ref{fsFig1}(b),(c) (circles) we plot the decay rates $\Gamma_{\xi}$ and energy shifts $J_{\xi}$ of each mode as indexed by the dominant wave vector, for $N=50$ atoms and both transverse and parallel polarizations, overlaid with the infinite lattice result. There is good agreement between them. For the decay rates of the finite chain, the points also correspond to those along a vertical cut in Fig.~\ref{fsFig3}(a), at the fixed value of $d/\lambda_0=0.2$.
The exact behavior of $\Gamma_\xi$ depends on the microscopic details, such as the polarization of the atoms. For instance, for two-level atoms, the smallest decay rate decreases monotonically as $d/\lambda_0\rightarrow 0$, while for transverse polarization it oscillates. Regardless of these details, however, the scaling with $N$ of the few lowest decay rates seems to show a universal behavior, going like $\Gamma_{\xi}/\Gamma_0 \sim \xi^2/N^{3}$. In Fig.~\ref{fsFig3}(c), we show the $1/N^3$ scaling for the three lowest eigenstates as a function of $N$, while in Fig.~\ref{fsFig3}(d) we show the $\xi^2$ scaling for fixed $N=20$. The scaling with $\xi$ is satisfied for all $\Gamma_{\xi}/\Gamma_0 \ll 1$. For transverse polarization, there is a particular value of lattice constant (that tends to $d/\lambda_0 \sim 0.25$ as the atom number increases), for which the decay rates do not follow exactly the scaling with $\xi$. We believe that this is related to the fact that for transverse polarization and $d/\lambda_0=0.25$ the band structure becomes flat at the edge of the Brillouin zone. Nevertheless, and as we discuss in Appendix~\ref{AppB}, the scaling $\Gamma_{\xi}\sim \xi^2/N^{3}$ seems to appear rather generically for finite-size, one-dimensional photonic crystal structures.\\
\begin{figure}
\centerline{\includegraphics[width=\linewidth]{fsFig4}}
\caption{Comparison between ansatz of Eq.(\ref{Eq:SingleParticleStates}) and exact single-excitation eigenstates in an atomic 1D chain. Here, we identify a selected number of modes based upon their dominant wave vector k, and compare the spatial wave-function coefficients $c^j$ with an ansatz built from the same wave vector: {\bf (a)} $k d=\pi N /(N+1)$ (most subradiant state), {\bf (b)} $k \sim k_0$ (close to light line) and {\bf (c)} $k d=\pi/(N+1)$ (most radiant state), for parallel atomic polarization. Blue and red circles denote the coefficients of the exact state and ansatz respectively, while the dashed red line indicates the function $\cos(k_n z)$ or $\sin(k_n z)$ associated with each mode. {\bf (d)} Error in overlap $\varepsilon$ between exact state and ansatz as a function of $k$, for parallel (blue) and transverse (black) atomic polarization. The error decreases far from the light line (denoted by black dashed line). {\bf (e)} Scaling of $\varepsilon$ with particle number $N$ for the most subradiant state. The lattice constant is set to $d/\lambda_0=0.3$ and [except in (e)] $N=20$.}
\label{fsFig4}
\end{figure}
\begin{figure*}
\centerline{\includegraphics[width=\linewidth]{fsFig5.pdf}}
\caption{Single-excitation collective modes in 2D square array of $50 \times 50$ two-level atoms. Collective decay rates as a function of predominant wave vector ($k_y$,$k_z$) associated with the mode, for different values of $d/\lambda_0$: {\bf (a)}, {\bf (b)} and {\bf (c)} are for parallel atomic polarization along $\hat{y}$ axis; {\bf (d)}, {\bf (e)} and {\bf (f)} are for transverse polarization. Subradiant and guided modes arise outside the circle defined by the light line ($|\textbf{k}| = k_0$), if $d/\lambda_0 < 1/\sqrt{2}$. For $d/\lambda_0 < 1$ and transverse atomic polarization, a different class of subradiant states emerge at $\textbf{k}=(0,0)$. {\bf (g)} Scaling of decay rates with $N$ ($N^2$ being the number of atoms) for two particular modes $(k_y,k_z) = (\pi/d,0)$ (top) and $(k_y,k_z) = (\pi/d,\pi/d)$ (bottom). Different colors are for different values of $d/\lambda_0 = 0.25,0.3,0.4,0.5,0.6,1/\sqrt{2}$ (blue, green, red, cyan, purple and brown, respectively). The dashed lines are guides to the eye for $N^{-3}$ and $N^{-6}$ scalings.}
\label{fsFig5}
\end{figure*}
{\bf \textit{Ansatz for single-excitation collective modes.--}}
If the chain is finite the single-excitation collective modes are not spin waves with pure wave vector $k_z$, and contrary to the infinite lattice case, they are not orthonormal in general. Nevertheless, as discussed earlier, the eigenstates can be characterized by a dominant wave vector $k$ that connects well with the infinite case. Furthermore, we find that the states far from the light line (including the most subradiant modes as well as those where $\textbf{k} \sim 0$) are almost orthonormal, and display a relatively simple spatial structure. This motivates us to find an orthonormal set of functions that approximates well these modes. For an even number of sites $N$, the wave-function coefficients $c_\xi^ j$ of the exact collective modes are close to the orthonormal set of functions defined by wave vector $k_n$:
\begin{align}
c_{{\rm ans,}k_n}^{j} &= \sqrt{2/(N+1)}\cos(k_n x_j) &\qquad{} \textrm{if $n$ odd}\notag\\
c_{{\rm ans,}k_n}^{j} &= \sqrt{2/(N+1))} \sin(k_n x_j) &\qquad{} \textrm{if $n$ even}.
\label{Eq:SingleParticleStates}
\end{align}
Here $k_n d = \pi n/(N+1)$, $n=1,2,...,N$ and the atomic positions $x_j = j d -x_0$ ($1\leq j \leq N$) and $x_0=(N+1) d/2$. Figure \ref{fsFig4}(a)-(c) shows the exact coefficients $c_{\xi}^j$ for the most subradiant state ($\xi=1$), a state with dominant wave vector near the light line, and the most radiant state ($\xi=N$), for $N=20$ atoms, together with the corresponding ansatz coefficients. The error between the exact wave function and ansatz can be quantified by considering the mismatch in overlap between the two states, $\varepsilon=1-|\Braket{\psi_\text{ans}}{\psi_\xi}|^2$. In Fig.~\ref{fsFig4}(d) this is plotted as a function of the wave-number $k$ associated to each of the modes. Generally, the error is negligible, except for states close to the light line. In Fig.~\ref{fsFig4}(e) we show that for the most subradiant state this error vanishes with chain length $N$ as $\varepsilon \sim N^{-2}$.
Moreover, this ansatz not only approaches the spatial pattern of the wave function, but its decay rate defined as $\Gamma_{\rm ans} = -(2/\hbar) \textrm{Im}\bra{\psi_\text{ans}} \mathcal{H}_\textrm{eff} \ket{\psi_\text{ans}}$ captures the same scaling with the index $\xi$ and $N$: $\Gamma_\textrm{ans} \propto \xi^2 / N^3$. The overall proportionality constant varies depending on the microscopic details (such as the polarization or the value of $d/\lambda_0$). For instance, for $d/\lambda_0=0.3$, $\Gamma_\text{ans}/\Gamma_{\xi=1} \approx 3/2$ (parallel polarization) and $\Gamma_\text{ans}/\Gamma_{\xi =1} \approx 8$ (transverse polarization). The comparison between $\Gamma_\xi$ and $\Gamma_\text{ans}$ is shown in Fig.~\ref{fsFig3}(c)-(d) (solid circles correspond to the ansatz).\\
{\bf{2D Square Array.}}\\
The previous results are not specific to the one-dimensional chain and can be extended to other lattice geometries. As an example, let us consider a finite square array of $N \times N$ atoms spanning the $\hat{y}$-$\hat{z}$ plane. Just like in the linear chain, we can diagonalize the block Hamiltonian with a single excitation and find the decay rates associated with the $N^2$ collective modes. We can also define an ansatz wave function with coefficients $c_{{\rm ans,}\textbf{k}}^\textrm{\bf j} \propto c_{{\rm ans,}k_y}^ {j_y} c_{{\rm ans,} k_z}^ {j_z}$, where $c_{{\rm ans,}k}^ {j}$ are the coefficients for the one-dimensional ansatz Eq.~(\ref{Eq:SingleParticleStates}). We can then associate with each of the collective modes a pair of values $\left( k_y, k_z \right)$, which lies within the first Brillouin zone, and for which the corresponding ansatz produces the highest overlap with the exact state.
In Fig.~\ref{fsFig5} we have plotted the decay rates as a function of $\left( k_y,k_z \right)$ after diagonalizing the Hamiltonian for an array of $50 \times 50$ atoms, for different values of $d/ \lambda_0 = \left\{1, 0.55, 0.3\right\}$. Figures \ref{fsFig5}(a)-(c) depict the case of polarization parallel to the array, and show the emergence of subradiant states (corresponding to wave vectors beyond the light line) for $d/\lambda_0 < 1/\sqrt{2}$. As it can be seen in this figure, the most subradiant modes correspond to those at the edges of the Brillouin zone, \textit{i.e.}, $\left( |k_y|,|k_z| \right) \sim (\pi/d,\pi/d)$. The wave-function amplitude of this mode is a generalization of the one shown in Fig.˜\ref{fsFig4}(a) for the 1D chain, where now the alternating plus and minus sign in the amplitude exhibits a checkerboard pattern. Figures \ref{fsFig5}(d)-(f) depict the decay rates for the case of transverse polarization. Here, one sees a set of subradiant states emerges beyond the light line for $d/\lambda_0 < 1/\sqrt{2}$ as before, and also a set of subradiant modes around $(k_x,k_y)=(0,0)$ for $d/\lambda_0 <1$, in agreement with the infinite lattice analysis.
While we can expect that in general the decay rate of the most subradiant modes will be suppressed with the system size, the scaling is more complex than for the linear chain. Nevertheless, we have numerically verified that for collective modes at the edge of the Brilllouin zone, and if $d/\lambda_0$ is small enough, the decay rate will scale as $\Gamma \sim N^{-\alpha}$. In particular, we find that for $\left( k_y,k_z \right) = (\pi/d,\pi/d)$ (most subradiant state), $\alpha = 6$ for $d/\lambda_0 <1/2$ and $\alpha = 3$ for $1/2 \leq d/\lambda_0 < 1/\sqrt{2}$, while for $\left( k_y,k_z \right) = (\pi/d,0)$, $\alpha = 3$ for $d/\lambda_0 < 1/2$. For ranges of $d/\lambda_0$ not included above, the decay rates are not suppressed with increasing system size, since in that case the wave vector $\left(k_x,k_y\right)$ of the two states lies within the light line. These scalings are shown for the two states in Fig.~\ref{fsFig5}(g).\\
\begin{figure}[t]
\centerline{\includegraphics[width=0.9\linewidth]{fsFig6.pdf}}
\caption{Decay rates of the $N^3$ modes at different lattice constants $d/\lambda_0$ for a cubic lattice of $10 \times 10 \times 10$ atoms with linear polarization along one of the lattice axes. The dashed red line corresponds to the particular value of $d/\lambda_0=\sqrt{3}/2$, where the light line touches the edge of the first Brillouin zone in 3D. In 3D, subradiant states can exist even beyond this value.} \label{fsFig6}
\end{figure}
{\bf{3D Cubic Array.}}\\
While the extension from 1D chains to 2D arrays is conceptually straightforward, it appears that three-dimensional lattices are governed by different physics. In particular, in infinite 1D and 2D arrays, Bloch modes diagonalize the system, and subradiant modes can be characterized as ``guided'' as the associated electromagnetic fields are evanescent in the spatial directions transverse to the array. In contrast, while Bloch modes still diagonalize the system in 3D, the associated fields are necessarily extended over space in all directions. Thus, for a finite-size system, it does not appear that subradiant states can be identified based on the infinite-system results, as was possible in 1D and 2D.
Nonetheless, for completeness, we can still numerically investigate the decay rates for a single excitation in a 3D finite-size lattice. In Fig.~\ref{fsFig6}, we plot the decay rates $\Gamma_{\xi}$ for the $N=10^3$ eigenstates associated with a $10 \times 10 \times 10$ lattice of two-level atoms, in the case that the polarization of the transition is aligned with one of the cubic axes. It can be seen that while decay rates are still most prominently suppressed for lattice constants $d \ll \lambda_0$, the effect of subradiance can survive even for lattice constants $d>\lambda_0$. It would be interesting to further explore the nature of subradiance in 3D systems in future work, and identify conceptual similarities it has to arrays in lower dimensions, if any.\\
{\bf{Atoms in a ring configuration.}}\\
The result that we have found for 1D linear chains, indicating that subradiant modes are guided and that radiation leakage is primarily from the system ends, motivates us to study the decay rates when the atoms form a closed configuration, since this might lead to a stronger suppression in the decay. In particular, we consider now that the atoms are sitting on a circular ring separated by an equal distance $d$ (see sketch in Fig.~\ref{fsFig7}) and with linear polarization transverse to the plane of the ring.
In Fig.~\ref{fsFig7}, for a distance of $d/\lambda_0=0.3$, we numerically diagonalize the single-excitation block Hamiltonian, and plot the decay rate $\Gamma_{\xi=1}$ of the most subradiant state versus atom number $N$. It can be seen that an exponential suppression emerges, $\Gamma_{\xi=1}\sim \exp(-N)$. For the chosen parameters, the minimum decay rate for a ring drops below that of an open chain for $N\gtrsim 20$ atoms. The subradiant modes of the ring can be interpreted as ``whispering gallery modes'', which weakly radiate into free space only via the finite radius of curvature. The exponential suppression with ring radius is analogous to the scaling of radiation losses in a conventional whispering gallery resonator \cite{BK03}.\\
\begin{figure}[t]
\centerline{\includegraphics[width=0.85\linewidth]{fsFig7.pdf}}
\caption{Decay rate of the most subradiant mode as a function of atom number $N$ in a circular ring of 2-level atoms (black circles). Two consecutive atoms are separated a distance $d$ as shown in the sketch (here $d/\lambda_0=0.3$), and the atomic polarization is transverse to the plane defined by the ring. For this geometry, there is an exponential suppression of the most subradiant decay rate with $N$. For comparison, the decay rate of the most subradiant mode of a linear chain with lattice constant $d$ is shown (blue circles), with only polynomial suppression with $N$.} \label{fsFig7}
\end{figure}
\begin{figure*}[t]
\centerline{\includegraphics[width=\linewidth]{fsFig10.pdf}}
\caption{Cavity in an atomic chain with slowly varying lattice constant. The chain along $\hat{z}$ is divided into three regions: in the left and right regions the lattice constant $d$ is uniform and equal to $d_{\rm max}$, while in the middle it changes slowly (from $d_{\rm max}$ to $d_{\rm min}=0.75 d_{\rm max}$). \textbf{(a)} Dispersion relation for two infinite lattices with constant $d_{\rm max}$ (cyan) and $d_{\rm min}$ (dark blue) vs wave vector along the chain (in units of the corresponding lattice constant). The light line is indicated by the vertical dashed lines with the same color. A bandwidth of frequencies (shaded pink) emerges wherein propagating modes exist for lattice constant $d_\text{min}$, but not for $d_{\rm max}$. This allows localized resonances to form within the non-uniform chain. An illustration of the chain is shown (top), where the red line represents how the lattice separation changes along the chain. \textbf{(b)} Excited state population vs atom position corresponding to the fundamental mode in the cavity, illustrated for $N=90$ atoms. \textbf{(c)} Decay rate of the fundamental mode vs atom number, showing an exponential suppression. We have chosen $d_{\rm max}/\lambda_0 = 0.4$ and atomic polarization along the chain axis.} \label{fsFig10}
\end{figure*}
{\bf{Localized resonance in an atomic chain.}}\\
Here, we also show how to achieve a spatially confined mode in a linear 1D chain of atoms, which also exhibits an exponential suppression of decay rate with atom number. This can be achieved by introducing a smooth, local variation in the lattice constant, in analogy to the principles that govern the design of a conventional photonic crystal cavity \cite{JMW95}.
To illustrate this, we consider the geometry schematically depicted in Fig.~\ref{fsFig10}(a). The atomic chain along $\hat{z}$ has been divided into three regions: in the left and right regions the lattice constant $d$ is uniform and equal to $d_{\rm max}$, while in the middle it changes slowly (from $d_{\rm max}$ to $d_{\rm min}=0.75 d_{\rm max}$) following the red line and creating a defect. The lattice constant in the middle is chosen to follow a sinusoidal variation, $d(z_i) = d_{\rm max}+(d_{\rm min}-d_{\rm max}) \sin^2(3\pi z_i/N)$, where $z_i$ ($i=N/3,...,2N/3$) denotes the atom position. In the same figure, the band structures for an infinite lattice with constant $d_{\rm max}=0.4\lambda_0$ and $d_\text{min}=0.3\lambda_0$ are plotted for the case of atomic polarization parallel to the chain. It can be seen that the smaller lattice constant $d_\text{min}$ supports propagating modes over a range of frequencies (pink shaded region) that lies within the band gap of lattice $d_{\rm max}$. Thus, for the system with slowly varying lattice constant, a set of localized resonances can appear in the middle region, with frequencies situated in the gap of lattice $d_{\rm max}$, and unable to propagate into the left and right regions.
The atomic excited state population associated with the fundamental localized mode (\textit{i.e.}, the mode with no nodes in the population) is illustrated in Fig.~\ref{fsFig10}(b), for a representative case of $N=90$ atoms. For a smooth variation in the lattice constant (here occurring over a region of size $N/3$), one expects that the fundamental mode will have a Fourier transform with an exponentially small weight of wave vectors lying within the light line. Likewise, the leakage of this mode through the left and right regions to the ends of the chain will be exponentially suppressed, leading to an overall exponential suppression in decay rate with increasing atom number $N$. The numerically calculated decay rate of this mode is plotted in Fig.~\ref{fsFig10}(c) as a function of $N$, and clearly confirms the expected behavior.
\begin{figure}[h!]
\centerline{\includegraphics[width=1\linewidth]{fsFig8.pdf}}
\caption{Two-excitation collective modes in a 1D chain of two-level atoms with polarization parallel to the chain. Probability $|c^{ij}|^2$ of atoms $i$ and $j$ to be excited (with $c^{ii}=0$) for: {\bf (a)} state resulting from occupying the most subradiant single-excitation mode twice, \textit{i.e.}, $(S_{\xi=1}^{\dagger} )^{2} \ket{g}^{\otimes N}$; {\bf (b)} most subradiant mode $c^{ij}_{\xi=1}$ (obtained by exact diagonalization); {\bf (c)} fermionic ansatz $c^{ij}_\textrm{ans}$. The two axes denote the excited atoms position ($i$ and $j$). {\bf (d)} Scaling of the collective decay $\Gamma^{(2)}$ with atom number corresponding to the states (a) (orange circles), (b) (blue open circles) and (c) (blue solid circles). The lines are polynomial fits (to the last five points), close to $\sim 1/N$ (a) and $\sim 1/N^3$ (b),(c). {\bf (e)} Decay rates as a function of the associated pair of quasi-momentum values $(k_1,k_2)$ of the two excitations (by construction $k_1\neq k_2$). Subradiant states arise when both $k_1,k_2>k_0$, where $k_0$ corresponds to the light line (dotted lines). {\bf (f)} Relative error $\delta \Gamma / \Gamma^{(2)} = |\Gamma^{(2)}_{\rm sum}-\Gamma^{(2)}| / \Gamma^{(2)}$ between the numerically exact decay rate, and the decay rate estimated from the sum of single-excitation decay rates $\Gamma^{(2)}_{\rm sum}$. {\bf (g)} Error in overlap between exact and antisymmetrized ansatz. In all plots $N$ is fixed to $50$ atoms [except (d)] and $d/\lambda_0 = 0.3$.}\label{fsFig8}
\end{figure}
\subsection{Multi-excitation modes}
\label{SecIIIC}
We now turn to the problem of multiple excitations stored in an atomic 1D lattice. As in the previous sub-section, we consider that the atoms are linearly polarized along the chain axis. However, similar results are found for the transverse polarization case. First, we note that if the Hamiltonian of Eq.(\ref{heff}) were composed of bosonic particles instead of spins (\textit{i.e.}, $\hge,\hat{\sigma}_{eg} \rightarrow a,a^{\dagger}$) the multiple excitation case would be trivial. In particular, the resulting Hamiltonian would be quadratic in the creation and annihilation operators, and a Fock state of $n$ excitations in a given mode would simply have a decay rate $n$ times that of a single excitation. The fact that we are dealing with spins, where a single spin cannot be excited twice ($\hat{\sigma}_{eg}^2=0$), leads to highly non-trivial properties of multiply excited states. In this section, we will analyze in detail the spatial properties and scaling of decay rates of multi-excitation subradiant states. While we will not explicitly utilize these states in later sections, these findings might help to provide some initial insight into how many-body physics can be encoded into subradiant manifolds.\\
Let us first consider the two-excitation manifold. A general state within this manifold can be written as $\ket{\psi^{(2)}} = \sum_{i<j} c^{ij} \ket{e_i e_j}$ where now $\ket{e_i e_j}= \hat{\sigma}_{eg}^{i} \hat{\sigma}_{eg}^{j}\ket{g}^{\otimes N}$ corresponds to the state whith atoms $i$ and $j$ excited while the rest remain in the ground state. Although it is necessary only to specify $c^{ij}$ for $i<j$ to define the wave function, in the following plots and for visual appeal we also assign values to $c^{ij}$ for $j<i$, by simply defining $c^{ij}=c^{ji}$. To illustrate that the spin system behaves differently than a bosonic system, we begin by considering the two-excitation state formed by occupying the same single-excitation mode twice. In particular, we construct the two-excitation state given by $\ket{ \psi^{(2)}_{\rm b}} \propto (S_{\xi=1}^{\dagger} )^2 \ket{g}^{\otimes N}$ (properly normalized). Here, $S_{\xi=1}^\dagger = N^{-1/2} \sum_j c^j_{\xi=1} \hat{\sigma}_{eg}^j$ is the collective operator that creates the most subradiant single excitation in a chain of $N$ atoms, when applied to the ground state. In Fig.~\ref{fsFig8}(a) we plot the corresponding probability density $|c_{\rm b}^{ ij}|^2$ for the case $N=50$. The two-excitation wave function appears relatively smooth, except for a sharp cut $c_{\rm b}^{ ii }=0$ along the diagonal, owing to the fact that a single spin cannot be excited twice. As a result of this feature, the Fourier transform of this state will be relatively broad, and in particular, will contain many components that lie within the light line and can subsequently radiate. Its decay rate, defined as $\Gamma^{(2)}_\text{b} = -(2/\hbar)\textrm{Im} \bra{\psi^{(2)}_\textrm{b}} \mathcal{H}_\textrm{eff} \ket{\psi^{(2)}_\textrm{b}}$ is only suppressed with the length of the chain as $\Gamma^{(2)}_\text{b} \sim N^{-1}$, in stark contrast to the single-excitation case. This scaling is shown in Fig.~\ref{fsFig8}(d) (orange circles).
Numerically, we now exactly diagonalize the Hamiltonian in the two-excitation manifold, and identify the most subradiant state. The scaling of its decay rate with $N$ is plotted in Fig.~\ref{fsFig8}(d) (blue open circles), and is seen to preserve the scaling $\Gamma \sim N^{-3}$ present in the single-excitation manifold. The probability density $|c^{ij}_{\xi=1}|^2$ for the case of $N=50$ is plotted in Fig.~\ref{fsFig8}(b). The wave function appears distinctly different than that of Fig.~\ref{fsFig8}(a), and in particular, it appears that the two excitations are smoothly repelled from one another.
From the previous considerations, it is apparent that the most subradiant states should simultaneously satisfy that they are composed predominantly of wave vectors beyond the light line, and that the excitations are smoothly repelled from one another in order to avoid sharp kinks in the wave function. This inspires us to try an antisymmetric (or fermionized) ansatz $\ket{\psi_{\rm ans}^{(2)}}$ for the wave function of the form $c_{{\rm ans,}k_1 k_2}^{ij} \propto c_{\text{ans},k_1}^i c_{\text{ans},k_2}^j - c_{\text{ans},k_2}^i c_{\text{ans},k_1}^j$ (properly normalized). Here $c_{\text{ans,} k_n}^i$ denote the coefficients of the single-excitation orthonormal ansatz Eq.(\ref{Eq:SingleParticleStates}) associated with the wave vector $k_n$. Such an ansatz naturally constructs a state that incorporates ``Pauli exclusion'', and a smooth separation of excitations in space. Taking $k_1d = \pi N/(N+1)$ and $k_2 d=\pi (N-1)/(N+1)$, \textit{i.e.}, building a two-excitation state from the two most subradiant single-excitation states, yields a wave function $c_{\rm{ans}}^{ij}$ that agrees well with the exact one, as seen in Fig.~\ref{fsFig8}(c) where the probability density is plotted. Moreover, the decay rate associated with this state scales again with the particle number as $\Gamma^{(2)}_\textrm{ans} \sim N^{-3}$, as it is shown in Fig.~\ref{fsFig8}(d) (blue solid circles).
We can then associate with each of the exact two-excitation collective states a pair of quasi-momentum values $\left\{ k_1, k_2 \right\}$ for which the wave-function overlap with the ansatz is maximum. The exact decay rates as a function of these values are plotted in Fig.~\ref{fsFig8}(e). This figure shows that when both $k_1, k_2 > k_0$ the decay rates are strongly suppressed, and we can identify this region as the one containing the subradiant states. In fact, the sum of decay rates of the single-excitation modes used to construct the ansatz, \textit{i.e.}, $\Gamma^{(2)}_\textrm{sum} = \Gamma_{\textrm{ans},k_1}+\Gamma_{\textrm{ans},k_2}$, is not far from the exact value. This is quantified by the relative error $\delta \Gamma \equiv |\Gamma^{(2)}-\Gamma^{(2)}_\textrm{sum}|/\Gamma^{(2)}$ and it is plotted in Fig.~\ref{fsFig8}(f). For completeness, we also show in Fig.~\ref{fsFig8}(g) the error in overlap between each two-excitation eigenstate and the best-matched ansatz state, $\varepsilon=1-|\Braket{\psi}{\psi^{(2)}_\text{ans}} |^2$. This error is very small in the subradiant region.
In the more general case of $n$ excitations, the most subradiant mode and its decay rate $\Gamma^{(n)}_{\xi = 1}$ can be found as in the previous cases by exactly diagonalizing the corresponding block Hamiltonian. For a low density of excitations, the scaling of the decay rate with the chain length is still as in the single and two-excitation manifolds, \textit{i.e.}, $\Gamma^{(n)}_{\xi = 1} \sim N^{-3}$, as shown in Fig.~\ref{fsFig9}(a) (open blue symbols). For comparison, we also show in the same figure (orange symbols) the decay rate of the state with $n$ excitations in the most subradiant mode, \textit{i.e.}, $\ket{ \psi^{(n)}_{\rm b}} \propto (S_{\xi=1}^{\dagger} )^n \ket{g}^{\otimes N}$, which scales as in the two-excitation case, $\Gamma_{\rm b}^{(n)} \sim N^{-1}$.
One can also numerically evaluate the error $\varepsilon$ in overlap between the most subradiant state and an ansatz state, $\ket{\psi_{\rm ans}^{(n)}} = \sum_{i_1<i_2<...<i_N} c_{\textrm{ans},k_1 k_2 ... k_N}^{i_1 i_2... i_N}$. Here the wave-function amplitudes $c_{ \textrm{ans},k_1 k_2...k_N}^{i_1 i_2... i_N}$ are generalized from the two-excitation case, and constructed from the Slater determinant of $n$ single-excitation wave-function ansatz coefficients $c_{\textrm{ans},k_n}^{i_n}$. For the most subradiant mode these correspond to the $n$ most subradiant single-excitation modes and in the large atom number limit the error is found to scale like $\varepsilon \sim N^{-2}$.
If the ansatz holds, then for $n$ excitations one expects that the decay rate for the most subradiant state scales like $\Gamma^{(n)}_{\xi =1}/\Gamma_0 \sim \sum_{m=1}^{n} m^2/N^3 \sim (n/N)^3$. In Fig.~\ref{fsFig9}(b), we compare this predicted scaling with numerically calculated values of $\Gamma_{\xi = 1}^{(n)}$, and find qualitatively good agreement for low excitation density $n/N\ll 1$. We note that the prediction of the ansatz also seems physically reasonable in that it can be extended to the thermodynamic limit, as it predicts a decay rate that only depends on the density $n/N$ of excitations.
\begin{figure}[b]
\centerline{\includegraphics[width=0.9\linewidth]{fsFig9.pdf}}
\caption{Multi-excitation states in a 1D chain of two-level atoms with polarization parallel to the chain. {\bf (a)} Scaling of the decay rate $\Gamma^{(n)}_{\xi = 1} / \Gamma_0$ of the most subradiant $n$-excitation state with the atom number $N$ (blue open symbols). For comparison, the decay rates of the Fock state constructed with $n$ excitations in the most subradiant single-excitation mode $\Gamma^{(n)}_{\rm b}$ (orange solid symbols), and for the fermionic ansatz $\Gamma^{(n)}_{\rm ans}$ (blue solid symbols) are shown ($n=2,3,4$ are denoted by circles, squares and diamonds). The lines are polynomial fits close to $\Gamma^{(n)}_{\xi=1} \sim N^{-3}$ and $\Gamma^{(n)}_{\rm b} \sim N^{-1}$, respectively. The sketches on top of the atomic chain represent the two-excitation density profile. {\bf (b)} Scaling of the decay rate $\Gamma^{(n)}_{\xi=1} / \Gamma_0$ of the most subradiant state with the excitation density $n/N$. The dashed line corresponds to the predicted scaling $\sim (n/N)^3$ valid at low excitation density ($n/N \ll 1$). Blue, red and black are for $N=10$, $N=15$ and $N=20$, respectively. All plots are for $d/\lambda_0=0.3$ and atomic polarization parallel to the chain.}\label{fsFig9}
\end{figure}
\section{Atoms coupled to a nanofiber: selectively-radiant states}
In the previous section we elucidated the nature of subradiant states in atomic arrays, whose long-lived nature arises from weak coupling to all propagating electromagnetic modes. Subradiant manifolds themselves might be useful for many purposes, for example, to accumulate interactions without dissipation in order to realize strongly correlated states. However, to realize an efficient atom-light interface, one would instead prefer to utilize a set of atomic states that strongly radiate into a desired electromagnetic mode (or set of modes) through constructive interference, while destructive interference simultaneously suppresses the emission rate into all undesired modes. We term states that satisfy this property to be ``selectively radiant," as the overall emission rate might not be small, but the branching ratio into desired versus undesired channels could be extremely high. It should be noted that such a definition of ``selectively radiant" is somewhat arbitrary -- for example, even a single isolated atom emitting into a dipole radiation pattern is selectively radiant, if the preferred optical mode is defined to be the dipole pattern itself. In practice, however, the collection efficiency of a dipole pattern with realistic optics is quite small \cite{EK00, EK01,DJD05,KCA08,WGH08,HSH11}, and a functionally useful definition should involve a mode (\textit{e.g.}, focused Gaussian beam or guided mode of a dielectric structure) that is generally accepted to be efficient to match to.
Here, we show that one natural way to realize and utilize selectively radiant states is to couple one-dimensional atomic chains to the guided modes of an optical nano-structure (such as an nanofiber). Qualitatively, for sufficiently small lattice constants $d<\lambda_0/2$, a set of spin-wave excitations with associated wave vector $|k_z|>k_0$ emerge, which inefficiently radiate into free space as the wave vector lies beyond the light line. However, as an optical mode guided by a high-index dielectric itself has a wave vector $|k_z|>k_0$, we show that it is possible that a set of spin-wave excitations simultaneously experiences an enhanced emission rate into the guided modes while being subradiant to free space. We will provide an explicit construction of a protocol where selectively radiant states are exploited, involving a quantum memory or photon storage. We find in particular that these states enable an exponential improvement in the error probability versus atom number, over previously known bounds.
This section is organized as follows. Section IV A describes the nanofiber and provides the Hamiltonian that governs the interactions between the atoms located in the vicinity of the nanostructure. We introduce the ``collective emission" model, which accounts for atom-atom interactions both through the guided and non-guided modes of the fiber. We also present the ``independent emission" model, in which atoms interact through the guided modes but coupling via free-space modes is neglected. This thus represents the ``standard model" of atom-light interactions specifically applied to nanofibers. In particular, it reproduces previously accepted bounds for fidelities of photon storage, against which the ``collective emission" model can be compared. Section IV B describes linear optical processes (\textit{i.e.}, single-photon transmission and reflection) for two-level atoms coupled to the fiber. We show how the conventional figure of merit, the optical depth, is not sufficient to characterize optical transport through the array when the collective emission into non-guided modes is taken into account. In Section IV C we study how selective radiance influences electromagnetically induced transparency \cite{HFI1990,HFI1991,FL00,L03,FIM05}, a phenomenon that is commonly used in photon storage protocols. In particular, we show that the bandwidth-delay product, which quantifies the number of photons that can be stored in the atomic medium, scales linearly with the number of atoms. The linear scaling is characteristic of ideal, non-lossy systems, and is in contrast to the independent emission model, which predicts a scaling that goes with the square root of the optical depth. In this sub-section, we also provide the first glimpse of improvement in photon storage beyond traditional bounds. Finally, in Section IV D we demonstrate how to achieve an exponential suppression with the atom number on the infidelity of a quantum memory.
\subsection{Description of the nanofiber}
\begin{figure}
\centerline{\includegraphics[width=\linewidth]{nFig1}}
\caption{Schematic of the setup under consideration: $N$ two-level atoms are located in the vicinity of a dielectric nanofiber of dielectric constant $\epsilon$ and radius $r$, at a distance $\rho_\text{a}$ from the center of the fiber, and at a constant distance $d$ from each other. For the calculations in this manuscript, we take $k_0r=1.2$, $\rho_\text{a}=1.5r$, and $\epsilon=4$. The atoms interact with each other not only through the guided mode, but also through non-guided photons. The single-atom emission rates into the fiber and into free space are $\Gamma_{\rm 1D}$ and $\Gamma'$, respectively.} \label{nFig1}
\end{figure}
The possibility of enhancing atom-light interactions through selective radiance should exist for any nanophotonic structure where atoms can be periodically trapped, including in nanofibers \cite{VRS10,GCA12,polzik,CGC16} and 1D and 2D photonic crystal waveguides \cite{DHH15,DHC15,GHH15,HGA16}. For complicated structures, however, the Green's function cannot be obtained analytically. Furthermore, while the Green's function can be calculated numerically \cite{S00}, to do so with sufficient accuracy appears quite challenging (in particular, it must be calculated with enough accuracy so that diagonalization correctly captures subradiant emission rates that scale like large inverse powers of $N$). Motivated by this observation, here we focus on a special geometry where the Green's function can be exactly obtained, which consists of a chain of atoms coupled to guided modes of an infinite, cylindrical nanofiber.
We consider that the chain of atoms lies parallel to the axis of a dielectric nanofiber oriented along $\hat{z}$, with radius $r$ and relative permittivity $\epsilon$ (or corresponding refractive index $n_\text{fiber}=\sqrt{\epsilon}$). As shown in Fig.~\ref{nFig1}, the distance between the atoms and the center of the nanofiber is $\rho_\text{a}$, and the orientation of their dipole transition is directed along $\hat{\rho}$, perpendicular to the axis of the nanofiber (\textit{i.e.}, $\db=\wp\hat{\rho}$). The Green's function for such a nanofiber can be found analytically \cite{S1941,YS08_2,KGB05,DSF10,LR14,QBH16,LR17}. In particular, we follow the work of Klimov and Ducloy ~\cite{KD04}. In the following, we provide a qualitative description of the derivation, while details are given in Appendix~\ref{AppC}. The first step in the derivation is to separate the Green's function into two terms, \textit{i.e.}, $\mathbf{G}(\rb_i,\rb_j,\omega_0)=\mathbf{G}_0(\rb_i,\rb_j,\omega_0)+\mathbf{G}_\text{sc}(\rb_i,\rb_j,\omega_0)$. Here $\mathbf{G}_0$ is the already-known vacuum Green's function given by Eq.~\eqref{Greens_def}, which corresponds to the field emitted by a dipole in free space, and $\mathbf{G}_\text{sc}$ is a general solution to the sourceless wave equation, which will physically correspond to the (thus far unknown) field scattered by the nanofiber. Exploiting the cylindrical symmetry of the problem, one can employ separation of variables and expand the vacuum and scattered Green's functions using a set of functions $f_{m, k_\parallel}(\rho)e^{ik_\parallel z+im\phi}$. Here $k_\parallel$ is the longitudinal wave vector and $m$ denotes angular momentum. The coefficients in the expansion of $\mathbf{G}_\text{sc}$ associated with each value of $k_\parallel$ and $m$ inside and outside the fiber are a priori unknown, but can be solved for through equations that enforce electromagnetic field continuity relations at the surface of the fiber.
The fiber supports a set of guided modes, \textit{i.e.}, electromagnetic modes that propagate along the nanostructure and are confined in the transversal direction. These modes are denoted by their angular momentum $m$, and their associated wave vectors $k_{m} (\omega_0)$ always satisfy $|k_m(\omega_0)|>\omega_0/c$, as the guiding mechanism is by total internal reflection [here we have dropped the ``$\parallel$" subscript associated with the guided mode wave vector $k_m(\omega_0)$, for notational simplicity]. In other words, these modes are evanescent, and their dispersion relations are situated beyond the light line. The number of guided modes is determined by the fiber radius and dielectric constant. We will restrict ourselves to a single-mode fiber (with $m=\pm 1$), which occurs for a sufficiently small fiber radius. Instead of working with $\mathbf{G}_0$ and $\mathbf{G}_\text{sc}$, for our purposes it is convenient to isolate the guided mode contribution and separate the Green's function $\mathbf{G}(\rb_i,\rb_j,\omega_0)=\mathbf{G}_\text{1D}(\rb_i,\rb_j,\omega_0)+\mathbf{G}'(\rb_i,\rb_j,\omega_0)$ into two terms: one that characterizes the excitation of the guided mode of the fiber, $\mathbf{G}_\text{1D}(\rb_i,\rb_j,\omega_0)$, and another that describes the non-guided electromagnetic modes, $\mathbf{G}'(\rb_i,\rb_j,\omega_0)$. In particular, the guided Green's function takes the form $\mathbf{G}_\text{1D}(\rb_i,\rb_j,\omega_0)=\mathbf{g}(\boldsymbol{\rho}_\text{a})\,e^{\ii k_\text{1D}|z_i-z_j|}$, where $\mathbf{g}(\boldsymbol{\rho}_\text{a})$ is a tensor that only depends on the radial and azimuthal position of the atoms (assumed to be identical), and $k_{\rm 1D}=|k_{\pm 1}(\omega_0)|$.
The dynamics of the atoms is governed by the non-Hermitian Hamiltonian $\mathcal{H}_{\rm eff}$ of Eq.~(\ref{heff}), which can be similarly split, \textit{i.e.}, $\mathcal{H}_\text{eff}=\mathcal{H}_{\rm 1D}+\mathcal{H}'$. From the form of $\mathbf{G}_\text{1D}$ given above, it follows that
\begin{align}\label{hguided}
\mathcal{H}_{\rm 1D}=-\ii\frac{\hbar\Gamma_{\rm 1D}}{2} \sum_{i,j=1}^Ne^{\iik_{\rm 1D} |z_i-z_j|} \hat{\sigma}_{eg}^{i}\hge^j,
\end{align}
where $\Gamma_{\rm 1D}=(2\mu_0\,\omega_0^2 \wp^2/\hbar) \,\text{Im}\,G^{\rm 1D}_{\rho \rho}(\rb_i,\rb_i,\omega_0)$ is the spontaneous emission rate of a single atom into the fiber guided mode. The plane-wave dependence reflects the fact that the guided photon propagates without diffraction between two atoms and thus produces an infinite-range interaction.
The non-guided term
\begin{align}\label{hprime}
\mathcal{H}'=-\frac{3\pi\hbar\Gamma_0}{k_0}\sum_{i,j=1}^N G'_{\rho\rho}(\rb_i,\rb_j,\omega_0)\,\hat{\sigma}_{eg}^i\hat{\sigma}_{ge}^j
\end{align}
accounts for the interaction through the remaining non-guided electromagnetic modes. Already for just a single atom, the self term of the non-guided Green's function $G'_{\rho \rho}(\rb_i,\rb_i,\omega_0)$ gives rise to both a frequency shift and a decay rate that we will denote as $J'-\ii\Gamma'/2=-(\mu_0\,\omega_0^2 \wp^2/\hbar) \,G'_{\rho \rho}(\rb_i,\rb_i,\omega_0)$. This self-term reflects the fact that the modification of electromagnetic modes by the nanofiber causes a single atom to have a resonance frequency $\omega_0+J'$ shifted from its vacuum value, and a decay rate into radiative modes $\Gamma'$ different than $\Gamma_0$. For many atoms, the above Hamiltonian accounts for collective emission into non-guided modes, as it takes into account atom-atom interactions that are not mediated by the guided mode. Unlike $\mathbf{G}_\text{1D}$, $\mathbf{G}'$ does not admit a simple form, and in what follows it will be evaluated numerically using the prescription detailed in Appendix~\ref{AppC}. Throughout this manuscript, we will refer to the dynamics generated by $\mathcal{H}_\text{1D}+\mathcal{H}'$ as the ``collective emission" model.
Whether in free space or a nanofiber (or other guided structures), exact collective effects involving modes that are not directly of interest (such as those captured by $\mathcal{H}'$) are typically difficult to treat in the context of applications of atomic ensembles. It is usually heuristically argued that photon-mediated interactions through these modes are not relevant, particularly for disordered or dilute atomic gases, and the ``standard'' model within quantum optics is to ignore such terms \cite{HSP10, GAF07,CJG12,QBH16}. Specifically, the terms of $\mathbf{G}'(\rb_i,\rb_j,\omega_0)$ involving two different atoms ($\rb_i\neq\rb_j$) are assumed to be zero, and the Hamiltonian accounting for emission into non-guided modes reduces to
\begin{align}\label{hindep}
\mathcal{H}'_\text{indep}=\hbar(J'-\ii\Gamma'/2)\sum_{j=1}^N\hat{\sigma}_{ee}^j.
\end{align}
In this approximation, the non-guided modes of the fiber introduce a modified Lamb shift due to the presence of the fiber surface, and more importantly, provide independent baths for each atom to emit into, at a rate $\Gamma'$. We will refer to the dynamics generated by $\mathcal{H}_\text{1D}+\mathcal{H}'_\text{indep}$ as the ``independent emission" model.
Before proceeding further, we digress to clarify the different usages of the terms super/sub-radiance in literature. Within the independent emission model, the concept of superradiance and subradiance has also been discussed, since $\mathcal{H}_\text{1D}$ alone yields a set of collective atomic states that radiate strongly or weakly into the waveguide \cite{CJG12,LFL13,AHC17}. Similar effects have also been pointed out in cavities (with collective states emitting strongly or weakly into the cavity mode) \cite{BCW12,RAK15,CFB15}. Protocols for photon generation and storage and other quantum information tasks have been built around the manipulation of these states \cite{CJG12,GVC15}. However, as these models still assume independent emission into free space, these protocols do not surpass conventional error bounds.
Throughout this manuscript, the nanofiber radius is taken to be $k_0r=1.2$, the distance between the atoms and the center of the nanofiber is $\rho_\text{a}=1.5r$, and the dielectric constant of the fiber is $\epsilon=4$ (as that of silicon nitride). As an illustration, for the D$_1$ line of Cesium (of resonance frequency $\omega_0=2\pi\times 335.1$ THz), the radius of the fiber would be $r\simeq 170$~nm, and the distance between the atoms and the fiber surface would be approximately $85$~nm. The wave vector of the photonic guided mode is found to be $k_{\rm 1D}\simeq 1.3 k_0$, larger than any wave vector within the light line, as the guided mode is confined. The single-atom decay rates are calculated to be $\Gamma_{\rm 1D}\simeq 0.4\Gamma_0$ to the guided mode, and $\Gamma'\simeq 1.3\Gamma_0$ to the non-guided modes. The modified Lamb shift due to the fiber is $J'\simeq-0.5\Gamma_0.$
In the following sub-sections, we will utilize the formalism introduced above to identify novel phenomena that emerge when collective emission is exactly accounted for, which cannot be predicted from the independent emission approximation, and will show how collective emission enables exponential improvement in performance for quantum memories of light.
\subsection{Linear optics for two-level atoms}
\begin{figure*}
\centerline{\includegraphics[width=\linewidth]{nFig2}}
\caption{Linear optics for a chain of $N=20$ atoms coupled to a nanofiber, for $d=\lambda_\text{1D}/4$ ($k_\text{1D}d=\pi/2$). \textbf{(a)} Transmittance, \textbf{(b)} reflectance, and \textbf{(c)} loss probability as a function of the atom-probe detuning. The blue curves are obtained by including the collective emission into free-space [see Eq.~\eqref{hprime}], and the red lines are produced within the ``independent emission" model, where it is assumed that free-space emission is a single-atom effect [see Eq.~\eqref{hindep}]. The parameters characterizing the nanofiber are given in Fig.~\ref{nFig1}.} \label{nFig2}
\end{figure*}
We begin by studying the transmission and reflection properties of a chain of atoms coupled to the fiber within the ``independent emission" model. The effective Hamiltonian that describes the atomic dynamics under a coherent-state guided-mode probe field of frequency $\omega_\text{p}$ reads $\mathcal{H}_\text{tot}=\mathcal{H}_\text{drive}+\mathcal{H}'_\text{indep}+\mathcal{H}_\text{1D}$. The Hamiltonians $\mathcal{H}_\text{1D}$ and $\mathcal{H}'_\text{indep}$ are defined in Eqs.~\eqref{hguided}, and \eqref{hindep}, respectively, and the driving term is given by
\begin{align}\label{hdrive}
\mathcal{H}_\text{drive}=-\hbar\Delta\sum_{i=1}^N\hat{\sigma}_{ee}^i-\hbar\sum_{i=1}^N\left(\Omega\, e^{\iik_{\rm 1D} z_i}\,\hat{\sigma}_{eg}^i+\text{h.c.}\right).
\end{align}
In the above expression, $\Delta=\omega_{\rm p}-\omega_0$ is the detuning between the probe field frequency and the atomic resonance frequency. We have also defined the Rabi frequency of the guided-mode probe field as $\Omega=\db^*\cdot\Eb_{\rm p}^+(\rho_\text{a})/\hbar$, where $\Eb_{\rm p}^+(\rb)\equiv\braket{\hat{\Eb}_{\rm p}^+(\rho_\text{a})}$ is the amplitude of a coherent probe field that implicitly contains the radial position $\rho_\text{a}$ of the atoms. For the remainder of this subsection, we consider that the probe field is weak and does not saturate the atoms. Therefore, all the calculations can be performed in the single-excitation manifold, the realm of classical linear optics.
In the single excitation subspace, the wave function of the atomic ensemble is written as the superposition $\ket{\psi(t)}= c_g (t)\ket{g}^{\otimes N}+ \sum_{j=1}^N c_e^{\, j} (t)\ket{e_j}$, where $\ket{e_j}=\hat{\sigma}_{eg}^j \ket{g}^{\otimes N}$. In the low saturation regime, with $c_g\simeq1$, the evolution equations for the amplitude of the $\ket{e}$ states are found to be
\begin{align}\label{sigmas}
\dot{c}_e^{\,j}(t)=&\ii\left(\Delta_{\rm}-J'+\ii\frac{\Gamma'}{2}\right)c_e^{\,j}(t)+\ii \,\Omega\, e^{\iik_{\rm 1D} z_j}\\\nonumber
-&\frac{\Gamma_{\rm 1D}}{2}\sum_{i=1}^N e^{\iik_{\rm 1D} |z_i-z_j|}\,c_e^{\,i}(t).
\end{align}
The generalized input-output expression of Eq.~\eqref{fielddef} allows us to calculate the guided-mode field at any point of the fiber, which reads
\begin{align}\label{iosimple}
\hat{\Eb}^+(\rb)&=\hat{\Eb}_{\rm p}^+(\rb)+\mu_0 \omega_0^2\sum_{j=1}^N {\bf G}_\text{1D}(\rb,\rb_j,\omega_0)\cdot\db\,\,\hge^j.
\end{align}
It is important to notice that the Green's function appearing in the field equation is not the total one, but just that of the guided mode, as it describes the propagation of the photonic guided field along the nanostructure. Due to the cylindrical symmetry of the fiber, the guided modes with angular momenta $m=\pm 1$ are degenerate. One can alternatively take superpositions of these to obtain quasilinearly-polarized $H$ and $V$ modes \cite{LR14,QBH16}. The polarization basis of the fiber can always be set so that the $H$ mode at the atomic positions has polarization components along the $\hat{\rho}$ and $\hat{z}$ directions. We will consider the case where the probe field is $H$-polarized, in which case the atoms scatter solely back into $H$, and the $V$-polarized mode de-couples from the problem.
We can thus project the input-output equation into 1D equations for the $H$-modes, and further separate the guided fields into left- and right-propagating components. The resulting equations are given by
\begin{subequations}\label{leftright}
\begin{align}\label{rightf}
\hat{E}_\text{1D,R}^+(z)&=\hat{E}_\text{in,R}^+(z)+\Omega e^{\ii k_{\rm 1D} z}\\\nonumber
&+\ii\frac{\Gamma_{\rm 1D}}{2} \sum_{j=1}^N \,e^{\ii k_{\rm 1D}(z-z_j)}\Theta(z-z_j)\hge^j,
\end{align}
\begin{align}\label{leftf}
\hat{E}_\text{1D,L}^+(z)&=\hat{E}_\text{in,L}^+(z)\\\nonumber
&+\ii\frac{\Gamma_{\rm 1D}}{2} \sum_{j=1}^N \,e^{\ii k_{\rm 1D}(z_j-z)}\Theta(z_j-z)\hge^j,
\end{align}
\end{subequations}
where $\hat{E}_\text{in,R(L)}^+(z)$ are the right(left)-going vacuum fluctuation fields, and $\Theta$ is the Heaviside function. The vacuum fluctuations do not contribute to any of our observables of interest. For convenience, we have re-scaled the fields so that the atomic parameters $\Omega$ and $\Gamma_{\rm 1D}$ directly appear.
In the quasistatic limit ($\dot{c}_e^{\,j}=0$), the solutions of Eq.~\eqref{sigmas} for the $\ket{e}$-state amplitudes are directly proportional to the probe field Rabi frequency $\Omega$. Together with Eqs.~(\ref{rightf},\ref{leftf}), this allows us to find the reflection and transmission coefficients for the guided field. For example, the transmittance is found by evaluating $T=\braket{\psi|\hat{E}_\text{1D,R}^-(z)\hat{E}_\text{1D,R}^+(z)|\psi}/\Omega^2$, where $z$ is a position immediately after the right-end of the atomic chain, and $\hat{E}_\text{1D,R}^-(z)$ is the Hermitian conjugate of $\hat{E}_\text{1D,R}^+(z)$. A similar expression can be found for the reflectance $R$. One can also calculate the loss probability due to scattering into free space, which is given by $\kappa=1-T-R$.
We choose the distance between the atoms to be $d=\lambda_\text{1D}/4$, with $\lambda_\text{1D}=2\pi/k_\text{1D}$ being the guided-mode wavelength. Any other separation except for the so-called mirror configuration, \textit{i.e.}, $d=\lambda_\text{1D}/2$ or integer multiples thereof \cite{CJG12}, would display qualitatively similar optical properties. In Appendix~\ref{AppD}, we analyze the linear optics of such a special configuration, which has been theoretically known and experimentally observed to become a very reflective mirror \cite{DSR95,SZC11,CJG12}, around which powerful protocols for quantum information processing can be built \cite{CJG12,GVC15,PKG16}. In the mirror configuration and within the independent emission model, there is only one atomic collective state that couples to the guided mode of the fiber, decaying superradiantly into it at a rate $N\Gamma_{\rm 1D}$.
In contrast, for any other separation, every atomic collective state is excited by the probe field, and contributes to light transmission and reflection. Therefore, the behavior of the atoms cannot be attributed to a single ``super-atom" of enhanced decay rate, and the transmission spectrum -- depicted by the red line of Fig.~\ref{nFig2}(a) -- differs significantly from a Lorentzian \cite{AHC17}. For large enough number of atoms, and for low single-atom coupling efficiency into the waveguide ($\Gamma_{\rm 1D}\lesssim \Gamma'$), the transmittance approximately follows the expression
\begin{align}\label{random}
T_\text{indep}\simeq \text{exp}\left[\frac{-D}{1+4(\Delta-J')^2/\Gamma'^2}\right],
\end{align}
in accordance with the result obtained for a free-space atomic gas \cite{AHC17}. On resonance (when $\Delta-J'=0$), the figure of merit that determines how much light is transmitted is the optical depth, $D=2N\Gamma_{\rm 1D}/\Gamma'$. For a chain of $N=20$ atoms, the expression of Eq.~\eqref{random} nicely reproduces the transmittance spectrum shown in Fig.~\ref{nFig2}(a). The corresponding reflectance spectrum is displayed by the red curve of Fig.~\ref{nFig2}(b), which shows a very small bump, as the distance $d=\lambda_\text{1D}/4$ minimizes reflection due to destructive interference \cite{CMS15}. As both transmission and reflection are very small close to resonance, the dominant process is photon loss due to atom-mediated scattering into free space. The loss probability $\kappa$ is shown in Fig.~\ref{nFig2}(c).
If the atoms are closely packed, the above calculations are no longer valid due to the atomic interactions mediated by non-guided modes. Nevertheless, the previous techniques can be straightforwardly modified to calculate the new transmission and reflection coefficients. Within the ``collective emission" model, the atoms evolve under the Hamiltonian $\mathcal{H}_\text{tot}=\mathcal{H}_\text{drive}+\mathcal{H}'+\mathcal{H}_\text{1D}$, where $\mathcal{H}'$ replaces $\mathcal{H}'_\text{indep}$. In the low saturation limit, the evolution equations for the $\ket{e}$ state amplitudes now read
\begin{align}
\dot{c}_e^{\,j}(t)=&\ii\Delta_{\rm}c_e^{\,j}(t)-\frac{\Gamma_{\rm 1D}}{2}\sum_{i=1}^N e^{\iik_{\rm 1D} |z_i-z_j|}\,c_e^{\,i}(t)\\\nonumber
&+\ii \,\Omega\, e^{\iik_{\rm 1D} z_j}+\ii\frac{3\pi\Gamma_0}{k_0}\sum_{i=1}^N G_{\rho\rho}'(\rb_i,\rb_j,\omega_\text{p})\,c_e^{\,i}(t).
\end{align}
Once again, we evaluate Eqs.~(\ref{rightf},\ref{leftf}) using the steady state solution for the atomic wave function, in order to reconstruct the electromagnetic field along the nanofiber. We overlay our results for the transmission, reflection, and loss probability spectra in Figures~\ref{nFig2}(a-c). The transmittance spectrum displays many sharp peaks that are not observed within the independent emission model, similar to what was found in Ref.~\cite{KSP16}. These peaks correspond to the interference between different collective atomic modes, whose response can be observed due to the diminished photon loss. Close to resonance, reflectance is significantly larger than that obtained within the independent emission model. As a matter of fact, accounting for cooperative emission into non-guided modes lowers significantly the probability $\kappa$ of photon scattering into free space, as can be observed in Fig.~\ref{nFig2}(c). However, this decrease in loss is not uniform for all detunings, and close to resonance this spectrum also showcases sharp peaks. Globally, the behavior is much more complex than that of a ``standard" atomic ensemble, such as given by Eq.~\eqref{random}.
One interesting question is how the loss scales with the number of atoms. We find that, far from resonance, $\kappa/\kappa_\text{indep}\sim N^{-1}$, where $\kappa_\text{indep}$ is the photon loss probability when the collective emission into non-guided modes is neglected. However, in Section III we have found that for a chain of atoms in free space the decay rate of the most subradiant mode scales as $N^{-3}$. Taken together, Figs.~\ref{nFig2}(a-c) suggest a simple reason for this apparent discrepancy. Based on previous arguments of Sec. IIIA, both the infinite atomic chain and the fiber have sets of perfectly guided modes, which experience zero radiation into free space. However, the dispersion relation of the effective medium formed by the fiber and the atomic chain is different from that of the bare fiber alone (\textit{i.e.}, for a given guided-mode frequency, there is a different wave vector). This impedance mismatch leads to large scattering loss at the interface between the two different systems (bare fiber versus fiber with atoms), in close analogy to what occurs between different conventional waveguides \cite{MDD1997,APL03}.
\subsection{Electromagnetically induced transparency}
\begin{figure}
\centerline{\includegraphics[width=\linewidth]{nFig3}}
\caption{Electromagnetically-induced transparency scheme. The $\ket{g}$ to $\ket{e}$ transition is coupled to the guided mode, and the $\ket{s}$ to $\ket{e}$ transition is driven by an external, classical control field of Rabi frequency $\Omega_\text{c}$. The distance between the atoms is a quarter of the guided mode wavelength, $d=\lambda_\text{1D}/4$.} \label{nFig3}
\end{figure}
\begin{figure*}
\centerline{\includegraphics[width=\linewidth]{nFig4}}
\caption{Electromagnetically induced transparency and photon storage efficiency, within the independent emission (red) and collective emission (blue) models. \textbf{(a)} Transmittance spectrum for a chain of $N=20$ atoms. The control field Rabi frequency is set to $\Omega_\text{c}=\Gamma_0$, while other system parameters can be found in the main text. \textbf{(b)} Scaling of the bandwidth-delay product with the number of atoms $N$. The circles represent the numerical calculation. The red line shows the theoretically predicted answer within the independent emission model, $\mathcal{P}=\sqrt{D}\simeq 0.76\sqrt{N}$, and the blue curve represents the best fit of the numerical data to a linear scaling, $\mathcal{P}\simeq 0.30 N$. The control field intensity is the same as in (a). \textbf{(c)} Infidelity in the retrieval of the spin excitation given by Eq.~\eqref{spinwave}. The circles show the numerics. The red line represents the expected scaling derived in Ref.~\cite{GAL07}, within the independent emission model, $\varepsilon=5.8/D\simeq 10/N$, and the blue line is the best fit of the numerical results to a $\propto N^{-2}$ scaling, and shows $\varepsilon\simeq26 /N^2$ (the range for the fit is $N\in[30,200]$).} \label{nFig4}
\end{figure*}
Having posited that scattering at the interface between the bare fiber and the atomic chain dominates the losses observed in the two-level case, we now attempt to reduce these losses by better matching the dispersion relations of the two regions, using three-level atoms under conditions of electromagnetically induced transparency (EIT) \cite{HFI1990,FL00,L03,FIM05}.
The system under consideration is illustrated in Fig.~\ref{nFig3}. In addition to the $\ket{g}$ to $\ket{e}$ transition studied earlier, a third metastable level $\ket{s}$ (of frequency $\omega_s$) is added. We assume that the $\ket{e}$ to $\ket{s}$ transition does not couple to the optical fiber (\textit{e.g.}, due to its dipole matrix element being orthogonal to the guided mode polarization), but can be addressed by an external classical control field of Rabi frequency $\Omega_c$ that propagates through free space. Through a two-photon interference effect mediated by the control field, a near-resonant guided photon interacting with an atom originally in state $\ket{g}$ can be coherently mapped to state $\ket{s}$, with minimal excitation of state $\ket{e}$. The lack of population in $\ket{e}$ and associated photon scattering causes the otherwise optically opaque medium to become transparent, and thus EIT nominally preserves the effective refractive index of the guided mode.
Again, we consider the case where atoms are separated by a distance $d=\lambda_\text{1D}/4$ ($k_{\rm 1D} d=\pi/2$), to guarantee minimal reflection. Any other distance of the form $d=n\lambda_\text{1D}/4$ (with $n$ being odd) also strongly suppresses reflection and should suffice, as long as $d$ fulfils the subradiance condition. In particular, as atoms nominally do not alter the effective index under EIT, the guided wave vector $k_{\rm 1D}$ itself should lie outside the light line. Without atoms this is clearly always true, as the fiber mode is guided. With atoms, however, one must ensure that $k_{\rm 1D}$ lies outside the light line when folded back into the first Brillouin zone. If we set the distance between the atoms to be such that $k_{\rm 1D} d=\pi/2$, then $k_{\rm 1D}$ lies within the first Brillouin zone and automatically satisfies this constraint, $k_0<k_{\rm 1D}\leq \pi/d$. However, if $k_{\rm 1D} d=3\pi/2$, the condition on the guided mode wave vector, $k_{\rm 1D}>3k_0$, becomes much more stringent. In fact, for the radius and dielectric constant of the fiber here considered, the subradiance condition is not met for $d=3\lambda_\text{1D}/4$.
We begin by solving for the characteristics of EIT under the independent emission model, which we find to reproduce previously derived and well-known results in free-space atomic ensembles. In particular, we consider the system evolving under the effective Hamiltonian $\mathcal{H}_\text{tot}=\mathcal{H}_\text{1D}+\mathcal{H}'_\text{indep}+\mathcal{H}_\text{drive}+\mathcal{H}_\text{c}$. The Hamiltonians $\mathcal{H}_\text{1D}$, $\mathcal{H}'_\text{indep}$, and $\mathcal{H}_\text{drive}$ are defined in Eqs.~\eqref{hguided}, \eqref{hindep}, and \eqref{hdrive}, respectively. $\mathcal{H}_\text{c}$ captures the interaction of the atoms with the control field, and is given by
\begin{align}
\mathcal{H}_{\rm c}=-\hbar\sum^N_{i=1}\Delta_{\rm s}\hat{\sigma}_{ss}^i-\hbar\sum^N_{i=1}\Omega_c (z_i)\left(\hat{\sigma}_{es}^i+\hat{\sigma}_{se}^i\right),
\end{align}
where $\Delta_{\rm s}=\omega_{\rm p}+\omega_c-\omega_{\rm s}$ is the two-photon detuning. We take the control field Rabi frequency $\Omega_c$ to be real, and allow for a possible spatial dependence. We have also assumed that $\ket{e}$ has a negligible decay rate into $\ket{s}$, as in the case of a dipole-forbidden transition or ladder system. For EIT within the independent emission model, this assumption is not necessary, as the emission rate from $\ket{e}$ to $\ket{s}$ can be incorporated into $\Gamma'$ and simply leads to a moderate decrease of optical depth $D$. Such a condition, however, becomes important when considering the collective emission case (see more detailed discussion about multi-level structure in Sec. V).
Within the single excitation manifold, the wave function of the atomic ensemble is $\ket{\psi(t)}=c_g (t)\ket{g}^{\otimes N}+ \sum_{j=1}^N c_e^{\, j} (t)\ket{e_j}+ \sum_{j=1}^N c_s^{\, j} (t)\ket{s_j}$, with $\ket{s_j}=\hat{\sigma}_{sg}^j \ket{g}^{\otimes N}$. For a uniform control field [$\Omega_\text{c}(z_i)\equiv\Omega_\text{c}$] and in the low saturation limit [$c_g(t)\simeq 1$], the equations for the evolution of the amplitudes of the $\ket{e}$ and $\ket{s}$ states read
\begin{subequations}\label{eqEIT}
\begin{align}
\dot{c}_e^{\,i}(t)&=\ii\left(\Delta-J'+\ii\frac{\Gamma'}{2}\right) c_e^{\,i}(t)+\ii \,\Omega\, e^{\iik_{\rm 1D} z_i}\\\nonumber
&+\ii\Omega_c c_s^{\,i}(t)-\frac{\Gamma_{\rm 1D}}{2}\sum_{j=1}^N e^{\iik_{\rm 1D} |z_i-z_j|}\,c_e^{\,j}(t),\\
\dot{c}_s^{\,i}(t)&=\ii\Delta_{\rm s}c_s^{\,i}(t)+\ii\Omega_c\,c_e^{\,i}(t).
\end{align}
\end{subequations}
We solve these equations in the steady state and readily find $c_s^{\,i}=-(\Omega_c/\Delta_s)\,c_e^{\,i}$, and
\begin{align}\label{eqEITstat}
\left(\Delta-J'-\frac{\Omega_c^2}{\Delta-J'}+\ii\frac{\Gamma'}{2}\right)&c_e^{\,i}+\Omega_i\,e^{\iik_{\rm 1D} z_i}\\\nonumber
+&\ii\frac{\Gamma_{\rm 1D}}{2}\sum_{j=1}^N e^{\iik_{\rm 1D} |z_i-z_j|}\,c_e^{\,j}=0.
\end{align}
Here, we have chosen $\Delta_{\rm s}=\Delta-J'$, which assures total transparency when the probe field is resonant with the (shifted) $\ket{e}-\ket{g}$ transition ($\Delta-J'=0$). Having found the steady state solution of the spin wave function, it is now possible to calculate the transmitted guided-mode field by means of the input-output expression of Eq.~\eqref{iosimple}, and, therefore, the transmission coefficient of the array under EIT, $t_{\rm EIT}$.
The transmission coefficient gives us enough information to calculate two key quantities that describe the EIT medium: the group velocity of the polariton, and the bandwidth-delay product, a parameter that quantifies how many spatially separate photons can be stored in the atomic ensemble \cite{CSH11}. After propagating along the atomic chain, the guided mode field acquires a phase $t_{\rm EIT}\equiv e^{\ii k_{\rm eff} Nd}$, where $k_{\rm eff}$ is a complex effective wave vector that encodes both light absorption and dispersion. Up to second order in the atom-probe detuning, the effective wave vector reads \cite{CSH11,AHC17}
\begin{align}\label{keff}
k_{\rm eff}&=k_\text{1D}+\frac{\Gamma_{\rm 1D}}{2d\Omega_c^2}(\Delta-J')\\\nonumber
&+\ii\frac{\Gamma_{\rm 1D}(\Gamma'+\eta\Gamma_{\rm 1D}/2N)}{4d\Omega_c^4}(\Delta-J')^2,
\end{align}
where $\eta=0 (1)$ for an even (odd) number of atoms. It can be seen that when $\Delta=J'$, the effective wave vector perfectly matches that of the bare fiber, $k_\text{eff}=k_\text{1D}$, as originally desired. From the above expression, the group velocity at the center of the transparency window is found to be $v_\text{g}=\left(\partial k_{\rm eff}/\partial \Delta\right)^{-1}=2\Omega_c^2d/\Gamma_{\rm 1D}$. The delay time, \textit{i.e.}, the time it takes for this slow polariton to traverse the medium is $\tau=Nd/v_\text{g}=N\Gamma_{\rm 1D}/2\Omega_c^2$. The bandwidth of the transparency window, which dictates how spectrally narrow a photon has to be to propagate with high transparency, is defined as
\begin{align}\label{deltaeit}
\Delta_\text{EIT}=2\delta=2\Omega_c^2\sqrt{2/N\Gamma_{\rm 1D}(\Gamma'+\eta\Gamma_{\rm 1D}/2N)},
\end{align}
where $\delta$ is the detuning for which $|t_\text{EIT}|^2=1/e$. Therefore, the bandwidth-delay product, $\mathcal{P}=\tau\Delta_\text{EIT}=\sqrt{2N\Gamma_{\rm 1D}/(\Gamma'+\eta\Gamma_{\rm 1D}/2N)}\simeq\sqrt{D}$, scales with the square root of the optical depth $D=2N\Gamma_{\rm 1D}/\Gamma'$, for realistic values of $\Gamma'$. This is the same scaling that is predicted in free-space atomic ensembles, when atoms are assumed to emit independently \cite{L03,FIM05}. In contrast, for some idealized system without loss (\textit{i.e.}, $\Gamma'=0$), the bandwidth-delay product scales simply with the number of atoms, $\mathcal{P}\sim N$ \cite{CSH11,SPP05}. This result does not follow from the perturbative expansion of Eq.~\eqref{keff}. Rather, one can perform an exact calculation of the optical band structure and the bandwidth, $\Delta_\text{EIT}\sim\Omega_c^2/\Gamma_{\rm 1D}$, all of which is usable in the absence of loss~\cite{CSH11}.
Figure~\ref{nFig4}(a) depicts a representative transmittance spectrum for a chain of $N=20$ atoms (red curve).
For $\Delta-J'=0$, \textit{i.e.}, when the probe field is in resonance with the (shifted) $\ket{e}-\ket{g}$ transition, the transmittance is perfect. However, total transparency is only exactly achieved at this precise frequency, decreasing with the detuning from resonance. The medium can be considered roughly transparent within a small window of bandwidth $\sim\Delta_\text{EIT}$, for which the transmittance $T>1/e$. The scaling of the bandwidth-delay product with the number of atoms is shown in Figure~\ref{nFig4}(b). The numerical results (red dots) are obtained by solving Eq.~\eqref{eqEITstat}, then calculating the transmission as a function of the atom-probe detuning, and finally numerically finding the values of $\Delta$ where the transmittance drops to $1/e$. The calculations follow perfectly the simple scaling $\mathcal{P}=\sqrt{D}$ derived above (continuous red line). We should note that the usual definition of $D$ --in terms of exponential reduction of transmittance on resonance-- does not apply any more to EIT. However, we maintain the definition $D=2N\Gamma_{\rm 1D}/\Gamma'$, as it represents a physical resource.
As the final step in our summary, we turn our attention to the problem of the efficiency of an EIT-based quantum memory. Qualitatively, the large bandwidth-delay product associated with an optically dense ensemble enables an incident pulse to become spatially compressed and localized completely within the ensemble, while propagating with a reduced group velocity $v_\text{g} \ll c$. The slow group velocity is associated with the photon mixing strongly with a collective spin excitation $\hat{\sigma}_{sg}$, to form a ``dark-state'' polariton. Once the pulse is completely inside, the pump field can be adiabatically decreased to zero ($\Omega_c=0$), in which case $v_\text{g}\rightarrow 0$ and the pulse becomes stored, while simultaneously the polariton becomes a pure spin excitation \cite{L03,FIM05}. This process of photon mapping can be reversed by ramping up the control field intensity at a later time, which allows for an ``on demand" retrieval of the stored photon. Gorshkov and co-workers demonstrated that, due to time reversal symmetry, the optimal efficiency of photon storage is identical to that of photon retrieval \cite{GAF07}. Therefore, our discussion will focus on the latter. Neglecting collective emission into non-guided modes, Gorshkov \textit{et al.} predicted that any smooth spin wave fitting inside the atomic medium should be retrieved with error $\varepsilon\sim1/D$ \cite{GAF07,GAL07}. The reason for such scaling is that the optical depth sets the branching ratio between emitting the photon into the desired channel (the guided mode of the fiber) and into the undesired reservoir (free space).
\begin{figure}
\centerline{\includegraphics[width=\linewidth]{nFig5}}
\caption{Selectively radiant states of the $s$-branch. \textbf{(a)} Guided [$\Gamma_{\rm 1D}(k_z)$, light blue] and non-guided [$\Gamma'(k_z)$, green] decay rates of the single-excitation eigenstates of $\mathcal{H}_\text{1D}+\mathcal{H}'+\mathcal{H}_\text{c}$ vs the dominant wave vector $k_z$ of each eigenstate, for a chain of $N=200$ atoms coupled to the fiber. The control field Rabi frequency is $\Omega_c=4\Gamma_0$, and the plot is restricted to eigenstates that consist mostly of population in the $s$ states (the s-branch). The gray shaded area represents the region within the light line, the dashed lines show the guided mode wave vector $\pmk_{\rm 1D}$, and the color lines are guides to the eye. \textbf{(b)} Scaling of the ratio $\Gamma_{\rm 1D}(k_z)/\Gamma'(k_z)$ with the number of atoms, at the wave vector $k_z$ where it is maximum. The dots are numerical results, and the curve represents the best quadratic fit, $\text{max}\{\Gamma_{\rm 1D}(k_z)/\Gamma'(k_z)\}\simeq 0.0053 N^2$.}
\label{nFig5}
\end{figure}
In order to demonstrate that our calculations match the previously known results, we initialize a single-excitation spin wave of the form
\begin{align}\label{spinwave}
\ket{\psi(t=0)}=\mathcal{N}\sum_{j=1}^N \,j\, \,e^{\ii k_{\rm 1D} z_j} \ket{s_j},
\end{align}
where $\mathcal{N}$ is a normalization constant, and the phase $e^{\ii k_{\rm 1D} z_j}$ guarantees retrieval of this excitation as a photon in the right-propagating guided mode, Eq.~\eqref{rightf}. This peculiar-shaped spin wave (in particular, the relative population of atom $j$ grows like $j^2$) presents a balance between the pulse being sufficiently smooth (such that its wave vector components fit within the transparency window), and the majority of the pulse sitting at the forward end of the medium, such that it does not accumulate propagation losses over a large distance. In the limit of large optical depth, such a polariton is predicted to be of the optimal shape to yield maximal retrieval efficiency (in particular, $\varepsilon\simeq 5.8/D$ \cite{GAL07}). At $t=0$ we switch on the control field and let the atomic wave function evolve under the effective Hamiltonian $\mathcal{H}_\text{1D}+\mathcal{H}'_\text{indep}+\mathcal{H}_\text{c}$ until no excitation is left in the atomic chain (having been emitted into the waveguide or free space). We calculate the infidelity in the photon retrieval in two different manners, which yield identical results. The first method consists in integrating over time the radiative emission into non-guided photonic modes. The error is thus $\varepsilon=\int_0^{\infty}\, dt\,\kappa'_\text{indep}(t)$, where $\kappa'_\text{indep}(t)=-(2/\hbar)\,\text{Im}\braket{\mathcal{H}'_\text{indep}}=\Gamma'\braket{\sum_{j=1}^N\hat{\sigma}_{ee}^j}$. The second way to calculate the infidelity is to realize that a successful retrieval occurs whenever the photon is emitted to the guided mode of the fiber. Then, the error is $\varepsilon=1-\int_0^{\infty}\, dt\,\kappa_\text{1D}(t)$, where the time-dependent decay rate into the guided mode is $\kappa_\text{1D}(t)=-(2/\hbar)\,\text{Im}\braket{\mathcal{H}_\text{1D}}$. Technically, the efficiency should be calculated only accounting for emission into the preferred (right-going) direction of the fiber, using Eq.~\eqref{rightf}. We have checked that this gives nearly an identical answer, as emission in the left-going direction is negligible. The scaling of the retrieval infidelity with the number of atoms is shown by red circles in Fig.~\ref{nFig4}(c). The numerical results agree very well with the expected scaling ($\varepsilon\simeq 5.8/D$ \cite{GAL07} \footnote{In Ref. \cite{GAL07} the scaling is $\varepsilon\simeq 2.9/D$, due to a factor of 2 difference in their definition of $D$.}, red line) for large number of atoms. In principle, the shape of the outgoing photon can be further tailored via a time-dependent control field, but we will not treat that case here.
Now that we have reviewed the basic parameters characterizing an EIT medium as well as the fidelity of a quantum memory, we analyze how collective emission into non-guided modes modifies the relevant figures of merit. In this case, the system evolves under the effective Hamiltonian $\mathcal{H}_\text{tot}=\mathcal{H}_\text{1D}+\mathcal{H}'+\mathcal{H}_\text{drive}+\mathcal{H}_\text{c}$, where now collective emission is taken into account through the $\mathcal{H}'$ term (instead of the previous $\mathcal{H}'_\text{indep}$). Before proceeding to the calculation of the optical properties, we would like to discuss the decay rates of the eigenstates of the system without guided-mode driving, \textit{i.e.}, of $\mathcal{H}_\text{1D}+\mathcal{H}'+\mathcal{H}_\text{c}$. Due to the presence of the $s$-states, the number of eigenstates in the single excitation subspace is $2N$. If the population of the $s$-states of a given eigenstate is larger than that of the $e$-states, we say that this eigenstate belongs to the ``$s$-branch", and vice versa. For any finite control field, there is mixing between the $e$- and $s$-branches, meaning that the eigenstates do not purely consist of $\ket{e}$ or $\ket{s}$ states.
In Fig.~\ref{nFig5}(a) we show the guided [$\Gamma_{\rm 1D}(k_z)=-(2/\hbar)\,\text{Im}\braket{\mathcal{H}_\text{1D}}$] and non-guided [$\Gamma'(k_z)=-(2/\hbar)\,\text{Im}\braket{\mathcal{H}'}$] decay rates of the numerically calculated eigenstates that belong to the $s$-branch, for a fixed number of atoms $N=200$. As in Section III A, we have performed a finite Fourier transform to associate an effective wave vector $k_z$ to each of the atomic spin eigenstates. As expected, the non-guided decay rates are negligible when the dominant wave vector $k_z$ lies beyond the light line. On the contrary, the guided decay rates peak strongly outside the light line, at $k_z=\pm k_{\rm 1D}$. It can also be seen that these same states experience a decay rate into free space of $\Gamma'(k_z)/\Gamma_0 \ll 1$, and are thus the ``selectively radiant" states that we previously anticipated. Some of the eigenstates with $|k_z|<k_0$ have a non-zero $\Gamma_{\rm 1D}(k_z)$ decay rate into the guided mode. This occurs because the eigenstates are not purely Bloch waves with a perfectly-determined $k_z$, but instead can have some finite contributions from all $k_z$. As a technical note, we remark that only when the chain of atoms is infinite can $\mathcal{H}_\text{1D}$ and $\mathcal{H}'$ be simultaneously diagonalized. For any finite number of atoms, the eigenstates of $\mathcal{H}_\text{1D}+\mathcal{H}'+\mathcal{H}_c$ are not simultaneously eigenstates of its guided and non-guided parts.
One can also consider the behavior of the selectively radiant states, as a function of atom number. In particular, of interest is the maximum possible branching ratio $\Gamma_{\rm 1D} (k_z)/\Gamma'(k_z)$ of all the eigenstates, as a function of $N$. We plot this quantity in Fig.~\ref{nFig5}(b), where we find an approximate scaling of max$\{\Gamma_{\rm 1D}(k_z)/\Gamma'(k_z)\}\propto N^2$. We find that this scaling is in fact independent of the magnitude of the control field, and is in contrast to the $\sim N$ scaling in the case of the independent emission model. We later show that this same scaling manifests itself in the photon storage/retrieval error probabilities.
Let's now calculate EIT transmittance spectra. Under the same conditions as before (low saturation, uniform control field), the evolution equations for the state amplitudes are found to be
\begin{subequations}\label{eqEIT2}
\begin{align}\nonumber
\dot{c}_e^{\,i}(t)&=\ii\Delta c_e^{\,i}(t)+\ii \Omega_i e^{\iik_{\rm 1D} z_i}-\frac{\Gamma_{\rm 1D}}{2}\sum_{j=1}^N e^{\iik_{\rm 1D} |z_i-z_j|}\,c_e^{\,j}(t)\\
&+\ii\Omega_cc_s^{\,i}(t)+\ii\frac{3\pi\Gamma_0}{k_0}\sum_{j=1}^N G_{\rho\rho}'(\rb_i,\rb_j,\omega_\text{p})\,c_e^{\,j}(t),\\
\dot{c}_s^{\,i}(t)&=\ii\Delta_{\rm s}c_s^{\,i}(t)+\ii\Omega_c\,c_e^{\,i}(t).
\end{align}
\end{subequations}
While analytical approximations are not as readily obtained, the numerical procedures follow exactly as presented for the case of independent emission. The blue curve in Fig.~\ref{nFig4}(a) shows how the transmittance spectrum is modified by selective radiance. The first noticeable consequence of collective suppression of the emission into non-guided modes is that the transparency window becomes wider, as expected if the loss becomes smaller. This is further confirmed in Fig.~\ref{nFig4}(b), which displays a linear scaling of the bandwidth-delay product with the atom number, in contrast to the conventional square root dependence. As mentioned before, such a scaling is characteristic of a system without losses \cite{SPP05}. This scaling, along with the conclusion in Sec. III that suppression of emission into free space can occur for low densities of excitations, suggests that it might be possible to store a number of photons in an atomic medium that scales linearly with atom number (in contrast to the $\sim\sqrt{D}$ scaling within the independent emission model).
Finally, Fig.~\ref{nFig4}(c) shows the improvement in the infidelity of retrieval of the spin wave given by Eq.~\eqref{spinwave}. The error is now calculated including collective emission as $\varepsilon=\int_0^{\infty}\, dt\,\kappa'(t)$, where
\begin{align}\label{kappa}
\kappa'(t)=-\frac{2}{\hbar}\,\text{Im}\braket{\mathcal{H}'}.
\end{align}
Again, this error matches the one calculated by taking into account the component of the photon that is released into the guided mode, \textit{i.e.}, $\varepsilon=1-\int_0^{\infty}\, dt\,\kappa_\text{1D}(t)$, with $\kappa_\text{1D}(t)=-(2/\hbar)\,\text{Im}\braket{\mathcal{H}_\text{1D}}$. As before, we have checked that emission in the left-going direction is negligible. We find that by exploiting collective emission, the error decreases with atom number like $\varepsilon\propto 1/N^2$. This result is consistent with the scaling of branching ratios for the most selectively radiant eigenstates, previously plotted in Fig.~\ref{nFig5}(b). Moreover, by varying the radial positions of the atoms over a limited range, thus modifying the ratio $\Gamma_{\rm 1D}/\Gamma'$, we are able to separate the contributions of the number of atoms and the optical depth to the infidelity. We obtain $\varepsilon\simeq 15/(ND)$, where the numerical prefactor is not necessarily universal, as it probably depends on the fiber properties.
\begin{figure*}
\centerline{\includegraphics[width=\linewidth]{nFig6}}
\caption{Retrieval of a Gaussian spin wave given by Eq.~\eqref{gaussianwave}, under a spatially uniform control field of Rabi frequency $\Omega_c=0.1\Gamma_0$, and within the collective emission model. \textbf{(a)} Evolution of the population in the $\ket{s}$ states (upper plot) and $\ket{e}$ states (lower plot) as a function of position and at several selected points in time $\tau$, for a chain of $N=200$ atoms, in arbitrary units. The time is normalized such that at $\tau=1$ all the spin population has completely decayed. \textbf{(b)} Instantaneous rate of photon scattering into free space $\kappa'(\tau)$. The solid line is a numerical calculation based upon Eq.~\eqref{kappa}, while the dashed line represents an estimate based upon taking a Fourier decomposition of the spin amplitude $c_e(z_j,t)$ and weighting each component by a wave vector-dependent decay rate. A large instantaneous scattering rate occurs when a large excited-state population is found at the end of the system [for example, at times $\tau=0.07$ and $\tau=0.9$ in figure (a)]. \textbf{(c)} Excited-state population $|c_e(k_z,\tau)|^2$ of the different wave vector components of the spin wave, for different evolution times [corresponding to the snapshots in (a)]. Only the region inside the light line is shown. \textbf{(d)} Scaling of the retrieval loss with the atom number $N$. The blue dots show the numerical calculation, whereas the blue line is the best fit to them and represents $\varepsilon=4.1/N$. The infidelities for the initial spin wave of Eq.~\eqref{spinwave} within the independent and collective emission models are shown by the dashed and dotted lines, respectively.} \label{nFig6}
\end{figure*}
An interesting question is why the error of photon storage improves `only' by a factor of $N$ (from $1/N$ to $1/N^2$). In particular, given that single excitations in a free-space chain can experience a suppression in the emission rate of up to $1/N^3$, one might have expected a greater suppression of errors of up to $1/N^4$ in photon storage. An initial -- but somewhat erroneous -- guess would be to attribute this ``bad scaling" to an unfavorable spatial profile of the initial spin wave. Perhaps surprisingly, although EIT nominally matches the effective guided mode indices of the bare fiber and the composite system of fiber and atomic chain, we show in the next subsection that the slight impedance mismatch away from perfect resonance $\Delta=J'$ is still responsible for the majority of scattering losses into free space. We thus present an improved impedance matching scheme, which allows for exponential improvement with $N$ of the quantum memory fidelity.
\subsection{Quantum memory with exponential fidelity}
The importance of residual impedance mismatch can be seen in a simple example, where one considers an initial Gaussian spin-wave profile,
\begin{align}\label{gaussianwave}
\ket{\psi(t=0)}=\mathcal{N}\sum_{j=1}^N\,e^{\ii k_{\rm 1D} z_j} \,e^{-(z_j-z_c)^2/2\sigma^2} \ket{s_j},
\end{align}
and investigates the dynamics of the retrieval process more carefully. In the above expression, $\mathcal{N}$ is a normalization constant, $z_c=(N-1)d/2$ is the center of the atom chain, and $\sigma=\sqrt{N} d$ is the standard deviation of the Gaussian spin wave. Figure~\ref{nFig6}(a) shows the evolution of the spin excitation at different times $\tau$, for a chain of $N=200$ atoms. Here, to aid in visualization, we have defined a re-scaled dimensionless time $\tau \in [0,1]$, where $\tau$ represents the total amount of atomic population that has decayed (\textit{i.e.}, at $\tau$=1 the spin wave has fully decayed and the photon has been completely released). We have plotted not only the population in the $\ket{s}$-state through the ensemble, which essentially matches population of the dark-state polariton, but also the excited state population, which is ultimately responsible for any emission into free space.
\begin{figure}
\centerline{\includegraphics[width=\linewidth]{nFig6b}}
\caption{Schematic of the spatial profile of the control field $\Omega_c(z_j)\sim\sqrt{N/(N+1-j)}$ (green curve, arbitrary units) for a chain of $N=200$ atoms. The initial spin wave, given by Eq.~\eqref{gaussianwave}, is overlaid in blue.} \label{nFig6b}
\end{figure}
\begin{figure*}
\centerline{\includegraphics[width=\linewidth]{nFig7}}
\caption{The same as Fig.~\ref{nFig6}, but for a spatially-varying control field of the form $\Omega_c(z_j)=0.005\sqrt{N/(N+1-j)}\Gamma_0$. In (a), the dotted lines show the analytical model. In (d), the blue line is a guide to the eye, and follows $\varepsilon=0.15 e^{-N/23}$.} \label{nFig7}
\end{figure*}
For a spatially uniform control field, and for a theory of EIT within a uniform medium (\textit{i.e.}, where the atomic density is treated as smooth rather than discrete points), it can readily be shown \cite{FL02} that the excited state population is proportional to $|\partial_z \hat{\sigma}_{gs}(z)|^2$ (also see Appendix~\ref{AppE}), a result that also agrees well with our numerical results. This excited state population is necessarily associated with a pulse of finite extent or bandwidth, and in complementary ways reflects the fact that perfect transparency in EIT only occurs at a single frequency, or that there are non-adiabatic corrections to the formation of a dark-state polariton. At certain times, such as at $\tau=0.07$ or $\tau=0.9$, the excited-state spin wave has a large amplitude at the edge of the atomic chain. Most of the error on retrieval occurs at these times, as can be seen in Fig.~\ref{nFig6}(b), which shows the instantaneous loss $\kappa'(\tau)$. Here, $\kappa'(\tau)$ is re-scaled as well, so that its integral provides the total loss, $\int_{0}^{1} d\tau \kappa'(\tau) = \varepsilon$.
A plausible cause of this behavior is the discontinuity of the excited state population at the system's end, which is associated with a large number of wave vector components $k_z$ that lie within the light line and couple to free space. To further confirm this intuition, we have developed a model for the time-dependent loss based upon the Fourier decomposition of the spatial profile of the excited state amplitude. At every time of the evolution, we calculate $c_e(k_z,\tau)$ by doing a finite Fourier transform of the excited state amplitudes, $c_e(z_j,\tau)\equiv c_e^{\,j}(\tau)$. Then, we find the Fourier-based instantaneous loss as $\tilde{\kappa}'(\tau)=(d/2\pi)\int_{-k_0}^{k_0} dk_z\,\Gamma'(k_z) |c_e(k_z,\tau)|^2$, where $\Gamma'(k_z)$ is obtained from classifying the decay rates of the eigenstates of $\mathcal{H}'$ according to their dominant wave vector. This calculation, represented by the dashed line in Fig.~\ref{nFig6}(b), shows good agreement with the numerics. Figure~\ref{nFig6}(c) depicts the components of $|c_e(k_z)|^2$ inside the light line for different times [corresponding to the snapshots of Fig.~\ref{nFig6}(a)]. For initial times (purple curve), the wave function has minimal population within the light line, suggesting it propagates with little loss down the atomic chain. The population drastically increases as the pulse hits the end of the chain (brown curve). A large population of wave vector components within the light line is correlated with increased instantaneous loss and sharp features in the profile of the excited state population at the system edge. Finally, Fig.~\ref{nFig6}(d) displays the scaling of the infidelity in photon retrieval with the atom number. The scaling is poor ($\varepsilon=4.1/N$), as large losses occur when the polariton hits the ends of the atomic chain.
We now describe how to smoothly reduce the excited state population at the end of the chain, by introducing a spatially-dependent control field. Heuristically, the idea is to increase the control field at the ends of the chain, as shown in Fig.~\ref{nFig6b}. As the EIT bandwidth is proportional to the control field intensity [see Eq.~\eqref{deltaeit}], the atomic medium becomes more transparent at the edges, where the excited state population is reduced. One can develop an effective continuum wave equation to predict the evolution of the populations in $\ket{s}$ and $\ket{e}$ during the retrieval process, in the presence of a spatially dependent control field (see Appendix~\ref{AppE}) \cite{FL02}. Similar to the case of a uniform control field presented earlier, in principle the scattering loss can then be estimated and minimized from these populations. This optimization process seems quite challenging in practice, however, as it depends on the initial spin wave, the control field profile, and on the integral of momentum components over the entire history of evolution. We will not do such an optimization here, but instead show that rather simple choices can already lead to significant improvements over the infidelity scalings that we found in the previous subsection.
For the initial spin wave of Eq.~\eqref{gaussianwave}, Fig.~\ref{nFig7}(a) shows the evolution of the e- and s-states populations of a chain of $N=200$ atoms for the spatially-dependent control field $\Omega_c(z_j)=\Omega_c^{(0)}\sqrt{N/(N+1-j)}$, which exhibits a rapid increase at the right edge of the chain [see Fig.~\ref{nFig6b}]. Since the control field is switched on suddenly in our simulations, its magnitude is taken to be very small ($\Omega_{c}^{(0)}=0.005\Gamma_0$) to minimize rapid, non-adiabatic accelerations of the spin wave that would artificially increase the losses at initial times. Both the $\ket{e}$- and $\ket{s}$-state populations exhibit smooth profiles at every time of the evolution, and in particular, one sees that the excited state population smoothly vanishes at the edge of the system. The dashed lines in the plots show the results from the analytical model developed in Appendix~\ref{AppE}, which agree well with the numerics. Both the instantaneous loss and the amount of spin-wave population lying within the light line are several orders of magnitude smaller than in the case of a uniform control field, as can be seen in Figs.~\ref{nFig7}(b) and (c), respectively. Moreover, the excited state population within the light line does not significantly increase from its initial values. Finally, Fig.~\ref{nFig7}(d) shows the exponential decrease of the retrieval infidelity $\varepsilon$ as a function of the atom number $N$ for this given profile. We anticipate that optimized initial spin waves and control field profiles will result in a much steeper exponential scaling. Nonetheless, we have demonstrated that a fairly trivial selection of those settings already exponentially improves previously-known bounds for photon storage.
\section{Physical implementations and possible challenges}
Having analyzed the physics of both subradiance and selective radiance, we devote this section to discussing suitable experimental platforms as well as the challenges that might be encountered to observe this physics.
\subsection{Physical implementations}
To potentially observe the physics that we described in Sec. III requires atoms to be regularly trapped, forming an ordered lattice. Moreover, among all the possible collective atomic states, one should be able to access the subradiant manifold. We shall start our discussion with possible physical platforms. As we have demonstrated, the minimal distance at which subradiant states appear depends on the dimensionality of the atomic array ($\lambda_0$ in 2D, and $\lambda_0/2$ in 1D). Standard free-space optical lattices \cite{B05} can achieve lattice constants of $d\sim \lambda_0/2$, and quantum gas microscopes \cite{BGP09} are able to generate single 2D arrays. In such systems, both bosonic \cite{BPT10} and fermionic \cite{GPM16} Mott insulator phases -- where the number of atoms per site can be limited to one -- have been realized.
Very recently, several experimental groups have built almost defect-free 1D \cite{EBK16} and 2D \cite{LLK15,BLL16} lattices in an atom-by-atom manner using optical tweezer arrays. While the inter-atomic distance achieved so far is still larger than the free space wavelength, due to the problem of interference between the tweezers at close distances, it could be possible that further improvements enable sub-wavelength distances to be reached. It might also be possible to employ a transition with a shorter wavelength for the trapping scheme, and use another of longer wavelength to explore subradiant phenomena. Finally, periodically patterned 1D or 2D dielectric structures can readily yield sub-wavelength trapping potentials, with the periodicity of the structure itself \cite{HMC13,DHC15}. While cold atoms can now routinely be trapped near dielectric structures \cite{VRS10,TTL13,GHY14,GHH15,HGA16}, the filling fractions remain quite low and new approaches (such as integration with tweezer arrays) must be developed to achieve near-perfect filling.
Overcoming the second requirement, that of exciting the subradiant manifold efficiently, is not trivial. As subradiant states are characterized by a wave vector that lies beyond the light line, they do not naturally couple to a laser beam that propagates through free space. There are several options to overcome this hurdle. An already-suggested possibility \cite{POR15} is to map superradiant states, which are easy to excite, to subradiant ones via magnetic field gradients. Specifically, a laser can efficiently excite a spin wave $\sim \sum_{j} e^{\ii k_z z_j} \ket{e_j}$ whose wave vector $k_z$ lies within the light line. In the case that the excited state is magnetic-field sensitive, a field gradient would then imprint a spatially dependent phase shift $\ket{e_j}\rightarrow e^{\ii (\beta t)z_j}\ket{e_j}$ in time, which then could allow the wave vector $k_z \rightarrow k_z+\beta t$ to be mapped outside of the light line. In 1D chains in free space, one might exploit the fact that the emission of subradiant states occurs primarily from the ends, and in a pattern that can be collected reasonably well with conventional optics. Note that the question of efficient excitation does not come up with selectively radiant states, as by definition they are well-coupled to a mode of interest.
Finally, we would like to stress that the exploration of both subradiance and selective radiance is not restricted to atoms. Molecules \cite{FTH14} and solid-state emitters should also exhibit these properties, although they pose a different set of challenges. The ability of deterministically placing quantum dots \cite{LSJ08,YTB15,LMS15}, rare earth ions \cite{KXK16,JGW16}, and color centers \cite{SES16} has significantly improved in the past years, thus putting ordered arrays within reach. However, one of the main appeals of atoms is that they are identical to each other and that their decay is purely radiative. An open question is how to translate these features into the domain of solid-state emitters, as it would require high homogeneity among them as well as a large emission into the zero-phonon line (minimizing non-radiative losses).
\subsection{Atomic level structure}
There are a number of potential imperfections that could limit the observation of subradiance and selective radiance, and the performance of protocols that exploit them. A number of these imperfections are conceptually clear (if not necessarily straightforward to analyze), such as disorder in atomic positions, classical and quantum motion, dephasing, and imperfect site filling. A more thorough investigation of these effects will be left to future work.
Here, we would like to discuss a more subtle issue, related to the complications associated with multi-level atomic structure. For most of this manuscript, we have assumed that atoms are two-level systems, with a single ground state and excited state. For atoms with hyperfine structure (and thus a ground state manifold), an effective two-level system is often achieved by exploiting a cycling transition \cite{MV99}, where an excited state of maximum angular momentum can only decay back into a single ground state, also of maximum angular momentum. Such a transition only responds to pure circularly polarized light. In the case of multiple atoms, an important issue is that light scattered from one atom does not display circular polarization globally in space. Thus, for example, the resulting dipole-dipole interactions can potentially drive other atoms to excited states outside of the cycling transition. This can be avoided in the specific case of a 1D chain (where the re-scattered field has the same polarization along the axis of the chain), but not in general.
Another possibility to avoid the full complexity of hyperfine structure is to use atoms without nuclear spin, such as bosonic Ytterbium or Strontium \cite{XLH03,SSK13}. In this case, there is a single ground state but three excited states with orthogonal dipole matrix elements (giving an isotropic optical response to the atoms). Then, one can exploit the fact that in 1D arrays, dipole-dipole interactions involving different excited states decouple from each other, to effectively yield two-level physics (similar to the case of circular polarization described above). In 2D arrays, the transitions involving a dipole matrix element out of the plane decouple, while the two in-plane transitions can hybridize, and calculating the band structure for an infinite system involves diagonalizing a $2\times 2$ matrix associated with the Fourier transform of the in-plane components of the Green's function, $\tilde{G}_{0,\alpha\beta}(\textbf{k})$, with $\{\alpha,\beta\}=\{y,z\}$. This solution qualitatively maintains the same properties as the two-level case analyzed in Sec. III [for example, see the discussion surrounding Eq.~\eqref{rate2D}].
In the presence of hyperfine structure, and excluding the special case of a 1D array described previously, the complication with regard to subradiance can be understood with the following simple example. Suppose that atoms are initialized in a single ground state $\ket{g}$, from which a single-excitation spin wave of the form $\ket{\psi}\sim \sum_j e^{\ii kz_j} \ket{e_j}$ is somehow generated. If the wave vector $k$ is beyond the light line, then as argued in Sec. III, collective dissipative interactions [such as those encoded in the $\hat{\sigma}_{eg}^{i}\hge^{j}$ term of Eq.~\eqref{ham}] will suppress emission of an excited state back into $\ket{g}$ through destructive interference. However, dipole-dipole interactions will also generally exist between that excited state and any other ground state $\ket{s}$ connected by a dipole-allowed transition, \textit{e.g.}, of the form $\hat{\sigma}_{es}^{i}\hat{\sigma}_{se}^{j}$. Since the initially prepared spin wave $\ket{\psi}$ does not contain any population in $\ket{s}$, there is no interference that prevents decay via this channel, and thus the spin wave would experience a decay rate into $\ket{s}$ equal to that of a single, isolated atom excited to $\ket{e}$. Interesting recent work suggests that it is possible to encode subradiance in a more complex initial state beyond the simple product state $\ket{g}^{\otimes N}$ \cite{HKO17}, which would be worth exploring further.
Within the context of the enhancement of EIT-based storage protocols through collective emission, studied in Sec. IV, this implies that the state $\ket{s}$ cannot be another state in the ground-state manifold that is directly connected to $\ket{e}$ by a dipole-allowed transition. Various possibilities to implement EIT and retain the desired collective interference effects include the use of a ladder scheme, with the state $\ket{s}$ being a long-lived excited state (\textit{e.g.}, a Rydberg level), or to use a state $\ket{s}$ in the ground-state manifold that is connected only through a two-photon transition \cite{ECZ97,PC08}.
\section{Summary and outlook}
In summary, we have shown that subradiant states acquire an elegant interpretation in 1D and 2D atomic arrays, in terms of optically guided modes whose decay rate is only limited by the system boundaries. We have provided a first glimpse into the nature of subradiance for multiple excitations, and introduced a new concept of selective radiance that should enable the construction of more efficient atom-light interfaces. As a concrete example, we have constructed a protocol for quantum memories for light using selectively radiant states in an optical nanofiber, whose infidelity decreases with atom number at a rate exponentially better than previously known bounds.
Even though memories are a very relevant quantum technology, the improvement in their performance is just an example of the bountiful possibilities spawned by subradiance and selective radiance. We anticipate that exploiting these phenomena could yield new error bounds and protocols for many applications of interest, ranging from nonlinear optics to metrology. At the same time, the nature of subradiance for multiple excitations or internal states could itself constitute a rich new many-body problem.
\vspace{10pt}
\textbf{Acknowledgments.}-- We are grateful to C. Regal, A. M. Rey, A. Gorshkov, E. Polzik, S. Yelin, M. Lukin, H. Ritsch, E. Shahmoon, and J. Muniz for stimulating discussions. H.J.K. funding is provided by the AFOSR Quantum Many-Body Physics with Photons and QuMPASS MURI, NSF Grant PHY-1205729, the Office of Naval Research (ONR) Award N00014-16-1-2399; the ONR QOMAND MURI; and the IQIM, an NSF Physics Frontiers Center. A.A.-G. was supported by the IQIM Postdoctoral Fellowship and the Global Marie Curie Fellowship LANTERN (655701). D.E.C. acknowledges support from Fundaci\'{o} Privada Cellex, Marie Curie CIG ATOMNANO, Spanish MINECO Severo Ochoa Programme SEV-2015-0522, MINECO Plan Nacional Grant CANS, CERCA Programme/Generalitat de Catalunya, and ERC Starting Grant FOQAL.
|
\section{Introduction}\label{Sec_Introduction}
\setcounter{equation}{0}
The notion of discrete-time quantum walks appears in numerous contexts
\cite{ADZ93,ABNVW01,GZ88,Gud88,Mey96,Wat01}. Among them, Gudder \cite{Gud88}, Meyer
\cite{Mey96}, and Ambainis et al. \cite{ABNVW01} introduced one-dimensional quantum
walks as a quantum mechanical counterpart of classical random walks. Nowadays, these
quantum walks and their generalisations have been physically implemented in various
ways \cite{MW14}. Versatile applications of quantum walks can be found in
\cite{COB15,IMSY15,Por13,Ven12} and references therein.
Recently, because of the controllability of their parameters, discrete-time quantum
walks have attracted attention as promising candidates to realise topological
insulators. In a series of papers \cite{KBFRBKADW12,KRBD10}, Kitagawa et al. have
shown that one and two dimensional quantum walks possess topological phases, and they
experimentally observed a topologically protected bound state between two distinct
phases. See \cite{Kit12} for an introductory review on the topological phenomena in
quantum walks. Motivated by these studies, Endo et al. \cite{EKH15} (see also
\cite{EEKST15,EEKST16}) have performed a thorough analysis of the asymptotic behaviour
of two-phase quantum walks, whose evolution is given by a unitary operator
$U_{\rm TP}=SC$ with $S$ a shift operator and $C$ a coin operator defined as a
multiplication by unitary matrices $C(x)\in\U(2)$, $x\in\Z$. When $C(x)$ is given by
\begin{equation} \label{eq_two_phase}
C(x)=
\begin{cases}
\frac1{\sqrt2}
\begin{pmatrix}
1 & \e^{i\sigma_+}\\
\e^{-i\sigma_+} & -1
\end{pmatrix}
& \hbox{if $x\ge0$}\bigskip\\
\frac1{\sqrt2}
\begin{pmatrix}
1 & \e^{i\sigma_-}\\
\e^{-i\sigma_-} & -1
\end{pmatrix}
& \hbox{if $x\le-1$}
\end{cases}
\end{equation}
with $\sigma_\pm\in[0,2\pi)$, the two-phase quantum walk with evolution operator
$U_{\rm TP}$ is called complete two-phase quantum walk, and when $C(x)$ satisfies the
alternative condition at $0$
\begin{equation}\label{eq_defect}
C(0)=\begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix},
\end{equation}
the quantum walk is called two-phase quantum walk with one defect. In
\cite{EEKST16,EKH15}, Endo et al. have proved a weak limit theorem \cite{Kon02,Kon05}
similar to the de Moivre-Laplace theorem (or the Central limit theorem) for random
walks, which describes the asymptotic behaviours of the two-phase quantum walk.
In the present paper and the companion paper \cite{RST_2}, we consider one-dimensional
quantum walks $U=SC$ with a coin operator $C$ exhibiting an anisotropic behaviour at
infinity, with short-range convergence to the asymptotics. Namely, we assume that
there exist matrices $C_{\ell},C_{\rm r}\in\U(2)$ and positive constants
$\varepsilon_\ell,\varepsilon_{\rm r}>0$ such that
\begin{equation}\label{eq_aniso_coin}
C(x)=
\begin{cases}
C_{\ell}+O\big(|x|^{-1-{\varepsilon_\ell}}\big) & \hbox{as $x\to-\infty$}\\
C_{\rm r}+O\big(|x|^{-1-{\varepsilon_r}}\big) & \hbox{as $x\to\infty$}.
\end{cases}
\end{equation}
We call this type of quantum walks quantum walks with an anisotropic coin or simply
anisotropic quantum walks. They include two-phase quantum walks with coins defined by
\eqref{eq_two_phase} and \eqref{eq_defect} and one-defect models
\cite{CGMV12,Kon10,KLS13,WLKGGB12} as special cases. In the case
$C_0:=C_{\ell}=C_{\rm r}$ and $\varepsilon_0:=\varepsilon_{\ell}=\varepsilon_{\rm r}$,
quantum walks with an anisotropic coin reduce to one-dimensional quantum walks with a
position dependent coin
$$
C(x)=C_0+O\big(|x|^{-1-\varepsilon_0}\big),\quad |x|\to\infty,
$$
for which the absence of the singular continuous spectrum was proved in \cite{ABJ15}
and for which a weak limit theorem was derived in \cite{Suz16}.
Quantum walks with an anisotropic coin are also related to Kitagawa's topological
quantum walk model called a split-step quantum walk \cite{Kit12,KBFRBKADW12,KRBD10}.
Indeed, if $R(\theta)\in\U(2)$ is a rotation matrix with rotation angle $\theta/2$,
$R(\Theta_j)$ the multiplication operator by
$R\big(\theta_j(\;\!\cdot\;\!)\big)\in\U(2)$ with $\theta_j:\Z\to[0,2\pi)$, $j=1,2$,
and $T_\downarrow,T_\uparrow$ shift operators satisfying
$S=T_\downarrow T_\uparrow=T_\uparrow T_\downarrow$, then the evolution operator of
the split-step quantum walk is defined as
$$
U_{\rm SS}(\theta_1,\theta_2)
:=T_\downarrow\;\!R(\Theta_2)\;\!T_\uparrow\;\!R(\Theta_1).
$$
Now, as mentioned in \cite{Kit12}, $U_{\rm SS}(\theta_1,\theta_2)$ is unitarily
equivalent to $T_\uparrow R(\Theta_1) T_\downarrow R(\Theta_2)$. Thus, our evolution
operator $U$ describes a quantum walk unitarily equivalent to the one described by
$U_{\rm SS}(\theta_1,\theta_2)$ if $\theta_1 \equiv 0$ and
$C(\;\!\cdot\;\!)=R\big(\theta_2(\;\!\cdot\;\!)\big)$ (see \cite{Ohn16,SS16} for the
definition of unitary equivalence between two quantum walks). In \cite{Kit12},
Kitagawa dealt with the case
$$
\theta_2(x)
:=\frac12(\theta_{2-}+\theta_{2+})+\frac12(\theta_{2+}-\theta_{2-})\tanh(x/3),
\quad \theta_{2-},\theta_{2+}\in[0,2\pi),~x\in\Z,
$$
which corresponds to taking the anisotropic coin \eqref{eq_aniso_coin} with
$C_{\ell}=R(\theta_{2-})$ and $C_{\rm r}=R(\theta_{2+})$, and which cannot be covered
by two-phase models.
The main goal of the present paper and \cite{RST_2} is to establish a weak limit
theorem for the the evolution operator $U$ of the quantum walk with an anisotropic
coin satisfying \eqref{eq_aniso_coin}. As put into evidence in \cite{Suz16}, in order
to establish a weak limit theorem one has to prove along the way the following two
important results:
\begin{itemize}
\item[(i)] the absence of singular continuous spectrum,
\item[(ii)] the existence of the asymptotic velocity.
\end{itemize}
In the present paper, we perform the spectral analysis of the evolution operator $U$
of quantum walks with an anisotropic coin. We determine the essential spectrum of $U$,
we show the existence of locally $U$-smooth operators, we prove the discreteness of
the eigenvalues of $U$ outside the thresholds, and we prove the absence of singular
continuous spectrum for $U$. In the companion paper \cite{RST_2}, we will develop the
scattering theory for the evolution operator $U$. We will prove the existence and the
completeness of wave operators for $U$ and a free evolution operator $U_0$, we will
show the existence of the asymptotic velocity for $U$, and we will finally establish a
weak limit theorem for $U$. Other interesting related topics such as the existence and
the robustness of a bound state localised around the phase boundary or a weak limit
theorem for the split-step quantum walk with $\theta_1\ne0$ are considered in
\cite{FFS1loc} and \cite{FFS1wlt}, respectively.
The rest of this paper is structured as follows. In Section 2, we give the precise
definition of the evolution operator $U$ for the quantum walk with an anisotropic coin
and we state our main results on the essential spectrum of $U$ (Theorem
\ref{thm_essential}), the locally $U$-smooth operators (Theorem
\ref{thm_U_smooth_walks}), and the eigen\-values and singular continuous spectrum of
$U$ (Theorem \ref{thm_spectrum_U_walks}). Section 3 is devoted to mathematical
preliminaries. Here we develop new commutator methods for unitary operators in a
two-Hilbert spaces setting, which are a key ingredient for our analysis and are of
independent interest. In Section 4, we prove our main theorems as an application of
the commutator methods developed in Section 3. In Subsection \ref{Sec_Mourre}, we
prove Theorem \ref{thm_essential} and we define in Lemma \ref{lemma_A_self} a
conjugate operator $A$ for the evolution operator $U$ built from conjugate operators
for the asymptotic evolution operators $U_\ell:=SC_\ell$ and $U_{\rm r}:=SC_{\rm r}$,
where $C_{\ell}$ and $C_{\rm r}$ are the constant coin matrices given in
\eqref{eq_aniso_coin}. Finally, in Subsection \ref{Sec_spectrum} we prove Theorems
\ref{thm_U_smooth_walks} and \ref{thm_spectrum_U_walks}.\\
\noindent
{\bf Acknowledgements.} The third author thanks the Graduate School of Mathematics of
Nagoya University for its warm hospitality in January-February 2017.
\section{Model and main results}\label{Sec_model}
\setcounter{equation}{0}
In this section, we give the definition of the model of anisotropic quantum walks that
we consider, we state our main results on quantum walks, and we present the main tools
we use for the proofs. These tools are results of independent interest on commutator
methods for unitary operators in a two-Hilbert spaces setting. The proofs of our
results on commutator methods are given in Section \ref{Sec_unitary} and the proofs of
our results on quantum walks are given in Section \ref{Sec_quantum}.
Let us consider the Hilbert space of square-summable $\C^2$-valued sequences
$$
\H:=\ell^2(\Z,\C^2)
=\left\{\Psi:\Z\to\C^2\mid\sum_{x\in\Z}\|\Psi(x)\|_2^2<\infty\right\},
$$
where $\|\cdot\|_2$ is the usual norm on $\C^2$. The evolution operator of the
one-dimensional quantum walk in $\H$ that we consider is defined by $U:=SC$, with $S$
a shift operator and $C$ a coin operator defined as follows. The operator $S$ is given
by
$$
(S\Psi)(x)
:=\begin{pmatrix}
\Psi^{(0)}(x+1)\\
\Psi^{(1)}(x-1)
\end{pmatrix},
\quad
\Psi
=\begin{pmatrix}
\Psi^{(0)}\\
\Psi^{(1)}
\end{pmatrix}\in\H,~x\in\Z,
$$
and the operator $C$ is given by
$$
(C\Psi)(x):=C(x)\Psi(x),\quad \Psi\in\H,~x\in\Z,~C(x)\in\U(2).
$$
In particular, the evolution operator $U$ is unitary in $\H$ since both $S$ and $C$
are unitary in $\H$.
Throughout the paper, we assume that the coin operator $C$ exhibits an anisotropic
behaviour at infinity. More precisely, we assume that $C$ converges with short-range
rate to two asymptotic coin operators, one on the left and one on the right in the
following way:
\begin{Assumption}[Short-range assumption]\label{ass_short}
There exist $C_\ell,C_{\rm r}\in\U(2)$, $\kappa_\ell,\kappa_{\rm r}>0$, and
$\varepsilon_\ell,\varepsilon_{\rm r}>0$ such that
\begin{align*}
&\big\|C(x)-C_\ell\big\|_{\B(\C^2)}
\le\kappa_\ell\;\!|x|^{-1-\varepsilon_\ell}\quad\hbox{if $x<0$}\\
&\big\|C(x)-C_{\rm r}\big\|_{\B(\C^2)}
\le\kappa_{\rm r}\;\!|x|^{-1-\varepsilon_{\rm r}}\quad\hbox{if $x>0$,}
\end{align*}
where the indexes $\ell$ and ${\rm r}$ stand for ``left" and ``right".
\end{Assumption}
This assumption provides us with two new unitary operators
\begin{equation}\label{eq_def_U_ell}
U_\ell:=SC_\ell\quad\hbox{and}\quad U_{\rm r}:=SC_{\rm r}
\end{equation}
describing the asymptotic behaviour of $U$ on the left and on the right. The precise
sense (from the scattering point of view) in which the operators $U_\ell$ and
$U_{\rm r}$ describe the asymptotic behaviour of $U$ on the left and on the right will
be given in \cite{RST_2}, and the spectral properties of $U_\ell$ and $U_{\rm r}$ are
determined in Section \ref{Sec_asymptotic}. Here, we just introduce the set
$$
\tau(U):=\partial\sigma(U_\ell)\cup\partial\sigma(U_{\rm r}),
$$
where $\partial\sigma(U_\ell),\partial\sigma(U_{\rm r})$ denote the boundaries in the
unit circle $\T:=\{z\in\C\mid|z|=1\}$ of the spectra
$\sigma(U_\ell),\sigma(U_{\rm r})$ of $U_\ell,U_{\rm r}$. In Section
\ref{Sec_asymptotic}, we show that $\tau(U)$ is finite and can be interpreted as the
set of thresholds in the spectrum of $U$.
Our main results on the operator $U$, proved in Sections \ref{Sec_Mourre} and
\ref{Sec_spectrum}, are the following three theorems on locally $U$-smooth operators
and on the structure of the spectrum of $U$. The symbols $\sigma_{\rm ess}(U)$,
$\sigma_{\rm p}(U)$ and $Q$ stand for the essential spectrum of $U$, the pure point
spectrum of $U$, and the position operator in $\H$, respectively (see \eqref{eq_def_Q}
for precise definition of $Q$).
\begin{Theorem}[Essential spectrum of $U$]\label{thm_essential}
One has $\sigma_{\rm ess}(U)=\sigma(U_\ell)\cup\sigma(U_{\rm r})$.
\end{Theorem}
\begin{Theorem}[$U$-smooth operators]\label{thm_U_smooth_walks}
Let $\G$ be an auxiliary Hilbert space and let $\Theta\subset\T$ be an open set with
closure $\overline\Theta\subset\T\setminus\tau(U)$. Then each operator $T\in\B(\H,\G)$
which extends continuously to an element of
$\B\big(\dom(\langle Q\rangle^{-s}),\G\big)$ for some $s>1/2$ is locally $U$-smooth on
$\Theta\setminus\sigma_{\rm p}(U)$.
\end{Theorem}
\begin{Theorem}[Spectrum of $U$]\label{thm_spectrum_U_walks}
For any closed set $\Theta\subset\T\setminus\tau(U)$, the operator $U$ has at most
finitely many eigenvalues in $\Theta$, each one of finite multiplicity, and $U$ has no
singular continuous spectrum in $\Theta$.
\end{Theorem}
To prove these theorems, we develop in Section \ref{Sec_unitary} commutator methods
for unitary operators in a two-Hilbert spaces setting: Given a triple $(\H,U,A)$
consisting in a Hilbert space $\H$, a unitary operator $U$, and a self-adjoint
operator $A$, we determine how to obtain commutator results for $(\H,U,A)$ in terms of
commutator results for a second triple $(\H_0,U_0,A_0)$ also consisting in a Hilbert
space, a unitary operator, and a self-adjoint operator. In the process, a bounded
identification operator $J:\H_0\to\H$ must also be chosen. The intuition behind this
approach comes from scattering theory which tells us that given a unitary operator $U$
describing some quantum system in a Hilbert space $\H$ there often exists a simpler
unitary operator $U_0$ in a second Hilbert space $\H_0$ describing the same quantum
system in some asymptotic regime.
Our main results in this context are the following. First, we present in Theorem
\ref{thm_fonctionrho} conditions guaranteeing that $U$ and $A$ satisfy a Mourre
estimate on a Borel set $\Theta\subset\T$ as soon as $U_0$ and $A_0$ satisfy a Mourre
estimate on $\Theta$ (equivalently, we present conditions guaranteeing that $A$ is a
conjugate operator for $U$ on $\Theta$ as soon as $A_0$ is a conjugate operator for
$U_0$ on $\Theta$). Next, we present in Proposition \ref{prop_C1_short} conditions
guaranteeing that $U$ is regular with respect to $A$ (that is, $U\in C^1(A)$) as soon
as $U_0$ is regular with respect to $A_0$ (that is, $U_0\in C^1(A_0)$). Finally, we
give in Assumption \ref{ass_eaa} and Corollaries
\ref{Corol_C1(A)}-\ref{Corol_est_supp} conditions guaranteeing that the most natural
choice for the operator $A$, namely $A=JA_0J^*$, is indeed a conjugate operator for
$U$ as soon as $A_0$ is a conjugate operator for $U_0$.
\section{Unitary operators in a two-Hilbert spaces setting}\label{Sec_unitary}
\setcounter{equation}{0}
In this section, we recall some facts on the spectral family of unitary operators, the
Cayley transform of a unitary operator, locally smooth operators for unitary
operators, and commutator methods for unitary operators in one Hilbert space. We also
present new results on commutator methods for unitary operators in a two-Hilbert
spaces setting.
\subsection{Cayley transform}\label{Sec_Cayley}
Let $\H$ be a Hilbert space with norm $\|\cdot\|_\H$ and scalar product
$\langle\;\!\cdot\;\!,\;\!\cdot\;\!\rangle_\H$ linear in the second argument, let
$\B(\H)$ be the set of bounded linear operators in $\H$ with norm
$\|\cdot\|_{\B(\H)}$, and let $\K(\H)$ be the set of compact linear operators in $\H$.
A unitary operator $U$ in $\H$ is a surjective isometry, that is, an element
$U\in\B(\H)$ satisfying $U^*U=UU^*=1$. Since $U^*U=UU^*$, the spectral theorem for
normal operators implies that $U$ admits exactly one complex spectral family $E_U$,
with support $\supp(E_U)\subset\T$, such that $U=\int_\C z\;\!E_U(\d z)$. The support
$\supp(E_U)$ is the set of points of non-constancy of $E_U$, which coincides with the
spectrum $\sigma(U)$ of $U$ \cite[Thm.~7.34(a)]{Wei80}. For each $s,t\in\R$, one has
the factorization
$$
E_U(s+it):=E_{\re(U)}(s)\;\!E_{\im(U)}(t),
$$
where $E_{\re(U)}$ and $E_{\im(U)}$ are the real spectral families of the bounded
self-adjoint operators
$$
\textstyle
\re(U):=\frac12\;\!(U+U^*)\quad\hbox{and}\quad
\im(U):=\frac1{2i}\;\!(U-U^*).
$$
One can associate in a canonical way a real spectral family $\widetilde E_U$, with
support $\supp(\widetilde E_U)\subset[0,2\pi]$, to the complex spectral family $E_U$
by noting that
$$
U=\int_\R\e^{i\lambda}\widetilde E_U(\d \lambda)
\quad\hbox{with}\quad
\widetilde E_U(\lambda):=
\begin{cases}
0 & \hbox{if }\lambda<0\\
E_U(\e^{i\lambda}) & \hbox{if }\lambda\in[0,2\pi)\\
1 & \hbox{if }\lambda\ge2\pi.
\end{cases}
$$
Since $\widetilde E_U$ is a real spectral family, the corresponding real spectral
measure $\widetilde E^U$ admits the decomposition
$$
\widetilde E^U=\widetilde E^U_{\rm p}+\widetilde E^U_{\rm sc}+\widetilde E^U_{\rm ac},
$$
with $\widetilde E^U_{\rm p}$, $\widetilde E^U_{\rm sc}$, $\widetilde E^U_{\rm ac}$
the pure point, the singular continuous, and the absolutely continuous components of
$\widetilde E_U$, respectively. The corresponding subspaces
$\H_{\rm p}(U):=\widetilde E^U_{\rm p}(\R)\H$,
$\H_{\rm sc}(U):=\widetilde E^U_{\rm sc}(\R)\H$,
$\H_{\rm ac}(U):=\widetilde E^U_{\rm ac}(\R)\H$ provide an orthogonal decomposition
$$
\H=\H_{\rm p}(U)\oplus\H_{\rm sc}(U)\oplus\H_{\rm ac}(U)
$$
which reduces the operator $U$. The sets
\begin{equation*}
\sigma_{\rm p}(U):=\sigma\big(U|_{\H_{\rm p}(U)}\big),\quad
\sigma_{\rm sc}(U):=\sigma\big(U|_{\H_{\rm sc}(U)}\big),\quad
\sigma_{\rm ac}(U):=\sigma\big(U|_{\H_{\rm ac}(U)}\big),
\end{equation*}
are called pure point spectrum, singular continuous spectrum, and absolutely spectrum
continuous of $U$, respectively, and the set
$\sigma_{\rm c}(U):=\sigma_{\rm sc}(U)\cup\sigma_{\rm ac}(U)$ is called the continuous
spectrum of $U$.
If $1\notin\sigma_{\rm p}(U)$, then the subspace $(1-U)\;\!\H$ is dense in $\H$, and
the Cayley transform of $U$ given by
\begin{equation}\label{Cayley_U}
H\varphi:=i(1+U)(1-U)^{-1}\varphi,\qquad\varphi\in\dom(H):=(1-U)\;\!\H,
\end{equation}
is a self-adjoint operator in $\H$ \cite[Thm.~8.4(b)]{Wei80}. Also, a simple
calculation shows that
\begin{equation}\label{Cayley_L}
U=(H-i)(H+i)^{-1}=\e^{iL}\quad\hbox{with}\quad L:=2\arctan(H)+\pi.
\end{equation}
Therefore, the points of the spectra $\sigma(L)\subset[0,2\pi]$ of $L$ and
$\sigma(U)\subset\T$ of $U$ are linked by the relation
$$
\theta\in\sigma(U)~\Leftrightarrow~
2\arctan\left(i\;\!\frac{1+\theta}{1-\theta}\right)+\pi\in\sigma(L)
$$
(in particular, the point $\theta=1$ in $\sigma(U)$ corresponds to the points
$\lambda=0$ and $\lambda=2\pi$ in $\sigma(L)$). In consequence, if $E^L$ denotes the
real spectral measure of $L$, one has for any Borel set $\Theta\subset\T$ the equality
\begin{equation}\label{eq_measures}
E^U(\Theta)=E^L(\Lambda)
\quad\hbox{with}\quad
\Lambda:=\left\{2\arctan\left(i\;\!\frac{1+\theta}{1-\theta}\right)+\pi
\mid\theta\in\Theta\right\}.
\end{equation}
This implies for each Borel set $\Lambda\subset[0,2\pi)$ that
$$
\widetilde E^U(\Lambda)=E^U(\e^{i\Lambda})=E^L\big(f(\Lambda)\big)
\quad\hbox{with}\quad
f(\lambda):=
\begin{cases}
0 & \hbox{if}~~\lambda=0\\
2\arctan\left(i\;\!\frac{1+\e^{i\lambda}}{1-\e^{i\lambda}}\right)+\pi
& \hbox{if}~~\lambda\in(0,2\pi).
\end{cases}
$$
But a simple calculation shows that $f(\lambda)=\lambda$ for each
$\lambda\in[0,2\pi)$. So, one has $\widetilde E^U(\Lambda)=E^L(\Lambda)$ for each
Borel set $\Lambda\subset[0,2\pi)$. Now, it is also clear from the definitions that
$\widetilde E^U(\Lambda)=E^L(\Lambda)$ for each Borel set
$\Lambda\subset\R\setminus[0,2\pi)$. So, one concludes that $\widetilde E^U=E^L$, and
thus that $U$ and $L$ possess the same spectral properties, up to the correspondence
$U=\e^{iL}$.
\subsection{Locally $U$-smooth operators}\label{Sec_smooth}
Let $U$ be a unitary operator in a Hilbert space $\H$, and let $\G$ be an auxiliary
Hilbert space. Then, an operator $T\in\B(\H,\G)$ is locally $U$-smooth on an open set
$\Theta\subset\T$ if for each closed set $\Theta'\subset\Theta$ there exists
$c_{\Theta'}\ge0$ such that
\begin{equation}\label{def_U_smooth}
\sum_{n\in\Z}\big\|T\;\!U^nE^U(\Theta')\varphi\big\|_\G^2
\le c_{\Theta'}\;\!\|\varphi\|_\H^2\quad\hbox{for each $\varphi\in\H$},
\end{equation}
and $T$ is (globally) $U$-smooth if \eqref{def_U_smooth} is satisfied with
$\Theta'=\T$. The condition \eqref{def_U_smooth} is invariant under rotation by
$\omega\in\T$ in the sense that if $T$ is $U$-smooth on $\Theta$, then $T$ is
$(\omega U)$-smooth on $\omega\Theta$ since
$$
\big\|T(\omega U)^nE^{\omega U}(\omega\Theta')\varphi\big\|_\G
=\big\|T\;\!U^nE^U(\Theta')\varphi\big\|_\G
$$
for each closed set $\Theta'\subset\Theta$ and each $\varphi\in\H$. An important
consequence of the existence of a locally $U$-smooth operator $T$ on $\Theta$ is the
inclusion $\overline{E^U(\Theta)T^*\G^*}\subset\H_{\rm ac}(U)$, with $\G^*$ the
adjoint space of $\G$ (see \cite[Thm.~2.1]{ABCF06} for a proof).
Local smoothness with respect to a self-adjoint operator $H$ in $\H$ with domain
$\dom(H)$ is defined in a similar way. An operator $T\in\B\big(\dom(H),\G\big)$ is
locally $H$-smooth on an open set $\Lambda\subset\R$ if for each compact set
$\Lambda'\subset\Lambda$ there exists $c_{\Lambda'}\ge0$ such that
\begin{equation}\label{def_T_smooth}
\int_\R\big\|T\e^{-itH}E^H(\Lambda')\varphi\big\|^2_\G\,\d t
\le c_{\Lambda'}\;\!\|\varphi\|_\H^2\quad\hbox{for each $\varphi\in\H$,}
\end{equation}
and $T$ is (globally) $H$-smooth if \eqref{def_T_smooth} is satisfied with
$\Lambda'=\R$. The condition \eqref{def_T_smooth} is invariant under translation by
$s\in\R$ in the sense that if $T$ is $H$-smooth on $\Lambda$, then $T$ is
$(H+s)$-smooth on $\Lambda+s$ since
$$
\big\|T\e^{-it(H+s)}E^{H+s}(\Lambda'+s)\varphi\big\|_\G
=\big\|T\e^{-itH}E^H(\Lambda')\varphi\big\|_\G
$$
for each compact set $\Lambda'\subset\Lambda$ and each $\varphi\in\H$. Also, the
existence of a locally $H$-smooth operator $T$ on $\Lambda\subset\R$ implies the
inclusion $\overline{E^H(\Lambda)T^*\G^*}\subset\H_{\rm ac}(H)$ (see
\cite[Cor~7.1.2]{ABG96} for a proof).
If $1\notin\sigma_{\rm p}(U)$, then the Cayley transform $H$ of $U$ and the operator
$L=2\arctan(H)+\pi$ are defined by \eqref{Cayley_U} and \eqref{Cayley_L}, and the
existence of locally $U$-smooth operators is equivalent to the existence of locally
$H$-smooth operators and locally $L$-smooth operators:
\begin{Lemma}
Let $U$ be a unitary operator in a Hilbert space $\H$ with $1\notin\sigma_{\rm p}(U)$,
let $\G$ be an auxiliary Hilbert space, let $T\in\B(\H,\G)$, and let $\Theta\subset\T$
be an open set. Then, the following are equivalent:
\begin{enumerate}
\item[(i)] $T$ is locally $U$-smooth on $\Theta$,
\item[(ii)] $T$ is locally $L$-smooth on
$\left\{2\arctan\big(i\;\!\frac{1+\theta}{1-\theta}\big)
+\pi\mid\theta\in\Theta\right\}$,
\item[(iii)] $T(H+i)$ is locally $H$-smooth on
$\left\{i\;\!\frac{1+\theta}{1-\theta}\mid\theta\in\Theta\right\}$.
\end{enumerate}
\end{Lemma}
The equivalence (i) $\Leftrightarrow$ (ii) in the case $\Theta=\T$ is due to T. Kato
(see \cite[Sec.~7]{Kat68}).
\begin{proof}
Assume that $T$ is locally $U$-smooth on $\Theta$, take a closed set
$\Theta'\subset\Theta$, and let
$$
\Lambda':=\left\{2\arctan\left(i\;\!\frac{1+\theta}{1-\theta}\right)+\pi
\mid\theta\in\Theta'\right\}.
$$
Then, Equations \eqref{Cayley_L}-\eqref{eq_measures} and Tonnelli's theorem imply for
each $\varphi\in\H$ that
\begin{align*}
\int_\R \big\|T\e^{-itL}E^L(\Lambda')\varphi\big\|^2_\G\,\d t
&=\sum_{n\in\Z}\int_0^1\big\|T\e^{-i(s+n)L}E^L(\Lambda')\varphi\big\|^2_\G\,\d s\\
&=\int_0^1 \sum_{n\in\Z}\big\|TU^{-n}E^U(\Theta')\e^{-isL}\varphi\big\|^2_\G\,\d s\\
&\le\int_0^1 c_{\Theta'}\;\!\big\|\e^{-isL}\varphi\big\|^2_\H\,\d s\\
&=c_{\Theta'}\;\!\|\varphi\|^2_\H\;\!.
\end{align*}
This shows the implication (i) $\Rightarrow$ (ii). The implication (ii) $\Rightarrow$
(i) is shown in a similar way.
To show the equivalence (i) $\Leftrightarrow$ (iii) we observe that
\eqref{def_U_smooth} is equivalent to
$$
\sup_{z\in\D,\,\psi\in\G,\,\|\psi\|_\G=1}\;\!
\left|\big\langle\psi,T\;\!\delta(U,z)E^U(\Theta')T^*\psi\big\rangle_\G\right|
<\infty,
$$
with $\D:=\{z\in\C\mid|z|<1\}$ and
$
\delta(U,z):=(1-zU^{-1})^{-1}-(1-\overline z^{-1}U^{-1})^{-1}
$
(this follows from the proof of \cite[Thm.~2.2]{ABCF06}), and we observe that
\eqref{def_T_smooth} is equivalent to
$$
\sup_{\omega\in\HH,\,\psi\in\G,\,\|\psi\|_\G=1}\;\!\left|\big\langle\psi,
T\im\big((H-\omega)^{-1}\big)E^H(\Lambda')T^*\psi\big\rangle_\G\right|
<\infty
$$
with $\HH:=\{z\in\C\mid\im(z)>0\}$ (this follows from \cite[Prop.~7.1.1]{ABG96}).
Also, we note that
$$
\delta(U,z)=\big(H^2+1\big)\im\left(\left(H-i\;\!\frac{1+z}{1-z}\right)^{-1}\right),
\quad z\in\D,
$$
and we recall that the map $\D\ni z\mapsto i\;\!\frac{1+z}{1-z}\in\HH$ (the Cayley
transform) is a bijection. So, for any closed set $\Theta'\subset\Theta$, we have
\begin{align*}
&\sup_{z\in\D,\,\psi\in\G,\,\|\psi\|_\G=1}\;\!
\left|\big\langle\psi,T\;\!\delta(U,z)E^U(\Theta')T^*\psi\big\rangle_\G\right|\\
&=\sup_{\omega\in\HH,\,\psi\in\G,\,\|\psi\|_\G=1}\;\!
\left|\big\langle\psi,T(H+i)\im\big((H-\omega)^{-1}\big)E^H(\Lambda')
\big(T(H+i)\big)^*\psi\big\rangle_\G\right|
\end{align*}
with $\Lambda'=\big\{i\;\!\frac{1+\theta}{1-\theta}\mid\theta\in\Theta'\big\}$, and
thus (i) and (iii) are equivalent (note that the operator
$$
T(H+i)\im\big((H-\omega)^{-1}\big)E^H(\Lambda')\big(T(H+i)\big)^*
$$
belongs to $\B(\G)$ for each $\omega\in\HH$ even if $1\in\Theta'$, that is, even if
$\Lambda'$ is not bounded).
\end{proof}
\subsection{Commutator methods in one Hilbert space}\label{Sec_one_Hilbert}
In this section, we present some results on commutator methods for unitary operators
in one Hilbert space $\H$. We start by recalling definitions and results borrowed from
\cite{ABG96,FRT13,Sah97_2}. Let $S\in\B(\H)$ and let $A$ be a self-adjoint operator in
$\H$ with domain $\dom(A)$. For any $k\in\N$, we say that $S$ belongs to $C^k(A)$,
with notation $S\in C^k(A)$, if the map
$$
\R\ni t\mapsto\e^{-itA}S\e^{itA}\in\B(\H)
$$
is strongly of class $C^k$. In the case $k=1$, one has $S\in C^1(A)$ if and only if
the quadratic form
$$
\dom(A)\ni\varphi\mapsto\big\langle A\;\!\varphi,S\varphi\big\rangle_\H
-\big\langle\varphi,SA\;\!\varphi\big\rangle_\H\in\C
$$
is continuous for the topology induced by $\H$ on $\dom(A)$. The operator
corresponding to the continuous extension of the form is denoted by $[A,S]\in\B(\H)$,
and it verifies
$$
[A,S]=\slim_{\tau\to0}\;\![A_\tau,S]
\quad\hbox{with}\quad A_\tau:=(i\tau)^{-1}\big(\e^{i\tau A}-1\big)\in\B(\H),
\quad\tau\in\R\setminus\{0\}.
$$
Three regularity conditions slightly stronger than $S\in C^1(A)$ are defined as
follows: $S$ belongs to $C^{1,1}(A)$, with notation $S\in C^{1,1}(A)$, if
$$
\int_0^1\big\|\e^{-itA}S\e^{itA}+\e^{itA}S\e^{-itA}-2S\big\|_{\B(\H)}
\,\frac{\d t}{t^2}<\infty.
$$
$S$ belongs to $C^{1+0}(A)$, with notation $S\in C^{1+0}(A)$, if $S\in C^1(A)$ and
$$
\int_0^1\big\|\e^{-itA}[A,S]\e^{itA}-[A,S]\big\|_{\B(\H)}\,\frac{\d t}t<\infty.
$$
$S$ belongs to $C^{1+\varepsilon}(A)$ for some $\varepsilon\in(0,1)$, with notation
$S\in C^{1+\varepsilon}(A)$, if $S\in C^1(A)$ and
$$
\big\|\e^{-itA}[A,S]\e^{itA}-[A,S]\big\|_{\B(\H)}
\le{\rm Const.}\;\!t^\varepsilon\quad\hbox{for all $t\in(0,1)$.}
$$
As banachisable topological vector spaces, the sets $C^2(A)$, $C^{1+\varepsilon}(A)$,
$C^{1+0}(A)$, $C^{1,1}(A)$, $C^1(A)$, and $C^0(A)=\B(\H)$, satisfy the continuous
inclusions \cite[Sec.~5.2.4]{ABG96}
$$
C^2(A)\subset C^{1+\varepsilon}(A)\subset C^{1+0}(A)\subset C^{1,1}(A)\subset C^1(A)
\subset C^0(A).
$$
Now, we adapt to the case of unitary operators the definition of two useful functions
introduced in \cite[Sec.~7.2]{ABG96} in the case of self-adjoint operators. For that
purpose, we let $U$ be a unitary operator with $U\in C^1(A)$, for $S,T\in\B(\H)$ we
write $T\gtrsim S$ if there exists an operator $K\in\K(\H)$ such that $T+K\ge S$, and
for $\theta\in\T$ and $\varepsilon>0$ we set
$$
\Theta(\theta;\varepsilon)
:=\big\{\theta'\in\T\mid|\arg(\theta-\theta')|<\varepsilon\big\}
\quad\hbox{and}\quad
E^U(\theta;\varepsilon):=E^U\big(\Theta(\theta;\varepsilon)\big).
$$
With these notations at hand, we define the functions
$\varrho^A_U:\T\to(-\infty,\infty]$ and $\widetilde\varrho^A_U:\T\to(-\infty,\infty]$
by
$$
\varrho^A_U(\theta)
:=\sup\big\{a\in\R\mid\exists\;\!\varepsilon>0~\hbox{such that}~E^U(\theta;\varepsilon)
U^{-1}[A,U]E^U(\theta;\varepsilon)\ge a\;\!E^U(\theta;\varepsilon)\big\}
$$
and
$$
\widetilde\varrho^A_U(\theta)
:=\sup\big\{a\in\R \mid\exists\;\!\varepsilon>0~\hbox{such that}~E^U(\theta;\varepsilon)
U^{-1}[A,U]E^U(\theta;\varepsilon)\gtrsim a\;\!E^U(\theta;\varepsilon)\big\}.
$$
In applications, the function $\widetilde\varrho^A_U$ is more convenient than the
function $\varrho^A_U $ since it is defined in terms of a weaker positivity condition
(positivity up to compact terms). A simple argument shows that
$\widetilde\varrho^A_U(\theta)$ can be defined in an equivalent way by
\begin{equation}\label{equivalent_def}
\widetilde\varrho^A_U(\theta)
=\sup\big\{a\in\R\mid\exists\;\!\eta\in C^\infty(\T,\R)~\hbox{such that}
~\eta(\theta)\ne0~\hbox{and}~\eta(U)U^{-1}[A,U]\eta(U)\gtrsim a\;\!\eta(U)^2\big\}.
\end{equation}
Further properties of the functions $\widetilde\varrho^A_U$ and $\varrho^A_U$ are
collected in the following lemmas. The first one corresponds to
\cite[Prop.~2.3]{FRT13}.
\begin{Lemma}[Virial Theorem for $U$]
Let $U$ be a unitary operator in $\H$ and let $A$ be a self-adjoint operator in $\H$
with $U\in C^1(A)$. Then,
$$
E^U(\{\theta\})U^{-1}[A,U]E^U(\{\theta\})=0
$$
for each $\theta\in\T$. In particular, one has
$\big\langle\varphi,U^{-1}[A,U]\varphi\big\rangle=0$ for each eigenvector
$\varphi\in\H$ of $U$.
\end{Lemma}
\begin{Lemma}\label{lemma_inequalities}
Let $U$ be a unitary operator in $\H$ and let $A$ be a self-adjoint operator in $\H$
with $U\in C^1(A)$. Assume there exist an open set $\Theta\subset\T$ and $a\in\R$ such
that $E^U(\Theta)U^{-1}[A,U]E^U(\Theta)\gtrsim a\;\!E^U(\Theta)$. Then, for each
$\theta\in\Theta$ and each $\eta>0$ there exist $\varepsilon>0$ and a finite rank
orthogonal projection $F$ with $E^U(\{\theta\})\ge F$ such that
$$
E^U(\theta;\varepsilon)U^{-1}[A,U]E^U(\theta;\varepsilon)
\ge(a-\eta)\big(E^U(\theta;\varepsilon)-F\big)-\eta F.
$$
In particular, if $\theta$ is not an eigenvalue of $U$, then
$$
E^U(\theta;\varepsilon)U^{-1}[A,U]E^U(\theta;\varepsilon)
\ge(a-\eta)E^U(\theta;\varepsilon),
$$
while if $\theta$ is an eigenvalue of $U$, one has only
$$
E^U(\theta;\varepsilon)U^{-1}[A,U]E^U(\theta;\varepsilon)
\ge\min\{a-\eta,-\eta\}E^U(\theta;\varepsilon).
$$
\end{Lemma}
\begin{proof}
The proof uses Virial Theorem for $U$ and is analogous to the proof of
\cite[Lemma~7.2.12]{ABG96} in the self-adjoint case. One just needs to replace in the
proof of \cite[Lemma~7.2.12]{ABG96} $[iH,A]$ by $U^{-1}[A,U]$, $E(J)$ by
$E^U(\Theta)$, $E(\{\lambda\})$ by $E^U(\{\theta\})$, and $E(\lambda;1/k)$ by
$E^U(\theta;1/k)$.
\end{proof}
\begin{Lemma}\label{lemma_properties}
Let $U$ be a unitary operator in $\H$ and let $A$ be a self-adjoint operator in $\H$
with $U\in C^1(A)$.
\begin{enumerate}
\item[(a)] The function $\varrho^A_U:\T\to(-\infty,\infty]$ is lower semicontinuous,
and $\varrho^A_U(\theta)<\infty$ if and only if $\theta\in\sigma(U)$.
\item[(b)] The function $\widetilde\varrho^A_U:\T\to(-\infty,\infty]$ is lower
semicontinuous, and $\widetilde\varrho^A_U(\theta)<\infty$ if and only if
$\theta\in\sigma_{\rm ess}(U)$.
\item[(c)] $\widetilde\varrho^A_U\ge\varrho^A_U$.
\item[(d)] If $\theta\in\T$ is an eigenvalue of $U$ and
$\widetilde\varrho^A_U(\theta)>0$, then $\varrho^A_U(\theta)=0$. Otherwise,
$\varrho^A_U(\theta)=\widetilde\varrho^A_U(\theta)$.
\end{enumerate}
\end{Lemma}
\begin{proof}
The proof is an adaptation of the proofs of Lemma 7.2.1, Proposition 7.2.3(a),
Proposition 7.2.6 and Theorem 7.2.13 of \cite{ABG96} to the case of unitary operators.
(a) The fact that $\varrho^A_U(\theta)<\infty$ if and only if $\theta\in\sigma(U)$
follows from the definition of $\varrho^A_U$ and the closedness of $\sigma(U)$. Let
$\theta_0\in\T$ and let $r\in\R$ be such that $\varrho^A_U(\theta_0)>r$. To show the
lower semicontinuity of $\varrho^A_U$ we must show that there is a neighbourhood of
$\theta_0$ on which $\varrho^A_U>r$. Since $\varrho^A_U(\theta_0)>r$, there exist
$a>r$ and $\varepsilon_0>0$ such that
$$
E^U(\theta_0;\varepsilon_0)U^{-1}[A,U]E^U(\theta_0;\varepsilon_0)
\ge a\;\!E^U(\theta_0;\varepsilon_0).
$$
Let $\varepsilon:=\varepsilon_0/2$ and $\theta\in\Theta(\theta_0;\varepsilon_0)$. By
multiplying on the left and on the right the preceding inequality by
$E^U(\theta;\varepsilon)$ and by using the fact that
$
E^U(\theta;\varepsilon)E^U(\theta_0;\varepsilon_0)=E^U(\theta;\varepsilon)
$,
one obtains
$$
E^U(\theta;\varepsilon)U^{-1}[A,U]E^U(\theta;\varepsilon)
\ge a\;\!E^U(\theta;\varepsilon).
$$
This implies that $\varrho^A_U(\theta)\ge a>r$ for all
$\theta\in\Theta(\theta_0;\varepsilon_0)$.
(b)-(c) The lower semicontinuity of $\widetilde\varrho^A_U$ is obtained similarly to
that of $\varrho^A_U$ in point (a), and the inequality
$\widetilde\varrho^A_U\ge\varrho^A_U$ is immediate from the definitions. For the last
claim, we use the fact that $\theta\notin\sigma_{\rm ess}(U)$ if and only if
$E^U(\theta;\varepsilon)\in\K(\H)$ for some $\varepsilon>0$. So
$\theta\notin\sigma_{\rm ess}(U)$ implies that $\widetilde\varrho^A_U(\theta)=\infty$.
Conversely, if $\widetilde\varrho^A_U(\theta)=\infty$, let
$m:=\|E^U(\theta;1)U^{-1}[A,U]E^U(\theta;1)\|_{\B(\H)}$ and $a>m$. Then, there is
$\varepsilon\in(0,1)$ such that
$$
E^U(\theta;\varepsilon)U^{-1}[A,U]E^U(\theta;\varepsilon)
\gtrsim a\;\!E^U(\theta;\varepsilon).
$$
On another hand, the inequality $m\ge E^U(\theta;1)U^{-1}[A,U]E^U(\theta;1)$ and the
fact that $E^U(\theta;\varepsilon)E^U(\theta;1)=E^U(\theta;\varepsilon)$ imply that
$$
m\;\!E^U(\theta;\varepsilon)\ge
E^U(\theta;\varepsilon)U^{-1}[A,U]E^U(\theta;\varepsilon).
$$
Thus $m\;\!E^U(\theta;\varepsilon)\gtrsim a\;\!E^U(\theta;\varepsilon)$, and there
exists $K\in\K(\H)$ such that $K\ge(a-m)\;\!E^U(\theta;\varepsilon)$. This implies by
heredity of compactness that $E^U(\theta;\varepsilon)\in\K(\H)$.
(d) If $\theta$ is not an eigenvalue of $U$, then Lemma \ref{lemma_inequalities}
implies that $\widetilde\varrho^A_U(\theta)\le\varrho^A_U(\theta)$, and so these two
numbers must be equal by point (c). Now assume that $\theta$ is an eigenvalue of $U$.
If $\widetilde\varrho^A_U(\theta)\le0$, then $a\le0$ in Lemma
\ref{lemma_inequalities}, hence $\min\{a-\eta,-\eta\}=a-\eta$ and we have the same
result as before. If $\widetilde\varrho^A_U(\theta)>0$, we may take $a>0$ in Lemma
\ref{lemma_inequalities}, which leads to the inequality $\varrho^A_U(\theta)\ge0$; the
opposite inequality $\varrho^A_U(\theta)\le0$ follows by using Virial theorem for $U:$
if $a<\varrho^A_U(\theta)$, there is $\varepsilon>0$ such that
$E^U(\theta;\varepsilon)U^{-1}[A,U]E^U(\theta;\varepsilon)
\ge a\;\!E^U(\theta;\varepsilon)$; hence
$0=E^U(\{\theta\})U^{-1}[A,U]E^U(\{\theta\})\ge a\;\!E^U(\{\theta\})$. Since
$E^U(\{\theta\})\ne0$, we must have $a\le0$.
\end{proof}
By analogy with the self-adjoint case, we say that $A$ is conjugate to $U$ at the
point $\theta\in\T$ if $\widetilde\varrho^A_U(\theta)>0$, and that $A$ is strictly
conjugate to $U$ at $\theta$ if $\varrho^A_U(\theta)>0$. Since
$\widetilde\varrho^A_U(\theta)\ge\varrho^A_U(\theta)$ for each $\theta\in\T$ by Lemma
\ref{lemma_properties}(c), strict conjugation is a property stronger than conjugation.
\begin{Theorem}[$U$-smooth operators]\label{thm_U_smooth}
Let $U$ be a unitary operator in $\H$, let $A$ be a self-adjoint operator in $\H$, and
let $\G$ be an auxiliary Hilbert space. Assume either that $U$ has a spectral gap and
$U\in C^{1,1}(A)$, or that $U\in C^{1+0}(A)$. Suppose also there exist an open set
$\Theta\subset\T$, a number $a>0$ and an operator $K\in\K(\H)$ such that
$$
E^U(\Theta)\;\!U^{-1}[A,U]\;\!E^U(\Theta)\ge aE^U(\Theta)+K.
$$
Then, each operator $T\in\B(\H,\G)$ which extends continuously to an element of
$\B\big(\dom(\langle A\rangle^s)^*,\G\big)$ for some $s>1/2$ is locally $U$-smooth on
$\Theta\setminus\sigma_{\rm p}(U)$.
\end{Theorem}
\begin{proof}
The claim follows by adapting the proof of \cite[Prop.~2.9]{FRT13} to locally
$U$-smooth operators $T$ with values in the auxiliary Hilbert space $\G$, taking into
account the results of Section \ref{Sec_smooth}.
\end{proof}
The last theorem of this section corresponds to \cite[Thm.~2.7]{FRT13}:
\begin{Theorem}[Spectrum of $U$]\label{thm_spec_U}
Let $U$ be a unitary operator in $\H$ and let $A$ be a self-adjoint operator in $\H$.
Assume either that $U$ has a spectral gap and $U\in C^{1,1}(A)$, or that
$U\in C^{1+0}(A)$. Suppose also there exist an open set $\Theta\subset\T$, a number
$a>0$ and an operator $K\in\K(\H)$ such that
$$
E^U(\Theta)\;\!U^{-1}[A,U]\;\!E^U(\Theta)\ge aE^U(\Theta)+K.
$$
Then, $U$ has at most finitely many eigenvalues in $\Theta$, each one of finite
multiplicity, and $U$ has no singular continuous spectrum in $\Theta$.
\end{Theorem}
\subsection{Commutator methods in a two-Hilbert spaces setting}\label{Sec_two_Hilbert}
From now on, in addition to the triple $(\H,U,A)$, we consider a second triple
$(\H_0,U_0,A_0)$ with $\H_0$ a Hilbert space, $U_0$ a unitary operator in $\H_0$, and
$A_0$ a self-adjoint operator in $\H_0$. We also consider an identification operator
$J\in\B(\H_0,\H)$. The existence of two such triples with an identification operator
is quite standard in scattering theory of unitary operators, at least for the pairs
$(\H,U)$ and $(\H_0,U_0)$ (see for instance the books \cite{BW83,Yaf92}). Part of our
goal in this section is to show that the existence of the conjugate operators $A$ and
$A_0$ is also natural, in the same way it is in the self-adjoint case \cite{RT13_2}.
In the one-Hilbert space setting, the unitary operator $U$ is usually a multiplicative
perturbation of the unitary operator $U_0$. In this case, if $U-U_0$ is compact, the
stability of the function $\widetilde\varrho_{U_0}^{A_0}$ under compact perturbations
allows one to infer information on $U$ from similar information on $U_0$ (see
\cite[Cor.~2.10]{FRT13}). In the two-Hilbert spaces setting, we are not aware of any
general result relating the functions $\widetilde\varrho_U^A$ and
$\widetilde\varrho_{U_0}^{A_0}$. The obvious reason for this being the impossibility
to consider $U$ as a direct perturbation of $U_0$ since these operators do not act in
the same Hilbert space. Nonetheless, the next theorem provides a result in that
direction. For two arbitrary Hilbert spaces $\H_1, \H_2$ and two operators
$S,T\in\B(\H_1,\H_2)$, we use the notation $T\approx S$ if $(T-S)\in\K(\H_1,\H_2)$.
\begin{Theorem}\label{thm_fonctionrho}
Let $(\H_0,U_0,A_0)$ and $(\H,U,A)$ be as above, let $J\in\B(\H_0,\H)$, and assume
that
\begin{enumerate}
\item[(i)] $U_0\in C^1(A_0)$ and $U\in C^1(A)$,
\item[(ii)] $JU_0^{-1}[A_0,U_0]J^*-U^{-1}[A,U]\in\K(\H)$,
\item[(iii)] $JU_0-UJ\in\K(\H_0,\H)$,
\item[(iv)] For each $\eta\in C(\C,\R)$, $\eta(U)(JJ^*-1)\eta(U)\in\K(\H)$.
\end{enumerate}
Then, one has $\widetilde\varrho_U^A\ge \widetilde\varrho_{U_0}^{A_0}$.
\end{Theorem}
An induction argument together with a Stone-Weierstrass density argument shows that
(iii) is equivalent to the apparently stronger condition
\begin{enumerate}
\item[(iii')] For each $\eta\in C(\C,\R)$, $J\eta(U_0)-\eta(U)J\in\K(\H_0,\H)$.
\end{enumerate}
Therefore, in the sequel, we will sometimes use the condition (iii') instead of (iii).
\begin{proof}
For each $\eta\in C(\C,\R)$, we have
\begin{equation}\label{eq_Eone}
\eta(U)U^{-1}[A,U]\eta(U)
\approx\eta(U)JU_0^{-1}[A_0,U_0]J^*\eta(U)\\
\approx J\eta(U_0)U_0^{-1}[A_0,U_0]\eta(U_0)J^*
\end{equation}
due to Assumption (i)-(iii). Furthermore, if there exists $a\in\R$ such that
$$
\eta(U_0)U_0^{-1}[A_0,U_0]\eta(U_0)\gtrsim a\;\!\eta(U_0)^2,
$$
then Assumptions (iii)-(iv) imply that
\begin{equation}\label{eq_Etwo}
J\eta(U_0)U_0^{-1}[A_0,U_0]\eta(U_0)J^*
\gtrsim a\;\!J\eta(U_0)^2J^*
\approx a\;\!\eta(U)JJ^*\eta(U)
\approx a\;\!\eta(U)^2.
\end{equation}
Thus, we obtain $\eta(U)U^{-1}[A,U]\eta(U)\gtrsim a\;\!\eta(U)^2$ by combining
\eqref{eq_Eone} and \eqref{eq_Etwo}. This last estimate, together with the definition
\eqref{equivalent_def} of the functions $\widetilde\varrho_{U_0}^{A_0}$ and
$\widetilde\varrho_U^A$, implies the claim.
\end{proof}
The regularity of $U_0$ with respect to $A_0$ is usually easy to check, while the
regularity of $U$ with respect to $A$ is in general difficult to establish. For that
purpose, various perturbative criteria have been developed for self-adjoint operators
in one Hilbert space, and often a distinction is made between so-called short-range
and long-range perturbations. Roughly speaking, the two terms of the formal commutator
$[A,U]=AU-UA$ are treated separately in the short-range case, while the commutator
$[A,U]$ is really computed in the long-range case. In the sequel, we discuss the case
of short-range type perturbations for unitary operators in a two-Hilbert spaces
setting. The results we obtain are analogous to the ones obtained in
\cite[Sec.~3.1]{RT13_2} for self-adjoint operators in a two-Hilbert spaces setting.
We start by showing how the condition $U\in C^1(A)$ and the assumptions (ii)-(iii) of
Theorem \ref{thm_fonctionrho} can be verified for a class of short-range type
perturbations. Our approach is to infer the desired information on $U$ from equivalent
information on $U_0$, which are usually easier to obtain. Accordingly, our results
exhibit some perturbative flavor. The price one has to pay is to impose some
compatibility conditions between $A_0$ and $A$. For brevity, we set
$$
B:=J U_0-U J\in\B(\H_0,\H)\quad\hbox{and}\quad B_*:=J U_0^*-U^*J\in\B(\H_0,\H).
$$
\begin{Proposition}\label{prop_C1_short}
Let $U_0\in C^1(A_0)$, assume that $\DD\subset\H$ is a core for $A$ such that
$J^*\DD\subset\dom(A_0)$, and suppose that
\begin{equation}\label{hyp1}
\overline{BA_0\upharpoonright\dom(A_0)}\in\B(\H_0,\H),
\quad \overline{B_*A_0\upharpoonright\dom(A_0)}\in\B(\H_0,\H)
\quad\hbox{and}\quad\overline{(JA_0J^*-A)\upharpoonright\DD}\in\B(\H).
\end{equation}
Then, $U\in C^1(A)$.
\end{Proposition}
\begin{proof}
For $\varphi\in\DD$, a direct calculation gives
\begin{align*}
&\big\langle A\varphi,U\varphi\big\rangle_\H
-\big\langle\varphi,UA\varphi\big\rangle_\H\\
&=\big\langle A\varphi,U\varphi\big\rangle_\H
-\big\langle\varphi,UA\varphi\big\rangle_\H
-\big\langle\varphi,J\;\![A_0,U_0]J^*\varphi\big\rangle_\H
+\big\langle\varphi,J\;\![A_0,U_0]J^*\varphi\big\rangle_\H\\
&=\big\langle\varphi,BA_0J^*\varphi\big\rangle_\H
-\big\langle B_*A_0J^*\varphi,\varphi\big\rangle_\H
+\big\langle U^*\varphi,(JA_0J^*-A)\varphi\rangle_\H
-\big\langle(JA_0J^*-A)\varphi,U\varphi\big\rangle_\H\\
&\quad+\big\langle\varphi,J\;\![A_0,U_0]J^*\varphi\big\rangle_\H.
\end{align*}
Furthermore, we have
$$
\big|\big\langle\varphi,BA_0J^*\varphi\big\rangle_\H
-\big\langle B_*A_0J^*\varphi,\varphi\big\rangle_\H\big|
\le{\rm Const.}\;\!\|\varphi\|_\H^2
$$
due to the first two conditions in \eqref{hyp1}, and we have
$$
\big|\big\langle U^*\varphi,(JA_0J^*-A)\varphi\big\rangle_\H
-\big\langle (JA_0J^*-A)\varphi,U \varphi\big\rangle_\H\big|
\le{\rm Const.}\;\!\|\varphi\|^2_\H
$$
due to the third condition in \eqref{hyp1}. Finally, since $U_0\in C^1(A_0)$ and
$J\in\B(\H_0,\H)$ we also have
$$
\big|\big\langle\varphi,J\;\![A_0,U_0]J^*\varphi\big\rangle_\H\big|
\le{\rm Const.}\;\!\|\varphi\|^2_\H.
$$
Since $\DD$ is a core for $A$, this implies that $U\in C^1(A)$.
\end{proof}
We now show how the assumption (ii) of Theorem \ref{thm_fonctionrho} is verified for a
short-range type perturbation. Note that the hypotheses of the following proposition
are slightly stronger than the ones of Proposition \ref{prop_C1_short}. Thus, $U$
automatically belongs to $C^1(A)$.
\begin{Proposition}\label{prop_com_compact}
Let $U_0\in C^1(A_0)$, assume that $\DD\subset\H$ is a core for $A$ such that
$J^*\DD\subset\dom(A_0)$, and suppose that
\begin{equation}\label{c123}
\overline{BA_0\upharpoonright\dom(A_0)}\in\B(\H_0,\H),
\quad\overline{B_*A_0\upharpoonright\dom(A_0)}\in\K(\H_0,\H)
\quad\hbox{and}\quad\overline{(JA_0J^*-A)\upharpoonright\DD}\in\K(\H).
\end{equation}
Then, the difference of bounded operators $JU_0^{-1}[A_0,U_0]J^*-U^{-1}[A,U]$ belongs
to $\K(\H)$.
\end{Proposition}
\begin{proof}
The facts that $U_0\in C^1(A_0)$ and $J^*\DD\subset\dom(A_0)$ imply the inclusions
$$
U_0J^*\DD\subset U_0\;\!\dom(A_0)\subset\dom(A_0).
$$
Using this and the last two conditions of \eqref{c123}, we obtain for $\varphi\in\DD$
and $\psi\in U^{-1}\DD$ that
\begin{align*}
&\big\langle\psi,\big(JU_0^{-1}[A_0,U_0]J^*-U^{-1}[A,U]\big)\varphi\big\rangle_\H\\
&=\big\langle\psi,B_*A_0U_0J^*\varphi\big\rangle_\H
+\big\langle B_*A_0J^*U\psi,\varphi\big\rangle_\H
+\big\langle(JA_0J^*-A)U\psi,U\varphi\big\rangle_\H
-\big\langle\psi,(JA_0J^*-A)\varphi\big\rangle_\H\\
&=\big\langle\psi,K_1U_0J^*\varphi\big\rangle_\H
+\big\langle K_1J^*U\psi,\varphi\big\rangle_\H
+\big\langle K_2U\psi,U\varphi\big\rangle_\H
-\big\langle\psi,K_2\varphi\big\rangle_\H
\end{align*}
with $K_1\in\K(\H_0,\H)$ and $K_2\in\K(\H)$. Since $\DD$ and $U^{-1}\DD$ are dense in
$\H$, it follows that the operator $JU_0^{-1}[A_0,U_0]J^*-U^{-1}[A,U]$ belongs to
$\K(\H)$.
\end{proof}
In the rest of the section, we particularize the previous results to the case where
$A=JA_0J^*$. This case deserves a special attention since it represents the most
natural choice of a conjugate operator $A$ for $U$ when a conjugate operator $A_0$ for
$U_0$ is given. However, one needs in this case the following assumption to guarantee
the self-adjointness of the operator $A:$
\begin{Assumption}\label{ass_eaa}
There exists a set $\DD\subset\dom(A_0J^*)\subset \H$ such that
$JA_0J^*\upharpoonright\DD$ is essentially self-adjoint, with corresponding
self-adjoint extension denoted by $A$.
\end{Assumption}
Assumption \ref{ass_eaa} might be difficult to check in general, but in concrete
situations the choice of the set $\DD$ can be quite natural (see for example Lemma
\ref{lemma_A_self} for the case of quantum walks or \cite[Rem.~4.3]{RT13_1} for the
case of manifolds with asymptotically cylindrical ends). The following two corollaries
follow directly from Propositions \ref{prop_C1_short}-\ref{prop_com_compact} in the
case Assumption \ref{ass_eaa} is satisfied.
\begin{Corollary}\label{Corol_C1(A)}
Let $U_0\in C^1(A_0)$, suppose that Assumption \ref{ass_eaa} holds for some set
$\DD\subset\H$, and assume that
$$
\overline{BA_0\upharpoonright\dom(A_0)}\in\B(\H_0,\H)
\quad\hbox{and}\quad\overline{B_*A_0\upharpoonright\dom(A_0)}\in\B(\H_0,\H).
$$
Then, $U$ belongs to $C^1(A)$.
\end{Corollary}
\begin{Corollary}\label{Corol_est_supp}
Let $U_0\in C^1(A_0)$, suppose that Assumption \ref{ass_eaa} holds for some set
$\DD\subset\H$, and assume that
$$
\overline{BA_0\upharpoonright\dom(A_0)}\in\B(\H_0,\H)
\quad\hbox{and}\quad \overline{B_*A_0\upharpoonright\dom(A_0)}\in\K(\H_0,\H).
$$
Then, the difference of bounded operators $JU_0^{-1}[A_0,U_0]J^*-U^{-1}[A,U]$ belongs
to $\K(\H)$.
\end{Corollary}
\section{Quantum walks with an anisotropic coin}\label{Sec_quantum}
\setcounter{equation}{0}
In this section, we apply the abstract theory of Section \ref{Sec_unitary} to prove
our results on the spectrum of the evolution operator $U$ of the quantum walk with an
anisotropic coin defined in Section \ref{Sec_model}. For this, we first determine in
Section \ref{Sec_asymptotic} the spectral properties and prove a Mourre estimate for
the asymptotic operators $U_\ell$ and $U_{\rm r}$. Then, in Section \ref{Sec_Mourre},
we use the Mourre estimate for $U_\ell$ and $U_{\rm r}$ to derive a Mourre estimate
for $U$. Finally, in Section \ref{Sec_spectrum}, we use the Mourre estimate for $U$ to
prove our results on $U$. We recall that the behaviour of the coin operator $C$ at
infinity is determined by Assumption \ref{ass_short}.
\subsection{Asymptotic operators $U_\ell$ and $U_{\rm r}$}\label{Sec_asymptotic}
For the study of the asymptotic operators $U_\ell$ and $U_{\rm r}$, we use the symbol
$\star$ to denote either the index $\ell$ or the index ${\rm r}$. Also, we introduce
the subspace $\H_{\rm fin}\subset\H$ of elements with finite support
$$
\H_{\rm fin}
:=\bigcup_{n\in\N}\big\{\Psi\in\H\mid\hbox{$\Psi(x)=0$ if $|x|\ge n$}\big\},
$$
the Hilbert space $\KK:=\ltwo\big([0,2\pi),\frac{\d k}{2\pi},\C^2\big)$, and the
discrete Fourier transform $\F:\H\to\KK$, which is the unitary operator defined as the
unique continuous extension of the operator
$$
(\F\Psi)(k):=\sum_{x\in\Z}\e^{-ikx}\Psi(x),\quad\Psi\in\H_{\rm fin},~k\in[0,2\pi).
$$
A direct computation shows that the operator $U_\star$ is decomposable in the Fourier
representation, namely, for all $f\in\KK$ and almost every $k\in[0,2\pi)$ we have
$$
(\F\;\!U_\star\;\!\F^*f)(k)=\widehat{U_\star}(k)f(k)\quad\hbox{with}\quad
\widehat{U_\star}(k)
:=\begin{pmatrix}
\e^{ik}&0\\
0&\e^{-ik}
\end{pmatrix}
C_\star\in\U(2).
$$
Moreover, since $\widehat{U_\star}(k)\in\U(2)$ the spectral theorem implies that
$\widehat{U_\star}(k)$ can be written as
$$
\widehat{U_\star}(k)=\sum_{j=1}^2\lambda_{\star,j}(k)\;\!\Pi_{\star,j}(k),
$$
with $\lambda_{\star,j}(k)$ the eigenvalues of $\widehat{U_\star}(k)$ and
$\Pi_{\star,j}(k)$ the corresponding orthogonal projections.
The next lemma furnishes some information on the spectrum of $U_\star$. To state it,
we use the following parametrisation for the matrices
$C_\star:$
\begin{equation}\label{eq_param_C}
C_\star=\e^{i\delta_\star/2}
\begin{pmatrix}
a_\star\e^{i(\alpha_\star-\delta_\star/2)}
& b_\star\e^{i(\beta_\star-\delta_\star/2)}\\
-b_\star\e^{-i(\beta_\star-\delta_\star/2)}
& a_\star\e^{-i(\alpha_\star-\delta_\star/2)}
\end{pmatrix}
\end{equation}
with $a_\star,b_\star\in[0,1]$ satisfying $a_\star^2+b_\star^2=1$, and
$\alpha_\star,\beta_\star,\delta_\star\in(-\pi,\pi]$. The determinant $\det(C_\star)$
of $C_\star$ is equal to $\e^{i\delta_\star}$. For brevity, we also set
\begin{align*}
\tau_\star(k)&:=a_\star\cos(k+\alpha_\star-\delta_\star/2),\\
\eta_\star(k)&:=\sqrt{1-\tau_\star(k)^2},\\
\varsigma_\star(k)&:=a_\star\sin(k+\alpha_\star-\delta_\star/2),\\
\theta_\star&:=\arccos(a_\star).
\end{align*}
\begin{Lemma}[Spectrum of $U_\star$]\label{lemma_spectrum_U_star}
\begin{enumerate}
\item[(a)] If $a_\star=0$, then $U_\star$ has pure point spectrum
$$
\sigma(U_\star)
=\sigma_{\rm p}(U_\star)
=\big\{i\e^{i\delta_\star/2},-i\e^{i\delta_\star/2}\big\}
$$
with each point an eigenvalue of $U_\star$ of infinite multiplicity.
\item[(b)] If $a_\star\in(0,1)$, then $\sigma_{\rm p}(U_\star)=\varnothing$ and
$$
\sigma(U_\star)
=\sigma_{\rm c}(U_\star)
=\big\{\e^{i\gamma}\mid
\gamma\in[\delta_\star/2+\theta_\star,\pi+\delta_\star/2-\theta_\star]
\cup[\pi+\delta_\star/2+\theta_\star,2\pi+\delta_\star/2-\theta_\star]\big\}.
$$
\item[(c)] If $a_\star=1$, then $\sigma_{\rm p}(U_\star)=\varnothing$ and
$\sigma(U_\star)=\sigma_{\rm c}(U_\star)=\T$.
\end{enumerate}
\end{Lemma}
\begin{proof}
Using the parametrisation for $C_\star$ given in \eqref{eq_param_C}, one gets
$$
\widehat{U_\star}(k)
=\e^{i\delta_\star/2}
\begin{pmatrix}
a_\star(k) & b_\star(k)\\
-\overline{b_\star(k)} & \overline{a_\star(k)}
\end{pmatrix}
$$
with
$$
a_\star(k):=a_\star\e^{i(k+\alpha_\star-\delta_\star/2)}
\quad\hbox{and}\quad
b_\star(k):=b_\star\e^{i(k+\beta_\star-\delta_\star/2)}.
$$
Therefore, the spectrum of $U_\star$ is given by
$$
\sigma(U_\star)=\big\{\lambda_{\star,j}(k)\mid j=1,2,~k\in[0,2\pi)\big\
$$
with $\lambda_{\star,j}(k)$ the solution of the characteristic equation
$$
\det\big(\widehat{U_\star}(k)-\lambda_{\star,j}(k)\big)=0,\quad j=1,2,~k\in[0,2\pi).
$$
In case (a), we obtain
$$
\lambda_{\star,1}(k)=i\e^{i\delta_\star/2}
\quad\hbox{and}\quad
\lambda_{\star,2}(k)=-i\e^{i\delta_\star/2}.
$$
In case (b), we obtain
$$
\lambda_{\star,j}(k)
=\e^{i\delta_\star/2}\big(\tau_\star(k)+i\;\!(-1)^{j-1}\eta_\star(k)\big),\quad j=1,2.
$$
Finally, in case (c) we obtain
$$
\lambda_{\star,1}(k)=\e^{i(k+\alpha_\star)}
\quad\hbox{and}\quad
\lambda_{\star,2}(k)=\e^{-i(k+\alpha_\star-\delta_\star)}.
$$
\end{proof}
We now exhibit normalised eigenvectors $u_{\star,j}(k)$ of $\widehat{U_\star}(k)$
associated with the eigenvalues $\lambda_{\star,j}(k)$ which are $C^\infty$ in the
variable $k:$
$$
\begin{cases}
u_{\star,j}(k):=\frac{\sqrt{\eta_\star(k)+(-1)^{j-1}\varsigma_\star(k)}}
{b_\star\sqrt{2\eta_\star(k)}}
\begin{pmatrix}
ib_\star(k)\\
\varsigma_\star(k)+(-1)^j\eta_\star(k)
\end{pmatrix}
& \hbox{if $a_\star\in[0,1)$}\medskip\\
u_{\star,1}(k):=\begin{pmatrix}1\\0\end{pmatrix}
\quad\hbox{and}\quad
u_{\star,2}(k):=\begin{pmatrix}0\\1\end{pmatrix}
& \hbox{if $a_\star=1$.}
\end{cases}
$$
We leave the reader check that $u_{\star,j}(k)$ are indeed normalised eigenvectors of
$\widehat{U_\star}(k)$ with eigenvalues $\lambda_{\star,j}(k)$. In addition, since for
$a_\star\in[0,1)$ one has $\eta_\star(k)>0$ and
$\eta_\star(k)+(-1)^{j-1}\varsigma_\star(k)>0$, we note that the $2\pi$-periodic map
$\R\ni k\mapsto u_{\star,j}(k)\in\C^2$ is of class $C^\infty$.
Our next goal is to construct a suitable conjugate operator for the operator
$U_\star$. For this, a few preliminaries are necessary. First, we equip the interval
$[0,2\pi)$ with the addition modulo $2\pi$, and for any $n\in\N$ we define the space
$C^n\big([0,2\pi),\C^2\big)\subset\KK$ as the set of functions $[0,2\pi)\to\C^2$ of
class $C^n$. In particular, we have $u_{\star, j}\in C^\infty\big([0,2\pi),\C^2\big)$,
and the space $\F\H_{\rm fin}\subset C^\infty\big([0,2\pi),\C^2\big)$ is the set of
$\C^2$-valued trigonometric polynomials.
Next, we define the asymptotic velocity operator for the operator $U_\star$. For
$j=1,2$, we let $v_{\star,j}:[0,2\pi)\to\R$ be the bounded function given by
\begin{equation}\label{def_small_v}
v_{\star,j}(k)
:=\frac{i\;\!\lambda_{\star,j}'(k)}{\lambda_{\star,j}(k)},\quad k\in[0,2\pi).
\end{equation}
Here, $(\;\!\cdot\;\!)'$ stands for the derivative with respect to $k$, and
$v_{\star,j}$ is real valued because $\lambda_{\star,j}$ takes values in the complex
unit circle. Finally, for all $f\in\KK$ and almost every $k\in[0,2\pi)$, we define the
decomposable operator $\widehat{V_\star}\in\B(\KK)$ by
\begin{equation}\label{def_big_V}
\big(\widehat{V_\star}f\big)(k):=\widehat{V_\star}(k)f(k)
\quad\hbox{where}\quad
\widehat{V_\star}(k):=\sum_{j=1}^2v_{\star,j}(k)\;\!\Pi_{\star,j}(k)\in\B(\C^2),
\end{equation}
and we call asymptotic velocity operator the operator $V_\star$ given as inverse
Fourier transform of $\widehat{V_\star}$, namely,
$$
V_\star:=\F^*\;\!\widehat{V_\star}\;\!\F.
$$
The basic spectral properties of $V_\star$ are collected in the following lemma.
\begin{Lemma}[Spectrum of $V_\star$]
Let $C_\star$ be parameterised as in \eqref{eq_param_C}.
\begin{enumerate}
\item[(a)] If $a_\star=0$, then $v_{\star,j}=0$ for $j=1,2$, and $V_\star=0$.
\item[(b)] If $a_\star\in(0,1)$, then
$v_{\star,j}(k)=\frac{(-1)^{j}\varsigma_\star(k)}{\eta_\star(k)}$ for $j=1,2$ and
$k\in[0,2\pi)$, $\sigma_{\rm p}(V_\star)=\varnothing$ and
$$
\sigma(V_\star)=\sigma_{\rm c}(V_\star)=[-a_\star, a_\star].
$$
\item[(c)] If $a_\star=1$, then $v_{\star,j}=(-1)^{j}$ for $j=1,2$, and $V_\star$
has pure point spectrum
$$
\sigma(V_\star)=\sigma_{\rm p}(V_\star)=\{-1,1\}
$$
with each point an eigenvalue of $V_\star$ of infinite multiplicity.
\end{enumerate}
\end{Lemma}
\begin{proof}
The claims follow from simple calculations using the formulas for
$\lambda_{\star,j}(k)$ in the proof of Lemma \ref{lemma_spectrum_U_star} and the
definition \eqref{def_small_v} of $v_{\star,j}(k)$.
\end{proof}
For any $\xi,\zeta\in C\big([0,2\pi),\C^2\big)$, we define the operator
$|\xi\rangle\langle\zeta|:C\big([0,2\pi),\C^2\big)\to C\big([0,2\pi),\C^2\big)$
by
$$
\big(|\xi\rangle\langle\zeta|f\big)(k)
:=\big\langle\zeta(k),f(k)\big\rangle_2\;\!\xi(k),
\quad f\in C\big([0,2\pi),\C^2\big),~k\in[0,2\pi),
$$
where $\langle\;\!\cdot\;\!,\;\!\cdot\;\rangle_2$ is the usual scalar product on
$\C^2$. This operator extends continuously to an element of $\B(\KK)$, with norm
satisfying the bound
\begin{equation}\label{eq_estimate}
\big\||\xi\rangle\langle\zeta|\big\|_{\B(\KK)}
\le\|\xi\|_{\linf([0,2\pi),\frac{\d k}{2\pi},\C^2)}
\;\!\|\zeta\|_{\linf([0,2\pi),\frac{\d k}{2\pi},\C^2)}.
\end{equation}
We also define the self-adjoint operator $P$ in $\KK$
$$
Pf:=-if',\quad f\in\dom(P):=\big\{f\in\KK\mid\hbox{$f$ is absolutely continuous,
$f'\in\KK$, and $f(0)=f(2\pi)$}\big\}.
$$
With these definitions at hand, we can prove the self-adjointness of an operator
useful for the definition of our future the conjugate operator for $U:$
\begin{Lemma}\label{lemma_X}
The operator
$$
\widehat{X_\star}f
:=-\sum_{j=1}^2\big(\big|u_{\star,j}\big\rangle\big\langle u_{\star,j}\big|P
-i\;\!\big|u_{\star,j}\big\rangle\big\langle u_{\star,j}'\big|\big)f,
\quad f\in\F\H_{\rm fin},
$$
is essentially self-adjoint in $\KK$, with closure denoted by the same symbol. In
particular, the Fourier transform $X_\star:=\F^*\widehat{X_\star}\F$ of
$\widehat{X_\star}$ is essentially self-adjoint on $\H_{\rm fin}$ in $\H$.
\end{Lemma}
\begin{proof}
The proof consists in checking the assumptions of Nelson's commutator theorem
\cite[Thm.~X.37]{RS2} applied with the comparison operator $N:=P^2+1$.
For this, we first note that the operator $N$ is essentially self-adjoint on
$\F\H_{\rm fin}$ because it is the Fourier transform of a multiplication operator
acting on functions with finite support (see \cite[Ex.~5.1.15]{Ped89}). Next, by
performing an integration by parts with boundary terms canceling each other out, we
verify that $\widehat{X_\star}$ is symmetric on $\F\H_{\rm fin}$. Then, by using the
definition of $\widehat{X_\star}$ and the estimate \eqref{eq_estimate}, we check that
the inequality $\|\widehat{X_\star}f\|_\KK\le{\rm Const.}\;\!\|Nf\|_\KK$ holds for
each $f\in\F\H_{\rm fin}$. Finally, a direct calculation shows that for all
$\xi,\zeta\in C^2\big([0,2\pi),\C^2\big)$ and $f\in\F\H_{\rm fin}$
\begin{align*}
&\big\langle Nf,|\xi\rangle\langle\zeta|\;\!f\big\rangle_\KK
-\big\langle f,|\xi\rangle\langle\zeta|Nf\big\rangle_\KK\\
&=\big\langle f,\big(|\xi''\rangle\langle\zeta|
-|\xi\rangle\langle\zeta''|
-2\;\!|\xi'\rangle\langle\zeta'|-2i\;\!|\xi'\rangle\langle\zeta|\;\!P
-2i\;\!|\xi\rangle\langle\zeta'|\;\!P\big)f\big\rangle_\KK.
\end{align*}
This, together with the definition of $\widehat{X_\star}$, implies that
$$
\big|\big\langle\widehat{X_\star}f,Nf\big\rangle_\KK
-\big\langle Nf,\widehat{X_\star}f\big\rangle_\KK\big|
\le{\rm Const.}\;\!\langle f,Nf\rangle_\KK.
$$
Thus, all the assumptions of Nelson's commutator theorem are verified, and the claim
is proved.
\end{proof}
The main relations between the operators introduced so far are summarized in the
following proposition. To state it, we need one more decomposable operator
$\widehat{H_\star}\in\B(\KK)$ defined for all $f\in\KK$ and almost every
$k\in[0,2\pi)$ by
$$
\big(\widehat{H_\star}f\big)(k):=\widehat{H_\star}(k)f(k)
\quad\hbox{where}\quad
\widehat{H_\star}(k):=-\sum_{j=1}^2v_{\star,j}'(k)\;\!\Pi_{\star,j}(k)\in\B(\C^2).
$$
We also need the inverse Fourier transform $H_\star:=\F^*\widehat{H_\star}\F$ of
$\widehat{H_\star}$.
\begin{Proposition}\label{prop_tech}
\begin{enumerate}
\item[(a)] One has the equality $[iX_\star,V_\star]=H_\star$ in the form sense on
$\H_{\rm fin}$.
\item[(b)] $U_\star$, $V_\star$ and $H_\star$ are mutually commuting.
\item[(c)] One has the equality $[X_\star,U_\star]=U_\star V_\star$ in the form sense
on $\H_{\rm fin}$.
\end{enumerate}
\end{Proposition}
\begin{proof}
(a) Let $f,g\in\F\H_{\rm fin}$. Then, a direct calculation using an integration by
parts (with boundary terms canceling each other out) implies that
$$
\big\langle\widehat{X_\star}f,i\;\!\widehat{V_\star}g\big\rangle_\KK
-\big\langle f,i\;\!\widehat{V_\star}\widehat{X_\star}g\big\rangle_\KK
=\big\langle f,\widehat{H_\star}g\big\rangle_\KK.
$$
Therefore, the claim follows by an application of the Fourier transform $\F$.
(b) The mutual commutativity of the operators $U_\star$, $V_\star$ and $H_\star$ is a
direct consequence of their boundedness and their definition in terms of the
orthogonal projections $\Pi_{\star,j}(k)$, $k\in[0,2\pi)$.
(c) As in point (a), the proof consists in computing for $f,g\in\F\H_{\rm fin}$ the
difference
$$
\big\langle\widehat{X_\star}f,\widehat{U_\star}g\big\rangle_\KK
-\big\langle f,\widehat{U_\star}\widehat{X_\star}g\big\rangle_\KK
$$
with an integration by parts, checking that this difference is equal to
$\big\langle g,\widehat{U_\star} \widehat{V_\star}f\big\rangle_\KK$, and applying the
Fourier transform $\F$.
\end{proof}
Since $X_\star$ is essentially self-adjoint on $\H_{\rm fin}$, Proposition
\ref{prop_tech}(a) implies that $V_\star\in C^1(X_\star)$. Therefore, the operator
$$
A_\star\Psi:=\frac12\big(X_\star V_\star+V_\star X_\star\big)\Psi,
\quad\Psi\in\dom(A_\star):=\big\{\Psi\in\H\mid V_\star\Psi\in\dom(X_\star)\big\},
$$
is self-adjoint in $\H$, and essentially self-adjoint on $\H_{\rm fin}$ (see
\cite[Lemma~2.4]{Tie16_2}). We can now state and prove the main results of this
section. We recall that $\Int(\Theta)$ and $\partial\Theta$ denote the interior and
the boundary of a set $\Theta\subset\T$. We also recall that the functions
$\varrho^{A_\star}_{U_\star}$ and $\widetilde{\varrho}^{A_\star}_{U_\star}$ have been
defined in Section \ref{Sec_one_Hilbert}.
\begin{Proposition}\label{prop_functions_rho}
\begin{enumerate}
\item[(a)] $U_\star\in C^1(A_\star)$ with $U_\star^{-1}[A_\star,U_\star]=V_\star^2$.
\item[(b)] $\varrho^{A_\star}_{U_\star}=\widetilde{\varrho}^{A_\star}_{U_\star}$, and
\begin{enumerate}
\item[(i)] if $a_\star=0$, then $\widetilde{\varrho}^{A_\star}_{U_\star}(\theta)=0$
for $\theta\in\big\{i\e^{i\delta_\star/2},-i\e^{i\delta_\star/2}\big\}$ and
$\widetilde{\varrho}^{A_\star}_{U_\star}(\theta)=\infty$ otherwise,
\item[(ii)] if $a_\star\in(0,1)$, then
$\widetilde{\varrho}^{A_\star}_{U_\star}(\theta)>0$ for
$\theta\in\Int\big(\sigma(U_\star)\big)$,
$\widetilde{\varrho}^{A_\star}_{U_\star}(\theta)=0$ for
$\theta\in\partial\sigma(U_\star)$, and
$\widetilde{\varrho}^{A_\star}_{U_\star}(\theta)=\infty$ otherwise,
\item[(iii)] if $a_\star=1$, then $\widetilde{\varrho}^{A_\star}_{U_\star}(\theta)=1$
for all $\theta\in\T$.
\end{enumerate}
\item[(c)]
\begin{enumerate}
\item[(i)] If $a_\star\in(0,1)$, then $U_\star$ has purely absolutely continuous
spectrum
$$
\sigma(U_\star)
=\sigma_{\rm ac}(U_\star)
=\big\{\e^{i\gamma}\mid
\gamma\in[\delta_\star/2+\theta_\star,\pi+\delta_\star/2-\theta_\star]
\cup[\pi+\delta_\star/2+\theta_\star,2\pi+\delta_\star/2-\theta_\star]\big\}.
$$
\item[(ii)] If $a_\star=1$, then $U_\star$ has purely absolutely continuous
spectrum $\sigma(U_\star)=\sigma_{\rm ac}(U_\star)=\T$.
\end{enumerate}
\end{enumerate}
\end{Proposition}
\begin{proof}
(a) A calculation in the forme sense on $\H_{\rm fin}$ using points (b) and (c) of
Proposition \ref{prop_tech} gives
$$
[A_\star,U_\star]
=\frac12\big(V_\star[X_\star,U_\star]+[X_\star,U_\star]V_\star\big)
=U_\star V_\star^2.
$$
Since $A_\star$ is essentially self-adjoint on $\H_{\rm fin}$, this implies that
$U_\star\in C^1(A_\star)$ with $U_\star^{-1}[A_\star,U_\star]=V_\star^2$.
(b) Take $\theta\in\T$ and $\varepsilon>0$. Then, using the result of point (a) and
\eqref{def_big_V}, we obtain for almost every $k\in[0,2\pi)$
\begin{align*}
\big(\F E^{U_\star}(\theta;\varepsilon)U_\star^{-1}[A_\star,U_\star]
E^{U_\star}(\theta;\varepsilon)\F^*\big)(k)
&=\big(\F E^{U_\star}(\theta;\varepsilon)\;\!V_\star^2
E^{U_\star}(\theta;\varepsilon)\F^*\big)(k)\\
&=E^{\widehat{U_\star}(k)}(\theta;\varepsilon)\;\!\widehat{V_\star}(k)^2
E^{\widehat{U_\star}(k)}(\theta;\varepsilon)\\
&\ge\min\big\{v_{\star,1}(k)^2,v_{\star,2}(k)^2\big\}
E^{\widehat{U_\star}(k)}(\theta;\varepsilon).
\end{align*}
Then, the definition \eqref{def_small_v} of $v_{\star,j}(k)$ shows that
$v_{\star,j}(k)=0$ if and only if $\lambda_{\star,j}'(k)=0$, which occurs when
$\lambda_{\star,j}(k)\in\partial\sigma(U_\star)$. Therefore, one gets
$\varrho^{A_\star}_{U_\star}=\widetilde{\varrho}^{A_\star}_{U_\star}$ by Lemma
\ref{lemma_properties}(d), and to conclude one just has to take into account the form
of the boundary sets $\sigma(U_\star)$ given in Lemma \ref{lemma_spectrum_U_star}.
(c) We know from point (a) that $U_\star\in C^1(A_\star)$ with
$U_\star^{-1}[A_\star,U_\star]=V_\star^2$, and Proposition \ref{prop_tech}(a) implies
that $V_\star\in C^1(A_\star)$. Thus, $U_\star\in C^2(A_\star)$. Therefore, if
$a_\star\in(0,1)$, we infer from point (b.ii) and Theorem \ref{thm_spec_U} that
$U_\star$ has no singular continuous spectrum in $\Int\big(\sigma(U_\star)\big)$.
This, together with Lemma \ref{lemma_spectrum_U_star}(b), implies the claim in the
case $a_\star\in(0,1)$. The claim in the case $a_\star=1$ is proved in a similar way.
\end{proof}
\subsection{Mourre estimate for $U$}\label{Sec_Mourre}
In this section, we use the Mourre estimate for the asymptotic operators $U_\ell$ and
$U_{\rm r}$ to derive a Mourre estimate for $U$. To achieve this, we apply the
abstract construction introduced in Section \ref{Sec_two_Hilbert}, starting by
choosing $\H_0:=\H\oplus\H$ as second Hilbert space and $U_0:=U_\ell \oplus U_{\rm r}$
as second unitary operator in $\H_0$.
The spectral properties of $U_0$ are obtained as a consequence of Lemma
\ref{lemma_spectrum_U_star}(a), Proposition \ref{prop_functions_rho}(c) and the direct
sum decomposition of $U_0:$
\begin{Lemma}[Spectrum of $U_0$]\label{lemma_spec_U_0}
One has $\sigma(U_0)=\sigma(U_\ell)\cup\sigma(U_{\rm r})$ and
$\sigma_{\rm sc}(U_0)=\varnothing$. Furthermore,
\begin{enumerate}
\item[(a)] if $a_\ell=a_{\rm r}=0$, then $U_0$ has pure point spectrum
$$
\sigma(U_0)
=\sigma_{\rm p}(U_0)
=\sigma_{\rm p}(U_\ell)\cup\sigma_{\rm p}(U_{\rm r})
=\big\{i\e^{i\delta_\ell/2},-i\e^{i\delta_\ell/2},i\e^{i\delta_{\rm r}/2},
-i\e^{i\delta_{\rm r}/2}\big\}
$$
with each point an eigenvalue of $U_0$ of infinite multiplicity,
\item[(b)] if $a_\ell=0$ and $a_{\rm r}\in(0,1]$, then
$\sigma_{\rm ac}(U_0)=\sigma_{\rm ac}(U_{\rm r})$ with $\sigma_{\rm ac}(U_{\rm r})$ as
in Proposition \ref{prop_functions_rho}(c), and
$$
\sigma_{\rm p}(U_0)
=\sigma_{\rm p}(U_\ell)
=\big\{i\e^{i\delta_\ell/2},-i\e^{i\delta_\ell/2}\big\}
$$
with each point an eigenvalue of $U_0$ of infinite multiplicity,
\item[(c)] if $a_\ell\in(0,1]$ and $a_{\rm r}=0$, then
$\sigma_{\rm ac}(U_0)=\sigma_{\rm ac}(U_\ell)$ with $\sigma_{\rm ac}(U_\ell)$ as in
Proposition \ref{prop_functions_rho}(c), and
$$
\sigma_{\rm p}(U_0)
=\sigma_{\rm p}(U_{\rm r})
=\big\{i\e^{i\delta_{\rm r}/2},-i\e^{i\delta_{\rm r}/2}\big\}
$$
with each point an eigenvalue of $U_0$ of infinite multiplicity,
\item[(d)] if $a_\ell,a_{\rm r}\in(0,1]$, then $U_0$ has purely absolutely continuous
spectrum
$$
\sigma(U_0)
=\sigma_{\rm ac}(U_0)
=\sigma_{\rm ac}(U_\ell)\cup\sigma_{\rm ac}(U_{\rm r})
$$
with $\sigma_{\rm ac}(U_\ell)$ and $\sigma_{\rm ac}(U_{\rm r})$ as in Proposition \ref{prop_functions_rho}(c).
\end{enumerate}
\end{Lemma}
Also, as intuition suggests and as already stated in Theorem \ref{thm_essential}, the
spectrum of $U_0$ coincides with the essential spectrum of $U$, namely,
$$
\sigma_{\rm ess}(U)=\sigma(U_\ell)\cup\sigma(U_{\rm r})=\sigma(U_0).
$$
\begin{proof}[Proof of Theorem \ref{thm_essential}]
The proof is based on an argument using crossed product $C^*$-algebras inspired from
\cite{GI02,Man02}.
Let $\A$ be the algebra of functions $\Z\to\B(\C^2)$ admitting limits at $\pm\infty$,
and let $\A_0$ be the ideal of $\A$ consisting in functions $\Z\to\B(\C^2)$ vanishing
at $\pm\infty$. Since $\A$ is equipped with an action of $\Z$ by translation, namely,
$$
\big(T_y\varphi\big)(x):=\varphi(x+y),\quad x,y\in\Z,~\varphi\in\A,
$$
we can consider the crossed product algebra $\A\rtimes\Z$, and the functoriality of
the crossed product implies the identities
\begin{equation}\label{eq_alg}
(\A\rtimes\Z)/(\A_0\rtimes\Z)
\cong(\A/\A_0)\rtimes\Z
=\big(\B(\C^2)\oplus\B(\C^2)\big)\rtimes\Z
=\big(\B(\C^2)\rtimes\Z\big)\oplus\big(\B(\C^2)\rtimes\Z\big),
\end{equation}
where the equality $\A/\A_0=\B(\C^2)\oplus\B(\C^2)$ is obtained by evaluation of the
functions $\varphi\in\A$ at $\pm\infty$.
Now, the algebras $\A\rtimes\Z$ and $\A_0\rtimes\Z$ can be faithfully represented in
$\H$ by mapping the elements of $\A$ and $\A_0$ to multiplication operators in $\H$
and the elements of $\Z$ to the shifts $T_z$. Writing $\mathfrak A$ and
$\mathfrak A_0$ for these representations of $\A\rtimes\Z$ and $\A_0\rtimes\Z$ in
$\H$, we can note three facts. First, $\mathfrak A_0$ is equal to the ideal of compact
operators $\K(\H)$. Secondly, the operator $U$ belongs to $\mathfrak A$, since
$$
U=SC
=T_1\begin{pmatrix}1&0\\0&0\end{pmatrix}C
+T_{-1}\begin{pmatrix}0&0\\0&1\end{pmatrix}C
$$
with $T_1,T_{-1}$ shifts and
$
\left(\begin{smallmatrix}1&0\\0&0\end{smallmatrix}\right)C,
\left(\begin{smallmatrix}0&0\\0&1\end{smallmatrix}\right)C
$
multiplication operators in $\H$. Thirdly, the essential spectrum of $U$ in
$\mathfrak A$ is equal to the spectrum of the image of $U$ in the quotient algebra
$\mathfrak A/\K(\H)=\mathfrak A/\mathfrak A_0$. These facts, together with
\eqref{eq_alg} and Lemma \ref{lemma_spec_U_0}, imply the equalities
$$
\sigma_{\rm ess}(U)
=\sigma \big(SC(-\infty)\oplus SC(+\infty)\big)
=\sigma\big(SC_\ell\oplus SC_{\rm r}\big)
=\sigma(U_\ell)\cup\sigma(U_{\rm r})
=\sigma(U_0),
$$
which prove the claim.
\end{proof}
Next, we define the identification operator $J\in\B(\H_0,\H)$ by
$$
J(\Psi_\ell,\Psi_{\rm r}):=j_\ell\;\!\Psi_\ell+j_{\rm r}\;\!\Psi_{\rm r},
\quad(\Psi_\ell,\Psi_{\rm r})\in\H_0,
$$
where
$$
j_{\rm r}(x):=
\begin{cases}
1 & \hbox{if $x\ge0$}\\
0 & \hbox{if $x\le-1$}
\end{cases}
\quad\hbox{and}\quad
j_\ell:=1-j_{\rm r}.
$$
The adjoint operator $J^*\in\B(\H,\H_0)$ satisfies
$$
J^*\Psi=(j_\ell\;\!\Psi,j_{\rm r}\;\!\Psi),\quad\Psi\in\H.
$$
Moreover, using the same notation for the functions $j_\ell,j_{\rm r}$ and the
associated multiplication operators in $\H$, one directly gets:
\begin{Lemma}\label{lemma_J}
$J^*J=j_\ell \oplus j_{\rm r}$ is an orthogonal projection on $\H_0$, and
$JJ^*=1_\H$.
\end{Lemma}
The first result of the next lemma is an analogue of Proposition
\ref{prop_functions_rho}(a) in the Hilbert space $\H_0$. To state it, we need to
introduce the operator $A_0:=A_\ell\oplus A_{\rm r}$ (which will be used as a
conjugate operator for $U_0$) and the operator $V_0:=V_\ell\oplus V_{\rm r}$.
\begin{Lemma}\label{lemma_B_compact}
\begin{enumerate}
\item[(a)] $U_0\in C^1(A_0)$ with $U_0^{-1}[A_0, U_0]=V_0^2$.
\item[(b)] $B:=JU_0-UJ\in\K(\H_0,\H)$ and $B_*:=JU_0^*-U^*J\in\K(\H_0,\H)$.
\end{enumerate}
\end{Lemma}
\begin{proof}
The proof of point (a) is similar to the proof of Proposition
\ref{prop_functions_rho}(a); one just has to replace the operators
$U_\star,A_\star,V_\star$ in $\H$ by the operators $U_0,A_0,V_0$ in $\H_0$. For point
(b), a direct computation with $(\Psi_\ell,\Psi_{\rm r})\in\H_0$ gives
\begin{align}
B (\Psi_\ell,\Psi_{\rm r})
&=\big(j_\ell\;\!U_\ell\Psi_\ell+j_{\rm r}\;\!U_{\rm r}\Psi_{\rm r}\big)
-U\big(j_\ell\Psi_\ell+j_{\rm r}\Psi_{\rm r}\big)\nonumber\\
&=\big([j_\ell,U_\ell]-(U-U_\ell)\;\!j_\ell\big)\Psi_\ell
+\big([j_{\rm r},U_{\rm r}]-(U-U_{\rm r})\;\!j_{\rm r}\big)\Psi_{\rm r}\nonumber\\
&=\big([j_\ell,S]C_\ell-S(C-C_\ell)\;\!j_\ell\big)\Psi_\ell
+\big([j_{\rm r},S]C_{\rm r}-S(C-C_{\rm r})\;\!j_{\rm r}\big)\Psi_{\rm r}.\label{eq_B}
\end{align}
Since we have $[j_\star,S]\in\K(\H)$ and $(C-C_\star)\;\!j_\star\in\K(\H)$ as a
consequence of Assumption \ref{ass_short}, it follows that $B\in\K(\H_0,\H)$. The
inclusion $B_*\in\K(\H_0,\H)$ is proved in a similar way.
\end{proof}
The next step is to define a conjugate operator $A$ for $U$ by using the conjugate
operator $A_0$ for $U_0$. For this, we consider the operator $JA_0J^*$ which is
well-defined and symmetric on $\H_{\rm fin}$. We have the equality
\begin{equation}\label{eq_J_A_0_J_star}
JA_0J^*=j_\ell\;\!A_\ell\;\!j_\ell+j_{\rm r}\;\!A_{\rm r}\;\!j_{\rm r}
\quad\hbox{on}\quad\H_{\rm fin},
\end{equation}
and $JA_0J^*$ is essentially self-adjoint on $\H_{\rm fin}:$
\begin{Lemma}[Conjugate operator for $U$]\label{lemma_A_self}
The operator $JA_0J^*$ is essentially self-adjoint on $\H_{\rm fin}$, with
corresponding self-adjoint extension denoted by $A$.
\end{Lemma}
\begin{proof}
The operator $\widehat{j_\star}:=\F j_\star\F^*\in\B(\KK)$ satisfies
$\widehat{j_\star}\;\!\dom(P)\subset\dom(P)$ and $[\widehat{j_\star},P]=0$ on
$\dom(P)$. Therefore, we have the following equalities on $\F\H_{\rm fin}$
\begin{align*}
\F j_\star A_\star\;\!j_\star\F^*
&=\tfrac12\F j_\star\big(X_\star V_\star+V_\star X_\star\big)\;\!j_\star\F^*\\
&=\tfrac12\;\!\widehat{j_\star}\big(\widehat{X_\star}\widehat{V_\star}
+\widehat{V_\star}\widehat{X_\star}\big)\;\!\widehat{j_\star}\\
&=\widehat{j_\star}\big(\widehat{V_\star}\widehat{X_\star}
-\tfrac{i}2\widehat{H_\star}\big)\;\!\widehat{j_\star}\\
&=-\sum_{j=1}^2\Big(\widehat{j_\star}\big|v_{\star,j}u_{\star,j}\big\rangle
\big\langle u_{\star,j}\big|\;\!\widehat{j_\star}\;\!P
-i\;\!\widehat{j_\star}\big|v_{\star,j}u_{\star,j}\big\rangle
\big\langle u_{\star,j}'\big|\;\!\widehat{j_\star}\Big)
-\tfrac{i}2\;\!\widehat{j_\star}\widehat{H_\star}\widehat{j_\star}.
\end{align*}
which give on $\F\H_{\rm fin}$
$$
\F JA_0J^*\F^*
=-\sum_{j=1}^2\sum_{\star\in\{\ell,{\rm r}\}}
\widehat{j_\star}\big|v_{\star,j}u_{\star,j}\big\rangle
\big\langle u_{\star,j}\big|\;\!\widehat{j_\star}\;\!P
+i\sum_{j=1}^2\sum_{\star\in\{\ell,{\rm r}\}}
\widehat{j_\star}\;\!\big|v_{\star,j}u_{\star,j}\big\rangle
\big\langle u_{\star,j}'\big|\;\!\widehat{j_\star}
-\tfrac{i}2\sum_{\star\in\{\ell,{\rm r}\}}
\widehat{j_\star}\widehat{H_\star}\widehat{j_\star}.
$$
The rest of the proof consists in an application of Nelson's commutator theorem
\cite[Thm.~X.37]{RS2} with the comparison operator $N:=P^2+1$. The estimates necessary
to apply the theorem are similar to the ones mentioned in the proof of Lemma
\ref{lemma_X}. As a consequence, it follows that $\F JA_0J^*\F^*$ is essentially
self-adjoint on $\F\H_{\rm fin}$, and thus that $JA_0J^*$ is essentially self-adjoint
on $\H_{\rm fin}$.
\end{proof}
We are thus in the setup of Assumption \ref{ass_eaa} with the set $\DD=\H_{\rm fin}$.
So, the next step is to show the inclusion $U\in C^1(A)$. For this, we use Corollary
\ref{Corol_C1(A)}. Using Corollary \ref{Corol_est_supp}, we also get an additional
compacity result:
\begin{Lemma}\label{lemma_last}
$U\in C^{1}(A)$ and $JU_0^{-1}[A_0, U_0]J^*-U^{-1}[A,U]\in\K(\H)$.
\end{Lemma}
\begin{proof}
First, we recall that $U_0\in C^1(A_0)$ due to Lemma \ref{lemma_B_compact}(a), and
that Assumption \ref{ass_eaa} holds with $\DD=\H_{\rm fin}$. Next, we note that the
expression for $B(\Psi_\ell,\Psi_{\rm r})$ with $(\Psi_\ell,\Psi_{\rm r})\in\H_0$ is
given in \eqref{eq_B}, and that
$$
B_*(\Psi_\ell,\Psi_{\rm r})
=\big(C^*[j_\ell,S^*]-(C^*-C_\ell^*)\;\!j_\ell S^*\big)\Psi_\ell
+\big(C^*[j_{\rm r},S^*]-(C^*-C_{\rm r}^*)\;\!j_{\rm r}S^*\big)\Psi_{\rm r}.
$$
Furthermore, we know from Lemma \ref{lemma_B_compact}(b) that $B,B_*\in\K(\H_0,\H)$.
In consequence, due to Corollaries \ref{Corol_C1(A)}-\ref{Corol_est_supp}, the claims
will follow if we show that $\overline{BA_0\upharpoonright\dom(A_0)}\in\B(\H_0,\H)$
and $\overline{B_*A_0\upharpoonright\dom(A_0)}\in\K(\H_0,\H)$. For this, we first note
that computations as in the proof of Lemma \ref{lemma_A_self} imply on $\H_{\rm fin}$
the equalities
\begin{align}
A_\star&=X_\star V_\star+\tfrac{i}2H_\star\nonumber\\
&=-\F^*\left\{P\sum_{j=1}^2\Big(\big|u_{\star,j}\big\rangle
\big\langle v_{\star,j}u_{\star,j}\big|+i\;\!\big|u'_{\star,j}\big\rangle
\big\langle v_{\star,j} u_{\star,j}\big|\Big)\right\}\F+\tfrac{i}2H_\star\nonumber\\
&=Q\;\!\F^*\left\{\sum_{j=1}^2\Big(\big|u_{\star,j}\big\rangle
\big\langle v_{\star,j} u_{\star,j}\big|+i\;\!\big|u'_{\star,j}\big\rangle\big\langle
v_{\star,j} u_{\star,j}\big|\Big)\right\}\F+\tfrac{i}2H_\star\label{eq_X_onleft}
\end{align}
with $Q$ the self-adjoint multiplication operator defined by
\begin{equation}\label{eq_def_Q}
\big(Q\;\!\Psi\big)(x)=x\;\!\Psi(x),
\quad x\in\Z,~\Psi\in\dom(Q):=\big\{\Psi\in\H\mid\|Q\;\!\Psi\|_\H<\infty\big\}.
\end{equation}
Therefore, since all the operators on the right of $Q$ in \eqref{eq_X_onleft} are
bounded, it is sufficient to show that
$$
\overline{B(Q\oplus Q)\upharpoonright\dom(Q)\oplus\dom(Q)}\in\B(\H_0,\H)
\quad\hbox{and}\quad
\overline{B_*(Q\oplus Q)\upharpoonright\dom(Q)\oplus\dom(Q)}\in\K(\H_0,\H).
$$
However, this can be deduced from the Assumption \ref{ass_short} once the following
observations are made: $\big[j_\star,S\big]=Sm_\star$ with $m_\star:\Z\to\B(\C^2)$ a
function with compact support, $[j_\star,S^*]=S^*n_\star$ with $n_\star:\Z\to\B(\C^2)$
a function with compact support, and $S^*Q=QS^*+b$ with
$b\in\linf\big(\Z,\B(\C^2)\big)$.
\end{proof}
We now recall that the set
$$
\tau(U):=\partial\sigma(U_\ell)\cup\partial\sigma(U_{\rm r}).
$$
has been introduced in Section \ref{Sec_model}. Due to Lemma
\ref{lemma_spectrum_U_star}, $\tau(U)$ contains at most $8$ values. Moreover, since we
show in the next proposition that a Mourre estimate holds on the set
$\{\sigma(U_\ell)\cup\sigma(U_{\rm r})\}\setminus\tau(U)$, it is natural to interpret
$\tau(U)$ as the set of thresholds in the spectrum of $U$.
\begin{Proposition}[Mourre estimate for $U$]\label{prop_Mourre}
We have $\widetilde\varrho_U^A\ge \widetilde\varrho_{U_0}^{A_0}$ with
$
\widetilde\varrho_{U_0}^{A_0}
=\min\big\{\widetilde\varrho_{U_\ell}^{A_\ell},
\widetilde\varrho_{U_{\rm r}}^{A_{\rm r}}\big\}
$
and $\widetilde\varrho_{U_\ell}^{A_\ell},\widetilde\varrho_{U_{\rm r}}^{A_{\rm r}}$
given in Proposition \ref{prop_functions_rho}. In particular,
$\widetilde\varrho_{U_0}^{A_0}(\theta)>0$ if
$\theta\in\{\sigma(U_\ell)\cup\sigma(U_{\rm r})\}\setminus\tau(U)$.
\end{Proposition}
\begin{proof}
The first claim follows from Theorem \ref{thm_fonctionrho}, with the assumptions of
this theorem verified in Lemmas \ref{lemma_J}-\ref{lemma_last}. The second claim
follows from Proposition \ref{prop_functions_rho} and standard results on the function
$\widetilde\varrho_{U_0}^{A_0}$ when $A_0$ and $U_0$ are direct sums of operators (see
\cite[Prop.~8.3.5]{ABG96} for a proof in the case of direct sums of self-adjoint
operators).
\end{proof}
\subsection{Spectral properties of $U$}\label{Sec_spectrum}
In order to go one step further in the study of $U$, a regularity property of $U$ with
respect to $A$ stronger than $U\in C^1(A)$ has to be established. This regularity
property will be obtained by considering first the operator $JU_0 J^*$, and then by
analysing the difference $U-JU_0J^*$. We note that $JU_0J^*$ and $U-JU_0J^*$ satisfy
the equalities
\begin{equation}\label{eq_J_U_0_J_star}
JU_0J^*=j_\ell\;\!U_\ell\;\!j_\ell+j_{\rm r}\;\!U_{\rm r}\;\!j_{\rm r}
\end{equation}
and
\begin{equation}\label{eq_difference}
U-JU_0J^*=j_\ell(U-U_\ell)\;\!j_\ell+j_{\rm r}(U-U_{\rm r})\;\!j_{\rm r}
+j_\ell\;\!U\;\!j_{\rm r}+j_{\rm r}\;\!U\;\!j_\ell.
\end{equation}
\begin{Lemma}\label{lemma_C2}
$JU_0J^*\in C^2(A)$.
\end{Lemma}
\begin{proof}
The proof is based on standard results from toroidal pseudodifferential calculus, as
presented for example in \cite[Chap.~4]{RT10_2}. The normalisation we use for the
Fourier transform differs from the one used in \cite{RT10_2}, but the difference is
harmless.
(i) First, we note that $\widehat{j_\star}$ is a toroidal pseudodifferential operator
on $\F\H_{\rm fin}$ with symbol in $S^0_{\rho,0}(\T\times\Z)$ for each $\rho>0$ (see
the definitions 4.1.7 and 4.1.9 of \cite{RT10_2} for details). Similarly, the equation
\eqref{eq_X_onleft} shows that $\widehat{A_\star}$ is a first order differential
operator on $\F\H_{\rm fin}$ with matrix coefficients in
$\M\big(2,C^\infty(\T)\big)\subset\M\big(2,S^0_{\rho,0}(\T\times\Z)\big)$ for each
$\rho>0$. In consequence, it follows from \cite[Thm.~4.7.10]{RT10_2} that the
commutator $\big[\widehat{j_\star},\widehat{A_\star}\big]$ on $\F\H_{\rm fin}$ is
well-defined and equal to a toroidal pseudodifferential operator with matrix
coefficients in $\M\big(2,S^{1-\rho}_{\rho,0}(\T\times\Z)\big)$ for each $\rho>0$.
This implies that $\big[\widehat{j_\star},\widehat{A_\star}\big]$ is bounded on
$\F\H_{\rm fin}$, and thus that $\widehat{j_\star}\in C^1(\widehat{A_\star})$ since
$\F\H_{\rm fin}$ is dense in $\dom(\widehat{A_\star})$. By Fourier transform, it
follows that $j_\star\in C^1(A_\star)$.
(ii) A calculation in the form sense on $\H_{\rm fin}$ using equations
\eqref{eq_J_A_0_J_star} and \eqref{eq_J_U_0_J_star} and the identities
$j_\ell\;\!j_{\rm r}=0=j_{\rm r}\;\!j_\ell$ gives
\begin{align}
\big[JU_0J^*,A\big]
&=\big[j_\ell\;\!U_\ell\;\!j_\ell,j_\ell\;\!A_\ell\;\!j_\ell\big]
+\big[j_{\rm r}\;\!U_{\rm r}\;\!j_{\rm r},j_{\rm r}\;\!A_{\rm r}\;\!j_{\rm r}\big]
\nonumber\\
&=\sum_{\star\in\{\ell,{\rm r}\}}j_\star\big(U_\star\;\!j_\star A_\star
-A_\star\;\!j_\star U_\star\big)\;\!j_\star\nonumber\\
&=\sum_{\star\in\{\ell,{\rm r}\}}j_\star\big(\big[U_\star,j_\star\big]A_\star
+\big[j_\star U_\star,A_\star\big]\big)\;\!j_\star.\label{eq_2_terms}
\end{align}
Since $j_\star U_\star\in C^1(A_\star)$ by Proposition \ref{prop_functions_rho}(a),
point (i) and \cite[Prop.~5.1.5]{ABG96}, the second term on the r.h.s. of
\eqref{eq_2_terms} belongs to $\B(\H)$. Furthermore, a calculation using the
definition of the shift operator $S$ shows that
$$
\big[U_\star,j_\star\big]
=\big[S,j_\star\big]C_\star
=B_\star m_\star
$$
with $B_\star\in\B(\H)$ and $m_\star:\Z\to\B(\C^2)$ a function with compact support.
It follows from \eqref{eq_X_onleft} that $\big[U_\star,j_\star\big]A_\star$ is bounded
on $\H_{\rm fin}$. Therefore, both terms on the r.h.s. of \eqref{eq_2_terms} are
bounded on $\H_{\rm fin}$, and thus we infer from the density of $\H_{\rm fin}$ in
$\dom(A)$ that $JU_0J^*\in C^1(A)$.
(iii) To show that $JU_0J^*\in C^2(A)$, one has to commute the r.h.s. of
\eqref{eq_2_terms} once more with $A$. Doing this in the form sense on $\H_{\rm fin}$
with the notation $\sum_{\star\in\{\ell,{\rm r}\}}j_\star D_\star\;\!j_\star$ with
$D_\star:=[U_\star,j_\star]A_\star+[j_\star U_\star,A_\star]$ for the r.h.s. of
\eqref{eq_2_terms}, one gets that $JU_0J^*\in C^2(A)$ if the operators
$[D_\star,A_\star]$, $[D_\star,j_\star]A_\star$ and $A_\star[D_\star,j_\star]$ defined
in the form sense on $\H_{\rm fin}$ extend continuously to elements of $\B(\H)$.
For the first operator, we have in the form sense on $\H_{\rm fin}$ the equalities
\begin{align}
[D_\star,A_\star]
&=\big[[U_\star,j_\star]A_\star+j_\star[U_\star,A_\star]
+[j_\star,A_\star]U_\star,A_\star\big]\nonumber\\
&=\big[[U_\star,j_\star]A_\star,A_\star\big]
+j_\star\big[[U_\star,A_\star],A_\star\big]
+[j_\star,A_\star][U_\star,A_\star]
+[j_\star,A_\star][U_\star,A_\star]
+\big[[j_\star,A_\star],A_\star\big]U_\star\;\!. \label{eq_first_com}
\end{align}
Then, simple adaptations of the arguments presented in points (i) and (ii) show that
the operators $[j_\star,A_\star],[U_\star,j_\star]\in\B(\H)$ can be multiplied in the
form sense on $\H_{\rm fin}$ by several operators $A_\star$ on the left and/or on the
right and that the resultant operators extend continuously to elements of $\B(\H)$.
Therefore, the first, the third, the fourth and the fifth terms in
\eqref{eq_first_com} extend continuously to elements of $\B(\H)$. For the second term,
we note from Propositions \ref{prop_tech}(a) and \ref{prop_functions_rho}(a) that
$U_\star,V_\star\in C^1(A_\star)$ with $[U_\star,A_\star]=-U_\star V_\star^2$. In
consequence, we have $U_\star V_\star^2\in C^1(A_\star)$ by \cite[Prop.~5.1.5]{ABG96}
and
$$
j_\star\big[[U_\star,A_\star],A_\star\big]
=-j_\star\big[U_\star V_\star^2,A_\star\big]
\in\B(\H).
$$
The proof that the operators $[D_\star,j_\star]A_\star$ and $A_\star[D_\star,j_\star]$
defined in the form sense on $\H_{\rm fin}$ extend continuously to elements of
$\B(\H)$ is similar. The only noticeable difference is the appearance of terms
$[U_\star V_\star^2,j_\star]A_\star$ and $A_\star [U_\star V_\star^2,j_\star]$.
However, by observing that $V_\star^2\in C^1(A_\star)$ and that $[V^2_\star,j_\star]$
is a toroidal pseudodifferential operator with matrix coefficients in
$\M\big(2,S^{-\rho}_{\rho,0}(\T\times\Z)\big)$ for each $\rho>0$, one infers that
$[U_\star V_\star^2,j_\star]A_\star$ and $A_\star [U_\star V_\star^2,j_\star]$ extend
continuously to elements of $\B(\H)$.
\end{proof}
In the next lemma, we prove that $U$ satisfies sufficient regularity with respect to
$A$, namely that $U\in C^{1+\varepsilon}(A)$ for some $\varepsilon\in(0,1)$. We recall
from Section \ref{Sec_one_Hilbert} that the sets $C^2(A)$, $C^{1+\varepsilon}(A)$,
$C^{1+0}(A)$ and $C^{1,1}(A)$ satisfy the continuous inclusions
$$
C^2(A)\subset C^{1+\varepsilon}(A)\subset C^{1+0}(A)\subset C^{1,1}(A).
$$
\begin{Lemma}\label{lemma_C_1_epsilon}
$U\in C^{1+\varepsilon}(A)$ for each $\varepsilon\in(0,1)$ with
$\varepsilon\le\min\{\varepsilon_\ell,\varepsilon_{\rm r}\}$.
\end{Lemma}
\begin{proof}
(i) Since $JU_0J^*\in C^2(A)$ by Lemma \ref{lemma_C2} and
$C^2(A)\subset C^{1+\varepsilon}(A)$, it is sufficient to show that
$U-JU_0J^*\in C^{1+\varepsilon}(A)$, with $U-JU_0J^*$ given by \eqref{eq_difference}.
Moreover, calculations as in the proof of Lemma \ref{lemma_C2} show that the last two
terms $j_\ell\;\!U\;\!j_{\rm r}$ and $j_{\rm r}\;\!U\;\!j_\ell$ of
\eqref{eq_difference} belong to $C^2(A)$. So, it only remains to show that
$
j_\ell(U-U_\ell)\;\!j_\ell+j_{\rm r}(U-U_{\rm r})\;\!j_{\rm r}
\in C^{1+\varepsilon}(A)
$.
(ii) In order to show the mentioned inclusion, we first observe from
\eqref{eq_def_U_ell} and \eqref{eq_J_A_0_J_star} that we have in the form sense on
$\H_{\rm fin}$ the equalities
\begin{align}
\big[j_\ell(U-U_\ell)\;\!j_\ell+j_{\rm r}(U-U_{\rm r})\;\!j_{\rm r},A\big]
&=\sum_{\star\in\{\ell,{\rm r}\}}
\big[j_\star(U-U_\star)\;\!j_\star,j_\star A_\star\;\!j_\star\big]\nonumber\\
&=\sum_{\star\in\{\ell,{\rm r}\}}\big(j_\star S(C-C_\star)
\;\!j_\star A_\star\;\!j_\star
-j_\star A_\star\;\!j_\star S(C-C_\star)\;\!j_\star\big).\label{eq_first_step}
\end{align}
Then, using Assumption \ref{ass_short}, the formula \eqref{eq_X_onleft} for $A_\star$
on $\H_{\rm fin}$, and a similar formula with the operator $Q$ on the right (recall
that $Q$ is the position operator defined in \eqref{eq_def_Q}), one obtains that the
operator on the r.h.s. of \eqref{eq_first_step} defined as
$$
D_\star:=j_\star S(C-C_\star)\;\!j_\star A_\star\;\!j_\star
-j_\star A_\star\;\!j_\star S(C-C_\star)\;\!j_\star
$$
extends continuously to an element of $\B(\H)$. Since $\H_{\rm fin}$ is dense in
$\dom(A)$, this implies that
$j_\ell(U-U_\ell)\;\!j_\ell+j_{\rm r}(U-U_{\rm r})\;\!j_{\rm r}\in C^1(A)$.
(iii) To show that
$
j_\ell(U-U_\ell)\;\!j_\ell+j_{\rm r} (U-U_{\rm r})\;\!j_{\rm r}
\in C^{1+\varepsilon}(A)
$,
it remains to check that
$$
\big\|\e^{-itA}D_\star\e^{itA}-D_\star\big\|_{\B(\H)}
\le{\rm Const.}\;\!t^\varepsilon\quad\hbox{for all $t\in(0,1)$.}
$$
But, algebraic manipulations as presented in \cite[p.~325-326]{ABG96} show that for
all $t\in(0,1)$
\begin{align*}
\big\|\e^{-itA}D_\star\e^{itA}-D_\star\big\|_{\B(\H)}
&\le{\rm Const.}\;\!\big(\|\sin(tA)D_\star\|_{\B(\H)}
+\|\sin(tA)(D_\star)^*\|_{\B(\H)}\big)\\
&\le{\rm Const.}\;\!\big(\|tA\;\!(tA+i)^{-1}D_\star\|_{\B(\H)}
+\|tA\;\!(tA+i)^{-1}(D_\star)^*\|_{\B(\H)}\big).
\end{align*}
Furthermore, if we set $A_t:=tA\;\!(tA+i)^{-1}$ and
$\Lambda_t:=t\langle Q\rangle(t\langle Q\rangle+i)^{-1}$, we obtain that
$$
A_t=\big(A_t+i(tA +i)^{-1}A\;\!\langle Q\rangle^{-1}\big)\Lambda_t
$$
with $A\langle Q\rangle^{-1}\in\B(\H)$ due to
\eqref{eq_J_A_0_J_star}-\eqref{eq_X_onleft}. Thus, since
$\big\|A_t+i(tA +i)^{-1}A\;\!\langle Q\rangle^{-1}\big\|_{\B(\H)}$ is bounded by a
constant independent of $t\in(0,1)$, it is sufficient to prove that
$$
\|\Lambda_t D_\star\|_{\B(\H)}+\|\Lambda_t(D_\star)^*\|_{\B(\H)}
\le{\rm Const.}\;\!t^\varepsilon\quad\hbox{for all $t\in(0,1)$.}
$$
Now, this estimate will hold if we show that the operators
$\langle Q\rangle^\varepsilon D_\star$ and $\langle Q\rangle^\varepsilon(D_\star)^*$
defined in the form sense on $\H_{\rm fin}$ extend continuously to elements of
$\B(\H)$. For this, we fix $\varepsilon\in(0,1)$ with
$\varepsilon\le\min\{\varepsilon_\ell,\varepsilon_{\rm r}\}$, and note that
$\langle Q\rangle^{1+\varepsilon}(C-C_\star)\;\!j_\star\in\B(\H)$. With this inclusion
and the fact that $\langle Q\rangle^{-1}A_\star$ defined in the form sense on
$\H_{\rm fin}$ extend continuously to elements of $\B(\H)$, one readily obtains that
$\langle Q\rangle^\varepsilon D_\star$ and $\langle Q\rangle^\varepsilon(D_\star)^*$
defined in the form sense on $\H_{\rm fin}$ extend continuously to elements of
$\B(\H)$, as desired.
\end{proof}
With what precedes, we can now prove our last two main results on $U$ which have been
stated in Section \ref{Sec_model}.
\begin{proof}[Proof of Theorem \ref{thm_U_smooth_walks}]
Theorem \ref{thm_U_smooth}, whose assumptions are verified in Proposition
\ref{prop_Mourre} and Lemma \ref{lemma_C_1_epsilon}, implies that each $T\in\B(\H,\G)$
which extends continuously to an element of
$\B\big(\dom(\langle A\rangle^s)^*,\G\big)$ for some $s>1/2$ is locally $U$-smooth on
$\Theta\setminus\sigma_{\rm p}(U)$. Moreover, we know from the proof of of Lemma
\ref{lemma_C_1_epsilon} that $\dom(Q)\subset\dom(A)$. Therefore, we have
$\dom(\langle Q\rangle^s)\subset\dom(\langle A\rangle^s)$ for each $s>1/2$, and it
follows by duality that
$
\dom(\langle A\rangle^s)^*
\subset\dom(\langle Q\rangle^s)^*
\equiv\dom(\langle Q\rangle^{-s})
$
for each $s>1/2$. In consequence, any operator $T\in\B(\H,\G)$ which extends
continuously to an element of $\B\big(\dom(\langle Q\rangle^{-s}),\G\big)$ some
$s>1/2$ also extends continuously to an element of
$\B\big(\dom(\langle A\rangle^s)^*,\G\big)$. This concludes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm_spectrum_U_walks}]
The claim follows from Theorem \ref{thm_spec_U}, whose hypotheses are verified in
Lemma \ref{lemma_C_1_epsilon} and Proposition \ref{prop_Mourre}.
\end{proof}
|
\section{Introduction}
The geometric rigidity estimate for gradient fields proved in~\cite{FJM} plays a crucial role in nonlinear elasticity theory. However, in the study of lattices with dislocations, a geometric rigidity estimate for incompatible fields (i.e., fields not arising from gradients) becomes necessary (cf. e.g.~\cite{MSZ} and~\cite{LL}). In~\cite{MSZ}, the authors proved a (\emph{scaling invariant}) version of the geometric rigidity theorem in~\cite{FJM} for incompatible fields in dimension $2$ for the critical exponent.\\
In this work we give a proof of the analogous result in dimension $\ge 3$, for the supercritical regime $p > 1^* = \frac{n}{n-1}$ (Theorem~\ref{thm:rig_LL_1}). The approach is to write down an incompatible field as the sum of a compatible term, for which we can use the classical geometric rigidity from~\cite{FJM} and a remainder, which is the $L^p$ norm of a weakly singular operator (the \emph{averaged linear homotopy operator}), whose derivative is a Calder\'on-Zygmund operator. This allows to give the bounds in the supercritical case. On the other hand, for the critical exponent we can still use the weak geometric rigidity estimate proved in~\cite{CDM} in order to find a scaling invariant estimate for the weak-$L^{1^*}$ norm (Theorem~\ref{thm:useless1}). From Theorem~\ref{thm:useless1}, we deduce directly in Proposition~\ref{prop:curl_bounds_D_SOn} that the $\Curl$ of a matrix field $A \in L^{1^*, \infty}(\Omega)^{n\times n}$ (where $\Omega$ is an open bounded set in $\mathbb{R}^n$) taking values in $SO(n)$ bounds its gradient.
\section{Notations and Preliminaries}
In what follows, $C$ will denote a (universal) constant whose value is allowed to change from line to line. We put $\widehat{x}:=\frac{x}{\modulus{x}}$, while $L^p(U, \Lambda^r)$ ($W^{m, p}(U, \Lambda^r)$) denotes the space of $r$-forms on $U$ whose coefficients are $L^p$ ($W^{m,p }$) functions. Moreover, recall that we can identify a tensor field $A\in L^1(\Omega)^{n\times n}$ with a vector of $1$-forms of length $n$, that is with $\omega:=\rB{\omega^i}_{i=1}^n$, $\omega^i = A^i_j \ensuremath{\mathrm{d}} x^j$, and its $\Curl$ with $\ensuremath{\mathrm{d}} \omega$ (or, more precisely, with $\rB{\star \ensuremath{\mathrm{d}} \omega}^{\flat}$), given by \[\displaystyle \ensuremath{\mathrm{d}} \omega^i = \sum_{j < k} \rB{\frac{\partial A^i_j}{\partial x^k} - \frac{\partial A^i_k}{\partial x^j}} \ensuremath{\mathrm{d}} x^j \wedge \ensuremath{\mathrm{d}} x^k .\]
We recall that a real-valued function $f$ from a measure space $(X, \mu)$ is in $L^{p, \infty}(X, \mu)$ or ($L^p_w(X, \mu)$) if
\[
\norm{f}_{L^{p, \infty}(X, \mu)}:=\sup_{t > 0} t \mu\rB{\cB{x \in X\biggr| \modulus{f(x)} > t}}^{\frac{1}{p}} < \infty.
\]
Is easy to check that $\norm{\cdot}_{L^{p, \infty}}(X, \mu)$ is only a quasi-norm, that is the triangle inequality holds just in the weak form
\[
\norm{f+g}_{L^{p, \infty}(X, \mu)} \le C_{p} \rB{\norm{f}_{L^{p, \infty}(X, \mu)} + \norm{g}_{L^{p, \infty}(X, \mu)}}.
\]
We write $L^{p, \infty}(\Omega)$ for $L^{p, \infty}(\Omega, \modulus{\cdot})$, when $\Omega\subset \mathbb{R}^n$ and $\modulus{\cdot}$ is the Lebesgue measure.\\
We recall the
\begin{definition}
Let $U \subset \mathbb{R}^n$ be a star-shaped domain with respect to the point $y \in U$. The \emph{linear homotopy operator} at the point $y$ is the operator
\[
k_y= k_{y, r} : \Omega^r(U) \to \Omega^{r-1}(U),
\]
defined as
\[
(k_y \omega) (x):=\int_0^1{s^{r-1} \omega(sx + (1-s)y)\zak (x-y)\ensuremath{\mathrm{d}} s},
\]
\end{definition}
where $(\omega(x)\zak v)\sB{v_1, \cdots v_{n-1}}:=\omega(x)\sB{v, v_1, \cdots, v_{n-1}}$. It is well known that the linear homotopy operator satisfies
\begin{equation}
\label{eq:lho1}
\omega = k_{y, r+1} \ensuremath{\mathrm{d}} \omega + \ensuremath{\mathrm{d}} k_{y, r} \omega\quad \forall \omega \in \Omega^r(U).
\end{equation}
In order to get more regularity, we consider the following \emph{averaged} linear homotopy operator on $B := B(0, 1)$, which coincides with the one introduced by Iwaniec and Lutoborski in~\cite{IL}, except for the choice of the weight function:
\[
T = T_r : \Omega^r(B) \to \Omega^{r-1}(B),
\]
\[
T\omega(x):=\int_B \phi(y) \rB{k_y\omega}(x) \ensuremath{\mathrm{d}} y,
\]
where $\phi \in \mathcal{C}^{\infty}_c(B(0, 2))$ is a positive cut-off function, with $\phi \equiv 1$ in $B$ and
\[\max\cB{\norm{\phi}_{L^{\infty}(\mathbb{R}^n)}, \norm{\nabla \phi}_{L^{\infty}(\mathbb{R}^n)}}\le 3.\]
Clearly,~\eqref{eq:lho1} holds for $T$ as well:
\begin{equation}
\label{eq:lho2}
\omega = T\ensuremath{\mathrm{d}} \omega + \ensuremath{\mathrm{d}} T \omega.
\end{equation}
An essential result is the rigidity estimate due to Friesecke, James and M\"uller:
\begin{theorem}[~\cite{FJM}]
\label{thm:FJM}
Let $\Omega\subset\mathbb{R}^n$ be a bounded Lipschitz domain, $n \ge 2$, and let $1<p<\infty$. There exists a constant $C = C(p, \Omega)$ such that for every $u \in W^{1, 2}(\Omega)$ there exists a rotation $R \in SO(n)$ such that
\[
\norm{\nabla u - R}_{L^p(\Omega)^{n\times n}} \le C\norm{\dist(\nabla u, SO(n))}_{L^p(\Omega)^{n\times n}}.
\]
\end{theorem}
For weak-$L^p$ estimate, we shall need the following theorem proved by Conti, Dolzmann and M\"uller:
\begin{theorem}[~\cite{CDM}]
\label{thm:cdm}
Let $p \in (1, \infty)$ and $\Omega\subset\mathbb{R}^n$ be a bounded connected domain. There exists a constant $C > 0$ depending only on $p, n$ and $\Omega$ such that for every $u \in W^{1, 1}(\Omega)^n$ such that $\dist(\nabla u, SO(n)) \in L^{p, \infty}(\Omega)^{n\times n}$ there exists a rotation $R \in SO(n)$ such that
\begin{equation}
\label{eq:cdm}
\norm{\nabla u - R}_{L^{p, \infty}(\Omega)^{n \times n}} \le C \norm{\dist(\nabla u, SO(n))}_{L^{p, \infty}(\Omega)^{n\times n}}.
\end{equation}
\end{theorem}
We also recall that, as proved in~\cite{IL}, $T$ satisfies (for smooth forms $\omega$) the pointwise bound
\begin{equation}
\label{eq:bound_LHO}
\modulus{T\omega(x)} \le C_{n, r} \int_B \frac{\modulus{\omega(y)}}{\modulus{x-y}^{n-1}} \ensuremath{\mathrm{d}} y.
\end{equation}
Indeed, for $\omega = \omega_{\alpha} \ensuremath{\mathrm{d}} x^{\alpha}\in \Omega^r(B)$ we have
\[
T\omega (x) = \rB{\int_B \ensuremath{\mathrm{d}} y \phi(y) \int_0^1 t^{r-1} \scal{x-y}{e_i} \omega_{\alpha}(tx+(1-t)y)} \ensuremath{\mathrm{d}} x^{\alpha} \zak e_i.
\]
We then make the substitution $\Phi(y, t):=\rB{tx+(1-t)y, \frac{t}{1-t}} \equiv (z(t, y), s(t))$, $\Phi: B(0, 1)\times (0, 1)\to B(0,1) \times (0, \infty)$, which gives
\[
\begin{split}
T\omega(x)
&= \rB{\int_B \ensuremath{\mathrm{d}} z \omega_{\alpha}(z)\frac{\scal{x-z}{e_i}}{\modulus{x-z}^n} \int_0^2 s^{r-1}(1+s)^{n-r}\phi(z+s\widehat{z - x})} \ensuremath{\mathrm{d}} x^{\alpha}\zak e_i \equiv\\
&\equiv \rB{\int_B K^i_r(z, x-z) \omega_{\alpha}(z) \ensuremath{\mathrm{d}} z} \ensuremath{\mathrm{d}} x^{\alpha} \zak e_i,
\end{split}
\]
where
\[
K^i_r(x, h):=\frac{\scal{h}{e_i}}{\modulus{h}^n} \int_0^2 s^{r-1} (1+s)^{n-r} \phi(x-s\widehat{h}) \ensuremath{\mathrm{d}} s,
\]
and we noticed that, since $\phi$ has compact support, the integral from $0$ to $\infty$ actually reduces to an integral over a finite interval. That is, we get~\eqref{eq:bound_LHO}. It also follows easily from~\eqref{eq:bound_LHO} that $T$ is a compact operator from $L^p(B, \Lambda^r)$ to $L^p(B, \Lambda^{r-1})$. Moreover, by density,~\eqref{eq:lho2} extends to every differential form $\omega \in W^{1, p}(B, \Lambda^r)$, and to every differential form $\omega \in L^1(B, \Lambda^r)$ whose differential is a bounded Radon measure, $\ensuremath{\mathrm{d}} \omega \in \mathcal{M}_b(B, \Lambda^{r+1})$.
\section{Proof of the Main Results}
Using the homotopy operator, we get the following weak-$L^p$ geometric rigidity estimate for incompatible fields:
\begin{theorem}
\label{thm:useless1}
Let $1^* = 1^*(n):=\frac{n}{n-1}$, and let $B \subset \mathbb{R}^n$ be the unit ball of $\mathbb{R}^n$.
There exists a constant $C = C(n) >0 $ such that for every $A \in L^{p^{*}}(B)$ whose $\Curl(A)$ is a vector measure on $U$ with bounded total variation and whose support is contained in $B$, i.e. $\spt \Curl(A) \Subset B$, there exist a rotation $R \in SO(n)$ such that
\[
\norm{A - R}_{L^{1^*, \infty}(B)} \le C\rB{\norm{\dist(A, SO(n))}_{L^{1^*, \infty}(B)} + \modulus{\Curl(A)}(B)}.
\]
\end{theorem}
\begin{proof}
Take any measurable subset $E \subset B$, and let $r > 0$ be such that $\modulus{B(0, r)} = \modulus{E}$. Then, using~\eqref{eq:bound_LHO} and the Hardy-Littlewood inequality
\[
\begin{split}
\int_E \ensuremath{\mathrm{d}} x \modulus{(T\omega)(x)} & \le C \int_E \ensuremath{\mathrm{d}} x \int_B \ensuremath{\mathrm{d}} y \frac{\modulus{\omega(y)}}{\modulus{x - y}^{n-1}} = \\
&=C \int_B \ensuremath{\mathrm{d}} y \modulus{\omega(y)} \int_E \frac{\ensuremath{\mathrm{d}} x}{\modulus{x - y}^{n-1}} \le\\
&\le C \int_B \ensuremath{\mathrm{d}} y \modulus{\omega(y)} \int_{\mathbb{R}^n} \rchi_{E - x}(y)\frac{\ensuremath{\mathrm{d}} y}{ \modulus{y}^{n-1}} \le\\
&\le C \int_B \ensuremath{\mathrm{d}} y \modulus{\omega(y)} \int_{\mathbb{R}^n} \rchi_{B(0, r)} \frac{\ensuremath{\mathrm{d}} y}{\modulus{y}^{n-1}} \le \\
&= C \int_B \ensuremath{\mathrm{d}} y \modulus{\omega(y)} \int_0^r \ensuremath{\mathrm{d}} t \int_{\partial B(0, t)} \frac{\ensuremath{\mathrm{d}} y}{t^{n-1}} = \\
&= C r \norm{\omega}_{L^1(B)} = C \modulus{E}^{\frac{1}{n}} \norm{\omega}_{L^1(B)}.
\end{split}
\]
This gives immediately
\[
\norm{T\omega}_{L^1(B)} \le C_n \norm{\omega}_{L^1(B)},
\]
and thus, using~\eqref{eq:lho2}, $\norm{A - T\ensuremath{\mathrm{d}} A}_{L^1(B)} \le C \norm{\ensuremath{\mathrm{d}} A}_{L^1(B)}$, which extends immediately by density in the case when $\ensuremath{\mathrm{d}} A$ is a vector measure with bounded total variation. Choosing $E = \cB{x \in B \biggr| \modulus{T\omega(x)} > t}$, for $t > 0$
\[
t \modulus{E} \le \int_E \modulus{T\omega(x)} \ensuremath{\mathrm{d}} x \le C \modulus{E}^{\frac{1}{n}} \modulus{\ensuremath{\mathrm{d}} A}(B).
\]
Passing to the supremum over $t > 0$, we find
\begin{equation}
\label{eq:weak_Lp_lho}
\norm{T\ensuremath{\mathrm{d}} A}_{L^{1^*, \infty}(B)} \le C_n \modulus{\ensuremath{\mathrm{d}} A}(B).
\end{equation}
Since $B$ is convex and $d(A - T\ensuremath{\mathrm{d}} A) = \ensuremath{\mathrm{d}}^2 TA = 0$, we can find a function $g$ such that $\ensuremath{\mathrm{d}} g = A - T\ensuremath{\mathrm{d}} A$. From the estimates proven, is possible to apply Theorem~\ref{thm:cdm} to $g$ and find
\[
\norm{\ensuremath{\mathrm{d}} g - R}_{L^{1^*, \infty}(B)} \le C \norm{\dist(\ensuremath{\mathrm{d}} g, SO(n))}_{L^{1^*, \infty}(B)}.
\]
But
\[
\norm{\ensuremath{\mathrm{d}} g - R}_{L^{1^*, \infty}(B)} \ge C \norm{A - R}_{L^{1^*, \infty}(B)} - \norm{T \ensuremath{\mathrm{d}} A}_{L^{1^*, \infty}(B)}
\]
and
\[
\norm{\dist(\ensuremath{\mathrm{d}} g, SO(n))}_{L^{1^*, \infty}(B)} \le \norm{\dist(A, SO(n))}_{L^{1^*, \infty}(B)} + \norm{T\ensuremath{\mathrm{d}} A}_{L^{1^*, \infty}(B)}.
\]
In particular,
\[
\norm{A - R}_{L^{1^*, \infty}(B)} \le C\rB{\norm{\dist(A, SO(n))}_{L^{1^*, \infty}(B)} + \modulus{\Curl(A)}(B)}.\qedhere
\]
\end{proof}
We now give another estimate for $L^p$ norms. It requires an $L^{\infty}$-bound on the matrix field $A$, which is natural in the context of the theory of elasticity.
\begin{theorem}
\label{thm:rig_LL_1}
Let $n \ge 3$, $1^*:=1^*(n):=\frac{n}{n-1}$, $p \in [1^*, 2]$ and fix $M > 0$. There exists a constant $C = C(n, M, p) > 0$, depending only on the dimension $n$, the exponent $p$ and the constant $M$, such that for every $A \in L^{\infty}(B)$, with $\norm{A}_{\infty} \le M$ and $\Curl(A) \in \mathcal{M}_b(B, \Lambda^2)$, $B:=B(0, 1)$, there exists a corresponding rotation $R \in SO(n)$ for which, if $p > 1^*$
\begin{equation}
\label{eq:rig_LL_p}
\int_B \modulus{A - R}^p \ensuremath{\mathrm{d}} x \le C\rB{\int_B\dist^{p}(A, SO(n))\ensuremath{\mathrm{d}} x + \modulus{\Curl(A)}^{1^*}(B)},
\end{equation}
while, if $p = 1^*$,
\begin{equation}
\label{eq:rig_LL}
\begin{split}
\int_B \modulus{A - R}^{1^*} \ensuremath{\mathrm{d}} x \le& C\int_B\dist^{1^*}(A, SO(n))\ensuremath{\mathrm{d}} x + \\
& + C\modulus{\Curl(A)}^{1^*}(B)\cB{\modulus{\log\rB{\modulus{\Curl(A)}(B)}}+1}.
\end{split}
\end{equation}
\end{theorem}
\begin{remark}
\label{rmk:everything_pointless}
The constant $C$ in~\eqref{eq:rig_LL} is \emph{not} scaling invariant in the critical regime $p = 1^*$.
\end{remark}
\begin{proof}[Proof of Theorem~\ref{thm:rig_LL_1}]
Without loss of generality, we can assume $T\ensuremath{\mathrm{d}} A$ not identically constant. Indeed, if $T\ensuremath{\mathrm{d}} A$ is identically constant, from the identity $T\ensuremath{\mathrm{d}} A = A - \ensuremath{\mathrm{d}} T A$, we see that $\ensuremath{\mathrm{d}} A = 0$, hence the result follows applying Theorem~\ref{thm:FJM}. As in the proof of Theorem~\ref{thm:useless1}, applying Theorem~\ref{thm:FJM} (and using $\modulus{a - b}^p \ge 2^{1-p}\modulus{a}^p - \modulus{b}^p$) we find a rotation $R \in SO(n)$ for which the inequality
\begin{equation}
\label{eq:useless11}
\int_B \modulus{A - R}^p \ensuremath{\mathrm{d}} x \le C_n\rB{\int_B\modulus{\dist(A, SO(n))}^p\ensuremath{\mathrm{d}} x + \int_B \modulus{T\ensuremath{\mathrm{d}} A(x)}^p \ensuremath{\mathrm{d}} x}
\end{equation}
holds. We then just need to estimate the last term in the right hand side of~\eqref{eq:useless11}. For, fix a $\Lambda > 1$ (to be chosen later), and define the integrals
\[
I:=\int_{\modulus{T\ensuremath{\mathrm{d}} A } > \Lambda} \modulus{T\ensuremath{\mathrm{d}} A}^p \ensuremath{\mathrm{d}} x,\qquad II:=\int_{\modulus{T\ensuremath{\mathrm{d}} A}\le \Lambda} \modulus{T \ensuremath{\mathrm{d}} A}^p \ensuremath{\mathrm{d}} x.
\]
We now give an estimate for $I$. Firstly, we recall that $T$ is a bounded operator from $L^p(B, \Lambda^r)$ into $W^{1, p}(B, \Lambda^{r+1})$, whenever $p\in (1, \infty)$ (cf. \cite[Proposition 4.1]{IL}). Moreover, $T\ensuremath{\mathrm{d}} A = A - \ensuremath{\mathrm{d}} TA $, and $\nabla T = S_1 + S_2$, where $S_1$ is a ``weakly'' singular operator which maps continuously $L^{\infty}$ into itself, while $S_2$ is a Calder\'on-Zygmund operator (cf. \cite[Proposition 4.1]{IL}). In particular,
\[
\norm{T \ensuremath{\mathrm{d}} A}_{\text{BMO}}\le C_n \norm{A}_{\infty} \le C_n M,
\]
where $C_n > 0$ is a constant depending only on the dimension. Now, we can write
\begin{equation}
\label{eq:useless777}
I = \Lambda^{p - \pst}\Lambda^{\pst} \modulus{\cB{\modulus{T \ensuremath{\mathrm{d}} A} > \Lambda}} + I',\qquad I':=\int_{\Lambda}^{\infty} \lambda^{p - 1}\modulus{\cB{\modulus{T\ensuremath{\mathrm{d}} A} > \lambda}} \ensuremath{\mathrm{d}} \lambda.
\end{equation}
Clearly,
\[
\Lambda^{\pst} \modulus{\cB{\modulus{T \ensuremath{\mathrm{d}} A} > \Lambda}} \le \norm{T\ensuremath{\mathrm{d}} A}_{L^{\pst, \infty}}^{\pst} \le C \modulus{\ensuremath{\mathrm{d}} A}(B)^{\pst}.
\]
We now take a Calder\'on-Zygmund decomposition of $F(x):=\modulus{T\ensuremath{\mathrm{d}} A(x)}^p$: namely, we find a function $g \in L^{\infty}$, with $\norm{g}_{\infty} \le 2^{-n}\Lambda^p$ and disjoint cubes $\cB{Q_j}_{j \ge 1}$ such that, if $b:=\sum_{j \ge 1} \rchi_{Q_j} F$,
\[
\begin{cases}
F = g + b,\\
2^{-n} \Lambda^p < \fint_{Q_j} F \ensuremath{\mathrm{d}} x\le \Lambda^p \quad \rB{\text{Jensen }\Rightarrow \modulus{\fint_{Q_j} T\ensuremath{\mathrm{d}} A(x) \ensuremath{\mathrm{d}} x} \le \Lambda},\\
\modulus{\bigcup_{j \ge 1} Q_j} < \frac{2^n}{\Lambda^p} \int \modulus{T \ensuremath{\mathrm{d}} A}^p \ensuremath{\mathrm{d}} x.
\end{cases}
\]
With such a decomposition, outside the cubes $Q_j$, $\modulus{T \ensuremath{\mathrm{d}} A}^p = \modulus{g(x)} \le 2^{-n}\Lambda^p \le \Lambda^p$. Hence, using the John-Nirenberg inequality and the elementary estimate
\[
\int_x^{\infty} \lambda^q e^{-\lambda} \ensuremath{\mathrm{d}} \lambda \le e^{-x}(1+x),\quad \forall q \le 1 \text{ and } x \ge 1,
\]
we find that (provided $p \le 2$)
\begin{equation}
\label{eq:useless888}
\begin{split}
I' &= \int_{\Lambda}^{\infty} \lambda^{p - 1} \sum_{j \ge 1} \modulus{\cB{x \in Q_j \biggr| \modulus{T\ensuremath{\mathrm{d}} A} > \lambda}} \ensuremath{\mathrm{d}} \lambda \le \\
&\le \int_{\Lambda}^{\infty} \lambda^{p - 1} \sum_{j \ge 1} \modulus{\cB{x \in Q_j\biggr| \modulus{T\ensuremath{\mathrm{d}} A(x) - \fint_{Q_j} T\ensuremath{\mathrm{d}} A \ensuremath{\mathrm{d}} x} > \lambda - \Lambda}} \ensuremath{\mathrm{d}} \lambda \le \\
&\le C_1 \int_{\Lambda}^{\infty} \lambda^{p - 1} \rB{\sum_{j \ge 1} \modulus{Q_j}} \exp\rB{-C_2 \frac{\lambda - \Lambda}{\norm{T\ensuremath{\mathrm{d}} A}_{\text{BMO}}}} \ensuremath{\mathrm{d}} \lambda < \\
&< C_1 \frac{2^n}{\Lambda^p} \rB{\int \modulus{T\ensuremath{\mathrm{d}} A}^p} e^{C_2 \frac{\Lambda}{\norm{T\ensuremath{\mathrm{d}} A}_{\text{BMO}}}} \rB{\frac{\norm{T\ensuremath{\mathrm{d}} A}_{\text{BMO}}}{C_2}}^{p} \int_{\frac{C_2}{\norm{T\ensuremath{\mathrm{d}} A}_{\text{BMO}}} \Lambda}^{\infty}\lambda^{p - 1} e^{-\lambda} \ensuremath{\mathrm{d}} \lambda \le \\
&\le C_1 \frac{2^n}{\Lambda^p} \rB{\int \modulus{T\ensuremath{\mathrm{d}} A}^p} \rB{\frac{\norm{T\ensuremath{\mathrm{d}} A}_{\text{BMO}}}{C_2}}^p \rB{1 + \frac{C_2}{\norm{T\ensuremath{\mathrm{d}} A}_{\text{BMO}}}\Lambda} \le \\
&\le C_{n, M} \rB{\int \modulus{T \ensuremath{\mathrm{d}} A}^p} \frac{1 + \Lambda}{\Lambda^p}.
\end{split}
\end{equation}
Hence, if we choose $\Lambda$ big enough (depending only on $n$ and $M$) in~\eqref{eq:useless888},
\begin{equation}
\label{eq:useless999}
I' \le \frac{1}{2} \int \modulus{T\ensuremath{\mathrm{d}} A}^p.
\end{equation}
Let us now estimate $II$. If $p > \pst$, we can write
\[
\begin{split}
\int_{\modulus{T\ensuremath{\mathrm{d}} A} \le \Lambda} \modulus{T\ensuremath{\mathrm{d}} A}^p \ensuremath{\mathrm{d}} x &= \int_{1 < \modulus{T\ensuremath{\mathrm{d}} A}\le \Lambda} \modulus{T\ensuremath{\mathrm{d}} A}^p \ensuremath{\mathrm{d}} x + \sum_{j \ge 0} \int_{2^{-j-1} < \modulus{T\ensuremath{\mathrm{d}} A} \le 2^{-j}} \le \\
&\le C \cB{\Lambda^p \modulus{\ensuremath{\mathrm{d}} A}^{\pst}(B) + \sum_{j \ge 0} 2^{-(j+1)p} \modulus{\cB{\modulus{T\ensuremath{\mathrm{d}} A} > 2^{-(j+1)}}} } \le \\
&\le C\modulus{\ensuremath{\mathrm{d}} A}^{\pst}(B)\rB{\Lambda^p + \sum_{j \ge 0} 2^{-j(\pst - p)}} \le \\
&\le C(n, p, M) \modulus{\ensuremath{\mathrm{d}} A}^{\pst}(B),
\end{split}
\]
which gives~\eqref{eq:rig_LL_p}. In the case $p = \pst$, we are going to make use of the increasing convex function $\Psi$, defined as the linear (convex) continuation of $t \mapsto t^p$ for $t \ge \Lambda$:
\[
\Psi(t):=\begin{cases}
t^{\pst} & \text{if } t \le \Lambda,\\
\pst \Lambda^{\pst - 1} t + (1-\pst) \Lambda^{\pst}& \text{if }t \ge \Lambda.
\end{cases}
\]
\begin{equation}
\label{eq:uselessB}
\begin{split}
II &\le \int_B \Psi(\modulus{T\ensuremath{\mathrm{d}} A(x)}) \ensuremath{\mathrm{d}} x \le \int_B \Psi\rB{\fint_B \frac{C\modulus{\ensuremath{\mathrm{d}} A}(B) \ensuremath{\mathrm{d}} \modulus{\ensuremath{\mathrm{d}} A}(y)}{\modulus{x-y}^{n-1}}} \le\\
&\le \int_B \fint \Psi\rB{\frac{C\modulus{\ensuremath{\mathrm{d}} A}(B)}{\modulus{x-y}^{n-1}}} \ensuremath{\mathrm{d}} \modulus{\ensuremath{\mathrm{d}} A}(y) \ensuremath{\mathrm{d}} x = \\
&= \fint_B \ensuremath{\mathrm{d}} \modulus{\ensuremath{\mathrm{d}} A}(y) \int_B \Psi\rB{\frac{C\modulus{\ensuremath{\mathrm{d}} A}(B)}{\modulus{x-y}^{n-1}}} \ensuremath{\mathrm{d}} x \le \\
&\le \int_{B(0, 2)} \Psi\rB{\frac{C\modulus{\ensuremath{\mathrm{d}} A}(B)}{\modulus{z}^{n-1}}} \ensuremath{\mathrm{d}} z = C\int_0^2 \ensuremath{\mathrm{d}} \rho \rho^{n-1} \Psi\rB{\frac{C\modulus{\ensuremath{\mathrm{d}} A}(B)}{\rho^{n-1}}} = \\
&= \int_0^{C\rB{\modulus{\ensuremath{\mathrm{d}} A}(B)\Lambda^{-1}}^{\frac{1}{n-1}}} \rho^{n-1}\rB{1^*\Lambda^{\pst - 1} \frac{C\modulus{\ensuremath{\mathrm{d}} A}(B)}{\rho^{n-1}} + (1 - 1^*)\Lambda^{\pst}}\ensuremath{\mathrm{d}} \rho + \\
&\quad+ C \int_{C\rB{\modulus{\ensuremath{\mathrm{d}} A}(B)\Lambda^{-1}}^{\frac{1}{n-1}}}^2 \frac{\modulus{\ensuremath{\mathrm{d}} A}(B)^{1^*}}{\rho} \ensuremath{\mathrm{d}} \rho \le \\
&\le C \modulus{\ensuremath{\mathrm{d}} A}(B)^{1^*}\rB{1 + \modulus{\log\rB{\modulus{\ensuremath{\mathrm{d}} A}(B)}}}.
\end{split}
\end{equation}
Combining together~\eqref{eq:useless777}, ~\eqref{eq:useless999} and~\eqref{eq:uselessB}, we obtain~\eqref{eq:rig_LL}.
\end{proof}
\begin{remark}
The same conclusions can be obtained considering the operator defined by an average on the sphere:
\[
\tilde{T}\omega(x):=\int_{\mathbb{S}^{n-1}} \ensuremath{\mathrm{d}} \mathcal{H}^{n-1}(y) k_y \omega(x).
\]
\end{remark}
\begin{remark}
Using Korn's inequality instead of Theorem~\ref{thm:FJM}, one can easily prove the linear counterpart of Theorem~\ref{thm:rig_LL_1}.
\end{remark}
\begin{proposition}
\label{prop:curl_bounds_D_SOn}
Let $\Omega \subset \mathbb{R}^n$ be a bounded open set, and suppose $A \in L^2(\Omega)$ and $\spt(A) \Subset \Omega$. Consider a tessellation of $\mathbb{R}^n$ with cubes $\cB{Q^{(\rho)}_i}_{i \ge 1} \equiv \cB{Q(x_i, \rho)}$ of side $\rho$, and define $A_{\rho}$ as the piecewise constant function
\begin{equation}
\label{eq:def_A_rho}
A_{\rho}:=\sum_{i \ge 1} R^{(\rho)}_i \rchi_{Q_{\rho,i}},
\end{equation}
where the rotations $R^{(\rho)}_i$ are the ones given by Theorem~\ref{thm:useless1} applied to $A$ on the balls $B(x_i, \frac{3}{2}\rho)$. There exists a constant $C = C(n) > 0$, depending only on the dimension $n$, such that
\begin{equation}
\label{eq:prop1_1}
\frac{1}{\rho}\norm{A - A_{\rho}}_{L^1(\Omega)} + \modulus{DA_{\rho}}(\Omega) \le C\rB{\rho^{\frac{n-2}{2}}\norm{\dist(A, SO(n))}_{L^2(\Omega)} + \modulus{\Curl(A)}(\Omega)}.
\end{equation}
In particular, if $A \in SO(n)$ almost everywhere,
\begin{equation}
\label{eq:prop1_2}
\modulus{DA}(\Omega) \le C \modulus{\Curl(A)}(\Omega).
\end{equation}
That is, $A \in BV(\Omega, SO(n))$ provided $\modulus{\Curl(A)}(\Omega)$ is finite.
\end{proposition}
\begin{proof}
By definition, the rotations $R^{(\rho)}_i$ in~\eqref{eq:def_A_rho} satisfy
\[
\norm{A - R^{(\rho)}_i}_{L^{1^*, \infty}(Q^{(\rho)}_i)} \le C_n\rB{\norm{\dist(A, SO(n))}_{L^{1^*, \infty}} + \modulus{\Curl(A)}(2Q^{(\rho)}_i)}.
\]
Let $\phi \in \mathcal{C}^1_c(\Omega)$. Then
\[
\modulus{\int A_{\rho}\text{div}(\phi) \ensuremath{\mathrm{d}} x} \le \sum_{\substack{i, j \text{ s.t. }\\ \partial Q^{(\rho)}_i \cap \partial Q^{(\rho)}_j \ne \emptyset}} \rho^{n-1} \modulus{R^{(\rho)}_i - R^{(\rho)}_j}.
\]
Now, for any two adjacent cubes $Q^{(\rho)}_i$ and $Q^{(\rho)}_j$, take the rotation $R'_{\rho, i}$ given applying Theorem~\ref{thm:useless1} to the cube $2Q^{(\rho)}_i$. Then
\[
\begin{split}
\modulus{R^{(\rho)}_i - R^{(\rho)}_j} \rho^{n - 1} &\le \rB{\modulus{R^{(\rho)}_i - R'_{\rho, i}} + \modulus{R'_{\rho, i} - R^{(\rho)}_j}} \rho^{n-1} \le \\
&\le C_n \rB{\norm{R^{(\rho)}_i - R'_{\rho, i}}_{L^{\pst, \infty}(Q^{(\rho)}_i)} + \norm{R'_{\rho, i} - R^{(\rho)}_j}_{L^{\pst, \infty}(Q^{(\rho)}_j)}} \le \\
&\le C_n \rB{\norm{A - R^{(\rho)}_i}_{L^{\pst, \infty}(Q^{(\rho)}_i)} + \norm{A - R^{(\rho)}_j}_{L^{\pst, \infty}(Q^{(\rho)}_j)} + \norm{A - R'_{\rho, i}}_{L^{\pst, \infty}(2Q^{(\rho)}_i)}} \le \\
&\le C_n\rB{\norm{\dist(A, SO(n))}_{L^{1^*, \infty}(4Q^{\rho}_i)} +\modulus{\Curl(A)}(4Q^{(\rho)}_i)} \\
&\le C_n\rB{\rho^{\frac{n-2}{2}}\norm{\dist(A, SO(n))}_{L^2(4Q^{\rho}_i)} +\modulus{\Curl(A)}(4Q^{(\rho)}_i)}.
\end{split}
\]
Taking the supremum over $\phi$, since the cubes $4Q^{(\rho)}_i$ overlap only finitely many times, we obtain
\[
\modulus{DA_{\rho}}(\Omega) \le C_n\rB{\rho^{\frac{n-2}{2}}\norm{\dist(A, SO(n))}_{\Omega}+ \modulus{\Curl(A)}(\Omega)}.
\]
Moreover, from the definition of weak-$L^1$:
\[
\frac{1}{\rho} \int_{Q^{(\rho)}_i} \modulus{A - A_{\rho}} \ensuremath{\mathrm{d}} x \le C_n \norm{A - A_{\rho}}_{L^{\pst, \infty}(Q^{(\rho)}_i)} \le C_n\rB{\norm{\dist(A, SO(n))}_{L^{1^*, \infty}(4Q^{\rho}_i)} + \modulus{\Curl(A)}(4Q^{(\rho)}_i)}.
\]
This gives in particular~\eqref{eq:prop1_1}. Moreover
\[
\begin{split}
\norm{A - A_{\rho}}_{L^1(\Omega)} &\le \sum_{i\ge 1} \norm{A - A_{\rho}}_{L^1(Q^{(\rho)}_i)} \le C_n \rho \sum_{i \ge 1}\rB{\norm{A - A_{\rho}}_{L^{\pst, \infty}(2Q^{(\rho)}_i)} + \modulus{\Curl(A)}(2Q^{(\rho)}_i) }\le\\
&\le C_n \rho \rB{\norm{A - A_{\rho}}_{L^{\pst, \infty}(\Omega)} + \modulus{\Curl(A)}(\Omega)} \xrightarrow[\rho \to 0]{} 0.
\end{split}
\]
That is, $A_{\rho} \to A$ strongly in $L^1$. Thus, if we let $\rho\to 0$, we obtain~\eqref{eq:prop1_2} provided $A \in SO(n)$ almost everywhere.
\end{proof}
|
\section{Introduction}
Deep networks have been stunningly successful in many machine learning domains since the area was reinvigorated by the work of \citet{krizhevsky2012imagenet}. Despite their success, neural networks in general, have two major limitations:
\begin{enumerate}
\item relatively little is known about them theoretically, although this is changing. In particular, for which problems are deep networks more effective than shallow learners; and
\item deciding on a particular architecture is still an empirical question.
\end{enumerate}
Compare using a deep network for a particular problem with a support vector machine (SVM). The choices for a deep network include: number of layers, number of neurons per layer, activation function, gradient descent algorithm, gradient descent algorithm parameters (e.g. learning rate), which learning tricks to use (dropout, batch normalisation etc), weight initialisation parameters, etc. Getting any one of these wrong can cause the network learning to fail. For an SVM, on the other hand, one needs to choose the kernel function and the kernel function parameters (usually few) and the slack variable parameter $C$. Similarly, other machine learning methods such as decision trees, boosting, random forests are much easier to use, albeit often with worse performance than the best neural network model.
This paper goes some way to addressing both of these limitations and offers a theoretical result and an applied result. The main theoretical result is a new upper bound for the number of neurons required in a deep network to approximate a radially symmetric function. The main applied result uses the theoretical result to construct a deep network approximation of an SVM that can be further trained using back propagation and which often results in improved performance.
\section{Related Work}
It is difficult to deny the empirical success of deep networks. It is still the case that relatively little is known theoretically about their fundamental abilities, although this has been changing over the last few years. Nevertheless, most work consists of existence proofs that do not lead directly to new methods for building or training networks. For example,
ReLU networks with $n_0$ inputs, $L$ hidden layers of width $n \ge n_0$ can compute functions that have $\Omega\left( (n/n_0)^{(L-1)n_0} n^{n_0} \right)$ linear regions compared to $\sum_{j=0}^{n_0} \binom{n}{j}$ for a shallow network \citep{montufar2014number}. More generally, \citet{telgarsky2016benefits} proved for semi-algebraic neurons (including ReLU, sigmoid etc), that networks exist with $\Theta(k^3)$ layers and $\Theta(k^3)$ nodes that require $\Omega(2^k)$ nodes to approximate with a network of $O(k)$ layers.
\citet{delalleau2011shallow} show that deep sum-product networks exist for which a shallow network would require exponentially more neurons to simulate. For convolutional arithmetic circuits (similar to sum-product networks), \citet{cohen2016expressive}, in an important result, show that ``besides a negligible (zero measure) set, all functions that can be realized by a deep network of polynomial size, require exponential size in order to be realized, or even approximated, by a shallow network.''
The above works, except for \citet{cohen2016expressive} focus on approximating deep networks with shallow networks, but do not indicate what problems are best attacked with deep networks.
For manifolds, \citet{basri2016efficient} show how deep networks can efficiently represent low-dimensional manifolds and that these networks are almost optimal, but they do not discuss limitations of shallow networks on the same problem. Somewhat similarly, \citet{shaham2016provable} show that depth-4 networks can approximate a function on a manifold where the number of neurons depends on the complexity of the function and the dimensionality of the manifold and only weakly on the embedding dimension. Again, they do not discuss the limitations of shallow networks for this problem. Importantly, both of these results are constructive and allow one to actually build the network.
\citet{szymanski2014deep} show that deep networks can approximate periodic functions of period $P$ over $\{0,1\}^N$ with $O(\log_2{N}-\log_2{P})$ parameters versus $O(P \log_2{N})$ for shallow. \citet{eldan2016power} show that 3-layer networks exist such that the network can approximate a radially symmetric function with $O(d^{19/4})$ neurons, that a 2-layer network requires at least $O(e^d)$ neurons.
Therefore evidence is building that deep networks are more powerful than their shallow counterparts in terms of the number of parameters or neurons required. Nevertheless, we shouldn't stop there because there is relatively little work linking this theory with practical applications of the same. In this work we directly extend the work of \citet{szymanski2014deep} and \citet{eldan2016power}. The latter work \citep{eldan2016power} is extended to deeper networks for approximating radially symmetric functions that require fewer parameters than their construction. The former \cite{szymanski2014deep} is extended by generalising their notion of folding transformations to work in multiple dimensions and more simply with ReLU networks. The proofs are constructive and allow us to build networks for approximating radially symmetric functions. These networks are used to approximate Gaussian kernel SVMs and the results show how we can further train these approximations to do better than the original SVM in many cases.
Our applied work is similar in spirit, but quite different in detail to other methods trying to combine deep learning and kernel learning. For example, \citet{wilson2016deep} use a deep network as the kernel in a Gaussian processes framework, but the choice of the deep network architecture remains with the practitioner. Our deep radial kernel could be used as the input kernel in that framework. We also note that multiple kernel learning \citep{lanckriet2004learning} and its descendants take a very different approach to the one we take here --- we make no attempt to learn a kernel matrix, nor to ensure positive semi-definiteness of the resulting network.
We start in Section \ref{sec-theory} with the main theoretical results, then proceed to approximating Gaussian kernel SVMs in Section \ref{sec-svm-approx}.
\section{Theory}
\label{sec-theory}
\subsection{Context and Notation}
A radially symmetric function is a function whose value is dependent on the norm of the input only. We are interested in $L$-Lipschitz functions $f$, $|f(x)-f(y)| \le L|x-y|$, as this covers many functions common in classification tasks. The number of dimensions of the input is $d$, and we assume that $f$ is constant outside a radius $R$. This is a similar context to that used by \citet{eldan2016power}. Further, we restrict ourselves to ReLU networks only, which is more restrictive than \citet{eldan2016power}, but allows us to explicitly construct the networks of interest.
Most proofs are only sketched in the main body of the paper. Detailed proofs are provided in the supplementary material.
\subsection{3 Layer Networks}
We start by stating a modified form of Lemma 18 from \citet{eldan2016power}:
\begin{restatable}{lemma}{lemep}
\label{lem18}%
\label{lem-3layer}
Modified form of Lemma 18 from \citet{eldan2016power}:\\
Let $\sigma(z) = \max(0,z)$. Let $f$ be an $L$-Lipschitz function supported on $[0, R]$. Then for any $\delta>0$, there exists a function $g$ expressible by a 3-layer network of width at most $\frac{6 d^2 R^2 + 3 R L}{\delta}$, such that
\begin{equation*}
\sup_{\myvec{x} \in \ensuremath{\mathbb{R}}^d}| g(\myvec{x})-f(||\myvec{x}||)| < \delta + L\sqrt{\delta}.
\end{equation*}
\end{restatable}
The proof follows the basic plan of \citet{eldan2016power} where the first layer is the input layer, the second layer approximates $x_i^2$ for each dimension $i$, and the third layer computes $\sum_i x_i^2$ and approximates $f$.
Since several sections of the second layer are doing the same thing (computing the square of their input), a weight-sharing corollary follows immediately where only one copy of the square approximation is needed.
\begin{corol}[3 Layer Weight Sharing]
Let $\sigma(z) = \max(0,z)$. Let $f$ be an $L$-Lipschitz function supported on $[0, R]$. Then for any $\delta>0$, there exists a function $g$ expressible by a 3-layer weight-sharing network with at most $\frac{6 d R^2 + 3 R L}{\delta}$ weights, such that
\begin{equation*}
\sup_{\myvec{x} \in \ensuremath{\mathbb{R}}^d}| g(\myvec{x})-f(||\myvec{x}||)| < \delta + L\sqrt{\delta}.
\end{equation*}
\end{corol}
\subsection{Deep Folding Networks}
\label{sec-folding}
In this section we show how folding transformations can be used to create a much deeper network with the same error, but many fewer weights than needed in Lemma \ref{lem-3layer}. A folding transformation is one in which half of a space is reflected about a hyperplane, and the other half remains unchanged. Figure \ref{fig-fold-2d} demonstrates how a sequence of folding transformations can transform a circle in 2D to a small sector. After enough folds, we can discard the almost zero coordinates to approximate the norm.
We will use this general idea to prove the following theorem:
\begin{restatable}{theorem}{main}
\label{thm:main}%
Let $\myvec{x} \in \ensuremath{\mathbb{R}}^d$, and $\sigma(z) = \max(0,z)$. Let $f$ be an $L$-Lipschitz function supported on $[0,R]$. Fix $L,\delta,R > 0$. There exists a function $g$ expressible by a $O(d\log_2(d) + \log_2(d)\log_2\left(\frac{R}{\sqrt{\delta}}\right))$ layer network where the number of weights, and number of neurons, $N_w, N_n = O(d^2 + d\log_2\left(\frac{R}{\sqrt{\delta}}\right) + \frac{3 RL}{\delta})$ , such that:
\begin{equation*}
\sup_{\myvec{x} \in \ensuremath{\mathbb{R}}^d} | g(\myvec{x}) - f(||\myvec{x}||)| < L\sqrt{\delta} + \delta
\end{equation*}
\end{restatable}
The approach taken here is a constructive one and specifies the architecture of the network needed to approximate $f$. In fact, all of the weights except those in the last layer are specified. The approach is somewhat different to that used to prove Lemma \ref{lem-3layer}. We build a sequence of layers to directly approximate $||\myvec{x}||$ and then approximate $f$ in the last layer. To build our layers, we need a few helper lemmas.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{Figs/fold2d_0.pdf}
\includegraphics[width=0.4\textwidth]{Figs/fold2d_1.pdf}\\
\includegraphics[width=0.4\textwidth]{Figs/fold2d_2.pdf}
\includegraphics[width=0.4\textwidth]{Figs/fold2d_3.pdf}\\
\includegraphics[width=0.4\textwidth]{Figs/fold2d_4.pdf}
\includegraphics[width=0.4\textwidth]{Figs/fold2d_5.pdf}
\caption{Series of folding transformations for 2D. \label{fig-fold-2d}}
\end{figure}
\begin{restatable}[2D fold]{lemma}{lemtwodfold}
There exists a function $g: \ensuremath{\mathbb{R}}^2 \to \ensuremath{\mathbb{R}}^2$, expressible by a ReLU network with 4 ReLU units and 2 sum units that can compute a folding transformation about a line through the origin, represented by the unit direction vector $\myvec{l}=(l_x, l_y)^T$. The function $g$ is of the form:
\begin{equation*}
g(\myvec{x}) = \begin{cases}
\myvec{x} & \myvec{l} \cdot \myvec{x^\perp} > 0 \\
\begin{bmatrix}
l_x^2-l_y^2 & 2 l_x l_y \\
2 l_x l_y & l_y^2-l_x^2
\end{bmatrix}
\myvec{x} & \text{otherwise.}
\end{cases}
\end{equation*}
\label{lem-2d-fold}
\end{restatable}
The requisite ReLU network is shown in Figure \ref{fig-relu-fold-2d}. Only one of the nodes labeled $x_{-}$ ($y_{-}$) and $x_{+}$ ($y_{+}$) are active at any one time. Therefore there are four possible cases depending on which two nodes are active. Note that $x_{-}$ is active when $\myvec{l} \cdot \myvec{x^\perp} < 0$ and $x_{+}$ is active when $\myvec{l} \cdot \myvec{x^\perp} > 0$.
To approximate the 2D norm, we simply stack layers of the type shown in Figure \ref{fig-relu-fold-2d} with suitable choice of $l_x, l_y$ at each layer. Note that the summation nodes aren't required since they can be incorporated into the summations and weights of the next ReLU layer.
These 2D folds can be used to estimate the norm of a vector as per the following lemma.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{Figs/deep-fold.pdf}
\caption{A network to produce a 2D fold. \label{fig-relu-fold-2d}}
\end{figure}
\begin{restatable}[Approximate $||\myvec{x}||, x \in \ensuremath{\mathbb{R}}^2, ||x||<R$]{lemma}{lemapproxtwod}
There exists a function $g$, expressible by a ReLU network with no more than $\log_2\left(R \frac{\pi}{\delta}\right)$ layers and $4$ nodes per layer such that:
\begin{equation*}
\sup_{\myvec{x}\in \ensuremath{\mathbb{R}}^2, ||\myvec{x}||\le R} | g(\myvec{x}) - ||\myvec{x}|| | \le \delta
\end{equation*}
\label{lem-approx-2d}
\end{restatable}
\begin{proof}
The proof is short and simple. After $f$ layers, each data point will be within an angle of $\frac{\pi}{2^{f}}$ of the $x$-axis. Simple geometry and appropriate approximations leads to:
\begin{align*}
\delta &=||\myvec{x}||-||\myvec{x}||\cos\left(\frac{\pi}{2^{f}}\right)\\
&\le R\left(1-\cos\left(\frac{\pi}{2^{f}}\right)\right)\\
&\le R\left(2 \sin\left(\frac{\pi}{2^{f+1}}\right)\right)\\
&\le R\left(\frac{\pi}{2^{f}}\right)\\
f &\le \log_2\left(R \frac{\pi}{\delta}\right)
\end{align*}
\end{proof}
The following lemma generalises this construction to folds in $d$ dimensions.
\begin{restatable}[Approximate $||\myvec{x}||, x \in \ensuremath{\mathbb{R}}^d$]{lemma}{lemapproxvecx}
\label{lem-approx-vecx}
There exists a function $g$, expressible by a ReLU network with:
\begin{align*}
\small
N_l \le &\log_2(d)\\
&\log_2\left(\frac{R \pi}{\delta} \left[\lemfive{(d+1)}{d} \right]\right) \\
N_n \le &4(d-1) \\
&\log_2\left(\frac{R \pi}{\delta} \left[\lemfive{(d+1)}{d} \right]\right) \\
N_w \le& 8(d-1)\\
&\log_2\left(\frac{R \pi}{\delta} \left[\lemfive{(d+1)}{d} \right]\right)
\end{align*}
such that:
\begin{equation*}
\sup_{\myvec{x}\in \ensuremath{\mathbb{R}}^d, ||\myvec{x}||\le R} | g(\myvec{x}) - ||\myvec{x}|| | \le \delta
\end{equation*}
\end{restatable}
We note that a fold in a 2D plane in $\ensuremath{\mathbb{R}}^d$ will leave all coordinates perpendicular to the plane unchanged. We can therefore apply the approximation of Lemma \ref{lem-approx-2d} to pairs of input coordinates to produce $d/2$ new coordinates. Then apply the same reduction to produce $d/4$ coordinates and continue on this way until there is only one coordinate left. In effect, we are calculating the norm via the following scheme:
{\small
\begin{align*}
&\sqrt{x_1^2 + x_2^2 + x_3^2 + \cdots + x_d^2} =\\
&\sqrt{\sqrt{ \sqrt{x_1^2 + x_2^2}^2 + \sqrt{x_3^2+x_4^2}^2}^2 \cdots \sqrt{\cdots + \sqrt{x_{n-1}^2 + x_{n-2}^2}^2}^2}
\end{align*}
}
Figure \ref{fig-multiple-folds} shows the resulting network. The proof is by induction and is rather long so is not produced here but appears in full in the supplementary material.
At this point we make use of Lemma 19 from \citet{eldan2016power} which we reproduce here:
\begin{restatable}[Lemma 19 from \citet{eldan2016power}]{lemma}{lemnineteen}
\label{lem19}
Let $\sigma(z) = \max(0,z)$ be the ReLU activation function, and fix $L,\delta,R>0$. Let $f: \ensuremath{\mathbb{R}} \to \ensuremath{\mathbb{R}}$ which is constant outside an interval $[-R,R]$. There exist scalars $a$, $\{\alpha_i,\beta_i\}_{i=1}^w$, where $w \le 3\frac{RL}{\delta}$, such that the function:
\begin{equation}
h(x) = a + \sum_{i=1}^w \alpha_i \sigma(x-\beta_i)
\end{equation}
is L-Lipschitz and satisfies:
\begin{equation}
\myover{\sup}{x \in \ensuremath{\mathbb{R}}} | h(x) - f(x) | \le \delta.
\end{equation}
Moreover, one has $|\alpha_i| \le 2L$ and $w \le 3 \frac{RL}{\delta}$.
\end{restatable}
We can now prove the main theorem (some steps left out for brevity):
\main*
\begin{proof}
From Lemma \ref{lem-approx-vecx} we can approximate $||\myvec{x}||$ to within $\sqrt{\delta}$ and using Lemma \ref{lem19}:
\begin{equation*}
f(||\myvec{x}||)+L\sqrt{\delta}-\delta \le g(||\myvec{x}||+\sqrt{\delta}) \le f(||\myvec{x}||)+L\sqrt{\delta}+\delta
\end{equation*}
and
\begin{equation*}
f(||\myvec{x}||)-L\sqrt{\delta}-\delta \le g(||\myvec{x}||+\sqrt{\delta}) \le f(||\myvec{x}||)-L\sqrt{\delta}+\delta
\end{equation*}
therefore:
\begin{equation*}
f(||\myvec{x}||)-L\sqrt{\delta}-\delta \le g(||\myvec{x}||+\sqrt{\delta}) \le f(||\myvec{x}||)+L\sqrt{\delta}+\delta
\end{equation*}
\noindent The number of weights and neurons required by Lemma \ref{lem19} is $3 \frac{RL}{\delta}$. The number of weights and neurons required to estimate $||\myvec{x}||$ is given by Lemma \ref{lem-approx-vecx} (substituting $\sqrt{\delta}$ for $\delta$). Stack the network from Lemma \ref{lem19} (1 layer) onto the end of the network from Lemma \ref{lem-approx-vecx} ($O\left(d\log_2(d) + \log_2(d)\log_2\left(R/\sqrt{\delta}\right)\right)$ layers), thus requiring a total number of neurons no more than:
\begin{align*}
N_n \le \bigg[ &4(d-1) \\
& \log_2{\left(\frac{R\pi}{\sqrt{\delta}} \left[ \lemfive{(d-1)}{d} \right]\right)} \bigg]\\
+ &\frac{3RL}{\delta} \\
N_n &= O\left(d^2 + d\log_2\left(\frac{R}{\sqrt{\delta}}\right) + \frac{3 RL}{\delta}\right)
\end{align*}
and a total number of weights no more than:
\begin{align*}
N_w \le \bigg[ &8(d-1) \\
& \log_2{\left(\frac{R\pi}{\sqrt{\delta}} \left[ \lemfive{(d-1)}{d} \right]\right)}
\bigg] \\
+ &\frac{3RL}{\delta}\\
N_w &= O\left(d^2 + d \log_2\left(\frac{R}{\sqrt{\delta}}\right) + \frac{3 RL}{\delta}\right)
\end{align*}
\end{proof}
Again, there is an obvious weight-sharing corollary:
\begin{corol}[Deep weight sharing network]
Let $\myvec{x} \in \ensuremath{\mathbb{R}}^d$, and $\sigma(z) = \max(0,z)$. Let $f$ be an $L$-Lipschitz function supported on $[0,R]$. Fix $L,\delta,R > 0$. There exists a function $g$ expressible by a network where the number of weights is at most $N_w = O\left(d + \log_2\left(\frac{R}{\sqrt{\delta}}\right) + \frac{3 RL}{\delta}\right)$ , such that:
\begin{equation*}
\sup_{\myvec{x} \in \ensuremath{\mathbb{R}}^d} | g(\myvec{x}) - f(||\myvec{x}||)| < L\sqrt{\delta} + \delta
\end{equation*}
\end{corol}
Comparing Theorem \ref{thm:main} to Lemma \ref{lem18}, both are of order $d^2$, however the folding network version is more efficient in terms of $R$ and $\delta$. This means the deeper network can be much more efficient when either $d$ or $R$ is large, or $\delta$ is small.
\section{Deep Radial Kernel Network (DRKN)}
\label{sec-svm-approx}
We have shown constructively that a deep network using folding transformations can more efficiently represent finite extent radially symmetric functions than a corresponding 3-layer network. This construction can therefore be used to approximate any system that makes use of radial functions, and could be particularly useful for approximating radially symmetric Gaussians. There are advantages and disadvantages to doing so. The main advantage is that it allows the power and flexibility of deep neural networks to be applied in a systematic way with the architecture specified by the problem at hand. The disadvantage is that it can be more computationally costly to evaluate the radial functions via a deep network compared to directly using the function itself. However, this cost is offset by the extra flexibility afforded by the deep structure. After initialisation, the network can be further trained, allowing it to adapt more towards the data and away from the radial function approximation. In this section we demonstrate this idea on support vector machines with Gaussian kernels and show empirically that such a construction tends to perform better than the corresponding SVM but does not appear to suffer greatly from overtraining.
The first step in creating a DRKN is to train a support vector machine. The method will work for any SVM (or support vector regression) that uses a radially symmetric kernel. The most popular is the Gaussian kernel and that is what we use here.
In this case, for multi-class problems, we use one-vs-rest SVMs. The decision function of the SVM is:
\begin{equation}
f(\myvec{X}) = \argmax_{1\le c \le N_C} \sum_{i=1}^{N_{C,V}} \alpha_{c,i} K(V_{c,i}, X),
\label{eqn-svm-decision}
\end{equation}
where $N_C$ is the number of classes, $N_{C,V}$ is the number of support vectors for class $C$, $\alpha_{c,i}$ is the coefficient for support vector $V_{c,i}$, and $K(\cdot,\cdot)$ is the kernel function. Note that $\alpha_{c,i}$ can be positive or negative as it incorporates the class label of the relevant binary classification problem (1 for the class of interest, and -1 for all other classes).
An SVM with a Gaussian kernel has two parameters: $\sigma$ which specifies the width of the kernel; and $C$, the trade-off between misclassification and decision surface smoothness. These parameters need to be estimated or specified to train an SVM. We make use of the Python scikit learn package \citep{scikit-learn} and a randomised search over an exponential distribution to estimate the optimal parameters for the SVM.
There are several ways to convert Equation \ref{eqn-svm-decision} into a deep network using the techniques of Section \ref{sec-folding}, but for this paper we use the most direct method. The majority of the network relates to approximating the kernel. This kernel is weight shared across all support vectors as in the support vector machine (one alternative is to have a different kernel for each support vector). The kernel used for training the SVM is a Gaussian kernel, however Section \ref{sec-folding} requires a kernel of finite support. For the DRKN we approximate the Gaussian kernel using the polynomial method of \cite{fornefett2001radial}[$Q_{3,1}$] first, then approximate the polynomial using the method of Section \ref{sec-folding}. The network implements the following decision function:
{\small
\begin{equation}
f'(\myvec{X}) = \argmax_{1\le c \le N_C} \left[ \frac{1}{2} + \frac{1}{2} \tanh\left(\sum_{i=1}^{N_{C,V}} \alpha_{c,i} F_n(V_{c,i}-X) \right) \right],
\label{eqn-fold-decision}
\end{equation}}
where $F_n$ is the fold network approximation. We use the cross-entropy softmax loss function and optimise over all weights in the fold network, the support vector centres, and the support vector weights. A conceptual diagram of part of the network for one class is given in Figure \ref{fig-deep-svm}.\footnote{Code for approximating an SVM and training a DRKN can be downloaded from \url{https://bitbucket.org/mccane/deep-radial-kernel-network}.} The cross-entropy objective function is used with stochastic gradient descent. The number of samples in each mini-batch varies depending on the problem, and ranges from 10 to 100.
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{Figs/deep-fold-nd.pdf}
\caption{A network to approximate the norm of a vector. Each fold network consists of multiple layers. \label{fig-multiple-folds}}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{Figs/deep-svm.pdf}
\caption{Part of a Deep Radial Kernel Network showing the sub-network for a single class for a 4-dimensional problem. There are also (coincidentally) 4 support vectors. Each fold network with the same number has identical weights as does the radial function approximation network $F_n$. Circle nodes indicate single units, while rectangular nodes indicate sub-networks consisting of potentially many neurons. The support vector units effectively behave as single input linear units with a bias (the support vector component). \label{fig-deep-svm}}
\end{figure*}
\subsection{Datasets}
Several standard datasets have been used to test the algorithm and these are listed in Table \ref{tab-datasets}. Most datasets were sourced from the UCI machine learning repository, with the covtype dataset coming via the Python module sklearn. If the dataset was already split into train and test sets, then our testing made use of these sets. If not, then the training set was split 70/30 into train and test sets except for the covtype dataset. In that case, there were too many samples for effective training of an SVM, so 100000 data points were randomly sampled from the set, and these were split 70/30 into train and test sets. In all cases, the SVM and deep network were trained and tested on the same data.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Name&Source&Dims&Training \#&Test \#\\
\hline
svmguide1&UCI&4&3090&400\\
whitewine&UCI&11&3428&1470\\
redwine&UCI&8&1112&480\\
breastcancer&UCI&10&490&210\\
sat&UCI&37&4435&2000\\
sensorless drive&UCI&48&46808&11702\\
segmentation&UCI&20&1617&694\\
covtype&sklearn&54&70000&30000\\
\hline
\end{tabular}
\caption{Datasets used to test the algorithm and compare with SVM. \label{tab-datasets}}
\end{table}
\subsection{Results}
Figure \ref{fig-results} shows the results for the 8 example problems with train and test error shown as the number of epochs increases. For reference we include the SVM test error and training and test error for a radial basis function (RBF) network that was initialised with the same kernel and support vectors as the SVM - essentially replacing the fold network and the function approximate in Figure \ref{fig-deep-svm} with a Gaussian RBF neuron. For the RBF network the Gaussian parameter, the support vectors and all the network weights are trainable.
The RBF network was included to test whether moving the support vectors and/or adjusting the width of the Gaussians were the factors producing improvement.
Note that there is a training anomaly in the covtype results for the DRKN. We think this is due the optimisation algorithm taking a misstep but quickly recovering. In any case, it has little effect on the long term results.
\begin{figure*}
\centering
\includegraphics[width=0.4\textwidth]{Figs/plot_svmguide1}
\includegraphics[width=0.4\textwidth]{Figs/plot_whitewine}
\includegraphics[width=0.4\textwidth]{Figs/plot_redwine}
\includegraphics[width=0.4\textwidth]{Figs/plot_breastcancer}
\includegraphics[width=0.4\textwidth]{Figs/plot_sat}
\includegraphics[width=0.4\textwidth]{Figs/plot_sensorless_drive}
\includegraphics[width=0.4\textwidth]{Figs/plot_segment}
\includegraphics[width=0.4\textwidth]{Figs/plot_covtype}
\caption{Errors and training epoch for each of the example problems. \label{fig-results}}
\end{figure*}
\section{Discussion}
We have derived a new upper bound for deep networks approximating radially symmetric functions and have shown that deeper networks are more efficient than the 3-layer network of \citet{eldan2016power}. The central concept in this construction is a space fold --- halving the volume of space that each subsequent layer needs to handle. We hypothesise that to take full advantage of deep networks we need to apply operations that work on multiple areas of the input space simultaneously, analogous to taking advantage of the multiple linear regions noted by \citet{montufar2014number}.
Folding transformations are one way to ensure that any operation applied in a later layer is simultaneously applied to many regions in the input layer. We believe there are many other possible transformations, but the reflections used here might be considered fundamental in a sense due to the Cartan–-Dieudonn\'{e}–-Scherk Theorem which states that all orthogonal transformations can be decomposed into a sequence of reflections. We are yet to fully investigate the consequences of this theorem.
Radial basis function (RBF) networks are universal approximators and therefore we can use deep radial approximations to approximate any function by decomposing the function into a sum of RBFs first (such as mixture of Gaussians). However, the efficiency of the method will decrease as the number of RBFs required increases. It remains an open question at which point it becomes more efficient to directly approximate the function with a shallow network.
We have used our theoretical construction to
build a deep radially symmetric function approximator and employed it to approximate Gaussian kernel SVMs. Of the 8 problems tested, the method performs much better on 3, moderately better on two, and similarly on 3. In contrast, an actual RBF network approximation has performed no better than the initial SVM.
It should be pointed out that there is no real loss, other than computation, to trying the DRKN on any particular problem. It requires no parameter tuning other than the choice of opimisation method and learning parameters, and if it performs better it can be adopted, but if it does not, one can always fall back to the original SVM.
The flexibility of the DRKN comes at a cost however, as the size of the network can be quite large, and hence training can be slow. However, the method is equally applicable to more scalable approximate SVM algorithms such as the Core Vector Machine \citep{tsang2005core} or SimpleSVM \citep{vishwanathan2002ssvm}, and future work will involve testing the method on these approximate algorithms.
|
\section*{Introduction}
There are many examples of dissipative systems that can be derived from conservative ones.
A classical example is the heat equation (or more generally the so-called ``porous medium"
equation) that can be derived from the Euler equations of isentropic gases. The derivation can be
done in many different ways, for example by adding a very strong friction term or by homogenization techniques
or by properly rescaling the time variable by a small
parameter (through the so-called ``parabolic scaling"). In the present paper, we will focus on a very straightforward
idea (that does not seem to be popular, to the best of our knowledge): just perform the
quadratic change of time
$t\rightarrow \theta=t^2/2$.
In \secref{examples},
we provide several examples: we start with the very simple example of conservative forces in classical mechanics (with the
Galileo model of falling bodies as a borderline case).
Next, we briefly retrieve from the Euler equation of isentropic gases
the Darcy law and the porous
medium equation, and, in particular the heat equation from the Euler equation of isothermal gases.
Our third example, at the interface of Geometry and High Energy Physics, starts with the
conservative evolution of classical strings according to the Nambu-Goto action,
from which we get, by quadratic change of time,
the dissipative geometric model of curve-shortening in $\mathbb{R}^d$,
which is the simplest example of mean-curvature flow with co-dimension higher than $1$:
\begin{equation}
\label{shortening}
\partial_\theta X
=\frac{1}{|\partial_s X|}\partial_s(\frac{\partial_s X}{|\partial_s X|}),
\end{equation}
where $s\rightarrow X(\theta,s)$ describes a time-dependent curve in $\mathbb{R}^d$ and $|\cdot|$ denotes
the Euclidean norm.
\\
\\
In \secref{main}, we will finally discuss the system of PDEs
$$
\partial_\theta B+\nabla\cdot \left(\frac{B\otimes P-P\otimes B}{\rho}\right)=0,\;\;\;\nabla\cdot B=0,
$$
$$
P=\nabla\cdot\left(\frac{B\otimes B}{\rho}\right),
\;\;\;\rho=|B|
$$
which turns out to be nothing but the ``Eulerian version" (in $\mathbb{R}^d$) of the curve-shortening
model (\ref{shortening}). Typically, in the case of a single loop $X$ subject to the curve-shortening flow,
$B$ would just be the singular
vector-valued measure
$$
(\theta,x)\rightarrow B(\theta,x)=\int_{\mathbb{R/Z}}\delta(x-X(\theta,s))\partial_sX(\theta,s) ds\in\mathbb{R}^d,
$$
for which the system of PDE makes sense since all nonlinearities are homogeneous of degree one.
(See Appendix $1$ for more details.)
These equations admit a ``non-conservative" version
\begin{equation}\label{eq:shortnc}
\partial_\theta b+(v\cdot\nabla)b=(b\cdot\nabla)v+b v^2,
\;\;\;v=(b\cdot\nabla)b,
\end{equation}
for the reduced variables $b=B/|B|$ and $v=P/|B|$.
For the conservative system,
we define a concept of ``dissipative solutions"
related to the work of P.-L. Lions for the Euler equation of incompressible fluids \cite{Li} (see also \cite{BDS,Vo}) or to
the work of L. Ambrosio, N. Gigli, G. Savar\'e \cite{AGS} for the heat equation
and, overall, quite similar to the one recently introduced by the first author in \cite{Br-topo}. We also refer to the works of
A. Tzavaras and collaborators \cite{DeStTz,LT}, E. Feireisl and collaborators \cite{Fe,FGSW} for related concepts of ``dissipative solutions".
The main point of the present paper is to show how to get the formulation right:
we start from the Eulerian version of the string equation, for which we can
use the ``relative entropy" method, quite classical in the theory of hyperbolic systems of conservation laws to get ``weak-strong" uniqueness
results (see \cite{Da}),
then, we apply the quadratic change of
time to get a good concept
of dissipative solutions for the curve-shortening flow, namely:
\begin{definition}\label{def:dddf1}
Let us fix $T>0$ and denote $\mathbb{T}^d=\mathbb{(R/Z)}^d$.
We say that $(B,P)$ with
$$
B\in C([0,T],C(\mathbb{T}^d
,\mathbb{R}^{d})'_{w^{*}}),\;\;\;P\in C([0,T]\times \mathbb{T}^d,\mathbb{R}^{d})'
$$
is a dissipative solution of the curve-shortening flow
with initial data $B_{0}\in C(\mathbb{T}^d,\mathbb{R}^{d})'$ if and only if:
\\
\\
i)
$\;\;\;B(0)=B_0$,
$\;\;\nabla\cdot B=0$
in sense of distributions;
\\
\\
ii) $B$ and $P$ are bounded, respectively in the spaces
$C^{1/2}([0,T],(C^{1}(\mathbb{T}^d))'_{w^*})$ and
$C([0,T]\times \mathbb{T}^d,\mathbb{R}^{d})',$
by constants depending only on $T$ and $\int_{\mathbb{T}^d}|B_0|$.
\\
\\
iii) For all $\lambda>0$, $\theta\in [0,T]$,
for all smooth trial functions $(b^*,v^*,A)$ valued in $\mathbb{R}^d$,
with $\|A\|_{\infty}\le \lambda$ and ${b^*}^2=1$, for all $r\ge c^*+\frac{\lambda^2}{2}+\lambda\|v^*\|_{\infty}$,
where $c^*$ is a constant depending
explicitly on $(b^*,v^*)$, we have:
\begin{multline}\label{dissipative-1}
e^{-r\theta}
\int \eta(\theta)
+\int_0^\theta e^{-r\sigma}
\left[\int P\cdot (A-L_3)+\left(r-c^*-\frac{A\cdot(A+2v^*)}{2}\right)\eta\right.\\
\left.-B\cdot \left(L_2+b^*\frac{A\cdot(A+2v^*)}{2}\right)\right]
(\sigma)d\sigma\le \int \eta(0).
\end{multline}
where
\begin{equation}\label{dissipative-4}
\eta=|B|-B\cdot b^*,
\end{equation}
\begin{equation}\label{dissipative-2}
\mathrm{L}_2=-\partial_\theta b^*-(v^*\cdot\nabla)b^*+(b^*\cdot\nabla)v^*+b^*{v^*}^2
-b^*(b^*\cdot\nabla)(b^*\cdot v^*),
\end{equation}
\begin{equation}\label{dissipative-3}
\mathrm{L}_3=-v^*+(b^*\cdot\nabla)b^*.
\end{equation}
\end{definition}
Here $C(\mathbb{T}^3,\mathbb{R}^d)'_{w^{*}}$ is metrizable space, we can equip a metric that is consistent with the weak-* topology. The ``weak compactness" of such solutions (i.e. any sequence of dissipative solutions has accumulations
points, in a suitable weak sense, and each of them is still a dissipative solution) directly follows from:
\begin{theorem}\label{theorem1}
For fixed initial condition $B_0$, the set of dissipative solutions, if not empty, is convex and compact
for the weak-* topology of measures.
\end{theorem}
Notice that it is more challenging to prove that the set of dissipative solutions is not empty. The standard strategy is as follows:
\\
i) construct
smooth approximate solutions $(B^\epsilon,P^\epsilon)$ with smooth approximate initial data $B^\epsilon_0$;
\\
ii) show that, the approximate solutions are relatively compact for the weak-* topology of measures, and, for any trial functions $(b^*,v^*,A)$, satisfy inequalities \eqref{dissipative-1} with some small error terms;
\\
iii) let $\epsilon$ go to zero, and prove that the limit $(B,P)$ is a dissipative solution.
\\
\\
To keep this paper simple, we leave this (important) step for a future work
\cite{XD}, in the more
general framework of the Born-Infeld theory \cite{BI,Br-BI}.
Finally,
we establish a ``weak-strong" uniqueness principle in the following sense:
\begin{theorem}\label{theorem2}
Let $(b,v)\in C^{1}([0,T]\times\mathbb{T}^d;\mathbb{R}^d\times \mathbb{R}^d)$,
be a smooth solution of the non-conservative
form of the curve-shortening flow (\ref{eq:shortnc}) with $b^2=1$.
Then any dissipative solution satisfies $B=|B| b$ and $P=|B| v,$ as soon as
$B(0)=|B(0)|b(0)$.
\end{theorem}
Notice that this is not a full uniqueness result: only the homogeneous variables $b=B/|B|$ and
$v=P/|B|$ get unique and a lot of room is left for the evolution of $|B|$ itself. So the concept of
dissipative solutions seems to suffer from the same type of ambiguity as the more general concept of Brakke solutions for mean-curvature flows
\cite{Bra}.
\subsection*{Acknowledgements}
This work has been partly supported by the contract : MAGA ANR-16-CE40-0014.
\\
The first author would like to thank
the Erwin Schr\"odinger Institute (ESI) for its hospitality when this work was started.
He is also very grateful to Dmitry Vorotnikov for explaining to him, at ESI in the summer of 2016, the possibility of deriving some mean-curvature motions
as gradient flows in optimal transportation style, in the spirit of \cite{JKO,Ot,Sa,Vi}. This information was very useful to understand that
the methods used for the Born-Infeld equations by the second author \cite{XD} could also be applied to the curve-shortening flow.
\section{Examples of quadratic change of time}
\label{examples}
\subsection
{Quadratic change of time of a simple dynamical system
}
Under the quadratic change of time $t\rightarrow \theta=t^2/2$, the ordinary dynamical system
$$
\frac{d^2 X}{dt^2}=-\nabla \varphi(X)
$$
becomes
$$
-\nabla \varphi(X)=\frac{d}{dt}(\frac{dX}{d\theta}\frac{d\theta}{dt})
=\frac{d}{dt}(t\frac{dX}{d\theta})
=\frac{dX}{d\theta}+
t\frac{d\theta}{dt}\frac{d^2 X}{d\theta^2}
=\frac{dX}{d\theta}+
2\theta\frac{d^2 X}{d\theta^2}
$$
with two asymptotic regimes as $\theta$ becomes either very small or very large:
\\
the ``gradient flow"
$$
\frac{dX}{d\theta}
=-\nabla \varphi(X),
$$
and the inertial motion
$$
\frac{d^2X}{d\theta^2}=0.
$$
Notice, in the first case, that only the initial position can be chosen freely, since
$$
\frac{dX}{dt}=\frac{dX}{d\theta}\frac{d\theta}{dt}
=t\frac{dX}{d\theta}
$$
necessarily vanishes at $t=0$.
Consistently, the conservation of energy in the original time variable reads
$$
\frac{d}{d t}(\frac{1}{2}|\frac{dX}{d t}|^2+\varphi(X))=0
$$
and becomes, with the new time variable $\theta=t^2/2$
$$
\frac{d}{d\theta}[\varphi(X)]+
\theta
\frac{d}{d\theta}|\frac{dX}{d\theta}|^2
=-|\frac{dX}{d\theta}|^2,
$$
leading
to the dissipation of energy
$$
\frac{d}{d\theta}[\varphi(X)]
=-|\frac{dX}{d\theta}|^2,
$$
in the asymptotic gradient flow regime.
Furthermore, we may compare the respective solutions $X(t)$ and $Z(\theta)$
of the dynamical system and the gradient flow,
with initial conditions
$$
X(t=0)=X_0=Z(\theta=0)
,\;\;\;\frac{dX}{dt}(t=0)=0,
$$
just by
monitoring
the ``modulated energy" (or ``relative entropy")
\begin{equation}\label{relative}
\frac{1}{2}|\frac{dX}{dt}-t\frac{dZ}{d\theta}|^2
+\varphi(X)-\varphi(Z)-\nabla \varphi(Z)\cdot(X-Z),
\end{equation}
provided $\varphi$ is strongly convex with bounded third derivatives. We get, after
elementary calculations,
\begin{equation}
\label{error}
|X(t)-Z(t^2/2)|^2+|\frac{dX}{dt}(t)-t\frac{dZ}{d\theta}(t^2/2)|^2\le Ct^5 ,\;\;\;\forall t\in [0,T],
\end{equation}
where $C$ is a constant that depends only on $T$, $Z$ and potential $\varphi$.
(Notice that the smallest expected error is $O(t^6)$ as shown by the example
$d=1$, $\varphi(x)=|x|^2/2$, for which $X(t)=X(0)\cos(t)$, while $Z(\theta)=X(0)\exp(-\theta)$.)
More details on the concept of "modulated energy" and the proof of (\ref{error}) can be found in an appendix at the
end of this paper.
\subsection*{Remark: the Galileo experiment}
The quadratic change of time $t\rightarrow \theta=t^2/2$ remarkably fits with the famous
experiment by Galileo, which was the starting point of modern classical mechanics:
a rigid ball descends a rigid ramp of constant slope,
with zero initial velocity and constant acceleration $G$, reaching position
$$
X=x_0+\frac{Gt^2}{2}
$$
at time $t$.
So, $X$ is just a linear function of the rescaled time $\theta$,
$X=x_0+\theta G$ and we not only get
$$
\frac{dX}{d\theta}+
{{2\theta\frac{d^2 X}{d\theta^2}}}=G
$$
but also $simultaneously$
$$
\frac{dX}{d\theta}=G,\;\;\;
{{\frac{d^2 X}{d\theta^2}}}=0,
$$
i.e. $both$ gradient flow $and$ inertial motion, with respect to the rescaled time $\theta$.
\subsection{From the Euler equations to the heat equation and the Darcy law}
Let us now move to a PDE example and explain how the
Darcy law and the ``porous medium" equation (and, in particular, the standard heat equation) can be recovered by quadratic change of time from
the Euler equations of isentropic compressible fluids. These equations read
\begin{equation}\label{}
\partial_t\rho+\nabla\cdot(\rho v)=0,\;\;\;
\partial_t (\rho v)+\nabla\cdot(\rho v\otimes v)=-\nabla p,
\end{equation}
where ${(\rho,p,v)\in\mathbb {R}^{1+1+3}}$
are the density, pressure and velocity fields of the fluid, $p$ being a given function
of $\rho$ (such as $p=\rho$, in the ``isothermal" case).
We set
\begin{equation}\label{}
t\rightarrow \theta=t^2/2,
\;\;\;\;
\rho(t,x)\rightarrow \rho(\theta,x),
\;\;\;\; v(t,x)\rightarrow v(\theta,x)\frac{d\theta}{dt}.
\end{equation}
(Notice the different scaling for $v$, enforcing
$v(t,x)dt\rightarrow v(\theta,x)d\theta$.)
This leads, after short calculations, to
$$
\partial_\theta \rho+\nabla\cdot(\rho v)
=0,\;\;\;
\rho v+
2\theta[\partial_\theta (\rho v)+\nabla\cdot(\rho v\otimes v)]=
-\nabla p(\rho).
$$
In the regime $\theta>>1$, we get the asymptotic model of ``pressureless" gas dynamics
$$
\partial_\theta \rho+\nabla\cdot(\rho v)
=0,\;\;\;
\partial_\theta (\rho v)+\nabla\cdot(\rho v\otimes v)=0,
$$
while, as ${\theta<<1}$, we recover
the Darcy law and the porous medium equation
$$
\rho v=-\nabla p(\rho),\;\;\;
\partial_\theta \rho=\bigtriangleup (p(\rho)),
$$
and, in the isothermal case $p=\rho$, the heat equation
$$
\partial_\theta \rho=\bigtriangleup \rho.
$$
\subsection{From string motion to curve-shortening}
Let us now move to a model at the interface of geometry and high energy physics.
We consider a surface
$$
(t,s)\in\Omega \subset{\mathbb{R}}^2\rightarrow (t,X(t,s))\in
{\mathbb{R}}\times{\mathbb{R}}^d,
$$
parameterized by
a sufficiently smooth (at least Lipschitz continuous)
function $X$ over a bounded open space-time cylinder $\Omega$.
According to classical string theory (see \cite{Po}, for instance),
this surface is a relativistic string if and only if
$X$ is a critical point, with respect
to all smooth perturbations, compactly supported in $\Omega$,
of the ``Nambu-Goto Action" defined by
$$
\int_\Omega \sqrt{\partial_s X^2(1-\partial_t X^2)
+(\partial_t X\cdot\partial_s X)^2}\;\;
dtds
$$
which is nothing but the area of the surface,
in the space ${\mathbb{R}}\times{\mathbb{R}}^d$,
with respect to the Minkowski metric $(-1,+1,\cdot\cdot\cdot,+1)$.
It is customary to regularize this setting by
viewing
$$
(t,s)\in\Omega \subset{\mathbb{R}}^2
\rightarrow (t,s,X(t,s))\in
{\mathbb{R}^2}\times{\mathbb{R}}^d,
$$
as a graph in the enlarged space ${\mathbb{R}}^2\times{\mathbb{R}}^d$
and considering its area
in the enlarged Minkowski space ${\mathbb{R}}^2\times{\mathbb{R}}^d$,
with (rescaled) Minkowski metric $(-1,+\epsilon^2,+1,\cdot\cdot\cdot,+1)$:
$$
\int_\Omega \sqrt{(\epsilon^2+\partial_s X^2)(1-\partial_t X^2)
+(\partial_t X\cdot\partial_s X)^2}\;\;
dtds.
$$
[Of course, we recover the previous setting just as the special (and degenerate) case
$\epsilon=0$.]
The variational principle implies that $X$ is a solution to
the following
first order partial differential system (of hyperbolic type as $\epsilon>0$):
\begin{equation}
\label{string}
\partial_t(F\partial_t X-G\partial_s X)
-\partial_s(G\partial_t X+H\partial_s X)=0,
\end{equation}
where
$$
F=\frac{\epsilon^2+\partial_s X^2}{S},\;\;\;
G=\frac{\partial_t X\cdot\partial_s X}{S},\;\;\;
H=\frac{1-\partial_t X^2}{S},
$$
$$
S=\sqrt{(\epsilon^2+\partial_s X^2)(1-\partial_t X^2)
+(\partial_t X\cdot\partial_s X)^2}.
$$
After performing the quadratic change of time $\theta=t^2/2$, as we did in the previous
subsections, while keeping
only the zeroth order terms with respect to $\theta$, we easily obtain, as asymptotic equation
the nonlinear equation of parabolic type:
\begin{equation}
\label{curve shortening graph}
((\epsilon^2+\partial_s X^2)\mathbb{I}-\partial_s X\otimes\partial_s X)\partial_\theta X
=\sqrt{\epsilon^2+\partial_s X^2}\;\partial_s(\frac{\partial_s X}{\sqrt{\epsilon^2+\partial_s X^2}}).
\end{equation}
(Notice that $(\epsilon^2+\partial_s X^2)\mathbb{I}-\partial_s X\otimes\partial_s X$
is an invertible symmetric matrix with eigenvalues larger or equal to $\epsilon^2$.)
In the limit case $\epsilon=0$ , we get
\begin{equation}
\label{curve shortening gauge}
(\partial_s X^2\mathbb{I}-\partial_s X\otimes\partial_s X)\partial_\theta X
=|\partial_s X|\partial_s(\frac{\partial_s X}{|\partial_s X|})
\end{equation}
which becomes an ambiguous evolution equation, since it leaves
$\partial_\theta X\cdot\partial_s X$ undetermined.
[As a matter of fact, this geometric equation is not modified by any smooth
time-independent change of parameterization of the curve $s\rightarrow \sigma(s)$.]
\\
However,
we may solve instead the simpler equation
\begin{equation}
\label{curve shortening}
\partial_\theta X
=\frac{1}{|\partial_s X|}\partial_s(\frac{\partial_s X}{|\partial_s X|}).
\end{equation}
Indeed, this is a consistent way of solving (\ref{curve shortening gauge}) since
$$
\partial_\theta X\cdot \partial_s X=
\frac{\partial_s X}{|\partial_s X|}
\cdot\partial_s(\frac{\partial_s X}{|\partial_s X|})
=\partial_s(\frac{\partial_s X\cdot\partial_s X}{2|\partial_s X|^2})=0.
$$
Finally, by doing so, we have just recovered the familiar model of ``curve-shortening" in the Euclidean space $\mathbb{R}^d$ (see \cite{De}, for instance).
\subsection{The Eulerian form of the curve-shortening flow}
The string equation (\ref{string}) admits a useful ``Eulerian" version
\begin{equation}\label{eq:string-1}
\partial_{t}B + \nabla\cdot\left(\frac{B\otimes P-P\otimes B}{\rho}\right)=0,
\end{equation}
\begin{equation}\label{eq:string-2}
\nabla\cdot B=0,\;\;\;
\rho=\sqrt{B^2+P^2},
\end{equation}
\begin{equation}\label{eq:string-3}
\partial_{t}P + \nabla\cdot\left(\frac{P\otimes P}{\rho}\right)
= \nabla\cdot \left(\frac{B\otimes B}{\rho}\right)
\end{equation}
(which reads, in coordinates,
$$
\partial_t B^i+\partial_j(\rho^{-1}(B^i P^j-B^j P^i))=0,\:\:\:\partial_iB^i=0,
$$
$$
\rho=\sqrt{B_i B^i+P_i P^i},\;\;\;
\partial_tP^i+\partial_j(\rho^{-1}(P^i P^j-B^j B^i))=0).
$$
[As a matter of fact, defining
$$
B(t,x)=\int_\mathbb{R/Z}\delta(x-X(t,s))\partial_s X(t,s)ds,
$$
(which automatically satisfies $\nabla\cdot B=0$), assuming $X$ to be smooth, not self-intersecting,
with $\partial_s X$ never vanishing,
we get, after elementary calculations (similar to the ones done for the curve-shortening flow in Appendix 2, below),
that $B$ solves equations (\ref{eq:string-1},\ref{eq:string-2},\ref{eq:string-3}) together with
$$
P(t,x)
=\int_\mathbb{R/Z}\delta(x-X(t,s))\frac{
\left(\partial_s X^2\mathbb{I}-\partial_s X\otimes\partial_s X\right)\partial_t X(t,s)
}{
\sqrt{\partial_s X^2(1-\partial_t X^2)
+(\partial_t X\cdot\partial_s X)^2}}
ds.]
$$
Importantly enough, this system admits an extra conservation law:
\begin{equation}\label{eq:string-4}
\partial_t\rho+\nabla\cdot P=\nabla\cdot\left(\frac{(P\cdot B)B}{\rho^2}\right),\;\;\;\rho=\sqrt{B^2+P^2},
\end{equation}
which describes the local conservation of energy.
[This is easy to check. Indeed, using coordinates, we find
$$
\partial_t\rho=\frac{B_i\partial_t B^i+P_i\partial_t P^i}{\rho}
=\frac{B_i}{\rho}\partial_j\left(\frac{B^jP^i-B^i P^j}{\rho}\right)+\frac{P_i}{\rho}\partial_j\left(\frac{B^jB^i-P^i P^j}{\rho}\right)
$$
and notice that the second and fourth terms of the right-hand side combine as:
$$
-P^j\partial_j\left(\frac{P^2+B^2}{2\rho^2}\right)-\frac{P^2+B^2}{\rho^2}\partial_jP^j=-\partial_jP^j
$$
(since $\rho^2=B^2+P^2$), while the first and third terms give:
$$
\partial_j\left(\frac{P^i B^j B_i}{\rho^2}\right)
$$
(using $\nabla\cdot B=0$), which leads to the ``entropy conservation law" (\ref{eq:string-4}).]
\\
\\
Let us now perform the quadratic change of time:
$$
t\rightarrow \theta=\frac{t^2}{2},\;\;\;
B\rightarrow B,\;\;\;
P\rightarrow \frac{d\theta}{dt}P,\;\;\;
$$
which leads, as $\theta<<1$, to the asymptotic system
\begin{equation}\label{eq:short-1}
\partial_{\theta}B + \nabla\cdot\left(\frac{B\otimes P-P\otimes B}{\rho}\right)=0,
\end{equation}
\begin{equation}\label{eq:short-2}
\nabla\cdot B=0,\;\;\;\rho=|B|,
\end{equation}
\begin{equation}\label{eq:short-3}
P= \nabla\cdot \left(\frac{B\otimes B}{\rho}\right).
\end{equation}
Notice that this implies $B\cdot P=0$, since, in coordinates,
$$
B\cdot P=B_i\partial_j\left(\frac{B^i B^j}{\rho}\right)=\rho\frac{B_i}{\rho}B^j\partial_j\left(\frac{B^i}{\rho}\right)
=\rho B^j\partial_j\left(\frac{B^2}{2\rho^2}\right)=0
$$
(using $\nabla\cdot B=0$ and $|B|=\rho$).
We also get the extra equation, derived from (\ref{eq:string-4}),
\begin{equation}\label{eq:short-4}
\partial_\theta\rho+\frac{P^2}{\rho}+\nabla\cdot P=0,\;\;\;\rho=|B|.
\end{equation}
[Indeed, $\sqrt{B^2+2\theta P^2}=|B|+ \frac{\theta P^2}{|B|}+O(\theta^2)$,
which leads to
$$\partial_\theta \sqrt{B^2+2\theta P^2}=\partial_\theta |B|+ \frac{P^2}{|B|}+O(\theta).$$
We also used $B\cdot P=0$.]
Notice that this equation is no longer in conservation form, due to the emergence of the
dissipation term $P^2/\rho$ after the quadratic change of time $t\rightarrow \theta$.
Equations (\ref{eq:short-1},\ref{eq:short-4}) also provide a priori bounds for
any smooth solutions $B$ and $P$ on the flat torus $\mathbb{T}^d$
(that we have already taken into account in Definition \ref{def:dddf1}):
$$
\int_{\mathbb{T}^d}
|B(\theta)|\le
\int_{\mathbb{T}^d} |B(0)|,\;\forall \theta\in [0,T],
\;\;\;\int_0^T\int_{\mathbb{T}^d} \frac{P^2}{|B|}\le \int_{\mathbb{T}^d}|B(0)|,
$$
$$
\int_{0}^{T}
\int_{\mathbb{T}^d} |P|
\le \sqrt{
\int_{0}^{T}
\int_{\mathbb{T}^d} \frac{P^2}{|B|}}
\sqrt{
\;\; \int_{0}^{T}
\int_{\mathbb{T}^d} |B|
}
\le \sqrt{T}\int_{\mathbb{T}^d}|B(0)|,
$$
and, for $0\le\theta_0\le\theta_1\le T$, and any smooth vector field $\varphi=\varphi(x)$,
$$
\left(\int_{\mathbb{T}^d} \left(B^i(\theta_1)-B^i(\theta_0)\right)\varphi_i \right)^2= \left(\int_{\theta_0}^{\theta_1}
\int_{\mathbb{T}^d} (\partial_j \varphi_i-\partial_i \varphi_j) \frac{P^j B^i}{|B|}\right)^2
$$
$$
\le {\rm{Lip}}(\varphi)^2(\theta_1-\theta_0)
\int_{\theta_0}^{\theta_1}
\left(\int_{\mathbb{T}^d} |P|\right)^2
$$
(where ${\rm{Lip}}(\varphi)$ denotes the Lipschitz constant of $\varphi$),
$$
\le {\rm{Lip}}(\varphi)^2(\theta_1-\theta_0)
\int_{\theta_0}^{\theta_1}
\left(\int_{\mathbb{T}^d} \frac{P^2}{|B|}\int_{\mathbb{T}^d}|B|\right)
$$
$$
\le {\rm{Lip}}(\varphi)^2(\theta_1-\theta_0)
\left(\int_{\mathbb{T}^d} |B(0)|\right)^2,
$$
which shows that $B$ is bounded in $C^{1/2}([0,T],(C^1(\mathbb{T}^d))'_{w^*})$
by a constant depending only on $\int_{\mathbb{T}^d}|B(0)|$ and $T$.
\\
\\
Equations (\ref{eq:short-1},\ref{eq:short-2},\ref{eq:short-3})
can also be written in non-conservative form
in terms of
$$
b=\frac{B}{\rho},\;\;\;v=\frac{P}{\rho},\;\;\;\rho=|B|.
$$
We already have $b^2=1$ and $b\cdot v=0$.
Using coordinates,
we first get from (\ref{eq:short-3})
$$
\rho v^i=P^i=\partial_j\left(\frac{B^i B^j}{\rho}\right)=\partial_j(\rho b^i b^j)=\rho b^j\partial_j b^i
$$
(since $\partial_j(\rho b^j)=\partial_jB^j=0$).
Next, (\ref{eq:short-1}) becomes
$$
\rho(\partial_\theta b^i+ v^j\partial_j b^i-b^j\partial_j v^i)=-b^i(\partial_\theta\rho+\partial_j(\rho v^j))
=b^i \rho v^2
$$
(thanks to (\ref{eq:short-4})).
So we have obtained
\begin{equation}
\partial_\theta b+(v\cdot\nabla)b=(b\cdot\nabla)v+b v^2,
\;\;\;v=(b\cdot\nabla)b,
\end{equation}
(which is consistent with $b^2=1$ and $b\cdot v=0$ as can be easily checked).
Notice that (\ref{eq:short-4}) can be written according to the non-conservative variables as
\begin{equation}\label{eq:short-5}
\partial_\theta\rho+\nabla\cdot(\rho v)=-\rho v^2,
\end{equation}
which is a linear equation in $\rho$.
\section{Analysis of the Eulerian curve-shortening flow}
\label{main}
\subsection{Relative entropy for the Eulerian equations for strings}
We start from the ``Eulerian'' version
(\ref{eq:string-1},\ref{eq:string-2},\ref{eq:string-3}) of the string equation
(\ref{string}).
This system belongs to the class of systems of conservation laws:
$$
\partial_t V+\nabla\cdot \mathcal{F}(V)=0,
$$
where $\mathcal{F}$ is a given function and $V$ is a vector-valued function (for us $V=(B,P)$).
If such a system admits an $extra$ conservation law
$$
\partial_t \mathcal{E}(V)+\nabla\cdot \mathcal{G}(V)=0,
$$
for a pair of functions $\mathcal{E,G}$, with $\mathcal{E}$ strictly convex, then the system is automatically
``hyperbolic" (i.e. well posed, at least for short time), under minor additional conditions \cite{Da}, and $\mathcal{E}$
is often called an ``entropy" for the system (although it should be called ``energy" for a large class of applications).
The system for strings (\ref{eq:string-1},\ref{eq:string-2},\ref{eq:string-3}) admits such an extra conservation law, namely
(\ref{eq:string-4}), with $V=(B,P)\rightarrow \mathcal{E}(V)=\sqrt{B^2+P^2}$ as entropy function.
[Notice, however, that the entropy $\rho$ is not strictly convex.]
Hyperbolic systems of conservation laws with convex entropy enjoy
a ``weak-strong uniqueness" principle \cite{Da},
based on the concept of ``relative entropy" (also called ``modulated energy" or ``Bregman divergence", depending
of the frameworks):
$$
\eta(V,V^*)=\mathcal{E}(V)-\mathcal{E}(V^*)-\nabla\mathcal{E}(V^*)\cdot(V-V^*),
$$
which is just the discrepancy between $\mathcal{E}$ at point $V$ and its linear approximation about a given point $V^*$.
(Observe that, as $\mathcal{E}$ is a convex function with Hessian bounded away from zero and infinity, the relative entropy behaves
as $|V-V^*|^2$.) Notice that the relative entropy is as convex as the entropy as a function of $V$ ($V^*$ being kept fixed)
since it differs just by an affine term.
\\
\\
In the case of system (\ref{eq:string-1},\ref{eq:string-2},\ref{eq:string-3}), the relative entropy density is defined, for $(B,P)\in\mathbb{R}^{d}\times \mathbb{R}^{d}$
and
$(b^*,v^*)\in\mathbb{R}^{d}\times \mathbb{R}^{d}$, by
$$
\eta=\frac{{|B-\rho b^*|}^2+{|P-\rho v^*|}^2}{2\rho}=
\rho\frac{1+{b^*}^2+{v^*}^2}{2}-B\cdot b^*-P\cdot v^*,\;\;\;\rho=\sqrt{B^2+P^2},
$$
which is convex in $(B,P)$. (Notice that, at this stage, we do not assume ${b^*}^2+{v^*}^2=1$, which would be natural
to define the relative entropy but would lead to contradictions after performing the quadratic change of time as will be done
in the next subsection.)
\\
Let us now consider a smooth, $\mathbb{Z}^d-$ periodic in space, solution $(B,P)(t,x)$ of equations
(\ref{eq:string-1},\ref{eq:string-2},\ref{eq:string-3})
and monitor the evolution, on a fixed time interval $[0,T]$, of
the integral of $\eta$ over ${\mathbb{(R/Z)}^d}$,
for some smooth trial functions:
$$
(t,x)\in[0,T]\times\mathbb{(R/Z)}^d
\rightarrow (b^*(t,x)\in\mathbb{R}^d,\;v^*(t,x)\in\mathbb{R}^d).
$$
After tedious and elementary calculations, we find
\begin{equation}\label{eq:in-1}
\begin{array}{r@{}l}
\displaystyle{\quad \frac{d}{dt}\int \eta\;=} & \displaystyle{
\; \int\frac{1}{2\rho}
(B_i-\rho b^*_i)(B_j-\rho b^*_j)(\partial_jv^*_{i}+\partial_iv^*_{j})
} \\
& \displaystyle{-\int\frac{1}{2\rho}
(P_i-\rho v^*_i)(P_j-\rho v^*_j)(\partial_jv^*_{i}+\partial_iv^*_{j})} \\
& \displaystyle{- \int\frac{1}{\rho}
(B_i-\rho b^*_i)(P_j-\rho v^*_j)
(\partial_jb^*_{i}-\partial_ib^*_{j}) }\\
& \displaystyle{+ \int \eta\cdot \mathrm{L}_1+ \int B\cdot \mathrm{L}_2+ \int P\cdot \mathrm{L}_3 -\int\frac{(P\cdot B)B}{\rho^2}\cdot\nabla\left(\frac{{b^*}^2+{v^*}^2}{2}\right) }\\
\end{array}
\end{equation}
where
$$
\mathrm{L}_1=\frac{\zeta^*}{1+{b^*}^2+{v^*}^2},$$
$$
\zeta^*=D_t^*({b^*}^2+{v^*}^2)-2b^*\cdot \nabla(b^*\cdot v^*),
\;\;\;D^*_t=(\partial_t+v^*\cdot\nabla),
$$
$$
\mathrm{L}_2=-D^*_t b^*+(b^*\cdot\nabla)v^*+\nabla(b^*\cdot v^*)+b^*{\rm{L}}_1
$$
$$
\;\;\;\mathrm{L}_3=-D^*_t v^*+(b^*\cdot\nabla)b^*+v^*{\rm{L}}_1
$$
\subsection{Relative entropy and quadratic change of time}
After the quadratic change of time,
$$
t\rightarrow \theta=\frac{t^2}{2},\;\;\;
\rho\rightarrow \rho,\;\;\;
B\rightarrow B,\;\;\;b^*\rightarrow b^*,\;\;\;
P\rightarrow \theta'(t)
P,\;\;\;
v^*\rightarrow \theta'(t)
v^*,
$$
we get
$\theta'(t)^2=2\theta,\;\;\;\theta''(t)=1$,
$$
\begin{array}{r@{}l}
\displaystyle{\quad \frac{d}{d\theta}\int \eta\;= }& \displaystyle{\;\int\frac{1}{2\rho}
(B_i-\rho b^*_i)(B_j-\rho b^*_j)(\partial_jv^*_{i}+\partial_iv^*_{j})
} \\
& \displaystyle{
-\int\frac{\theta}{\rho}
(P_i-\rho v^*_i)(P_j-\rho v^*_j)(\partial_jv^*_{i}+\partial_iv^*_{j})} \\
& \displaystyle{- \int\frac{1}{\rho}
(B_i-\rho b^*_i)(P_j-\rho v^*_j)
(\partial_jb^*_{i}-\partial_ib^*_{j}) }\\
& \displaystyle{ + \int \eta\cdot \mathrm{L}_1+ \int B\cdot \mathrm{L}_2+ \int P\cdot \mathrm{L}_3 -\int\frac{(P\cdot B)B}{\rho^2}\cdot\nabla\left(\frac{{b^*}^2+2\theta{v^*}^2}{2}\right), }\\
\end{array}
$$
where
$$
\mathrm{L}_1=\left(1+{b^*}^2+2\theta{v^*}^2\right)^{-1}\zeta^*,$$
$$
\zeta^*=D_\theta({b^*}^2+2\theta{v^*}^2)-2b^*\cdot \nabla(b^*\cdot v^*)
,\;\;\;
D^*_\theta=(\partial_\theta+v^*\cdot\nabla),
$$
$$
\mathrm{L}_2=-D^*_\theta b^*+(b^*\cdot\nabla)v^*+\nabla(b^*\cdot v^*)+b^*{\rm{L}}_1
$$
$$
\;\;\;\mathrm{L}_3=-v^*-2\theta D^*_\theta v^*+(b^*\cdot\nabla)b^*
+2\theta v^*{\rm{L}}_1.
$$
Now, in order to address the Eulerian curve-shortening system,
we want to drop the terms of order $O(\theta)$ and limit ourself to the case when ${b^*}^2=1$.
However, we have to be very careful about all terms involving $\partial_\theta$.
This happens first in the definition of $\zeta^*$, because of the term
$$
D_\theta({b^*}^2+2\theta{v^*}^2)=D_\theta({b^*}^2)+2{v^*}^2+O(\theta).
$$
So, in the limit $\theta=0$, with ${b^*}^2=1$,
we find
$$
\zeta^*=2{v^*}^2-2b^*\cdot \nabla(b^*\cdot v^*)
$$
and, therefore,
$$
\mathrm{L}_1={v^*}^2-b^*\cdot \nabla(b^*\cdot v^*).
$$
Similarly, we have to take care of
$$
\frac{d}{d\theta}\int \eta.
$$
where
$$
\eta=\frac{1+{b^*}^2+2\theta{v^*}^2}{2}\rho-B\cdot b^*-2\theta P\cdot v^*
$$
and
$$
\rho=\sqrt{B^2+2\theta P^2}
=|B|+\theta\frac{P^2}{|B|}+O(\theta^2)
$$
We get
$$
\partial_\theta \rho
=\partial_\theta|B|+\frac{P^2}{|B|}+O(\theta),
$$
$$
\partial_\theta \eta
={v^*}^2\rho +\partial_\theta(\frac{1+{b^*}^2}{2}\rho)-
\partial_\theta(B\cdot b^*)-2P\cdot v^*+O(\theta)
$$
This leads, as ${b^*}^2=1$, to
$$
\rho=|B|+O(\theta),\;\;\;\eta=|B|-B\cdot b^*+O(\theta),
$$
$$
\partial_\theta \eta
={v^*}^2\rho +\partial_\theta|B|+\frac{P^2}{|B|}
-\partial_\theta(B\cdot b^*)-2P\cdot v^*+O(\theta)
$$
$$
=\partial_\theta(|B|-B\cdot b^*)
+\frac{(P-|B|v^*)^2}{|B|}
+O(\theta)
=\partial_\theta(\rho-B\cdot b^*)
+\frac{(P-\rho v^*)^2}{\rho}
+O(\theta).
$$
Finally, after dropping the terms of order $O(\theta)$ and limiting ourself to the case when ${b^*}^2=1$,
we have found
$$
\mathrm{L}_1={v^*}^2-b^*\cdot \nabla(b^*\cdot v^*),
$$
$$
\mathrm{L}_2=-D^*_\theta b^*+(b^*\cdot\nabla)v^*+\nabla(b^*\cdot v^*)+b^*{\rm{L}}_1
$$
$$
=-D^*_\theta b^*+(b^*\cdot\nabla)v^*+\nabla(b^*\cdot v^*)+b^*({v^*}^2-b^*\cdot \nabla(b^*\cdot v^*)),
$$
$$
D^*_\theta=(\partial_\theta+v^*\cdot\nabla),\;\;\;\mathrm{L}_3=-v^*+(b^*\cdot\nabla)b^*,
$$
and, for all smooth trial field $b^*$ such that ${b^*}^2=1$,
\begin{equation}\label{eq:in-1}
\begin{array}{r@{}l}
\displaystyle{\quad \frac{d}{d\theta}\int \eta
+\int\frac{(P-\rho v^*)^2}{\rho}\;=} & \displaystyle{\;
\int\frac{1}{2\rho}
(B_i-\rho b^*_i)(B_j-\rho b^*_j)(\partial_jv^*_{i}+\partial_iv^*_{j})
} \\
& \displaystyle{- \int\frac{1}{\rho}
(B_i-\rho b^*_i)(P_j-\rho v^*_j)
(\partial_jb^*_{i}-\partial_ib^*_{j}) }\\
& \displaystyle{
+ \int \eta\; \mathrm{L}_1+ \int B\cdot \mathrm{L}_2+ \int P\cdot \mathrm{L}_3 },\\
\end{array}
\end{equation}
where
$$
\rho=|B|,\;\;\;\eta=\rho-B\cdot b^*=|B|-B\cdot b^*=\frac{(B-|B|b^*)^2}{2|B|}=\frac{(B-\rho b^*)^2}{2\rho}.
$$
Using the Cauchy-Schwarz inequality, we can find a constant $c^*$ depending only on $b^*$ and $v^*$ such that
$$
\begin{array}{r@{}l}
& \displaystyle{\quad \frac{d}{d\theta}\int \eta
+\int\frac{|P-\rho v^*|^2}{2\rho}
\le c^* \int \eta\;dx
+ \int B\cdot \mathrm{L}_2+ \int P\cdot \mathrm{L}_3 }.\\
\end{array}
$$
This implies, for any constant $r\ge c^*$,
$$
(-r+\frac{d}{d\theta})\int \eta
+\int\frac{|P-\rho v^*|^2}{2\rho}+ (r-c^*) \int \eta\;dx
\le \int B\cdot \mathrm{L}_2+ \int P\cdot \mathrm{L}_3
$$
and, after multiplying this inequality by $e^{-r\theta}$ and integrating in time
$\sigma\in[0,\theta]$,
\begin{equation}\label{dissipative-0}
e^{-r\theta
}\int \eta(\theta)
+\int_0^\theta e^{-r \sigma}
\left((r-c^*)\int \eta+\int \frac{|P-\rho v^*|^2}{2\rho}-R\right)
(\sigma)d\sigma\le \int \eta(0),
\end{equation}
where
$$
R=\int B\cdot \mathrm{L}_2+ \int P\cdot \mathrm{L}_3,
$$
$$
\mathrm{L}_2=-D^*_\theta b^*+
(b^*\cdot\nabla)v^* +b^*{v^*}^2 -b^*(b^*\cdot\nabla)(b^*\cdot v^*),
\;\;\;\mathrm{L}_3=-v^*+(b^*\cdot\nabla)b^*.
$$
We can write
\begin{equation*}
\begin{array}{r@{}l}
\displaystyle{ \frac{|P-\rho v^*|^2}{2\rho}\; }& \displaystyle{=\; \sup_{A}\;(P-\rho v^*)\cdot A -\rho\frac{A^2}{2}}\\
&\displaystyle{=\; \sup_{A} P\cdot A -(\eta + B\cdot b^*)\left(v^*\cdot A+\frac{A^2}{2}\right)}
\end{array}
\end{equation*}
(since $\eta=\rho-B\cdot b^*$)
and substitute for inequality (\ref{dissipative-0})
the family of inequalities
\begin{multline}
\label{dissipative-1-bis}
e^{-r\theta}
\int \eta(\theta)
+\int_0^\theta e^{-r\sigma}
\left[\int P\cdot (A-L_3)+\left(r-c^*-\frac{A\cdot(A+2v^*)}{2}\right)\eta\right.\\
\left.-B\cdot \left(L_2+b^*\frac{A\cdot(A+2v^*)}{2}\right)\right]
(\sigma)d\sigma\le \int \eta(0).
\end{multline}
Observe that these inequalities are convex in $(B,P)$ as long as $r$ is chosen so that
$$
r\ge c^*+\sup_{\theta,x}\frac{A\cdot(A+2v^*)}{2}.
$$
However, this creates a problem, since $r$ must depend on $A$.
This is why we input
a cut-off parameter
$\lambda>0$ and assume that the trial functions $A$ are chosen with $|A(\theta,x)|\le \lambda$.
By doing this, the advantage is that we maintain the convexity of inequality as long as $r$ is chosen
big enough only as a function of $b^*,v^*$ and $\lambda$,
namely:
$$
r\ge c^*+\frac{\lambda^2}{2}+\lambda\|v^*\|_{\infty}.
$$
The price to pay is that we cannot fully recover
$$
\frac{|P-\rho v^*|^2}{2\rho}
$$
by taking the supremum over all $A$ such that $|A|\le \lambda$, but only the $\lambda-$approximation
$
K_\lambda(\rho,P-\rho v^*),
$
where
$$
K_\lambda(\rho,Z)
=\sup_{|A|\le \lambda} Z\cdot A-\rho\frac{A^2}{2}
=\frac{|Z|^2}{2\rho}-\frac{(|Z|-\lambda\rho)_+^2}{2\rho}\ge 0.
$$
Observe that, by doing so, we keep a good control of the distance between $P$ and $\rho v^*$, since
(as can be easily checked)
\begin{equation}
\label{K-inequality}
K_\lambda(\rho,P-\rho v^*)\ge \min\left(\frac{(P-\rho v^*)^2}{2\rho},\frac{\lambda|P-\rho v^*|}{2}\right).
\end{equation}
So, the supremum of inequalities (\ref{dissipative-1-bis})
over all trial functions $A$ such that $|A|\le \lambda$,
is $equivalent$ to
\begin{equation}
\label{dissipative-3-bis}
e^{-r\theta}
\int \eta(\theta)
+\int_0^\theta e^{-r\sigma}
(\int K_\lambda(\rho,P-\rho v^*)+(r-c^*)\eta
-R)
(\sigma)d\sigma\le \int \eta(0).
\end{equation}
Now let us consider $(B,P)$ not only as functions but also as vector-valued Borel measures,
for which (\ref{dissipative-1-bis}) is still well-defined. The $\lambda-$approximation $K_\lambda(\rho,P-\rho v^*)$ can be interpreted as a function of measures \cite{DeTe} and (\ref{dissipative-3-bis}) is equivalent to (\ref{dissipative-1-bis}) in the sense that,
$$\int_0^\theta e^{-r\sigma}\int K_\lambda(\rho,P-\rho v^*)=\sup_{A\in C^{0}\atop\|A\|_{\infty}\le\lambda} \int_0^\theta e^{-r\sigma}\int (P-\rho v^*)\cdot A -\rho\frac{A^2}{2}.$$
Notice that, due to the convexity of $K_\lambda$,
we have
\begin{equation}\label{K-in}
\int_0^\theta e^{-r\sigma}\int K_\lambda(\rho,P-\rho v^*)(\sigma)d\sigma
\geq e^{-r\theta} \int_0^\theta K_\lambda\left(\int \rho(\sigma),\int |P-\rho v^*|(\sigma)\right)d\sigma.
\end{equation}
With these calculations, we have recovered
the concept of dissipative solutions as given in Definition \ref{def:dddf1}.
Then,
the proof of our main results becomes
straightforward.
\subsubsection*{Proof of Theorem \ref{theorem1}}
We just have to show that, for fixed initial conditions $B_0$, the set of dissipative solutions,
as defined by Definition \ref{def:dddf1}, if not empty, is convex and compact
for the weak-* topology of measures. The convexity of the set of solutions is almost free. It follows directly from the convexity of inequalities (\ref{dissipative-1}). Let's focus on the compactness. Our goal is to prove that, if $\{(B_n,P_n)\}_{n\in\mathbb{N}}$ is a sequence of dissipative solutions with initial data $B_0$, then up to a subsequence, it converges in the weak-* topology of measures to a dissipative solution $(B,P)$ with the same initial data. This follows from the inequalities (\ref{dissipative-1}) and suitable bounds that we assume for $B_n$ and $P_n$. To see this, let's first show that, $\sup_{\theta}\int|B_n(\theta)|$ is uniformly bounded. (Indeed, let's take $b^*=(1,0,\ldots,0)$, $v^*=A=0$ in (\ref{dissipative-1}). Then we have $\int |B_n(\theta)|-B^1_{n}(\theta)\leq C\int |B_0|$, $\forall\theta\in[0,T]$. Since $B_n$ is bounded in $C^{1/2}([0,T],(C^{1}(\mathbb{T}^d))'_{w^*})$, there exists a constant $C'$ such that for any $n,\theta$, $|\int B^1_{n}(\theta)-\int B_0^1|\le C'$. So we get a uniform upper bound of $\int |B_n(\theta)|$.) Therefore, for any $\theta\in[0,T]$, the set $\{B_n(\theta)\}_{n\in\mathbb{N}}$ is relatively compact for the weak-* topology of $C(\mathbb{T}^d
,\mathbb{R}^{d})'$. Next, we look at the map $[0,T]\rightarrow C(\mathbb{T}^d
,\mathbb{R}^{d})'_{w^*}$, $t\rightarrow B(t)$. This map is equicontinuous because of the assumption on ${B_n}$. Then, by Arzel\`{a}-Ascoli's theorem, there exists $B\in C([0,T],C(\mathbb{T}^d,\mathbb{R}^{d})'_{w^{*}})$, such that, up to a subsequence, $B_n(\theta)\rightharpoonup^{*}B(\theta)$, $\forall\theta\in[0,T]$. Now, since $\iint |P_n|$ is bounded, there exists $P\in C([0,T]\times \mathbb{T}^d,\mathbb{R}^{d})'$, such that $P_n\rightharpoonup^{*}P$. Then because inequalities (\ref{dissipative-1}) are stable under weak-* convergence, we can prove that the limit $(B,P)$ satisfies all the requirements in Definition \ref{def:dddf1}, therefore, it is also a dissipative solution with initial data $B_0$.
\subsubsection*{Proof of Theorem \ref{theorem2}}
Let $(b,v)$ be a smooth solution of the non-conservative
form of the curve-shortening flow (\ref{eq:shortnc}) with $b^2=1$, which directly implies
$b\cdot v=0$.
We have to show that any dissipative solution satisfies $B=|B| b$ and $P=|B| v,$ as soon as
$B(0)=|B(0)|b(0)$.
The proof is quite straightforward:
we already have $\eta(0)=0$ since $B(0)=|B(0)|b(0)$.
Next, we set $b^*=b$, $v^*=v$, $A=0$ and fix $\lambda>0$ in definition (\ref{dissipative-1}).
Since we have (\ref{eq:shortnc}) and $b\cdot v=0$, we get ${\rm{L}}_2={\rm{L}}_3=0$.
Since $\eta\ge0$, the inequality \ref{dissipative-1} directly implies $\eta=0$, $\forall\theta\in[0,T]$,
and, therefore $B=|B|b$. Now, let's go back to the inequality \ref{dissipative-1} which is already simplified since $\eta=0$. By taking the supremum over all $A$ such that $\|A\|_{\infty}\le \lambda$, we get
$$\int_0^T e^{-r\sigma}\int K_\lambda(\rho,P-|B| v)\le 0.$$
Using (\ref{K-in}) we deduce
$$
\int_0^T K_\lambda\left(\int \rho(\sigma),\int \big|P-|B|v\big|(\sigma)\right)d\sigma=0,
$$
and, therefore, $P=|B|v$
(because of (\ref{K-inequality})), which completes the proof.
\section{
Appendix 1: modulated energy and dissipative solutions for ordinary dynamical equations
}
In this appendix, we
explain, in the very elementary case of our dynamical system, the concepts of ``modulated energy" (also called ``relative entropy") and ``dissipative formulation",
which will later be used and extended to the dissipative setting.
\\
Here, we crucially assume that the potential $\varphi$ is convex and, in order to keep
the presentation simple, we assume that the spectrum of the symmetric matrix $D^2\varphi(x)$ is
uniformly contained in some fixed interval $[r,r^{-1}]$ for some constant $r\in (0,1/2)$. We further
assume that the third derivatives of $\varphi$ are bounded.
The total energy of a curve $t\rightarrow X(t)$ is defined by
$$
\frac{1}{2}|X'(t)|^2+\varphi(X(t))\;\;\;({\rm where}\;\;X'(t)=\frac{dX}{dt}\;)
$$
and is a constant as $X$ is solution to the dynamical system
$$
X''(t)=-\nabla\varphi(X(t)).
$$
Given a smooth curve $t\rightarrow Y(t)$, we define the ``modulated energy" (or ``relative entropy") of $X$ at time $t$ with respect to $Y$
by expanding the energy about $Y$ at $X$:
$$
\eta[t,X,Y]=\frac{1}{2}|X'(t)-Y'(t)|^2+\varphi(X(t))-\varphi(Y(t))-\nabla\varphi(Y(t))\cdot(X(t)-Y(t)).
$$
Because of the assumption we made on $\varphi$, $\eta$ is a perfect substitute for the
squared distance between $(X,X')$ and $(Y,Y')$:
$$
r\le \frac{2\eta[t,X,Y]}{|X-Y|^2+|X'-Y'|^2}\le r^{-1}.
$$
We get
$$
\frac{d}{dt}\eta[t,X,Y]=(X'-Y')\cdot(
X''-Y'')
+\nabla\varphi(X)\cdot X'
-\nabla\varphi(Y)\cdot Y'
$$
$$
-\nabla\varphi(Y)\cdot(X'-Y')
-Y'\cdot D^2\varphi(Y)\cdot (X-Y)
$$
$$
=(X'-Y')\cdot(X''+\nabla\varphi(X)-Y''-\nabla\varphi(Y))+Y'\cdot(\nabla\varphi(X)-\nabla\varphi(Y)-D^2\varphi(Y)\cdot (X-Y)).
$$
We first observe that
$$
|\nabla\varphi(X)-\nabla\varphi(Y)-D^2\varphi(Y)\cdot (X-Y)|\le C|X-Y|^2\le C\eta[t,X,Y]
$$
where, from now on,
$C$ is a generic constant that depends only on $\varphi$ or $Y$.
So,
$$
\frac{d}{dt}\eta[t,X,Y]-
(X'-Y')\cdot(X''+\nabla\varphi(X)-Y''-\nabla\varphi(Y))
\le C\eta[t,X,Y]
$$
and then, after integration in time for $t\in [0,T]$, $T>0$ being an arbitrarily chosen fixed time,
\begin{equation}
\label{ineq}
\eta[T,X,Y]
-\int_0^T (X'(t)-Y'(t))\cdot (\omega_X(t)-\omega_Y(t))e^{(T-t)C}dt\le \eta[0,X,Y]e^{CT},
\end{equation}
where
$$
\omega_Z(t)=Z''(t)+\nabla\varphi(Z(t))
$$
\\
Let us exploit inequality (\ref{ineq}) in several different ways.
\\
First, we see that for a curve $X$ it is equivalent to be solution of the dynamical system, i.e. $\omega_X=0$
or to satisfy
\begin{equation}
\label{solution dissipative}
\eta[T,X,Y]
+\int_0^T (X'(t)-Y'(t))\cdot \omega_Y(t)e^{(T-t)C}dt\le \eta[0,X,Y]e^{CT},\;\;\;\forall T>0,
\end{equation}
for any smooth curve $Y$, where $C$ is a constant depending only on $Y$ (up to time $T$) and $\varphi$.
Indeed, by taking as $Y$ the unique solution of the dynamical system with initial conditions $Y(0)=X(0)$, $Y'(0)=X'(0)$
provided by the Cauchy-Lipschitz theorem on ODEs,
we get both $\omega_Y=0$ and $\eta[0,X,Y]=0$. Thus inequality (\ref{solution dissipative}) just says $\eta[T,X,Y]=0$ for
all $T>0$,
which means $X=Y$ and, therefore, $X$ is indeed a solution to the dynamical system.
Thus, we can take (\ref{solution dissipative}) as an alternative
notion of solution, that we call ``dissipative solution". This inequality has the advantage to be convex in $X$,
as the initial conditions $X(0),X'(0)$ are fixed, and therefore preserved under weak convergence of $(X,X')$.
\\
\\
Next, we use (\ref{solution dissipative}) to compare a solution $X$ of the dynamical system with zero initial velocity,
i.e. $X'(0)=0$,
to the solution $Z$ of the gradient flow equation
$$
Z'(\theta)+\nabla\varphi (Z(\theta))=0,
$$
with initial condition $Z(0)=X(0)$. Indeed, let us set $Y(t)=Z(\theta)$, $\theta=t^2/2$.
Then $Y'(t)=tZ'(\theta)$, $Y'(0)=0$, $Y(0)=Z(0)=X(0)$, $Y''(t)=Z'(\theta)+t^2 Z''(\theta)$, which implies
$ \eta[0,X,Y]=0$ and $\omega_Y(t)=t^2 Z''(\theta)$. So, (\ref{solution dissipative}) gives
$$
\eta[T,X,Y]
+\int_0^T (X'(t)-Y'(t))\cdot t^2 Z''(t^2/2)e^{(T-t)C}dt\le 0,
$$
which implies (by Cauchy-Schwarz inequality and by definition of $\eta$)
$$
\eta[T,X,Y]
\le C\int_0^T (\eta[t,X,Y]+t^4)
dt,
$$
where $C$ is a generic constant depending only on $T$, $\varphi$ and $Z$. By Gronwall's lemma,
we conclude that $\eta[t,X,Y]\le Ct^5$ which implies, by definition of $\eta$,
$$
|X(t)-Z(t^2/2)|^2+|\frac{dX}{dt}(t)-t\frac{dZ}{d\theta}(t^2/2)|^2\le Ct^5,\;\;\;\forall t\in [0,T],
$$
as already claimed, at the beginning of this subsection. (See (\ref{error}).)
(Notice that the smallest expected error is $O(t^6)$ as shown by the example
$d=1$, $\varphi(x)=|x|^2/2$, for which $X(t)=X(0)\cos(t)$, while $Z(\theta)=X(0)\exp(-\theta)$.)
\section{Appendix 2: direct recovery of the Eulerian curve-shortening flow}
For the sake of completeness, let us check that system (\ref{eq:short-1},\ref{eq:short-2},\ref{eq:short-3})
indeed describes the curve-shortening
flow in $\mathbb{R}^d$, for a continuum of non intersecting curves.
Let us do the calculation in the case of a single smooth time-dependent loop,
$s\in\mathbb{R/Z}\rightarrow X(\theta,s)$, that we assume to be
non self-intersecting at every fixed time $\theta$, and such that
$\partial_s X$ never vanishes.
We introduce (as a distribution, or, if one prefers, as a ``$1-$current")
$$
B(\theta,x)=\int_\mathbb{R/Z}\delta(x-X(\theta,s))\partial_s X(\theta,s)ds,
$$
which automatically satisfies $\nabla\cdot B=0$.
Since $X$ is smooth, not self-intersecting, and $\partial_s X$ never vanishes,
by assumption,
we may find a smooth vector field $v(\theta,x)$
such that
$$
\partial_\theta X(\theta,s)=v(\theta,X(\theta,s))
$$
that we can interpret as the ``Eulerian velocity field" attached to the loop
evolution.
We also introduce the nonnegative field
$$
\rho(\theta,x)
=\int_\mathbb{R/Z}\delta(x-X(\theta,s))|\partial_s X(\theta,s)|ds
$$
which can also be interpreted as $|B(\theta,x)|$ since $X$ is supposed to be non self-intersecting.
We get (using indices $i,j,k\in\{1,\cdot\cdot\cdot,d\}$ with implicit summation on repeated indices)
$$
\partial_\theta B^i(\theta,x)=\int_\mathbb{R/Z}[
-(\partial_j\delta)(x-X(\theta,s))\partial_\theta X^j(\theta,s)\partial_s X^i(\theta,s)
$$
$$
+\delta(x-X(\theta,s))\partial^2_{s\theta} X^i(\theta,s)]ds
$$
(in distributional sense)
$$
=-\int_\mathbb{R/Z}
(\partial_j\delta)(x-X(\theta,s))
[\partial_\theta X^j(\theta,s)\partial_s X^i(\theta,s)
-\partial_\theta X^i(\theta,s)\partial_s X^j(\theta,s)]ds
$$
(after integration by part in $s\in \mathbb{R/Z}$ of the second term)
$$
=-\partial_j \int_\mathbb{R/Z}\delta(x-X(\theta,s))
[\partial_\theta X^j(\theta,s)\partial_s X^i(\theta,s)
-\partial_\theta X^i(\theta,s)\partial_s X^j(\theta,s)]ds.
$$
So
$$
\partial_\theta B(\theta,x)=-\nabla\cdot \int_\mathbb{R/Z}\delta(x-X(\theta,x))
(\partial_s X(\theta,s)\otimes \partial_\theta X(\theta,s)
-\partial_\theta X(\theta,s)\otimes \partial_s X(\theta,s))ds.
$$
Then we can write
$$
\partial_\theta B(\theta,x)+\nabla\cdot \int_\mathbb{R/Z}\delta(x-X)
(\partial_s X(\theta,s)\otimes v(\theta,X(\theta,s))
-v(\theta,X(\theta,s))\otimes \partial_s X(\theta,s))ds=0
$$
which exactly means, by definition of $B$,
\begin{equation}
\label{induction}
\partial_\theta B+\nabla\cdot(B\otimes v-v\otimes B)=0.
\end{equation}
Since $X$ is assumed to be non-intersecting,
by definition of $v$, we may write
$$
(|B|v)(\theta,x)=\int_\mathbb{R/Z}\delta(x-X(\theta,s))|\partial_s X(\theta,s)|v(\theta,X(\theta,s))ds
$$
So far, we have not used equation (\ref{curve shortening}), namely
$$
\partial_\theta X
=\frac{1}{|\partial_s X|}\partial_s(\frac{\partial_s X}{|\partial_s X|}),
$$
Let us do it now:
$$
(\rho v^i)(\theta,x)=(|B|v^i)(\theta,x)=\int_\mathbb{R/Z}\delta(x-X(\theta,s))\partial_s(\frac{\partial_s X^i}{|\partial_s X|})ds
$$
$$
=\int_\mathbb{R/Z}(\partial_j\delta)(x-X(\theta,s))\frac{\partial_s X^j\partial_s X^i}{|\partial_s X|}ds
$$
(after integrating by part in $s\in \mathbb{R/Z}$)
$$
=\partial_j\int_\mathbb{R/Z}\delta(x-X(\theta,s))\frac{\partial_s X^j\partial_s X^i}{|\partial_s X|}ds
$$
that we can interpret as
$$
\rho v=\nabla\cdot \frac{B\otimes B}{|B|}=\nabla\cdot \frac{B\otimes B}{\rho}.
$$
Finally we can write (\ref{induction}) as
$$
\partial_\theta B+\nabla\cdot(\frac{B\otimes P-P\otimes B}{\rho})=0,\;\;\;\nabla\cdot B=0,\;\;\;P=\nabla\cdot \frac{B\otimes B}{\rho},\;\;\;\rho=|B|,
$$
where $P$ stands for $\rho v$
and $(B,\rho,P)$ solves equations (\ref{eq:short-1},\ref{eq:short-2},\ref{eq:short-3}).
So far, our claim has been justified only in the case of a single loop.
We next argue that, due to its homogeneity of degree 1, equations (\ref{eq:short-1},\ref{eq:short-2},\ref{eq:short-3}),
in spite of their nonlinearity, enjoy a nice superposition
principle, in the sense that we may still get a solution by superposing several
smooth curves subject to curve-shortening as long as they do not intersect and we may even build
smooth solutions by using a continuum of such curves.
This concludes the proof of our claim
that equations (\ref{eq:short-1},\ref{eq:short-2},\ref{eq:short-3}),
are the ``Eulerian formulation" of the curve-shortening flow.
Notice that similar calculations can also be performed to justify
the Eulerian version
(\ref{eq:string-1},\ref{eq:string-2},\ref{eq:string-3}) of the string equation
(\ref{string}).
|
\section{Introduction}
\label{sec:intro}
A Markov chain Monte Carlo (MCMC) method is a powerful numerical tool
to study a wide variety of statistical mechanical
problems\,\cite{LandauB2005,NewmanB1999}. Many kinds of non-trivial
phases and phase transitions, both in classical and quantum systems,
have been uncovered as its applications. The essence of the method is
to construct a transition kernel as a series of {\em local} kernels
acting on local state variables and to achieve eventual sampling from
an arbitrary target distribution even in a huge number of dimensions
(or degrees of freedom) of a state space.
Because next state (sample) is generated from the previous state, one
has to care correlation of samples, which can be described by an
autocorrelation function\,\cite{LandauB2005,NewmanB1999}:
\begin{equation}
A_{\hat{\mathcal O}} (t) = \frac{\langle {\mathcal O}_{i+t}{\mathcal O}_i \rangle - \langle \hat{\mathcal O} \rangle^2}{\langle \hat{\mathcal O}^2
\rangle - \langle \hat{\mathcal O} \rangle^2},
\label{A}
\end{equation}
where ${\mathcal O}_s$ is an observable, or a realization of a random
variable, at the $s$-th Monte Carlo step, the bracket $\langle \cdot
\rangle$ denotes the Monte Carlo average, and $\hat{\mathcal O}$ is an
estimator of physical quantity ${\mathcal O}$, e.g., energy or
susceptibility.
The autocorrelation function eventually becomes independent of $i$ in
Eq.~(\ref{A}) after distribution convergence (thermalization or
burn-in). In many cases, the function dumps exponentially for large
$t$:
\begin{equation}
A_{\hat{\mathcal O}}(t) \sim e^{ - t / \tau_{{\rm exp}, {\hat{\mathcal O}}}} \label{A_dump},
\end{equation}
where
\begin{equation}
\tau_{{\rm exp},{{\hat{\mathcal O}}}} = \limsup_{t \to \infty} \frac{ t }{ - \log | A_{\hat{\mathcal O}}(t)| }
\end{equation}
is the exponential autocorrelation time. Here we assume $\tau_{{\rm
exp}, {\hat{\mathcal O}}}$ is finite, which is the case for finite
size systems we study in the present paper. A thermalization period in
a Monte Carlo simulation should be, at least, several times as long as
the exponential autocorrelation time.
Meanwhile, autocorrelation reduces the effective number of Monte Carlo
samples to $M_{\rm eff} \approx M / 2 \tau_{{\rm int},{\hat{\mathcal
O}}}$, where $M$ is the number of samples in simulation and
\begin{equation}
\tau_{{\rm int}, {\hat{\mathcal O}}}= \frac{1}{2} + \sum_{t=1}^{\infty} A_{\hat{\mathcal O}} (t) \label{tau_int}
\end{equation}
is the integrated autocorrelation time. The constant $\frac{1}{2}$
comes from the discrete nature of Monte Carlo steps. Then needed
computational time is proportional to these autocorrelation times:
$\tau_{{\rm exp}, {\hat{\mathcal O}}}$ and $\tau_{{\rm int},
{\hat{\mathcal O}}}$, which may differ among estimators and update
methods.
MCMC methods can be applied to many kinds of phase transitions in
principle, but distribution-convergence rate and sampling efficiency
can become quite poor in some cases, such as critical slowing
down\,\cite{HohenbergH1977,Sokal1997}. For example, in the case of the
Ising model on the square lattice, the Metropolis algorithm for the
single spin update suffers from the rapid growth of the
autocorrelation times: $\tau_{\rm exp} \sim \tau_{\rm int} \sim \xi^z
\sim L^{z}$ with $z \approx
2.17$\,\cite{WangH1997,NightingaleB2000,MuraseI2008,LiuPS2014}, where
$L$ is the system length. Here $z$ is called the dynamic critical
exponent, which may differ between $\tau_{\rm exp}$ and $\tau_{\rm
int}$ depending on estimator. Nevertheless, the exponent is expected
to be universal among many types of updates used for MCMC
sampling\,\cite{NightingaleB2000}. This slowing down hampers accurate
or precise analysis of phase transition. It is thus crucial to devise
an efficient update method that alleviates or avoids the slowing down.
In the case of unfrustrated models, the cluster update, such as the
Swendsen-Wang\,\cite{SwendsenW1987} and the Wolff\,\cite{Wolff1989}
algorithm, reduces the dynamic critical exponents
significantly\,\cite{TamayoBK1990,CoddingtonB1992,LiuPS2014}: for
example, $z \approx 0.3$ for the Ising model in two dimensions.
The size of a cluster corresponds to the correlation length in
systems, and the flip of clusters, which can be performed with
probability one, achieves an efficient global (non-local) spin
update. Forming such an efficient cluster, however, is non-trivial or
impractical in general cases. Application of a non-local update is
thus limited to specific cases so far.
In the meantime, the worm algorithm has been one of the most versatile
techniques in the worldline quantum Monte Carlo
method\,\cite{ProkofievST1998,SyljuasenS2002}. In quantum cases, a
naive local update is often not allowed because of a certain
conservation law; for example, the total magnetization is conserved in
the $XXZ$ quantum spin model including the $XY$ and Heisenberg models
as special cases. The worm algorithm works well especially for systems
with such a conservation law. The key point of the algorithm is to
extend physical state space and allow configurations with kinks
breaking the conservation law, a pair of which is called a worm.
A whole procedure of the worm algorithm is described by the repetition
of the following processes: (i) A pair of kinks is inserted at a
randomly chosen position of a system. (ii) These kinks move on the
system in a stochastic way and update configurations. (iii) When the
kinks meet each other, the conservation law is again satisfied and an
original physical configuration is sampled. Although each worm move is
local, a non-local update is achieved in terms of the original state
space after a worm update (from insertion to removal of a worm). We
review the detail of the algorithm in Sec.~\ref{alg1}.
The worm algorithm has been proposed also for classical
systems\,\cite{ProkofievS2001}. It aims at a random walk of kinks on
sites of a lattice. Remarkably, in spite of the local nature of the
update, the worm algorithm greatly reduces the dynamic critical
exponents in the case of the Ising models and other classical
models\,\cite{DengGS2007,LiuDG2011}.
It is critical to optimize the stochastic worm update for efficient
computation. The original algorithm (hereafter called the classical
algorithm) uses the simple Metropolis algorithm; the next site is
chosen at random among nearest sites and the site shifting process is
accepted or rejected according to the distribution weights. The
detailed balance is satisfied in every worm shifting process.
Meanwhile, the efficient directed-loop algorithm was proposed in the
worldline quantum Monte Carlo method\,\cite{SyljuasenS2002}. The
directed worm does not satisfy the detailed balance for each worm
scattering process but does for one worm update from insertion to
removal. Here the worm backscattering process canceling the previous
update, which is thus a rejection process, should be averted for
efficient sampling.
In the present paper, we propose a modified worm update moving on
bonds instead of sites. The worm scattering probability is optimized
using the geometric allocation\,\cite{SuwaT2010,Suwa2014}, which is a
universal approach for probability optimization under the
non-negativity condition. One can find solutions satisfying global
balance even without detailed balance. In our modified worm update, we
minimize the worm backscattering rate and maximize the probability of
going {\em straight} on lattices for reducing the diffusive nature of
worm random walk. The reduction of the diffusive behavior is confirmed
in the probability distribution of worm position. The present
algorithm is detailed in Sec.~\ref{alg2}.
We demonstrate, in Sec.~\ref{result}, that the computational efficiency
quantified by exponential autocorrelation time and an asymptotic variance
is approximately 25 times as high as that of the classical worm for
the simple-cubic-lattice Ising model at the critical temperature. Our
algorithm is even more efficient than the Wolff algorithm. We stress
that sampling efficiency of an MCMC method should be compared in
asymptotic variance, which is discussed in Sec.~\ref{mc}. There is no
extra computational cost in the present algorithm, compared with the
classical algorithm. Our approach is applicable to generic classical
models, such as the $| \phi|^4$ model, the Potts
model\,\cite{MercadoEG2012}, the O($n$) loop
model\,\cite{JankeNS2010,LiuDG2011,ShimadaJK2014}, and the lattice
QCD\,\cite{AdamsC2003}. The present paper is summarized in
Sec.~\ref{sd}.
\section{Classical Algorithm}
\label{alg1}
We review the classical, or conventional, worm
algorithm\,\cite{ProkofievS2001} for the Ising model in this
section. Let the model be represented by $-H/T=K\sum_{\langle ij \rangle}
\sigma_i \sigma_j$, where $H$ is the Hamiltonian, $T$ is a
temperature, and $\sigma_i = \pm 1$ is the Ising spin variable at site
(vertex) $i$ of the lattice (graph). The partition function of the
canonical ensemble is expanded into the Taylor series as
\begin{align}
Z&=\sum_{\sigma_i = \pm 1} e^{K \sum_{\langle ij \rangle} \sigma_i \sigma_j}=\sum_{\sigma_i = \pm 1} \prod_{b=\langle ij \rangle} e^{K \sigma_i \sigma_j} \nonumber \\
&= \sum_{\sigma_i= \pm 1} \prod_{b=\langle ij \rangle} \cosh(K) \sum_{n_b=0,1} [\sigma_i \sigma_j \tanh K]^{n_b} \nonumber \\
&= 2^{N} \cosh(K)^{N_b^{tot}}\sum_{\{n_b\}}^{loops} [\tanh K]^{\ell} ,
\label{Z}
\end{align}
where $n_b$ denotes the bond variable on bond $b$, the identity $e^{K
\sigma_i \sigma_j}= \cosh(K) \sum_{n_b=0,1}[ \sigma_i \sigma_j \tanh
K]^{n_b}$ is used in the second line, and $N$ and $N_b^{tot}$ are
the total number of sites and bonds of a lattice, respectively. In the
last line, the sum runs over bond configurations forming loops whose
total length is denoted by $\ell \equiv \sum_{b} n_b$. The terms of
other (non-loop) configurations cancel with each others in tracing out
the spin degrees of freedom. Then the bond variables on the constraint
of loop structure are sampled by means of the MCMC method. Any set of
bond variables can be used as the initial state in a simulation as
long as the loop constraint is satisfied. We choose the state with all
bond variables deactivated ($n_b=0$ $\forall b$) as the initial one.
The worm algorithm is an efficient update method to sample under such
a constraint or a conservation law. The main idea is to extend the
sampled state space and allow configurations with kinks breaking the
constraint. Let us consider inserting two kinks, and move one of them
in a stochastic way. The moving kink is called the head of a worm and
the other is the tail of it. The classical worm algorithm is then
described as follows:
\begin{enumerate}
\item Choose a site $i_0$ at random as the starting point and set $i \leftarrow i_0$. Insert the head and tail of the worm at $i_0$. Go to step 2.
\item Choose a nearest neighbor site, $j$, at random among nearest
neighbor sites of site $i$ and shift the head of the worm from $i$
to $j$ with probability $p=[\tanh K]^{1-n_b}$, where $n_b=0,1$ is
the bond variable on $b=\langle i j \rangle$ before shifting. If the shift
is accepted, update $n_b$ as $0 \leftrightarrow 1$ and set $i
\leftarrow j$. If $j = i_0$, go to step 3. Otherwise, repeat step
2.
\item Measure observables. Go to step 1 after removing the worm with probability $p_{\rm move}$, or go to step 2 with probability $1-p_{\rm move}$.
\end{enumerate}
The probability $p_{\rm move}$ can be set an arbitrary value in
$(0,1]$: $p_{\rm move}=1/2$ in Ref.\,\citenum{ProkofievS2001}.
As for measurement, energy can be measured by the total number of
activated bonds:
\begin{align}
E &= - d \log Z / d\beta \nonumber \\
&= - N_b^{total} - N \left( \frac{1}{\tanh K} - \tanh K \right) \langle \ell \rangle, \label{e}
\end{align}
where $\beta=1/T$ is the inverse temperature. The correlation
function, $G_{ij} \equiv {\rm tr} [\sigma_i \sigma_j e^{-\beta H}] /
Z$, can be calculated by $\langle N_{ij} \rangle / \langle N_j \rangle$, where
$N_{ij}$ is the number of times when the head is at $i$ and the tail
is at $j$ in the interval between measurements (at above Step 3), and
$N_j$ is the number of times when both the head and tail are at $j$ at
Step 3. Susceptibility, $\chi \equiv \frac{\beta}{N}\sum_{ij}G_{ij}$,
can be calculated by $\beta \langle \ell_{\rm worm} \rangle$, where $\ell_{\rm
worm}$ is the worm length, or the total number of worm shifting
processes (Step 2) in the interval between measurements. It is
straightforward to calculate the Fourier transformed (or the Fourier
series of) correlation function and the correlation length by use of
the moment method\,\cite{SuwaT2015}, where one only needs to consider
a phase factor depending on worm position.
In several cases, the worm update produces a much smaller dynamic
critical exponent than a naive local spin update does: in
Ref.\,\citenum{DengGS2007}, $z \approx 0.379 $ and $0.174$ for the
Ising model on the square and the cubic lattices, respectively. The
worm algorithm is efficient not only for the Ising model but for many
fundamental physical systems: the Potts model\,\cite{MercadoEG2012},
the $| \phi |^4$ model, the O($n$) loop
model\,\cite{JankeNS2010,LiuDG2011,ShimadaJK2014}, the Lattice
QCD\,\cite{AdamsC2003}, and so on.
Meanwhile, one can use the worm algorithm for dual variables on a dual
lattice\,\cite{ProkofievS2001,HitchcockSA2004,Wang2005,RakalaD2017}. The
dual worm algorithm samples domain walls of original spin variables;
in other words, it samples ``unsatisfied'' bonds increasing
energy. While the classical worm explained above is formulated in the
high-temperature expansion, the dual worm is in the low-temperature
expansion with dual inverse temperature $\beta' = - \frac{1}{2} \ln
\tanh \beta$\,\cite{KramersW1941,Kogut1979}. One of the advantages of
the dual worm is that it is applicable also to frustrated cases while
the original worm suffers from the negative sign
problem\,\cite{Wang2005,RakalaD2017}.
Contrary to the high-temperature expansion, dual variables have the
constraint that the winding numbers of unsatisfied bonds are even in
the case of the periodic boundary condition. If the constraint is
ignored, nevertheless, the free energy difference between the periodic
and the antiperiodic boundaries can be calculated from the winding
number histogram\,\cite{HitchcockSA2004}: $e^{-\beta ( F_{\rm AP} -
F_{\rm P} )} = Z_{\rm AP} / Z_{\rm P} = \langle N_{\rm AP} \rangle / \langle
N_{\rm P} \rangle$, where $F_{\rm AP}$ and $F_{\rm P}$ are the free
energies, $Z_{\rm AP}$ and $Z_{\rm P}$ are the partition functions,
and $N_{\rm AP}$ and $N_{\rm P}$ are the numbers of times when the
winding numbers are even (odd) in one (the other) direction, and both
are even, respectively. Since the square lattice has the self-duality,
the dual worm update at the critical temperature is exactly the same
with the original update except for the winding number constraint. The
free energy difference is then available in both the formalisms.
\section{Present Approach}
\label{alg2}
We present a modified directed-worm update in this section. The worm
backscattering (rejection) probability is minimized using the
geometric allocation approach, and our algorithm is indeed free from
rejection at the critical temperature of the Ising model on the square
and the cubic lattices. As a result, the random walk behavior of the
kink is successfully suppressed, which is evident in the probability
distribution of worm position. We also explain how to measure relevant
physical quantities, such as energy and susceptibility.
\subsection{Worm on Bonds}
\label{bw}
We adopt the same representation of the partition function
[Eq.~(\ref{Z})] with the classical worm algorithm. Our goal is to
sample bond variables $\{n_b\}$ efficiently under the loop
constraint. We insert a worm, namely a pair of kinks, at a bond of a
lattice, and move the head, namely one of the pair, in a stochastic
way. When the head meets a site, it scatters onto another (or possibly
same) bond with a certain probability. This scattering process
continues until the head comes back to the tail, namely the other of
the pair.
In the case of the square lattice, a typical example of configurations
with the present worm after several scattering processes is
illustrated in Fig.~\ref{fig:example}. In our algorithm, the head and
tail breaking the loop constraint are located at the {\em center} of
bonds, or edges, of a lattice. As a result, bond variables can take
$n_b=0, \frac{1}{2}$, or $1$. The head has a moving direction in the
same sense of the directed-loop algorithm\,\cite{SyljuasenS2002}.
\begin{figure}
\begin{center}
\includegraphics[width=5cm]{fig-1.jpg}
\caption{Example of a configuration with the present worm on the
square lattice. The solid lines show activated bonds with $n_b=1$
and the broken lines show deactivated bonds with $n_b=0$. The solid
circles indicate the head ($h$) and the tail ($t$) of the worm
breaking the loop constraint of activated bonds.}
\label{fig:example}
\end{center}
\end{figure}
Let us move the head in Fig.~\ref{fig:example} upward and scatter it
around the next vertex. The head is going to scatter onto a bond
connecting to the vertex, which we define as a worm scattering
process. The four candidate states after scattering are shown as $b$,
$c$, $d$, and $e$ in Fig.~\ref{fig:ws}. The next state is chosen with
a certain (optimized) probability, which we will discuss in
Sec.~\ref{go}. After the worm scattering, the bond variables are
updated as shown in Fig.~\ref{fig:ws}; the halves of bonds are updated
($n_b=0,\frac{1}{2},1$) since the kink is assumed at the center of a
bond. We repeat this worm scattering process until the head comes back
to the tail.
The whole algorithm of the present method is described as follows:
\begin{enumerate}
\item Choose a bond $b_0$ at random as a starting point and $b
\leftarrow b_0$. Insert the head and tail of the worm at
$b_0$. Choose a moving direction at random. Go to step 2.
\item Choose the next bond $c$ according to a set of probabilities (optimized using the geometric allocation). If $b \neq c$, update bond variables $n_b$ and $n_c$, and set $b \leftarrow c$. if $b = b_0$, go to step 3. Otherwise, repeat step 2.
\item Measure observables and go to step 1 after removing the worm.
\end{enumerate}
Compared with the classical worm algorithm, we simply set $p_{\rm move}=1$ in our procedure.
As an advantage of our approach, it is straightforward to optimize the
worm scattering probability and improve efficiency. In the MCMC
method, transition probabilities are set under global balance. The
worm shifting probability at a site, in the classical algorithm, is
determined by other shifting processes at the nearest neighboring
sites. The processes at the nearest neighboring sites are then
affected by the next nearest sites. Thus, it is non-trivial to write
down the balance condition in a closed form. The Metropolis scheme
reduces the condition to a local form, but no room for optimization is
left except for increasing candidate states. In our approach, on the
other hand, the balance condition of the worm scattering process is
represented in a closed form. This simple structure of the balance
condition leaves much room for optimization we discuss in the next
subsection, compared to the previous directed-loop
approaches\,\cite{HitchcockSA2004}.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig-2.jpg}
\caption{Example of the worm scattering process on the square lattice
in the present approach. The solid circles show the worm head, and
the arrows show the direction of the head. When the head meets a
vertex (state $a$), the next state is chosen with a certain
probability between $b$, $c$, $d$, and $e$.}
\label{fig:ws}
\end{center}
\end{figure}
\subsection{Geometric Optimization}
\label{go}
We detail the transition-probability optimization in the geometric
allocation approach\,\cite{SuwaT2010,Suwa2014}. To satisfy the global
balance or the detailed balance, raw stochastic flow is allocated in a
geometric way instead of solving simultaneous algebraic equations. Let
us denote a raw stochastic flow from state $i$ to $j$ as
$v_{ij}:=\pi_i \, p_{i\to j}$, where $\pi_i$ is the weight, or the
measure, of state $i$ apart from the normalization of a target
distribution, and $p_{i \to j}$ is the transition probability from $i$
to $j$. The law of probability conservation and the global balance
condition are expressed by
\begin{equation}
\pi_i = \sum_{j=1}^n v_{ij} \qquad \forall i
\label{pc}
\end{equation}
and
\begin{equation}
\pi_j = \sum_{i=1}^n v_{ij} \qquad \forall j ,
\label{bc}
\end{equation}
respectively, where $n$ is the number of candidate states. The average
rejection (worm backscattering) rate is written as $\sum_i v_{ii} /
\sum_i \pi_i$.
We optimize the flows so that the average rejection rate is
minimized. To reduce further the worm diffusive behavior, we maximize
the average probability that the worm head goes {\em straight} on the
square or the cubic lattice under the condition of rejection
minimization. This preference at local transition is expected to
reduce the variance of the worm length, namely the variance of return
time for the head to come back to the same position.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8cm]{fig-3.jpg}
\caption{(Color online) Geometric allocation for one case of the
square lattice. In the upper panel, the solid (broken) lines show
the activated (deactivated) bonds, and the solid circles show the
worm head. In the lower panel, the weight, or the measure, of each
state is denoted by $\pi_i$ ($i=1,2,3,4$) apart from the
normalization of the target distribution, and allocated raw
stochastic flow from $i$ to $j$ is denoted by $v_{ij}$. Although the
ratio of the weights, $\pi_4 / \pi_1$, varies according to
temperature in a simulation, it is possible to allocate flows in a
similar way for $T < 2 / \ln 2$.}
\label{fig:go1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig-4.jpg}
\caption{(Color online) Geometric allocation for the other case of the
square lattice. The notations are the same with Fig.~\ref{fig:go1}.}
\label{fig:go2}
\end{center}
\end{figure}
In our approach, raw stochastic flows are optimized using the
geometric allocation. The optimized flows in the case of the square
lattice are illustrated in Figs.~\ref{fig:go1} and~\ref{fig:go2}. It
is easy to confirm that Eqs.~(\ref{pc}) and~(\ref{bc}) are both
satisfied: the area of each weight (color) is conserved, which is
nothing but the probability conservation, and the whole {\em box}
shape is intact after the allocation, which ensures the global
balance. As for the probability that the head goes straight, the value
$v_{12}+v_{21}+v_{34}+v_{43}$ is maximized. The ratio of the weights,
$\pi_4 / \pi_1=\tanh K$, varies according to temperature in
simulation. The rejection free condition is, in general, $\pi_1 \leq
\sum_{i=2}^n \pi_i$: that is, $\tanh K \geq 1/3 \iff T \leq 2 / \ln 2
\simeq 2.88539$ in the case of Fig.~\ref{fig:go1}. This condition is
always satisfied in the case of Fig.~\ref{fig:go2}. Thus, our update
is rejection free at the critical temperature, $T_{\rm
c}=2/\ln(1+\sqrt{2}) \simeq 2.269$\,\cite{KramersW1941}.
In our implementation, we calculate all transition probabilities $p_{i
\to j}= v_{ij} / \pi_i$ before sampling, and prepare a look-up table
storing probabilities. We then choose the next state at each worm
scattering by using the Walker's method of alias in $O(1)$
computational time\,\cite{FukuiT2009,HoritaST2017}. The advantage of
this method is that the computational cost does not increase with the
number of candidates $n$, while the binary search does in $O(\log
n)$. There is no extra computational cost in our algorithm, compared
with the classical worm algorithm.
We choose a set of flows satisfying detailed balance, which is
expressed by $v_{ij}=v_{ji}$ in the allocation approach. It is easy to
find many, or actually infinite, solutions to satisfy the required
conditions [Eqs.~(\ref{pc}) and~(\ref{bc})] thanks to the geometric
picture. Even solutions without detailed balance can be readily
found\,\cite{SuwaT2010}, but the performance is not different much as
far as we checked for the Ising model, compared with the present
choice of flows.
\begin{figure}
\begin{center}
\includegraphics[width=8.2cm]{fig-5.jpg}
\caption{(Color online) Geometric allocations for the simple-cubic-lattice
case in three dimensions. The six candidate states are indexed so
that (1,\,2), (3,\,4), and (5,\,6) are pairs of states before and
after worm going straight in a similar way to the square-lattice
case. It is possible to allocate flows likewise for $T <
2/\ln(3/2)$.}
\label{fig:go3}
\end{center}
\end{figure}
In the simple-cubic-lattice case, we choose a set of flows
as illustrated in Fig.~\ref{fig:go3}. The six candidate states are
indexed so that (1,\,2), (3,\,4), and (5,\,6) are pairs of states
where the worm head is on a bond in the same direction; that is, the
pairs are states before and after worm going straight in a similar way
to the square-lattice case. All possible cases of the simple cubic lattice
are shown in Fig.~\ref{fig:go3}. The rejection-free condition is
satisfied for $T \leq 2/\ln(3/2) \simeq 4.9326$ including the critical
temperature, $T_{\rm c} \approx 4.511$\,\cite{DengB2003}. In addition
to the conditions of the rejection minimization and the
going-straight-probability maximization, we here put a further
condition to find the unique solution; the variance of the going
straight flow, which is $\sum_{k=1,3,5} ( v_{k \, k+1} -
\overline{v})^2$, where $\overline{v}= \frac{1}{3} \sum_{k=1,3,5} v_{k
\, k+1}$, is minimized. Other solutions, nevertheless, are expected
to work as well as our choice as long as the rejection rate is
minimized and the going straight flow is maximized. Our worm update
optimized by the geometric allocation can be generalized to many
classical models, such as the $|\phi|^4$ model, the Potts model, and
the O($n$) loop model. It is promising to improve computational
efficiency, in general, as well as the case of the Ising model in the
present paper.
\begin{figure}
\begin{center}
\includegraphics[width=8.4cm]{fig-6.eps}
\caption{(Color online) Probability distribution of the distance
between two kinks of the worm (head and tail) for the square-lattice
model with $L=128$ in the classical (upper panel) and the present
(lower panel) approaches. The distance was measured after 256 worm
shifting or scattering processes. The contours show the coordinates
where $P(x,y)=0.5, 1.0, 1.5, 2.0$, and $2.5 \times 10^{-4}$. Because
the worm is removed when the two kinks meet each other, the
distribution is lowered around the center, which is more significant
in the case of the classical algorithm. Removed worms are not shown
here but counted in the distribution normalization.}
\label{fig:P}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=8.4cm]{fig-7.eps}
\caption{(Color online) Tails of the probability distributions of the
distance between two kinks for the square-lattice model with $L=128$
in the classical (open) and the present (solid) algorithms after 64
(triangles), 128 (squares), and 256 (circles) worm shifting or
scattering processes. The distribution was measured at $r=|{\mathbf
r}|$, where ${\mathbf r = (x,y)}$ and $|x|=|y|$. The tails are
fitted to Gaussian distributions: $ P(r) \propto e^{-r^2 / 2
\sigma^2 }$, where $\sigma^2$ is a parameter. The inset shows the
linear scaling of the estimated parameter in the case of the
classical (circles) and the present (squares) algorithms as a
function of the number of worm shifting or scattering processes
($s$). The variance in the present algorithm is approximately 6
times as large as that in the classical algorithm.}
\label{fig:P_r}
\end{center}
\end{figure}
\subsection{Reduction of Diffusive Behavior}
\label{dist}
We demonstrate here that the present worm algorithm indeed reduces the
diffusive behavior of worm position. Figure~\ref{fig:P} shows the
probability distribution of the distance between two kinks (the head
and tail of the worm) for the square-lattice case with $L=128$ after
256 worm shifting or scattering processes since the insertion. The
present worm exhibits the much broader distribution than the classical
worm does. Worms removed at $(x,y)=(0,0)$ before 256 worm processes
are not shown in Fig.~\ref{fig:P} but counted in the distribution
normalization.
The tail of the distribution of kink distance is well approximated by
a Gaussian distribution. We estimate the variance of the Gaussian
distribution after 64, 128, and 256 worm shifting or scattering
processes for $L=128$ on the square lattice, and find the linear
growth of the variance as a function of the number of processes, as
shown in Fig.~\ref{fig:P_r}: $\sigma^2 \propto s$, where $\sigma^2$ is
the estimated variance and $s$ is the number of processes. In
Fig.~\ref{fig:P_r}, although some faster decay is seen in the
distribution after 64 scattering processes of the present worm, the
tails of the distributions after 128 and 256 processes are well fitted
to the Gaussian form until longer distance. The distribution of
distance in the present algorithm is much broader: the variance in the
present approach grows approximately 6 times as fast as that in the
classical approach does, as a function of the number of worm shifting
or scattering processes. These observations clearly show that the
present method successfully reduces the diffusive nature of the worm
random walk, which is expected to make sampling more efficient.
\subsection{Estimators}
\label{obs}
The present worm can measure all physical quantities accessible by the
classical worm. Energy can be calculated by the same
formula~(\ref{e}). Meanwhile, quantities related to worm position are
measured in a different way. We explain how to measure the spin
correlation function and the susceptibility here.
In the classical worm algorithm, the sampled configuration space is
extended including configurations with two kinks on sites of a
lattice. Then the number of times when the two kinks are at sites $i$
and $j$ naturally becomes an estimator of the spin correlation between
$i$ and $j$, as mentioned in Sec.~\ref{alg1}. On the other hand, the
present worm never visits sites, so we need a different estimator for
measuring the correlation.
Let us here consider a {\it virtual} Monte Carlo step to shift two
kinks from bonds to sites. There are four choices of sites because
each bond connects two sites. Let us then choose a pair of sites at
random and consider the Metropolis algorithm for accepting or
rejecting the virtual shift. Since we assume the kink is at the center
of a bond in the present algorithm, the weight of configurations will
be changed by this kink shift according to the change in $\ell =
\sum_b n_b$. If this virtual shift would be accepted, we would count
one for measuring the corresponding spin correlation in a similar way
to the classical worm algorithm. Thus, we can use the acceptance
probability as the reweighting factor from bond-kink configurations to
site-kink configurations. To calculate the susceptibility, we simply
need to sum up the reweighting factor during the one-time worm update
from insertion to removal.
To sum up the argument of the above virtual shift, the susceptibility
estimator in the present algorithm is expressed by
\begin{equation}
\hat{\chi} = \frac{2 \beta}{4z} \sum_{\rm worm} f_{\rm rew}, \label{chi}
\end{equation}
where $\beta$ is the inverse temperature, $z$ is the coordination
number (the number of bonds connecting to a site, which is four on the
square lattice and six on the cubic lattice),
\begin{equation}
f_{\rm rew} = \left( s + \frac{1}{s} \right) f_h \label{f-rew}
\end{equation}
is the reweighting factor after a worm scattering process, and $s
\equiv \sqrt{\tanh K}$. In Eq.~(\ref{f-rew}), $f_h$ is $2/s$ if the
head is on an activated bond, $2s$ on a deactivated bond, and
$s+\frac{1}{s}$ on a half-activated and half-deactivated bond. In
other words, $f_h$ takes $2/s$ or $2s$ if the head comes back to the
tail, $s+\frac{1}{s}$ otherwise. The summation in Eq.~(\ref{chi})
means that $f_{\rm rew}$ is calculated after each worm scattering
process and summed in the interval between measurements. Note that
\begin{equation}
\hat{\chi} \sim \frac{ \beta }{ 2z } \left( s + \frac{1}{s} \right)^2 \ell_{\rm worm} ,
\label{chi-sim-ell}
\end{equation}
where $\ell_{\rm worm}$ is the worm length, because $f_h$ takes
$s+\frac{1}{s}$ except for the case where the head comes back to the
tail.
Regarding the prefactor in Eq.~(\ref{chi}), we assume the present worm
has the extra weight, $1/2$, for accepting the worm insertion and
removal with probability one, and need to consider it for estimators
related to the extended configuration. The value of the extra weight
comes from the fact that there are two possible directions for the
worm head to go in. Then, in the susceptibility estimator~(\ref{chi}),
the reweighting factor is multiplied by two, which is the inverse of
the extra weight the worm carries. In addition, it is divided by four
because we need the average of four reweighting factors, and divided
by the coordination number because of multiple counts caused by the
virtual shift from bonds to sites. It is straightforward to calculate
other quantities, such as the Fourier transformed correlation function
and the correlation length. Note that while the extra weight worm
carries is two in the case of the Ising model, it depends on model and
worm variant used in simulation.
\subsection{Possible Bias}
\label{cwa}
Before closing this section on the methodology, we note possible bias
caused by fixed-time computation of the worm algorithm as well as the
Wolff cluster algorithm\,\cite{Wolff1989}. In these methods,
computational time cost in a Monte Carlo step fluctuates depending on
worm length, or cluster size. The mean worm length, which corresponds
to susceptibility, is usually a decreasing function of
temperature. Energy, on the other hand, is an increasing
function. Thus, when a configuration is at a higher energy, the
computational time spent at the subsequent Monte Carlo step will be
shorter on average. As a result, given a simulation time, say, one
hour, high-energy configurations tend to be sampled more times than
low-energy configurations. Therefore, such a fixed-time computation
causes bias. For example, if parallel simulations are run using
different random numbers and the average among them is naively
calculated after some run-time, the estimator has a bias. To avoid it,
we have to fix the number of Monte Carlo steps instead of run-time and
calculate the average among runs of the same number of Monte Carlo
steps. The bias we discuss here is caused in quantum Monte Carlo
simulations\,\cite{ProkofievST1998,SyljuasenS2002} as well as
classical simulations. Although the systematic error might be
negligible compared to statistical error, our simulations are
carefully done avoiding the bias.
\section{How to Compare MCMC Samplers}
\label{mc}
We discuss how to quantify the computational efficiency of an MCMC
sampler. There are two points to care\,\cite{Suwa2014}: thermalization
(namely distribution convergence or burn-in) and sampling
efficiency. On the former, since Monte Carlo samples are taken after
thermalization process, faster distribution convergence to a target
one allows for sampling from an earlier Monte Carlo step; on the
latter, more efficient sampling yields a smaller statistical
error. The mean squared error of an estimator in an MCMC method is
proportional to the inverse of the number of samples (Monte Carlo
steps) according to the central limit
theorem\,\cite{RobertC2004}. Efficiency then should be compared in the
prefactor of the scaling, that is, the asymptotic
variance\,\cite{Suwa2014}. We explain here how to measure these
quantities.
Thermalization rate is quantified by exponential autocorrelation
time. An autocorrelation function exponentially decays in large Monte
Carlo steps as shown in Eq.~(\ref{A_dump}), which is the case for
finite-size systems we study in the present paper. We calculate the
function by running independent simulations and estimate the
exponential autocorrelation time as a fitting parameter. The error of
an estimate is calculated by
bootstrapping\,\cite{DavisonH1997,SenSS2015}. We simply choose a
single exponential function as the fitting function.
In the worm algorithm, we consider a Monte Carlo step in a simulation
to be a one-time worm update. In other words, the number of Monte
Carlo steps is equal to the number of times when the head comes back
to the tail. Here, one Monte Carlo step should be renormalized in
units of the number of sites. Then an autocorrelation time
($\tau'_{\rm exp}$) estimated by regression is rescaled:
\begin{align}
\tau_{\rm exp} = \tau'_{\rm exp} \frac{ \langle \ell_{\rm worm} \rangle }{N} ,
\label{tau_exp_worm}
\end{align}
where $ \langle \ell_{\rm worm} \rangle$ is the mean worm length, and $N$ is
the number of sites. The mean worm length differs between the
classical and the present algorithms since the sampled space is
extended in different manners.
Sampling efficiency is, on the other hand, related to an integrated
autocorrelation time. It can be estimated by
\begin{equation}
\tau_{\rm int}' = \frac{\sigma^2}{2 \bar{\sigma}^2} ,
\label{tau_int_est}
\end{equation}
where $\sigma^2$ is a mean squared error, namely the square of a
statistical error, calculated by the binning analysis using a much
larger bin size than the exponential autocorrelation time, and
$\bar{\sigma}^2$ is calculated without binning. The above
estimator~(\ref{tau_int_est}) will give an exact integrated
autocorrelation time in the large Monte-Carlo-step
limit\,\cite{LandauB2005}. In a similar way to
Eq.~(\ref{tau_exp_worm}), it is also rescaled as
\begin{align}
\tau_{\rm int} = \tau'_{\rm int} \frac{ \langle \ell_{\rm worm} \rangle }{N}
\label{tau_int_worm}
\end{align}
for comparison.
Although the integrated autocorrelation time is useful to study Monte
Carlo dynamics, we stress that sampling efficiency should be compared
in asymptotic variance that is the prefactor of the scaling:
\begin{equation}
\sigma_{\hat{\mathcal O}}^2 \approx \frac{ v_{{\rm asymp},{\hat{\mathcal O}}} }{M} \label{clt},
\end{equation}
where $\sigma_{\hat{\mathcal O}}^2$ is the mean squared error of an
estimator $\hat{\mathcal O}$, $v_{{\rm asymp}, {\hat{\mathcal O}}}$ is
the asymptotic variance of $\hat{\mathcal O}$, and $M$ is the number
of Monte Carlo steps used for sampling. Here we assume $\hat{\mathcal
O}$ is an unbiased estimator of a physical quantity ${\mathcal O}$:
$\langle \hat{\mathcal O} \rangle = {\mathcal O}$. Then the asymptotic
variance is represented by:
\begin{equation}
v_{{\rm asymp},{\hat{\mathcal O}}} = 2 \tau_{{\rm int}, {\hat{\mathcal O}}} v_{\hat{{\mathcal O}}} \label{v_asymp},
\end{equation}
where $v_{\hat{\mathcal O}}= \langle \hat{\mathcal O}^2 \rangle - \langle
\hat{\mathcal O} \rangle ^2$ is the variance of $\hat{\mathcal O}$. In
other words, if $M$ Monte Carlo samples are all independent with each
others, the mean squared error of $\hat{\mathcal O}$ is estimated to
be $\sigma_{\hat{\mathcal O}}^2 \approx v_{\hat{\mathcal O}} / M$. In
actual MCMC simulations, samples are correlated, so $ v_{{\rm
asymp},{\hat{\mathcal O}}} \neq v_{\hat{\mathcal O}}$; the
effective number of samples are reduced to $M_{\rm eff} \approx M / 2
\tau_{\rm int, {\hat{\mathcal O}}}$. Note $2\tau_{\rm int,
{\hat{\mathcal O}}}>1$ in most cases: $2\tau'_{\rm int,
{\hat{\mathcal O}}}>1$ in the present paper. Because different
estimators may have different variances, an integrated autocorrelation
time may not be adequate for comparison of sampling efficiency between
MCMC methods. We, therefore, use asymptotic variances to compare MCMC
samplers.
In the present paper, according to Eqs.~(\ref{tau_int_est}),
(\ref{tau_int_worm}), (\ref{clt}) and~(\ref{v_asymp}), we estimate the
variances by using the jackknife method\,\cite{Berg2004} in the
following ways:
\begin{align}
v_{{\rm asymp},\hat{\mathcal O}} &= M \frac{\sigma_{\hat{\mathcal O}}^2}{\mu_{\hat{\mathcal O}}^2} \frac{ \langle \ell_{\rm worm} \rangle }{N} \label{v_asymp_est}\\
v_{\hat{\mathcal O}} &= M \frac{\bar{\sigma}_{\hat{\mathcal O}}^2}{\mu_{\hat{\mathcal O}}^2}, \label{v_obs_est}
\end{align}
where $M$ is the number of Monte Carlo steps used for sampling,
$\mu_{\hat{\mathcal O}}$ is the average, and $\sigma_{\hat{\mathcal
O}}^2$ and $\bar{\sigma}_{\hat{\mathcal O}}^2$ are the mean
squared errors of an estimator $\hat{\mathcal O}$ with and without
binning, respectively. The squared coefficient of variation
($\sigma^2/\mu^2$) is used here for avoiding the dependence of the
definition of estimator: for example, an asymptotic variance estimated
by Eq.~(\ref{v_asymp_est}) is the same for total energy and energy
density. Note that the rescaling factor $\langle \ell_{\rm worm} \rangle / N$
in Eqs.~(\ref{tau_exp_worm}), (\ref{tau_int_worm}), and
(\ref{v_asymp_est}) is necessary for the worm and also the Wolff
algorithms, but not for the single spin update.
\section{Results}
\label{result}
We investigate the performance of our worm algorithm for the Ising
model on the simple cubic lattice, focusing on the critical
slowing down at the transition temperature.
We compare the present algorithm to the classical worm
algorithm\,\cite{ProkofievS2001} and the Wolff cluster
algorithm\,\cite{Wolff1989}.
The ensemble used in simulations is represented by the Boltzmann
distribution at the critical temperature: $1/T_{\rm c} \approx
0.22165455$\,\cite{DengB2003}. The boundary condition is periodic in
all the spatial directions. We optimize the worm scattering
probability as illustrated in Fig.~\ref{fig:go3}. More than $2^{24}$
Monte Carlo samples were taken, in total, after $2^{16}$
thermalization steps.
For fair comparison, autocorrelation times are rescaled in units of
the number of sites as shown in Eqs.~(\ref{tau_exp_worm})
and~(\ref{tau_int_worm}). In the case of the Wolff algorithm, we
rescale integrated autocorrelation time by using cluster size:
$\tau_{\rm int} = \tau'_{\rm int} \langle \ell_{\rm cl} \rangle / N$, where
$\tau'_{\rm int}$ is estimated using Eq.~(\ref{tau_int_est}), $\langle
\ell_{\rm cl} \rangle$ is the mean cluster size, and $N$ is the number of
sites. The mean worm length of the classical algorithm is proportional
to the susceptibility: $\langle \ell_{\rm classical worm} \rangle = \chi /
\beta \propto L^{\gamma/\nu}$, where $\gamma$ and $\nu$ are the
critical exponents of the susceptibility and the correlation length,
respectively\,\cite{LandauB2005}. We found the relation of the worm
length between the present and the classical algorithms: $\langle
\ell_{\rm present \ worm} \rangle \approx 1.765 \langle \ell_{\rm classical
\ worm} \rangle$ for $L \geq 16$.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig-8.eps}
\caption{(Color online) Integrated autocorrelation time (top panel),
variance (middle panel), and asymptotic variance (bottom panel) of
the energy estimator for the Ising model on the simple cubic lattice,
calculated by the Wolff (triangles), the classical worm (circles),
or the present worm algorithm (squares). The dynamic critical
exponents are estimated to be $z \approx 0.28$, $0.31$, and $0.27$
for $\tau_{{\rm int}, \hat{E}}$, $z \approx -2.75$ in common for
$v_{\hat{E}}$, and $z \approx -2.44$, $-2.35$, and $-2.46$ for
$v_{{\rm asymp}, \hat{E}}$, in the use of the Wolff, the classical
worm, and the present worm algorithms, respectively. The inset of
the bottom panel shows the asymptotic variances of the classical
worm (diamonds) and the Wolff algorithm (pentagons) in units of the
asymptotic variance of the present algorithm, which are
approximately 27 and 2.2 for relatively large system sizes,
respectively.}
\label{fig:e-3D}
\end{center}
\end{figure}
The integrated autocorrelation time, the variance, and the asymptotic
variance of the energy estimator are shown in
Fig.~\ref{fig:e-3D}. These quantities were calculated in the manner
explained above in Sec.~\ref{mc}. In the Wolff algorithm, energy is
simply calculated from spin configuration. In Fig.~\ref{fig:e-3D}, the
present algorithm gives the shortest integrated autocorrelation time
and the smallest asymptotic variance. Meanwhile, the variance of the
estimator is almost the same for the three algorithms. It is
interesting that the variance decays fast as the system size
increases, which is sort of a self-averaging effect. The shorter
correlation time of the present algorithm allowed us to calculate the
larger system size.
By fitting data to a power-law form ($\propto L^z$), we estimate the
dynamic critical exponents to be $z \approx 0.28$, $0.31$, and $0.27$
for $\tau_{{\rm int}, \hat{E}}$, $z \approx -2.75$ in common for
$v_{\hat{E}}$, and $z \approx -2.44$, $-2.35$, and $-2.46$ for
$v_{{\rm asymp}, \hat{E}}$, in the use of the Wolff, the classical
worm, and the present worm algorithms, respectively. We expect three
algorithms produce the same exponent asymptotically. Nevertheless, the
present algorithm yields approximately 27 and 2.2 times as small
asymptotic variance as the classical worm and the Wolff algorithm do,
respectively, as shown in the inset of the bottom panel of
Fig.~\ref{fig:e-3D}.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig-9.eps}
\caption{(Color online) Integrated autocorrelation time (top panel),
variance (middle panel), and asymptotic variance (bottom panel) of
the susceptibility estimator for the Ising model on the simple cubic
lattice, calculated by the Wolff (triangles), the classical
(circles), or the present worm algorithm (squares). In the Wolff
algorithm, two estimators using spins (upper triangles) or cluster
size (lower triangles) are compared (see the main text for the
detail of the estimators). The dynamic critical exponents are
estimated to be $z=0.150(9)$, $-0.50(1)$, $-0.731(7)$, and
$-0.679(4)$ for $\tau_{{\rm int}, \hat{\chi}}$, $z \approx 0.01$,
$0.58$, $0.92$, $0.85$ for $v_{\hat{\chi}}$, and $z \approx 0.18$,
$0.18$, $0.22$, and $0.18$ for $v_{{\rm asymp}, \hat{\chi}}$, in the
use of spins in the Wolff, cluster size in the Wolff, the classical
worm, and the present algorithm, respectively. The inset of the
bottom panel shows the asymptotic variances of the classical worm
(diamonds) and the Wolff algorithm (pentagons) in units of the
asymptotic variance of the present algorithm, which are
approximately 23 and 1.6 for relatively large system sizes,
respectively.}
\label{fig:sus-3D}
\end{center}
\end{figure}
The quantities of the susceptibility estimator are shown in
Fig.~\ref{fig:sus-3D} in a similar way to the energy estimator. In the
case of the Wolff algorithm, we compare two estimators: $\hat{\chi} =
\beta M_z^2 /N$, where $M_z$ is the total magnetization of spins, and
$\hat{\chi} = \beta \ell_{\rm cl} $, where $\ell_{\rm cl} $ is cluster
size. The dynamic critical exponents are estimated to be
$z=0.150(9)$, $-0.50(1)$, $-0.731(7)$, and $-0.679(4)$ for $\tau_{{\rm
int}, \hat{\chi}}$; $z \approx 0.01$, $0.58$, $0.92$, and $0.85$
for $v_{\hat{\chi}}$; and $z \approx 0.18$, $0.18$, $0.22$, and $0.18$
for $v_{{\rm asymp}, \hat{\chi}}$, in the use of spins in the Wolff,
cluster size in the Wolff, the classical worm, and the present
algorithms, respectively. The numbers in parentheses hereafter
indicate statistical uncertainty, one standard deviation, on the
preceding digit. Interestingly, while the scalings of $\tau_{{\rm
int}, \hat{\chi}}$ and $v_{\hat{\chi}}$ are quite different
between the estimators and algorithms, the dynamic critical exponent
of $v_{{\rm asymp}, \hat{\chi}}$ is likely common. Particularly, in
the case of the Wolff algorithm, $v_{{\rm asymp}, \hat{\chi}}$ is
almost the same between the estimators using spins and cluster
size. The present worm, nonetheless, yields approximately 23 and 1.6
times as small asymptotic variance as the classical worm and the Wolff
cluster do, respectively, as shown in the inset of the bottom panel of
Fig.~\ref{fig:sus-3D}.
We emphasize sampling efficiency should be compared in asymptotic
variance. The integrated autocorrelation time of the susceptibility
estimator in the classical worm algorithm is orders of magnitude
shorter than that measured by spins in the Wolff algorithm, showing
the quite different size scaling. However, the asymptotic variances of
the two estimators show almost the same scaling. In fact, the
asymptotic variance is much larger in the classical worm algorithm
because of the different scalings of the estimator variances. If you
only look at integrated autocorrelation time, it may lead to an
incorrect conclusion about efficiency.
Compared to the classical worm, the variance of the susceptibility
estimator is significantly reduced by the present worm update as shown
in the middle panel of Fig.~\ref{fig:sus-3D}. Because the worm length
is, exactly in the case of the classical worm and almost in the case
of the present worm as shown in Eq.~(\ref{chi-sim-ell}), proportional
to the susceptibility estimator, also the variance of the worm length
is reduced by the present algorithm in a similar way to the middle
panel of Fig.~\ref{fig:sus-3D}. We expect the performance improvement
to be attributed to the variance reduction of the worm length.
\begin{figure}
\begin{center}
\includegraphics[width=8.4cm]{fig-10.eps}
\caption{(Color online) Autocorrelation functions of the energy (left
panel) and the susceptibility (right panel) estimators for $L=4$
(open) and $8$ (solid) on the simple cubic lattice in the use of the
classical worm (circles) and the present algorithm (squares). The
horizontal axis is the renormalized number of Monte Carlo steps in
units of $L^3$ worm shifting or scattering processes.}
\label{fig:A}
\end{center}
\end{figure}
We compare not only sampling efficiency but also thermalization
rate. The autocorrelation functions~(\ref{A}) of the energy and the
susceptibility estimators for $L=4$ and $8$ are shown in
Fig.~\ref{fig:A}, calculated from more than $2^{30}$ independent
Markov chains (sample paths). The function of the energy estimator
shows the almost single exponential decay.
As for the susceptibility estimator, on the other hand, some fast and
slow decays are observed. While $\tau_{{\rm int}, \hat{\chi}}$
decreases with $L$ as shown in the top panel of Fig.~\ref{fig:sus-3D},
the autocorrelation function for $L=8$ clearly has slower decay than
that for $L=4$ does, in the case of the classical algorithm, as shown
in the right panel of Fig.~\ref{fig:A}. The reason why $\tau_{{\rm
int}, \hat{\chi}}$ decreases with $L$ is that the prefactor of the
slowly decaying mode decreases, which is also seen in
Fig.~\ref{fig:A}.
The scaling of the exponential autocorrelation time is shown in
Fig.~\ref{fig:tau_exp-3D}, calculated by bootstrapping as mentioned in
Sec.~\ref{mc}. We found $\tau_{{\rm exp}, \hat{\chi}} \approx
\tau_{{\rm exp}, \hat{E}}$, which is also seen in Fig.~\ref{fig:A}.
Since the autocorrelation function of the energy estimator is well
approximated by a single exponential function, it is expected that
$\tau_{{\rm exp}, \hat{E}} \approx \tau_{{\rm int}, \hat{E}}$, which
we indeed confirmed. Therefore, we expect the asymptotic scaling:
$\tau_{{\rm exp}, \hat{\chi}} \approx \tau_{{\rm exp}, \hat{E}}
\approx \tau_{{\rm int}, \hat{E}} \propto L^{0.27}$, where the power
was estimated from the plots in Fig.~\ref{fig:e-3D}.
The ratio of the exponential autocorrelation times between the
classical and the present algorithms is approximately 26 as shown in
the inset of Fig.~\ref{fig:tau_exp-3D}, which is consistent with the
ratio regarding the integrated autocorrelation time or the asymptotic
variance of the energy estimator. Note that the summation of the
autocorrelation function with respect to the renormalized time is
somewhat different from the renormalized integrated autocorrelation
time~(\ref{tau_int_worm}) because of the constant $1/2$ in the
definition~(\ref{tau_int}). Nonetheless, the asymptotic scaling with
respect to system size should be the same for the two quantities.
A lesson to learn from these analyses is that we must be careful for
estimating $\tau_{\rm exp}$ and needed thermalization (burn-in)
period. Because $\tau_{\rm int}$ is usually easier to estimate than
$\tau_{\rm exp}$, people roughly estimate $\tau_{\rm exp}$ from
$\tau_{\rm int}$ in some (or maybe many) cases. Assume that an
autocorrelation function is approximated by a single exponential
function and $\tau_{\rm exp} \gg 1$, then $\tau_{\rm int} \sim
\tau_{\rm exp}$. When an autocorrelation function has a slow decay
with prefactor $c$, $\tau_{\rm int} \sim c \, \tau_{\rm exp}$. Thus
$\tau_{\rm exp}$ can be much larger than $\tau_{\rm int}$ possibly in
orders of magnitude as we have estimated $\tau_{{\rm int}, \hat{\chi}}
\propto L^{-0.731}$, but $\tau_{{\rm exp}, \hat{\chi}} \propto
L^{0.27}$ for the classical algorithm.
\begin{figure}
\begin{center}
\includegraphics[width=8.4cm]{fig-11.eps}
\caption{(Color online) Size dependence of the exponential
autocorrelation time of the energy (solid) and the susceptibility
(open) estimators in the use of the classical (circles) and the
present (squares) algorithms for the Ising model on the simple cubic
lattice. The inset shows the ratio of the times of the energy
estimators between the two algorithms.}
\label{fig:tau_exp-3D}
\end{center}
\end{figure}
\section{Summary}
\label{sd}
We have proposed the modified worm algorithm for classical models and
demonstrated the performance improvement for the Ising model at the
critical temperature. The kinks of the present worm are located on
bonds instead of sites on a lattice. The probabilities at worm
scattering are optimized using the directed-loop framework and the
geometric allocation approach so that the bounce (rejection)
probability is minimized and indeed reduced to zero in a wide range of
temperature including the critical point. Furthermore, the probability
for the worm head to go straight is maximized to reduce the diffusive
nature of worm position.
The reduction of the diffusive behavior is confirmed in the
probability distribution of worm position as the rapid growth of the
variance of the distribution as a function of scattering processes in
Figs.~\ref{fig:P} and~\ref{fig:P_r}. As a result, the variance of worm
length, as well as the variance of the susceptibility estimator, is
significantly reduced as seen in the middle panel of
Fig.~\ref{fig:sus-3D}, which is expected to lead to efficient
sampling.
We have discussed how to compare Monte Carlo samplers and measure
relevant quantities. While thermalization rate is quantified by
exponential autocorrelation time, sampling efficiency, as we have
stressed, should be compared in asymptotic variance.
The exponential autocorrelation time and the asymptotic variance in
the present method are approximately only 4\% as large as those in the
classical algorithm for the Ising model on the simple cubic lattice, shown in
Figs.~\ref{fig:e-3D},~\ref{fig:sus-3D}, and~\ref{fig:tau_exp-3D}. Our
algorithm is remarkably even more efficient than the Wolff cluster
algorithm. The compared algorithms likely produce the same dynamic
critical exponents of the exponential autocorrelation time and the
asymptotic variance. The exponent of the longest time scale in the
worm update is estimated to be $ z \approx 0.27$, which is somewhat
larger than the previous estimate in the classical worm algorithm, $z
\approx 0.174$\,\cite{DengGS2007}, but consistent with the recent
estimate in the Wolff update, $z=0.24(2)$\,\cite{LiuPS2014}. This
agreement suggests that the worm and the Wolff algorithms share the
same exponent of the exponential autocorrelation time as well as the
asymptotic variance.
Meanwhile, we observed the integrated autocorrelation time of the
susceptibility estimator shows the quite different scaling from the
exponential autocorrelation time: $\tau_{{\rm exp}, \hat{E}} \approx
\tau_{{\rm int}, \hat{E}} \approx \tau_{{\rm exp}, \hat{\chi}} \gg
\tau_{{\rm int}, \hat{\chi}}$. These findings clearly indicate that
asymptotic variance is more adequate than integrated autocorrelation
time for comparison of sampling efficiency.
The present approach can be generalized to a wide range of physical
models, such as the $| \phi|^4$ model, the Potts
model\,\cite{MercadoEG2012}, the O($n$) loop
model\,\cite{JankeNS2010,LiuDG2011,ShimadaJK2014}, and the lattice
QCD\,\cite{AdamsC2003}. The directed-loop framework and the geometric
allocation approach are expected to reduce computational cost
significantly for many kinds of systems. In particular, it is
promising to apply to frustrated models in combination with the dual
worm formalism\,\cite{RakalaD2017}. The performance of the present
worm algorithm in the case of other models needs to be investigated in
the future.
\acknowledgements {The author is grateful to Synge Todo for discussion
on measurement and the distribution of worm position. Some
simulations have been done using computational resources of the
Supercomputer Center at the Institute for Solid State Physics, the
University of Tokyo. The author acknowledges support by KAKENHI
under Grant No.\,16K17762 from JSPS.}
|
\section{Introduction}
\label{intro}
Relativistic wave equations namely Dirac equation and KG equation have great importance in the efforts to determine dynamics of a relativistic particle in relativistic quantum mechanics.
Solution of the KG equation explains the behavior of a spinless particle of rest mass $m$ at high energies and velocities comparable to the speed of light. Bound state solutions of the KG equation have been studied by many authors in the literature, see \cite{yasuk,egrifes,diao,akbarieh,motavalli,antia} and references therein. Different methods can be used to obtain exact or approximate solutions of KG equations which are written for various potential functions. The Nikiforov-Uvarov (NU) method, Supersymmetric Quantum Mechanics, Factorization Method and Asymptotic Iteration Method are the most frequently used methods \cite{ikot}.
Using the theory of fractional calculus which is based on non integer order differentiations and integrations, a lot of physical phenomena can be described successfully \cite{ganji,tenreiro,sier,podlubny}. Consequently differential equations which describe physical systems are handled in the fractional domain. Various definitions have been proposed for the fractional order differential and integral operators.
Predominant definitions are Riemann-Liouville and Caputo definitions \cite{miller}.
\begin{eqnarray}
\label{rl}
^{RL}D_{a}^{\mu}f(x)=\frac{1}{\Gamma(n-\mu)}\frac{d^{n}}{dx^{n}}\int_{a}^{x}(x-t)^{n-\mu -1}f(t)dt, \nonumber \\
\label{caputo}
^{C}D_{a}^{\mu}f(x)=\frac{1}{\Gamma(n-\mu)}\int_{a}^{x}(x-t)^{n-\mu -1}f^{(n)}(t)dt.\nonumber
\end{eqnarray}
where $\mu \in R$, $n-1 \leq \mu < n$ and the superscripts $RL$ and $C$ stand for Riemann-Liouville and Caputo respectively. These are non local operators and do not satisfy classical properties such as chain, product and quotient rules which allow us to achieve analytical solution in the standard calculus. In 2014 a local form of fractional derivative operator is defined by Khalil et al. \cite{khalil}. This is the most natural fractional order derivative operator which provides the above mentioned rules. Thereafter conformable fractional calculus theory \cite{abdaljawad}, applicability of this definition in quantum mechanics \cite{anderson}, solution of fractional Schr\"odinger equation \cite{eslami} are studied in the view of this local fractional derivative definition. Though the definition fails some properties which are pointed out by Ortigueira and Machado \cite{ort}, it is more suitable for applications as compared with Riemann-Liouville or Caputo fractional derivative operators for real physical problems \cite{iyola}.
In a recent work, we have derived conformable fractional form of the NU method and solved local fractional Schr\"odinger equation for harmonic oscillator potential, Hulthen potential and Woods-Saxon potential in order to present accuracy of the method \cite{karayer}.
The aim of the present work is to solve fractional order one dimensional time independent KG equation for the generalized Hulthen potential in the scalar coupling scheme using the conformable fractional NU method. The manuscript is organized as follows; In Sec.~\ref{sec:1} the formalism of KG equation for the generalized Hulthen potential is briefly outlined. In Sec.~\ref{sec:2} the definition of conformable fractional derivative operator and conformable fractional NU method are reviewed. In Sec.~\ref{sec:3} we have presented analytical solution of fractional KG equation for the generalized Hulthen potential. Finally, conclusions are discussed in the last section.
\section{Formalism of the KG equation with the generalized Hulthen potential in scalar scheme}
\label{sec:1}
One dimensional time independent KG equation for a spinless particle of rest mass $m$ in the presence of vector and scalar potentials is given by;
\begin{equation}
\psi''(x)+\frac{1}{\hbar^2 c^2}\bigg[\big(E-V(x)\big)^2 -\big(mc^2+S(x)\big)^2\bigg]\psi(x)=0,
\end{equation}
where $V(x)$ and $S(x)$ are vector and scalar potentials respectively. For the existence of bound state solutions it is required that $S(x)>V(x)$ \cite{egrifes}. When $V(x)=0$, the one dimensional KG equation for a given scalar potential $S(x)$ is reduced to the following form;
\begin{equation}\label{skg}
\psi''(x)+\frac{1}{\hbar^2 c^2}\bigg[E^2 -\big(mc^2+S(x)\big)^2\bigg]\psi(x)=0,
\end{equation}
In this case Eq.(\ref{skg}) can be transformed to a second order Schr\"odinger-like differential equation;
\begin{equation}\label{slkg}
\psi''(x)+\frac{2m}{\hbar^2 }\bigg[E_{eff} -U_{eff}(x) \bigg]\psi(x)=0,
\end{equation}
where $E_{eff}$ and $U_{eff}$ are effective energy and effective potential given by;
\begin{equation}
E_{eff}=\frac{E^2-m^2c^4}{2mc^2},\,\,\,\,\,\,U_{eff}(x)=\frac{S^2(x)}{2mc^2}+S(x).
\end{equation}
Therefore Eq.(\ref{skg}) can be rewritten in the following form for $\hbar=c=1$ \cite{egrifes};
\begin{equation}\label{slkg1}
\psi''(x)+\big[-S^2(x)-2mS(x)-(m^2-E^2)\big]\psi(x)=0.
\end{equation}
In order to specify dynamics of a relativistic particle in a scalar potential, the potential function $S(x)$ is inserted in this equation. Here the potential function is chosen as generalized Hulthen function which is given by;
\begin{equation}\label{ghp}
S(x)=-S_{0}\frac{e^{-\alpha x}}{1-qe^{-\alpha x}}
\end{equation}
where $q$ is a deformation parameter. This potential transforms to exponential potential, standard Hulthen potential and Woods-Saxon potential for $q=0$, $q=1$ and $q=-1$ respectively. Substituting the potential function given by Eq.(\ref{ghp}) in Eq.(\ref{slkg1}) and using a transformation $z=S_{0}e^{-\alpha x}$ the following hypergeometric type differential equation is obtained \cite{egrifes};
\begin{eqnarray}\label{cfkghp}
\psi_{q}''(z)+\frac{S_{0}-qz}{z(S_{0}-qz)}\psi_{q}'(z)+
\frac{1}{\big[z(S_{0}-qz)\big]^{2}}\big[-(\gamma^2+\nonumber \\
q\beta^2+q^2\epsilon^2)z^2+S_{0}(\beta^2+2q\epsilon^2)z-S_{0}^2 \epsilon^2\big]\psi_{q}(z)=0
\end{eqnarray}
for which
\begin{equation}\label{qntts}
\gamma^{2}=\frac{S_{0}^{2}}{\alpha^{2}}, \,\,\,\,\,\,\beta^{2}=\frac{2mS_{0}}{\alpha^{2}},\,\,\,\,\, \epsilon^{2}=\frac{1}{\alpha^{2}}(m^{2}-E^{2}).
\end{equation}
\section{A brief review of the conformable fractional derivative operator and conformable fractional NU method}
Local fractional derivative operator which is a natural extension of the standard derivative definition, is introduced by Khalil et al. for the first time;
\label{sec:2}
\begin{eqnarray}\label{eq1}
D^{\mu}[f(t)]&=& \lim\limits_{\epsilon \rightarrow 0}\frac{f(t+\epsilon t^{1-\mu})-f(t)}{\epsilon}, \,\,\,\,\,t>0\\
\label{rightlimit}
f^{(\mu)}(0)&=& \lim\limits_{t \rightarrow 0^{+}}f^{(\mu)}(t)
\end{eqnarray}
where $0<\mu \leq 1$ and $D^{\mu}$ is the local fractional derivative operator \cite{khalil}. This operator provides the basic rules such as product, quotient, chain rules which are valid in standard calculus;
\begin{eqnarray}
D^{\mu}[af+bg]=aD^{\mu}[f]+bD^{\mu}[g] & & linearity \nonumber \\
D^{\mu}[fg]=fD^{\mu}[g]+gD^{\mu}[f] & & product \,\, rule \nonumber \\
D^{\mu}[f(g)]=\frac{df}{dg}D^{\mu}[g] & & chain \,\, rule \nonumber \\
D^{\mu}[f]=t^{1-\mu}f' & & {where} \,\,\,\, f'=\frac{df}{dt}. \nonumber
\end{eqnarray}
The last property is called as key property of the definition. If $f$ is differentiable, then $\mu th$ order derivative of $f$ is equal to product of its first order derivative with $t^{1-\mu}$.
Since fractional order differential equations have great importance in describing physical systems with a realistic approach, some appropriate methods are derived to solve these equations. The NU method is a well known method which gives exact solutions of second order linear differential equations. In quantum mechanics the method has been used to solve Schr\"odinger-like differential equations for various potentials. The method is based on reducing the handled equation to a hypergeometric type second order differential equation;
\begin{equation}
\label{eq:nu}
\psi''(z)+\frac{\widetilde{\tau}(z)}{\sigma(z)}\psi'(z)+\frac{\widetilde{\sigma}(z)}{\sigma^{2}(z)}\psi(z)=0
\end{equation}
where $\widetilde{\tau}(z)$ is a polynomial of at most first-degree, $\sigma(z)$ and $\widetilde{\sigma}(z)$ are polynomials of at most second-degree and $\psi(z)$ is a function of hypergeometric-type \cite{Nikiforov}.
Then the reduced equation which is called as basic equation of the method, can be solved systematically by means of special orthogonal functions and eigenstate solutions can be achieved completely \cite{Nikiforov,1999a,egrifes b,ikhdair}.
Conformable fractional form of this method is introduced in order to solve conformable fractional order Schr\"odinger equation and presented in our recent work \cite{karayer}. In the case of the conformable fractional NU method fractional orders are inserted in the basic equation. Then using the key property of the conformable fractional derivative operator one can reach to the following second order differential equation;
\begin{equation}
\label{eq:fnukd}
\psi''(z)+\frac{\widetilde{\tau}_{f}(z)}{\sigma_{f}(z)}\psi'(z)+\frac{\widetilde{\sigma}(z)}{\sigma_{f}^{2}(z)}\psi(z)=0.
\end{equation}
where $\widetilde{\tau}_{f}(z)=(1-\mu)z^{-\mu}\sigma(z)+\widetilde{\tau}(z)$ and $\sigma_{f}(z)=z^{1-\mu}\sigma(z)$ and the subscript $f$ stands for fractional. Boundary conditions of the conformable fractional NU method is determined by the degrees of the coefficients in the basic equation of the method given by Eq.(\ref{eq:fnukd}). Here $\widetilde{\tau}_{f}(z)$ is a function of at most $\mu th$ degree (which means that this function can also be equal to a constant), $\sigma_{f}(z)$ is a function of at most $(\mu+1) th$ (i.e. the degree of this function can also be equal to $1$) and $\widetilde{\sigma}(z)$ is a function of at most $2\mu th$ degree (i.e. the degree of this function can also be equal to $0$ or $\mu$). If any fractional order differential equation is reduced to the basic equation using the key property of local fractional derivative operator, then it can be solved analytically by the conformable fractional NU method.
After determining the following newly defined functions related to the initial functions in the basic equation, eigenvalue and eigenfunction solution of Eq.(\ref{eq:fnukd}) can be obtained;
\begin{equation}
\label{eq:fpi}
\pi_{f}(z)=\frac{\sigma_{f}'(z)-\widetilde{\tau}_{f}(z)}{2}\pm\sqrt{(\frac{\sigma_{f}'(z)-\widetilde{\tau}_{f}(z)}{2})^{2}-\widetilde{\sigma}(z)+k(z)\sigma_{f}(z)}.
\end{equation}
Recall that $\pi_{f}(z)$ is a function of at most $\mu th$ degree. Providing this condition the expression under the square root sign must be the square of a $\mu th$ order function. Thus the function $k(z)$ under the square root sign must be chosen properly.
\begin{equation}
\label{eq:fh}
\tau_{f}(z)=\widetilde{\tau}_{f}(z)+2\pi_{f}(z).
\end{equation}
\begin{equation}
\label{eq:fk}
\lambda(z)=k(z)+\pi_{f}'(z).
\end{equation}
\begin{equation}
\label{eq:fp}
\lambda_{n}(z)=-n\tau_{f}'(z)-\frac{n(n-1)}{2}\sigma_{f}''(z) \,\,\,\,\,\,\,\, (n=0,1,2,...).
\end{equation}
In order to obtain eigenvalue solution, the function $ \lambda(z) $ in Eq.(\ref{eq:fk}) is taken equal to $ \lambda_{n}(z) $ in Eq.(\ref{eq:fp}). For the eigenfunction solution, functions $\phi(z)$ and $y_{n}(z)$ given by;
\begin{equation}
\label{fi}
\frac{\phi'(z)}{\phi(z)}=\frac{\pi_{f}(z)}{\sigma_{f}(z)},
\end{equation}
\begin{equation}
\label{rho}
(\sigma_{f}(z)\rho(z))'=\tau_{f}(z)\rho(z).
\end{equation}
\begin{equation}
\label{rod}
y_{n}(z)=\frac{B_{n}}{\rho(z)}\frac{d^{n}}{dz^{n}}[\sigma_{f}^{n}(z)\rho(z)],
\end{equation}
are inserted in $\psi(z)=\phi(z)y(z)$.
\section{Solution of conformable fractional KG equation for the generalized Hulthen potential}
\label{sec:3}
Conformable fractional form of one dimensional KG equation for the generalized Hulthen potential given by Eq.(\ref{cfkghp}) is written by replacing integer orders with fractional orders;
\begin{eqnarray}\label{cfkghp1}
D^{\mu}D^{\mu}\psi_{q}(z)+\frac{S_{0}-qz^{\mu}}{z^{\mu}(S_{0}-qz^{\mu})}D^{\mu}\psi_{q}(z)+
\frac{1}{\big[z^{\mu}(S_{0}-qz^{\mu})
\big]^{2}}\big[-(\gamma^2+\nonumber \\
q\beta^2+q^2\epsilon^2)z^{2\mu}+S_{0}(\beta^2+2q\epsilon^2)z^{\mu}-S_{0}^2 \epsilon^2\big]\psi_{q}(z)=0
\end{eqnarray}
Using the key property of the conformable fractional derivative definition, Eq.(\ref{cfkghp1}) can be transformed to a second order differential equation;
\begin{eqnarray}\label{cfkghp2}
\psi_{q}''(z)+\frac{(S_{0}-qz^{\mu})(2-\mu)}{z(S_{0}-qz^{\mu})}\psi_{q}'(z)+
\frac{1}{\big[z(S_{0}-qz^{\mu})\big]^{2}}\big[-(\gamma^2
+q\beta^2+ \nonumber \\
q^2\epsilon^2)z^{2\mu}+S_{0}(\beta^2+2q\epsilon^2)z^{\mu}-S_{0}^2 \epsilon^2\big]\psi_{q}(z)=0.
\end{eqnarray}
Comparing this equation with the basic equation of the method, the parameters in Eq.(\ref{eq:fnukd}) are determined as;
\begin{eqnarray}\label{prmtr}
\widetilde{\tau}_{f}(z)=(S_{0}-qz^{\mu})(2-\mu) \nonumber \\
\sigma_{f}(z)=z(S_{0}-qz^{\mu})\nonumber \\
\widetilde{\sigma}(z)=-(\gamma^2+q\beta^2+q^2\epsilon^2)z^{2\mu}+S_{0}(\beta^2+2q\epsilon^2)z^{\mu}-S_{0}^2 \epsilon^2.
\end{eqnarray}
Since $\widetilde{\tau}_{f}(z)$, $\sigma_{f}(z)$ and $\widetilde{\sigma}(z)$ are $\mu th$, $(\mu +1)th$ and $2\mu th$ order, conformable fractional NU method can be used in order to obtain the bound state solutions of local fractional KG equation for the generalized Hulthen potential. After substituting the parameters given by Eq.(\ref{prmtr}) into Eq.(\ref{eq:fpi}), the function $\pi_{f}(z)$ can be obtained as;
\begin{eqnarray}
\label{eq:fpihp}
\pi_{f}(z)=\frac{1}{2}\big\{(\mu -1)S_{0}-qz^{\mu}(2\mu-1)\pm \nonumber \\
\big[[q^{2}(2\mu-1)^{2}+4(\gamma^{2}+q\beta^{2}+q^{2}\epsilon^{2})-4k_{\mu}q]z^{2\mu}+ [-2(\mu-1)(2\mu-1)qS_{0}- \nonumber \\ 4S_{0}(\beta^{2}+2q\epsilon^{2})+4k_{\mu}S_{0}]z^{\mu}+4S_{0}^{2}\epsilon^{2}+(\mu-1)^{2}S_{0}^{2}\big]^{\frac{1}{2}}\big\}.
\end{eqnarray}
For the requirement of $\pi_{f}(z)$ to be a $\mu th$ degree function, parameter $k_{\mu}$ which is given by $k=k_{\mu}z^{\mu -1}$ must be chosen properly;
\begin{equation}\label{kmu}
k_{\mu_{1,2}}=\frac{1}{2}[2\beta^{2}+\mu(\mu-1)q\pm \sqrt{(\mu^2 q^2+4\gamma^2)((\mu-1)^2+4\epsilon^2)}]
\end{equation}
Taking into account the $\pm$ signs in Eq.(\ref{kmu}), four different forms of $\pi_{f}(z)$ are obtained. The function $\pi_{f}(z)$ which is chosen as the function $\tau_{f}(z)$ given by Eq.(\ref{eq:fh}) has a negative derivative for physical validity \cite{Nikiforov}.This condition is provided by;
\begin{equation}\label{kmu2}
k_{\mu}=\frac{1}{2}[2\beta^{2}+\mu(\mu-1)q - \sqrt{(\mu^2 q^2+4\gamma^2)((\mu-1)^2+4\epsilon^2)}].
\end{equation}
Using the chosen $k_{\mu}$ in Eq.(\ref{kmu2}) the function $\pi_{f}(z)$ is obtained as;
\begin{eqnarray}
\pi_{f}(z)=\frac{1}{2}\bigg[S_{0}\big(\mu-1+\sqrt{(\mu-1)^{2}+4\epsilon^{2}}\big)-\nonumber \\ \big(q(2\mu-1+\sqrt{(\mu -1)^{2}+4\epsilon^{2}})+\sqrt{\mu^{2}q^{2}+4\gamma^{2}}\big)z^{\mu}\bigg].
\end{eqnarray}
After determining $\pi_{f}(z)$, one can obtain the functions $\tau_{f}(z)$, $\lambda(z)$ and $\lambda_{n}(z)$ from Eq.(\ref{eq:fh}), Eq.(\ref{eq:fk}) and Eq.(\ref{eq:fp}) respectively;
\begin{equation}
\tau_{f}(z)=S_{0}\big(1+\sqrt{(\mu-1)^{2}+4\epsilon^{2}}\big)-\big(q(\mu+1+\sqrt{(\mu-1)^{2}+4\epsilon^{2}})+\sqrt{\mu^{2}q^{2}+4\gamma^{2}}\big)z^{\mu},
\end{equation}
\begin{eqnarray}
\lambda(z)=\frac{1}{2}\big[2\beta^{2}-\mu^{2}q-\sqrt{(\mu^{2}q^{2}+4\gamma^{2})((\mu-1)^{2}+4\epsilon^{2})}-\nonumber \\
\mu \sqrt{\mu^{2}q^{2}+4\gamma^{2}}-\mu q \sqrt{(\mu-1)^{2}+4\epsilon^{2}}\big]z^{\mu-1},
\end{eqnarray}
\begin{equation}
\lambda_{n}(z)=n\mu\big[q(\mu +1+\sqrt{(\mu-1)^{2}+4\epsilon^{2}})+\sqrt{\mu^{2}q^{2}+4\gamma^{2}}+\frac{(n-1)(\mu+1)q}{2}\big]z^{\mu-1}.
\end{equation}
For $\lambda (z)= \lambda_{n}(z)$, eigenvalue spectra of the problem is established by recalling the equalities given by Eq.(\ref{qntts});
\begin{eqnarray}
E^{2}-m^{2}=\frac{\alpha^{2}}{4}\bigg\{(\mu-1)^{2}- \nonumber \\
\big[\frac{4mS_{0}-\mu^{2}\alpha^{2}q-\mu \alpha\sqrt{\mu^{2}\alpha^{2}q^{2}+4S_{0}^{2}}(1+2n)-\mu(\mu+1)n(n+1)q\alpha^{2})}{\mu \alpha^{2}q(2n+1)+\alpha\sqrt{\mu^{2}\alpha^{2}q^{2}+4S_{0}^{2}}}\big]^{2}\bigg\}.
\end{eqnarray}
In order to obtain eigenfunction solution, the function $\phi(z)$ is determined by using Eq.(\ref{fi});
\begin{equation}\label{phir}
\phi(z)=z^{\frac{1}{2}(\mu-1+\sqrt{(\mu-1)^{2}+4\epsilon^{2}})}(S_{0}-qz^{\mu})^{\frac{1}{2\mu q}(\mu q+\sqrt{\mu^{2} q^{2}+4\gamma^{2}})}.
\end{equation}
Then, the functions $\rho(z)$ and $y_{n}(z)$ are obtained from Eq.(\ref{rho}) and Eq.(\ref{rod});
\begin{equation}
\rho(z)=z^{\sqrt{(\mu-1)^{2}+4\epsilon^{2}}}(S_{0}-qz^{\mu})^{\frac{1}{\mu q}(\sqrt{\mu^{2} q^{2}+4\gamma^{2}})}.
\end{equation}
\begin{eqnarray}\label{rodr}
y_{n}(z)=B_{n}z^{-\sqrt{(\mu-1)^{2}+4\epsilon^{2}}}(S_{0}-qz^{\mu})^{-\frac{1}{\mu q}(\sqrt{\mu^{2} q^{2}+4\gamma^{2}})}\nonumber \\
\frac{d^{n}}{dz^{n}}\big[z^{n+\sqrt{(\mu-1)^{2}+4\epsilon^{2}}}(S_{0}-qz^{\mu})^{n+\frac{1}{\mu q}(\sqrt{\mu^{2} q^{2}+4\gamma^{2}})}\big].
\end{eqnarray}
The right hand sides of Eq.(\ref{phir}) and Eq.(\ref{rodr}) are inserted in the transformation $\psi(z)=\phi(z)y_{n}(z)$:
\begin{eqnarray}
\psi(z)=B_{n}z^{\frac{1}{2}(\mu-1+\sqrt{(\mu-1)^{2}+4\epsilon^{2}})}(S_{0}-qz^{\mu})^{\frac{1}{2\mu q}(\mu q+\sqrt{\mu^{2} q^{2}+4\gamma^{2}})}\nonumber \\
z^{-\sqrt{(\mu-1)^{2}+4\epsilon^{2}}}(S_{0}-qz^{\mu})^{-\frac{1}{\mu q}(\sqrt{\mu^{2} q^{2}+4\gamma^{2}})}\nonumber \\
\frac{d^{n}}{dz^{n}}\big[z^{n+\sqrt{(\mu-1)^{2}+4\epsilon^{2}}}(S_{0}-qz^{\mu})^{n+\frac{1}{\mu q}(\sqrt{\mu^{2} q^{2}+4\gamma^{2}})}\big].
\end{eqnarray}
Consequently the eigenvalue and the eigenfunction spectra of a spinless particle in the generalized Hulthen potential which are identical to results in Ref.\cite{egrifes} and Ref.\cite{ikh} for $\mu=1$, have been obtained completely in view of conformable fractional calculus.
\section{Results and Discussion}
Fractionalization of the relativistic wave equations have been widely studied by using Riemann-Liouville or Caputo fractional derivative operator in general. Since all fractional derivative operators have nonlocal character and they do not satisfy Leibniz rule, the wave equations including these operators are so complicated in order to obtain an analytical solution related to the fractional dimension of the space. Herein, a local fractional derivative operator is needed to arrive an exact solution. Local fractional form of the KG equation is proposed in order to describe the dynamics of a relativistic particle moving in the generalized Hulthen potential by means of conformable fractional derivative operator. Therefore variation of the energy and the wavefunction spectra with respect to the fractional order can be obtained in a more realistic manner.
In the presented figures, evolution of ground state energy of a spinless particle in deformed Hulthen potential is represented as a function of the fractional order $\mu$ for three different values of the deformation parameter $q$ and for three different values of the range parameter $\alpha$, namely $0.5$, $1$ and $2$. It can be seen that the curves increase more rapidly with increasing $q$ to a particular value of $\mu$. Then they decrease to the well known values of ground state energy at $\mu=1$ when green line in Figure~(\ref{fig:1}) is excluded.
In Figure~(\ref{fig:2}) and Figure~(\ref{fig:3}) the initial values of the curves start at $\mu \neq 0$ for all values of $q$. Moreover maximum values of the curves are in evidence when $\mu$ reaches to the value $1$. On the whole, all curves intersect two by two at different points which correspond to the different values of $\mu$ and the curves reach maximum values more rapidly with increasing $\alpha$.
Moreover ground state energy for the deformed Hulthen potential is given numerically for fixed $S_{0}=0.25$ and given $\alpha$ in Table~(\ref{tab:1}), Table~(\ref{tab:2}) and Table~(\ref{tab:3}).
|
\section*{Introduction}
\mylabel{Introduction}
A central result in the classification of algebraic surfaces is that
there are only four types of algebraic surfaces $S$ with first Chern class $c_1(S)=0$.
These are the abelian surfaces, the bielliptic surfaces,
the K3 surfaces and the Enriques surfaces. They are distinguished by their second Betti numbers,
which take the respective values
$$
b_2(S)=6,\quad b_2(S)=2,\quad b_2(S)=22\quadand b_2(S)=10.
$$
This holds regardless of the characteristic. Over the complex numbers,
the bielliptic surfaces are quotients of abelian surfaces
by a finite free group action, and that Enriques surfaces $Y=X/G$ are quotients of K3 surfaces
$X$ by free involutions.
This no longer holds true literally when the characteristic $p\geq 0$ of
the ground field $k$ divides the order of the group $G$.
A fundamental insight of Bombieri and Mumford \cite{Bombieri; Mumford 1977}
was that much carries over if one replaces finite groups by finite group schemes. This comes with
the price that singularities appear on the covering $X$, although the
quotient $Y=X/G$ is smooth. In some sense, the non-smoothness of $X$ and $G$ cancel each other.
For example, I have used this effect
to give the correct Kummer construction in characteristic $p=2$,
by replacing an abelian surface by the self-product of a rational cuspidal curve
\cite{Schroeer 2007}.
For Enriques surfaces $Y$, the numerically trivial part $P=\Pic^\tau_{Y/k}$
of the Picard scheme is a group scheme of order two.
By the Tate--Oort classification of finite group schemes of prime order
\cite{Tate; Oort 1970}, there are three possibilities in characteristic $p=2$.
In the case where the group scheme $P$ is unipotent, one says that $Y$ is
a \emph{simply-connected Enriques surface}.
These surface come along with a torsor $\epsilon:X\ra Y$ with respect
to the Cartier dual $G=\underline{\Hom}(P,\GG_m)$, which here indeed is a local group scheme.
The torsor takes over the role
of the universal covering, and is therefore called the \emph{K3-like covering}.
This is very adept terminology, because the K3-like covering has the same cohomological
properties of a K3 surface, yet is never a smooth surface.
The goal of this paper is to study K3-like coverings $X$, and the ensuing
simply connected Enriques surfaces $Y$.
These where studied, among others, by Bombieri and Mumford \cite{Bombieri; Mumford 1976},
Blass \cite{Blass 1982}, Lang \cite{Lang 1983, Lang 1988},
Cossec and Dolgachev \cite{Cossec; Dolgachev 1989}, Ekedahl and Shepherd-Barron
\cite{Ekedahl; Shepherd-Barron 2004} in the non-normal case,
Ekedahl, Hyland and Shepherd Barron \cite{Ekedahl; Hyland; Shepherd-Barron 2012} in the normal case,
Katsura and Kondo \cite{Katsura; Kondo 2014}, and Liedtke \cite{Liedtke 2015}.
There are, however, many open foundational questions.
In this paper, we shall concentrate on \emph{normal K3-like coverings}, although
we take care that our theory works in general.
An open question was whether non-rational singularities could appear on a K3-like covering $X$.
One main result of this paper is that they do:
\begin{maintheorem}
{\rm (Compare Theorem \ref{existence K3-like II}.)}
There are normal K3-like coverings $X$
containing an elliptic double point, which is obtained from the contraction of a rational cuspidal $(-1)$-curve.
\end{maintheorem}
We actually give a systematic way to produce normal K3-like coverings, not bases on
explicit equations, but by a procedure of Frobenius base-change combined with flops.
If a non-rational singularity appears, the situation is rather special:
\begin{maintheorem}
{\rm (Compare Theorem \ref{unique elliptic singularity}.)}
If a normal K3-like covering $X$ contains a non-rational singularity,
then there are no other singularities on $X$.
\end{maintheorem}
Each elliptic fibration $\varphi:Y\ra\PP^1$ on the Enriques surface
comes with a jacobian fibration, which is a rational elliptic surface $J\ra\PP^1$.
According to the result of Liu, Lorenzini and Raynaud \cite{Liu; Lorenzini; Raynaud 2005}, the Kodaira types
of all fibers $Y_b$ and $J_b$ coincide. The latter were classified by
Persson \cite{Persson 1990} and Miranda \cite{Miranda 1990} over the complex numbers,
which was extended by Lang in \cite{Lang 1994, Lang 2000} to characteristic two.
There are 110 families with only reduced fibers.
We can show that, with one strange undecided case,
all these arise from simply-connected Enriques surfaces:
\begin{maintheorem}
{\rm (Compare Theorem \ref{good mutations exist}.)}
For each rational elliptic surface $J\ra\PP^1$
with only reduced fibers, with the possible exception of the case when a fiber of Lang type 10C is present,
there is a simply-connected Enriques surface $Y$ whose K3-like covering $X$
is birational to the Frobenius pullback $X'=J^\fpb$,
and having an elliptic fibration $\varphi:Y\ra\PP^1$ whose jacobian is $J\ra\PP^1$.
\end{maintheorem}
This relies on Shioda's theory of Mordell--Weil lattices \cite{Shioda 1990}
and their classification for rational elliptic surfaces by Oguiso and Shioda \cite{Oguiso; Shioda 1991}.
The passage
$$
X'\longleftarrow S\lra X
$$
from the Frobenius pullback $X'=J^\fpb$
to a K3-like covering $X$ is the fundamental idea of the paper.
It should be seen as a \emph{flop}, a notion coming from the minimal model program for the
classification of higher-dimensional algebraic scheme. Roughly speaking, it describes
the passage from one minimal model to another, without changing the canonical class.
Ekedahl, Hyland and Shepherd Barron showed
in their groundbreaking paper \cite{Ekedahl; Hyland; Shepherd-Barron 2012}
that for simply-connected Enriques surfaces $Y$
whose K3-like covering $X$ has only rational double points,
the tangent sheaf $\Theta_{X/k}=\underline{\Hom}(\Omega^1_{X/k},\O_X)$ is locally free,
in fact isomorphic to $\O_{X}\oplus\O_X$. We generalize this to arbitrary normal K3-like coverings.
Indeed, the main reason for the flop $X'\leftarrow S\ra X$ is to make the tangent space
trivial.
It turns out that in the restricted Lie algebra $\ideal g= H^0(X,\Theta_{X/k})$,
each vector is $p$-closed, that is, each line is invariant under the $p$-map
$x\mapsto x^{[p]}$. Using the correspondence between finite-dimensional restricted Lie algebras
$\ideal l$
and finite height-one groups schemes $G=\Spec(U^{[p]}(\ideal l)^\vee)$, we get
a rational map
$$
\xymatrix{
\PP(\ideal g) \ar@{-->}[r] & \shM_\Enr, & \ideal l\ar@{|->}[r] & X/G
}
$$
from the \emph{twistor curve} $\PP(\ideal g)$ into the moduli stack of Enriques surfaces $\shM_\Enr$.
It was introduced in \cite{Ekedahl; Hyland; Shepherd-Barron 2012}
to study this moduli stack.
We like to call it the \emph{twistor construction}, because it resembles the
2-sphere of complex hyperk\"ahler manifolds obtained by rotating the
complex structure while keeping the underlying topological space.
Whether this is more than an analogy remains to be seen.
The twistor map is not everywhere defined.
This is explained by our general theory of \emph{Zariski singularities},
which is crucial to understand the non-smoothness of the K3-like coverings
$\epsilon:X\ra Y$ and their generalizations.
We therefore develop a theory of Zariski singularities in full generality for arbitrary dimensions
$n\geq 2$ and characteristics $p>0$.
These are hypersurface singularities $A=R/(g)$, where $R=k[[x_1,\ldots,x_n,z]]$
where the power series can be chosen of the form $g=z^p-f(x_1,\ldots,x_n)$.
Using methods from commutative algebra involving free resolutions and projective dimensions,
we get the following criterion for the freeness of the tangent module:
\begin{maintheorem}
{\rm (See Theorem \ref{theta and pd}.)}
An hypersurface singularity $A=R/(g)$ has free tangent module $\Theta_{A/k}$ of rank $n$ if
and only if $A$ is geometrically reduced, and the module of first-oder deformations
$T^1_{A/k}$ has projective dimension
$\pd\leq 2$.
\end{maintheorem}
This automatically holds for the 2-dimensional geometrically reduced Zariski singularities.
Note that it is usually difficult to recognize a Zariski singularity, if it is not
given in the usual form $g=z^p-f(x_1,\ldots,x_n)$.
We therefore develop a homological characterization of Zariski singularities, which
is analogous to Serre's characterization of regularity in local rings.
Roughly speaking:
An hypersurface singularity $A=R/(g)$ Zariski singularity if and only if there
is an $\alpha_2$-action to that the structure sheaf $\O_Z$ of the
closed orbit $Z\subset\Spec(A)$ has finite projective dimension.
Furthermore, we develop easy tools, involving Tjurina numbers, local class groups, local fundamental groups,
length formulas, and Eisenbud's matrix factorizations \cite{Eisenbud 1980} to recognize Zariski singularities.
The following are the Zariski singularities among the rational double points:
$$
A_1, D_{2n}^0, E_7^0, E_8^0\; (p=2),\quad A_2, E_6^0, E_8^0\; (p=3),\quad
A_4, E_8^0\; (p=5),\quad A_{p-1}\; (p\geq 7).
$$
Finally, we develop a theory of \emph{canonical coverings} for arbitrary proper algebraic
schemes $Y$ with $h^0(\O_Y)=1$,
using the Raynaud Correspondence $ H^1(Y_\fppf, G_Y) = \Hom(\hat{G},\Pic^\tau_{Y/k})$.
There is indeed a canonical choice: Let $G$ be the Cartier dual to the Frobenius kernel
$\Pic_{Y/k}[F]$. To understand the singularities on the resulting canonical coverings
$\epsilon:X\ra Y$, we restrict the covering to closed subscheme $A\subset Y$
that have certain singularities, or on which the $G$-torsor trivializes.
For example, for ADE-configurations $A$ on smooth surfaces $Y$ we introduce
the \emph{fundamental singular locus}, which consist of the singular locus
of the fundamental cycle, and show:
If $G$ is non-reduced, then the $G$-torsor $X\ra Y$ is singular over
the fundamental singular locus of each ADE-configuration.
Under the assumption that $X$ is normal, this gives some crucial restrictions on the geometry
of the curves, in particular if $Y$ is an Enriques surface.
\medskip
The paper is organized as follows:
In Section \ref{Zariski singularities}, we develop the
theory of Zariski singularities in arbitrary dimensions. Section \ref{Homological characterization}
contains a homological characterization of such singularities.
This is applied to rational double points in Section \ref{Rational double points}.
In Section \ref{Canonical coverings}, we give a general theory of canonical coverings,
based on the Raynaud correspondence.
In Section \ref{Simply-connected enriques}, we review the notion of
simply-connected Enriques surfaces and discuss their basic properties.
In Section \ref{Trivial tangent sheaf}, we introduced six conditions
that normal K3-like coverings satisfy, and that indeed allow the construction of
K3-like coverings. The resulting twistor curves are discussed in Section \ref{Twistor curves}.
In Section \ref{K3-like coverings fibrations}, we relate K3-like coverings $X$
to elliptic fibrations on Enriques surfaces $Y$ and rational surfaces $J$.
To put this to work, we use Ogg's Formula to understand the behavior
of Kodaira types under Frobenius base-change in Section \ref{Ogg's formula}.
In Section \ref{Tate algorithm}, we give a geometric interpretation of the Tate Algorithm,
which computes minimal Weierstra{\ss} equations, in order to describe certain
elliptic singularities.
All this is put into action in Section \ref{Frobenius pullback}, which deals
with the Frobenius pullback of rational elliptic surfaces.
In Section \ref{Lang classification}, we discuss Lang's Classification of
rational elliptic surfaces in characteristic $p=2$, which will be crucial for our goals.
Section \ref{K3-like with edp} the construction of normal K3-like coverings with
an elliptic singularity.
The absence of other singularities is established in Section \ref{Uniqueness edp}.
In Section \ref{K3-like with rdp} we study normal K3-like coverings with rational singularities.
The last Section \ref{Canonical covering fibers} contains a discussion of
the canonical coverings of fibers of Kodaira type $\II$, $\III$ and $\IV$.
\begin{acknowledgement}
This paper resulted from many stimulating discussions with
Shigeyuki Kondo and Toshiyuki Katsura, to whom the
author is very grateful. Large parts of the paper were
written during two stays of the author at the Nagoya University, financed by
Shigeyuki Kondo's grant
JSPS, Grant-in-Aid for Scientific Research (S) No.\ 15H05738.
The author greatly benefited from the ensuing mathematical interchange with Shigeyuki Kondo,
and wishes to thank the Graduate School of Mathematics, Nagoya University for its hospitality.
\end{acknowledgement}
\section{Zariski singularities}
\mylabel{Zariski singularities}
Let $k$ be a ground field, and $A$ be a complete local noetherian ring
such that the canonical map $k\ra A/\maxid_A$ is bijective. We then get an identification $k=A/\maxid_A$ of
the ground field with the residue field.
Let $\dim(A)\leq \edim(A)$ be the dimension and embedding dimension.
Equality holds if and only if the
local noetherian ring $R$ is regular.
We are interested in the following situation:
\begin{proposition}
The two conditions below are equivalent:
\begin{enumerate}
\item
The ring $A$ is equidimensional, contains no embedded associated primes, and we have
$\edim(A)\leq\dim(A)+1$.
\item
The ring $A$ is isomorphic to $R/(g)$, with $R=k[[x_1,\ldots,x_n,z]]$ and some power series
$g\in R$ that is neither a unit nor zero.
\end{enumerate}
If this holds, then $ \dim(A)=n$, and the ring $A$ is Cohen--Macaulay and Gorenstein.
\end{proposition}
\proof
If (ii) holds, then Krull's Principal Ideal Theorem ensures that every
irreducible component of $\Spec(A)$ is $n$-dimensional.
Moreover, since $R$ is Cohen--Macaulay and Gorenstein, the same holds for $A=R/(g)$.
In particular, $A$ contains no embedded associated prime.
To see (i)$\Rightarrow$(ii), set $n=\dim(A)$.
Using the completeness and the assumption on the embedding dimension, we write
$A=R/\ideal a$ for some ideal $\ideal a$ in the formal power series ring
$R=k[[x_1,\ldots,x_n,z]]$, which is neither the zero nor the unit ideal. According to \cite{EGA IVd}, Proposition 21.7.2
the subscheme $\Spec(A)\subset\Spec(R)$ is a Weil divisor.
Then $\ideal a$ is generated by a non-zero non-unit $g\in R$, because the ring $R$ is factorial.
\qed
\medskip
It is convenient to say that the local scheme $\Spec(A)$,
and also the local ring $A$, is a \emph{hypersurface singularity} if
the equivalent conditions of the Proposition hold.
Note that this includes the case where the local ring $A$ is regular, when
we may choose $g=z$.
An isomorphism $A\simeq R/(g)$
as above is called a \emph{presentation}.
Note also that in light of Condition (i),
the ring $R'=R\otimes_kk'$ is a hypersurface singularity if and only if $R$ is a hypersurface
singularity, where $k\subset k'$ is a finite field extension.
Let us write
$$
J=(g_{x_1},\ldots,g_{x_n},g_z)\subset A
$$
for the \emph{jacobian ideal} in $A=R/(g)$ generated by residue classes of the partial derivatives
$g_{x_i}=\partial g/\partial x_i$, $1\leq i\leq n$ and $g_z=\partial g/\partial z$.
The primes $\primid\subset A$ not containing the jacobian ideal are exactly those
for which the local noetherian rings $A_\primid$ are geometrically regular.
By abuse of language, we call the closed subset $\Spec(A/J)\subset\Spec(A)$ the \emph{locus of non-smoothness}.
Note that
$$
A/J=R/(g,g_{x_1},\ldots,g_{x_n},g_z),
$$
where the ideal on the right is the \emph{Tjurina ideal} in $R$.
The ring $A$ is reduced or integral if and only
if the power series $g\in R$ is squarefree or irreducible, respectively. This holds because
$R=k[[x_1,\ldots,x_n,z]]$ is factorial.
The ring $A$ is regular if and only if one of the partial derivatives of $g(x_1,\ldots,x_n,z)$ is a unit.
We say that the $R$ is a \emph{geometrically isolated hypersurface singularity} if
$\dim(A/J)=0$, such that $\Spec(A)$ is geometrically regular outside the closed point.
Now suppose that the ground field $k$ has positive characteristic $p>0$.
Choose some algebraically closed extension field $\Omega$,
and let $k'=k^{1/p}$ be the intermediate extension comprising all elements $\omega\in \Omega$
with $\omega^p\in k$. Then $k\subset k'$ is a purely inseparable extension of height one,
and the vector space dimension
$$
\pdeg(k) =
\dim_k(\Omega^1_{k'/k}) \in \NN\cup\left\{\infty\right\}
$$
is called the \emph{$p$-degree} of $k$. This is also known as the \emph{degree of imperfection},
and can also be seen as the cardinality of
a $p$-basis for $k\subset k^{1/p}$ or equivalently $k^p\subset k$.
If $k$ is the field of fractions for a polynomial ring or formal power series ring
in $d$ indeterminates over a perfect field $k_0$, then $\pdeg(k)=n$.
In what follows, we suppose that \emph{our ground field $k$ has finite $p$-degree}, which
holds in particular for perfect fields.
This assumption will save us some troubles when it comes to K\"ahler differentials
of formal power series, which become apparent as follows:
Let $kR^p\subset R$ be the join subring generated by the subrings $k$ and $R^p$.
Clearly, we have an equality
$R^p=k^p[[x_1^p,\ldots,x_n^p,z^p]]$ inside $R$.
Since the field extension $k^p\subset k$ is finite, the canonical maps
$$
k\otimes_{k^p} k^p[[x_1^p,\ldots,x_n^p,z^p]]\lra kR^p\subset k[[x_1^p,\ldots,x_n^p,z^p]]
$$
are bijective. It follows that the module of K\"ahler differentials
$$
\Omega^1_{R/k} = \Omega^1_{R/kR^p}
$$
is finitely generated, in fact free, with basis $dx_1,\ldots,dx_n,dz$.
For ground fields of characteristic zero, or of infinite $p$-degree, the module of K\"ahler differentials is usually
not finitely generated, not even $\maxid_R$-adically separated.
In our situation, the standard exact sequence $gR/g^2R\ra\Omega^1_{R/k}\otimes_RA\ra\Omega^1_{A/k}\ra 0$
becomes
$$
A\lra A^{n+1}\lra\Omega^1_{A/k}\lra 0,
$$
where the map on the right sends the standard basis vectors to the differentials $dx_1,\ldots,dx_n,dz$
and the map on the left is an $(n+1)\times 1$-matrix of partial derivatives given by the transpose of
$(g_{x_1},\ldots,g_{x_n},g_z)$.
The latter is non-zero if and only if the ring $A$ is geometrically reduced. Since $A$ contains no embedded
associated primes,
the map is then injective. Dualizing the short exact sequence we get an exact sequence
\begin{equation}
\label{theta sequence}
0\lra\Theta_{A/k}\lra A^{\oplus n+1}\lra A\lra\Ext^1(\Omega^1_{A/k},A) \lra 0.
\end{equation}
The term on the right is also called $T^1_{A/k}=\Ext^1(\Omega^1_{A/k},A)$,
the \emph{module of first oder deformations}. By the exact sequence, this module coincides with the residue class ring $A/J$,
whose spectrum is the locus of non-smoothness, and the tangent module
$\Theta_{A/k}=\Hom(\Omega^1_{A/k},A)$ on the left is a syzygy for $A/J$.
This reveals that the locus of non-smoothness does not depend on the presentation $A=R/(g)$,
and carries a canonical scheme structure. We also immediately get:
\begin{theorem}
\mylabel{theta and pd}
Suppose that $A$ is a hypersurface singularity.
Then the tangent module $\Theta_{A/k}$
is free of rank $n=\dim(A)$ if and only if $A$ is geometrically reduced and the $A$-module $T^1_{A/k}=A/J$
has projective dimension $\leq 2$.
\end{theorem}
The above property is rather peculiar: Write $A=R/(g)$ with some
power series $g(x_1,\ldots,x_n,z)$. Consider the
\emph{Frobenius power} $J^{[p]}=(g_{x_1}^p,\ldots, g_{x_n}^p,g_z^p)\subset A$ of the jacobian ideal
$J=(g_{x_1}, \ldots,g_{x_n},g_z)\subset A$.
The following condition was already used in \cite{Schroeer 2008} in dimension two
with the computer algebra system Magma \cite{Magma}:
\begin{proposition}
\mylabel{theta surface}
Suppose that $A$ is geometrically isolated hypersurface singularity
of dimension $n=\dim(A)$.
Then the tangent module $\Theta_{A/k}$ is
free of rank $n$ if and only if $n\leq 2$ and the length formula $\length(A/J) = p^n\cdot\length(A/J^{[p]})$ holds.
\end{proposition}
\proof
For geometrically isolated hypersurface singularities, the jacobian ideal $J\subset A$ is $\maxid_A$-primary.
According to a result of Miller (\cite{Miller 2003}, Corollary 5.2.3),
an $\maxid_A$-primary ideal $\ideal a\subset A$ has finite projective dimension if
and only if the length formula $\length(A/\ideal a) = p^n\cdot\length(A/\ideal a^{[p]})$ holds.
It follows that our condition is sufficient: If $n\leq 2$ and the length formula holds, $A/J$ has finite projective
dimension, which a priori is $\pd(A/J)\leq n\leq 2$, hence $\Theta_{A/k}$ is free of rank $n$ by Proposition \ref{theta and pd}.
The condition is also necessary: If the tangent module
$\Theta_{X/k}$ is free, then the module $A/J$ has finite projective dimension $\leq 2$,
and the length formula holds by Miller's result. Moreover, the Auslander--Buchsbaum Formula
$\pd(A/J) +\depth(A/J) =\depth(A) $
immediately gives $n=\depth(A)=\pd(A/J)\leq 2$.
\qed
\medskip
We say that our complete local ring $A$ is a \emph{Zariski singularity} if it is a hypersurface
singularity admitting a presentation $A=R/(g)$ with a power series of the form
$$
g(x_1,\ldots,x_n,z)= z^p-f(x_1,\ldots,x_n),
$$
where $f\in k[[x_1,\ldots,x_n]]$ is not a unit. This term was coined by
Ekedahl, Hyland and Shepherd-Barron \cite{Ekedahl; Hyland; Shepherd-Barron 2012}
in the case of isolated surface singularities,
apparently referring to the theory of \emph{Zariski surfaces}, compare Blass and J.\ Lang \cite{Blass; Lang 1987}.
Clearly, the Zariski singularity $A$ is reduced if and only if $f$ is not a $p$-power.
Using \cite{AC 8-9}, Chapter VIII, \S7, No.\ 4, Proposition 7, we immediately get:
\begin{proposition}
\mylabel{hilbert-samuel multiplicity}
The Hilbert--Samuel multiplicity $e(A)$ of a Zariski singularity satisfies the
inequalities $1\leq e(A)\leq p$.
\end{proposition}
Now set $A_0=k[[x_1,\ldots,x_n]]$ and $J_0=(f_{ x_1},\ldots,f_{x_n})\subset A_0$.
Then $J=J_0A$ is the jacobian ideal for $A$. This has the following consequence:
\begin{proposition}
\mylabel{theta and depth}
Suppose that $A$ is a geometrically reduced Zariski singularity.
Then the tangent module $\Theta_{A/k}$
is free of rank $n=\dim(A)$ if and only if the ring
$A_0/J_0$ has depth $\geq n-2$.
\end{proposition}
\proof
Clearly, $A$ is a finite flat algebra over $A_0$ of degree $p$, so the projective dimension of
$A/J=A/J_0A=A_0/J_0\otimes_{A_0}A$ as $R$-module
coincides with the projective dimension of $A_0/J_0$ as an $A_0$-module.
By the Auslander--Buchsbaum Formula, we have
$$
\pd(A_0/J_0) + \depth(A_0/J_0) = \depth(A_0) = n,
$$
and the result follows from Proposition \ref{theta and pd}.
\qed
\medskip
In dimension $n=2$, the condition becomes vacuous, and we get:
\begin{corollary}
\mylabel{theta for zariski surface}
The tangent module $\Theta_{A/k}$ of any two-dimensional geometrically reduced Zariski singularity is free of rank two.
\end{corollary}
If $A$ is a geometrically isolated hypersurface singularity, that is, the $A$-module $T^1_{A/k}$ has finite length,
its length $\tau=\length T^1_{A/k}$ is called the the \emph{Tjurina number}.
This is also the colength of the jacobian ideal $J\subset A$, or equivalently the colength of the
Tjurina ideal in $R$. For Zariski singularities, we get:
\begin{proposition}
\mylabel{tjurina for isolated zariski}
If $A$ is a geometrically isolated Zariski singularity given by a formal power series
$g=z^p-f(x_1,\ldots,x_n)$, then the Tjurina number satisfies
$$
\tau=p\cdot \length k[[x_1,\ldots,x_n]]/(f_{x_1},\ldots, f_{x_n}).
$$
In particular, the Tjurina number $\tau\geq 0$ is a multiple of the characteristic $p>0$.
\end{proposition}
Finally, we state some useful facts on Zariski-singularities pertaining to fundamental groups
and class groups:
\begin{proposition}
\mylabel{coverings for normal zariski}
Let $A$ be a normal Zariski singularity, and
$A\subset B$ be a finite ring extension, with $\Spec(B)$ normal and connected. Assume that $\Spec(B)\ra\Spec(A)$
is \'etale in codimension one.
Then there is a finite separable field extension $k\subset k'$ with $B=A\otimes_kk'$.
\end{proposition}
\proof
Write $A=R/(g)$ as usual, and set $A_0=k[[x_1,\ldots,x_n]]$.
The resulting morphism $\Spec(A)\ra\Spec(A_0)$ is a universal homeomorphism that is
finite and flat of degree $p$.
Let $U_0\subset\Spec(A_0)$ be the open subset corresponding to the regular locus $U\subset\Spec(A)$.
According to \cite{SGA 1}, Expose IX, Theorem 4.10 every finite \'etale covering $U'\ra U$ is the
pullback of some finite \'etale covering $U_0'\ra U_0$.
The latter extends to $\Spec(A_0)$, by the Zariski--Nagata Purity Theorem
\cite{SGA 2}, Expose X, Theorem 3.4.
Since the complete local ring $A_0$ is henselian, the extension must have the form $A_0'=A_0\otimes_kk'$
for some finite separable field extension $k\subset k'$. Consider the pullback $A'=A\otimes_{A_0} A_0'=A\otimes_kk'$.
Then $\Spec(A')$ is regular in codimension one and Cohen--Macaulay, thus normal.
By construction, the two morphisms $\Spec(B)\ra\Spec(A)$ and $\Spec(A')\ra\Spec(A)$ coincide
over $U\subset\Spec(A)$, whose complement has codimension $d\geq 2$.
Now \cite{Hartshorne 1994}, Theorem 1.12 ensures that there is an isomorphism $B\simeq A'$ of $A$-algebras.
\qed
\begin{corollary}
\mylabel{components rational}
Suppose that $A$ is a normal two-dimensional Zariski singularity, with separably closed ground field $k$
and resolution of singularities $X\ra \Spec(A)$. Then each irreducible component $E_i$
of the exceptional divisor $E\subset X$ is a rational curve.
\end{corollary}
\proof
Fix some prime $l\neq 0$, and suppose that some $E_i$ is non-rational.
Then one finds an invertible sheaf $\shL_E\not\simeq\O_E$ on $E$ with $\shL_E^{\otimes l}\simeq\O_E$.
One easily sees that it successively extends to the infinitesimal neighborhoods $nE\subset X$,
and finally gets an invertible sheaf $\shL$ on $X$, restricting to $\shL_E$, with $\shL^{\otimes l}\simeq \O_X$.
Then the ensuing $\O_X$-algebra $\shA=\O_X\oplus\shL\oplus\ldots\oplus\shL^{\otimes 1-l}$ of rank $l>1$
defines a ring extension $A\subset B$ of degree $l$ as in the Proposition.
However, since $k$ is separably closed, we must have $B=A$, contradiction.
\qed
\begin{proposition}
\mylabel{class group for normal zariski}
Let $A$ be a normal Zariski singularity.
Then the class group $\Cl(A)$ is an abelian $p$-group,
and every element is annihilated by $p^{d+n-1}$, where $d=\pdeg(k)$ and $n=\dim(A)$.
\end{proposition}
\proof
Write $A=k[[x_1,\ldots,x_n,z]]/(g)$ for some power series $g=z^p-f(x_1,\ldots,x_n)$,
and consider the subring $A_0=[[x_1,\ldots,x_n]]$ and the finite ring extension $A_0\subset B$
with $B=k^{1/p}[[x_1^{1/p},\ldots,x_n^{1/p}]]$. The inclusion $A_0\subset B$ is finite
and flat, of degree $p^{n+d}$. Clearly, the power series $f\in A_0$ admits a $p$-th root $h\in B$,
which gives an integral homomorphism $A\ra B$ of $A_0$-algebras via $z\mapsto h$. This map must be injective,
because $A$ is integral and $\dim(A)=\dim(B_0)$. The composite inclusion $A_0\subset A\subset B$
has degree $p^{d+n}$, and the first inclusion has degree $p$, whence $A\subset B$ has
degree $d=p^{d+n-1}$. Let $U\subset\Spec(A)$ be the regular locus, and $V\subset \Spec(B)$ be its
preimage. Then $\Cl(A)=\Pic(U)$ and $ \Cl(B)=\Pic(V)=\Pic(B)=0$.
The composition $\Pic(U)\ra\Pic(V)$ with the norm map $\Pic(V)\ra\Pic(U)$
is multiplication by $\deg(V/U)=p^{d+n-1}$, and the result follows.
\qed
\begin{corollary}
\mylabel{intersection matrix for surface zariski}
Suppose that $A$ is a normal two-dimensional Zariski singularity,
with resolution of singularities $X\ra \Spec(A)$. Let $E\subset S$ be the exceptional divisor,
$E=E_1\cup\ldots\cup E_r$ be the irreducible components, and $\Phi=(E_i\cdot E_j)$ the
intersection matrix. If the ground field
$k$ is algebraically closed, then the Smith group $\ZZ^r/\Phi\ZZ^r$ is annihilated by $p$.
In particular, the absolute value of $\det(\Phi)$ is a $p$-power.
\end{corollary}
\proof
The Smith group $\ZZ^r/\Phi\ZZ^r$ is finite, because the intersection matrix is negative-definite.
Any effective Weil divisor on $\Spec(A)$ can be seen as an effective Weil divisor
$D\subset X$ that contains no exceptional curve $E_i$,
and we get intersection numbers $(D\cdot E_i)$.
This yields a well-defined homomorphism $\Cl(A)\ra\ZZ^r/\Phi\ZZ^r$.
In light of the Proposition, we merely have to check that the map is surjective.
Each element of the Smith group is represented by a difference of vectors with
non-negative entries.
Suppose we have a vector of non-negative integers $(n_1,\ldots,n_r)\in\ZZ^r$.
As the ground field $k$ is algebraically closed, there is an effective Cartier divisor $D_0\subset E$
such that $\shL_0=\O_E(D)$ has intersection numbers $(\shL_0\cdot E_i)=n_i$.
Since $A$ is complete and in particular henselian, we may extend it to some invertible sheaf $\shL$ on $X$
by \cite{EGA IVd}, Corollary 21.9.12.
This yields an element in $\Cl(A)$ having the same class as $(n_1,\ldots,n_r)$
in the Smith group $\ZZ^n/\Phi\ZZ^r$.
\qed
\section{Homological characterization}
\mylabel{Homological characterization}
Since there are many different ways to specify a complete local noetherian ring $A$,
it is a priori not clear how to verify that $A$ a Zariski singularity,
even if it is given in the form $A=R/(g)$ as in the previous section.
In this section, we establish a homological characterization
using $p$-closed derivations and actions of height one group schemes
that is both intrinsic and practical.
We start by recalling the equivalence between finite groups schemes of height one
and finite-dimensional restricted Lie algebras. For details, we refer to the monograph of
Demazure and Gabriel \cite{Demazure; Gabriel 1970}, Chapter II, \S7.
A very brief summary is contained in \cite{Schroeer 2007}, Section 1.
Fix a ground field $k$ of characteristic $p>0$.
Let $\ideal g$ be a \emph{restricted Lie algebra}, with Lie bracket $[x,y]$ and
$p$-map $x^{[p]}$, as in \cite{Demazure; Gabriel 1970}, Chapter II, Definition 3.3.
Let $U(\ideal g)$ be the universal enveloping algebra,
$U^{[p]}(\ideal g)$ be the quotient by the ideal generated by
the elements $x^p-x^{[p]}\in U(\ideal g)$ with $x\in\ideal g$, and $U^{[p]}(\ideal g)^\vee$ the dual vector space.
The diagonal map $\ideal g\ra \ideal g\oplus \ideal g$
induces the comultiplication $\Delta: U^{[p]}(\ideal g)\ra U^{[p]}(\ideal g)\otimes U^{[p]}(\ideal g)$
and dually a multiplication in $U^{[p]}(\ideal g)^\vee$. The latter becomes a commutative $k$-algebra,
and we may form
$$
G=\Spec(U^{[p]}(\ideal g)^\vee).
$$
As explained in loc.\ cit.\ Proposition 3.9, this acquires the structure of a group scheme of height one,
with $\Lie(G)=\ideal g$. Its $R$-valued points $G(R)$ is the group of elements
$g\in U^{[p]}(\ideal g)\otimes_k R$ satisfying $\Delta(g)=g\otimes g$.
Such elements $g$ are also called \emph{group-like}. Furthermore, the correspondences
$$
G\longmapsto \Lie(G)\quadand \ideal g\longmapsto \Spec(U^{[p]}(\ideal g)^\vee)
$$
are adjoint equivalences between the categories of finite groups schemes
of height one and the category of finite-dimensional restricted Lie algebras,
by loc.\ cit.\ Proposition 4.1.
If $X$ is a scheme, the right $G$-actions
$X\times G\ra X$ correspond to
the homomorphisms $\ideal g=\Lie(G)\ra\Der_k(\O_X,\O_X)$
of restricted Lie algebras. Here
$$
\Der_k(\O_X,\O_X)=\Hom_{\O_X}(\Omega^1_{X/k},\O_X) = H^0(X,\Theta_{A/k})
$$
is the restricted Lie algebra of derivations $D:\O_X\ra \O_X$, where the
$p$-map is the $p$-fold composition $D^p=D\circ\ldots\circ D$ in the associative algebra
of all differential operators $\O_X\ra\O_X$.
In the case that $X=\Spec(A)$ is affine, this boils down to specify a derivation $D:A\ra A$.
If the restricted Lie algebra $\ideal g$ is one-dimensional, with basis vector
$x\in\ideal g$, then the Lie bracket must be trivial,
and the $p$-map is determined via $x^{[p]}=\lambda x$ by some unique scalar $\lambda\in k$.
Let us write $\ideal g_\lambda$ and $G_\lambda$ for the resulting restricted Lie algebra
and height-one group scheme.
The right actions of $ G_\lambda$ on the scheme $X$ thus correspond to
derivations $D:A\ra A$ satisfying $D^p=\lambda D$. Such derivations are called
\emph{$p$-closed}.
To make this more explicit, write $U^{[p]}(\ideal g_\lambda)= k[x]/(x^p-\lambda x)$, and
choose as a vector space basis the divided powers $u_i=x^i/i!$ with $0\leq i\leq p-1$.
This basis has the advantage that the comultiplication is given by $\Delta(u_i)=\sum_{r+s=i}u_r\otimes u_s$.
Denote the dual basis by $t_i\in U^{[p]}(\ideal g_\lambda)^\vee$.
Inside the ring $U^{[p]}(\ideal g_\lambda)^\vee$, we get $t_r\cdot t_s=t_{t+s}$, and in particular
$t_1^p=0$. With $t=t_1$, one gets
$U^{[p]}(\ideal g_\lambda)^\vee=k[t]/(t^p)$.
Consequently, the action $ X\times G_\lambda\ra X$ takes the explicit form
\begin{equation}
\label{action formula}
A\lra A\otimes_k U^{[p]}(\ideal g_\lambda)^\vee,\quad
f\longmapsto \sum_{i=0}^{p-1}\frac{D^i(f)}{i!}\otimes t^i.
\end{equation}
Now suppose we have a right $G_\lambda$-action on an affine scheme $X=\Spec(A)$,
and let $a\in\Spec(A)$ be a $k$-rational point, corresponding to
a maximal ideal $\maxid \subset A$. Let
$\ideal a\subset A$ be the intersection of the kernels for the $k$-linear maps
\begin{equation}
\label{action kernels}
\maxid \stackrel{D^i}{\lra} A\lra A/\maxid ,\quad 1\leq i\leq p-1.
\end{equation}
One easily checks that this vector subspace $\ideal a\subset A$ is an ideal. We call it the \emph{orbit ideal}
for the $G_\lambda$-action, or the corresponding $p$-closed derivation $D:A\ra A$.
\begin{proposition}
\mylabel{fixed and free points}
The closed subscheme
$X'=\Spec(A/\ideal a)$ is a $G_\lambda$-invariant Artin scheme
consisting only of the point $a\in X$, and its length is either $l=1$ or $l=p$.
In the latter case, $X'$ is a trivial $G_\lambda$-torsor over $\Spec(k)$, and their is
an open neighborhood $U\subset X$ of the point $a\in X$ so that the $G_\lambda$-action is free.
\end{proposition}
\proof
For each $f\in\maxid $, we obviously have $f^p\in\ideal a$, whence $X'=\{a\}$ holds as a set.
From \eqref{action formula} we infer that $X'\subset X$ is invariant under the $G_\lambda$-action.
To compute its length, we may assume that $\ideal a=0$, by replacing $X$ with $X'$.
Then the composite map
$$
A\lra k\otimes U^{[p]}(\ideal g_\lambda)^\vee=k[t]/(t^p)
$$
obtained from \eqref{action formula} is injective, and thus $\length(A)\leq p$.
Furthermore, we may regard $A$ as a subalgebra of $B=k[t]/(t^p)$.
If $\maxid_A=0$, then $\length(A)=1$.
Now suppose that $\maxid_A\neq 0$. Fix some non-zero
$f\in\maxid_A$. Since $A\subset B$, there is some $0\leq i\leq p-1$ so that $D^i(f)$
does not vanish in $A/\maxid_A=k$. Hence $D^i(f)\in A$ is a unit.
This already ensures that $\Spec(A)\ra\Spec(A^D)$ is a torsor, for example
by \cite{Restuccia; Schneider 2003}, Theorem 4.1.
Here $A^D$ is the kernel of the derivation $D:A\ra A$, which is a $k$-subalgebra $A^D\subset A$.
It follows that $\length(A)\geq p$. In light of the inclusion $A\subset B$,
we must have $\length(A)=p$.
Now suppose that $\length(A)=p$. Then $X\times G\ra X$ is a $G$-torsor
containing a rational point, so the torsor is trivial.
To show that there is an open neighborhood on which the action is free,
we revert to the original situation $X=\Spec(A)$, with $X'=\Spec(A/\ideal a)$.
Let $A^D=\Kernel(D)$ be the kernel of the additive map $D:A\ra A$, which is a subring of $A$.
Since $\ideal a\neq\maxid$, there is some $f\in\maxid$ so that $D^i(f)$ is non-zero in $A/\maxid$,
whence a unit in the local ring $A_\primid$. Replacing $A$ by some suitable localization
$A_g$, we may assume that $D^i(f)\in A$ is a unit.
This already ensures that $\Spec(A)\ra\Spec(A^D)$ is a torsor, again
by \cite{Restuccia; Schneider 2003}, Theorem 4.1.
\qed
\medskip
We call the subring $A^D\subset A$ the \emph{ring of invariants}.
Then $X/G_\lambda=\Spec(A^D)$ is a categorical quotient in the category $(\Aff/k)$ of
all affine schemes, and in the category of all schemes $(\Sch/k)$ as well.
If the action is free,
that is, the canonical map $X\times X\ra X\times_{X/G_\lambda} X$ given by $(a,g)\mapsto (a,ag)$
is an isomorphism, and
the canonical morphism $X\ra X/G_\lambda$ is a $G_\lambda$-torsor.
We say that a rational point $a\in X=\Spec(A)$ is a \emph{fixed point} if
the orbit ideal $\ideal a\subset A$ is maximal. Otherwise, the $G_\lambda$-action is free in an open neighborhood.
Being an $A$-module, the orbit ideal has a projective dimension $\pd(\ideal a)\geq 0$, which may be finite or infinite.
From this we get the desired \emph{homological characterization} of Zariski singularities:
\begin{theorem}
\mylabel{homological characterization}
Suppose that $A$ is a complete local noetherian ring with residue field $A/\maxid_A=k$.
Then $A$ is a Zariski singularity if and only if there
is a $p$-closed derivation $D:A\ra A$ satisfying $D^p=\lambda D$ with some scalar of the form $\lambda=\mu^{p-1}$,
such that the orbit ideal $\ideal a\subset A$ is not the maximal ideal, and has finite projective dimension.
\end{theorem}
\proof
The condition is necessary: If $A$ is a Zariski singularity, say given by the
power series $g=z^p-f(x_1,\ldots,x_n)$,
we may take the standard derivation $D=D_z$, which satisfies $D^p=0$.
Clearly, $x_1,\ldots,x_n,z^p$ are contained in the orbit ideal $\ideal a\subset A$,
but not $z$, because $D(z)=1$ is nonzero in $A/\maxid_A$.
Thus the orbit ideal is not maximal, and by Proposition \ref{fixed and free points} we must have
$\ideal a=(x_1,\ldots,x_n,z^p)$.
This ideal is induced from the maximal ideal for the finite flat extension $k[[x_1,\ldots,x_n]]\subset A$, whence
has finite projective dimension.
The condition is also sufficient:
If $\lambda=\mu^{p-1}$ is non-zero, we replace $D$ by $\mu D$.
The upshot is that either $D^p=D$ or $D^p=0$.
Then the corresponding height one group scheme is $G=\mu_p$ or $G=\alpha_p$, respectively.
According to Proposition \ref{fixed and free points}, the $G$-action on $X$ is free.
Since the ring of invariants $A^D$ contains the ring of $p$-th powers $A^p$,
the projection $X\ra X/G$ is a universal homeomorphism, whence the ring of invariants $A^D$ is local.
According to the Eakin--Nagata Theorem, or equivalently by faithful flatness, it is also noetherian.
It then also must be complete. It follows from \cite{Schroeer 2007}, Proposition 2.2
that the invariant ring is also regular.
Chose some regular system of parameters $x_1,\ldots,x_n\in A^D$, and write the invariant ring as
$A^D=k[[x_1,\ldots,x_n]]$.
Now set $S=\Spec(A^D)$, and suppose that $G=\mu_p$.
From the short exact sequence $1\ra G\ra\GG_m\stackrel{F}{\lra}\GG_m\ra 1$,
we get an exact sequence
$$
H^0(S,\GG_m)\lra H^1(S,G)\lra H^1(S,\GG_m).
$$
The term on the right vanishes, because $\Pic(S)=0$. Whence the total space $X=\Spec(A)$ of the torsor
takes the form $A=A^D[z]/(z^p-f)$
for some formal power series $f(x_1,\ldots,x_n)$. Its constant term $\alpha\in k$
must be a $p$-power, because $A^D\subset A$ has trivial residue field extension.
Replacing $z$ by $ z-\beta$ with $\beta^p=\alpha$ reveals that $A$ is a Zariski singularity.
In the case $G=\alpha_p$, one uses the short exact sequence
$0\ra G\ra\GG_a\stackrel{F}{\ra}\GG_a\ra 0$ and argues similarly.
\qed
\medskip
According to Proposition \ref{theta for zariski surface}, the tangent module
of a 2-dimensional geometrically reduced Zariski singularity is free.
It is possible to give an explicit basis:
\begin{proposition}
\mylabel{basis tangent module}
Let $A=k[[x,y,z]]/(g)$ be a two-dimensional geometrically isolated Zariski singularity,
given by an formal power series $g=z^p-f(x,y)$. Then the derivations
\begin{equation}
\label{basis tangent}
D_z\quadand f_yD_x - f_xD_y
\end{equation}
form an $A$-basis for the tangent module $\Theta_{A/k}=\Der_k(A,A)=\Hom(\Omega^1_{A/k},A)$.
\end{proposition}
\proof
In our situation, the exact sequence \eqref{theta sequence} becomes
$$
0\lra\Theta_{A/k}\lra A^{\oplus 3}\stackrel{(f_x,f_y,0)}{\lra} A.
$$
Clearly, the transpose of the vectors $(0,0,1)$ and $(f_y,-f_x,0)$ lie in the kernel of the map on the right,
so the derivations in \eqref{basis tangent} indeed can be regarded as elements of the tangent module $\Theta_{A/k}$.
Moreover, $D_z\in\Theta_{A/k}$ generates a free direct summand,
and the kernel of the linear map $(f_x,f_y):A^2\ra A$ gives a complement.
By assumption, $A$ is not geometrically regular, such that $f_x,f_y\in\maxid_A$.
Regard $A$ as a finite flat algebra over $A_0=k[[x,y]]$. Since the jacobian ideal $(f_x,f_y)$
is $\maxid_{A_0}$-primary, the elements $f_x,f_y\in A_0$ form
a regular sequence. Thus the Koszul complex
$$
0\lra A_0\stackrel{\binom{f_y}{-fx}}{\lra} A_0^2 \stackrel{(f_x,f_y)}{\lra} A_0
$$
is exact. Tensoring with the flat algebra $A$ shows that the kernel of $(f_x,f_y):A^2\ra A$
is free of rank one, with generator $f_yD_x - f_xD_y$.
\qed
\begin{proposition}
\mylabel{free actions}
Let $A=k[[x,y,z]]/(g)$ be a two-dimensional geometrically isolated Zariski singularity,
given by an formal power series $g =z^p-f(x,y)$. Let
$$
D=uD_z+v(f_yD_x - f_xD_y)\in\Theta_{A/k}
$$
be a $p$-closed derivation, for some coefficients $u,v\in A$.
Then the corresponding $G$-action on $\Spec(A)$ is free if and only if $u\in A^\times$.
If this is the case, the orbit ideal is given by $\ideal a=(x,y)$, and the ring of invariants
$A^D$ is regular.
\end{proposition}
\proof
The orbit ideal $\ideal a\subset A$ obviously contains $z^p$.
Since the local ring $A$ is geometrically singular, we have $f_x,f_y\in\maxid_A$. Thus also $x,y\in\ideal a$.
If the coefficient $u\in A$ is not invertible, that is, $u\in\maxid_A$, we also have $z\in\ideal a$,
and the $G$-action on $\Spec(A)$ is not free.
Now suppose that $u\in A^\times$. Then $D(z)=u$ does not vanish in the residue field, and thus $z\not\in\ideal a$.
Consequently, the $G$-action on $\Spec(A)$ is free, and the inclusion $(x,y,z^p)\subset\ideal a$ is an equality.
In light of the relation $z^p=f(x,y)$, we get $\ideal a=(x,y)$, and this has finite projective dimension.
It follows from \cite{Schroeer 2007}, Proposition 2.2
that the invariant ring $A^D$ is regular.
\qed
\medskip
This is quite remarkable, because the ideal $\ideal a=(x,y)$ in the ring $A$ neither depends on
the choice of the regular system of parameters $x,y,z\in R$, nor the choice of the formal power series $g=z^p-f(x,y)$,
nor the choice of the $p$-closed derivation $D\in\Theta_{A/k}$ giving a free group scheme action.
We simply call it the \emph{orbit ideal}
of the two-dimensional geometrically reduced Zariski singularity. It defines a
canonical zero-dimensional subscheme of embedding dimension one on the singular scheme $X=\Spec(A)$,
called the \emph{orbit subscheme}.
In turn, it yields a unique tangent vector
\begin{equation}
\label{orbit direction}
\Spec k[\epsilon]\subset \Spec k[[x,y,z]]/(x,y,z^p)\subset\Spec(A)
\end{equation}
that indicates the direction of any $p$-closed vector field corresponding to a free action.
We call it the
\emph{orbit direction}. Here and throughout, $\epsilon $ denotes
an indeterminate satisfying $\epsilon^2=0$.
\begin{theorem}
\mylabel{canonical line}
Let $A$ be a two-dimensional geometrically reduced Zariski singularity that is not regular,
and $\ideal g\subset\Theta_{A/k}$ be a restricted Lie subalgebra with $\dim_k(\ideal g)=2$
for which every element is $p$-closed,
and $A\ideal g=\Theta_{A/k}$. Then there is a line $\ideal l\subset \ideal g$ so that
for each element $D\in\ideal g$, the corresponding group scheme
action on $\Spec(A)$ is free if and only if $D\not\in\ideal l$.
If these conditions holds, the invariant ring $A^D$ is regular.
\end{theorem}
\proof
Suppose $A$ is given by the formal power series $g=z^p-f(x,y)$. Choose a $k$-basis $D_1,D_2\in \ideal g$.
Then $D_1,D_2\in \Theta_{A/k}$ form an $A$-basis as well.
According to Proposition \ref{free actions}, the $D_z,f_yD_x - f_xD_y\in\Theta_{A/k}$ yield another
$A$-basis. Write
$$
D_z= r_{11}D_1 + r_{12}D_2\quadand f_yD_x - f_xD_y= r_{21}D_1 + r_{22}D_2
$$
for some base-change matrix $(r_{ij})\in\GL_2(A)$.
Let $\bar{r}_{ij}\in A/\maxid_A=k$ be the residue classes of the matrix entries,
and let $\ideal l\subset\ideal g$ be the line generated by the vector $\bar{r}_{21}D_1 + \bar{r}_{22}D_2$.
Each derivations $D=\lambda D_1+\mu D_2\in\ideal g$ with coefficients $\lambda,\mu\in k$ can be written as
$D=uD_z+v(f_yD_x - f_xD_y)$ with coefficients $u,v\in A$.
Obviously, $D\not\in\ideal l$ means that $u\in A^\times$.
The assertion thus follows from Proposition \ref{free actions}.
\qed
\medskip
In the above situation, we call $\ideal l\subset\ideal g$ the
\emph{canonical line}.
This also can be regarded in the following way:
The composite map $\ideal g \subset\Theta_{A/k}\ra \Theta_{A/k}\otimes k$ is bijective.
Under the canonical map
$$
\Theta_{A/k}\otimes k=\Hom_A(\Omega^1_{A/k},A)\otimes_Ak \lra \Hom_k(\Omega^1_{A/k}\otimes k,k) = \Hom_A(\Omega^1_{A/k},k),
$$
the elements $D\in\ideal g$ are turned into derivations $D:A\ra k$,
which correspond to homomorphisms $A\ra k[\epsilon]$.
Note, however, that the vector spaces above have dimension two and three,
and that the canonical line $\ideal l\subset\ideal g$ is the kernel of the above linear map.
Hence the image is one-dimensional, and indeed induces the orbit direction
$\Spec k[\epsilon] \subset\Spec(A)$.
In Theorem \ref{canonical line}, the 2-dimensional restricted Lie algebra $\ideal g$
has the property that every vector is $p$-closed.
Up to isomorphism, there are exactly two such restricted Lie algebras of dimension two:
\begin{lemma}
\mylabel{classification lie algebras}
Suppose $k$ is algebraically closed.
Then the 2-dimensional restricted Lie algebras $\ideal g$ for which every vector is $p$-closed
are up to isomorphism given by $\ideal g=kx\oplus ky$ with
$$
[x,y] = x^{[p]}=y^{[p]}=0 \qquad\text{or}\qquad [x,y]=y,\, x^{[p]}=x,\, y^{[p]}=0.
$$
\end{lemma}
\proof
This follows from the classification of 2-dimensional restricted Lie algebras \cite{Wang 2013}, Proposition A3.
\qed
\section{Rational double points}
\mylabel{Rational double points}
Let $k$ be an algebraically closed ground field of characteristic $p>0$, and $A$ be a complete local ring
that is normal and of dimension two. We say that $A$ is a \emph{rational double point}
if it is a rational singularity of Hilbert--Samuel multiplicity $e(A)=2$.
Equivalently, it is a rational Gorenstein singularity.
Recall that in characteristic $p=2$, the \emph{rational double points} are classified into
$$
A_n,\quad D_{2n}^r, D_{2n+1}^r,\quad E_6^0,E_6^1,\quad E_7^0,\ldots,E_7^3,\quadand E_8^0,\ldots, E_8^4,
$$
with $0\leq r\leq n-1$, according to Artin's analysis \cite{Artin 1977}.
We now can determined which of them are Zariski:
\begin{theorem}
\mylabel{zariski rdp}
In characteristic $p=2$, the rational surface singularities that are Zariski singularities are
precisely the rational double points of type $A_1$, $D_{2m}^0$, $E_7^0$ and $E_8^0$.
\end{theorem}
\proof
Being a hypersurface singularity, any Zariski singularity is Gorenstein.
Thus a surface singularity that is both rational and Zariski must be a rational double point.
The rational double points of type $A_1$, $D_{2m}^0$, $E_7^0$ and $E_8^0$
are Zariski is obvious from Artin's normal forms for rational double points in characteristic two
\cite{Artin 1977}.
The converse is not completely obvious. Let $A=k[[x,y,z]]/(g)$ be a rational double point that
is a Zariski singularity, given by the formal power series $g=z^2-f(x,y)$.
According to Proposition \ref{intersection matrix for surface zariski},
the intersection matrix $\Phi=(E_i\cdot E_j)$ for the exceptional divisor
on a resolution of singularities has $\det(\Phi)=\pm 2^\nu$ for some $\nu\geq 0$,
whence only type $A_n$ with $n =2^\nu-1$, $D_n$, $E_7$ and $E_8$ are possible.
Clearly, the module of K\"ahler differentials $\Omega^1_{A/k}$ is generated by $dx,dy,dz$,
and $Adz\subset\Omega_{A/k}$ is an invertible direct summand.
According to \cite{Schroeer 2008}, proof of Theorem 6.3 such an invertible direct summand
does not exist for $D_n^r$ with $r>0$. Likewise, in the proof for loc.\ cit.\ Theorem 6.4
the same is shown for the singularities $E_7^r$ and $E_8^r$ with $r>0$.
It remains to argue that neither $A_n$ with $n=2^\nu-1\geq 2$, nor $D_n^0$ with $n=2m+1$ are Zariski.
In the former case, the singularity is given by the polynomial $g=z^{2m}+xy$ for some $m\geq 2$.
Using Proposition \ref{theta and pd}, one sees that the tangent module $\Theta_{A/k}$ is free.
As in Proposition \ref{basis tangent module}, the derivations $D_z,xD_x+yD_y$ form a basis.
Given a linear combination $D=uD_z+v( xD_x+yD_y)$,
it transforms the generators as
$$
x\longmapsto vx,\quad
y\longmapsto vy\quadand
z\longmapsto u
$$
whence $x,y,z^2\in R$ lie in the kernel for the composition $A\stackrel{D}{\ra} A\ra A/\maxid_A$.
Now suppose that $A_n$ is Zariski. In light of Theorem \ref{homological characterization}, we may
choose $D$ so that $D^2=0$ and that
the orbit of the closed point on $\Spec(A)$ for the corresponding action of $G=\alpha_2$
is defined by an ideal $\ideal a\subset A$ of colength two so that $\bar{A}=A/\ideal a$ has finite
projective dimension. It follows that $\Tor_3(\bar{A},M)=0$ for any $A$-module $M$.
The ideal $\ideal a$ contains $(x,y,z^2)\subset R$. Since the latter already has colength two, we must have
$\ideal a=(x,y,z^2)$.
The matrix
$$
\varphi=
\begin{pmatrix}
z^m & x\\
y & z^m
\end{pmatrix}
\in\Mat_2(k[[x,y,z]])
$$
yields a matrix factorization $\varphi\circ \varphi=(\begin{smallmatrix}g&0\\0&g\end{smallmatrix})$,
and thus gives a periodic free resolution
$$
...\stackrel{\varphi}{\lra} A^2 \stackrel{\varphi}{\lra} A^2 \stackrel{\varphi}{\lra} A^2 \lra M\lra 0
$$
for $M=\Cokernel(\varphi)$, according to \cite{Eisenbud 1980}, Corollary 6.3.
Use this free resolution to compute Tor-groups: Since $m\geq 2$, the matrix $G\otimes_A\bar{A}$
is the zero matrix, and thus $\Tor_i(\bar{A},M)=\bar{A}^2\neq 0$ for all $i\geq 1$, contradiction.
It remains to threat the case $D_n^0$ with $n=2m+1$ odd. Here the defining polynomial is $g=z^2+x^2y+y^mz$.
Again it turns out that $\Theta_{A/k}$ is free, and the derivations
$$
D_x\quadand y^mD_y + (x^2+my^{m-1}z)D_z
$$
form a basis. For each linear combination $D$, the kernel $\ideal a$ for the composite map
$A\stackrel{D}{\ra}A\ra A/\maxid_A$ contains $(x^2,y,z)$. Now suppose that $D_n^0$ is Zariski.
Then we may choose $D$ with $D^2=0$ so that $\ideal a=(x^2,y,z)$.
Using the matrix factorization
$\varphi\circ \varphi=(\begin{smallmatrix}g&0\\0&g\end{smallmatrix})$
with
$$
\varphi=
\begin{pmatrix}
z & y\\
x^2+y^{m-1}z & z
\end{pmatrix}
\in\Mat_2(k[[x,y,z]]),
$$
we get a contradiction as above.
\qed
\medskip
With analogous arguments, one shows:
\begin{proposition}
\mylabel{zariski rdp other p}
In characteristic $p=3$, the rational surface singularities that are Zariski singularities
are the rational double points of type $A_2, E_6^0,E_8^0$.
In characteristic $p=5$, these are $A_4$ and $E_8^0$.
For $p\geq 7$, there is only $A_{p-1}$.
\end{proposition}
A rational double point that is also a Zariski singularity is called a \emph{Zariski rational double point}.
Going through their normal form \cite{Artin 1977}, one obtains the following observation:
\begin{proposition}
\mylabel{tjurina rdp}
The Tjurina numbers for Zariski rational double points
of type $A_n$, $D_n$ or $E_n$ in characteristic $p>0$ are $\tau=pn/(p-1)$.
In particular, $\tau=2n$ in characteristic two.
\end{proposition}
\section{Canonical coverings of algebraic schemes}
\mylabel{Canonical coverings}
The goal of this section is to establish some general facts about ``canonical coverings'' of proper algebraic schemes.
We will apply this later to the K3-like covering of simply-connected Enriques surface, but the
underlying principles have a general relevance.
Fix a ground field $k$, and let $Y$ be a proper $k$-scheme
with $k=H^0(Y,\O_Y)$, and $\Pic_{Y/k}$ be its Picard scheme.
Let $G$ be a finite commutative group scheme, and $\hat{G}=\ul \Hom(G,\GG_m)$ be the Cartier dual.
This is also a finite commutative group scheme, of the same order $h^0(\O_G)=h^0(\O_{\hat{G}})$,
and the biduality map
$$
G\lra\ul\Hom(\ul\Hom(G,\GG_m),\GG_m),\quad g\longmapsto (f\mapsto f(g))
$$
is an isomorphism of group schemes.
Set $S=\Spec(k)$, let $f:Y\ra S$ be the structure morphism, and write $G_Y=f^*(G)$ for the induced
relative group scheme over $Y$.
According to \cite{Raynaud 1970}, Proposition 6.2.1 applied to the Cartier dual $\hat{G}$, we have a a canonical bijection
$$
R^1f_*(G_Y) \lra \underline{\Hom}(\hat{G},\Pic_{Y/k}) = \underline{\Hom}(\hat{G},\Pic^\tau_{Y/k})
$$
of abelian sheaves on the site $(\text{Sch}/k)$ of all $k$-schemes, endowed with the fppf-topology.
Taking global sections yields an identification
$$
H^0(S, R^1f_*(G_Y)) = \Hom(\hat{G},\Pic^\tau_{Y/k}).
$$
The term on the left sits in the exact sequence coming from the Leray--Serre spectral sequence
$$
0\lra H^1(S, f_*(G_Y) ) \lra H^1(Y, G_Y) \lra H^0(S, R^1f_*(G)) \lra H^2(S, f_*(G_Y)).
$$
If the ground field $k$ is algebraically closed, every non-empty $k$-scheme of finite type has a rational point, by
Hilbert's Nullstellensatz, and it follows that $H^r(S,F)=0$ for every degree $r\geq 1$ and every abelian fppf-sheaf $F$.
Thus the outer terms in the preceding sequence vanish. This yields:
\begin{proposition}
\mylabel{raynaud correspondence}
If $k$ is algebraically closed, then the canonical map
$$
H^1(Y, G_Y) \lra \Hom(\hat{G},\Pic^\tau_{Y/k})
$$
is bijective. In other words, the isomorphism classes of $G$-torsors $\epsilon:X\ra Y$ correspond to homomorphisms of
group scheme $\hat{G}\ra\Pic_{Y/k}$.
\end{proposition}
Often there are \emph{canonical choices} for $G$ with respect to $Y$ and the resulting numerical
trivial part $P=\Pic^\tau_{Y/k}$.
For example, if $Y$ is Gorenstein with numerically trivial dualizing sheaf $\omega_Y$,
we may take for $\hat{G}\subset P$ the discrete subgroup generated by the class of $\omega_Y$.
One may call the resulting $\epsilon:X\ra Y$ the \emph{$\omega$-canonical covering}.
If the Picard scheme is zero-dimensional, then one may choose $G$ with $\hat{G}=P$
or $\hat{G}=P^0$,
and the resulting torsors are called the \emph{$\tau$-canonical covering} and \emph{$0$-canonical covering},
respectively.
For every $n\neq 0$ that is prime to the characteristic exponent $p\geq 1$,
we furthermore get the \emph{$n$-canonical coverings} with $\hat{G}=P[n]$,
the kernel of the multiplication by $n$-map.
In characteristic $p>0$, we may also use the kernel of the relative Frobenius map $F:P\ra P^{(p)}$,
and get the \emph{$F$-canonical covering}.
Later, we will apply this to simply-connected Enriques surfaces $Y$. Then we have
$\Pic^0_{Y/k}=\Pic^\tau_{Y/k}=\Pic_{Y/k}[F]$, and the ensuing torsor $\epsilon:X\ra Y$ is called
the \emph{K3-like covering}.
In what follows, we will assume that the ground field $k$ is algebraically closed,
such that we have $H^1(Y, G_Y)=\Hom(\hat{G},\Pic^\tau_{Y/k})$.
This identification is natural in $Y$ and $G$, and the following obvious consequence will play a key role throughout:
\begin{lemma}
\mylabel{pullback trivial}
Let $\epsilon:X\ra Y$ be a $G$-torsor, corresponding to a homomorphism $\hat{G}\ra\Pic_{Y/k}$.
Let $Z\ra Y$ be a proper morphism.
If $h^0(\O_Z)=1$, and the composite homomorphism $\hat{G}\ra\Pic_{X/k}\ra\Pic_{Z/k}$ is zero,
then the induced torsor $X\times_Y Z\ra Z$ is trivial.
\end{lemma}
\proof
The condition $h^0(\O_Z)=1$ ensures that $H^1(Z,G)=\Hom(\hat{G},\Pic_{Z/k})$,
and the naturality of this identification implies that the induced torsor becomes trivial.
\qed
\medskip
As a consequence, we get a statement on the singularities of the total space of torsors:
\begin{proposition}
\mylabel{singularity total space}
Let $\epsilon:X\ra Y$ be be a $G$-torsor with $G$ non-reduced, $Z\subset Y$ be a Weil divisor with
$h^0(\O_Z)=1$, and assume that the composite map $\hat{G}\ra\Pic^\tau_{Y/k}\ra\Pic_{Z/k}$ is zero.
Let $x\in X$ be a closed point mapping to a point $z\in Z$ where the local ring $\O_{Z,z}$ is singular.
Then $\O_{X,x}$ is singular as well.
\end{proposition}
\proof
Set $n=\dim(\O_{X,x})$, such that the local ring $\O_{Z,z}$ on the Weil divisor has dimension $n-1$.
Since this local ring is not regular, its embedding dimension is $\edim(\O_{Z,z})\geq n$.
The induced torsor is $T\simeq Z\times G$. The local ring at the origin $0\in G$
has embedding dimension $\edim(\O_{G,0})\geq 1$, because the group scheme $G$ is non-reduced. Thus we have
$$
\edim(\O_{X,x})\geq \edim(\O_{T,x})=\edim(\O_{Z,z})+\edim(\O_{G,0})\geq n+1,
$$
hence the local ring $\O_{X,x}$ is not regular.
\qed
\medskip
Sometimes, one may verify the assumption by looking at cohomology groups:
\begin{corollary}
\mylabel{singularity via cohomology}
Assumptions as in the previous proposition. Suppose further $\hat{G}$ is local,
that $Y$ is normal of dimension $n\geq 1$, and that the canonical mapping
$H^{n-1}(Y,\omega_Y)\ra H^{n-1}(Y,\omega_Y(Z))$ is injective.
Let $x\in X$ be a closed point mapping to a point $z\in Z$ where $\O_{Z,z}$ is singular.
Then $\O_{X,x}$ is singular as well.
\end{corollary}
\proof
The short exact sequence $0\ra\O_Y(-Z) \ra\O_Y\ra\O_Z\ra 0$ gives a long exact sequence
$$
H^1(\O_X(-Z))\lra H^1(X,\O_X) \lra H^1(Z,\O_Z).
$$
The map on the left hand side is surjective, because it is Serre dual to
the injection $H^{n-1}(Y,\omega_Y)\ra H^{n-1}(Y,\omega_Y(Z))$.
Whence the map on the right is zero. The latter is the tangent map
for the restriction $\Pic_{Y/k}\ra\Pic_{Z/k}$.
Suppose for a moment that $\hat{G}$ has height one. Then this group scheme is
determined by its restricted Lie algebra
$\Lie(\hat{G})$,
the composite map $\hat{G}\ra\Pic_{Y/k}\ra\Pic_{Z/k}$ is trivial, and the assertion follows from the
proposition. In the general case, we may filter $\hat{G}$ with subquotients of height one,
and then check by induction on the length of the filtration that $\hat{G}\ra\Pic_{Z/k}$ is zero.
\qed
\medskip
The following connection to singularity theory follows immediately
from Theorem \ref{homological characterization}, our homological characterization of Zariski singularities:
\begin{proposition}
\mylabel{canonical covering zariski}
Suppose that $Y$ is smooth, and that $\hat{G}$ is unipotent of order $p$,
in other words either $\hat{G}=\ZZ/p\ZZ$ or $\hat{G}=\alpha_p$. Then for each closed point
$x\in X$, the complete local ring $\O_{X,x}^\wedge$ is a Zariski singularity.
\end{proposition}
Now suppose that $Y$ is a smooth surface.
A curve $E\subset Y$ is called \emph{negative-definite} if the intersection
matrix $\Phi=(E_i\cdot E_j)$ attached to the integral components $E_1,\ldots,E_r\subset E$
is negative-definite.
This are precisely the \emph{exceptional divisors} for some contraction $Y\ra Y'$,
where $Y'$ is a normal 2-dimensional algebraic space.
To each such negative-definite curve we have the \emph{fundamental cycle} $Z=\sum n_iE_i$
with certain coefficients $n_i\geq 1$. It can be defined as the smallest such cycle
for which all intersection numbers are $Z\cdot E_i\leq 0$, and we have Artin's Algorithm
\cite{Artin 1966}
to compute it from the intersection matrix $\Phi$.
We call the singular locus $\Sing(Z)$ of the fundamental cycle
the \emph{fundamental-singular locus} of $E\subset Y$. Note that
$$
\Sing(Z) = \bigcup_{i\neq j} (E_i\cap E_j)\quad \cup \quad\bigcup_{n_i>1} E_i.
$$
This coincides with the singular locus of the reduced curve $E$ if the
fundamental cycle is reduced, but usually is much larger.
We define the \emph{fundamental-genus} of the negative-definite curve $E\subset Y$
as the natural number $h^1(\O_Z)\geq 0$.
A negative-definite curve $E\subset Y$ is called of \emph{rational type} if
it has fundamental-genus $h^1(\O_Z)=0$. In other words, it contracts
to a rational singularity. In the special case of rational double points,
the intersection matrix $\Phi$ then
corresponds to one of the Dynkin diagrams $A_n$, $D_n$, $E_6$, $E_7$ or $E_8$,
and we call the curve an \emph{ADE-configuration}.
\begin{proposition}
\mylabel{singularity over fundamental-singular}
Let $Y$ be a smooth surface, and $\epsilon:X\ra Y$ be a $G$-torsor for some non-reduced $G$.
Then the local ring $\O_{X,x}$ is singular for each point $x\in X$ mapping
to the fundamental-singular locus $Z\subset X$ of negative-definite curve of rational type $A\subset Y$.
\end{proposition}
\proof
Since $G$ is non-reduced, there is a subgroup scheme $H\subset G$ so that
$G'=G/H$ is either $\mu_p$ or $\alpha_p$.
Set $X'=X/H$. By construction, $X\ra X'$ is a $H$-torsor, and
$X'\ra Y$ is a $G'$-torsor, in fact coming from the
composite $\hat{G'}\subset \hat{G}\ra\Pic_{Y/k}$.
According to \cite{EGA IVc}, Proposition 6.7.4 it thus suffices to treat the case $G=G'$.
Then the Cartier dual $\hat{G}$ is isomorphic to $\ZZ/p\ZZ$ or $\alpha_p$, respectively. The Picard scheme
of the fundamental cycle is of the form $\Pic_{Z/k}=\ZZ^{\oplus r}$, where $r\geq 1$ is the number
of irreducible components of $Z$. In turn, the homomorphism $\hat{G}\ra\Pic_{Z/k}$ must be trivial,
and the assertion follows from Proposition \ref{singularity total space}.
\qed
\medskip
We say that a curve $A$ is \emph{seminormal}
at some closed point $a\in A$ if
the irreducible components in $\Spec(\O_{A,a}^\wedge)$ are normal and meet
like coordinate axis in $\AA^n$.
For example, this holds if $\O_{A,a}^\wedge$ is isomorphic to $k[[x,y]]/(xy)$,
that is, if the curve has \emph{normal crossings} at $a\in A$.
\begin{proposition}
\mylabel{singularity over nc}
Let $Y$ be a normal surface, and $\epsilon:X\ra Y$ be a $G$-torsor for some non-reduced $G$.
Let $A\subset Y$ be a curve whose normalization is a disjoint union of copies of $\PP^1$.
Then the local ring $\O_{X,x}$ is singular for each point $x\in X$
that maps to a point in $a\in A$ where $A$ is singular but semi-normal.
\end{proposition}
\proof
Again it suffices to treat the case that $G$ is either $\mu_p$ or $\alpha_p$.
Clearly, we may assume that each irreducible component of $A$ contains the point $a\in A$.
Let $A'\ra A$ be the morphism that is the normalization outside $a\in A$, and an isomorphism
over an open neighborhood of $a\in A$. We have $h^0(\O_{A'})=1$, because the curve $A'$ is connected
and reduced. Moreover, its Picard scheme then sits in a short exact sequence
$$
0\lra\GG_m^{\oplus r}\lra\Pic_{A'/k}\lra\ZZ^{\oplus s}\lra 0
$$
for some integers $r,s\geq 0$. Since $\hat{G}$ is isomorphic to $\ZZ/p\ZZ$ or $\alpha_p$,
there is only the zero homomorphism $\hat{G}\ra\Pic_{A'/k}$, whence the torsor $X\times_YA'\ra A'$
is trivial. It follows that the torsor $X\times_YA\ra A$ is trivial on some open neighborhood of $a\in A$.
By Proposition \ref{singularity total space}, the local ring $\O_{X,x}$ is singular.
\qed
\begin{proposition}
\mylabel{singularity over line}
Suppose $Y$ is a normal Gorenstein surface, and $\epsilon:X\ra Y$ is a $G$-torsor for some non-reduced $G$.
Let $A\subset Y$ be a curve isomorphic to $\PP^1$. If the invertible sheaf $\omega_X$ is nef,
then the preimage $\epsilon^{-1}(A)$ passes through some point $x\in X$ where the local ring
$\O_{X,x}$ is not factorial, in particular not regular.
\end{proposition}
\proof
It suffices to treat the case that $G$ is either $\mu_p$ or $\alpha_p$. Since $\Pic_{A/k}$ is torsion free and reduced,
the torsor is trivial over $A$, whence the preimage $\epsilon^{-1}(A)$ is isomorphic to $A\times G$.
In turn, $C=\epsilon^{-1}(A)_\red$ is isomorphic to $\PP^1$.
Seeking a contradiction, we assume that this preimage lies in the locus where $X$ is locally factorial.
Then $C\subset X$ is a Cartier divisor with $C^2=0$, because its conormal sheaf $\shI/\shI^2$ is trivial.
The Adjunction Formula gives $-2=K_X\cdot C= K_X\cdot C + C^2$, contradicting $K_X\cdot C\geq 0$.
\qed
\medskip
We finally turn to some global invariants.
Let us formulate the following fact:
\begin{proposition}
\mylabel{properties covering}
Suppose that $Y$ is normal and proper, and that $P=\Pic^\tau_{Y/k}$ is a finite unipotent
group scheme of order $p$. Let $\epsilon:X\ra Y$ be the $G$-torsor with $\hat{G}=P$.
Then
$$
h^0(\O_X)=1,\quad \chi(\O_X)=p\chi(\O_Y)\quadand \omega_X=\epsilon^*(\omega_Y).
$$
\end{proposition}
\proof
The assumption on $P=\Pic^\tau_{Y/k}$ means that the group scheme $G$ of order $p$ is local.
In turn, $\epsilon:X\ra Y$ is a universal homeomorphism. Let $\eta\in Y$ be the generic point.
Then $\O_{X,\eta}$ is an Artin local ring, and has degree $p$ as an algebra over the function field
$k(Y)=\O_{Y,\eta}$. Moreover, the field extension $k(Y)\subset k(X)$ is purely inseparable,
hence its degree $d\geq 1$ is either $d=1$ or $d=p$.
Seeking a contradiction, we assume $h^0(\O_X)>1$. Since the ground field $k$ is algebraically closed, the
algebra $H^0(X,\O_X)$ and whence the scheme $X$ is non-reduced.
Since $X$ contains no embedded associated points, the Artin local ring $\O_{X,\eta}$ is non-reduced.
We infer that $d=1$, such that $X_\red\ra Y$ is birational.
By Zariski's Main Theorem, the finite birational $X_\red\ra Y$ is an isomorphism, hence
the $G$-torsor $\epsilon:X\ra Y$ admits a section, contradiction.
The statement on the Euler characteristic is a special case of \cite{Mumford 1970},
Theorem 2 on page 121.
It remains to establish the assertion on the dualizing sheaves.
The relative dualizing sheaf for the finite dominant morphism $\epsilon:X\ra Y$
is given by $\omega_{X/Y}=\underline{\Hom}_{\O_Y}(\O_X,\O_Y)$.
In light of $\omega_X=\epsilon^*(\omega_Y)\otimes\omega_{X/Y}$, it suffices
to show that $\omega_{X/Y}\simeq\O_X$.
Since $\epsilon:X\ra Y$ is a $G$-torsor, the finite flat $\O_Y$-algebra $\O_X$ of degree $p$ locally admits
a $p$-basis consisting of a single element.
It follows that the sheaf of K\"ahler differentials $\Omega^1_{X/Y}$ is invertible.
According to \cite{Kiehl; Kunz 1965}, Satz 9 we have $\omega_{X/Y}=(\Omega^1_{X/Y})^{\otimes (1-p)}$, so it suffices
to check that $\Omega^1_{X/Y}\simeq\O_X$.
Recall from \cite{EGA IVd}, Section 16.3 that the sheaf of K\"ahler differentials can be defined by the exact sequence
$$
0\lra\Omega^1_{X/Y} \lra\O_{X^{(1)}}\lra\O_X\lra 0,
$$
where $X=X\times_XX$ is the diagonal inside $X\times_YX$, and $X^{(1)}$ is its first infinitesimal neighborhood.
Since the morphism $\can:X\times G\lra X\times_YX$ given by $(x,g)\mapsto (x,xg)$ is an isomorphism,
we get a commutative diagram
$$
\begin{CD}
X @>>> X\times G\\
@V\Delta VV @VV\can V\\
X\times_XX @>>> X\times_YX,
\end{CD}
$$
where the upper map is given by $x\mapsto (x,e)$. Using that the underlying scheme of $G$
is isomorphic to the spectrum of $k[t]/(t^p)$, we infer
that the conormal sheaf of $X\subset X\times_YX$ is isomorphism to $\O_X$,
in other words $\Omega^1_{X/Y}\simeq\O_X$.
\qed
\begin{corollary}
\mylabel{invariants covering}
Assumptions as in the proposition.
Suppose furthermore that $Y$ is Gorenstein, of even dimension $n\geq 2$, with $\omega_Y$ numerically trivial,
and such that
$$
\sum_{i=3}^n (-1)^i h^i(\O_Y)\geq 0\quadand \sum_{j=2}^{n-1}(-1)^jh^j(\O_X)\leq 0.
$$
Then the above inequalities are equalities, and we have $p=2$, $\omega_X=\O_X$, $h^1(\O_X)=0$ and $h^0(\O_X)=h^n(\O_X)=1$.
\end{corollary}
\proof
The invertible sheaf $\omega_X=\epsilon^*(\omega_Y)$ is numerically trivial, so Serre Duality gives $h^n(\O_X)\leq 1$.
If the Picard scheme is non-smooth, we have $\Pic^\tau_{Y/k}=\alpha_p$, thus $h^1(\O_Y)=1$ and
the obstruction group $H^2(\O_Y)$ must be non-zero, according to \cite{Mumford 1966},
Corollary on page 198.
If the Picard scheme is smooth, we have $\Pic^\tau_{Y/k}=\ZZ/p\ZZ$, consequently $h^1(\O_Y)=0$.
In both case, the inequality $1-h^1(\O_Y)+h^2(\O_Y)\geq 1$ holds. This gives a chain
of inequalities
$$
p\leq p(1-h^1(\O_Y)+h^2(\O_Y))\leq p\chi(\O_Y) =\chi(\O_X) \leq 1-h^1(\O_X) + h^n(\O_X) \leq 2-h^1(\O_X).
$$
The only possibility is $p=2$, $h^1(\O_X)=0$ and $h^n(\O_X)=1$. The latter shows that $\omega_X$ has
a non-zero global section $s:\O_X\ra\omega_X$. This map is necessarily bijective, because $\omega_X$ is numerically trivial.
\qed
\section{Simply-connected Enriques surfaces and K3-like coverings}
\mylabel{Simply-connected enriques}
Fix an algebraically closed ground field $k$ of characteristic $p=2$,
and let $Y$ be an \emph{Enriques surface}. By definition, this is a smooth proper connected surface
with
$$
c_1(Y)=0\quadand b_2(Y)=10.
$$
This implies that $Y$ is minimal, and that the dualizing sheaf $\omega_Y$ is numerically trivial.
They form one of the four classes of surfaces with $c_1=0$, the
other being the abelian, bielliptic and K3-surfaces, which respective Betti number
$b_2=6$, $b_2=2$ and $b_2=22$, according to the classification of surfaces.
For Enriques surfaces, the numerically trivial part $P=\Pic^\tau_{Y/k}$ is a group scheme of order two, and its group
of rational point $P(k)$ is generated by $\omega_Y$.
The Enriques surfaces come in three types: \emph{Ordinary, classical and supersingular},
which means that $P$ is isomorphic to the respective group schemes
$\mu_2$, $\ZZ/2\ZZ$ and $\alpha_2$. Note that there are several other
designations in the literature.
Let $G=\ul\Hom(P,\GG_m)$ be the Cartier dual, such that $\hat{G}=P$,
and denote by $\epsilon:X\ra Y$ the resulting $G$-torsor, as discussed in Section \ref{Canonical coverings}.
If $P=\mu_2$ is diagonalizable, this is an \'etale covering,
and $X$ is a K3 surface endowed with an free involution, which can also be viewed
as the universal covering of $Y$.
If $P=\ZZ/2\ZZ,\alpha_2$ is unipotent, then the Cartier dual $G=\mu_2,\alpha_2$
is local. In this case we say that $Y$ is a \emph{simply-connected Enriques surface},
and the $G$-torsor $\epsilon:X\ra Y$, which is a universal homeomorphism, is called the \emph{K3-like covering}.
Let us say that a connected reduced surface $X$ is a \emph{K3-like covering} if it is isomorphic
to the K3-like covering of some Enriques surface $Y$.
We record:
\begin{proposition}
\mylabel{enriques quotient}
Suppose that a simply-connected reduced surface $X$ with numerically trivial $\omega_X$
admits a free action of a local group scheme $G$
of order two whose quotient $Y=X/G$ is smooth. Then $Y$ is an Enriques surface,
the quotient map $\epsilon:X\ra Y$ is the K3-like covering, and $\Pic^\tau_{Y/k}\simeq\hat{G}$.
\end{proposition}
\proof
First note that since $X$ is reduced, the $G$-torsor $\epsilon:X\ra Y$ is non-trivial,
hence the corresponding homomorphism $\hat{G}\ra \Pic^\tau_{Y/k}$ is non-zero.
This already ensures that $Y$ is not a K3 surface.
The scheme $X$ is Cohen--Macaulay and Gorenstein, because this holds for $Y$,
and $\epsilon:X\ra Y$ is a $G$-torsor. In particular, $\omega_X$ is invertible.
The proof for Proposition \ref{properties covering} shows that $\epsilon^*(\omega_Y)=\omega_X$,
so $\omega_Y$ is numerically trivial.
By the classification of surface, $Y$ is either Enriques, abelian or bielliptic.
Since $G$ is local, the map $\epsilon:X\ra Y$ is
a universal homeomorphism, consequently $Y$ is simply-connected.
Therefore, $Y$ is neither abelian nor bielliptic, and the only remaining possibility
is that $Y$ is an Enriques surface.
Thus $P=\Pic^\tau_{Y/k}$ has order two. It follows that the non-zero homomorphism $\hat{G}\ra P$
is an isomorphism, and $\epsilon:X\ra Y$ must be the K3-like covering.
\qed
\medskip
From now on, we assume that $Y$ is a simply-connected Enriques surface.
We then have the following well-known facts on the K3-like covering $X$:
\begin{proposition}
\mylabel{facts on X}
The scheme $X$ is integral, the complete local rings $\O_{X,x}^\wedge$ at the closed points $x\in X$
are reduced Zariski singularities, and we have
$$
\omega_X=\O_X,\quad h^0(\O_X)=h^2(\O_X) =1\quadand h^1(\O_X)=0.
$$
Moreover, the scheme $X$ is not smooth.
\end{proposition}
\proof
The first statement follows from Proposition \ref{canonical covering zariski},
the second from Corollary \ref{invariants covering}.
Suppose $X$ would be smooth. By the classification of surfaces, $X$ is either an abelian surface
or a K3 surface, thus the second Betti number is either $b_2=22$ or $b_2=6$.
However, by the topological invariance of Betti numbers, we have
$b_2(X)=b_2(Y)=10$, contradiction.
\qed
\medskip
An integral curve $A\subset Y$ with $A^2=-2$ is called
a \emph{$(-2)$-curve}. By the Adjunction Formula, this are the curves on the Enriques surface isomorphic to $\PP^1$.
An integral curve $C\subset Y$ with $C^2=0$ and $\Pic^\tau_{C/k}=\GG_a$
is called a \emph{rational cuspidal curve}. This are the curves on $Y$ that are isomorphic
to
$$
\Spec(k[t^2,t^3])\cup\Spec(k[t^{-1}]).
$$
An integral curve $F\subset Y$ with $F^2=0$ and $\Pic^\tau_{F/k}=\GG_m$ is called
a \emph{rational nodal curve}. This are the curves on $Y$ isomorphic to
$$
\Spec(k[s,t]/(st))\cup\Spec(k[s^{-1},t^{-1}]/([s^{-1}-t^{-1})).
$$
\begin{proposition}
\mylabel{curves images singularities}
Each $(-2)$-curve $A\subset Y$, each rational cuspidal curve $C\subset Y$
and each rational nodal curve $F\subset Y$
passes through the image of $\Sing(X)$.
Each negative-definite curve $E\subset Y$ is an ADE-configuration,
and its fundamental-singular locus is contained in the image of $\Sing(X)$.
\end{proposition}
\proof
To see that $E\subset Y$ is an ADE-configuration, let $Z$ be its fundamental cycle,
and consider the exact sequence
$$
H^1(Y,\O_Y)\lra H^1(Z,\O_Z)\lra H^2(Y,\O_Y(-Z))\lra H^2(Y,\O_Y).
$$
The map on the right is Serre dual to $H^0(Y,\omega_Y)\ra H^0(Y,\omega_Y(Z))$.
Any non-zero global section of $\omega_Y(Z)$ would define a curve $Z'\subset Y$ numerically
equivalent to $Z\subset Y$. Since the latter is negative definite, the only possibility is $Z'=Z$,
whence $\omega_Y\simeq\O_Y$. In this case, we conclude that the inclusion $H^0(Y,\omega_Y)\subset H^0(Y,\omega_Y(Z))$
is bijective, again because $Z\subset Y$ is negative-definite.
The upshot is that in the above exact sequence, the map on the right is injective,
whence the map on the left is surjective.
If $Y$ is ordinary then $h^1(\O_Y)=0$, hence $h^1(\O_Z)=0$ and $E\subset Y$ is an ADE-configuration.
If $Y$ is supersingular, we have $h^1(\O_Y)=1$ and $\Pic^\tau_{Y/k}=\alpha_2$.
Seeking a contradiction, we assume that the surjection $H^1(Y,\O_Y)\ra H^1(Z,\O_Z)$ is non-zero, hence
bijective. Using that curves have smooth Picard groups, we infer that $\Pic^\tau_{E/k}=\GG_a$.
Each irreducible component $E_i\subset E$ must be a $(-2)$-curve, whence $E_i\simeq\PP^1$ is smooth.
It follows that there are two irreducible components $E_i\neq E_j$ with intersection number
$n=E_i\cdot E_j\geq 2$. Then the intersection matrix $\begin{pmatrix} -2&n\\n&-2\end{pmatrix}$ is
not negative-definite, contradicting that $E\subset Y$ is negative-definite.
Again we see that $h^1(\O_Z)=0$, and the curve $E\subset Y$ is an ADE-configuration.
The assertions on the curves $A,F,E$ and the image of $\Sing(X)$
follow from Propositions \ref{singularity over line}, \ref{singularity over nc}.
and \ref{singularity over fundamental-singular}, respectively.
It remains to treat the rational cuspidal curve $C$, which has $\Pic^\tau_{C/k}=\GG_a$.
If the restriction map $\Pic^\tau_{Y/k}\ra\Pic^\tau_{C/k}$ is trivial,
we may apply Proposition \ref{singularity total space}.
Now suppose that the restriction map is non-zero. Then $\tilde{C}=\epsilon^{-1}(C)$
is a non-trivial $G$-torsor over $C$. It becomes trivial after pulling-back
along the normalization $\nu:\PP^1\ra C$, according to Proposition \ref{pullback trivial}.
We thus get a cartesian diagram
$$
\begin{CD}
\PP^1\oplus\O_{\PP^1} @>>> \tilde{C}\\
@VVV @VVV\\
\PP^1 @>>> C.
\end{CD}
$$
where the horizontal arrows are birational. Consequently, the Weil divisor $\tilde{C}\subset X$
is of the form $\tilde{C}=2B$, where $B=\tilde{C}_\red$ is an irreducible Weil divisor,
and the morphism $B\ra C$ is birational.
Since $C$ is the rational cuspidal curve, the morphism $B\ra C$ is either the normalization or the identity map.
To see that the letter does not happen, note that the $G$-torsor $\epsilon:X\ra Y$ becomes
trivial after pulling-back to itself. Since $h^0(\O_X)=1$, this means
that the restriction map $\Pic^\tau_{Y/k}\ra\Pic^\tau_{X/k}$ is trivial. It follows that
the composite map $\Pic^\tau_{Y/k}\ra\Pic^\tau_{B}$ is trivial, contradiction.
Thus we must have $B=\PP^1$. Consider the short exact sequence
$$
0\lra\shN\lra\O_{\tilde{C}}\lra \O_B\lra 0.
$$
The sheaf $\shN$ is torsion free, and therefore invertible when regarded as module over $\O_B$.
By Riemann--Roch, its degree is
$$
\deg(\shN) = \chi(\shN) - \chi(\O_B) = \chi(\O_{\tilde{C}}) - 2\chi(\O_B) = 2\chi(\O_C) -2\chi(\O_B) =-2.
$$
Seeking a contradiction, we now assume that the surface $X$ is locally factorial along $B$.
Then $B\subset X$ is a Cartier divisor isomorphic to $\PP^1$ with selfintersection number $B^2=-\deg(\shN) = 2$.
Riemann--Roch yields $-2=\deg(K_B)=(K_X+B)\cdot B = 2$, contradiction.
\qed
\medskip
We now furthermore assume that the K3-like covering $X$ has only isolated singularities, that
is, the surface $X$ is normal.
Let $r:S\ra X$ be a resolution of singularities.
The \emph{geometric genus} of a singularity $x\in X$ is the length
$p_g(\O_{X,x}) = \length R^1r_*(\O_S)_x$.
A singularity $x\in X$ is called \emph{rational} if its geometric genus is $p_g=0$,
and we call it \emph{elliptic} if $p_g=1$. Note that this term has various meanings
in the literature. For example, Wagreich \cite{Wagreich 1970} uses it for singularities
with arithmetic genus $p_a=1$, which is the largest integer of the form $1-\chi(\O_Z)$,
where $Z\subset S$ ranges over the exceptional divisors.
According to loc.\ cit., this includes all singularities with fundamental genus $p_f=1$,
which is defined as $p_f(\O_{X,x})=h^1(\O_Z)$, where $Z\subset S$ is the fundamental divisor
\cite{Artin 1966}. A related class of singularities are the minimally elliptic singularities
studied by Laufer \cite{Laufer 1977}.
A singularity with Hilbert--Samuel multiplicity
$e(\O_{X,x})=2$ is called a \emph{double point}.
\begin{proposition}
\mylabel{singularities on X}
All rational singularities on the K3-like covering $X$ are rational double points,
and the only possible types are $A_1$, $E_7^0$, $E_8^0$ and $D_{2n}^0$.
There is at most one non-rational singularity $x\in X$. If present, it is an elliptic double point.
\end{proposition}
\proof
The singularities on the K3-like covering are Zariski singularities, whence
have multiplicity $e=2$, by Proposition \ref{hilbert-samuel multiplicity}. Furthermore, they
are Gorenstein, thus all rational singularities must be rational double points.
The only rational double points that are Zariski singularities are of type $A_1, D_{2n}^0, E_7^0,E_8^0$,
according to Proposition \ref{zariski rdp}.
Next, consider the Leray--Serre spectral sequence for the minimal resolution of singularities $r:S\ra X$.
It gives an exact sequence
\begin{equation}
\label{long exact sequence}
0\lra H^1(X,\O_X)\lra H^1(S,\O_S)\lra H^0(X,R^1r_*\O_S)\lra H^2(X,\O_X).
\end{equation}
The term on the left vanishes, and the term on the right is one-dimensional. Moreover, we have
$K_S=K_{S/X}$, which is a Cartier divisor supported on the exceptional curve with coefficients $\leq 0$.
Now suppose that there is a non-rational singularity $x\in X$. Then $K_{S/X}< 0$,
and Serre Duality gives $H^2(S,\O_S)=0$. Thus the Picard scheme is smooth, and the connected component
$\Pic^0_{S/k}$ is an abelian variety.
On the other hand, each integral curve $E\subset S$ contained in the exceptional locus
is a curve of genus zero,
so the restriction map $\Pic^0_{S/k}\ra\Pic^0_{E/k}$ is trivial. In turn, the Albanese morphism
$S\ra A$ contracts each such $E\subset S$, and thus factors over $Y$. Using that $\Pic_{Y/k}$
is 0-dimensional,
we conclude that $\dim(A)=0$, and thus $H^1(S,\O_S)=0$. It now follows from the above exact sequence
that $R^1r_*(\O_S)$ has length $\leq 1$. In turn, there is precisely one such singularity, and its geometric genus is one.
\qed
\medskip
We will later see that if there is an elliptic singularity $x\in X$, there is no further
singularities. In any case, the structure of the smooth surface $S$ depends
on the nature of the singularities on $X$:
\begin{proposition}
\mylabel{numbers for S}
The smooth surface $S$ is a K3 surface with $\rho(S)=b_2(S)=22$
if the K3-like covering $X$ contains only rational singularities,
and is a rational surface with $\rho(S)=b_2(S)\geq 11$ otherwise.
\end{proposition}
\proof
If all singularities are rational, the exact sequence \eqref{long exact sequence}
shows that $h^1(\O_S)=0$. Furthermore, we have $\omega_{S/X}=\O_S$, whence $\omega_S=\O_S$.
By the classification of algebraic surfaces, $S$ is a K3 surface.
It then has Betti number $b_2=22$, whereas our Enriques surface $Y$ and its
K3-like covering have $b_2(X)=b_2(Y)=10$. It follows that
the exceptional divisor for $r:S\ra X$ consist of $12=b_2(S)-b_2(X)$ irreducible components.
It follows that $\rho(S)=b_2(S)$.
If there is a non-rational singularity $x\in X$, then $-K_S=-K_{S/X}$ is an effective Cartier divisor,
and we still have $H^1(S,\O_S)=0$. By the classification of surfaces, $S$ is rational.
In any case, the Picard number satisfies $\rho(S)>\rho(X)\geq \rho(Y)=10$, because there is at least one singularity,
and thus $b_2(S)\geq 11$.
\qed
\medskip
Every Enriques surface $Y$ admits at least one genus-one fibration $\varphi:Y\ra \PP^1$.
It has one or two multiple fibers $Y_b=2F$, and their multiplicity is $m=2$.
Two multiple fibers occur precisely for the classical Enriques surfaces, and we then have
$\omega_Y=\O_Y(C_1-C_2)$, where $C_1,C_2\subset Y$ are the two half-fibers.
Two fibrations $\varphi,\varphi':Y\ra\PP^1$ are called \emph{orthogonal}
if for the respective half-fibers $F,F'\subset Y$ have intersection number $(C\cdot C')=1$.
For our simply-connected Enriques surface $Y$ whose K3-like covering $X$ is normal,
there are some strong restrictions on the nature of these fibrations.
The following were already observed by Cossec and Dolgachev
(see \cite{Cossec; Dolgachev 1989}, Proposition 5.7.3 and its Corollary):
\begin{theorem}
\mylabel{properties genus-one fibrations}
For every genus-one fibration $\varphi:Y\ra \PP^1$, the following holds:
\begin{enumerate}
\item
The fibration $\varphi:Y\ra \PP^1$ is elliptic, and not quasielliptic.
\item
Each singular fiber $Y_b=\varphi^{-1}(b)$ is of Kodaira type $\I_n$ for some $1\leq n\leq 9$,
or of Kodaira type $\II$, $\III$ or $\IV$.
\item
There are at least two other elliptic fibrations $\varphi',\varphi'':Y\ra\PP^1$ so that
$\varphi,\varphi',\varphi''$ are mutually orthogonal.
\end{enumerate}
\end{theorem}
\proof
Suppose there is some quasielliptic fibration $\varphi:Y\ra\PP^1$. Then almost all fibers
$Y_a$, $a\in\PP^1$ are rational cuspidal curves. According to Proposition \ref{curves images singularities},
the normal surface $X$ must contain infinitely many singularities, contradiction.
The Kodaira types in (ii) correspond to the curves of canonical type that are reduced.
If there would be another Kodaira type, the corresponding fiber contains
an ADE-configuration of type $D_4$. Its fundamental cycle is non-reduced, and again
$X$ contains a curve of singularities, contradiction.
The last property holds for all Enriques surfaces with a few exceptions, according
to \cite{Cossec; Dolgachev 1989}, Theorem 3.5.1. The exceptions are called \emph{extra special} in loc.\ cit.,
and each of them contains a genus-one fibration containing a fiber of type $\II^*$,
$I_4^*$ or $\III^*$, which are non-reduced. This contradicts Property (ii).
\qed
\medskip
A integral curve $E\subset Y$ is called \emph{non-movable} if $h^0(\O_Y(E))=1$.
\begin{proposition}
\mylabel{preimage elliptic}
Let $E\subset Y$ be a non-movable elliptic curve.
Then the schematic preimage $\epsilon^{-1}(E)\subset X$ is an elliptic curve, and
the K3-like covering $X$ is smooth along this preimage.
\end{proposition}
\proof
The curve $2E$ is movable, and defines an elliptic fibration $\varphi:Y\ra\PP^1$ for which
$E\subset Y_b$ is a half-fiber. Suppose first that $Y$ is classical.
Then there is another half-fiber $C$, and $\omega_Y=\O_Y(E-C)$ has order two in $\Pic(Y)$.
According to \cite{Cossec; Dolgachev 1989}, Theorem 5.7.2 multiple fibers are not wild,
so the restriction $\omega_Y|E=\O_E(E)$ has order two in $\Pic(E)$.
It follows that the restriction map $\Pic^\tau_{Y/k}\ra\Pic^\tau_{E/k}$ is injective.
Now suppose that $Y$ is supersingular, such that $\omega_Y=\O_Y$. The short exact sequence
$0\ra\O_Y(-E)\ra \O_Y\ra\O_E\ra 0$ yields a long exact sequence
$$
H^1(Y,\O_Y)\lra H^1(E,\O_E)\lra H^2(Y,\O_Y(-E))\lra H^2(Y,\O_Y).
$$
The map on the right is Serre dual to $H^0(Y,\O_Y)\ra H^0(Y,\O_Y(E))$.
The latter is bijective, because the curve $E\subset Y$ is non-movable.
In turn, the map on the left $H^1(Y,\O_Y)\ra H^1(E,\O_E)$ is surjective,
and again we conclude that the restriction map $\Pic^\tau_{Y/k}\ra\Pic^\tau_{E/k}$ is injective.
In both cases, we see that the $G$-torsor $E'=\epsilon^{-1}(E)\ra E$ is non-trivial.
According to Proposition 4.9, the curve $E'$ is integral, with $h^0(\O_{E'})=h^1(\O_{E'})=1$.
This curve is normal. If not, the normalization $E''\ra E'$ must have $h^1(\O_{E''})<1$,
giving an integral surjection $\PP^1\ra E$, contradiction.
Thus $E'$ is a smooth Cartier divisor in the surface $X$, and we conclude that $X$ is smooth
along $E'$.
\qed
\medskip
Let us write $x_1,\ldots,x_m\in X$ for the singularities on the K3-like coverings.
The complete local rings $\O_{X,x_i}$ are
Zariski singularity, thus given by a formal power series of the form $g_i=z^2-f_i(x,y)$.
Recall that the \emph{Tjurina number} satisfies
$$
\tau_i = 2\cdot\length k[[x,y]]/(f_i,\partial f_i/\partial x,\partial f_i/\partial y),
$$
by Proposition \ref{tjurina for isolated zariski}.
Note that this is always an even number.
\begin{proposition}
\mylabel{theta and tjurina}
We have $\Theta_{X/k}=\O_X^{\oplus 2}$,
and the Tjurina numbers for the singularities satisfy $\sum_{i=1}^m \tau_i = 24$.
In particular, if $S$ is a K3 surface, then the exceptional divisor for
the resolution of singularities $r:S\ra X$ has 12 irreducible components.
\end{proposition}
\proof
If $S$ is K3, this is due to Ekedahl, Hyland and Shepherd-Barron \cite{Ekedahl; Hyland; Shepherd-Barron 2012},
and their arguments generalize as follows:
Since $\epsilon:X\ra Y$ is a torsor for the Cartier dual $G=\Hom(\Pic^0_{Y/k},\GG_m)$,
we have a short exact sequence
$$
0\lra\O_X\lra\Theta_{X/k}\lra \shL\lra 0
$$
for some invertible sheaf $\shL$. The map on the left corresponds to
the $p$-closed vector field defining the $G$-action on $X$.
Using that $\det(\Theta_{X/k})=\omega_X=\O_X$,
we conclude that $\shL\simeq\O_X$. The preceding extension splits, because
we have $\Ext^1(\O_X,\O_X)=H^1(X,\O_X)=0$. Summing up, $\Theta_{X/k}=\O_X^{\oplus 2}$.
Now suppose that $S$ is a K3 surface.
It then has Betti number $b_2=22$, whereas our Enriques surface $Y$ and its
K3-like covering have $b_2(X)=b_2(Y)=10$. It follows that
the exceptional divisor for $r:S\ra X$ consist of $12=b_2(S)-b_2(X)$ irreducible components.
Since the resulting singularities are rational double points, we must have $\sum \tau_i = 24$.
Finally suppose that $S$ is rational.
Here we have Betti number $b_2(S)=10-K_S^2$
and Chern number $c_2(S)=2 + b_2(S) = 12-K_S^2$. The Chern numbers are related,
according to \cite{Ekedahl; Hyland; Shepherd-Barron 2012}, Proposition 3.12 and Corollary 3.13,
in the following way:
$$
c_2(S) = c_2(X) + \gamma \quadand c_2(\Theta_{X/k}) = c_2(X) - \tau.
$$
Here $c_2(X)=c_2(L_{X/k}^\bullet)$ is defined as the second Chern class of the cotangent complex
$L_{X/k}^\bullet$, which in our situation is a complex of length one comprising locally free sheaves
of finite rank.
It can be defined with the help of an embedding $X\subset\PP^n$ into some projective $n$-space.
Moreover, $\tau=\sum \tau_i$, and the other correction term $\gamma$ is given by the Local Noether Formula
$$
-1=-\length R^1g_*(\O_S) = \frac{K_S^2 + \gamma}{12},
$$
from \cite{Ekedahl; Hyland; Shepherd-Barron 2012}, Proposition 3.12.
In other words $\gamma= -12 - K_S^2$. Combining these equations, one gets
$$
0=c_2(\Theta_{X/k}) = c_2(S) -\gamma -\tau = (12-K_S^2) + (12 +K_S^2) -\tau = 24-\tau.
$$
Again, we have $\sum \tau_i=24$.
\qed
\section{Normal surfaces with trivial tangent sheaf}
\mylabel{Trivial tangent sheaf}
Let $k$ be an algebraically closed ground field of characteristic $p=2$.
In order to understand and construct simply-connected Enriques surface $Y$ and
their K3-like coverings $X$, we now impose
seven axiomatic conditions on certain elliptic fibrations.
Let $h:S\ra\PP^1$ be a smooth elliptic surface, whose total space $S$ is either
a rational surface or a K3 surface.
An irreducible curve $C\subset S$ is called \emph{vertical} if it is contained
in some fiber $h^{-1}(a)$, $a\in\PP^1$. Otherwise, it is called \emph{horizontal}.
Let $S\ra S'\ra\PP^1$ be the relative minimal model, obtained by successively contracting
vertical $(-1)$-curves. Note that if $S$ is a K3 surface, there are no $(-1)$-curves
at all, such that $S=S'$.
Let $E\subset S$ be a vertical negative-definite curve, and $r:S\ra X$ be its contraction.
Then $X$ is a proper normal surface, with an induced elliptic fibration $f:X\ra\PP^1$, satisfying $h=f\circ r$.
We denote by $x_1,\ldots,x_r\in X$ the images of the connected components $E_1,\ldots,E_r\subset E$,
and assume that each local ring $\O_{X,x_i}$ is singular.
Then $r:S\ra X$ is a resolution of singularities, but we do not assume that it is the minimal resolution.
Our seven conditions are:
\newcounter{ENr}
\setcounter{ENr}{0}
\newcommand{\En}[1]{\refstepcounter{ENr}\textbf{(E\arabic{ENr})} \label{#1} }
\newcommand{\eref}[1] {\text{\rm (E\ref{#1})}}
\medskip
\begin{list}{-mm}{\leftmargin2em\itemsep1em}
\item[\En{CI}]
\emph{The singular local rings $\O_{X,x_i}$, $1\leq i\leq r$ are complete intersections
with free tangent modules.}
\item[\En{Tjurina}]
\emph{The Tjurina numbers $\tau_i$ for the local rings $\O_{X,x_i}$ add up to
$\sum\tau_i =24$.}
\item[\En{Length}]
\emph{If $S$ is a rational surface, the skyscraper sheaf $R^1r_*(\O_S)$ has length one;
it is supported by the multiple fiber if $h:S\ra\PP^1$ has a multiple fiber.}
\item[\En{Nonfixed}]
\emph{If $S$ is a rational surface, then $\omega_S=\shN(D)$, where
$D$ is a divisor supported on the exceptional divisor $E\subset S$,
and the invertible sheaf $\shN$ is non-fixed.}
\item[\En{Elliptic}]
\emph{There is a horizontal Cartier divisor $F\subset X$ that is an elliptic curve.}
\item[\En{Zariski}]
\emph{For each singularity $x_i\in X$, the local rings $\O_{X,x_i}$ are Zariski singularities.}
\item[\En{Cuspidal}]
\emph{There is a horizontal Cartier divisor $C\subset X$ that is a rational cuspidal curve,
and the singular point on $C$ is a regular point on $X$.}
\end{list}
\medskip
Here the invertible sheaf $\shN\in\Pic(S)$ is called \emph{non-fixed}
if $h^0(\shN)\geq 1$, and equality holds if and only if $\shN=\O_S$.
Note that \eref{CI} is a consequence of \eref{Zariski}, according to Proposition \ref{theta for zariski surface}.
However, it is useful to have \eref{CI} as a separate condition, because it will be handy when
the Enriques surface acquires rational double points.
Our conditions indeed help to analyze and construct simply-connected Enriques surface,
because we have:
\begin{proposition}
\mylabel{conditions necessary}
Suppose $Y$ is a simply-connected Enriques surface whose K3-like covering $X$ is normal,
$r:S\ra X$ is a resolution of singularities, $E\subset S$ is the exceptional divisor,
and $h:S\ra \PP^1$ is the fibration
induced from some elliptic fibration $\varphi:Y\ra \PP^1$.
Then conditions \eref{CI}--\eref{Zariski} do hold, whereas condition \eref{Cuspidal} does not hold.
\end{proposition}
\proof
Condition \eref{Nonfixed} is true, because $\omega_X=\O_X$, and thus $\omega_S=\O_S(K_{S/X})$,
where the relative canonical divisor $K_{S/X}$ is supported by the exceptional divisor.
Property \eref{Zariski} and thus also \eref{CI} hold according to Theorem \ref{homological characterization}.
Proposition \ref{theta and tjurina} gives \eref{Tjurina}, and Theorem \ref{properties genus-one fibrations} ensures
\eref{Elliptic}.
Condition \eref{Length} pertains to the case that $S$ is rational.
According to Proposition \ref{numbers for S} together with Proposition
\ref{singularities on X}, there is exactly one elliptic singularity $x_1\in X$,
whence $R^1r_*(\O_S)$ has length one.
Now suppose that the relatively minimal rational elliptic surface $S'\ra\PP^1$ has a multiple fiber
$S'_a$. Write $C'$ for the underlying indecomposable curve of canonical type.
Then $K_{S'}=-C'$, according to \cite{Cossec; Dolgachev 1989}, Proposition 5.6.1.
Seeking a contradiction, we assume that the elliptic singularity $x_1\in X$ does not lie on $X_a$.
Without restriction, we may assume that $r:S\ra X$ is the minimal resolution of singularities.
Then $K_S=-E$, where $E\subset S$ is a negative-definite curve mapping to $x_1\in X$.
In turn, $K_{S'}= -E'$, where $E'\subset S'$ is the image of $E$.
Thus the curves $E'\neq C'$ are linearly equivalent. But since $S'_a$ is a multiple fiber,
the indecomposable curve of canonical type $C'$ is non-movable, contradiction.
Finally, we have to verify that Condition \eref{Cuspidal} does not hold.
Suppose to the contrary that there is a Cartier divisor $C\subset X$
that is a rational cuspidal curve such that the local ring $\O_{X,c}$ is regular,
where $c\in C$ is the singular point.
The Adjunction Formula $\deg(\omega_C) = (K_X+C)\cdot C$ implies $C^2=0$.
Consequently, the image $D=\epsilon(C)$ on the Enriques surface
is an integral rational curve with $D^2=0$. It follows that $D$ is a rational cuspidal curve.
By Proposition \ref{curves images singularities}, there is
a singular point $x\in X$ mapping to $D$.
But for each closed point $x\in C$, the local ring $\O_{X,x}$ is regular, because
either $x=c$ or $\O_{C,x}$ is regular, contradiction.
\qed
\medskip
The goal of this section is to establish a converse for Proposition \ref{conditions necessary}.
Let us write $\O_X(n)$ for the preimage of the invertible sheaves $\O_{\PP^1}(n)$ under
the elliptic fibration $f:X\ra\PP^1$, and likewise we write $\O_S(n)$ and $\O_{S'}(n)$.
The following is the key step in producing Enriques surfaces with normal K3-like coverings:
\begin{theorem}
\mylabel{conditions for theta}
If $r:S\ra X$ satisfies the conditions \eref{CI}--\eref{Nonfixed}, then the dualizing sheaf
is $\omega_X=\O_X$, and the tangent sheaf is given by
$\Theta_{X/k}= \O_X(n)\oplus\O_X(-n)$ for some integer $0\leq n\leq 2$.
If furthermore condition \eref{Elliptic} or \eref{Cuspidal} holds, we have $n=0$ or $n\neq 0$, respectively.
In the latter case, the integer satisfies $nd=4$, where $d\geq 1$ is the degree of the induced morphism $f:C\ra\PP^1$.
\end{theorem}
\proof
Our first step is to show that $\omega_X=\O_X$.
The case that $S$ is K3 is obvious: Every negative-definite curve on $S$
produces a rational double point on $X$, thus our $X$ is Gorenstein with $\omega_X=\O_X$.
The more interesting case is that $S$ is rational: Then $H^1(S,\O_S)=H^2(S,\O_S)=0$,
and the Leray--Serre spectral sequence for
the resolution of singularities $r:S\ra X$ gives
an exact sequence
$$
0\ra H^1(X,\O_X)\ra H^1(S,\O_S)\ra H^0(X,R^1r_*(\O_S))\ra H^2(X,\O_X)\ra H^2(S,\O_S).
$$
It follows that $h^1(\O_X)=0$ and $h^2(\O_X)=1$, the latter by Condition \eref{Length}.
Serre duality gives $h^0(\omega_X)=1$.
Suppose that $\omega_X$ is non-trivial. Then $\omega_X=\O_X(C)$ for some Cartier divisor
$C\subset X$ that is not linearly equivalent to any other effective Cartier divisor.
According to Condition \eref{Nonfixed}, we have $\omega_S=\shN(D)$, where
$D\subset S$ is supported by the exceptional divisor $E\subset S$ and $\shN$ is non-fixed.
If $\shN=\O_S$, then $\omega_X$ is trivial outside the singularities,
whence everywhere trivial, contradiction. Thus there are two curves $A\neq B$ on $S$
both giving $\shN$. Write $A=A_1+A_2$ and $B=B_1+B_2$ where $A_2$, $B_2$ are the parts
supported by the exceptional divisor $E\subset S$. Using $h^0(\omega_X)=1$,
we infer that both $A_1,B_1\subset S$ map to $C\subset X$, hence $A_1=B_1$. It follows that
$A_2\neq B_2$ are linearly equivalent and supported by the exceptional divisor $E\subset S$. This contradicts the
fact that the intersection form $(E_i\cdot E_j)$ is negative-definite.
Summing up, we have $\omega_X=\O_X$.
The tangent sheaf $\Theta_{X/k}$ is locally free of rank two, according to condition \eref{CI}.
Note that the dual of $\Theta_{X/k}$ coincides with the bidual of $\Omega^1_{X/k}$. Consequently
we have $\det(\Theta_{X/k})=\omega_X^\vee=\O_X$.
The next step is to verify that the second Chern class $c_2(\Theta_{X/k})$ vanishes.
The case of K3 surfaces was already treated in \cite{Ekedahl; Hyland; Shepherd-Barron 2012}, Section 3.
We proceed in a similar way: In both cases on has
$$
c_2(\Theta_{X/k}) = c_2(X) - \tau\quadand c_2(X) = c_2(S) - \nu,
$$
where $\tau=\sum\tau_i=24$ is the sum of Tjurina numbers, and $\nu$ is a correction term
$$
\nu = c_1^2(X) - c_1^2(S) -12\cdot\length R^1r_*(\O_S).
$$
according to the Local Noether Formula
\cite{Ekedahl; Hyland; Shepherd-Barron 2012}, Proposition 3.12. In our situation $c_1^2(X)=c_1^2(\Theta_X)=0$.
If $S$ is K3, then $c_1^2(S) =0$, $c_2(S)=24$ and $\nu=0$, which gives $c_2(\Theta_{X/k})=0$.
If $S$ is rational, we have $c_2(S) = 2 + b_2=12-c_1^2(S)$ and $\nu= c_1^2(S) -12$,
also giving
$$
c_2(\Theta_{X/k})= c_2(S) -\tau -\nu = (12 - c_1^2(S)) - 24 - (-c_1^2(S) -12) = 0.
$$
Our next step is to show that the tangent sheaf splits as a sum of two line bundles.
Consider the short exact sequence
$f^*(\Omega^1_{\PP^1/k})\ra \Omega^1_{X/k} \ra\Omega^1_{X/\PP^1}\ra 0$.
The map on the left is injective, because the function field extension $k(\PP^1)\subset k(X)$ is separable
and $X$ has no embedded components. Thus we have a short exact sequence
$$
0\lra f^*(\Omega^1_{\PP^1/k})\lra \Omega^1_{X/k} \lra\Omega^1_{X/\PP^1}\lra 0.
$$
The sheaf $\Omega^1_{X/\PP^1}$ is invertible at each point
$x\in X$ at which the fiber $X_{f(x)}$ is regular.
The bidual $\shN=(\Omega^1_{X/\PP^1})^{\vee\vee}$ is
a reflexive sheaf of rank one. Taking biduals for the above exact sequence thus gives an exact sequence
$$
0\lra\shL\lra\Theta_{X/k}^\vee\lra\shF\lra 0,
$$
where $\shL=f^*(\Omega^1_{\PP^1/k})=\O_X(-2)$, and $\shF\subset\shN$ is a subsheaf of rank one
whose inclusion is an equality on the complement of a finite set $Z=\Supp(\shN/\shF)$.
On this open subset, we have $\omega^\vee_X=\det(\Theta_{X/k}) = \shL\otimes\shF$,
and it follows that $\shN=\shL^\vee$. In particular, the reflexive rank one sheaf $\shN$
is invertible, and our exact sequence takes the form
$$
0\lra\shL\lra\Theta_{X/k}^\vee\lra\shI\shL^\vee\lra 0
$$
for some ideal $\shI\subset\O_X$ defining a closed subscheme supported by the finite set $Z\subset X$.
This gives
$$
0=c_2(\Theta_{X/k}^\vee) = -c_1^2(\shL) + \length(\O_X/\shI).
$$
Since $\shL$ comes from $\PP^1$, we have $c_1^2(\shL)=0$ and whence $\shI=\O_X$.
Summing up, the tangent sheaf is an extension of $\shL^\vee$ by $\shL$. The extension class
lies in
$$
\Ext^1(\shL^\vee,\shL) = H^1(X,\shL^{\otimes 2}).
$$
The latter can be computed with the Leray--Serre spectral sequence, together with the Projection Formula:
We get an exact sequence
$$
0\lra H^1(\PP^1,\O_{\PP^1}(-4))\lra H^1(X,\shL^{\otimes 2}) \lra H^0(\PP^1, R^1f_*(\O_X)(-4)).
$$
The term on the right vanishes: The higher direct image sheaf $R^1f_*(\O_X)$ commutes with base-change.
Since $S\ra\PP^1$ has no wild fibers, according to \cite{Cossec; Dolgachev 1989}, Proposition 5.6.1,
if follows that the map $t\mapsto h^1(\O_{X_t})$
takes the constant values one, and thus $R^1f_*(\O_X)$ is invertible.
The Leray--Serre spectral sequence yields
$$
\chi(R^1f_*(\O_X)) = \chi(\O_{\PP^1}) - \chi(\O_X) = 1 - 1 =0,
$$
thus $R^1f_*(\O_X)=\O_{\PP^1}(-1)$, and consequently $R^1f_*(\O_X)(-4)$ has no non-zero global
section.
We infer that the extension of $\shL^\vee$ by $\shL$ giving $\Theta_{X/k}$ comes
from an extension of $\O_{\PP^1}(-2)$ by $\O_{\PP^1}(2)$,
in particular $\Theta_{X/k}$ is the preimage of some locally free sheaf $\shE$ of rank two
on $\PP^1$. But all such sheafs on $\PP^1$ are sums of line bundles, according to
Grothendieck (see for example \cite{Okonek; Schneider; Spindler 1980}).
The splitting type of $\shE$ is of the form $(-n,n)$ for some
integer $n\geq 0$, because $c_1(\shE)=0$. Since $\shE$ surjects onto $\O_{\PP^1}(2)$,
only the three possibilities $n=0,1,2$ exists.
Now suppose that Condition \eref{Elliptic} holds:
Consider the horizontal Cartier divisor $F\subset X$ that is an elliptic curve,
and let $d>0$ be the degree of the map $F\ra\PP^1$.
We have an exact sequence
$$
0\lra \O_F(-F)\lra \Omega^1_{X/k}|_F\lra\Omega^1_{F/k}\lra 0.
$$
The scheme $X$ is regular at the regular Cartier divisors, so the above
sheaves are locally free on $F$.
The sheaves $\Omega^1_{X/k}$ and $\Theta_{X/k}^\vee$ have the
same restriction to $F$, which is thus of form $\shM\oplus\shM^\vee$,
where $\deg(\shM)=nd$. Seeking a contradiction, we assume $n\neq 0$,
such that $\deg(\shM)>0$. This implies that there is only the zero map $\shM\ra\Omega^1_{F/k}=\O_F$.
It follows that $\shM^\vee\ra\Omega^1_{F/k}$ is surjective, thus bijective, contradicting
$\deg(\shM^\vee)<0$.
Finally, suppose that Condition \eref{Cuspidal} holds: Consider the horizontal Cartier divisor $C\subset X$ that is a
rational cuspidal curve,
and let $d>0$ be the degree of the map $C\ra\PP^1$.
Now we have an exact sequence
$$
0\lra \O_C(-C)\lra \Omega^1_{X/k}|_C\lra\Omega^1_{C/k}\lra 0.
$$
To compute the term on the right, write the rational cuspidal curve with two affine charts as
$$
C=\Spec k[t^2,t^3] \cup \Spec k[t^{-1}].
$$
On the first chart, the K\"ahler differentials are generated by $dt^2, dt^3$ modulo the
relation $t^4dt^2$. On the second chart, $dt^{-1}$ is a generator. On the overlap,
we have $dt^3=t^2dt= t^4\cdot dt^{-1}$.
It follows that the coherent sheaf $\shN=\Omega^1_{C/k}/\text{\rm Torsion}$
is invertible with $\deg(\shN)=-4$. As in the preceding paragraph, $\Omega^1_{X/k}|_C = \shM\oplus\shM^\vee$
with $\deg(\shM)=nd$. Hence the map $\shM\ra\shN$ vanishes,
and $\shM^\vee\ra\shN$ is surjective, thus bijective. It follows that $nd=4$,
therefore $n\neq 0$.
\qed
\begin{proposition}
\mylabel{p-closed}
Suppose that $S\ra X$ satisfies conditions \eref{CI}--\eref{Elliptic}.
Then the canonical map of restricted Lie algebras
$H^0(X,\Theta_{X/k})\ra H^0(\PP^1,\Theta_{\PP^1/k})$ is injective,
and ever non-zero vector in $H^0(X,\Theta_{X/k})$ is $p$-closed.
\end{proposition}
\proof
We saw in the proof for Theorem \ref{conditions for theta} that the
canonical map for K\"ahler differentials $f^*(\Omega^1_{\PP^1/k})\ra\Omega^1_{X/k}$
induces a short exact sequence
$$
0\lra \O_X(-2) \lra\Theta_{X/k}\lra f^*(\Theta_{\PP^1/k})\lra 0.
$$
In turn, we get an exact sequence
$0\ra \O_{\PP^1}(-2) \ra f_*(\Theta_{X/k}) \ra \Theta_{\PP^1/k}$,
and thus an inclusion $H^0(X,\Theta_{X/k})\subset H^0(\PP^1,\Theta_{\PP^1/k})$.
Since $\Theta_{\PP^1/k}=\O_{\PP^1}(2)$,
the restricted Lie algebra $\ideal h = H^0(\PP^1,\Theta_{\PP^1/k})$ is 3-dimensional.
It can be seen as a semidirect product $\ideal h = \ideal a\rtimes \ideal b$,
where $\ideal a=k^{\oplus 2}$ with trivial Lie bracket and $p$-map, and $\ideal b = kb$
with $b^{[p]} = b$. The semidirect product structure comes from the homomorphism
$\ideal b\ra\mathfrak{gl}(\ideal a)$ given by $b\mapsto \id_{\ideal a}$,
compare the discussion in \cite{Schroeer 2007}, Section 3.
It is easy to see that every non-zero vector in $\ideal a\rtimes\ideal b$
is $p$-closed, whence the same holds for the restricted Lie algebra $H^0(X,\Theta_X)$.
\qed
\medskip
Now suppose that $S\ra X$ satisfies the condition \eref{CI}--\eref{Zariski}.
Then the restricted Lie algebra $\ideal g=H^0(X,\Theta_{X/k})$ is two-dimensional,
and every non-zero vector is $p$-closed.
Moreover, the singularities $x_1,\ldots,x_r\in X$ are Zariski.
Let $A_i=\O_{X,x_i}^\wedge$ be the corresponding complete local ring.
Then $A_i\ideal g =\Theta_{A_i/k}$. In turn, Proposition \ref{canonical line}
yields canonical lines $\ideal l_i\subset \ideal g$, for $1\leq i\leq r$.
\begin{theorem}
\mylabel{conditions sufficient}
Suppose $S\ra X$ satisfies conditions \eref{CI}--\eref{Zariski}.
Then $X$ is a K3-like covering. More precisely,
for each vector field $D\in\ideal g= H^0(X,\Theta_{X/k})$ not contained in
the union of the canonical lines $\ideal l_i\subset \ideal g$, the ensuing quotient $Y=X/G$
is a simply-connected Enriques surface.
The Picard group $\Pic^\tau_{Y/k}$ is the Cartier dual of $G$,
and the projection $\epsilon:X\ra Y$ is the K3-like covering.
\end{theorem}
\proof
Since the vector field $D$ avoids the canonical lines, the $G$-action is free and
the quotient $Y=X/G$ is smooth, according to Proposition \ref{canonical line}.
Furthermore, $\omega_X=\O_X$ by Theorem \ref{conditions for theta}.
In the resolution of singularities $r:S\ra X$, the smooth surface $S$ is either K3 or rational,
and in both cases the algebraic fundamental group $\pi_1(S)$ vanishes.
It follows that $\pi_1(X)$ vanishes as well.
Now Proposition \ref{enriques quotient} ensures that $Y$ is a simply-connected Enriques surface,
the quotient morphism $\epsilon:X\ra Y$ is the K3-like covering,
and $\Pic^\tau_{Y/k}$ is the Cartier dual of $G$.
\qed
\section{The twistor curves in the moduli stack}
\mylabel{Twistor curves}
Let $X$ be a normal K3-like covering. The results in the preceding section immediately
give:
\begin{proposition}
\mylabel{k3-like theta}
The tangent sheaf $\Theta_{X/k}$ is isomorphic to $\O_X^{\oplus 2}$,
every vector in the 2-dimensional restricted Lie algebra $\ideal g=H^0(X,\Theta_{X/k})$
is $p$-closed, and for each point $x\in X$, we have $\O_{X,x}\ideal g=\Theta_{X/k,x}$.
Moreover, for each elliptic fibration $f:X\ra\PP^1$, we get an inclusion
of restricted Lie algebras $\ideal g\subset H^0(\PP^1,\Theta_{\PP^1/k})$.
\end{proposition}
\proof
Choose an Enriques surface $Y$ so that $\epsilon:X\ra Y$ is the K3-like covering,
and some elliptic fibration $\varphi:Y\ra\PP^1$.
According to Proposition \ref{conditions necessary}, Conditions
\eref{CI}--\eref{Zariski} hold. The claim on $\Theta_{X/k}$ now follows from
Theorem \ref{conditions for theta}. Moreover, every vector space basis $D_1,D_2\in\ideal g$ yields an isomorphism
$(D_1,D_2):\O_X^{\oplus 2}\ra \Theta_{X/k}$, whence an $\O_{X,x}$-basis $D_1,D_2\in\Theta_{X/k,x}$
at each point $x\in X$.
The assertions on the restricted Lie algebra
$\ideal g$ come from Proposition \ref{p-closed}.
\qed
\medskip
Let $x_1,\ldots,x_r\in X$ be the singularities.
The corresponding local rings $A_i=\O_{X,x_i}$ are Zariski singularities,
and we have $A_i\ideal g=\Theta_{A_i/k}$.
According to Proposition \ref{canonical line}, this gives
canonical lines $\ideal l_i\subset\ideal g$, $1\leq i\leq r$
in the two-dimensional restricted Lie algebra $\ideal g=\Theta_{X/k}$.
We call the projective line $\PP(\ideal g)\simeq \PP^1$ the \emph{twistor curve}
attached to the K3-like covering $X$, in analogy to twistor spaces coming from
the two-spheres of complex structures on complex hyperk\"ahler manifolds.
The closed points $t_i\in\PP(\ideal g)$
corresponding to the canonical lines ${\ideal l}_i\subset\ideal g_i$ are called the \emph{boundary points}.
We write
$$
\PP(\ideal g)^\circ= \PP(\ideal g)\smallsetminus \{t_1,\ldots,t_r\}
$$
for the complementary open subset of \emph{interior points}.
We saw in the last section that each interior point $t\in\PP(\ideal g)$ of the
twistor line gives a simply connected Enriques surface $Y_t=X/G_t$, where
$G_t$ is the height-one group scheme corresponding to the non-zero vector field $D\in H^0(X,\Theta_{X/k})$
giving the interior point $t\in\PP(\ideal g)$.
It is easy to see that this construction extends to a flat family
$$
\formal Y\lra\PP(\ideal g)^\circ
$$
of Enriques surfaces $\formal Y_t=X/G_t$. This family can be seen as a morphism
$$
\PP(\ideal g)^\circ\lra \shM_\Enr
$$
into the moduli stack $\shM_\Enr$ of Enriques surface.
We now seek to understand this map more closely, in particular how to verify that
this map is non-constant.
The elliptic fibration $f:X\ra \PP^1$ induces a family of elliptic fibrations
$\formal Y\ra \PP^1\times\PP(\ideal g)^\circ$
over the interior of the twistor line. Moreover, the nonzero vector field
$$
D\in\ideal g=H^0(X,\Theta)\subset H^0(X,\Theta_{\PP^1/k})
$$
induce non-zero vector fields on $\PP^1$. In turn, we get a commutative diagram
\begin{equation}
\label{fibration diagram}
\begin{CD}
X @>\epsilon>> X/G\\
@VfVV @VV\varphi V\\
\PP^1 @>>> \PP^1/G.
\end{CD}
\end{equation}
The vertical map on the right is the induced elliptic fibration
$\varphi:Y\ra\PP^1$ on the simply-connected Enriques surface $Y=X/G$.
Since $\Theta_{\PP^1/k}=\O_{\PP^1}(2)$,
this induced vector field vanishes along a divisor $A\subset\PP^1$ of degree two.
Geometrically speaking, the global vector field $D\in H^0(X,\Theta_{X/k})$
is vertical along the fiber $X_a\subset X$ for the points $a\in A$.
The divisor $A\subset \PP^1$ of degree $\deg(A)=2$ consists either of one or two points.
\begin{proposition}
\mylabel{classical or supersingular}
If $A\subset\PP^1$ consist of two points, then the simply-connected Enriques surface $Y=X/G$ is
classical. If this divisor consists of only one point, $Y$ is supersingular.
\end{proposition}
\proof
Write $\tilde{\ideal g}=H^0(\PP^1,\Theta_{\PP^1/k})$ for the three-dimensional restricted Lie algebra
that contains $\ideal g = H^0(X,\Theta_{X/k})$.
Writing $\PP^1=\Spec k[t] \cup\Spec k[t^{-1}]$, we may regard
$$
D_0=\partial/\partial t, \quad D_1=t\partial/\partial t\quadand D_2= t^2\partial/\partial t
$$
as a vector space basis for $\tilde{\ideal g}$.
Given a linear combination $D=\sum\lambda_iD_i$, one computes $D(t)=\sum\lambda_it^i$.
A direct computation shows that $D^2=0$ holds if and only if
$\lambda_1=0$. The condition $D^2=0$ means that $G=\alpha_p$, in other words that the
Enriques surface $Y$ is supersingular.
On the other hand, the condition $\lambda_1=0$ means that $D$, regarded as a section of $\Theta_{\PP^1/k}=\O_{\PP^1}(2)$,
vanishes at a single point.
\qed
\begin{proposition}
\mylabel{multiple fibers}
The fibers $Y_b\subset Y$ over the images $b\in \PP^1=\PP^1/G$ of the points
points $a\in A\subset \PP^1$ are precisely the multiple fibers for $\varphi:Y\ra\PP^1$.
\end{proposition}
\proof
Write $B\subset\PP^1/G$ for the image of $A\subset\PP^1$, and consider the complementary
open subsets $U=\PP^1\smallsetminus A$ and $V=\PP^1/G\smallsetminus B$.
The $G$-action on $U$ is free, and the projection $U\ra V$ is a $G$-torsor.
The diagram \eqref{fibration diagram}, which is $G$-equivariant,
yields a morphism
$X_U\ra U\times_V Y_V$
of $G$-torsors over $Y_V$. This must be an isomorphism, by the general fact that
the categories of torsors are groupoids.
Now let $Y_b$ be some multiple fiber, for some closed point $b\in\PP^1/G$.
Then for each closed point $y\in Y_b$, the local ring has $\edim(\O_{Y_b,y})=2$.
Consequently, the fiber product
$\PP_1\times_{\PP^1/G} Y_b$ has embedding dimension $\geq 3$ at each closed point.
In light of the preceding paragraph, it follows that $b\not\in V$, because
our K3-like covering $X$ is normal.
Summing up, we have shown that multiple fibers for $\varphi:X\ra\PP^1$
may occur at most over points $b\in B$.
If $A\subset\PP^1$ consists of only one point, the $Y$ is supersingular
by Proposition \ref{classical or supersingular}, and there
is precisely one multiple fiber. If $A$ consists of two points, then $Y$ is classical, and there
are precisely two multiple fibers. In both case, we conclude that for all points $b\in B$,
the fiber $Y_b$ must be multiple.
\qed
\medskip
It follows that if the restricted Lie algebra $\ideal g\subset H^0(\PP^1,\Theta_{\PP^1/k})$ contains
global vectors fields $D$ with $D^2=0$ and $D^2\neq 0$, and the former defines an interior point
in the twistor curve,
then the flat family $\formal Y\ra\PP(\ideal g)^\circ$ of simply-connected Enriques surfaces
contains both supersingular and classical members, whence the morphism
$\PP(\ideal g)^\circ\ra \shM_\Enr$ is non-constant.
Moreover, one may use the position of the points $b\in \PP^1/G$ where $Y_b$ is multiple,
relative to other points $b'\in\PP^1/G$ where $Y_{b'}$ is singular,
to deduce further results on the non-constancy of the twistor construction
$\PP(\ideal g)^\circ\ra \shM_\Enr$.
\section{K3-like coverings and elliptic fibrations}
\mylabel{K3-like coverings fibrations}
Let $k$ be an algebraically closed ground field of characteristic $p=2$,
and $Y$ be a simply-connected Enriques surface, with K3-like covering
$\epsilon:X\ra Y$. The latter is a torsor for the Cartier dual $G$ of
$\Pic^\tau_{X/k}$. Assume that $X$ has only isolated singularities.
We now want to study the geometry and the singularities of $X$ in more detail, by choosing an
elliptic fibration $\varphi:Y\ra\PP^1$.
\begin{proposition}
\mylabel{stein factorization}
The Stein factorization for the composite map $\varphi\circ\epsilon:X\ra \PP^1$ is
the Frobenius map $F:\PP^1\ra\PP^1$.
\end{proposition}
\proof
This could be deduced from the diagram \eqref{fibration diagram}. Let us give a direct, independent argument:
Write $C$ for the Stein factorization, such that we have a commutative diagram
$$
\begin{CD}
X @>\epsilon>> Y\\
@V\psi VV @VV\varphi V\\
C @>>s> \PP^1.
\end{CD}
$$
The surjection $s:C\ra\PP^1$ is radical,
because this holds for $\epsilon:X\ra Y$. Whence $C=\PP^1$, and the morphism
is a power of the relative Frobenius map $F:\PP^1\ra\PP^1$. Since $\deg(\epsilon)=2$,
we either have $s=F$ or $s=\id$. To rule out the latter is suffices to check
that $\shA=\varphi_*(\epsilon_*(\O_X))$ has rank $\geq 2$ as $\O_{\PP^1}$-module.
Consider first the case that $Y$ is classical, such that $G=\mu_2$.
Then $\epsilon_*(\O_X)=\O_Y\oplus\shL$ for some invertible sheaf $\shL\in\Pic(Y)$ of order two,
whence $\shL=\omega_Y=\O_Y(C_1-C_2)$, where $C_1,C_2\subset Y$ are the two half-fibers.
It immediately follows that $\rank(\shA)\geq 2$.
Finally, suppose that $Y$ is supersingular, such that $G=\alpha_2$.
Then we have a short exact sequence
$0\ra\O_Y\ra\epsilon_*(\O_X)\ra\O_Y\ra 0$,
according to \cite{Ekedahl 1988}, Proposition 1.7. In turn, we get a long exact sequence
$$
0\lra\O_{\PP^1}\lra\shA\lra \varphi_*(\O_Y)\lra R^1\varphi_*(\O_Y).
$$
Write $R^1\varphi_*(\O_Y)=\shL\oplus\shT$ for some invertible sheaf $\shL$ and some torsion sheaf $\shT$.
The Canonical Bundle Formula (\cite{Bombieri; Mumford 1977}, Theorem 2) ensures that $\shL=\O_{\PP^1}(-2)$.
It follows that the coboundary map $\O_{\PP^1}=\varphi_*(\O_Y)\ra R^1\varphi_*(\O_Y)$ has nontrivial kernel,
and again $\rank(\shA)\geq 2$.
\qed
\medskip
Write $f:X\ra \PP^1$ for the Stein factorization, which sits in the commutative diagram
$$
\begin{CD}
X @>\epsilon>> Y\\
@VfVV @VV\varphi V\\
\PP^1 @>>_F> \PP^1.
\end{CD}
$$
We thus get a canonical morphism $X\ra Y^\fpb$ from the K3-like covering to the \emph{Frobenius pullback}
$Y^\fpb=Y\times_{\PP^1}\PP^1$.
\begin{proposition}
\mylabel{normalization}
The morphism $X\ra Y^\fpb$ is the normalization.
Its ramification locus consists of the preimages $\epsilon^{-1}(Y_b)\subset X$
of the multiple fibers $Y_b$, $b\in \PP^1$.
\end{proposition}
\proof
Both maps $X\ra Y$ and $Y^\fpb\ra Y$ have degree two, whence $X\ra Y^\fpb$ is birational.
It thus must be the normalization, because we assume throughout that $X$ is normal.
For each closed point $y\in Y_b$ lying on a multiple fiber, the embedding dimension is
$\edim(\O_{Y_b,y})=2$, so the local rings for each closed point on the preimage of $Y_b$
in the Frobenius pullback $Y^\fpb$
has embedding dimension three. If follows that the preimage $\epsilon^{-1}(Y_b)\subset X$ belongs to the
ramification locus. On the other hand, $Y^\fpb$ is smooth at each point mapping to a point
$y\in Y$ where the morphism $\varphi:Y\ra\PP^1$ is smooth.
Since all fibers of $\varphi:Y\ra\PP^1$ are reduced by Theorem \ref{properties genus-one fibrations},
it follows that the ramification locus for $X\ra Y^\fpb$ consists precisely of the preimages of the multiple fibers.
\qed
\medskip
Next, consider the jacobian fibration $J\ra\PP^1$ attached to the elliptic fibration
$\varphi:Y\ra\PP^1$, which is a relatively minimal elliptic fibration endowed with a section $O\subset J$.
According to \cite{Cossec; Dolgachev 1989}, Theorem 5.7.2 the smooth surface $J$ is rational. Moreover,
all such rational elliptic surfaces $J\ra\PP^1$ were classified by Lang \cite{Lang 2000}.
In some sense, we have full computational control on such jacobian fibrations.
Our strategy throughout is to relate the rational elliptic surface $J\ra\PP^1$
to the fibration $f:X\ra\PP^1$ on the K3-like covering.
\begin{proposition}
\mylabel{fibers jacobian}
For each closed point $b\in \PP^1$, the fibers $Y_b$ and $J_b$ have the same Kodaira type.
In particular, the rational elliptic surface $J\ra\PP^1$ has only reduced
fibers, and the possible Kodaira types
for the singular fibers
are $\I_n$ with $1\leq n\leq 9$, $\II$, $\III$ or $\IV$.
\end{proposition}
\proof
If $b\in \PP^1$ is a closed point whose fiber $Y_b$ is non-multiple, then
$J\ra \PP^1$ admits a section over the formal completion $\Spec(\O_{\PP^1,b}^\wedge)$,
which yields an isomorphisms
$$
J\otimes_{\PP^1}\Spec(\O_{\PP^1,b}^\wedge) \simeq Y\otimes_{\PP^1}\Spec(\O_{\PP^1,b}^\wedge).
$$
For the multiple fibers $Y_b$, one can at least say that
$J_b$ and $Y_b$ have the same Kodaira type,
according to a general result of
Liu, Lorenzini and Raynaud \cite{Liu; Lorenzini; Raynaud 2005}, Theorem 6.6.
The second assertion follows from Theorem \ref{properties genus-one fibrations}.
\qed
\medskip
To proceed, consider for the rational elliptic surface $J\ra\PP^1$ the Frobenius pullback
$$
X'=J^\fpb = J\times_{\PP^1}\PP^1.
$$
According to Proposition \ref{frobenius pullback zariski} below, this is a normal surface, having only isolated
Zariski singularities, and the dualizing sheaf is $\omega_X=\O_X$.
Clearly, the elliptic fibration $X'\ra\PP^1$ admits a section.
Write $\eta\in \PP^1$ for the generic point.
\begin{proposition}
\mylabel{birational sections}
We have $X'_\eta\simeq X_\eta$ if and only if the fibration $f:X\ra\PP^1$ admits
a section.
\end{proposition}
\proof
The condition is obviously necessary. Conversely, suppose that $f:X\ra\PP^1$ admits a section.
Then $X_\eta\simeq\Pic^0_{X_\eta/\eta}$, whence it suffices to check $X'_\eta\simeq\Pic^0_{X_\eta/\eta}$.
Using Proposition \ref{normalization}, we infer $X_\eta=(Y^\fpb)_\eta$.
By definition, $J_\eta=\Pic^0_{Y_\eta/\eta}$, and in turn
$X'_\eta=\Pic^0_{(Y^\fpb)_\eta/\eta}$.
The assertion follows.
\qed
\medskip
We need to understand this condition better, in order to exploit the connection
between the K3-like covering $X$ of the Enriques surface $Y$
and the Frobenius pullback $X'$ of the rational elliptic surface $J$.
Let us call a curve $A\subset Y$ on the Enriques surface a \emph{two-section} if
$A$ is an irreducible curve that is horizontal with respect to the elliptic fibration $\varphi:Y\ra\PP^1$,
of relative degree two. We say that $A\subset Y$ is a \emph{radical two-section}
if the surjection $A\ra\PP^1$ is radical, hence a universal homeomorphism.
\begin{proposition}
\mylabel{two-section = section}
Let $A\subset Y$ be a radical two-section for $\varphi:Y\ra\PP^1$. Then
$B=\epsilon^{-1}(A)_\red$ is a section for $f:X\ra\PP^1$.
Conversely, the image $A=\epsilon(B)$ for any section $B\subset X$
is a radical two-section. Moreover, such $A$ and $B$ exists if and only if $X'_\eta\simeq X_\eta$.
\end{proposition}
\proof
The first assertion is obvious, because $X$ is the normalization of $Y^{(2/\PP^1)}$
and the two morphism $F:\PP^1\ra\PP^1$ and $\varphi:A\ra\PP^1$ coincide generically.
Conversely, suppose that $B\subset X$ is a section. Its image $A=\epsilon(B)$
is an integral Cartier divisor, and the composition
$B\ra A\ra\PP^1$ is radical of degree two.
We thus have either $\deg(B/A)=1$ and $\deg(A/\PP^1)=2$, or $\deg(B/A)=2$ and $\deg(A/\PP^1)=1$.
The latter case is impossible, because the elliptic fibration $\varphi:Y\ra \PP^1$ has no sections.
Thus $A\subset Y$ must be a two-section.
The last statement follows from Proposition \ref{birational sections}.
\qed
\medskip
A curve $A\subset Y$ is called \emph{rational} if it is reduced, and its
normalization is isomorphic to the projective line $\PP^1$.
\begin{proposition}
\mylabel{radical = rational}
A two-section $A\subset Y$ is radical if and only if the curve $A$ is rational.
\end{proposition}
\proof
The condition is clearly necessary. Conversely, suppose that the two-section $A\subset Y$
is rational, and let $\PP^1\ra A$ be the normalization map. The induced $G$-torsor
$X_{\PP^1}=\PP^1\times G$ is trivial. It follows that the Weil divisor $\epsilon^{-1}(A)\subset X$
is non-reduced, thus its reduction $B$ yields a section for $f:X\ra \PP^1$.
According to Proposition \ref{two-section = section}, the image $A=\epsilon(B)$ is a radical two-section.
\qed
\begin{proposition}
\mylabel{two-section = half-fiber}
Suppose there is another elliptic fibration $\varphi':Y\ra\PP^1$ that is orthogonal
to $\varphi:Y\ra\PP^1$, and has a singular half-fiber $C\subset\varphi'^{-1}(b')$.
Then some irreducible component $A\subset C$ is a radical two-section
for $\varphi:Y\ra\PP^1$.
\end{proposition}
\proof
By the classification of singular fibers, each irreducible component $A\subset C$ is a rational curve.
Let $F=\varphi^{-1}(b)$ be a fiber, such that $C\cdot F=2$. Since $F$ is movable,
there is some irreducible component $A\subset C$ with $1\leq A\cdot F\leq 2$.
We must have $A\cdot F=2$, because $\varphi:Y\ra\PP^1$ admits no section.
The assertion now follows form Proposition \ref{radical = rational}.
\qed
\medskip
For the applications we have in mind, it is permissible to replace the chosen
elliptic fibration $\varphi:Y\ra\PP^1$ be another, more suitable one:
\begin{proposition}
\mylabel{radical two-section}
Suppose that the simply-connected Enriques surface $Y$ contains a $(-2)$-curve $A\subset Y$,
or a rational cuspidal curve $C$ that is not movable. Then
there is an elliptic fibration $\varphi:Y\ra\PP^1$ admitting a radical two-section.
\end{proposition}
\proof
According to \cite{Lang 1988}, Theorem A.3 each $(-2)$-curve $A\subset Y$ is
a two-section for some genus-one fibration $\varphi:Y\ra\PP^1$.
Now suppose that there is a non-movable rational cuspidal curve $C\subset Y$.
Then it is a half-fiber of some genus-one fibration $\psi:Y\ra\PP^1$.
According to Theorem \ref{properties genus-one fibrations}, there is an orthogonal fibration $\varphi:Y\ra\PP^1$.
In both cases, the two-sections are radical, by Proposition \ref{radical = rational}.
\qed
\medskip
We do not know if the conditions of the preceding proposition holds for
\emph{all} simply-connected Enriques surfaces. Note that there are Enriques surfaces $Y$ without $(-2)$-curves, according to
\cite{Lang 1983}, Theorem 4.3. In Section \ref{K3-like with edp} we shall give concrete examples for this.
In any case, the exceptions must be very special surfaces:
\begin{proposition}
\mylabel{no radical two-section}
Suppose the simply-connected Enriques surface $Y$ has no elliptic fibration admitting a radical two-section.
Then $Y$ is supersingular, every elliptic fibration $\varphi:Y\ra \PP^1$ has
relative $j$-invariant $j=0$,
every non-multiple singular fiber has Kodaira type $\II$,
and the half-fiber is smooth. Moreover, there is at least one and at most three
non-multiple singular fibers.
\end{proposition}
\proof
In light of Proposition \ref{radical two-section} and Proposition \ref{two-section = half-fiber}, the Enriques surface
$Y$ contains neither $(-2)$-curves nor non-movable rational cuspidal curves.
In turn, only $\I_0$ or $\II$ are possible Kodaira types, and the half-fibers are smooth.
According to Proposition \ref{fibers jacobian}, the Kodaira types of the fibers $Y_b$ and $J_b$ coincide, for all closed points
$b\in \PP^1$.
Looking at Lang's classification \cite{Lang 2000} of rational elliptic surfaces,
we see that there is at least one and at most three singular fibers of type $\II$ on $J$.
Furthermore, the relative $j$-invariant for $J\ra\PP^1$ is $j=0$.
Let $E\subset Y_b$ be a half-fiber.
We claim that also $j(E)=0$. Suppose this would not be the case, such that the elliptic curve $E$ is ordinary.
Choose some prime $l\neq p$. Then there is an invertible sheaf $\shL_E$ on $E$ of order $l$ in $\Pic(E)$.
Consider the exact sequence
$$
H^1(E,\O_E(E))\lra \Pic(2E)\lra \Pic(E) \lra H^2(E,\O_E(E)).
$$
The term on the right vanishes, and the group on the left is annihilated by $p=2$.
Thus we may extend $\shL_E$ to an invertible sheaf $\shL_{2E}$ on the fiber $Y_b=2E$, of the same order $l\neq 2$ in $\Pic(2E)$.
Continuing in this way and using Grothendieck's Existence Theorem, we get an invertible sheaf
$\shL$ on the $\formal Y=Y\otimes_{\PP^1}\Spec(\O_{\PP^1_b}^\wedge)$ of order $l$.
It is not isomorphic to $\O_{\formal Y}(iY_b)$, because the latter is trivial or has order $p=2$.
It follows that the restriction of $\shL$ to the
generic fiber $\formal Y\smallsetminus \formal Y_b$
is an invertible sheaf of order $l\neq 2$
on an elliptic curve with $j$-invariant $j=0$, contradiction.
Summing up, we have $j(E)=0$, and the elliptic curve $E$ is supersingular.
Thus $\Pic(E)$ contains no elements of order two.
Then the conormal sheaf $\O_E(-E)$, whose order divides two, must be trivial, so
the fiber is wild. According to \cite{Cossec; Dolgachev 1989}, Theorem 5.7.2 the simply-connected Enriques surface $Y$
is not classical, hence it must be supersingular.
\qed
\medskip
Let us now assume that $\varphi:Y\ra\PP^1$ admits a radical two-section.
Then we have an isomorphism $X_\eta=X'_\eta$, an in particular
the normal surfaces $X$ and $X'$, which both have trivial canonical class,
are birational. Using terminology from the \emph{minimal model program} in dimension three and higher,
one may regard $X$ as a \emph{flop} of $X'$, because their canonical neither became more negative
nor more positive.
Choose a common resolution of singularities $X'\stackrel{r'}{\leftarrow} S\stackrel{r}{\rightarrow} X$.
The induced projection $h:S\ra X'\ra\PP^1$ is an elliptic fibration on the smooth surface $S$,
and we write $S\ra J'$ for the relative minimal model. Locally, $J'$ is obtained
by applying the Tate Algorithm to the Frobenius base-change of the Weierstra\ss{} equation for $J$.
Summing up, we have the following commutative diagram:
$$
\xymatrix{
& & S\ar[ddll]\ar[dd]_{r'}\ar[dr]^r \\
& & & X\ar[rr]^\epsilon\ar[dd]_(.3)f & & Y\ar[dd]^\varphi\\
J' & & X'\ar[dr]\ar[rr]|\hole & & J\ar[dr] & \\
& & & \PP^1\ar[rr]_F & & \PP^1
}
$$
If $X$ and $X'$ have only rational singularities, then their minimal resolution of singularities
coincides with the relatively minimal model of $X_\eta\simeq X_\eta'$, whence both
are given by $S$. It follows that $S$ is a K3-surface, and
the morphism $S\ra J'$ is
an isomorphism. In this case, $X'\leftarrow S\rightarrow X$ are both crepant resolutions.
On the other hand, if $X$ has an elliptic singularity, such that $S$ and hence $J'$ is a rational surface,
we have $K_{J'}=-F$ for some fiber $F\subset J'$. Now one should regard $J'$ as
a \emph{flip} of both $X'$ and $X$, because the canonical class became more negative.
We observe:
\begin{lemma}
\mylabel{dual graph}
For all closed points $a\in \PP^1$, with Frobenius image $b=F(a)\in\PP^1$,
the four curves $Y_b$, $X_a$, $X'_a$ and $J_b$ have the same dual graph.
\end{lemma}
\proof
The curves $X_a$ and $Y_b$ have the same dual graph, because the K3-like covering $\epsilon:X\ra Y$ is
a universal homeomorphism. The same argument applies to the curves $X'_a$ and $J_b$.
Finally, the curves $Y_b$ and $J_b$ have the same dual graph
according \cite{Liu; Lorenzini; Raynaud 2005}, Theorem 6.6.
\qed
\medskip
The singularities on the K3-like covering $X$ of the Enriques surface $Y$ are closely related to the
singularities on the Frobenius pull-back $X'$ of the rational elliptic surface $J$,
at least outside the multiple fibers:
\begin{proposition}
\mylabel{formal isomorphism non-multiple}
For each closed point $a\in \PP^1$ whose Frobenius image $b\in\PP^1$ has
a non-multiple fiber $Y_b$, there is an isomorphism of two-dimensional schemes
$$
X\times_{\PP^1} \Spec(\O_{\PP^1,a}^\wedge) \simeq X'\o\times_{\PP^1}\Spec(\O_{\PP^1,a}^\wedge).
$$
\end{proposition}
\proof
Write $R=\O_{\PP^1,b}^\wedge$ for the complete local ring of the point $b\in\PP^1$.
Since $Y_b$ is non-multiple, the induced projection $Y\times_{\PP^1}\Spec(R)\ra\Spec(R)$
admits a section, which in turn
induces an identification $Y\times_{\PP^1}\Spec(R) = J\times_{\PP^1}\Spec(R)$.
In light of Proposition \ref{normalization}, taking the fiber product with the Frobenius morphism $F:\PP^1\ra\PP^1$
yields the assertion.
\qed
\section{Ogg's Formula and Frobenius base-change}
\mylabel{Ogg's formula}
In this section let let $k$ be an algebraically closed ground field of arbitrary characteristic $p>0$,
and $J\ra B$ be a smooth elliptic surface that is jacobian and relatively minimal.
For simplicity, we assume that $B$ is a smooth proper connected curve, such that $J$ is a smooth proper connected surface,
although the analysis also applies to local situations as well.
Let us write $B^{(p)}$ for the scheme $B$, endowed with the new structure morphism
$B^{(p)}\ra\Spec(k)\stackrel{F}{\ra}\Spec(k)$, obtained by transport of structure with
the absolute Frobenius of $k$. Then the absolute Frobenius on $B$ becomes the \emph{relative
Frobenius morphism} $B^{(p)}\ra B$. The cartesian square
$$
\begin{CD}
J^{(p/B)} @>>> J\\
@VVV @VVV\\
B^{(p)} @>>> B\\
\end{CD}
$$
defines an integral proper connected surface $J^{(p/B)}$ endowed with an elliptic fibration.
We call $J^{(p/B)}$ the \emph{Frobenius pullback}. Let us first record:
\begin{lemma}
\mylabel{frobenius pullback zariski}
The singularities on the Frobenius pullback $J^{(p/B)}$ are Zariski singularities.
\end{lemma}
\proof
Fix a closed point $x\in J^{(p/B)}$, and write $a\in B^{(p/B)}$, $y\in J$ and $b\in B$ for its images.
Choose a uniformizer $\pi\in\O_{B,b}$, and write $z=\pi$ for the resulting uniformizer $z\in\O_{B^{(p)},a}$.
Then the morphism of complete local $k$-algebras
$$
k[[z]]\lra \O_{B^{(p)},b}^\wedge,\quad \sum\lambda_iz^i\longmapsto\sum\lambda_i^pz^i
$$
is bijective. Using this identification, we may regard the extension $\O_{B,b}^\wedge\subset\O_{B^{(p)},a}^\wedge$
as $k[[\pi]]\subset k[[\pi,z]]/(z^p-\pi)$.
In turn, we get
$$
\O_{J^{(p/B)},x}^\wedge = \O_{J,y}^\wedge[z]/[z^p-\pi],
$$
where the unifomizer $\pi\in\O_{B,b}^\wedge$ becomes a non-unit $\pi\in \O_{J,y}^\wedge$.
Since the local ring $\O_{J,y}$ is regular, the local ring $\O_{J^{(p/B)},x}^\wedge$ is a Zariski singularity.
\qed
\medskip
Let $S\ra J^{(p/B)}$ be a resolution of singularities, which inherits an elliptic fibration
$S\ra B$, and write $S\ra J'$ for the
contraction to the relative minimal model. Then we have a new
smooth surface $J'$ endowed with an elliptic fibration $J'\ra B^{(p)}$, which is jacobian and relatively minimal.
We call $J'$ the \emph{smooth Frobenius pullback}.
Note that in general, the Kodaira dimension of the surfaces $J$ and $J'$ are different.
Fix a section $O\subset J$ for the jacobian elliptic fibration $J\ra B$, and let $W\ra\PP^1$
be the ensuing Weierstra\ss{} fibration. The normal surface $W$ is obtained by contracting
all vertical irreducible curves disjoint to the section.
Locally at each point $b\in B$,
it is given by some minimal Weierstra\ss{} equation
$$
y^2+a_1xy+a_3y = x^3+a_2x^2+a_4x+a_6 \qquad a_i\in\O_{B,b}.
$$
We denote by $v_b\geq 0$ the \emph{valuation of the discriminant} $\Delta_b\in\O_{B,b}$
for any such minimal Weierstra\ss{} equation,
by $m_b\geq 1$ the \emph{number of irreducible components} in the fiber $J_b$,
and by $\delta_b\geq 0$ the \emph{wild part of the
conductor} for the Galois representation on the geometric generic $l$-torsion points.
Here $l\neq p$ is any prime different from the characteristic.
If the fiber $J_b$ is of semistable, that is of type $\I_m$, then $\delta_b=0$ and $v_b=m$.
If the fiber is unstable, Ogg's Formula \cite{Ogg 1967} gives
\begin{equation}
\label{ogg formula}
v_b = 2+\delta_b + (m_b-1).
\end{equation}
The closed point $b\in B^{(p)}$ may also be seen as a closed point on the scheme $B^{(p)}$.
To avoid confusion, we denote it by $a\in B^{(p)}$.
This gives numerical invariants $v_a,\delta_a,m_a$ for the smooth Frobenius pullback $J'\ra B$.
Now observe that we may obtain $J'$ from the base-changed Weierstra\ss{} fibration
$W^{(p/B)}$ as follows: Run the Tate Algorithm \cite{Tate 1972}
until the Weierstra\ss{} equation becomes minimal and then do the minimal resolution of singularities.
This gives another invariant: The \emph{length of the Tate Algorithm}
$\lambda_a\geq 0$, that is, the number of repetitions that are necessary to finish the Tate Algorithm.
The case $\lambda_a=0$ means that the Weierstra\ss{} equation is already minimal.
If the fiber is semistable, we have $\lambda_a=0$ and $m_a=pm_b$.
Indeed, a local computation shows that
on $W^{(p/B)}$ we get rational double points of type $A_{p-1}$ lying over the
singular points of the semistable fiber $J_b$.
The situation is more interesting at the unstable fibers:
\begin{lemma}
\mylabel{unstable fiber}
Suppose the fiber $J_b$ is unstable. Then $J'_a$ is unstable as well,
and its numerical invariants are given by the formulas
$$
v_a=pv_b-12\lambda_a,\quad \delta_a=\delta_b\quadand m_a=pm_b+(p-1)(\delta_b+1) -12\lambda_a.
$$
In characteristic two, this means $v_a=2v_b-12\lambda_a$ and $m_a=2m_b+\delta_b+1-12\lambda_a$.
\end{lemma}
\proof
Among all function field extensions that achieve semistable reduction, there
is a smallest one, and this smallest one is a Galois extension, which follows from the N\'eron--Ogg--Shafarevich Criterion
(see for example \cite{Bosch; Luetkebohmert; Raynaud 1990}, Section 7.4, Theorem 5).
In particular, unstable fibers stay unstable under purely inseparable field extensions.
Each round of the Tate Algorithm reduces the valuation of the discriminant by 12.
The wild part of the conductor depends on a Galois representation, and is thus not
affected by purely inseparable extensions. The equation for the number $m_a\geq 1$ of irreducible components
is now a consequence of Ogg's Formula \eqref{ogg formula}.
\qed
\medskip
By construction, the Weierstra\ss{} model $W$ is normal, and the fibers
$W\ra B$ contain at most one non-smooth point, which is a rational double point.
In turn, the base-change $W^{(p/B)}$ is normal, and contains fiber-wise at most
one singularity.
\begin{proposition}
\mylabel{singularity rational}
We have $\lambda_a>0$ if and only if the normal surface $W^{(p/B)}$ contains a non-rational singularity
lying over the point $a\in B^{(p)}$.
\end{proposition}
\proof
There is nothing to prove if $W^{(p/B)}$ is smooth along the fiber. Now suppose that a singularity
$x\in W^{(p/B)}$ is present, such that the fiber $C=W_a^{(p/B)}$ is a rational cuspidal curve.
Let $S\ra W^{(p/B)}$ be the minimal resolution of this singularity.
It the singularity is rational, then $S$ is the smooth model at $a\in B^{(b)}$,
thus the Weierstra\ss{} equation was minimal, whence $\lambda_a=0$.
Conversely, suppose that the singularity is non-rational. Then $W^{(p/B)}$
cannot be a Weierstra\ss{} model, so the Weierstra\ss{} equation was not minimal,
that is, $\lambda_a>0$.
\qed
\begin{proposition}
\mylabel{rational elliptic}
If both $J$ and $J'$ are rational surfaces, then we have $\sum\lambda_a=p-1$,
where the sum runs over all closed points $a\in B^{(p)}=\PP^1$.
\end{proposition}
\proof
Let us write $v=\sum v_b$, $v'=\sum v_a$ and $\lambda=\sum\lambda_a$ for the sums of local invariants.
Since $J$ and $J'$ are rational, $v=v'=12$ must hold, according to \cite{Lang 2000},
Lemma 0.1. On the other hand, we have $v'=pv-12\tau$, which forces $\tau=p-1$.
\qed
\medskip
In particular, in characteristic $p=2$ there is precisely one fiber that
does not stay minimal, and it becomes minimal after one round of the Tate Algorithm.
We shall analyze this particular situation in the next section.
\section{Geometric interpretation of the Tate Algorithm}
\mylabel{Tate algorithm}
The \emph{Tate Algorithm} turns arbitrary Weierstra\ss{} equations over discrete valuation rings
into minimal Weierstra\ss{} equations \cite{Tate 1972}. This algorithm is of paramount importance
for the arithmetic theory of elliptic curves, as well as the structure of elliptic surface.
The goal of this section
is to describe the geometry behind it. This is perhaps well-known, but we could
not find a suitable reference.
Let $R$ be a discrete valuation ring with residue field $k=R/\maxid_R$ and field of
fractions $F=\Frac(R)$, and choose a uniformizer $\pi\in R$.
Suppose we have a Weierstra\ss{} equation
$$
y^2 +a_1xy + a_3y = x^3 + a_2x^2 + a_4x + a_6
$$
with coefficients $a_i\in R$ and discriminant $\Delta\neq 0$.
Setting $y=X/Z$, $x=X/Z$ and multiplying with $Z^3$ gives its homogenization
$$
Y^2Z + a_1XYZ + a_3YZ^2 = X^3 + a_2 X^2Z + a_4XZ^2 + a_6Z^3,
$$
which defines a relative cubic $J\subset \PP^2_R=\Proj R[X,Y,Z]$
endowed with a section $O\subset J$ given by the equations $X=Z=0$.
\begin{theorem}
\mylabel{tate geometric}
Let $S\ra J$ be the blowing-up with center given by $\pi=X=Z=0$,
and $S\ra J'$ be the blowing-down of the strict transform of the closed fiber $J_0$.
Then $J'\ra\Spec(R)$ is a relative cubic given by the Weierstra\ss{} equation
\begin{equation}
\label{nonminimal Weierstrass}
y^2 + a_1'xy + a'_3y = x^3 + a_2'x^2 + a_4'x + a'_6,
\end{equation}
whose coefficients are $a'_i=a_i\pi^i$.
\end{theorem}
\proof
Write the closed fiber of the blowing-up as $S_0=C\cup E$, where $C$ is the strict transform of $J_0$,
and $E$ is the $(-1)$-curve.
In light of the exact sequence
$$
1\lra \O_{S_0}^\times\lra \O_C^\times\oplus\O_E^\times\lra \kappa(s)^\times\lra 1,
$$
where $s\in C\cap E$ is the intersection point,
we easily see that the restriction map $\Pic(S_0)\ra\Pic(C)\oplus\Pic(E)$
is injective. Using Riemann--Roch and the Theorem on Formal Functions, one infers that every
invertible sheaf $\shL\in\Pic(S)$ with $(\shL\cdot C)=2$ and $(\shL\cdot E)=1$
is semiample, with $R^1h_*(\shL)=0$ and $h_*(\shL)$ free of rank three.
We thus get a morphisms $S\ra \PP^2_R$, whose image is another relative cubic $X\subset\PP^2_R$.
The morphism $S\ra X$ induces an an open embedding on the complement of $C$,
sends $E$ isomorphically onto a line, and maps $C$ to a double line.
Me may compute the cubic equation for $X\subset\PP^2_R$ by regarding $X$ as
a closure of some affine cubic.
Consider first the affine open subset
$D_+(Y)\subset\PP^2_R$. Here the relative cubic $J$ is given by the dehomogenized cubic equation
$$
\frac{Z}{Y} +
a_1\frac{X}{Y}\frac{Z}{Y} +
a_3\left(\frac{Z}{Y}\right)^2
=
\left(\frac{X}{Y}\right)^3 +
a_2 \left(\frac{X}{Y}\right)^2\frac{Z}{Y} +
a_4\frac{X}{Y}\left(\frac{Z}{Y}\right)^2 +
a_6\left(\frac{Z}{Y}\right)^3.
$$
The blowing-up $S\ra J$ is covered by
two affine charts, and we look at the $\pi$-chart given by the indeterminates $ X/\pi Y,Z/\pi Y$ over $R$,
subject to the relation
\begin{multline*}
\frac{Z}{\pi Y} +
a_1\pi\frac{X}{\pi Y}\frac{Z}{\pi Y} +
a_3\pi\left(\frac{Z}{\pi Y}\right)^2
=\\
\pi^2\left(\frac{X}{\pi Y}\right)^3 +
a_2\pi^2\left(\frac{X}{\pi Y}\right)^2 \frac{Z}{\pi Y} +
a_4\pi^2 \frac{X}{\pi Y} \left(\frac{Z}{\pi Y}\right)^2 +
a_6\pi^2\left(\frac{Z}{\pi Y}\right)^3.
\end{multline*}
Via the substitutions $X/\pi Y=U/V $ and $Z/\pi Y=W/V$, we may view this
as a closed subscheme inside $D_+(V)\subset \PP^2_R=\Proj R[U,V,W]$.
Its closure in $\PP^2_R$ is given by the homogeneous cubic equation
\begin{equation}
\label{homogeneous cubic}
V^2W + a_1\pi UVW + a_2\pi VW^2 = \pi^2 U^3 + \pi^2a_2U^2W + \pi^2a_4UW^2 + \pi^2a_6W^3.
\end{equation}
The closed fiber is thus given $\pi=V^2W=0$, which is the union of the line $W=0$
and the double line $V^2=0$ in $\PP^2_k$. Indeed, since $S\smallsetminus C$ coincides with the $\pi$-chart of the
blowing-up, and $S\smallsetminus C \ra X$ is an open embedding, the above closure equals $X$.
Now consider the automorphism
$$
f=\begin{pmatrix}
\pi^{-2} \\
& \pi^{-2}\\
&& 1
\end{pmatrix}
\in\PGL_3(F) =\Aut(\PP^2_F).
$$
Over the function field, it gives a new cubic $f(J_F)\subset\PP^2_F$. Applying $f$
to the cubic equation \eqref{homogeneous cubic} and multiplying by $\pi^4$, we get the homogeneous equation
$$
V^2W + a_1'UVW + a_3'VW = U^3 + a_2'U^2W + a_4'UW^2 + a_6'W^3,
$$
where $a_i'=\pi^ia_i$. Dehomogenizing the above homogeneous equation with $x=U/W$ and $y=V/W$
gives the desired Weierstra\ss{} equation \eqref{nonminimal Weierstrass}.
Thus the relative cubic $J'\subset \PP^2_R$ defined by the above homogeneous equation
is the schematic closure of $f(J_F)\subset\PP^2_F$.
Note that the closed fiber $J'_k$ is a rational cuspidal curve,
whence $J'$ is normal.
Clearly, the rational map $f:\PP^2_R\dashrightarrow \PP^2_R$ is defined on the generic fiber $\PP^2_K$
and the standard affine open subsets $D_+(U)$ and $D_+(V)$, whence it is defined everywhere
except at the point $z=(0:0:1)\in\PP^2_k$.
In turn, the rational map $f:J\dashrightarrow J'$ is defined outside $z\in J$.
The latter is nothing but the
intersection of the line $\pi=W=0$ and the double line $\pi=V^2=0$ on the closed fiber.
Every section $(a:b:0)$ for $J\ra\Spec(R)$ is mapped to itself under the rational
map $f:\PP^2_R\dashrightarrow\PP^2_R$, whence $f:J\dashrightarrow J'$ is a bijective morphism on the complement
of the double line $\pi=V^2=0$. Now choose a modification $S'\ra J$ with center $z\in J$ so that
the rational map $J\dashrightarrow J'$ comes from a morphism $S'\ra J'$.
Let $E\subset S'_0$ be an irreducible component different from the strict transform
of the line $\pi=W=0$. By Zariski's Main Theorem, it cannot map to a regular point on $J'_0$, because
these are already images of points form the line $\pi=W=0$.
In turn, it must map to the singular point of the rational cuspidal curve $J'_0$.
Thus the morphism $S'\ra J'$ factors over $J$, and furthermore the induced morphism $J\ra J'$
contracts the double line $\pi=V^2=0$.
It follows that $S\ra J'$ is the contraction of $C\subset S$.
\qed
\medskip
This observation gives some information on elliptic singularities, which turns out very useful.
Let us call a Weierstra\ss{} equation
\begin{equation}
\label{almost minimal equation}
y^2 +a'_1xy + a'_3y = x^3 + a'_2x^2 + a'_4x + a'_6
\end{equation}
\emph{almost minimal} if its discriminant $\Delta\in R$ is nonzero,
and the Weierstra\ss{} equation becomes minimal after one round in the Tate Algorithm.
\begin{proposition}
\mylabel{almost minimal}
Every almost minimal Weierstra\ss{} equation
defines an elliptic singularity, for which the exceptional divisor on the minimal resolution
of singularities coincides with the closed fiber on the regular minimal model $J\ra\Spec(R)$.
The intersection numbers on the exceptional divisor and the fiber divisor coincide,
except that for one component that has
multiplicity one in the fiber, the intersection number $0$ or $-2$ become
$-1$ respective $-3$ on the exceptional divisor.
\end{proposition}
\proof
Running through the Tate Algorithm, we may assume
that $\pi^i| a'_i$. Now Theorem \ref{tate geometric} describes how the singularity on $J'$
arises from $J$ via the birational correspondence $J\leftarrow S\rightarrow J'$.
The statement on the exceptional divisor and the intersection numbers follow.
It remains to check that the singularity is indeed elliptic.
Write the closed fiber as $S_0=E+C+D$, where $E$ is the $(-1)$-curve,
$C$ is the irreducible component that intersects $E$, and $D$ be the union of the remaining
irreducible components. The contraction $r':S\ra J'$ factors over the contraction $S'\ra J'$ of $D$,
which introduces a rational double points. In light of the Leray--Serre spectral sequence,
it suffices to show that for
the induced map $g:S'\ra J'$ the higher direct image $R^1g_*(\O_{S'})$ has length one.
Changing notation, we write the closed fiber $S'_0=E+C$, where $E=\PP^1$
and $C$ is a rational cuspidal curve. The intersection point $E\cap C$ is smooth in $S'$,
whence both $E,C\subset S'$ are Cartier divisors, with intersection matrix
$\Phi=(\begin{smallmatrix}-1&1\\1&-1\end{smallmatrix})$. By the Theorem on Formal Functions,
we have to show that $h^1(\O_{nC})=1$ for all $n\geq 1$.
This is obvious for $n=1$. We proceed by induction, using the long exact sequence
attached to the short exact sequence
$$
0\lra \O_C(-nC)\lra \O_{nC}\lra \O_{(n-1)C}\lra 0,
$$
together with the fact that $\shL=\O_C(-nC)$ has degree $n\geq 1$, whence $H^1(C,\shL)=0$
by Riemann--Roch.
\qed
\medskip
Note that blowing-ups of elliptic singularities are usually non-normal.
Furthermore, the sequence of normalized blowing-ups of reduced singular points
usually produces a non-minimal resolution. This makes it often so difficult
to compute a resolution of singularities explicitly. The above result
avoids all these complications and makes it unnecessary to deal with
equations.
For our goals, the cases of Kodaira type $\II$ is particularly important:
\begin{corollary}
\mylabel{almost minimal II}
Every almost minimal Weierstra\ss{} equation
of type $\II$ defines an elliptic singularity, for which the exceptional divisor $E$ on the
minimal resolution of singularities is a rational cuspidal curve with intersection matrix
$\Phi=(-1)$.
\end{corollary}
\medskip
This is one of the simplest cases of Wagreich's classification of
elliptic double points \cite{Wagreich 1970}, Theorem 3.5.
It belongs to the family of elliptic double points, whose exceptional divisors
are symbolically described by Wagreich in the following way:
\begin{center}
\begin{tikzpicture}
\node (a) at (1,1) [circle, draw, label=below:{$\scriptstyle{[C]}$}] {};
\node (a') at (1,1) [] {$\scriptscriptstyle{-1}$};
\node (b) at (3,1) [circle, draw] {};
\node (b') at (3.7,1) {};
\node (c') at (4.3,1) {};
\node (c) at (5,1) [circle, draw] {};
\node (d) at (4,1) [below] {$\underbrace{\phantom{mmmmmm}}_{k}$};
\draw (a) -- (b);
\draw [dashed] (b) -- (c);
\draw (b) -- (b');
\draw (c) -- (c');
\end{tikzpicture}
\end{center}
Here chain of length $k\geq 0$ to the right consists of $(-2)$-curves.
Our singularity is the special case $k=0$.
We simply say that it is the \emph{elliptic singularity obtained by contracting
a cuspidal $(-1)$-curve}.
One also describe it
in terms of the minimal good resolution,
where the exceptional divisor has simple normal crossings.
It is obtained from the minimal resolution $S$ by three further
blowing-ups, such that $E$ has four irreducible components, which
are smooth rational curves. The dual graph has the form
\begin{center}
\begin{tikzpicture}
[node distance = 2cm]
\node (a) at (1,1) [circle, draw, label=below:$-2$] {};
\node (b) [right of=a] [circle, draw, label=below:$-1$] {};
\node (c) [right of=b] [circle, draw, label=below:$-3$] {};
\node (d) [above of=b] [below=.5cm, circle, draw, label=left:$-6$] {};
\draw (a)--(b)-- (c);
\draw (b) -- (d);
\end{tikzpicture}
\end{center}
\section{Frobenius pullback of rational elliptic surfaces}
\mylabel{Frobenius pullback}
Let $k$ be an algebraically closed ground field of arbitrary characteristic $p>0$.
Suppose that $f:J\ra\PP^1$ is a rational elliptic surface that is relatively minimal and jacobian.
In this section, we examine the Frobenius pullback $J^{(p/\PP^1)}$, which is defined by the cartesian diagram
$$
\begin{CD}
J^{(p/\PP^1)} @>>> J\\
@VgVV @VVfV\\
\PP^1 @>>F> \PP^1.
\end{CD}
$$
Here we collect some elementary general facts, although eventually we are mainly interested in characteristic two.
First note that the singularities on $J^{(p/B)}$ are Zariski, according to Proposition \ref{frobenius pullback zariski}.
In particular, the Frobenius pullback is locally of complete intersection, in particular Gorenstein,
and its tangent sheaf is locally free.
Let us write $\O_J(n)$ and $\O_{J^{(p/\PP^1)}}(n)$ for the invertible sheaves $f^*\O_{\PP^1}(n)$ and $g^*\O_{\PP^1}(n)$,
respectively.
\begin{proposition}
\mylabel{dualizing sheaf}
The dualizing sheaf is $\omega_{J^{(p/\PP^1)}}=\O_{J^{(p/\PP^1)}}(p-2)$.
\end{proposition}
\proof
By assumption, $f:J\ra\PP^1$ is a rational elliptic surface that
is relatively minimal and jacobian.
According to \cite{Cossec; Dolgachev 1989}, Proposition 5.6.1 we have $\omega_J=\O_J(-1)$.
Using $\omega_{\PP^1}=\O_{\PP^1}(-2)$
one gets $\omega_{J/\PP^1}=\O_J(1)$. Relative dualizing sheaves commute with
flat base change.
It follows that the relative dualizing sheaf
for $J^{(p/\PP^1)}\ra\PP^1$ is the preimage of $F^*\O_{\PP^1}(1)=\O_{\PP^1}(p)$,
and the result follows.
\qed
\begin{proposition}
\mylabel{normal reduced fibers}
The singular locus on the surface $J^{(p/\PP^1)}$ corresponds
to the locus on the surface $J$ where the fibration
$J\ra\PP^1$ is not smooth. In particular, $J^{(p/\PP^1)}$ is normal
if and only if $J\ra\PP^1$ has only reduced fibers.
\end{proposition}
\proof
Clearly, $ J^{(p/\PP^1)}$ is smooth over each $y\in J$ at which the fiber $F=J_{f(y)}$ is smooth.
Now let $y\in J$ be a closed point at which the fiber $F$ has embedding dimension $\edim(\O_{F,y})\geq 2$.
The preimage of the fiber in $J^{(p/\PP^1)}$ is a scheme isomorphic to $F'=F\otimes_kk[T]/(T^p)$,
which has $\edim(\O_{F',x})\geq 3$. In turn, the Frobenius pullback acquires a singularity there.
\qed
\begin{proposition}
\mylabel{cohomology vanishes}
We have $H^1(J^{(p/\PP^1)}, \O_{J^{(p/\PP^1)}}) =0$.
\end{proposition}
\proof
The Leray--Serre spectral sequence for $g:J^{(p/\PP^1)}\ra\PP^1$ gives an exact sequence
\begin{equation}
\label{cohomology exact sequence}
H^1(\PP^1,\O_{\PP^1})\lra H^1(J^{(p/\PP^1)}, \O_{J^{(p/\PP^1)}}) \lra H^1(J^{(p/\PP^1)}, R^1g_*\O_{J^{(p/\PP^1)}}),
\end{equation}
and the term on the left vanishes.
Forming higher direct images commute with flat base-change.
So to understand the term on the right, we merely have to compute $R^1f_*(\O_J)$.
Choose a section $O\subset J$. Using $\omega_J=\O_J(-1)$ and Adjunction Formula, one gets
$(O\cdot O)=-1$. The short exact sequence $0\ra\O_J\ra\O_J(O)\ra\O_{\PP^1}(-1)\ra 0$
induces an exact sequence
$$
\O_{\PP^1}(-1)\lra R^1f_*(\O_J)\lra R^1f_*(\O_J).
$$
The terms on the left is invertible, the term in the middle has rank one,
and the term on the right vanishes, consequently $R^1f_*(\O_J)=\O_{\PP^1}(-1)$.
Thus $R^1g_*\O_{J^{(p/\PP^1)}}=\O_{\PP^1}(-p)$, whose first cohomology vanishes.
The assertion thus follows from the exact sequence \eqref{cohomology exact sequence}.
\qed
\medskip
Since $J$ is a rational surface with Picard number $\rho=10$,
there are some \emph{rulings} $r:J\ra\PP^1$, that is, fibrations whose generic fiber
is a projective line.
\begin{proposition}
\mylabel{ruling degree two}
For each closed fiber $D=r^{-1}(t)$ of some ruling, the induced
map $f:D\ra\PP^1$ has degree two.
\end{proposition}
\proof
Regarding this degree as an intersection number, we easily reduce to the
case that the fiber $D$ is smooth. Then $D=\PP^1$. Now the Adjunction Formula gives
$-2=(D\cdot\omega_S)=-(D\cdot F)$, and the assertion follows.
\qed
\medskip
Now suppose that the elliptic fibration $f:J\ra\PP^1$ has only reduced fibers, such that $J^{(p/\PP^1)}$ is normal.
Let $r:J\ra\PP^1$ be some ruling. Consider the Stein factorization $J^{(p/\PP^1)}\ra C$ given by
the commutative diagram
$$
\begin{CD}
J^{(p/\PP^1)} @>>> J\\
@VVV @VVrV\\
C @>>> \PP^1.
\end{CD}
$$
\begin{proposition}
\mylabel{fibrations from rulings}
The morphism $C\ra\PP^1$ is an isomorphism, and
the generic fiber of $J^{(p/\PP^1)}\ra C$ is a regular curve $R$ with
$h^1(\O_R)=p-1$. If $p\geq 3$, the geometric generic fiber is a rational curve
with two singularities.
In characteristic two, $J^{(p/\PP^1)}\ra C$ is a quasielliptic fibration.
\end{proposition}
\proof
Let $D=r^{-1}(t)$ be a smooth fiber, thus $D\simeq\PP^1$. The restriction $f:D\ra\PP^1$ is an affine morphism
of degree two, and thus given by the relative spectrum of $\shA=\O_{\PP^1}\oplus\O_{\PP^1}(-1)$.
The relative Frobenius $F:\PP^1\ra\PP^1$ is given by the relative spectrum
of $\shB=\O_{\PP^1}\oplus\O_{\PP^1}(-1)^{\oplus p-1}$.
In turn, the preimage of $D$ in $J^{(2/\PP^1)}$ is the spectrum of $\shC=\shA\otimes\shB$,
and we see $h^0(\shC)=1$ and $h^1(\shC)=p-1$.
For $p$ odd, the double covering $\PP^1=D\ra\PP^1$ is ramified at two points,
so the preimage of $D$ has precisely two singularities.
In light of $h^0(\shC)=1$, the Theorem on Formal Functions
ensures that the composite mapping $J^{(p/\PP^1)}\ra \PP^1$
equals its own Stein factorization, thus $C=\PP^1$.
Its generic fiber is normal, because this holds for $J^{(p/\PP^1)}$,
and its arithmetic genus equals $h^1(\shC)=p-1$.
In characteristic $p=2$, this means that the fibration is quasielliptic.
In odd characteristics, the non-smooth locus of the generic fiber for $J^{(p/\PP^1)}\ra C$ consist of two points,
because this holds for almost all closed fibers.
\qed
\section{Lang's classification and Frobenius pullback}
\mylabel{Lang classification}
For rational elliptic surfaces over the complex numbers, the possible configurations of Kodaira types were classified by
Persson \cite{Persson 1990} and Miranda \cite{Miranda 1990}.
This classification was extended by Lang \cite{Lang 2000} to characteristic $p=2$,
including for each case examples of global Weierstra\ss{} equations. These equations show that
all numerically possible cases indeed exist. This builds on the normal forms for
unstable fibers obtained in \cite{Lang 1994}, Section 2A.
Lang's results are the key
to understand and construct simply-connected Enriques surfaces and their K3-like coverings.
He introduced very useful short-hand notation for configurations of Kodaira types:
For example, $\II + 21^6$ stands for the configuration comprising one fiber of type $\II$, one fiber of type $\I_2$,
and six fibers of type $\I_1$.
It turns out that there are 147 possible configurations of Kodaira types. Here we are only interested in
cases where all fibers are reduced, for which there are 110 configurations.
There are 35 cases where all fibers of $J$ are semistable, 58 cases with semistable fibers and
precisely one unstable fiber, and 17 cases with only unstable fibers.
The latter is equivalent to the condition that the global $j$-invariant is $j(J/\PP^1)=0$.
Lang used alpha-numerical symbols for
normal forms of \emph{local} Weierstra\ss{} equations for unstable fibers.
Let us call them the \emph{Lang types}. One may tabulate the Lang types for reduced fibers as follows:
$$
\begin{array}[t]
{l | c c c | c c | c | c c c | c c c | c}
\text{Lang type} & {\rm 1A}& {\rm 1B}& {\rm 1C}& {\rm 2A}& {\rm 2B}& {\rm 3}& {\rm 9A} & {\rm 9B}& {\rm 9C}& {\rm 10A} & {\rm 10B}& {\rm 10C}& {\rm 11}\\
\hline &&&&&&&&&&&&&\\[-2ex]
\text{Kodaira type} & \II & \II & \II & \III & \III & \IV & \II & \II & \II & \III & \III & \III & \IV\\
m & 1 & 1 & 1 & 2 & 2 & 3 & 1 & 1 & 1 & 2 & 2 & 2 & 3\\
v & 4 & 6 & 7 & 4 & 6 & 4 & 4 & 8 & 12 & 4 & 8 & 12 & 4\\
\delta & 2 & 4 & 5 & 1 & 3 & 0 & 2 & 6 & 10 & 1 & 5 & 9 & 0\\
\end{array}
$$
\vspace{3em}
\centerline{Table \stepcounter{table}\arabic{table}: \text{The reduced unstable Lang types}}
\vspace{-3.5em}
\noindent
Here $m\geq 1$ is the number of irreducible components in the fiber,
$v\geq 1$ is the valuation of a minimal Weierstra\ss{} equation,
and $\delta $ is the wild part of the conductor, which satisfies $\delta= v -m-1$ according
to Ogg's Formula.
In what follows, we work over an algebraically closed ground field $k$ of characteristic $p=2$.
Fix a rational elliptic surface $J\ra \PP^1$ that is jacobian, with only reduced fibers.
Hence the Frobenius pullback $X'=J^\fpb$ is normal.
Choose a section $O\subset J$, and let $W\ra\PP^1$ be the resulting Weierstra\ss{} fibration.
In what follows, fibers of Lang type 9C play an exceptional role.
They have Kodaira type $\II$ and numerical invariants $v=12$, $m=1$ and $\delta=10$.
If present, there are no other singular fibers, hence the contraction $J\ra W$ to
the Weierstra\ss{} model is an isomorphism, and the whole of $J$ is
given by the global Weierstra\ss{} equation
\begin{equation}
\label{9C}
y^2+t^3\gamma_0 y = x^3+t\gamma_1x^2+ t\gamma_3x + t\gamma_5,
\end{equation}
according to \cite{Lang 1994}, Section 2A. Here $\gamma_i\in k[t]$ are polynomials of degree $\leq i$, and furthermore
$t\nmid\gamma_0,\gamma_5$. We regard the base of the fibration as the projective line
$\PP^1=\Spec k[t]\cup\Spec k[t^{-1}]$. The singular fiber is located over $t=0$.
\begin{theorem}
\mylabel{nonrational singularity}
The Frobenius pullback $X'=J^\fpb$ contains a non-rational singularity $x_1'\in X'$ if and only
if $J\ra\PP^1$ contains a singular fiber of Lang type {\rm 9C}, and the coefficient
$\gamma_3\in k[t]$ in the above Weierstra\ss{} equation is divisible by $t\in k[t]$.
In this case, the singularity is an elliptic double point,
the exceptional divisor on the minimal resolution
of singularities is a rational cuspidal curve with intersection matrix $\Phi=(-1)$,
and there are no other singularities on $X'$.
\end{theorem}
\proof
We start by collecting some general observations.
Clearly a singularity $x\in J^\fpb$ is rational if
the local Weierstra\ss{} equations for the corresponding fiber in the Weierstra\ss{}
model $W\ra\PP^1$ of $J\ra\PP^1$ remains minimal after Frobenius pullback.
This automatically holds for the semistable fibers.
Let $J_a\subset J$ be an unstable fiber.
The invariants for the minimal model of $J^{(2/\PP^1)}$ satisfy $2v-12\lambda = 2+\delta + (m'-1)$,
where $\lambda\geq 0$ is the number of repetitions in the Tate Algorithm, as explained
in Proposition \ref{unstable fiber}. Thus
$12\lambda\leq 2v-1-\delta$.
The right hand side is $<12$, thus $\lambda=0$ and the corresponding singularities
on $X'$ are rational for all types, except for 9C and 10C,
by the table of Lang types.
According to \cite{Lang 2000}, Lemma 0.1, we have $\sum_{a\in\PP^1}v_a=12$.
Fibers of Lang type 9C and 10C have $v_a=12$, whence $J\ra\PP^1$ has exactly one
singular fiber. Their Kodaira types are $\II$ or $\III$, whence the projection $J\ra\PP^1$
is smooth outside a single point, so $\Sing(X')$ contains exactly one point.
To verify the equivalence of the conditions, it suffices to treat these two special cases.
Let $x'_1\in X'$ be the unique singularity. Note that if this singularity is rational,
it must be a rational double point of type $D_{12}$, and the corresponding singular fiber on the minimal model
has Kodaira type $\I_8^*$.
Suppose first that $J\ra\PP^1$ contains a fiber of Lang type 10C.
By \cite{Lang 2000}, the Weierstra\ss{} equation is of the form
\begin{equation}
\label{10C}
y^2+t^3\gamma_0 y = x^3+t\gamma_1x^2+ t\gamma_3x + t^2\gamma_4,
\end{equation}
where $\gamma_i\in k[t]$ are polynomials of degree $\leq i$, subject to
$t\nmid\gamma_0,\gamma_3$. The singular fiber is again located over $t=0$.
Going through the Tate Algorithm and using $t\nmid\gamma_3$, we see that the
Frobenius pullback of the Weierstra\ss{} equation remains minimal, and
the singular fiber on $X'$ has Kodaira type $\I_n^*$ for some $n\geq 0$.
We have $v_a=2v_b=24$ and $\delta_a=\delta_b=9$, so Ogg's Formula $v_a=2+\delta_a+(m_a-1)$
gives $m_a=14$, thus the fiber is of type $I_9^*$.
Now suppose that $J\ra\PP^1$ contains a fiber of Lang type 9C
as in \eqref{9C}. Again going through the Tate Algorithm, we see that the
the singular fiber on $X'$ has Kodaira type $\I_n^*$ for some $n\geq 0$ if and only if $t\nmid \gamma_3$.
We then argue as in the preceding paragraph to
deduce that the fiber is of Kodaira type $I_8^*$.
This establishes the equivalence of the two conditions.
Now suppose that there is a non-rational singularity $x'_1\in X'$, hence $J\ra\PP^1$ contains
a fiber of Lang type 9C. The local Weierstra\ss{} equation for $X'$ must have $v'=2\cdot 12-12$,
whence is almost minimal, and the singularity is elliptic, by Proposition \ref{almost minimal}.
The minimal model of $X'$ has $v'=2\cdot v-12=12$. Ogg's Formula $12=2+(m'-1)+10$ yields
$m'=1$, so the Kodaira type is $\II$.
The statement on the exceptional divisor now follows from Corollary \ref{almost minimal II}.
\qed
\medskip
We now turn to the global properties:
Let $r':S\ra X'$ be the minimal resolution of singularities, and write
$h:S\ra \PP^1$ for the induced elliptic fibration.
\begin{lemma}
The smooth surface $S$ is either a K3 surface or a rational surface.
The latter holds if and only if $X'$ contains a non-rational singularity.
\end{lemma}
\proof
According to Proposition \ref{dualizing sheaf} and Proposition \ref{cohomology vanishes},
we have $\omega_{X'}=\O_{X'}$ and $h^1(\O_{X'})=0$.
Now suppose that $X'$ has only rational singularities. For the minimal resolution of singularities,
this means $\omega_S=\O_S$ and $h^1(\O_S)=0$. Furthermore, the rational elliptic surface $J$
has Betti number $b_2(J)=10$, thus $b_2(S)\geq 11$. From the classification of surfaces we infer
that $S$ is a K3 surface.
Now suppose that $X'$ contains a non-rational singularity $x'_1\in X'$.
Then $K_{S/X'} = -D'$ for some negative-definite curve $D'\subset S$ contracted
by $r:S\ra X'$. Consequently the plurigenera $P_m(S)=h^0(\omega_S^{\otimes n})$ vanish for all $n\geq 1$.
In particular, we have $h^2(\O_S)=h^0(\omega_S)=0$, so the Picard scheme $\Pic_{S/k}$ is smooth.
Now consider the Albanese map $S\ra \Alb_{S/k}$.
This map factors over $X'$, because the exceptional curve for the resolution of singularities $r:S\ra X'$
are rational curves by Proposition \ref{components rational}.
Since $X'\ra J$ is a universal homeomorphism, and $J$ is covered by rational
curves, the Albanese map must be zero, and we infer $h^1(\O_S)=0$.
The Castelnuovo Criterion now ensures that the surface $S$ is rational.
\qed
\medskip
Thus we are precisely in the situation studied in Section \ref{Trivial tangent sheaf},
and determine which of the seven condition \eref{CI}--\eref{Cuspidal} from Section \ref{Trivial tangent sheaf} do hold:
\begin{proposition}
\mylabel{conditions fpb}
The elliptic fibration $h':S\ra\PP^1$ and $r':S\ra X'$ satisfies Conditions
\eref{CI}--\eref{Cuspidal}, except Condition \eref{Elliptic},
and the tangent sheaf is given by $\Theta_{X'}=\O_{X'}(2)\oplus \O_{X'}(-2)$.
\end{proposition}
\proof
According to Proposition \ref{fibrations from rulings}, each ruling $r:J\ra\PP^1$ induce
a quasielliptic fibration $X'\ra\PP^1$, hence \eref{Cuspidal} holds.
The singularities on the Frobenius pullback $X'$ are Zariski singularities by Lemma \ref{frobenius pullback zariski},
so Condition \eref{Zariski} and in particular \eref{CI} hold.
By assumption, $J\ra\PP^1$ admits a section. Thus the same holds for $h':S\ra X'$,
hence the latter admits no multiple fiber.
If $S$ is rational, then
by Theorem \ref{nonrational singularity} the surface contains a single singularity, which is
an elliptic singularity, so $R^1r'_*(\O_S)$ has length one.
In other words, Condition \eref{Length} holds.
Furthermore, we have $\omega_S=\O_S(-C')$ for some negative-definite curve
supported on the exceptional divisor for $r':S\ra X'$, so Condition \eref{Nonfixed} holds.
It remains to check Condition \eref{Tjurina} on the total Tjurina number.
Let $v_b,m_b,\delta_b$ be the numerical invariants for
$J\ra\PP^1$, and $v_a,m_a,\delta_a$ be the ensuing invariants
for the minimal model $J'\ra\PP^1$ of the Frobenius pullback $X'$. Recall that $\delta_a=\delta_b$.
Since $J$ is a rational surface, we have $\sum v_b=12$, according to
\cite{Lang 2000}, Lemma 0.1.
Suppose first that $X'$ has only rational singularities.
Then $v_a=2v_b$, and the total number of exceptional divisors for the
resolution of singularities $J'\ra X'$ is $\sum (m_a-m_b)$.
In light of Proposition \ref{tjurina rdp}, we have to verify
$\sum(m_a-m_b)=12$.
If the fiber $J_b$ is semistable, the same holds for $J_a$, and Ogg's Formula gives
$m_a-m_b = v_a - v_b =v_b$.
If the fiber $J_b$ is unstable, also $J_a$ is unstable, and likewise get
$$
m_a-m_b = (v_a -1 - \delta_a) - (v_b -1 - \delta_b) = =2v_b-v_b= v_b.
$$
In any case, we obtain $\sum(m_a-m_b)=\sum v_b=12$.
Now suppose that $X'$ contains an elliptic singularity.
Then $J\ra\PP^1$ is given by the Weierstra\ss{} equation \eqref{9C},
and its Frobenius pullback is described by
$$
y^2+t^6\gamma_0^2 y = x^3+t^2\gamma_1^2x^2+ t^2\gamma_3^2x + t^2\gamma_5^2,
$$
where $t\nmid\gamma_0$.
The Tjurina ideal is generated by the given Weiererstra\ss\ polynomial $y^2 +t^6\gamma_0^2 y-\ldots$, together with
the partial derivatives $t^6\gamma_0^2$ and $x^2+t^2\gamma_3^2$.
Using that $\gamma_0$ is a unit in $k[[t]]$, and regarding the generators as integral equations in
$t$, $x$ and $y$ of degree $d=6,2,2$, we infer that
$$
k[[t,x,y]]/( t^6, x^2+t^2\gamma_3^2,y^2 +t^6\gamma_0^2 y-\ldots)
$$
has length $\tau =6\cdot 2\cdot 2=24$.
Finally, we compute the tangent sheaf. Choose a ruling $J\ra \PP^1$, and let
$X'\ra\PP^1$ be the resulting quasielliptic fibration.
For each smooth fiber $D\subset J$, the preimage $C\subset X'$ has degree two
with respect to the elliptic fibration $X'\ra\PP^1$, according to Proposition \ref{ruling degree two}.
Now Theorem \ref{conditions for theta} gives $\Theta_{X'/k}=\O_{X'}(-2)\oplus\O_{X'}(2)$.
\qed
\section{K3-like coverings with elliptic double point}
\mylabel{K3-like with edp}
In this section, we shall construct simply-connected Enriques surfaces $Y$
whose K3-like covering $X$ is normal and contains an elliptic double point.
The idea is simple: We start with a rational elliptic surface $J\ra\PP^1$
containing a singular fiber of Lang type 9C, chosen
so that the Frobenius pullback $X'=J^\fpb$ stays rational.
It comes with an induced elliptic fibration $X'\ra\PP^1$.
Its relatively minimal smooth model $J'$ can be seen as a flip $X'\leftarrow S\rightarrow J'$,
and the desired K3-like covering arises as a flop $X'\leftarrow S\rightarrow X$.
The crucial step is the construction of the flop $X$, starting from the flip $J'$,
so that the flop removes all quasielliptic fibrations, introduce another elliptic fibration,
and yet does not change the nature of the Zariski singularities.
For this, we need as key ingredients Shioda's theory of Mordell--Weil lattices
\cite{Shioda 1990} and their classification for rational elliptic surface, which is due
to Oguiso and Shioda \cite{Oguiso; Shioda 1991}.
We start by recalling this. Let
$J\ra B$ be an elliptic surface over a smooth proper curve $B$, with smooth total space
and endowed with zero section $0\subset J$.
The N\'eron--Severi group $\NS(J)=\Pic(J)/\Pic^0(J)$ is endowed with
the intersection pairing $(P\cdot Q)$.
The group of sections for $J\ra B$ is called the \emph{Mordell--Weil group} $\MW(J/B)=\Pic^0(J_\eta)$.
The \emph{trivial sublattice}
$T=T(J/B)$ inside the N\'eron--Severi group $\NS(J)$
is the subgroup generated by the zero-section $O\subset J$, together
with all irreducible components $\Theta\subset J_b$, $b\in B$ that are disjoint from the zero-section.
Thus we may identify $\MW(J/B)$, up to its torsion subgroup, as a subgroup of
the orthogonal complement $T^\perp\subset\NS(J)\otimes\QQ$.
Consequently, it acquires a non-degenerate bilinear $\QQ$-valued form
$$
\langle P,Q\rangle=-(\bar{P}\cdot\bar{Q})
$$
called the \emph{height pairing},
which was extensively studied by Shioda \cite{Shioda 1990}. The sign is a customary convention,
introduced to have a positive-definite height pairing,
and $P\mapsto \bar{P}$ denotes the orthogonal projection onto $T^\perp$.
One also calls $\MW(J/B)$ the \emph{Mordell--Weil lattice}.
The subgroup comprising all sections $P\subset J$ that pass through the
same fiber components as the zero section $O\subset J$ is called the \emph{narrow Mordell--Weil lattice}.
The explicit formula for the height pairing is
$$
\langle P,Q\rangle = \chi(\O_J) + (P\cdot O) + (Q\cdot O) - (P\cdot Q) - \sum\contr_b(P,Q),
$$
according to \cite{Shioda 1990}, Theorem 8.6.
The summands in the middle are the usual intersection numbers,
and the \emph{local contributions} $\contr_b(P,Q)\in\QQ_{\geq 0}$
are certain rational numbers tabulated in \cite{Shioda 1990}, page 229. They reflect how the sections $P,Q$ pass through
the irreducible components of the fiber $J_b$, in relation to the zero-section $O\subset J$.
The corresponding quadratic form can be written as
$$
\langle P,P\rangle = 2\chi(\O_J) + 2(P\cdot O) - \sum\contr_b(P),
$$
where $\contr_b(P)=\contr_b(P,P)$.
For each point $b\in B$, we denote by $\Phi_b$ the group of components in the N\'eron model
of $J\otimes\O_{B,b}$. Its order is the number of irreducible components in $S_b$ with multiplicity
$m=1$ in the fiber. The \emph{narrow Mordell--Weil lattice} is the subgroup of $\MW(J/B)$
of elements $P$ with $P\cong O$ in $\Phi_b$ for all $b\in B$.
Note that if $\Phi\cong O$ in $\Phi_b$, then the local contributions $\contr_b(P,Q)$
in the explicit formula for the height pairing vanish.
Suppose now that $J\ra\PP^1$ be a jacobian rational elliptic surface.
Then we have $\chi(\O_J)=1$.
Oguiso and Shioda \cite{Oguiso; Shioda 1991} tabulated the possible isomorphism classes of Mordell--Weil
lattices $\MW(J/\PP^1)$ in terms of the trivial sublattices $T=T(J/\PP^1)$, a list containing 74 cases.
The arguments are purely lattice-theoretical, and apply to arbitrary characteristics.
Following their convention, we regard $T$ as a positive-definite lattice, by introducing a sign.
Let us give an example: If the configuration of Kodaira types is $\III+ 32^21$, the trivial lattice becomes
$A_1^{\oplus 3}\oplus A_2$, which appears as No.\ 23 in the Oguiso--Shioda list; we then
read off that the Mordell--Weil lattice is given by
$$
\MW(J/\PP^1) = A_1^\vee \oplus \frac{1}{6}\begin{pmatrix}2 & 1\\1&2\end{pmatrix}.
$$
Here the second summand is a Gram matrix, and the first summand is the dual $A_1^\vee=(1/2)$ for
the ADE-lattice $A_1=(2)$.
The following observation will be useful:
\begin{proposition}
\mylabel{disjoint section}
Suppose that $J\ra\PP^1$ has only reduced fibers. Then
there is a section $P\subset J$ that is disjoint from the zero-section $O\subset J$
and has $P\not\cong O$ in the group of components $\Phi_b$ for some point $b\in \PP^1$.
Furthermore, every torsion section $P\neq O$ is disjoint from the zero-section.
\end{proposition}
\proof
According to \cite{Oguiso; Shioda 1991}, Theorem 2.5, the Mordell--Weil group is generated by sections
that are disjoint from the zero-section.
So we first check that the Mordell--Weil group is non-zero. According to the Oguiso--Shioda list, only No.\ 62
has $\MW(J/\PP^1)=0$. In this case, the vertical lattice is $T=E_8$.
It follows that $J\ra\PP^1$ has a fiber of Kodaira type $\II^*$, which is nonreduced, contradiction.
Going through the Oguiso--Shioda list, one observes that in all other cases the
narrow Mordell--Weil lattice is a proper sublattice, so the first assertion follows.
The last statement is \cite{Oguiso; Shioda 1991}, Proposition 5.4.
\qed
\medskip
In what follows, $k$ denotes an algebraically closed ground field of characteristic $p=2$.
Choose a Weierstra\ss{} equation of Lang type 9C, as given in \eqref{9C} with $t\mid\gamma_3$.
The resulting rational elliptic surface $J\ra\PP^1$ then contains only one
singular fiber, which is located over the origin $b\in\PP^1$ and of Lang type 9C, and the
Frobenius pullback $X'=J^\fpb$ is a normal rational surface containing a unique
singularity, which is an elliptic double point obtained by contracting
a cuspidal $(-1)$-curve, and is located over the point $a\in \PP^1$
corresponding to $b$.
Note that the exceptional divisor on the minimal resolution of singularities
is a rational cuspidal curve with self-intersection $-1$.
Let $r':S\ra X'$ be the non-minimal resolution obtained by blowing-up further a closed regular point
on the rational cuspidal curve. Write $h:S\ra\PP^1$ for the induced elliptic fibration.
The fiber fiber $S_a=h^{-1}(a)$ has three irreducible components $D_1,D_2,D_3$,
where $D_2$ is the rational cuspidal curve and $D_1,D_3$ are smooth rational curves.
The dual graph, with the selfintersection numbers, is
\begin{center}
\begin{tikzpicture}
\tikzstyle{vertex}=[circle, draw, inner sep=0pt, minimum size=2ex]
\node[vertex] (a) at (2,1) [label=below:{$-1$}, label=above:{$D_1$}] {};
\node[vertex] (b) at (4,1) [label=below:{$-2$}, label=above:{$D_2$}] {};
\node[vertex] (c) at (6,1) [label=below:{$-1$}, label=above:{$D_3$}] {};
\node (d) at (10,1) {with $D_2$ rational cuspidal.};
\draw (a) -- (b) -- (c);
\end{tikzpicture}
\end{center}
\begin{proposition}
\mylabel{picard number non-minimal}
The smooth rational surface $S$ has Picard number $\rho(S)=12$.
\end{proposition}
\proof
The elliptic fibration $h:S\ra\PP^1$ is relatively minimal, except at the fiber $S_a=D_1+D_2+D_3$.
The curves $D_1,D_3$ are disjoint $(-1)$-curves.
Contracting them yields the relatively minimal model.
The latter has Picard number ten, by \cite{Cossec; Dolgachev 1989}, Proposition 5.6.1,
so our surface $S$ has Picard number twelve.
\qed
\medskip
We choose the indices for the irreducible components $S_a=D_1+D_2+D_3$ so that $r':S\ra X'$ contracts $E'=D_2+D_3$,
and that $S\ra J'$ contracts the $(-1)$-curves $D_1+D_3$. Since $J'$ is a rational elliptic surface,
we have $K_{J'}=-J'_a$ whence $K_S=-D_2$.
Now let $r:S\ra X$ be the contraction of $E=D_1+D_2$.
This yields another normal surface $X$ with $\omega_X=\O_X$, containing
a unique elliptic singularity obtained by contracting a cuspidal $(-1)$-curve.
\begin{theorem}
\mylabel{existence K3-like II}
The normal surface $X$, with its elliptic singularity obtained by contracting a cuspidal $(-1)$-curve,
is a K3-like covering.
\end{theorem}
\proof
In light of Theorem \ref{conditions sufficient}, we have to verify that
the elliptic fibration $h:S\ra\PP^1$ and the contraction $r:S\ra X$ of the exceptional divisor $E=D_1+D_2$ satisfies
Conditions \eref{CI}--\eref{Zariski} of Section \ref{Trivial tangent sheaf}.
Recall that $x_1\in X$ and $x'_1\in X'$ are the unique singular points.
It follows from Lemma \ref{isomorphic singularities} below
that the complete local rings $\O_{X,x_1}^\wedge$ and $\O^\wedge_{X',x'_1}$ are isomorphic.
Thus all conditions but \eref{Elliptic} follow from
Proposition \ref{conditions fpb}.
It remains to show that condition \eref{Elliptic} holds, that is, we must construct on $X$
another elliptic fibration.
For this we first consider the flip $J'$, which is a smooth rational elliptic surface,
whose sole singular fiber is of type $\II$. Choose a zero-section $O\subset J'$,
and consider the resulting Mordell--Weil lattice $\MW(J'/\PP^1)$.
Since the trivial lattice is $T=0$, the Mordell--Weil lattice
is $E_8$, according to \cite{Oguiso; Shioda 1991}. For each section $P\subset J'$ corresponding to a root $\alpha\in E_8$,
the height pairing gives
$2=\langle P,P\rangle = 2 +2 (P\cdot 0)$, whence $P$ is disjoint from the zero-section.
Now let $\alpha,\beta\in E_8$ be two adjacent simple roots, and let
$P,P'\subset J'$ be the sections corresponding to the roots $\alpha,\alpha+\beta\in E_8$,
such that $\alpha\cdot(\alpha+\beta)=2-1=1$.
In terms of the height pairing, this means
$$
1=\langle P,P'\rangle = 1 +(P\cdot O) + (P'\cdot O) - (P\cdot P') = 1-(P\cdot P'),
$$
thus $O,P,P'\subset J'$ are pairwise disjoint. From this we infer that
at least one of the resulting three strict transforms on $S$
is a section for $h:S\ra\PP^1$ passing through the irreducible component $D_2\subset S_a=h^{-1}(a)$.
After changing the zero-section $O\subset J$, we may assume that the strict transform $Q\subset S$
of the zero-section satisfies $Q\cdot D_2=1$.
The dual graph takes the form
\begin{center}
\begin{tikzpicture}
[node distance=2cm]
\tikzstyle{vertex}=[circle, draw, inner sep=0mm, minimum size=2ex]
\node[vertex] (b) at (2,3) [label=left:{$-2$}, label=45:{$D_2$}] {};
\node[vertex] (a) [below of=b] [above=.5cm, label=left:{$-1$}, label=right:{$D_1$}] {};
\node[vertex] (c) [above of=b] [below=.5cm, label=left:{$-1$}, label=right:{$D_3$}] {};
\node[vertex] (d) [right of=b] [right=.5cm, label=below:{$-1$}, label=above:{$Q$}] {};
\draw (a) -- (b) -- (c);
\draw (b) -- (d);
\node (d) [right of=d] [right=.5cm] {with $D_2$ rational cuspidal.};
\end{tikzpicture}
\end{center}
Consider the negative-semidefinite curve $C=D_1+D_2+ Q\subset S$ of arithmetic genus $h^1(\O_C)=1$.
Then some $\O_S(nC)$, with $n\geq 1$
induces another genus-one fibration $g:S\ra\PP^1$.
We claim that the curve $C$ itself is not movable. Seeking a contradiction,
we assume it were. Then the curve $C\subset S$ is \emph{d-semistable},
which means that the normal sheaf is $\O_C(-C)=\O_C$, such that
$\O_{D_2}(-D_2-D_1-Q)=\O_{D_2}$. In light of the fibration
$h:S\ra\PP^1$ defined by the movable curve $D_1+D_2+D_3$, we have $\O_{D_2}(-D_2)=\O_{D_2}(D_1+D_3)$.
It follows that $\O_{D_2}(D_3-Q)=\O_{D_2}$. By Riemann--Roch, this does not
hold for the two points $D_3\cap D_2\neq Q\cap D_2$ on the rational cuspidal curve $D_2$.
We conclude that $C$ is not movable.
However, the curve $2C $ is movable. To see this, set $\shL=\O_S(C)$. Since $\Pic^0(C)=k$ is 2-torsion and
the sheaf $\shL_C$ is numerically trivial, we must have $\shL^{\otimes 2}_C=\O_C$. Consider the short exact sequence
$0\ra\shL\ra\shL^{\otimes 2}\ra \O_C\ra 0$
and the ensuing long exact sequence
$$
H^0(S,\shL^{\otimes 2})\lra H^0(C,\O_C)\lra H^1(S,\shL).
$$
It thus suffices to check that the term on the right vanishes.
By Serre Duality, $h^2(\shL)=h^0(\omega_S\otimes\shL^\vee)=0$,
because $K_S=-D_2$. Riemann--Roch gives
$$
1-h^1(\shL)=\chi(\shL)= C\cdot (C + D_2)/2 + \chi(\O_S) = 1,
$$
thus $h^1(\shL)=0$.
Summing up, $2C\subset S$ is movable, and thus defines a genus-one fibration
$g:S\ra\PP^1$. The intersection number
$$
C\cdot (D_1+D_2+D_3) = C\cdot D_3 = D_2\cdot D_3= 1
$$
shows that $2C$ and thus also all other fibers of $g:S\ra\PP^1$ have degree two with
respect to the original elliptic fiber $h:S\ra\PP^1$.
According to Proposition \ref{picard number non-minimal}, the rational surface $S$ has Picard number $\rho(S)=12$.
By construction, the irreducible components $D_1,Q\subset C$ are disjoint $(-1)$-curves on $S$.
Contracting them thus gives the relatively minimal model for the genus-one fibration $g:S\ra\PP^1$.
We conclude that all $g$-fibers but $C\subset S$ are minimal, contain at most two irreducible components,
and are thus of Kodaira type $\II$ or $\III$.
It remains to check that our genus-one fibration is elliptic.
Seeking a contradiction, we assume that $g:S\ra\PP^1$ is quasielliptic.
Now we use the fact that the fibers of $h:S\ra\PP^1$, except the sole singular fiber $h^{-1}(a)=D_1+D_2+D_3$,
contain no $(-2)$-curves.
It follows that every component of every fiber of $g:S\ra\PP^1$ is horizontal with
respect to $h:S\ra\PP^1$, with the exception of $D_1$ and $D_2$. Moreover, every fiber other than $2C$
has Kodaira type $\II$ or $\III$, and its irreducible components necessarily pass through $D_3\subset h^{-1}(a)$,
because $D_1$ and $D_2$ are contained in the fiber $2C$.
Since $S$ is rational, all other fibers beside $2C$ are non-multiple (\cite{Cossec; Dolgachev 1989}, Proposition 5.6.1),
whence there type coincides with the type of the corresponding fiber on the
associated jacobian quasielliptic fibration.
According to Ito's analysis of jacobian quasielliptic fibrations in characteristic two
(\cite{Ito 1994}, Proposition 5.1), there is at least one reducible fiber, which in our situation
must be of type $\III$. In fact, the Mordell--Weil lattice for a quasi-elliptic fibration is 2-torsion,
whence the trivial lattice has rank ten, so there are eight fibers of Kodaira type $\III$.
Each irreducible component $R,R'\subset S$ of such a fiber is a section for $h:S\ra\PP^1$,
both pass through $D_3\subset S$, and they meet tangentially somewhere.
Consequently, their images define two sections $P,P'\subset J'$ with
intersection matrix $(\begin{smallmatrix}-1&3\\3&-1\end{smallmatrix})$.
Obviously, $P,P'$ are disjoint from the zero-section $O\subset J'$, which is the image of $Q\subset S$.
To derive a contradiction, we compute the Gram matrix for the height paring on $J'$.
As noted above, the sections $P,P'\neq O$ are disjoint from the zero-section.
Thus $\langle P,P\rangle = 2 + 2(P\cdot O) = 2$, likewise $\langle P',P'\rangle =2$, and finally
$$
\langle P,P'\rangle = 1 + (P\cdot O)+(P'\cdot O) - (P\cdot P')= -2.
$$
So the Gram matrix with respect to the height pairing is $(\begin{smallmatrix}2&-2\\-2&2\end{smallmatrix})$,
which has zero determinant. On the other hand, the Mordell--Weil lattice is non-degenerate,
contradiction. Thus $g:S\ra\PP^1$ and the induced fibration on $X$ are elliptic
rather than quasielliptic fibrations.
This means that the condition \eref{Elliptic} indeed holds.
\qed
\medskip
In the preceding proof, we have used the following observation:
\begin{lemma}
\mylabel{isomorphic singularities}
Let $R$ be a discrete valuation ring, and $J\ra\Spec(R)$ a relatively minimal
genus-one fibration, endowed with two disjoint sections $P_1,P_2\subset J$.
Let $S\ra J$ be the blowing-up whose center consists of the two closed points
$z_i\in P_i$, and let $E_i\subset S$ be their preimage.
Write $S\ra X_i$ for the blowing-down so that $E_i$ surjects onto the closed fiber of $X_i$,
and let $x_i\in X_i$ be the resulting singular point. Then the two local rings
$\O_{X_1,x_1}$ and $\O_{X_2,x_2}$ are isomorphic.
\end{lemma}
\proof
Let $f:J_\eta\ra J_\eta$ be the isomorphism that sends the the generic fiber of $P_1$
to the generic fiber of $P_2$. By relative minimality, it extends to
an isomorphism $f:J\ra J$. It has $f(z_1)=z_2$, and by the universal property
of blowing-ups induced an isomorphism $f:S\ra S$ with $f(E_1)=E_2$.
In turn, it maps the exceptional locus for $S\ra X_1$ to the exceptional locus for $S\ra X_2$.
By the universal property of contractions \cite{EGA II}, Lemma 8.11.1,
it yields the isomorphism between the two local rings in question.
\qed
\section{Uniqueness of the elliptic double point}
\mylabel{Uniqueness edp}
Let $Y$ be a simply-connected Enriques surface over an algebraically closed
ground field $k$ of characteristic $p=2$, with normal
K3-like covering. The goal of this section is to establish the
following uniqueness result:
\begin{theorem}
\mylabel{unique elliptic singularity}
If there an an elliptic singularity $x_1\in X$, then
there are no other singularities on $X$.
\end{theorem}
\proof
Seeking a contradiction, we assume that there is another singularity $x_2\in X$.
Let $y_1,y_2\in Y$ be the images of the two singularities $x_1,x_2\in X$.
First, we show that there are no elliptic fibrations $\varphi:Y\ra\PP^1$ admitting
a radical two-section $A\subset Y$. Seeking a contradiction, we assume that such a fibration exists.
Let $J\ra\PP^1$ be its jacobian fibration, which is a rational elliptic surface,
and write $X'=J^\fpb$ for the Frobenius pullback.
By Proposition \ref{two-section = section},
the normal surfaces $X'$ and $X$ are birational. It follows that the surface $X'$ is rational,
whence contains an elliptic singularity. According to Theorem \ref{nonrational singularity},
$J\ra\PP^1$ must contain a fiber of Lang type 9C, and this is the only singular fiber $J_b$.
In turn, the Frobenius base-change $X'=J^\fpb$ has but one singularity, lying on the fiber $X'_a$,
where $a\in\PP^1$ corresponds to $b\in\PP^1$ under the relative Frobenius morphism.
Now let $2F\subset Y$ be a multiple fiber. If $F$ is smooth, then the K3-like covering $X$
is smooth along $\epsilon^{-1}(F)$, according to Proposition \ref{preimage elliptic}.
If $Y_b$ is not a multiple fiber, it follows that all half-fibers are smooth,
and there is but one singularity on $X$, namely the point
corresponds to the singular point in $Y_b$,
contradiction. Thus $Y_b$ is a multiple fiber.
Write $ Y_b=2F$ for the corresponding half-fiber $F$. By the above discussion, the two singularities
$x_1,x_2\in X$ map to $F$.
This analysis applies not only to $\varphi:Y\ra\PP^1$, but also to all other elliptic fibrations.
According to Theorem \ref{properties genus-one fibrations}, there is another elliptic fibration $\varphi':Y\ra\PP^1$
with $F\cdot F'=1$. But this intersection number contradicts $x_1,x_2\in F\cap F'$.
Consequently, there is no elliptic fibration on $Y$ admitting a radical two-section.
According to Proposition \ref{radical two-section}, there are no $(-2)$-curves on $Y$.
Now choose again an elliptic fibration $\varphi:Y\ra\PP^1$.
Then all fibers are irreducible, according to Proposition \ref{radical two-section}.
By the result of Liu, Lorenzini and Raynaud \cite{Liu; Lorenzini; Raynaud 2005}, Theorem 6.6
the same holds for the associated jacobian fibration $J\ra\PP^1$.
Suppose all elliptic fibrations on $Y$ would have only one singular fiber $Y_b$.
Arguing as above, this fiber $Y_b=2F$ is multiple, and we find two orthogonal elliptic fibrations with
$F\cdot F'=1$ and thus $y_1,y_2\in F\cap F'$, contradiction.
Consequently, we may choose an elliptic fibration $\varphi:Y\ra\PP^1$ so
that its jacobian $J\ra\PP^1$ contains at least two singular fibers.
Since all fibers are irreducible, the trivial lattice $T\subset \NS(J)$ is zero, and
the Mordell--Weil group $\MW(J/\PP^1)$ is isomorphic to $E_8$.
If follows that the Picard group of the generic fiber $Y_\eta$ is a free abelian group of rank nine.
Since $\epsilon:X\ra Y$ is a universal homeomorphism, the same holds for
the generic fiber $X_\eta$. Now let $S\ra X$ be the resolution of singularities,
and $S\ra S'$ be the relative minimal model over $\PP^1$.
Then $S'\ra\PP^1$ is a rational elliptic surface, which is not necessarily jacobian.
In turn, $\Pic(S')$ has rank ten.
However, we just saw $\Pic(S'_\eta/\eta)$ has rank nine, whence
all fibers of $S'\ra\PP^1$ are irreducible.
Again by the result of Liu, Lorenzini and Raynaud \cite{Liu; Lorenzini; Raynaud 2005}, Theorem 6.6
the same holds for the smooth model $J'\ra\PP^1$ of the Frobenius pullback $X'=J^\fpb$.
If a singular fiber $J_a'$ is semistable, then also the corresponding
fiber $J_b$ is semistable, and $m_a=2m_b\geq 2$, contradiction. Thus all singular fibers $J'_a$
are of Kodaira type $\II$, with numerical invariant $m_a=1$.
According to Theorem \ref{nonrational singularity}, the presence of two
singular fibers on $J$ ensures that all singularities on the Frobenius pullback $X'=J^\fpb$
are rational. By Lemma \ref{unstable fiber}, each singular fiber $J'_a$ has numerical invariant
$m_a=2m_b+\delta +1 \geq 3$, contradiction.
\qed
\section{K3-like coverings with rational double points}
\mylabel{K3-like with rdp}
Let $k$ be an algebraically closed ground field of characteristic $p=2$.
In this section, we construct K3-like coverings $X$ that are normal with
only rational singularities.
The main result is that for each but one of the 110 configurations
of Kodaira types from Lang's classification \cite{Lang 2000} of rational elliptic surfaces
with reduced fibers
actually appears as a configuration on some K3-like covering $X$ and
the ensuing Enriques surfaces $Y=X/G$.
The case we do not understand yet is the rational elliptic surface having
a unique degenerate fiber of Lang type 10C, which already appeared in the proof
of Theorem \ref{existence K3-like II}.
Indeed, we start with a rational elliptic surface $J\ra\PP^1$ that is jacobian with
only reduced fibers and assume that the Frobenius pullback $X'=J^\fpb$
has only rational singularities. In other words, the minimal resolution
of singularities $r':S\ra X'$ is a K3 surface.
The case leading to an elliptic singularity was already
treated in Section \ref{K3-like with edp}.
Let $E'\subset S$ be the exceptional divisor for the minimal resolution of singularities $r':S\ra X'$.
The task now is to find a suitable flop $X'\leftarrow S\rightarrow X$
by contracting another vertical negative-definite curve $E\neq E'$ on $S$.
Roughly speaking, the new contraction must
introduce another elliptic fibration, destroy all quasielliptic fibrations,
yet keep the Kodaira type of the singular fibers.
To achieve this task, we introduce for each closed point $a\in \PP^1$ the following
conditions relating the vertical negative-definite curves
$E_a$ and $ E'_a$ on the K3-surface $S$:
\newcounter{MNr}
\setcounter{MNr}{0}
\newcommand{\Mu}[1]{\refstepcounter{MNr}\textbf{(M\arabic{MNr})} \label{#1} }
\newcommand{\mref}[1] {\text{\rm (M\ref{#1})}}
\medskip
\begin{list}{-mm}{\leftmargin2em\itemsep1em}
\item[\Mu{Semistable}]
\emph{If the fibers $X'_a$ and $S_a$ are semistable, then $E_a$ either coincides with $E'_a$, or it is
the strict transform of $X'_a$.}
\item[\Mu{II}]
\emph{If the fiber $X'_a$ has type $\II$ and $S_a$ has type $\I_n^*$ for some $n\geq 0$, then either $E_a$ coincides
with $E_a'$, or it is the union
of all irreducible components $\Theta\subset S_a$ except the terminal component
nearest to the strict transform of $X'_a$.}
\item[\Mu{III}]
\emph{If the fiber $X'_a$ has type $\III$ and $S_a$ has type $\I_n^*$ for some $n\geq 0$, then either $E_a$ coincides
with $E_a'$, or it is the union
of all irreducible components $\Theta\subset S_a$ except the two terminal components not contained in $E_a'$. }
\item[\Mu{IV}]
\emph{If the fiber $X'_a$ has type $\IV$ and $S_a$ has type $\IV^*$, then either $E_a$ coincides
with $E_a'$, or it is the strict transform of $X'_a$ together with the central component $\Theta\subset S_a$.}
\item[\Mu{Other}]
\emph{In all other cases, $E_a$ coincides with $E'_a$.}
\end{list}
\medskip
By abuse of notation, we here say that the fiber $X_a$ on the normal surface $X$
has a certain Kodaira type if this is the Kodaira type of $J_a$ on the smooth surface $J$.
Moreover, a \emph{terminal irreducible component} on $S_a$ means a component corresponding to
a vertex in the dual graph that is adjacent to only one other vertex.
Note that in Condition \mref{II}, in the boundary case $n=0$ we may take
for $\Theta\subset S_a$ any of the three terminal components different from the strict transform of $X'_a$.
\begin{definition}
\mylabel{mutations}
In the preceding situation, we say that $X'\leftarrow S\rightarrow X$ is a \emph{mutation}
if the conditions \mref{Semistable}--\mref{Other} on the vertical negative-definite curves
$E,E'\subset S$ holds for all closed points $a\in\PP^1$.
A mutation is a \emph{good mutation} if in addition
there is a horizontal Cartier divisor $F\subset X$ that is an elliptic curve.
\end{definition}
Good mutations lead to the desired flops:
\begin{proposition}
\mylabel{mutations K3-like coverings}
Suppose $X'\leftarrow S\ra X$ is a good mutation of the Frobenius pullback $X'=J^\fpb$.
Then the normal surface $X$ is a K3-like covering with only rational singularities.
The induced elliptic fibration $\varphi:Y\ra\PP^1$ on the simply-connected
Enriques surface $Y=X/G$ has the same configuration of Kodaira types
as the original rational elliptic surface $J\ra\PP^1$.
\end{proposition}
\proof
In our situation, $S$ is a K3-surface.
In order to apply Theorem \ref{conditions sufficient}, it suffices
to check that Conditions \eref{Tjurina}, \eref{Elliptic}
and \eref{Zariski}
form Section \ref{Trivial tangent sheaf} hold.
Condition \eref{Elliptic} holds by definition on good mutations.
According to Proposition \ref{conditions fpb},
the Frobenius pullback $X'=J^\fpb$ satisfies
Conditions \eref{Tjurina} and \eref{Zariski}.
It remains to check that a mutation $X'\leftarrow S\rightarrow X$
produces on $X$ only Zariski singularities, and that the total Tjurina does not change.
This is obvious if we are in the case \mref{Semistable}, when $S_a$ is semistable.
In case $\mref{IV}$, the fiber $X'_a$ has type $\IV$ and $S_a$ has type $\IV^*$, and $X$ acquires four
rational double points of type $A_1$ lying over $a$, whereas $X'$ contains
a rational double point of type $D_4$. So in this situation, the singularities are different Zariski singularities,
but the total Tjurina number is unchanged.
It remains to verify the cases \mref{II} and \mref{III}.
Then $S_a$ is a fiber of Kodaira type $I^*_n$ for some $n\geq 0$.
The following argument shows that the ensuing singularities on $X'$ and $X$ are formally isomorphic.
Set $R=\O_{\PP^1,a}$ and $F=\Frac(R)$. For each terminal component $\Theta,\Theta'\subset S_a$,
we may choose formal sections $A,A'\subset S\otimes \hat{R}$ passing through $\Theta$ and $\Theta'$,
respectively.
Since the generic fiber $U=S\otimes\hat{F}$ is an elliptic curve, there is a translation automorphism
$\tau:U\ra U'$ sending the $\hat{F}$-rational point $A\cap U$ to $A'\cap U$. By the functoriality of the relative minimal
model, this extends to an automorphism of $S\otimes\hat{R}$ sending $\Theta$ to $\Theta'$.
In turn, the singularity arising from contractions of all components but $\Theta$ is
formally isomorphic to the singularity given by the contraction of all components but $\Theta'$.
This settles the case \mref{II}, and the case \mref{III} follows analogously.
\qed
\medskip
It is very easy to specify mutations. The challenge, however, is to introduce another elliptic
fibration as well. Rather than giving an elliptic curve, we shall construct its degenerate fibers.
This boils down to finding sections for $J'\ra\PP^1$ with suitable intersection behavior.
Using the following, we shall infer that the resulting fibration is elliptic rather then quasi-elliptic:
\begin{proposition}
\mylabel{good mutations}
Let $X'\leftarrow S\rightarrow X$ be a mutation with respect to the vertical negative-definite $E\subset S$.
Suppose that there is curve of canonical type $F\subset S$ with the following properties:
\begin{enumerate}
\item The type of the curve $F$ is either $\I_n$ or $\I_{2n+1}^*$ or $\IV$ or $\IV^*$.
\item It has degree $>0$ with respect to the elliptic fibration $S\ra\PP^1$.
\item Each irreducible component of $E$ is either contained in or disjoint from $F$.
\end{enumerate}
Then $X'\leftarrow S\rightarrow X$ is a good mutation.
\end{proposition}
\proof
The curve of canonical type $F\subset S$ induces a genus-one fibration $g:S\ra\PP^1$ on the K3 surface $S$.
Condition (i) ensures that the new fibration is elliptic, in light of \cite{Rudakov; Safarevic 1978},
the Proposition on page 150.
Condition (ii) implies that this fibration is different from our original elliptic fibration $S\ra\PP^1$.
By the last condition, the new elliptic fibration on $S$ induces a new elliptic fibration on $X$.
One of its closed fibers is an elliptic curve $E\subset X$ that is horizontal with respect
to the old elliptic fibration $X\ra\PP^1$. Thus the mutation $X'\leftarrow S\rightarrow X$ is good.
\qed
\medskip
We recall from Lang's Classification that if $J\ra\PP^1$ contains a fiber of Lang type $10C$,
this fiber has Kodaira type $\III$, and there are no other singular fibers.
Now we can state the main result of this section:
\begin{theorem}
\mylabel{good mutations exist}
Suppose that the rational elliptic surface $J\ra\PP^1$ contains no fiber of Lang type 10C.
Then there exists a good mutation $X'\leftarrow S\rightarrow X$, and the resulting $X$ is a K3-like
covering.
\end{theorem}
\proof
The second statement follows from Theorem \ref{mutations K3-like coverings}.
The strategy for the first statement is to use Lang's classification for $J\ra\PP^1$ together with
the Oguiso--Shioda list for the Mordell--Weil lattice
$\MW(J/\PP^1)$, in order to guess some vertical negative-definite curve $E\subset S$ that
yields a mutation, as well as
to produce suitable sections $P,P',Q,\ldots\subset J$
that allow to construct a curve of canonical type $F\subset S$
for which Proposition \ref{good mutations} applies. Throughout, $O\subset J$ denotes
the zero-section. Recall from Proposition \ref{disjoint section} that in any case there is a section
disjoint from the zero-section $O$.
To simplify notation, we now write $a,b\in\PP^1$ for closed points, and
$J_a,X'_a, S_a$ for the fibers for the fibrations $S\ra\PP^1$, $X'\ra\PP^1$ and $J\ra\PP^1$.
We have to examine five cases:
\newcounter{CasesForJ}
\setcounter{CasesForJ}{0}
\medskip
{\bf Case (\stepcounter{CasesForJ}\roman{CasesForJ}):}
\emph{Suppose there are at least two semistable fibers $J_a$ and $J_b$.}
Then define $E_a\subset S_a$ and $E_b\subset S_b$
as the strict transform of the fibers $X'_a$ and $X'_b$, respectively.
Let $E\subset S$ be the curve obtained from $E'$ by replacing $E'_a+E'_b$ by $E_a+ E_b$.
This yields a mutation $X'\leftarrow S\ra X$.
Now choose a section $P\subset Y$ disjoint from the zero-section $O\subset Y$.
Together with suitable chains of irreducible components inside $J_a$ and $J_b$, they form
a cycle of projective lines. Their preimage $F\subset S$ becomes a cycle
of $(-2)$-curves, that is, a curve of canonical type $\I_n$ for some $n\geq 4$.
By construction, each irreducible component of $E$ is either disjoint from or contained in $F$.
Thus Proposition \ref{good mutations} applies.
\medskip
{\bf Case (\stepcounter{CasesForJ}\roman{CasesForJ}):}
\emph{Suppose there is exactly one semistable fiber $J_b$.}
According to \cite{Lang 2000}, Section 1 and 2, there are four possibilities for the configuration of singular fibers:
$$
\II+8,\quad \II+5,\quad \III+ 8,\quad \III+6.
$$
The respective Lang types of the unstable fiber $J_a$ are $1A,1C,2A,2B$.
Write the semistable fiber $J_b=\sum\Theta_i$ with cyclic indices in the natural way, that is,
$\Theta_0\cdot O=1$ and $\Theta_i\cdot\Theta_{i+1}=1$.
First suppose that the configuration is $\II+8$.
The fiber $J_a$ has numerical invariants
$v=4$, $m=1$, $\delta=2$. On $S$ the invariants become
$v'=8$ and $m'=5$, whence $S_a$ has Kodaira type $\I^*_0$, and
the corresponding singularity on $X'$ is of type $D_4$.
Fix some terminal irreducible component in $S_a$ different from the strict transform of $X'_a$,
and define $E_a$ as the union of the remaining components $\Theta\subset S_a$.
Let $E\subset S$ be the curve obtained from $E'$ by replacing $E'_a$ by $E_a$.
This yields a mutation $X'\leftarrow S\rightarrow X$.
The strict transforms of the zero-section $O,\Theta_0$ together with $E_a$
and the its two adjacent components in the
semistable fiber $S_b$ form curve $F\subset S$ of canonical type $I^*_3$,
and each irreducible component of $E$ is either disjoint from or contained in $F$.
The case that the configuration is $\III+8$ is analogous and left to the reader.
Next suppose that the configuration is $\II+5$.
Then the fiber $J_a$ has numerical invariants
$v=7$, $m=1$, $\delta=5$.
The invariants for $S_a$ become $v'=14$ and $m'=8$, hence the Kodaira type is $\III^*$ or $\I^*_3$.
The corresponding singularity on $X'$ is then $E_7$ or $D_7$. The latter is not Zariski, by Theorem \ref{zariski rdp}.
Therefore the singularity is $E_7$ and the Kodaira type is $\III^*$.
Let $E_b\subset S$ be the strict transform of the semistable fiber $X'_b$.
Define $E\subset S$ as the curve obtained from $E'$ by replacing $E'_b$ by $E_b$.
This yields a mutation $X'\leftarrow S\ra X$.
It remains to specify the curve $F\subset S$.
The trivial lattice for $J\ra\PP^1$ is $T=A_4$, whence we are in No.\ 8 of the Oguiso--Shioda list,
and we read-off that the narrow Mordell--Weil lattice is $A_4$. Hence there are two orthogonal simple roots $\beta,\beta'$,
and $\alpha=\beta+\beta$ has $\alpha^2=4$. In turn, there is
a section $P\subset J$ passing through the same fiber components as the zero-section $O\subset J$,
with
$$
4=\langle P,P\rangle = 2 + 2 (P\cdot O).
$$
Thus $(P\cdot O)=1$. The strict transforms of $O,P,\Theta_0$ form a curve $F\subset S$ of canonical type
$I_3$ or $\IV$, depending whether or not the intersection point $O\cap P$ lies on $\Theta_0$.
By construction, each irreducible component of $E$ is contained in or disjoint from $F$.
Finally suppose we have the configuration $\III+6$.
The invariants for $J_a$ are $v=6$, $m=2$, $\delta=3$,
whence $S_a$ has invariants $v'=12$, $m'=8$. The Kodaira type is either $\III$ or $\I^*_3$,
and the corresponding singularity on $X'$ is $E_6$ or $D_6$.
The former is not a Zariski singularity, by Theorem \ref{zariski rdp}.
Consequently, the singularity is $D_6$ and the Kodaira type is $I^*_3$.
Let $E_b\subset S_b$ be the strict transform of the semistable fiber $X_b$, and
let $E\subset S$ be obtained from $E'$ by replacing $E'_b$ by $E_b$.
This yields a mutation $X'\leftarrow S\rightarrow X$. It remains to define
the curve of canonical type $F\subset S$.
The trivial lattice of $J$ is $T=A_1\oplus A_5$.
This is one of the few cases where there are two possibilities for
the Mordell-Weil group: No.\ 28 and No.\ 29,
with Mordell--Weil lattice
either $(1/2)\oplus(1/6)$ or $A_2^\vee\oplus\ZZ/2\ZZ$.
Both lattices contain an element $\alpha$ with $\alpha^2=2/3$, thus
there is a section $P\subset J$ with
$$
2/3=\langle P,P\rangle = 2+2(P\cdot O) -\contr_a(P)-\contr_b(P).
$$
This gives $(P\cdot O)=0$, with the two possibilities $\contr_a(P)=0,\contr_b(P)=8/6$ or $\contr_a(P)=3/6,\contr_b(P)=5/6$.
Let $P'\subset J$ be the section corresponding to $-P\in\MW(J/\PP^1)$.
Using
$$
-4/6=\langle P,P'\rangle=1-(P\cdot P') -\contr_a(P,P')-\contr_b(P,P'),
$$
and checking the two possibilities, one infers that only $\contr_a(P)=0,\contr_b(P)=8/6$ is possible,
thus $P\cong P'$ in the group of components $\Phi_a$, and
$(P\cdot\Theta_2)=(P'\cdot\Theta_4)=1$, which finally gives $(P\cdot P')=1$.
Thus $P,\Theta_2,\Theta_3,\Theta_4,P'$ form a cycle, and its preimage
$F\subset S$ becomes a curve of canonical type $\I_7$.
Each irreducible component of $E$ is contained in or disjoint from $F$.
\medskip
\emph{From now on, we assume that all singular fibers $J_a$ are unstable.}
\medskip
{\bf Case (\stepcounter{CasesForJ}\roman{CasesForJ}):}
\emph{Suppose there is a fiber $J_a$ of Kodaira type $\II$, whose Lang type is 9A or 9C.}
This means $v=4r$, $m=1$ and $\delta=4r-2$, with respective values $r=1, 3$.
In turn, the invariants on $S$ become $v'=8r$ and $m'=4r+1$.
The only possibility is that the fiber $S_a$ has type $\I^*_{4r-4}$,
and the singularity on $X'=J^{(2/\PP^1)}$ is of type $D_{4r}$.
Fix the terminal component of $S_a$ nearest to the strict
transform of $X'_a$, and let $E_a\subset S_a$ be the union of the remaining irreducible
components.
Let $E\subset S$ be the curve obtained from $E'\subset S$ by replacing
$E'_a$ by $E_a$. This yields a mutation $X'\leftarrow S\rightarrow X$.
Let $P\subset J$ be a section disjoint from the zero-section $O\subset J$.
Together with the $4r$ components of $E_a$, they induce a canonical curve $F\subset S$
of type $\I^*_{4r-3}$. By construction, each irreducible component from $E$ is either disjoint
from or contained in $F$.
\medskip
{\bf Case (\stepcounter{CasesForJ}\roman{CasesForJ}):}
\emph{Suppose there is a fiber $J_a$ of Kodaira type $\III$ with Lang type 10A.}
The fiber $J_a$ then has invariants $v=4$, $m=2$ and $\delta=1$. In turn,
the surface $S$ gets a fiber with $v'=8$ and $m'=6$,
which must be of type $\I_1^*$, and the singularity on $X'$ is of type $D_4$.
Let $\Theta\subset S_a$ be the terminal component that is nearest to the strict transform
of $X'_a$, and let
$E_a$ as the union of the other components.
Define $E\subset S$ as the curve obtained from $E'\subset S$ by replacing
$E'_a$ by $E_a$. Its contraction yields a mutation $X'\leftarrow S\rightarrow X'$.
According to Lang's classification, we have the following seven possibilities for the configuration
of Kodaira types on $J$:
$$
\III+\II+\II,\,
\III+\III+\II,\,
\III+\III+\III,\,
\III+\II+\IV,\,
\III+\IV+\IV,\,
\III+\II,\,
\III+\III.
$$
The corresponding trivial lattices are of the form $A_1^{\oplus n}$ with $n=1,2,3$
and $A_1\oplus A_2^{\oplus m}$ with $m=1,2$.
In the Oguiso--Shioda classification, this are No.\ 2,4,7,6,20.
One sees that in all cases,
the narrow Mordell--Weil lattice contains a vector $\alpha$ with $\alpha^2=2$.
In turn, there is a section $P\subset J$ intersecting the same fiber components as the
zero-section $O\subset J$, with
$$
2=\langle P,P\rangle = 2+ 2(P\cdot O).
$$
It follows that $O,P$ are disjoint. Consequently $O,P$ and the four components of $E_a$ induce a curve
$F\subset S$ of canonical type $I_1^*$.
By construction, each irreducible component of $E$ is either contained in or disjoint from $F$.
\medskip
{\bf Case (\stepcounter{CasesForJ}\roman{CasesForJ}):}
\emph{Suppose there is a fiber $J_a$ of Kodaira type $\IV$.}
This fiber has Lang type 11, with invariants $v=4$, $m=4$, $\delta=0$.
Whence the fibers $S_a,S_b$ have invariants $v'=8$, $m'=7$.
Their possible Kodaira types are $\IV$ or $\I^*_2$.
The respective singularities on $X'$ are $D_4$ or $A_4$.
The latter is not Zariski by Theorem \ref{zariski rdp},
hence the singularity is $D_4$ and the Kodaira type is $\IV$.
We first consider the situation that there is also a fiber of different Kodaira type.
According to Lang's Classification, the possible configuration are
$$
\IV+\II+\II,\,
\IV+\III+\II,\,
\IV+\III+\III,\,
\IV+\II,\,
\IV+\III.
$$
The corresponding trivial lattices are of the form $A_2\oplus A_1^{\oplus n}$ with $n=0,1,2$.
In the Oguiso--Shioda Classification, this are the cases No.\ 3, 6,12.
In each case, one finds two simple orthogonal roots $\beta,\beta'$ in the narrow Mordell--Weil lattice.
In turn, $\alpha=\beta+\beta'$ has $\alpha^2=4$.
Thus there is a section $P\subset J$ passing through the same fiber components as the zero-section, with
$$
4=\langle P,P\rangle = 2 + 2(P\cdot O).
$$
In turn, $P\cdot O=1$.
Therefore, we define $E_a$ to be the strict transform of $X'_a$, together with the central component
$\Theta\subset S_a$.
Let $E\subset S$ be the curve obtained from $E'$ by replacing $E'_a$ by $E_a$.
This yields a mutation $X'\leftarrow S\ra X$.
Write $\Theta\subset J_a$ for the component intersecting $O,P$.
The strict transforms of $P,O,\Theta_0$ on $S$ form a curve of canonical type $I_3$ or $\IV$, depending
whether or not the intersection point $P\cap O$ does not or does lie on $\Theta_0$.
By construction, each irreducible component of $E$ is either contained in or disjoint from $F$.
It remains to understand the case that all singular fibers are of Kodaira type $\IV$.
According to \cite{Lang 2000}, the configuration of singular fibers is $\IV+\IV+\IV$.
Let $J_a,J_b$ be two of the singular fibers, and define $E_a,E_b$ as the
strict transform of $X'_a,X'_b$, together with the central components
of $S_a,S_b$. Let $E\subset S$ be the curve obtained from $E'$ by replacing $E'_a+E'_b$
by $E_a+E_b$. This defines a mutation $X'\leftarrow S\ra$.
Choose a section $P\subset J$ that is disjoint from the zero-section $O\subset J$.
Then $O,P$ together with suitable chains from $X'_a,X'_b$ form a cycle.
Its preimage on $S$ form a curve of canonical type $I_n$ for some $n\geq 0$.
By construction, each component of $E$ is either disjoint from or contained in $E$.
\medskip
This completes the proof, in light of Lang's classification:
Cases (i) covers applies if all fibers are semistable, by \cite{Lang 2000}, Section 1.
It also applies to all but four cases with $j(J/\PP^1)\neq 0$, by \cite{Lang 2000}, Section 2.
These four extra cases are treated in (ii). The remaining cases have global
$j$-invariant $j(J/\PP^1)=0$, which is treated in \cite{Lang 2000}, Section 3. Then all fibers are unstable,
and their is a fiber $J_a$ of Lang type 9A, 9C, 10A or 11.
These are treated in (iii)--(v).
The only exception are the rational elliptic surfaces $J\ra\PP^1$ containing
a fiber of Lang type 10C, which we had to exclude
from consideration.
\qed
\section{Canonical coverings for fibers of type $\II$, $\III$ and $\IV$}
\mylabel{Canonical covering fibers}
Let $k$ be an algebraically closed ground field of characteristic $p>0$. The
rational cuspidal curve $C$ has as Picard scheme the additive group $\Pic^0_{C/k}=\GG_a$,
and its Frobenius kernel is $\alpha_p$. In turn, we have a canonical $\alpha_p$-torsor
$\tilde{C}\ra C$. The goal of this section is to describe this torsor explicitly for $p=2$.
Note that the rational cuspidal curve $C$ occurs as fibers of type II
in genus-one fibrations. We will construct canonical torsor for fibers of type III and IV
as well. This illustrates our general result in Section \ref{Canonical coverings}.
The following terminology of Bayer and Eisenbud \cite{Bayer; Eisenbud 1995} is useful:
A \emph{ribbon} on a scheme $X$ is a closed embedding $X\subset Z$
where the ideal sheaf $\shI\subset\O_Z$ satisfies $\shI^2=0$
and is invertible when viewed as a module over $\O_X=\O_Z/\shI$.
Over a ground field $k$, isomorphism classes of such infinitesimal extensions with
fixed $\shL=\shI$ correspond to extension classes in
$$
\Ext^1(\Omega^1_{X/k},\shL),
$$
as explained in \cite{Bayer; Eisenbud 1995}. The zero element corresponds to the
class of the \emph{split ribbon}, which one may denote by $X[\shL]$.
For $X=\PP^1$ and $\shL=\O_{\PP^1}(-2)$, the Ext-group vanishes,
and we see that every ribbon $\PP^1\subset Z$ with ideal $\O_{\PP^1}(-2)$ is split.
To determine the tangent sheaf of the ribbon, we write $\PP^1\subset Z$ in explicit coordinates as follows:
$$
Z = \Spec(k[s,\epsilon]) \cup \Spec(k[s^{-1},\eta])
$$
with $\epsilon^2=\eta^2=0$ and $\eta=s^{-2}\epsilon$. The sheaf $\Omega^1_{Z/k}$ is if freely generated on the first chart by
$ds,d\epsilon$ and on the second chart by $ds^{-1},d\eta$,
which transition matrix given by
$ds^{-1} = s^{-2}dt$ and $d\eta=s^{-2}d\epsilon$. In turn, the tangent sheaf $\Theta_{Z/k}$ is given
by $D_s,D_\epsilon$ and $D_{s^{-1}}=s^2D_s, D_\eta=s^2D_\epsilon$, whence
the locally free sheaf comes from the cocycle $(\begin{smallmatrix} s^2\\0&s^2\end{smallmatrix})$.
In turn, we have a decomposition
$$
\Theta_{Z/k} = \shL\oplus\shL,
$$
where $\shL$ is invertible with $\shL|\PP^1=\O_{\PP^1}(2)$.
Let us record the following consequence:
\begin{proposition}
Let $G$ be either $\alpha_p$ or $\mu_p$.
Then there is no free $G$-action on $Z$ so that the quotient $Z/\alpha_p$ is regular.
Moreover, there is no free $G$-action on $Z$ that leaves $Z_\red=\PP^1$ invariant.
\end{proposition}
\proof
A free action correspond to a $p$-closed global vector field $D\in H^0(Z,\Theta_{Z/k})$
that does not vanish anywhere. In other words, if we decompose $D=D_1+D_2$
according to the decomposition $\Theta_{Z/k}=\shL\oplus\shL$, then $D_1,D_2$ have no common zero.
The condition that $Z/\alpha_p$ is regular means that the orbits $Gz\subset Z$
have finite projective dimension. Since $\deg(\shL)>0$, each summand must vanish somewhere.
Let $z\in Z$ be a point where $D_2$ vanishes, and write $\O_{Z,z}^\wedge=k[[u,v]]/v^2$.
Then the orbit ideal has the form $\ideal a=(u^p,v)$.
But the residue class ring $\bar{A}=k[[u,v]]/(u^p,v)$ of $A=k[[u,v]]/(v^p)$
has infinite projective dimension, by considering the matrix factorization with
$\varphi=(\begin{smallmatrix}v&0\\0&v\end{smallmatrix})$ and
$\psi=(\begin{smallmatrix}v^{p-1}&0\\0&v^{p-1}\end{smallmatrix})$,
the ensuing free resolution
$$
\ldots\stackrel{\varphi}{\lra} A^2\stackrel{\psi}{\lra}A^2\stackrel{\varphi}{\lra} A^2\lra M\lra 0
$$
and the higher Tor-groups $\Tor_i(M,\bar{A})$, as for Theorem \ref{zariski rdp}.
Now suppose we have a free action that leaves $Z_\red\subset Z$ invariant.
This means that $D_2=0$. Since $D_1$ has a zero, the action is not free, contradiction.
\qed
\medskip
In what follows, we assume $p=2$, and consider the derivation
$$
D=D_s + s^2D_\epsilon=s^{-2}D_{s^{-1}}+D_\eta\in H^0(Z,\Theta_{Z/k}),
$$
viewed as a global vector field on the ribbon $Z$. Then we have $D^2=0$,
thus the derivation corresponds to an $\alpha_2$-action on the ribbon.
\begin{proposition}
The $\alpha_2$-action on the ribbon $Z$ is free, and the quotient $Z/\alpha_2$ is the rational cuspidal curve.
\end{proposition}
\proof
Since the coefficients of $D$ do not vanish simultaneously, the subsheaf $D\O_Z\subset\Theta_{Z/k}$
is locally a direct summand, thus the $\alpha_2$-action is free.
At the origin $z\in Z$, which is given by $s=0$, the second coefficient of the derivation $D$ vanishes,
which implies that the orbit ideal is generated by $s^2,\epsilon\in k[s,\epsilon]/(\epsilon^2)$,
whose projective dimension is infinite. Thus the quotient acquires a singularity at the image.
Indeed, the elements $s^2, s^3+\epsilon\in k[s,\epsilon]$ lie in the kernel of $D$,
and the corresponding ring extension is finite flat of degree two. It follows that
$k[s^2,s^3+\epsilon]=k[s,\epsilon]^D$.
At all other points $z\in Z$, the tangent vector points into the ``infinitesimal direction'',
thus $\alpha_2z\subset Z$ is a Cartier divisor. Moreover, the element
$s^{-1}+s^{-2}\eta$, which coincides with $s^{-1}$ up to the invertible factor $1+s^{-1}\epsilon$,
generated the invariant ring $k[s^{-1},\eta]^D$.
Setting $t^2=s^2,t^3=s^3+\epsilon$ and $t^{-1}=s^{-1}+s^{-2}\eta$, we see that the quotient scheme
$Z/\alpha_2$ equals the rational cuspidal curve, with normalization $\PP^1=\Spec(k[t])\cup\Spec(k[t^{-1}])$.
\qed
\medskip
We now see that the ribbon $Z$ is the total space of an $\alpha_2$-torsor on the rational cuspidal curve $C$.
Using the uniqueness of the torsor, we conclude:
\begin{corollary}
The canonical covering $\epsilon:Z\ra C$ of the rational cuspidal curve $C$
defined by the inclusion of $\alpha_2=\Pic_{C/k}[F]$
into the Picard scheme
is given by the ribbon $\PP^1\subset Z$ with ideal $\shI=\O_{\PP^1}(-2)$.
\end{corollary}
The rational cuspidal curve appears as fiber of Kodaira type II in genus-one fibrations.
Let us also treat the other reduced unstable fibers.
Consider the union $A\cup A'$ of two projective lines meeting transversely, and its infinitesimal extension
$A\cup A'\subset Z$ given by two affine charts with coordinate rings
$$
k[s,s']/(ss')^2\quadand k[s^{-1},s']/(s'^2)\times k[s'^{-1},s]/(s^2).
$$
One easily sees that $\Omega_{Z/k}^1$ is locally free of rank two.
The derivation
$$
D=D_s+D_{s'} = (s^{-2}D_{s^{-1}}+D_{s'},s'^{-2}D_{s'^{-1}}+D_s)
$$
defines a global vector field $D\in H^0(Z,\Theta_{Z/k})$
with $D^2=0$, whence an $\alpha_2$-action.
\begin{proposition}
The $\alpha_2$-action on $Z$ is free, and the quotient $Z/\alpha_2$
is the union $C=B\cup B'$ of two copies of $\PP^1$ meeting tangentially.
\end{proposition}
\proof
At no closed point the coordinates for $D$ vanish simultaneously, thus the
$\alpha_2$-action is free.
In the first coordinate ring $A=k[s,s']/(ss')^2$, the elements
$a=s^2$, $b=s'^2$ and $c=s+s'$ are
invariant and satisfy the relation $c^2=a+b$ and $ab=0$.
The ring extension $k[s^2,s'^2,s+s']\subset A$ is freely generated
as a module by $1,s$, which has $D(s)=1$, and we infer that $a,b,c\in A$
indeed generate the full invariant ring.
We have the relations $c^2=a+b$ and $ab=0$. Eliminating $b$, we see
that this defines the union of two affine lines meeting tangentially.
The canonical map
$$
k[a,c]/(ac^2+a^2)=k[a,b,c]/(ab, c^2+a+b)\lra k[s^2,s'^2,s+s']=A^D
$$
is surjective and birational. It is bijective, because the ring on the left is a complete intersection,
thus contains no embedding components.
It follows that $Z/\alpha_2=B\cup B'$, the union of two projective lines meeting tangentially.
\qed
\medskip
One may view $A\cup A'\subset Z$ as the nonsplit ribbon on the union of two projective lines
meeting transversely, with ideal $\shI=\omega_{A\cup A'}$.
the latter is the invertible sheaf corresponding to the class $(-1,-1)\in\Pic(A\cup A')=\Pic(A)\oplus\Pic(A')$.
Clearly, the projection $A\cup A'=Z_\red\ra Z/\alpha_2=B\cup B'$ is not isomorphism,
although the induced maps $A\ra B$, $A'\ra B'$ are isomorphisms.
This determines the canonical covering of a fiber of Kodaira type III:
\begin{corollary}
The canonical covering $\epsilon:Z\ra C$ of the union $C=B\cup B'$ of two copies of $\PP^1$ meeting tangentially
defined by the inclusion of $\alpha_2=\Pic_{C/k}[F]$ into the Picard scheme
is the nonsplit ribbon $A\cup A'\subset Z$ on the transverse union of $A=B$, $A'=B'$
with ideal $\shI=\omega_{A\cup A'}$.
\end{corollary}
We finally come to fibers of type III. Now consider the union $A\cup A'\cup A''$ of three copies of $\PP^1$
meeting at a point with embedding dimension three, and the infinitesimal thickening $A\cup A'\cup A''\subset Z$ defined
by two affine charts with coordinate rings
$$
k[x,y,z]/(xyz,yz+xz+xy)
$$
and
$$
k[x^{-1},y,z]/(yz,x+y)\times k[y^{-1},x,z]/(xz,x+z) \times k[z^{-1},x,y]/(xy,x+y).
$$
The derivation
$$
D=D_x+D_y+D_z = (x^{-2}D_{x^{-1}}+D_y, y^{-2}D_{y^{-1}}+D_z, z^{-2}D_{z^{-1}}+D_x)
$$
defines a global vector field $D\in H^0(Z,\Theta_{Z/k})$ with $D^2=0$,
whence an $\alpha_2$-action.
\begin{proposition}
The $\alpha_2$-action on $Z$ is free, and $Z/\alpha_2$ is isomorphic to the
union $B\cup B'\cup B''\subset\PP^2$ of three lines meeting in a common point.
\end{proposition}
\proof
The coefficients in $D$ do not vanish simultaneously, thus the $\alpha_2$-action is free.
In the first coordinate ring $A=k[x,y,z]/(xyz,yz+xz+xy)$, the elements
$a=y+z$, $b=x+z$, $c=y+z$ are invariant and satisfy the relations $a+b+c=0$ and $abc=0$.
The ring extension $k[a,b]\subset A$ is freely generated as a module by $1,z$.
Since $D(z)=1$ the invariant subring is generated by $a,b$.
The canonical surjection
$$
k[a,b]/(ab(a+b))\lra A^D,\quad a\longmapsto y+z,\quad b\longmapsto x+z
$$
is surjective and birational. It is bijective, because the ring on the left is a complete intersection,
whence contains no embedded components. From this we conclude that
$Z/\alpha_2$ is the union $B\cup B'\cup B''\subset\PP^2$ of three lines meeting at a common point.
Note that the composite maps $A\ra B$, $A'\ra B'$ and $A''\ra B''$ are isomorphisms.
\qed
\medskip
Since $Z_\red\ra B\cup B'\cup B''$ is not an isomorphism, the
torsor $Z\ra Z/\alpha_2$ is not trivial.
Whence:
\begin{corollary}
The canonical covering of the union $C=B\cup B'\cup B''$ of three lines meeting at a common point
given by the inclusion $\alpha_2=\Pic_{C/k}[F]$ is the morphism $\epsilon:Z\ra C$ constructed above.
\end{corollary}
The total spaces of the torsors considered above have the following properties:
\begin{proposition}
Let $C$ be a proper curve whose integral components are rational curves. Suppose that
$h^0(\O_C)=h^1(\O_C)=1$ and
$\omega_C=\O_C$ and $\Pic^0_{C/k}=\GG_a$. Let $\epsilon:Z\ra C$ be the canonical covering corresponding
to $\alpha_2=\Pic_{C/k}[F]$. Then we also have $h^0(\O_Z)=h^1(\O_Z)=1$ and $\omega_Z=\O_Z$ and $\Pic^0_{Z/k}=\GG_a$.
\end{proposition}
\proof
We have a short exact sequence
\begin{equation}
\label{extension for covering}
0\lra\O_C\lra\O_Z\lra\shL\lra 0
\end{equation}
for some invertible $\O_C$-module $\shL$, because $Z\ra C$ is finite and flat of degree two.
Let $C_1,\ldots,C_r\subset C$ be the irreducible components. Since the $\alpha_2$-torsor
becomes trivial on the normalizations $\PP^1\ra C_i$, we must have $\deg(\shL|C_i)=0$
for all irreducible components, and thus $\chi(\shL)=\chi(\O_C)=0$.
Consequently $\chi(\O_Z)=0$.
The next task is to verify $h^0(\O_Z)=1$. Suppose first that $\shL\neq\O_C$.
If there is a nonzero global section $s\in H^0(C,\shL)$, its image in $H^0(C_i,\shL|C_i)$
is nonzero for some irreducible component $C_i$. Since $C_i$ is reduced,
we must have $\deg(\shL|C_i)>0$, contradiction.
Now suppose that $\shL=\O_C$, and choose a trivializing section $s\in H^0(C,\shL)$, which is also a
vector space basis.
Consider the exact sequence
$$
H^0(C,\O_Z)\lra H^0(C,\shL)\lra H^1(C,\O_C).
$$
If the extension \eqref{extension for covering} does not split,
the element $s$ does not lie in the image of the restriction map $H^0(C,\O_Z)\ra H^0(C,\shL)$,
whence the coboundary map is injective, and it follows $h^0(\O_Z)=1$.
Now suppose that the extension does split. Then $\O_Z=\O_C\oplus\O_Cs$, and the algebra structure
is given by $s^2=as+b$ for some global sections $a,b\in H^0(C,\O_C)$. We have $a=0$,
because otherwise the projection $Z\ra C$ would be \'etale somewhere.
Now $Z=C\otimes_kR$
for the algebra $R=k[s]/(s^2- b)$. In turn, $Z_\red\subset Z$ defines a section for the torsor $Z\ra C$.
But the torsor is non-trivial, because the defining map $\alpha_2\ra\Pic_{C/k}$ is nonzero.
Summing-up, we showed $h^0(\O_Z) =1$, and thus also $h^1(\O_Z)=1$.
Since $h^2(\O_Z)=0$ and $h^1(\O_Z)=1$, the Picard scheme $\Pic_{Z/k}$ is smooth
and of dimension one. The map $Z\ra C$ is a homeomorphisms, so the map
$\Pic^0_{C/k}\ra\Pic^0_{Z/k}$ induces an isomorphism on the multiplicative part
\cite{Bosch; Luetkebohmert; Raynaud 1990}.
It follows that $\Pic^o_{Z/k}$ is unipotent.
According to \cite{SGA 3b}, we must have $\Pic_{Z/k}=\GG_a$.
Since $Z\ra C$ is a $G$-torsor for some finite group scheme,
the morphism $Z\ra C$ is of complete intersection, and the relative canonical class must be
$\omega_{Z/C}=\O_Z$. It follows that $\omega_Z=\omega_{Z/C}\otimes \epsilon^*(\omega_Z)=\O_Z$ as well.
\qed
|
\section*{Acknowledgememnts}
This article grew out of a discussion with Duco van Straten who raised
the question of the possible splitting of the sequence
\eqref{eq:Pshortexseq}, see Theorem \ref{thm-calcExt}. We would like
to express our sincere thanks to him and to Christian Sevenheck for
generously sharing their ideas.
\section{Introduction}
During the 1980s {Gel{$'$}fand, Graev, Kapranov and
Zelevinski\u\i\ introduced a class of systems of complex partial
differential equations which are a vast generalization of the
Gau\ss\ hypergeometric equation and which are nowadays known as
\emph{$A$-hypergeometric} (or GKZ) \emph{systems} (\emph{cf.}
\cite{GGZ,GKZ} and a string of other articles of that period). Such
an $A$-hypergeometric system has a hybrid combinatorial and algebraic
flavor, its initial datum being an integer matrix $A$ and a
parameter vector $\beta$ in the column space of $A$. This determines
a left ideal $H_A(\beta)$ in the Weyl algebra $D$ and the
$A$-hypergeometric system with respect to $A$ and $\beta$ is then
the cyclic left $D$-module $M_A^\beta := D/H_A(\beta)$. From this
definition it is far from clear that these systems have any
geometric interpretation.
The analytic behavior of $M_A^\beta$ (as a system of PDEs) is highly
dependent on the parameter vector $\beta$. A technique to study this
dependence, the \emph{Euler--Koszul functor}, was developed by
Matusevich, Miller and the second author in \cite{MMW05}. This is a
functor from the category of toric modules, which are a mild
generalization of $\mathbb{Z} A$-graded $\mathbb{C}[\mathbb{N} A]$ modules, to the category
of complexes of $D$-modules. The construction of this functor
generalizes the Euler--Koszul complex on the semi-group ring $\mathbb{C}[\mathbb{N}
A]$ (already known to {Gel{$'$}fand, Kapranov and
Zelevinski\u\i, \cite{GKZ}) and was inspired by it.
The Euler Koszul complex provides a $D$-resolution of the
corresponding $A$-hypergeometric system provided $\beta$ does not lie
in the \emph{$A$-exceptional locus}, defined via the local cohomology
of $\mathbb{C}[\mathbb{N} A]$. An important step in the geometric
interpretation of $A$-hypergeometric systems was achieved by Schulze
and the second author in \cite{SchulzeWalther-ekdi}, generalizing work
of Gel{$'$}fand et al. There they
showed, using the Euler--Koszul complex, that the Fourier--Laplace
transform of $M_A^\beta$ can be identifed with the direct image of
a twisted structure sheaf on a torus under a monomial map (depending
on $A$) to affine space whenever $\beta$ is outside the set of \emph{strongly
resonant parameters}.
If $A$ is homogeneous, \emph{i.e.}\ if (1,...,1) is in its row span,
then this embedding descends to an embedding of a torus of dimension
one less into projective space. It was realized by Brylinski
\cite{Brylinski} that the Fourier--Laplace transform of a $D$-module
on affine space which is constant on all punctured lines through the
origin can be expressed by a Radon transform of the corresponding
$D$-module on the projective space.
Using this Radon transform the first author showed in \cite{Reich2}
that homogeneous $A$-hypergeometric systems with not strongly
resonant, but integer parameter vector $\beta$ carry the structure of a
mixed Hodge module. Furthermore, there exists a morphism to $M_A^\beta$
from the Gau\ss--Manin system of the maximal
family of Laurent polynomials with Newton polytope equal to the convex
hull of the columns of $A$. This map
has $\mathcal{O}$-free
kernel and cokernel,
and is
compatible with the natural mixed Hodge module structure on the
Gau\ss--Manin system and on $M_A^\beta$ respectively.
Since $M_A^\beta$ is the terminal Euler--Koszul homology of the
semigroup ring $\mathbb{C}[\mathbb{N} A]$ one wonders whether the Euler--Koszul
homology of other toric modules (for example, $A$-graded ideals of
$\mathbb{C}[\mathbb{N} A]$) carry a natural mixed Hodge module structure as well.
In this paper we consider the maximal graded ideal of $\mathbb{C}[\mathbb{N} A]$
and prove that its terminal Euler--Koszul homology is isomorphic to the
Gau\ss--Manin system of a map whose fibers are the complement of the
fibers of the Laurent polynomial alluded to above.
We now give a short overview of the content of this article. In the
first section we review the definition of $A$-hypergeometric systems,
of the Euler--Koszul complex, and several functors on $D$-modules. In
the following section we compute the restriction of $M_A^\beta$ to the
origin, its de Rham cohomology, and the groups
$\Ext^\bullet_D(\mathcal{M}_A^\beta,\mathcal{O})$. A novel feature of this article
is that we work throughout over any field of characteristic zero,
rather than specifically over $\mathbb{C}$. In this more general setting we
(re)prove that for not strongly resonant parameter the
Fourier--Laplace transformed $A$-hypergeometric system can be viewed
as the direct image of a twisted structure sheaf under a torus
embedding. In the third section we show that the long exact Euler--Koszul
homology sequences induced by the inclusion of the maximal graded
ideal in $\mathbb{C}[\mathbb{N} A]$ is isomorphic to certain Gau\ss--Manin systems
coming from a family of Laurent polynomials and compare this sequence
with the sequence obtained in \cite{Reich2}.
\subsection{$A$-hypergeometric systems}
We introduce here the main notation and review some basis facts on
$A$-hypergeometric systems and the Euler--Koszul functor. We refer to
\cite{MMW05,SchulzeWalther-ekdi} for more details.
\begin{ntn}
Throughout, we work over the field $\mathbbm{k}$ of characteristic zero.
In general we adopt the convention that we denote a sheaf by a
calligraphic letter such as $\mathcal{M}$, a module by an Italic letter such
as $M$, and categories and functors by Roman letters such as $\textrm{M}$.
\end{ntn}
\begin{ntn}\label{ntn-A}
Throughout, $A$ will be an integer matrix that we assume to be
pointed: there should be a $\mathbb{Z}$-linear functional on the column space of
$A$ that evaluates positively on each column of $A$.
For any integer matrix $A$, let $\mathbbm{k}^A$ be a vector space with basis
corresponding to the columns $\{{\mathbf{a}}_j\}_j$ of $A$. Let $R_A$
(resp.\ $O_A$) be the polynomial ring over $\mathbbm{k}$ generated by the
variables $\partial_A=\{\partial_j\}_j$ (resp.\ $x_A=\{x_j\}_j$) corresponding
to $\{{\mathbf{a}}_j\}_j$; we read $R_A$ as coordinate ring on the variety
$X_A:=\mathbbm{k}^A$. Further, let $D_A$ be the ring of $\mathbbm{k}$-linear differential
operators on $O_A$, where we identify $\frac{\partial}{\partial x_j}$ with
$\partial_j$ so that both $R_A$ and $O_A$ are subrings of $D_A$.
For any (semi)ring of coefficients $C$ we write $CA$ for the set
of $C$-linear combinations of the columns of $A$. In particular,
$\mathbbm{k} A$ is a vector space.
\end{ntn}
\begin{dfn}\label{dfn-A}
Let $A$ be an integer matrix with independent rows whose $\mathbb{Z}$-ideal of
maximal minors equals $\mathbb{Z}$.
For the parameter $\beta\in\mathbbm{k} A$ let $H_A(\beta)$ be the
$D_A$-ideal generated by the \emph{homogeneity equations}
\begin{eqnarray*}
\{E_i\bullet\phi
&=&\beta_i\cdot\phi\}_i
\end{eqnarray*}
together with the \emph{toric} partial differential \emph{equations}
\begin{eqnarray*}
\{(\partial^{{\mathbf{v}}_+}_A-\partial^{{\mathbf{v}}_-}_A)\bullet \phi=0&\mid&
A\cdot {\mathbf{v}}=0\},
\end{eqnarray*}
using (throughout) multi-index notation.
Here, with ${\mathbf{0}}_A=(0,\ldots,0)$ in $\mathbbm{k}^A$, we write $E_i:=\sum_ja_{ij}x_j\partial_j$ and ${\mathbf{v}}_+=\max({\mathbf{v}},{\mathbf{0}}_A)$,
${\mathbf{v}}_-=-\min({\mathbf{v}},{\mathbf{0}}_A)$.
We put
\[
M_A^\beta:=D_A/H_A(\beta).
\]
\hfill\ensuremath{\diamond}
\end{dfn}
We have
\begin{eqnarray*}
x^{\mathbf{u}} E_i-E_ix^{\mathbf{u}} &=& -(A\cdot {\mathbf{u}})_ix^{\mathbf{u}},\\
\partial^{\mathbf{u}} E_i-E_i\partial^{\mathbf{u}} &=& (A\cdot {\mathbf{u}})_i\partial^{\mathbf{u}}.
\end{eqnarray*}
The \emph{$A$-degree function} (with values in $\mathbb{Z} A$) on $R_A$ and
$D_A$ is:
\[
-\deg_A(x_j):={\mathbf{a}}_j=:\deg_A(\partial_j).
\]
We denote $\deg_{A,i}(-)$ the degree function associated to the weight
given by the $i$-th row of $A$. Then $E_iP=P(E_i-\deg_{A,i}(P))P$ for
$A$-graded $P$.
Let
\[
\varepsilon_A:=\sum_j\deg_A(\partial_j)=\sum_j{\mathbf{a}}_j.
\]
Let $M$ be an $A$-graded $D_A$-module.
There are
commuting $D_A$-linear endomorphisms $E_i$ via
\[
E_i\circ m:=(E_i+\deg_i(m))\cdot m.
\]
for $A$-graded $m\in M$. In particular, if $N$ is an $A$-graded
$R_A$-module one obtains commuting sets of $D_A$-endomorphisms on the
left $D_A$-module $D_A\otimes_{R_A}N$ by
\[
E_i\circ (P\otimes Q):=(E_i+\deg_i(P)+\deg_i(Q))P\otimes Q.
\]
The \emph{Euler--Koszul complex} $K_\bullet(N;E-\beta)$ of the
$A$-graded module $N$ is the homological Koszul complex induced by
$E-\beta:=\{(E_i-\beta_i)\circ\}_i$ on $D_A\otimes_{R_A}N$. In
particular, the terminal module $D_A\otimes_{R_A}N$ sits in
cohomological degree zero. We denote $\mathcal{K}_\bullet(N;E-\beta)$ the
corresponding complex of quasi-coherent sheaves. The cohomology
objects are $H_\bullet(N;E-\beta)$ and $\mathcal{H}_\bullet(N;E-\beta)$
respectively.
If $N(\alpha)$
denotes the usual shift-of-degree functor on the category of
graded $R_A$-modules, then $K_\bullet(N;E-\beta)(\alpha)$ and
$K_\bullet(N(\alpha);E-\beta+\alpha)$ are identical.
Identifying $\mathbb{Z} A$ with $\mathbb{Z}^{\rk(A)}$ we get coordinates $\{t_i\}_i$
on $T_A=\Spec(\mathbbm{k}[\mathbb{Z} A])=\Spec(\mathbbm{k}[\{t_i^\pm\}_i])$ and then an embedding
\begin{gather}
h_A\colon T_A\to \Spec(\mathbb{C}[\{\partial_j\}_j])=\mathbbm{k}^A
\end{gather}
induced by
the monomial morphism}
\begin{gather}
\label{eq-monomial-map}
t:=\{t_i\}_i\to \{\prod_i t_i^{{\mathbf{a}}_{ij}}\}_j=:t^A
\end{gather}
The closure of the image of $h_A$ in
$X_A$ becomes a toric variety via $h_A$ and is defined by the
$R_A$-ideal $I_A$ given as the kernel of
\eqref{eq-monomial-map} and generated by all binomials
$\partial_A^{{\mathbf{v}}_+}-\partial_A^{{\mathbf{v}}_-}$ where $A{\mathbf{v}}=0$. We denote the
semigroup ring
\[
S_A:=R_A/I_A\simeq \mathbbm{k}[\mathbb{N} A].
\]
We denote $\widetilde{\mathbb{N} A}$ the saturation of $\mathbb{N} A$ and by $\tilde
S_A$ the associated semigroup ring, identical with the normalization of
$S_A$.
The \emph{faces} $\tau$ of the rational polyhedral cone $\mathbb{R}_+A$,
\emph{i.e.} the subsets of (the columns of) $A$ that minimize (over
$A$) some linear functional $\mathbb{Z} A\to \mathbb{Z}$, correspond to $A$-graded
prime ideals $I_A^\tau$ of $R_A$ with
$I_A^\tau=I_A+R_A(\{\partial_j\}_{j\not\in \tau})$. We let $R_A\partial_A$ be
the unique $A$-graded maximal $R_A$-ideal.
An $R_A$-module $N$ is \emph{toric} if it is $A$-graded, and if it has
a (finite) $A$-graded composition chain
\[
0=N_0\subsetneq N_1\subseteq N_2\cdots\subsetneq N_k=N
\]
such that each composition factor $N_i/N_{i-1}$ is isomorphic as
$A$-graded $R_A$-module to a face ring $R_A/I_A^\tau$ or one
of its shifts by an element of $\mathbb{Z} A$.
For a finitely generated $A$-graded $R_A$-module
$N=\bigoplus_{\alpha\in\mathbb{Z} A}
N_\alpha$, let
\begin{eqnarray*}
\deg_A(N)&=&\{\alpha\in\mathbb{Z} A\mid N_\alpha\not=0\},\\
\qdeg_A(N)&=&\bar{\deg_A(N)}^{Zar},
\end{eqnarray*}
the latter being the Zariski closure of the former in
$\mathbbm{k} A=\mathbbm{k}\otimes_\mathbb{Z}\ZZ A$. For unions of such modules, degrees as
well as quasi-degrees are defined to be the corresponding unions, compare
\cite{SchulzeWalther-ekdi}.
Let $N=\mathbbm{k}(-\alpha)$ be the graded $R_A$-module whose module structure
is that of $R_A/I_A^\emptyset=R_A/R_A\partial_A\simeq \mathbbm{k}$, and
which lives entirely inside degree $\alpha\in\mathbb{Z}^d$.
Then $K_\bullet(N;E-\beta)$ is an exact complex if $\beta\neq
\alpha$, and its differentials are zero otherwise.
\begin{dfn}
If the row span of $A$ contains ${\mathbf{1}}_A$ we call $A$
\emph{homogeneous}. Homogeneity is equivalent to $I_A$ defining a
projective variety, and to the system $H_A(\beta)$ having only regular
singularities \cite{SchulzeWalther-Duke}.
\hfill\ensuremath{\diamond}\end{dfn}
\subsection{$\mathcal{D}$-module functors}
Let $X$ be a smooth algebraic $\mathbbm{k}$-variety of dimension $d_X$. We
denote by $\mathcal{D}_X$ the sheaf of algebraic differential operators and
by $D_X$ its ring of global sections. For $X = \AA^n$ we sometimes
write $D_n$ . We denote by ${\mathrm{Mod}}(\mathcal{D}_{X})$ the Abelian category of
left $\mathcal{D}_{X}$-modules. The full triangulated subcategories of the
derived category $\mathrm{D}^b(\mathcal{D}_X):=\mathrm{D}^b({\mathrm{Mod}}(\mathcal{D}_X))$ consisting of
objects with $\mathcal{O}_{X}$-quasi-coherent (resp.\ holonomic) cohomology
are denoted by $\mathrm{D}^b_{qc}(\mathcal{D}_X)$ (resp. $\mathrm{D}^b_{h}(\mathcal{D}_X))$.
We recall the notation for cohomological shifting a complex
$C^\bullet$: $C^\bullet[1]$ is the complex $C^\bullet$ shifted one
step left, $(C^\bullet[1])^i=C^{i+1}$, with corresponding shift of the
morphisms.
Let $f: X \rightarrow Y$ be a map between smooth algebraic
varieties.
Let $\mathcal{M} \in \mathrm{D}^b_{qc}(\mathcal{D}_X)$ and $\mathcal{N} \in
\mathrm{D}^b_{qc}(\mathcal{D}_Y)$, then we denote by
\[
f_+ \mathcal{M} := {\mathrm{R}} f_*(\mathcal{D}_{Y \leftarrow X} \overset{L}{\otimes} \mathcal{M})\quad
\text{ and }\quad f^+ \mathcal{N}:= \mathcal{D}_{X \rightarrow Y} \overset{L}{\otimes}
f^{-1}\mathcal{N} [d_X - d_Y]
\]
the direct and inverse image functors for $\mathcal{D}$-modules; both
preserve holonomicity and if $f$ is non-characteristic with respect to $\mathcal{N}$ then $f^+$ is exact (up to a shift),
(see e.g. \cite[Def. 2.4.2 \& Thm 2.4.6]{HottaTakeuchiTanasaki}).
\noindent We denote by
\begin{align}
\mathbb{D}: \mathrm{D}^b_h(\mathcal{D}_X) &\longrightarrow (\mathrm{D}^b_h(\mathcal{D}_X))^{opp} \notag \\
\mathcal{M} &\mapsto {\mathrm{R}}\sHom(\mathcal{M},\mathcal{D}_X \otimes_{\mathcal{O}_X} \Omega^{\otimes-1}_X)[d_X]\notag
\end{align}
the duality functor, which also preserves holonomicity.
We additionally define the functors
\[
f_\dag := \mathbb{D} \circ f_+
\circ \mathbb{D} \quad\text{ and }\quad f^\dag := \mathbb{D} \circ f^+ \circ \mathbb{D}.
\]
If $X$ is an affine variety, we have an equivalence of categories
\begin{align}
{\mathrm{Mod}}(\mathcal{D}_X) &\longrightarrow {\mathrm{Mod}}(D_X) \notag \\
\mathcal{M} &\mapsto M:= \Gamma(X,\mathcal{M})
\end{align}
where ${\mathrm{Mod}}(D_X)$ is the category of left $D_X$-modules.
\begin{dfn}\label{dfn-FL}
Let
\[
\langle -,-\rangle: \AA^{\ell} \times \hat{\AA}^{\ell} \rightarrow \AA^1, \qquad
(\lambda_1,\ldots, \lambda_\ell, \mu_1,\ldots \mu_\ell)
\mapsto \sum_{i=1}^\ell \lambda_i \mu_i.
\]
(Here, and elsewhere, $\hat{\AA}^\ell$ denotes an affine space of
dimension $\ell$; we use the ``hat'' to keep apart source and range
of the two functors defined in \eqref{eq-FL-dfn} below).
Now define two $\mathcal{D}_{\AA^\ell\times\hat{\AA}^\ell}$-modules by
\[
\mathcal{L}:= \mathcal{O}_{\AA^\ell \times \hat{\AA}^\ell} e^{\langle\cdot ,\cdot
\rangle},\qquad \overline{\mathcal{L}}:= \mathcal{O}_{\AA^\ell \times \hat{\AA}^\ell}
e^{-\langle\cdot ,\cdot \rangle}.
\]
We refer to \cite[Section 5]{KaSc} for details on these sheaves.
Denote by $p_1: \AA^\ell \times \hat{\AA}^\ell \rightarrow \AA^\ell$ and $p_2: \AA^\ell \times \hat{\AA}^\ell \rightarrow \hat{\AA}^\ell$ the
projection to the first and second factors respectively. The
\emph{Fourier--Laplace transform} is defined by
\begin{align}\label{eq-FL-dfn}
\FL: \mathrm{D}^b_{qc}(\mathcal{D}_{\AA^\ell}) &\longrightarrow \mathrm{D}^b_{qc}(\mathcal{D}_{\hat{\AA}^\ell}) \notag \\
M &\mapsto p_{2+} (p_1^+ M \overset{L}\otimes \mathcal{L})
\end{align}
and
\begin{align}
\FL^{-1}: \mathrm{D}^b_{qc}(\mathcal{D}_{\hat{\AA}^\ell}) &\longrightarrow \mathrm{D}^b_{qc}(\mathcal{D}_{\AA^\ell}) \notag \\
M &\mapsto p_{1+} (p_2^+ M \overset{L}\otimes \overline{\mathcal{L}})
\end{align}
Then $ \FL^{-1} \circ \FL(M) \simeq \iota^+ M$ where $\iota$ is given
by $\lambda \mapsto -\lambda$, and we set
\[
\hat{\mathcal{M}}^\beta_A := \FL^{-1} (\mathcal{M}^\beta_A)
\]
with global sections $\hat M_A^\beta$.
\hfill\ensuremath{\diamond}\end{dfn}
\begin{ntn}\label{ntn-FL-stuff}
If $\AA^\ell$ and $\hat{\AA}^\ell$ are an $\FL$-pair with
$\AA^\ell=\mathbbm{k}^A$ for some matrix $A$, we shall denote by $\hat R_A,
\hat O_A, \hat S_A,\ldots$ the $A$-graded objects on
$\hat{\AA}^\ell$ corresponding to the $A$-graded objects $R_A, O_A,
S_A,\ldots$ on $\AA^\ell$.
\hfill\ensuremath{\diamond}\end{ntn}
\section{Restriction and de Rham functors of Euler--Koszul complexes}
In this section we make some computations considering certain functors
on the class of (generalized) hypergeometric systems.
\subsection{Local cohomology}
Relevant in several ways are the local cohomology functors
$H^\bullet_{\partial_A}(-)$ given as the higher derived functors of the
$\partial_A$-torsion functor
\[
\Gamma_{\partial_A}(M):=\{m\in M\mid
\partial_i^k \cdot m=0\, \forall k\gg 0, \forall i\},
\]
a subfunctor of the
identity functor on the category of $R_A$-modules. If $M$ is
$A$-graded, so are all $H^i_{\partial_A}(M)$ since the support ideal is
$A$-graded. See \cite{24h} for details and background.
\begin{lem}
\label{lem-lc-tor}
For any $R_A$-module $N$ there is a
functorial isomorphism
\[
{\mathrm{R}}\Gamma_{\partial_A}(N)[\dim(X_A)]=(D_A/\partial_AD_A)\otimes^L_{R_A}N
\]
so that $H^\bullet_{\partial_A}(N)=
\Tor^{R_A}_{\dim(X)-\bullet}(D_A/\partial_AD_A,N)$. Any $R_A$-grading
$\deg(-)$ on $N$ makes this isomorphism graded if the right side is
shifted by $\sum_j\deg(\partial_j)$.
\end{lem}
\begin{proof}
One representative for ${\mathrm{R}}\Gamma_{\partial_A}(-)$ is the \v Cech
(\emph{i.e.}, stable Koszul) complex ${\check
C}_A^\bullet(-)=(-)\otimes_{R_A}\bigotimes_j(R_A\to R_A[1/\partial_j])$.
On $R_A$, this returns a $D_A$-complex with unique cohomology group,
in cohomological degree $\dim X_A$, given by
$\bigoplus_{{\mathbf{v}}<0}\mathbbm{k}\cdot \partial^{\mathbf{v}}$ where ${\mathbf{v}}$ is
componentwise negative. The $D_A$-isomorphism of this module with
$D_A/D_A\partial_A$ that identifies the coset of $1/\prod_j \partial_j$ in
the former with the coset of $1$ in the latter (is $A$-graded of
degree $\varepsilon_A$ and) shows that (up to this shift in degree) this is
the injective hull of
$R_A/R_A\cdot\partial_A$ over $R_A$.
The anti-automorphism
induced by $x^{\mathbf{u}}\partial^{\mathbf{v}}\to \partial^{\mathbf{v}}(-x)^{\mathbf{u}}$ allows to
view ${\check C}_A^\bullet$ as complex of right $D_A$-modules without
affecting the $R_A$-structure. Then $H^{\dim X_A}_{\partial_A}(R_A)=
D_A/\partial_AD_A$ is the canonical module of $R_A$ with its natural right
$D_A$-structure.
The modules in ${\check C}_A^\bullet$ are flat, so ${\check
C}_A^\bullet\otimes_{R_A}N[\dim X_A]=D_A/\partial_A
D_A\otimes_{R_A}^LN$. If $N$ is graded, then---since $\partial_A$ is
monomial---$D_A/\partial_AD_A$ and its flat resolution ${\check
C}_A^\bullet$ are also graded. Hence ${\check
C}_A^\bullet\otimes_{R_A}N$ has graded cohomology. The
identification $H^{\dim X_A}({\check C}_A^\bullet)[\dim X_A]\simeq
D_A/\partial_AD_A$ shifts the grading by the degree of the socle
element $1/\prod_j\partial_j$ of the left hand side.
\end{proof}
\subsection{Strongly resonant parameters}
We recall from \cite{MMW05,SchulzeWalther-ekdi} the following important
sets. The \emph{exceptional locus} ${\mathcal{E}}_A$ is
\[
{\mathcal{E}}_A:=\qdeg_A\left(\bigoplus_{k>\dim X_A-\dim
T_A}\Ext^k_{R_A}(S_A,R_A)\right)=\bigcup_{k<\dim T_A} \left(\bar{\deg_A
H^k_{\partial_A}(S_A)}^{Zar}\right).
\]
A larger
interesting set is
\[
\sRes(A):=\bigcup_j\qdeg_A(H^1_{\partial_j}(S_A)),
\]
the \emph{strongly resonant} parameters of $A$.
For $\mathbbm{k}=\mathbb{C}$ the following results were shown in
\cite{MMW05,SchulzeWalther-ekdi}. A parameter is
in ${\mathcal{E}}_A$ if and only if the complex $\mathcal{K}_\bullet(S_A;E-\beta)$
fails to be a resolution of $\mathcal{M}_A^\beta$; it is in $\sRes(A)$ if
and only if $\mathcal{K}_\bullet(S_A;E-\beta)$ fails to resolve the
Fourier--Laplace transform of $h_{A+}(\mathcal{O}_{T_A}^\beta)$ where
\[
\mathcal{O}_{T_A}^\beta=\mathcal{D}_{T_A}/\mathcal{D}_{T_A}(\{\partial_{t_i}t_i+\beta_i\}_i),
\]
or
alternatively if and only if $h_{A+}(\mathcal{O}_{T_A}^\beta)$ disagrees with
$\hat\mathcal{M}_A^\beta$. We are interested in these results over $\mathbbm{k}$:
\begin{thm}
Let $\mathbbm{k}$ be an arbitrary field of characteristic zero. For each $j$,
the following are equivalent:
\begin{enumerate}
\item $\beta\not\in\sRes_j(A):=\qdeg_A(H^1_{\partial_j}(S_A))$;
\item left-multiplication by $\partial_{x_j}$ is a quasi-isomorphism on
$K_\bullet(E-\beta;S_A)$.
\end{enumerate}
\end{thm}
\begin{cor}\label{cor:SW}
Over any coefficient field $\mathbbm{k}$ of characteristic zero, the following
are equivalent:
\begin{enumerate}
\item $\beta\not\in\sRes(A)$;
\item $\mathcal{K}_\bullet(E-\beta;S_A)$ represents the Fourier--Laplace transform of
$h_{A+}\mathcal{O}_{T_A}^\beta
\item $\mathcal{M}_A^\beta$ is naturally isomorphic to the Fourier--Laplace transform
of $\mathcal{H}^0 h_{A+}\mathcal{O}_{T_A}^\beta$
\end{enumerate}
\end{cor}
Inspection shows that, apart from formal computations that do not
depend on $\mathbbm{k}$, there are the following logical dependencies in
\cite{SchulzeWalther-ekdi}.
\begin{itemize}
\item \cite[Cor.~3.7]{SchulzeWalther-ekdi} needs \cite[Thm.~3.5,
Cor.~3.1,Prop.~2.1]{SchulzeWalther-ekdi} and
\cite[Prop.~5.3]{MMW05}, and the fact that higher Euler--Koszul
homology is $(\prod_j\partial_j)$-torsion.
\item \cite[Thm.~3.5]{SchulzeWalther-ekdi} needs
\cite[Lem.~3.2]{SchulzeWalther-ekdi} and
\cite[Prop.~5.3]{MMW05}.
\item \cite[Lem.~3.2]{SchulzeWalther-ekdi} is completely formal and
independent of the field $\mathbbm{k}$.
\item \cite[Cor.~3.1]{SchulzeWalther-ekdi} needs the left and right
\O re properties of $D_{A}$, and
\cite[Prop.~2.1]{SchulzeWalther-ekdi}.
\item \cite[Prop.~2.1]{SchulzeWalther-ekdi} needs that direct images
over $\mathbbm{k}$ are formally the same for all $\mathbbm{k}$ (which they are),
plus $\mathcal{D}$-affinity of tori, plus various formal computations
contained in \cite{Borel}, namely an identification of a direct
image module in VI.7.3, the chain rule in VI.4.1, exactness of
direct images for affine closed embeddings in VI.8.1, and equality of direct
images under open embeddings in the $\mathcal{D}$- and $\mathcal{O}$-category in
VI.5.2.
\end{itemize}
Tori are $\mathcal{D}$-affine since they are $\mathcal{O}$-affine. Higher
Euler--Koszul homology is $(\prod_j\partial_j)$-torsion since localizing
every $\partial_{j}$ leads to Euler--Koszul homology of the quasi-toric
Cohen--Macaulay module $\mathbbm{k}[\mathbb{Z} A]$ (compare
\cite{SchulzeWalther-ekdi} for quasi-toricity). The \O re properties
of $D_A$ rely on the Leibniz rule and are unaffected by $\mathbbm{k}$. Any
closed embedding over $\mathbbm{k}$ can be base-changed to a closed embedding
(and hence to an affine faithful map) over $\mathbb{C}$, by viewing $\mathbbm{k}$
(algebraically) as a subfield of $\mathbb{C}$. Since $\mathbb{C}$ is fully faithful
over $\mathbbm{k}$ and affine faithful maps over $\mathbb{C}$ yield exact direct
image functors, so do they over $\mathbbm{k}$. Direct images for open
embeddings agree over $\mathcal{D}$ and $\mathcal{O}$ more or less by definition.
It therefore remains to inspect \cite[Prop.~5.3]{MMW05} and exactness
of the Euler--Koszul complex on maximal Cohen--Macaulay input over
$\mathbbm{k}$.
Using superscripts to indicate base fields,
$K_A^\mathbbm{k}(D_A^\mathbbm{k};E-\beta)\otimes_\mathbbm{k}\mathbb{C} = K_A^\mathbb{C}(D_A^\mathbb{C};E-\beta)$ as
long as $\beta\in\mathbbm{k} A$. The notion of a toric module
is formally independent of $\mathbbm{k}$, and so the categories of toric modules
and their Euler--Koszul complexes embed into one another for
containments of fields. In particular, the formal mechanisms are
identical and scale from one field to another faithfully.
The required part of \cite[Prop.~5.3]{MMW05} is the equivalence
(3)$\Leftrightarrow$(4). The proof passes through the equivalences
(2)$\Leftrightarrow$(3) and (2)$\Leftrightarrow$(4). For both we
need, modulo formal computations involving toric composition chains,
only to check the equivalence of conditions (1) and (3) in
\cite[Lem.~4.9]{MMW05}. The implication (1)$\Rightarrow$(3) is linear
algebra over any field. The reverse follows by contradiction from base
change to $\mathbb{C}$.
Finally, if $M$ is a maximal Cohen--Macaulay toric module over $\mathbbm{k}$
then vanishing of higher Euler--Koszul homology follows like over
$\mathbb{C}$ from the spectral sequence \cite[Thm.~6.3]{MMW05} since the
existence of the spectral sequence is abstract homological
nonsense. However, this use of the spectral sequence requires the
concept of holonomicity: one would like to use that Euler--Koszul
homology modules are holonomic and that therefore their duals are modules.
The Euler--Koszul homology modules induce $\mathcal{D}_A$-modules on affine
space. On that class, (dimension, and hence) holonomicity can be
defined over all fields, via the theory of good filtrations. That
holonomic modules have holonomic modules as their duals was proved by Roos, see
the Bernstein notes \cite[Thm.~3.15]{Bernstein-notes}.
\subsection{Restriction to the origin}
Let $\rho$ be the restriction functor to ${\mathbf{0}}_A\in\mathbbm{k}^A$,
\[
\rho(-):=(D_A/x_AD_A)\otimes^L_{R_A}(-)
\]
from the category of
($A$-graded) $\mathcal{D}_A$- or $D_A$-modules to the category of ($A$-graded)
$\mathbbm{k}$-vector spaces. Denote $\rho_k(-)$ its $k$-th homology.
We start with a topological observation derived from
\cite{SchulzeWalther-ekdi}. By $H_{dR}^\bullet(-;\mathbbm{k})$ we mean the
algebraic de Rham cohomology in the sense of Grothendieck
\cite{Grothendieck-dR}.
\begin{lem}\label{lem-restr-sRes}
If $\beta\not\in\sRes(A)$ then $\rho(\mathcal{M}_A^\beta)$ is
naturally identified with the homology of the local system to
$\mathcal{O}_{T_A}^\beta$ on the torus $T_A$:
\[
H_j(\rho(\mathcal{M}_A^\beta))=\rho_j(\mathcal{M}_A^\beta)\simeq \left\{\begin{array}{cr}
H_{dR}^j (T_A;\mathbbm{k})&\text{if } \beta\in\mathbb{Z} A\smallsetminus \sRes(A);\\
0&\text{if }\beta\in\mathbbm{k} A\smallsetminus(\mathbb{Z} A\cup\sRes(A)).
\end{array}\right.
\]
\end{lem}
\begin{proof}
By \cite{SchulzeWalther-ekdi} and \ref{cor:SW}, if
$\beta\not\in\sRes(A)$ then $\FL^{-1}(\mathcal{M}_A^\beta)\simeq
h_{A+}(\mathcal{O}_{T_A}^\beta)$. Under Fourier--Laplace, restriction
$\rho$ converts to the functor $(D_A/\partial_AD_A)\otimes^L_{D_A}(-)$. On
the affine space $X_A=\mathbbm{k}^A$ this is the $D$-module direct image under
the map to a point. Hence with $\beta\not\in\sRes(A)$,
$\rho(\mathcal{M}_A^\beta)$ represents the direct image of $\mathcal{O}_{T_A}^\beta$
under projection to a point---in other words, the cohomology of the
local system.
\end{proof}
We next extend this lemma by identifying algebraically
$\rho(M_A^\beta)$ with $\bigwedge(\mathbbm{k} A)$ for non-exceptional
$\beta$. (We view the exterior algebra as an abstract copy of the
cohomology of $T_A$). Note that in this case the Euler--Koszul complex resolves
$\mathcal{M}_A^\beta$ but is not necessarily a representative for
$\FL^{-1} h_{A+}(\mathcal{O}_{T_A}^\beta)$. Studying restrictions of Euler--Koszul
complexes turns out to be very down to earth.
\begin{lem}\label{lem-restr-EK}
If $\phi\colon N\to N'$ is an $A$-graded morphism (of degree zero) of
$A$-graded $R_A$-modules then
\begin{asparaenum}
\item
the restriction $\rho(\mathcal{K}_\bullet(N;E-\beta))$ is naturally
$\bigwedge (\mathbbm{k} A)\otimes_\mathbbm{k} N_\beta$, in the sense that
\item
the induced morphism
$\rho(\mathcal{K}_\bullet(N;E-\beta))\to \rho(\mathcal{K}_\bullet(N';E-\beta))$ is
identified with the morphism $\bigwedge(\mathbbm{k} A)\otimes_\mathbbm{k} N_\beta \to
\bigwedge (\mathbbm{k} A)\otimes_\mathbbm{k} N'_\beta$ induced from $\phi_\beta$.
\end{asparaenum}
\end{lem}
\begin{proof}
We extend the domain of the Euler--Koszul functor to modules of the
form $Q\otimes_{R_A}N$ where $N$ is an $A$-graded $R_A$-module and $Q$
a right $A$-graded $D_A$-module by setting $E_i\circ(q\otimes
\nu)=q(E_i+\deg_{A,i}(\nu))\otimes \nu$.
Morally, $E_i\circ(-)$ remains right-multiplication by $E_i$ and
(since multiplications on the left and right commute) one easily
checks that there is an isomorphism of functors
$\rho(\mathcal{K}_\bullet(-;E-\beta))=\mathcal{K}_\bullet(\rho(D_A\otimes_{R_A}(-));E-\beta)=
\mathcal{K}_\bullet((D_A/x_AD_A)\otimes^L_{R_A}(-);E-\beta)$ from the category of
$A$-graded $R_A$-modules to the category of $A$-graded vector
spaces.
As right $R_A$-module, $(D_A/x_AD_A)\otimes_{R_A}N=N$ for any
$A$-graded $N$. Hence
$(D_A/x_AD_A)\otimes^L_{R_A}(-)=(D_A/x_AD_A)\otimes_{R_A}(-)$. The
$E_i$-action is then $E_i\circ(1\otimes \nu)=(\deg_{A,i}(\nu))\otimes
\nu$. In particular, the Euler--Koszul complex of $E-\beta$ on
$(D_A/x_AD_A)\otimes N$ is in degree $\alpha\in\mathbb{Z} A$ the Koszul
complex on $N_\alpha$ induced by the numbers
$\{\alpha_i-\beta_i\}_i$. If $\alpha=\beta$ then this Koszul complex
is $\bigwedge(\mathbbm{k} A)\otimes_\mathbbm{k} N_\alpha$ with zero differential. If
$\alpha\not=\beta$ then this Koszul complex is the Koszul complex
(over $\mathbb{Z}$) of a set of generators of the unit ideal and hence exact.
The final claim is clear from the construction.
\end{proof}
\begin{cor}
If $\beta\not\in{\mathcal{E}}_A$ then
\[
\rho_j(\mathcal{M}_A^\beta)\simeq \left\{\begin{array}{ccc}
H_{dR}^j (T_A;\mathbbm{k})&\text{ if }& \beta\in\mathbb{N} A;\\
0 &\text{ if }& \beta\not\in\mathbb{N} A.\end{array}\right.
\]
\end{cor}
\begin{proof}
If $\beta\not\in{\mathcal{E}}_A$ then
$\rho(\mathcal{M}_A^\beta)=\rho(\mathcal{K}_\bullet(S_A;E-\beta))$. Now use Lemma
\ref{lem-restr-EK}.
\end{proof}
The natural morphism $\rho(\mathcal{K}_\bullet(S_A;E-\beta))\to
\rho(\mathcal{M}_A^\beta)$ need not be an isomorphism:
\begin{exa}
Let $A=\begin{pmatrix}1&1&1&1\\0&1&3&4\end{pmatrix}$ and take
$\beta=(1,2)$, the only parameter with higher Euler--Koszul homology
for this $A$ (by \cite{SturmfelsTakayama98}).
The $\mathbbm{k}$-dimension vectors for $\rho(\mathcal{M}_A^\beta)$ and
$\rho(\mathcal{K}_\bullet(S_A;E-\beta))$ are $(0,0,1,0,0)$ and
$(0,0,0,0,0)$ respectively.
\hfill\ensuremath{\diamond}\end{exa}
In order to better understand the relationship between the
restrictions of the $A$-hypergeometric system
and the Euler--Koszul complex respectively, we consider the $3$-rd
quadrant spectral sequence
\[
E^2_{-i,-j}=\rho_j(\mathcal{H}_i(-;E-\beta))\Longrightarrow
(\rho(\mathcal{K}_\bullet(-;E-\beta)))_{i+j}.
\]
The $k$-th differential is $d_k\colon E^k_{-p,-q}\to
E^k_{-p-k+1,-q+k}$. A toric map $N\to N'$ induces a morphism of
corresponding spectral sequences.
All our experiments indicate that if $\beta\in\mathbb{N} A$ then
$\rho_j(\mathcal{M}_A^\beta)=H_{dR}^j(T_A)$, irrespective of exceptionality. While
we cannot show that, we have a one-way estimate:
\begin{lem}
If $\beta\in\mathbb{N} A$ then there is a natural inclusion $\bigwedge(\mathbbm{k} A)\hookrightarrow
\rho(\mathcal{M}_A^\beta)$.
\end{lem}
\begin{proof}
Consider the morphism of spectral sequences attached to the inclusion
$S_A\hookrightarrow \tilde S_A$ of $S_A$ into its normalization.
For any $\beta\in\mathbb{N} A$, the
induced map of abutments ${\rho(\mathcal{K}_\bullet(S_A;E-\beta))\to
\rho(\mathcal{K}_\bullet(\tilde S_A;E-\beta))}$ is an isomorphism
by Lemma \ref{lem-restr-EK}. Since $\tilde S_A$ is Cohen--Macaulay, it
has no higher Euler--Koszul homology and so the abutment
$\rho(\mathcal{K}_\bullet(\tilde S_A;E-\beta))$ is stored in the $i=0$ column
of the $E^2$-term. It follows that the isomorphism on abutments must
be coming from the map of the $i=0$ column, for $k\gg 0$. But
$E^k_{0,-j}$ is a submodule of $E^2_{0,-j}$ for $k\geq 2$. In
particular, $\rho(\mathcal{K}_\bullet(\tilde S_A;E-\beta))\simeq \bigwedge (\mathbbm{k} A)$
is contained in $\rho(\mathcal{K}_\bullet(S_A;E-\beta))=\rho(\mathcal{M}_A^\beta)$.
\end{proof}
\subsection{De Rham cohomology}
We consider now the effect of $(D/\partial_AD_A)\otimes_{D_A}^L(-)$ on
$M_A^\beta$ and on the Euler--Koszul complex. This behaves differently
since $(D_A/\partial_AD_A)\otimes_{D_A}N$ is not $N$ for most $A$-graded
$R_A$-modules $N$.
\begin{dfn}
If $\beta$
is in $\deg_A({\mathrm{R}}\Gamma_{\partial_A}(S_A))=\bigcup_{k<\dim
T_A}\deg_A(H^k_{\partial_A}(S_A))\subseteq {\mathcal{E}}_A$ it is called
\emph{strongly $A$-exceptional}.
\hfill\ensuremath{\diamond}\end{dfn}
\begin{thm}\label{thm-deRham-EK}
For any $A$-graded $R_A$-module $N$,
$(D/\partial_AD_A)\otimes_{D_A}^L K_\bullet(N;E-\beta)$ vanishes
whenever $\beta$ is not an $A$-degree of
${\mathrm{R}}\Gamma_{\partial_A}(N))$. More precisely,
\[
(D/\partial_AD_A)\otimes_{D_A}^L K_\bullet(N;E-\beta)
\simeq
\left(\bigoplus_i H^i_{\partial_A}(N)\right)_\beta\otimes_\mathbbm{k}
\bigwedge(\mathbbm{k} A)[\dim X_A].
\]
As in Lemma \ref{lem-restr-EK}, an $A$-graded map $N\to N'$ induces a
map of de Rham complexes that is identified with
$\left({\mathrm{R}}\Gamma_{\partial_A}(N)\to
{\mathrm{R}}\Gamma_{\partial_A}(N')\right)_\beta\otimes_\mathbbm{k} \bigwedge(\mathbbm{k} A)[\dim
X_A]$.
If $\beta$ is not strongly exceptional (\emph{e.g}, if $S_A$ is Cohen--Macaulay),
then
\[
\Tor_\bullet^{D_A}(D/\partial_AD_A,K_\bullet(S_A;E-\beta))=
H_{dR}^{\bullet+\dim X_A}(T_A;\mathbbm{k})
\]
if $\beta$ is in $\deg_A(H^{\dim T_A}_{\partial_A}(S_A))$ and
zero otherwise.
\end{thm}
\begin{proof}
As in the proof of Lemma \ref{lem-restr-EK},
we extend the action of the Euler operators to the
quotient $(D_A/\partial_AD_A)\otimes_{R_A}N$ for any $A$-graded $N$. Hence
$(D_A/\partial_AD_A)\otimes^L_{D_A}K_\bullet(N;E-\beta)=
K_\bullet((D_A/\partial_AD_A)\otimes^L_{R_A}N;E-\beta)$
for any $A$-graded $R_A$-module $N$.
Recall that $\varepsilon_A=\sum_j{\mathbf{a}}_j$ and that its components
$\varepsilon_{A,i}$ satisfy $E_i+\varepsilon_{A,i}=\sum_ja_{ij}\partial_jx_j$. Take now
a free $A$-graded $R_A$-resolution $F_\bullet$ for $N$. Then for any
$A$-graded element $P\otimes f\in (D_A/\partial_AD_A)\otimes_{R_A}F_k$ the
cosets of $(E_i-\beta_i)\circ(P\otimes f)$, of
$(E_i-\beta_i+\deg_{A,i}(P\otimes f))P\otimes f$ and of
$(-\varepsilon_{A,i}-\beta_i+\deg_{A,i}(P\otimes f))P\otimes f$ coincide. So,
as in Lemma \ref{lem-restr-EK}, the Euler--Koszul complex on
$(D_A/\partial_AD_A)\otimes F_\bullet$ is in degree $\alpha$ the Koszul
complex on $D_A/\partial_AD_A\otimes F_\bullet$ induced by the numbers
$\{-\varepsilon_{A_i}-\beta_i+\alpha_i\}$. Hence
$K_\bullet((D_A/\partial_AD_A)\otimes^L_{R_A}N;E-\beta)$ can only have
cohomology when $(D_A/\partial_AD_A)\otimes^L_{R_A}N$ has a cohomology
class in degree $\beta+\varepsilon_A$. By Lemma \ref{lem-lc-tor} this is
equivalent to $\beta$ being the degree of a nonzero cohomology class
in ${\check C}_A^\bullet\otimes_{R_A}N$ which proves the first claim.
If ${\mathrm{R}}\Gamma_{\partial_A}(N)$ is non-exact in degree $\beta$ then
$(D_A/\partial_AD_A)\otimes^L_{R_A}(K_\bullet(N;E-\beta-\varepsilon_A))$ is
$(H^\bullet_{\partial_A}(N))_\beta$ tensored with a Koszul complex (shifted by $\dim X_A$) on
$\dim (T_A)$ maps $\mathbbm{k}\to\mathbbm{k}$ each of which is the zero map. Hence in
this case, the resulting cohomology is
$(H^\bullet_{\partial_A}(N))_\beta\otimes_\mathbbm{k} \bigwedge(\mathbbm{k} A)[\dim X_A]$.
The indicated naturality condition is clear from the discussion.
If $\beta$ is not strongly exceptional then
$({\mathrm{R}}\Gamma_{\partial_A}(S_A))_\beta\simeq (H^{\dim
T_A}_{\partial_A}(S_A))_\beta$. The latter is a subquotient of $\mathbbm{k}[\mathbb{Z}
A]$ and hence its $A$-graded Hilbert function takes values in
$\{0,1\}$. The final claim follows.
\end{proof}
\begin{exa}
Let $A=\begin{pmatrix}1&1&1&1\\0&1&3&4\end{pmatrix}$. Then
$(H^2_{\partial_A}(S_A))_\beta$ is nonzero exactly if $\beta$ is an
interior lattice
point of $-\mathbb{R}_+ A$, while $H^1_{\partial_A}(S_A)$ is a $1$-dimensional
vector space concentrated in degree $(1,2)$.
It follows that
$\Tor_\bullet^{D_A}(D/\partial_AD_A,K_\bullet(S_A;E-\beta))$ is
$H^{\bullet+4}(T_A;\mathbbm{k})$ when $\beta$ supports $H^2_{\partial_A}(S_A)$; it is
the shifted $H^{\bullet+4}(T_A;\mathbbm{k})[1] =H^{\bullet+5}(T_A;\mathbbm{k})$ when $\beta=(1,2)$; it is zero in all
other cases.
In particular, no simple general formula (not appealing to local
cohomology modules) for
$\Tor_\bullet^{D_A}(D/\partial_AD_A,K_\bullet(S_A;E-\beta))$ comes to
mind.
\hfill\ensuremath{\diamond}\end{exa}
\begin{cor}\label{cor-TorExt-zero}
$\Tor_\bullet^{D_A}(D/\partial_AD_A,M_A^\beta)$ and
$\Ext^\bullet_{D_A}(O_A,M_A^\beta)$ are nonzero only if
\[
\beta\in
{\mathcal{E}}_A\cup(\deg_A(H^{\dim(S_A)}_{\partial_A}(S_A))).
\]
\end{cor}
\begin{proof}
Note first that resolving $O_A$ over $D_A$ and dualizing the
resolution gives a resolution of (a cohomologically shifted)
$D_A/\partial_AD_A$, so that the Ext- and Tor-claims are equivalent.
By \cite{MMW05}, the Euler--Koszul complex resolves
$M_A^\beta$ whenever $\beta\not\in{\mathcal{E}}_A$. So, for such $\beta$
not in $\deg_A(H^{\dim(S_A)}_{\partial_A}(S_A))$, the
indicated Ext- and Tor-groups vanish by Theorem \ref{thm-deRham-EK}.
\end{proof}
\begin{dfn}
Let $N_A$ be the \emph{interior ideal} of $S_A$, generated by the
monomials whose degrees are in the topological interior of $\mathbb{R}_+A$.
\hfill\ensuremath{\diamond}\end{dfn}
\begin{cor}\label{cor-deRham-GKZ}
If $S_A$ is normal, then
\begin{align}
\Tor_\bullet^{D_A}(D/\partial_AD_A,M_A^\beta)&=
\Tor_\bullet^{D_A}(D/\partial_AD_A,K_\bullet(S_A;E-\beta)) \notag \\
&= \left\{\begin{array}{cc}H_{dR}^{\bullet + \dim X_A}(T_A;\mathbbm{k})&\text{ if }-\beta\in \deg_A(N_A);
\\0&\text{else}.\end{array}\right. \notag
\end{align}
\end{cor}
\begin{proof}
The exceptional locus is here empty. The interior ideal is
the canonical module $\omega_{S_A}$ in the $A$-graded category by
\cite[Cor.~6.3.6]{BrunsHerzog} while also in the $A$-graded category
$\omega_{S_A}=\Ext^{\dim X_A-\dim T_A}_{R_A}(S_A,\omega_{R_A})$,
\cite[Prop.~3.6.12]{BrunsHerzog}. Then graded local duality
\cite[Thm.~3.6.19]{BrunsHerzog} yields that
$\deg_A(H^{\dim T_A}_{\partial_A}(S_A))=-\deg_A(N_A)$. Now use Theorem
\ref{thm-deRham-EK}.
\end{proof}
For our applications, it is interesting to know that $\mathbb{N} A$ does not
meet the strongly $A$-exceptional locus where for $\tau\subseteq A$ a
face we write $\partial_\tau$ for $\{\partial_j\}_{j\in\tau}$:
\begin{lem}\label{lem-deg-lc}
For any face $\tau$ of $A$, no element of $\mathbb{N} A$ is a degree of ${\mathrm{R}}
\Gamma_{\partial_\tau}(S_A)$.
\end{lem}
\begin{proof}
For $j\in\tau$, $\bigotimes_\tau(S_A\rightarrow S_A[\partial_j^{-1}])\simeq
(S_A\rightarrow S_A[\partial_j^{-1}])\otimes \bigotimes_{\tau\ni j'\not
=j}(S_A\rightarrow S_A[\partial_{j'}^{-1}])$. The corresponding double complex
spectral sequence starts on the $E^1$-page with modules of the form
$S_A[(\partial_j\cdot\prod_{j'\in\tau'}\partial_{j'})^{-1}]/S_A[(\prod_{j'\in\tau'}\partial_{j'})^{-1}]$
for all possible $\tau'\subseteq \tau\smallsetminus \{j\}$.
The $\mathbbm{k}$-dimension of $A$-graded localizations of $S_A$ in each $A$-degree
is zero or one, and $S_A$ is a domain,. So,
$S_A[(\partial_j\cdot\prod_{j'\in\tau'}\partial_{j'})^{-1}]/S_A[(\prod_{j'\in\tau'}\partial_{j'})^{-1}]$
is of dimension zero in each degree $\beta\in\mathbb{N} A$. Hence the same
holds for the abutment.
\end{proof}
In contrast, elements of $\mathbb{N} A$, including the origin $0$, can indeed
be \emph{quasi}-degrees of lower local cohomology (and hence exceptional
parameters):
\begin{exa}
Let
$A=\begin{pmatrix}2&1&0&1&0\\0&1&1&0&1\\0&0&0&1&1\end{pmatrix}$. The
exceptional locus is the line $\mathbbm{k}\cdot {\mathbf{a}}_1$.
\hfill\ensuremath{\diamond}\end{exa}
The following corollary will be used in Section \ref{sec-4term}
\begin{cor}
Suppose $S_A$ is Cohen--Macaulay,
and put $M=H_0(S_A\partial_A;E-\beta)$.
Then for $\beta=0$,
\[
\Tor_i^{D_A}(D_A/\partial_AD_A,M)=\left\{\begin{array}{ccl}
\mathbbm{k}^{\dim T_A}&\text{ if }&i=\dim X_A;\\
\mathbbm{k}&\text{ if }&i=\dim X_A -1;\\
0&\text{ if }&i<\dim X_A -1;\\\end{array}\right.
\]
while all Tor-groups vanish if $0\neq \beta\in\mathbb{N} A$.
\end{cor}
\begin{proof}
Consider the toric sequence $0\to S_A\partial_A\to S_A\to \mathbbm{k}\to
0$. Cohen--Macaulayness ensures, by \cite{MMW05}, that the
Euler--Koszul functor produces an exact sequence
\begin{gather}\label{eq-4term-toric}
0\to H_1(\mathbbm{k};E-\beta)\to H_0(S_A\partial_A;E-\beta)\to M_A^\beta\to
H_0(\mathbbm{k};E-\beta)\to 0.
\end{gather}
For $\beta\not =0$, the outer modules are zero. For $\beta=0$, the
right module is $O_A$ and the left is $O_A^{\dim T_A}$. The
claim then follows from Theorem \ref{thm-deRham-EK} and Lemma
\ref{lem-deg-lc}:
apply $\Tor_\bullet^{D_A}(D_A/\partial_AD_A,-)$ to $0\to M/O_A^{\dim T_A}\to
M_A^\beta\to O_A\to 0$ and
$0\to O_A^{\dim
T_A}\to M\to M/ O_A^{\dim T_A}\to 0$.
\end{proof}
\subsection{Ext and the polynomial solution functor}
Dualizing a $D_A$-resolution of $O_A$ gives a resolution of (a
cohomologically shifted) $D_A/\partial_AD_A$. Hence, up to shift by
$\varepsilon_A$ in the $A$-grading,
$\Ext^\bullet_{D_A}(O_A,M_A^\beta)=\Tor^{D_A}_{\dim X_A-\bullet}(D_A/\partial_AD_A,M_A^\beta)$.
In particular, the vanishing results in the previous section apply
to $\Ext^\bullet_{D_A}(O_A,M_A^\beta)$.
In this subsection we consider the behavior of the solution functor
$\Hom_{D_A}(-,O_A)$ with values in the ring of polynomials on the
class of $A$-hypergeometric modules $M_A^\beta$. It is immediately
clear that $\Hom_{D_A}(M_A^\beta,O_A)$ can only be nonzero if
$\beta\in\mathbb{N} A$, and it is an old result that $\beta\in\mathbb{N} A $ implies
that $\Hom_{D_A}(M_A^\beta,O_A)$ is 1-dimensional, see
\cite[Prop.~3.4.11]{SST}. We investigate here the derived polynomial
solution functor and prove
\begin{thm}\label{thm-ext-gkz-O}
If $\beta\not\in{\mathcal{E}}_A$ (for example, if $S_A$ is
Cohen--Macaulay) then
\[
\Ext^i_{D_A}(M_A^\beta,O_A)=\left\{\begin{array}{ccc}
H_{dR}^i(T_A;\mathbbm{k})&\text{ if }&\beta\in\mathbb{N} A;\\
0&&else.\end{array}\right.
\]
\end{thm}
(All experiments indicate this
to be true even if $\beta\in{\mathcal{E}}_A$.)
\begin{proof}
Write $\tau(-)$ for the transposition $x^{\mathbf{u}}\partial^{\mathbf{v}}\mapsto
\partial^{\mathbf{v}} (-x)^{\mathbf{u}}$ on $D_A$. Let $F_\bullet$ be an $A$-graded
$R_A$-free resolution of $S_A$ and denote
$\omega_{R_A}=D_A/\partial_AD_A[\dim X_A]$.
Then we have the following equalities, where $(-)^\vee$ is the vector
space dual:
\begin{eqnarray*}
\left({\mathrm{R}}\Hom_{D_A}(M_A^\beta,O_A))\right)^\vee&\stackrel{\text{(a)}}{\simeq}&
{\mathrm{R}}\Hom_{D_A}(O_A,\mathbb{D} M_A^\beta)\\
&\stackrel{\text{(b)}}{\simeq}&\omega_{R_A}\otimes^L_{D_A}\mathbb{D} M_A^\beta\\
&\stackrel{\text{(c)}}{=}&\omega_{R_A}\otimes^L_{D_A}\mathbb{D} K_\bullet(S_A;E-\beta)\\
&\stackrel{\text{(d)}}{=}&\omega_{R_A}\otimes^L_{D_A}\mathbb{D} K_\bullet(F_\bullet;
E-\beta)\\
&\stackrel{\text{(e)}}{=}&\omega_{R_A}\otimes^L_{D_A} K_\bullet(\Hom_{R_A}(F_\bullet,R_A);E+\beta+\varepsilon_A)\\
&\stackrel{\text{}}{\simeq}&\omega_{R_A}\otimes^L_{D_A} K_\bullet({\mathrm{R}}\Hom_{R_A}(S_A,R_A);E+\beta+\varepsilon_A)\\
&\stackrel{\text{(f)}}{\simeq}&K_\bullet(\omega_{R_A}\otimes^L_{R_A}{\mathrm{R}}\Hom_{R_A}(S_A,R_A);E+\beta+\varepsilon_A)\\
&\stackrel{\text{(g)}}{\simeq}&({\mathrm{R}}\Gamma_{\partial_A}{\mathrm{R}}\Hom_{R_A}(S_A,R_A))_{-\beta-\varepsilon_A}\otimes\bigwedge
(\mathbbm{k} A)\\
&\stackrel{\text{(h)}}{\simeq}&\left(({\mathrm{R}}\Hom_{R_A}({\mathrm{R}}\Hom_{R_A}(S_A,R_A),R_A))_{\beta}\otimes
\bigwedge (\mathbbm{k} A)\right)^\vee\\
&=&\left((S_A)_\beta\otimes
\bigwedge (\mathbbm{k} A)\right)^\vee.
\end{eqnarray*}
The following notes justify the above transformations:
\begin{enumerate}[label=$\bullet$(\alph*)]
\item Duality gives
${\mathrm{R}}\Hom_{D_A}(M,M')\simeq \left({\mathrm{R}}\Hom_{D_A}(\mathbb{D} M',\mathbb{D} M)\right)^\vee$, \cite[\S
2.6]{HottaTakeuchiTanasaki}.
\item Resolve $O_A$ and dualize the resolution, incurring a
cohomological shift.
\item By \cite{MMW05}, the hypergeometric
system is resolved by the Euler--Koszul complex as long as $\beta$
is not exceptional.
\item The Euler--Koszul functor can be applied to any $A$-graded
resolution.
\item $K_\bullet(F_\bullet;E-\beta)$ is a free complex.
Applying $\Hom_{D_A}(-,D_A)$ and the transposition $\tau$ turns
$D_A\otimes_{R_A}F_\bullet$ into
$D_A\otimes_{R_A}\Hom_{R_A}(F_\bullet,R_A)$ and the Euler--Koszul
complex on $\beta$ into that on $-\beta-\varepsilon_A$ since $x_j\partial_j$ turns
into $-\partial_jx_j$.
\item As in the proof of Theorem \ref{thm-deRham-EK}.
\item Theorem \ref{thm-deRham-EK} works for $A$-graded complexes just
as well.
\item Apply local $A$-graded duality (responsible for the dual).
\end{enumerate}
\end{proof}
\section{Three four-term sequences}\label{sec-4term}
\begin{notn}
From now on, $A$ is a $(d+1)\times (n+1)$ matrix
and $\mathbb{N} A$ is assumed to be saturated, in addition to the
conventions in Notation \ref{ntn-A} and Definition \ref{dfn-A}.
All products of $\mathbbm{k}$-schemes are by default over $\mathbbm{k}$.
\hfill\ensuremath{\diamond}\end{notn}
Consider the exact toric sequence $0\to S_A\partial_A\to S_A\to \mathbbm{k}\to
0$. Normality ensures, by \cite[Prop.~5.3, Thm.~6.6]{MMW05},
that the Euler--Koszul functor produces the exact sequence
\eqref{eq-4term-toric},
and, for $i\geq 1$, isomorphisms
\begin{gather}\label{eq-EK-kk}
H_i(\mathbbm{k};E-\beta) \simeq \begin{cases}O_A^{\binom{d+1}{i}} & \text{for}
\; \beta = 0; \\ 0 & \text{else.} \end{cases}
\end{gather}
In this section we will show that the sequence \eqref{eq-4term-toric}
has a geometric interpretation when $A$ is homogeneous. Our approach
is inspired by \cite{Stienstra98}, where Stienstra defined on the
torus $T_A$ a family $F$ of Laurent polynomials using the matrix
$A$. He showed that one term in the long exact cohomology sequence of
the pair ($T_A$, fiber of $F$) could be naturally identified with a
fiber in the $A$-hypergeometric system $M_A^0$ when $F$ is smooth. We
will extend this identification to the non-smooth fibers of $F$.
We will proceed as follows. First we identify the second term of
\eqref{eq-4term-toric} as a concatenation of (proper) direct image
functors applied to the structure sheaf $\mathcal{O}_{T_A}$. The third term
already has such an interpretation by
Corollary \ref{cor:SW} above. The remaining terms are
identified as the cohomology of the cone of a natural adjunction
morphism between the second and third term.
As a second step we show in Lemma \ref{lem:FLsidelift} that the
sequence \eqref{eq-4term-toric} is part of a long exact sequence
coming from a triangle of elementary $\mathcal{D}$-modules on the line
$\hat{\AA}^1$. We also show in Proposition \ref{prp-FL-qplus} that
the Fourier--Laplace transform of \eqref{eq-4term-toric} is induced by
the FL-transformed triangle of elementary $\mathcal{D}$-modules from Lemma
\ref{lem:FLsidelift}. This enables us to give a geometric
interpretation of the exact sequence in Theorem \ref{thm-GM-toric} in
terms of Gau\ss--Manin systems of the pair ($T_A$, fiber of $F$) as
alluded to above.
As a preparatory result we begin with an identification of two
functors on certain sheaves.
\subsection{Quasi-equivariant bundles}
Denote $\mathbb{G}_m$ the scheme of units of $\mathbbm{k}$. A
\emph{$\mathbb{G}_m$-action on the variety $Y$} is a multiplicative morphism
$\mu\colon \mathbb{G}_m\times Y\to Y$ where $1\in\mathbb{G}_m$ acts as
identity. That is, $\mu$ is a morphism, $\mu(g,\mu(g',y))=\mu(gg',y)$
and $\mu(1,y)=y$.
Let $X$ be an affine smooth variety and $\pi: E = X \times \AA^n \rightarrow
X$ be a trivial vector bundle on $X$. Write $E^\ast = E \setminus (X
\times \{0\})$ and let $E_x$ be the fiber over $x\in X$.
The zero section is identified with $X$ as closed subvariety via the
embedding
\[
i\colon X\hookrightarrow E.
\]
\begin{dfn}
A $\mathbb{G}_m$-action $\mu\colon\mathbb{G}_m\times E\to E$ on $E$ is \emph{fibered} if
\begin{asparaenum}
\item $\mu$ preserves fibers, $\mu\colon \mathbb{G}_m\times E_x\to E_x$;
\item $\mu$ is the restriction of a morphism $\mu\colon \AA^1\times E\to E$
under $\mathbb{G}_m\hookrightarrow \AA^1$;
\item $0\in\AA^1$ multiplies into the zero section, $\mu\colon
0\times E_x\to i(X)$
\item $\AA^1$ fixes the zero section, $\mu\colon \AA^1\times
i(X)\to i(X)$
\end{asparaenum}
\hfill\ensuremath{\diamond}\end{dfn}
\begin{dfn}
Let $\mu: \mathbb{G}_m \times E \rightarrow E$ be a fibered $\mathbb{G}_m$-action on $E$.
Write $\pr\colon \mathbb{G}_m\times E\to E$ for the projection and denote
by $\mu'$ and $\pr'$ the restrictions of $\mu$ and $\pr$ to $\mathbb{G}_m
\times E^*$.
A holonomic $\mathcal{D}_{E}$-module $\mathcal{M}$ is called
\emph{$\mathbb{G}_m$-quasi-equivariant} if $(\mu')^+ \mathcal{M}_{\mid E^*}
\simeq (\pr')^+ \mathcal{M}_{\mid E^*}$.
\hfill\ensuremath{\diamond}\end{dfn}
We consider the derived category of bounded complexes of
$\mathcal{D}_E$-modules with holonomic and quasi-equivariant cohomology.
\begin{lem}\label{lem:equivpoint}
Let $\pi: E \rightarrow X$ be fibered and denote $i: X \rightarrow E$ the inclusion
of the zero section. For every $\mathbb{G}_m$-quasi-equivariant $\mathcal{D}_E$-module
$\mathcal{M}$,
\begin{align}
\pi_+ \mathcal{M} \simeq i^\dag \mathcal{M} \qquad \text{and} \qquad \pi_\dag \mathcal{M} \simeq i^+ \mathcal{M} \notag
\end{align}
\end{lem}
\begin{proof}
By duality, it suffices to prove the first claim.
Denote by $j: E^\ast \rightarrow E$ the open embedding of the complement of
the zero section and let $\pi$ be the projection to the base
$X$.
We
have the exact triangles
\begin{gather}
j_\dag j^{-1} \mathcal{M} \longrightarrow \mathcal{M} \longrightarrow i_+ i^\dag \mathcal{M}
\overset{+1}{\longrightarrow}\\
\label{eqn-pi+j!1/j}
\pi_+ j_\dag j^{-1} \mathcal{M} \longrightarrow \pi_+\mathcal{M} \longrightarrow i^\dag \mathcal{M} \overset{+1}{\longrightarrow}
\end{gather}
and the Cartesian diagram
\[
\xymatrix{\mathbb{G}_m \times E^* \ar[r]^{j'} \ar[d]_{\mu'} & \AA^1_m \times E
\ar[d]_\mu \\ E^* \ar[r]^j & E}
\]
where $\mu'$ is the restriction of $\mu$ to $\mathbb{G}_m \times E^*$ and
$j'$ is the canonical inclusion. The morphism $s: E \rightarrow \AA^1
\times E$ with $s(x) = (1,x)$ is a section of
$\mu$. Thus, the composition (induced by the natural
transformation $\id_E\to\mu_+\mu^\dagger$)
\[
\pi_+ j_\dag j^{-1} \mathcal{M} \rightarrow \pi_+ \mu_+ \mu^\dag j_\dag j^{-1} \mathcal{M} \rightarrow \pi_+ \mu_+ s_+ s^\dag \mu^\dag j_\dag j^{-1} \mathcal{M} = \pi_+ j_\dag j^{-1} \mathcal{M}
\]
is an isomorphism by \eqref{eqn-pi+j!1/j}; hence it is enough to prove
$\pi_+ \mu_+ \mu^\dag j_\dag j^{-1} \mathcal{M} = 0$. By base change,
\[
\pi_+ \mu_+ \mu^\dag j_\dag j^{-1} \mathcal{M} \simeq \pi_+ \mu_+ j'_\dag (\mu')^\dag j^{-1} \mathcal{M}.
\]
Since $\mathcal{M}$ is $\mathbb{G}_m$-quasi-equivariant, we have
\[
(\mu')^\dag j^{-1}\mathcal{M} \simeq \pr^\dag j^{-1}\mathcal{M} \simeq
\mathcal{O}_{\mathbb{G}_m} \boxtimes j^{-1}\mathcal{M}\, .
\]
Therefore (letting $a\colon \AA^1 \rightarrow
\{pt\}$ be the map to the point) we get
\begin{align*}
\pi_+ \mu_+ j'_\dag(\mu')^\dag j^{-1} \mathcal{M} &\simeq \pi_+ \mu_+ j'_\dag(\mathcal{O}_{\mathbb{G}_m} \boxtimes j^{-1}\mathcal{M}) \\
&\simeq \pi_+ \mu_+ (j_{1\dag} \mathcal{O}_{\mathbb{G}_m} \boxtimes j_{\dag}j^{-1}\mathcal{M}) \\
&\simeq \pi_+ \pr_+(j_{1\dag} \mathcal{O}_{\mathbb{G}_m} \boxtimes j_{\dag}j^{-1}\mathcal{M}) \\
&\simeq a_+ j_{1\dag} \mathcal{O}_{\mathbb{G}_m} \boxtimes \pi_+ j_{\dag}j^{-1}\mathcal{M}
\end{align*}
where $j_1 :\mathbb{G}_m\rightarrow \AA^1$ is the canonical inclusion.
Since $a_+ j_{1\dag} \mathcal{O}_{\mathbb{G}_m} =
0$ in $\mathrm{D}^b(\mathcal{D}_{\{pt\}})$ we have $\pi_+ \mu_+\mu^\dag j_\dag j^{-1}
\mathcal{M} = 0$.
\end{proof}
Recall that $A$ is a $(d+1)\times(n+1)$ matrix.
Let $T_A = \Spec(\mathbbm{k}[t_0^\pm,\ldots,t_d^\pm])$ and consider
the ring homomorphism
\begin{align}
\mathbbm{k}[y_0,\ldots,y_n] &\longrightarrow \mathbbm{k}[t_0^{\pm},\ldots,t_d^{\pm}] \notag \\
y_i &\mapsto t^{{\mathbf{a}}_i} \notag
\end{align}
which gives rise to a morphism
\[
h_A: T_A \rightarrow
\hat{\AA}^{n+1},
\]
where $\hat{\AA}^{n+1} = \Spec(\mathbbm{k}[y_0,\ldots,y_n])$.
We factorize this embedding as
\[
T_A \overset{h_1}{\longrightarrow} \hat{\AA}^{n+1} \setminus \{0\} \overset{h_2}{\longrightarrow} \hat{\AA}^{n+1}\, .
\]
We are now ready to show a useful property of $A$-hypergeometric
systems and their Fourier--Laplace transforms.
\begin{lem}\label{lem:hotequiv}
The $\mathcal{D}_{\hat{\AA}^{n+1}}$-module $h_{A+} \mathcal{O}_{T_A}$ is
$\mathbb{G}_m$-quasi-equivariant.
\end{lem}
\begin{proof}
We view ${\hat A}^{n+1}$ as trivial bundle over itself.
Since $A$ is pointed, there is ${\mathbf{u}}\in\mathbb{Z}^{d+1}$ with
${\mathbf{v}}={\mathbf{u}}^T\cdot A$ componentwise positive.
Let $\mu'\colon \mathbb{G}_m\times
\hat{\AA}^{n+1}\to\hat{\AA}^{n+1}$ be the monomial action induced by
${\mathbf{v}}$ and let $\tilde\mu\colon \mathbb{G}_m\times T_A\to T_A$ be
the action induced by ${\mathbf{u}}$. (Compare the discussion on the Euler
space in \cite{RSW}.)
Consider the Cartesian diagram
\[
\xymatrix{ \mathbb{G}_m\times T_A \ar[r]^-{h_1\times id}
\ar@<0.5ex>[d]^{\tilde{\mu}} \ar@<-0.5ex>[d]_{\pr} &
\mathbb{G}_m\times \hat{\AA}^{n+1}\setminus \{0\} \ar@<0.5ex>[d]^{\mu'}
\ar@<-0.5ex>[d]_{\pr} & \\ T_A \ar[r]^-{h_1} &
\hat{\AA}^{n+1} \setminus \{0\} \ar[r]^-{h_2} & \hat{\AA}^{n+1}}
\]
and note that the positivity of ${\mathbf{v}}$ allows to extend $\mu'$ to
$\AA^1\times \hat{\AA}^{n+1}$.
Then
\[
\pr^+ h_{1+} \mathcal{O}_{T_A} \simeq (h_1 \times id)_+ \pr^+
\mathcal{O}_{T_A} \simeq (h_1 \times id)_+ \tilde{\mu}^+
\mathcal{O}_{T_A} \simeq (\mu')^+ h_{1+} \mathcal{O}_{T_A}
\]
and so $\pr^+ h_{1+}\mathcal{O}_{T_A} \simeq (\mu')^+
h_{1+}\mathcal{O}_{T_A}$.
\end{proof}
\subsection{The four-term sequence in terms of direct images}
Recall Notation \ref{ntn-FL-stuff} regarding Fourier--Laplace
transforms and consider the inverse Fourier--Laplace transformation of
the sequence \eqref{eq-4term-toric}:
\begin{gather}\label{eq-4term-toric-FL}
0 \longrightarrow H_1(\mathbbm{k};\hat{E}+\beta) \longrightarrow H_0(\hat{S}_A \cdot
y_A;\hat{E}+\beta) \longrightarrow \hat{M}^\beta_A \longrightarrow H_0(\mathbbm{k};\hat{E}+\beta)
\longrightarrow 0.
\end{gather}
\begin{dfn}
Let $\mathcal{B}_{0}$ be the unique simple
$\mathcal{D}_{\hat{\AA}^{n+1}}$-module supported in
$0\in\hat{\AA}^{n+1}$.
\hfill\ensuremath{\diamond}\end{dfn}
\begin{prp}\label{prop:FL4termexseq}
For $\beta=0$ there is an isomorphism of exact 4-term sequences
\[
\xymatrix{0 \ar[r] & \mathcal{H}_1(\mathbbm{k};\hat{E}) \ar[r] & \mathcal{H}_0(
\hat{S}_A \cdot y_A;\hat E) \ar[r] & \hat{\mathcal{M}}^0_A \ar[r] &
\mathcal{H}_0(\mathbbm{k};\hat{E}) \ar[r] & 0 \\ 0 \ar[r]& {\mathcal{B}_0}^{d+1}
\ar[u]^\simeq \ar[r] & \mathcal{H}^0(h_{2\dag} h_{1+}
\mathcal{O}_{T_A}) \ar[u]^\simeq \ar[r] &
\mathcal{H}^0(h_{+}\mathcal{O}_{T_A}) \ar[u]^\simeq \ar[r] & \mathcal{B}_0
\ar[r] \ar[u]^\simeq & 0}
\]
\end{prp}
\begin{proof}
By Corollary \ref{cor:SW}, $h_{A+} \mathcal{O}_{T_A}
\simeq \hat{\mathcal{M}}^0_A$. By \eqref{eq-EK-kk},
$\mathcal{H}_i(\mathbbm{k};\hat E)\simeq
{\mathcal{B}_0}^{{d+1\choose i}}$.
Restricting to $\hat{\AA}^{n+1} \setminus
\{0\}$ we see that
\[
\mathcal{H}_0(\hat{S}_A \cdot y_A;\hat E)_{\mid \hat{\AA}^{n+1} \setminus
\{0\}} \simeq (\hat{\mathcal{M}}^0_A)_{\mid \hat{\AA}^{n+1} \setminus
\{0\}} \simeq h_{1+} \mathcal{O}_{T_A} \qquad \text{in}\quad
{\mathrm{Mod}}_h(\mathcal{D}_{\hat{\AA}^{n+1}})\, .
\]
Since $\mathcal{H}_{>0}(S_A \cdot y_A;\hat E)_{\mid \hat{\AA}^{n+1}
\setminus \{0\}} = 0$, we have
\[
\mathcal{H}_0(S_A \cdot y_A;\hat E)_{\mid \hat{\AA}^{n+1} \setminus \{0\}}
\simeq \mathcal{K}_\bullet(S_A \cdot y_A;\hat E)_{\mid \hat{\AA}^{n+1}
\setminus \{0\}} \qquad \text{in} \quad
\mathrm{D}^b_h(\mathcal{D}_{\hat{\AA}^{n+1}})\, .
\]
By adjunction this gives a morphism
\[
h_{2 \dag}h_{1+} \mathcal{O}_{T_A} \overset{\simeq}\longrightarrow h_{2 \dag}
h_{2}^{-1}\mathcal{K}_\bullet(\hat{S}_A \cdot y_A;\hat E) \longrightarrow
\mathcal{K}_\bullet(\hat{S}_A \cdot y_A;\hat E)
\]
and so induces a morphism
$\mathcal{H}^0 (h_{2 \dag}h_{1+} \mathcal{O}_{T_A}) \longrightarrow
\mathcal{H}_0(\hat{S}_A \cdot y_A;\hat E)$
such that the center and right squares in our diagram commute. We
need to prove that the morphism
\begin{equation}\label{eq:adjEuKos}
h_{2 \dag} h_2^{-1} \mathcal{K}_\bullet(\hat{S}_A \cdot
y_A;\hat E) \longrightarrow \mathcal{K}_\bullet(\hat{S}_A \cdot y_A;\hat E)
\end{equation}
is an isomorphism. (While we know that $\mathcal{H}_1(\mathbbm{k};\hat E)$ and
${\mathcal{B}_0}^{d+1}$ are isomorphic, it is not yet clear that
$\mathcal{H}^0 (h_{2\dag} h_{1+}
\mathcal{O}_{T_A}) \rightarrow \mathcal{H}_0( \hat{S}_A \cdot
y_A;\hat E)$ induces such isomorphism.)
In order to prove that the morphism \eqref{eq:adjEuKos} is an
isomorphism we have to show that the third term in the adjunction
triangle
\begin{equation}\label{eq:adjEuKosTriangle}
h_{2 \dag} h_2^{-1} \mathcal{K}_\bullet(\hat{S}_A \cdot y_A;\hat E) \longrightarrow
\mathcal{K}_\bullet(\hat{S}_A \cdot y_A;\hat E) \longrightarrow i_+ i^{\dag}
\mathcal{K}_\bullet(\hat{S}_A \cdot y_A;\hat E) \overset{+1}\longrightarrow
\end{equation}
vanishes. By Kashiwara equivalence it is enough to show that $i^{\dag}
\mathcal{K}_\bullet(\hat{S}_A \cdot y_A;\hat E)$ is isomorphic to zero.
Since $\mathcal{K}_\bullet(\widehat{S}_A \cdot y_A;\hat E)_{\mid
\hat{\AA}^{n+1} \setminus \{0\}} \simeq (h_{A+} \mathcal{O}_T)_{\mid
\hat{\AA}^{n+1} \setminus \{0\}}$, the complex $\mathcal{K}_\bullet(\widehat{S}_A
\cdot y_A;\hat E)$ is $\mathbb{G}_m$-quasi-equivariant. By Lemma
\ref{lem:equivpoint}, $i^\dag \mathcal{K}_\bullet(\widehat{S}_A \cdot
y_A;\hat E) \simeq a_+ \mathcal{K}_\bullet(\widehat{S}_A \cdot y_A;\hat E)$
where $a$ is the map to a point. Now $a_+ \mathcal{K}_\bullet(\widehat{S}_A
\cdot y_A;\hat E)$ is dual to $\rho\mathcal{K}_\bullet(S_A\partial_A;E)$ which
allows us to use Lemma \ref{lem-restr-EK} to conclude.
\end{proof}
\subsection{The four-term sequence with Gau\ss--Manin systems}
\begin{notn}
From now on, in addition to the assumptions in Notation \ref{ntn-A}
and Definition \ref{dfn-A} as
well as normality, we assume that the matrix $A$ is
homogeneous, \emph{i.e.} that $(1,\ldots,1)$ is in the row span of $A$.
Furthermore, for the remainder of the paper, $\beta=0$.
\hfill\ensuremath{\diamond}\end{notn}
One may put $A$ into the following shape by elementary row
operations
\begin{equation}\label{def:Atilde}
A = ({\mathbf{a}}_1, \ldots, {\mathbf{a}}_n) =\left(\begin{matrix}
1 & 1 & \ldots & 1 \\
0 & a_{11} & \dots & a_{1n} \\
\vdots & \vdots & &\vdots \\
0 & a_{d1} & \dots &a_{dn}\end{matrix} \right) =
\left(\begin{matrix}
1 & 1 & \ldots & 1\\
0 & & & \\
\vdots & & B & \\
0 & & &\end{matrix} \right)
\end{equation}
where $B= ({\mathbf{b}}_1,\ldots,{\mathbf{b}}_n)$ is the $d \times
n$-matrix with entries $(a_{ij})_{1\leq i \leq d, 1 \leq j\leq n}$.
Using this homogeneity assumption, we will here give a geometric
interpretation to our 4-term sequence \eqref{eq-4term-toric}. For
this we will need a variant of a comparison theorem of d'Agnolo and
Eastwood \cite{AE}, between the Radon and Fourier--Laplace transform,
and several other preparatory statements.
Set $T_B := \Spec(\mathbbm{k}[t_1^\pm,\ldots,t_d^\pm])$ and $\hat{\AA}^1 :=
\Spec(\mathbbm{k}[t_0])$. We will identify $T_A$ with $T_B \times (\hat{\AA}^1
\setminus \{0\})$. From the ring homomorphism
\begin{align}
\mathbbm{k}[y_0,\ldots,y_n] &\longrightarrow \mathbbm{k}[t_0,t_1^\pm,\ldots,t_d^\pm]\notag
\\ (y_0,\ldots,y_n) &\mapsto (t_0,t_0 t^{{\mathbf{b}}_1},\ldots,t_0
t^{{\mathbf{b}}_n}) \notag
\end{align}
we get a map
\begin{gather}\label{eq-map-k}
k\colon T_B \times \hat{\AA}^1 \longrightarrow \hat{\AA}^{n+1}
\end{gather}
whose restriction to $T_A$ is just our
old morphism $h_A$. Let $\tilde{k}$ be the closed embedding
\[
\tilde{k}:= (id_{T_B} \times k) :T_B
\times \hat{\AA}^1 \rightarrow T_B \times \hat{\AA}^{n+1},
\]
let $j,i$ be the embedding and inclusion
\[
j\colon T_A=T_B \times (\hat{\AA}^1
\setminus \{0\})\rightarrow T_B \times \hat{\AA}^1,\qquad
i\colon T_B \times \{0\} \rightarrow T_B \times \hat{\AA}^1.
\]
Then there is a commutative diagram
\[
\xymatrix{\{0\} \ar[d]_{i_0} & \ar[l]_a T_B \ar[d]_i \ar[dr]^{k_0} & \\
\hat{\AA}^1 & T_B \times \hat{\AA}^1 \ar[r]^k \ar[l]_{p} & \hat{\AA}^{n+1} \\
\hat{\AA}^1 \setminus\{0\} \ar[u]^{j_0}& T_A \ar[l]_{~p'} \ar[ur]^{h_A} \ar[u]^j \ar[r]^{h_1~} & \hat\AA^{n+1}\setminus \{0\} \ar[u]_{h_2}}
\]
where $p\colon T_B \times \hat{\AA}^1 \rightarrow \hat{\AA}^1$ is the projection and
$k_0$ sends $T_B$ to the origin. Define the following $\mathcal{D}$-modules on
$\hat{\AA}^1$:
\[
\mathcal{D}_{\hat{\AA}^1} \bullet 1 := \mathcal{D}_{\hat{\AA}^1}/(\partial_t), \qquad
\mathcal{D}_{\hat{\AA}^1} \bullet \mathfrak{H} := \mathcal{D}_{\hat{\AA}^1}/(t \partial_t),
\]
\[
\mathcal{D}_{\hat{\AA}^1} \bullet 1/t := \mathcal{D}_{\hat{\AA}^1}/(\partial_t t),
\qquad \mathcal{D}_{\hat{\AA}^1} \bullet \delta := \mathcal{D}_{\hat{\AA}^1} / (t).
\]
(The module $\mathcal{D}_{\hat{\AA}^1} \bullet \mathfrak{H}$ encodes the
Heaviside distribution).
\begin{lem}\label{lem:FLsidelift
We have the following isomorphisms:
\begin{align}
k_+ \mathcal{O}_{T_B \times \hat{\AA}^1} &\simeq k_+ p^+(\mathcal{D}_{\hat{\AA}^1}
\bullet 1), & h_{A+} \mathcal{O}_{T_A} &\simeq
k_+p^+(\mathcal{D}_{\hat{\AA}^1} \bullet 1/t), \notag \\ k_{0+}\mathcal{O}_{T_B} &\simeq
k_+ p^+ (\mathcal{D}_{\hat{\AA}^1} \bullet \delta), & h_{2 \dag} h_{1+}
\mathcal{O}_{T_A} &\simeq k_+ p^+(\mathcal{D}_{\hat{\AA}^1} \bullet \mathfrak{H}).
\notag
\end{align}
The adjunction morphism $h_{2 \dag} h_{1+} \mathcal{O}_{T_A} \rightarrow
h_{A+} \mathcal{O}_{T_A}$ is induced by the adjunction morphism
$\mathcal{D}_{\hat{\AA}^1}\bullet \mathfrak{H} \rightarrow \mathcal{D}_{\hat{\AA}^1}\bullet 1/t$.
\end{lem}
\begin{proof}
The first three isomorphisms follow from
\begin{align}
k_+ \mathcal{O}_{T_B \times \hat{\AA}^{1}} &\simeq k_+ p^+
\mathcal{O}_{\hat{\AA}^1} = k_+ p^+ (\mathcal{D}_{\hat{\AA}^1} \bullet 1)\notag,
\\ h_{A+} \mathcal{O}_{T_A} &\simeq k_+ j_+ \mathcal{O}_{T_A}
\simeq k_+ j_+ {p'}^+ \mathcal{O}_{\hat{\AA}^1 \setminus \{0\}} \simeq k_+ p^+
j_{0+}\mathcal{O}_{\hat{\AA}^1 \setminus \{0\}} \simeq k_+ p^+
(\mathcal{D}_{\hat{\AA}^1} \bullet 1/t), \notag \\ k_{0+} \mathcal{O}_{T_B} &\simeq k_+
i_+ \mathcal{O}_{T_B} \simeq k_+ i_+ a^+ \mathcal{O}_{\{0\}} \simeq k_+ p^+ i_0
\mathcal{O}_{\{0\}} \simeq k_+ p^+ (\mathcal{D}_{\hat{\AA}^1} \bullet \delta).
\notag
\end{align}
For the last one we have
\[
k_+ p^+ (\mathcal{D}_{\hat{\AA}^1} \bullet \mathfrak{H}) \simeq k_+ p^+ j_{0\dag}
\mathcal{O}_{\hat{\AA}^1 \setminus \{0\}} \simeq k_+ j_\dag {p'}^+
\mathcal{O}_{\hat{\AA}^1 \setminus \{0\}} \simeq k_+ j_\dag
\mathcal{O}_{T_A}.
\]
So it remains to prove that $k_+ j_\dag \mathcal{O}_{T_A} \simeq
h_{2\dag}h_{1+} \mathcal{O}_{T_A}$. For this consider the diagram
with Cartesian squares
\[
\xymatrix{T_B \ar[d]^i \ar[r]^{p_0} & \{0\} \ar[d]_{i_0} \\ T_B \times
\AA^1 \ar@/^/[u]^p\ar[r]^k & \AA^{n+1}
\ar@/_/[u]_\pi\\ T_A \ar[u]_j \ar[r]^{h_1} &
\AA^{n+1}\setminus \{0\} \ar[u]_{h_2}}
\]
Base change in the lower square gives
$h_{1+} \mathcal{O}_{T_A} \stackrel{\simeq}{\rightarrow}\, h_{1+} j^{-1}j_\dag
\mathcal{O}_{T_A} \stackrel{\simeq}{\rightarrow} h^{-1}_2 k_+
j_{\dag}\mathcal{O}_{T_A}$. Adjunction yields a morphism $h_{2
\dag} h_{1+} \mathcal{O}_{T_A} \rightarrow k_+ j_{\dag}
\mathcal{O}_{T_A}$. In order to prove that this is an isomorphism,
it is hence enough to show that $h_{2\dag}h_{1+}\mathcal{O}_{
T_A}\simeq h_{2\dag}h_2^{-1} k_+ j_\dag
\mathcal{O}_{T_A} \rightarrow k_+ j_\dag \mathcal{O}_{T_A}$ is an
isomorphism. Using the triangle
\[
h_{2 \dag}h_2^{-1} k_+ j_\dag \mathcal{O}_{T_A} \longrightarrow k_+ j_\dag
\mathcal{O}_{T_A} \longrightarrow i_{0+}i_0^\dag k_+ j_\dag
\mathcal{O}_{T_A} \overset{+1}\longrightarrow\,,
\]
it remains to show that $i_{0+}i_0^\dag k_+ j_\dag \mathcal{O}_{T_A}$ is
zero. For this we observe that
\[
h_2^+ k_+ j_\dag \mathcal{O}_{T_A} \simeq h_{1+} j^+ j_\dag
\mathcal{O}_{T_B} \simeq h_{1+} \mathcal{O}_{T_B}
\]
is the restriction of a quasi-equivariant module.
This shows, via Lemma \ref{lem:hotequiv}, that $k_+ j_\dag
\mathcal{O}_{T_A}$ is $\mathbb{G}_m$-quasi-equivariant. We therefore have
\[
i_{0+}i_0^\dag k_+ j_\dag \mathcal{O}_{T_A} \simeq i_{0+} \pi_+
k_+ j_\dag \mathcal{O}_{T_A} \simeq i_{0+}p_{0+} p_+j_\dag
\mathcal{O}_{T_A} \simeq i_{0+}p_{0+} i^{\dag} j_\dag \mathcal{O}_{T_A}\,,
\]
using Lemma \ref{lem:equivpoint} to substitute $p_+$ by $i^!$.
Since $i^\dag j_\dag \mathcal{O}_{T_A}$ is zero, the claim
follows.
\end{proof}
Consider the diagram
\[
\xymatrix{\AA^1 & T_B \times \AA^{n+1} \ar[l]_-F \ar[r]^-q & \AA^{n+1}}
\]
where $q$ is the projection and
\[
F(t_1,\ldots,t_n,\lambda_0,\ldots,\lambda_n)
= \lambda _0 + \sum_{i=1}^n \lambda_i t^{{\mathbf{b}}_i}.
\]
Denote $\Gamma=\Var(F)$ and write
\[
i_\Gamma\colon \Gamma \subset T_B \times \AA^{n+1},\quad j_U\colon U
\rightarrow T_B \times \AA^{n+1}
\]
for the inclusion of $\Gamma$ and its complement $U$. The
Gau\ss--Manin system $q_{U+}\mathcal{O}_U$ is of interest since it carries a
mixed Hodge structure by Saito's work in \cite{Saito-MHM-Kyoto90}.
Our article gives evidence to our belief that many $D$-modules
arising from Euler--Koszul complexes also carry such structure, and
that they relate to interesting geometric information.
\begin{prp}\label{prp-FL-qplus}
With $u = 1,\delta, 1/t, \mathfrak{H}$ and $\hat{u} = \delta,1,\mathfrak{H},1/t$,
and with $k$ as in \eqref{eq-map-k} we have the following isomorphisms
\[
\FL(k_+ p^+ (\mathcal{D}_{\hat{\AA^1}} \bullet \hat{u})) \simeq q_+ F^+
(\mathcal{D}_{\AA^1} \bullet u).
\]
\end{prp}
\begin{proof}
Consider the diagram
\[
\xymatrix@C=4pc@R=3pc{ \hat{\AA}^1 & T_B \times \hat{\AA}^1 \ar[l]_{p}
\ar[r]^{k} & \hat{\AA}^{n+1} \\ \hat{\AA}^1 \times \AA^1
\ar[u]_{p_1} \ar[d]_{p_2} & T_B \times \hat{\AA}^1 \times \AA^{n+1}
\ar[u]_{p_{12}} \ar[l]_{id_{\hat{\AA}^1} \times F} \ar[d]^{p_{13}}
\ar[r]^{k\times id_{\AA^{n+1}}} & \hat{\AA}^{n+1} \times \AA^{n+1}
\ar[u]_{q_{1}} \ar[d]^{q_{2}} \\ \AA^1 & T_B \times \AA^{n+1} \ar[l]_F
\ar[r]^q& \AA^{n+1}}
\]
where $p_{ij}$ are the projections to the factors $i$ and $j$. Recall
the Fourier--Laplace sheaf $\mathcal{L}$ on $\AA^{n+1}$ from Definition
\ref{dfn-FL} and denote $\mathcal{L}_1$ the Fourier--Laplace sheaf on
$\AA^1\times\AA^1$. Then
\begin{align*}
\FL(k_+p^+ (\mathcal{D}_{\hat{\AA}^1}\bullet \cdot \hat{u})) &= q_{2+}\left((q_1^+
k_+ p^+(\mathcal{D}_{\hat{\AA}^1} \bullet \hat{u})) \otimes
\mathcal{L}\right)\\
&\simeq q_{2+}((k\times id)_+\; p_{12}^+\; p^+
(\mathcal{D}_{\hat{\AA}^1} \bullet \hat{u}) \otimes \mathcal{L}) \\
&\simeq q_{2+}(k
\times id)_+(p_{12}^+ p^+(\mathcal{D}_{\hat{\AA}^1} \bullet \hat{u}) \otimes
(k \times id)^+\mathcal{L})\\
&\simeq q_+ p_{13+}((id \times F)^+ p_1^+
(\mathcal{D}_{\hat{\AA}^1} \bullet \hat{u}) \otimes (k \times
id)^+\mathcal{L})\\
&\simeq q_+ p_{13+}((id \times F)^+ p_1^+
(\mathcal{D}_{\hat{\AA}^1} \bullet \hat{u}) \otimes (id \times
F)^+\mathcal{L}_1)\\
&\simeq q_+ p_{13+}(id \times F)^+( p_1^+
(\mathcal{D}_{\hat{\AA}^1} \bullet \hat{u}) \otimes \mathcal{L}_1) \\
&\simeq q_+
F^+ p_{2+} p_1^+ (\mathcal{D}_{\hat{\AA}^1} \bullet \hat{u}) \otimes
\mathcal{L}_1)\\
&\simeq q_+ F^+ (\mathcal{D}_{\AA^1} \bullet u).
\end{align*}
\end{proof}
Now consider the diagram
\[
\xymatrix{\{0\} \ar[d]_{i_0} & \Gamma \ar[l]_a \ar[d]^{i_\Gamma}
\ar[dr]^{q_\Gamma} & \\ \AA^1 & T_B \times \AA^{n+1} \ar[l]_F \ar[r]^q
& \AA^{n+1} \\ \AA^1 \setminus \{0\} \ar[u]^{j_0} & U \ar[u]_{j_U}
\ar[l]_<<<<<<{F|_U} \ar[ur]_{q_U} &}
\]
where $q$ is the projection. We have, writing $F$ for $F_U$,
\begin{align*}
q_+ F^+ (\mathcal{D}_{\AA^1} \bullet 1/t) &\simeq q_+ F^+ j_{0+} \mathcal{O}_{\AA^1
\setminus \{0\}} \simeq q_+ j_{U +} F^+ \mathcal{O}_{\AA^1 \setminus
\{0\}} \simeq q_{U +} \mathcal{O}_U, \\
q_+ F^+ (\mathcal{D}_{\AA^1} \bullet
\mathfrak{H}) &\simeq q_+ F^+ j_{0\dag} \mathcal{O}_{\AA^1 \setminus \{0\}} \simeq
q_+ j_{U \dag} F^+ \mathcal{O}_{\AA^1 \setminus \{0\}} \simeq q_+ j_{U \dag}
\mathcal{O}_U, \\
q_+F^+ (\mathcal{D}_{\AA^1}\bullet \delta) &\simeq q_+ F^+ i_{0+}
\mathcal{O}_{\{0\}}\simeq q_+ i_{\Gamma +}a^+ \mathcal{O}_{\{0\}} \simeq q_+
i_{\Gamma +} \mathcal{O}_\Gamma \simeq q_{\Gamma+} \mathcal{O}_\Gamma, \\
q_+ F^+(\mathcal{D}_{\AA^1} \bullet 1) &\simeq q_+ \mathcal{O}_{T_B \times \AA^{n+1}},
\end{align*}
where the second isomorphism in the second line follows from the
smoothness of $F$.
\begin{ntn}
If $W$ is a $\mathbbm{k}$-space (for example, $H^i_{dR}(T_B;\mathbbm{k})$) then
$\underline W$ denotes the trivial vector bundle $W \otimes_\mathbbm{k}
\mathcal{O}_{\AA^{n+1}}$.
\hfill\ensuremath{\diamond}\end{ntn}
Consider the following exact sequence of $\mathcal{D}_{\hat{\AA}^1}$-modules
\begin{align}\label{eq-I}
0 \longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet \delta \longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet
\mathfrak{H} \longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet 1/t \longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet
\delta \longrightarrow 0
\end{align}
\begin{thm}\label{thm-GM-toric}
The exact sequence \eqref{eq-I} induces an isomorphism of
exact sequences
\[
\xymatrix{0\ar[r] &
\mathcal{H}_1(\mathbbm{k};E)\ar[r] &
\mathcal{H}_0(S_A\partial_A;E)\ar[r] &
\mathcal{M}_A^0\ar[r] &
\mathcal{H}_0(\mathbbm{k};E)\ar[r] & 0 \\
0 \ar[r]&
{\begin{array}{c}\underline{H^d_{dR}(T_B;\mathbbm{k})}\\
\oplus\\
\underline{H^{d-1}_{dR}(T_B;\mathbbm{k})}\end{array}}\ar[r] \ar[u]_\simeq &
\mathcal{H}^0 (q_{U+} \mathcal{O}_U) \ar[r] \ar[u]_\simeq &
\mathcal{H}^0 (q_+ j_{U \dag} \mathcal{O}_U) \ar[r] \ar[u]_\simeq &
\underline{H^d_{dR}(T_B;\mathbbm{k})} \ar[u]_\simeq \ar[r] & 0}
\]
\end{thm}
\begin{proof}
The sequence \eqref{eq-I} is part of the long exact sequence coming from
the triangle
\[
j_{0 \dag}j_0^{-1} \mathcal{O}_{\hat{\AA}^1} \longrightarrow j_{0 +} j_0^{-1} \mathcal{O}_{\hat{\AA}^1} \longrightarrow i_{0+} i_0^\dag j_{0 +} j_0^{-1} \mathcal{O}_{\hat{\AA}^1} \overset{+1}\longrightarrow
\]
which is isomorphic to
\[
\mathcal{D}_{\hat{\AA}^1} \bullet \mathfrak{H} \longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet 1/t \longrightarrow
(\mathcal{D}_{\hat{\AA}^1} \bullet \delta) \oplus (\mathcal{D}_{\hat{\AA}^1} \bullet
\delta[1]) \overset{+1}\longrightarrow
\]
Applying the concatenated functor $\textup{FL} \circ k_+ p^+$ to the
triangle above and using Lemma \ref{lem:FLsidelift}, Proposition
\ref{prop:FL4termexseq}, and the fact that $\mathcal{H}^i(k_{0+}\mathcal{O}_{T_B})=
{\mathcal{B}_0}^{\binom{d}{i}}$ we obtain the upper sequence in the
theorem. (Recall that $k_0$ sends $T_B$ to the origin in
$\hat{\AA}^{n+1}$). Applying $q_+ F^+ \circ \FL$ instead gives the lower
sequence.
\end{proof}
If one applies $q_+ F^+ \circ \FL$ to the exact sequence
\begin{align}\label{eq-II}
0 \longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet 1 \longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet 1/t
\longrightarrow \mathcal{D}_{\hat{\AA}^1} \bullet \delta \longrightarrow 0
\end{align}
one obtains as a part of the resulting long exact sequence the piece
\begin{equation}\label{eq:seqII}
0 \longrightarrow \underline{H^{d-1}_{dR}(T_B;\mathbbm{k})} \longrightarrow \mathcal{H}^0(q_{\Gamma+}
\mathcal{O}_\Gamma) \longrightarrow \mathcal{H}^0(q_+ j_{U \dag} \mathcal{O}_U) \longrightarrow
\underline{H^d_{dR}(T_B;\mathbbm{k})} \longrightarrow 0,
\end{equation}
We now determine how this sequence relates to the two sequences
in Theorem \ref{thm-GM-toric}.
\begin{prp}
The exact sequence \eqref{eq:seqII}
is the quotient of the exact sequence
\[
0 \longrightarrow \underline{H^{d}_{dR}(T_B;\mathbbm{k})} \oplus
\underline{H^{d-1}_{dR}(T_B;\mathbbm{k})} \longrightarrow \mathcal{H}^0(q_{U+} \mathcal{O}_U) \longrightarrow
\mathcal{H}^0(q_+ j_{U \dag} \mathcal{O}_U) \longrightarrow \underline{H^d_{dR}(T_B;\mathbbm{k})} \longrightarrow 0
\]
by the exact sequence
\[
0 \longrightarrow \underline{H^{d}_{dR}(T_B;\mathbbm{k})} \longrightarrow \underline{H^{d}_{dR}(T_B;\mathbbm{k})} \longrightarrow 0 \longrightarrow 0 \longrightarrow 0.
\]
\end{prp}
\begin{proof}
Consider the Fourier--Laplace transforms of the sequences \eqref{eq-I}
and \eqref{eq:seqII}. We get a commutative diagram with exact rows and
columns:
\[
\xymatrix{ & 0 \ar[r] & \mathcal{D}_{\AA^1} \bullet 1 \ar[r] \ar[d] & \mathcal{D}_{\AA^1} \bullet 1 \ar@{^{(}->}[d] \ar[r] & 0 \ar[r] \ar[d] & 0 \ar[r] \ar[d] & 0 \\
& 0 \ar[r] & \mathcal{D}_{\AA^1} \bullet 1 \ar[r] \ar[d] & \mathcal{D}_{\AA^1} \bullet 1/t \ar[r] \ar@{->>}[d] & \mathcal{D}_{\AA^1} \bullet \mathfrak{H} \ar[r] \ar[d] & \mathcal{D}_{\AA^1} \bullet 1 \ar[r] \ar[d] & 0 \\
& 0 \ar[r] & 0 \ar[r] & \mathcal{D}_{\AA^1} \bullet \delta \ar[r] & \mathcal{D}_{\AA^1} \bullet \mathfrak{H} \ar[r] & \mathcal{D}_{\AA^1} \bullet 1 \ar[r] & 0}
\]
and morphisms of triangles
\[
\xymatrix{ \mathcal{D}_{\AA^1} \bullet 1 \ar[r] \ar[d] & 0 \ar[r] \ar[d] & (\mathcal{D}_{\AA^1} \bullet 1) [1] \ar[r]^-{+1} \ar[d] & \\
\mathcal{D}_{\AA^1} \bullet 1/t \ar[r] \ar[d] & \mathcal{D}_{\AA^1} \bullet \mathfrak{H} \ar[r] \ar[d] & (\mathcal{D}_{\AA^1}\bullet 1) \oplus (\mathcal{D}_{\AA^1} \bullet 1) [1] \ar[d] \ar[r]^-{+1} & \\
\mathcal{D}_{\AA^1} \bullet \delta \ar[r] & \mathcal{D}_{\AA^1} \bullet \mathfrak{H} \ar[r] & \mathcal{D}_{\AA^1} \bullet 1 \ar[r]^-{+1}&}
\]
From this, we get an
exact sequence of exact rows
\[
\xymatrix{0 \ar[r] & \underline{H^{d}_{dR}(T_B;\mathbbm{k})} \ar[r] \ar@{^{(}->}[d]& \underline{H^{d}_{dR}(T_B;\mathbbm{k})} \ar[r] \ar@{^{(}->}[d] & 0 \ar[r] \ar[d] & 0 \ar[r] \ar[d] & 0 \\
0 \ar[r]& \underline{H^{d}_{dR}(T_B;\mathbbm{k})} \oplus \underline{H^{d-1}_{dR}(T_B;\mathbbm{k})} \ar[r] \ar@{->>}[d] & \mathcal{H}^0(q_{U+} \mathcal{O}_U) \ar[r] \ar@{->>}[d] & \mathcal{H}^0(q_+ j_{U \dag} \mathcal{O}_U) \ar[r] \ar[d]& \underline{H^d_{dR}(T_B;\mathbbm{k})} \ar[r] \ar[d]& 0 \\
0 \ar[r] & \underline{H^{d-1}_{dR}(T_B;\mathbbm{k})} \ar[r] & \mathcal{H}^0(q_{\Gamma+} \mathcal{O}_\Gamma) \ar[r] & \mathcal{H}^0(q_+ j_{U \dag} \mathcal{O}_U) \ar[r] & \underline{H^d_{dR}(T_B;\mathbbm{k})} \ar[r] & 0}
\]
The lower middle maps are surjective since $F^+$ is
exact and $q_+$ is right exact.
\end{proof}
We now introduce a family of Laurent polynomials defined on $T_B \times
\AA^n$ using the columns of the matrix $B$. For this, recall
Definition
\ref{def:Atilde} and consider the ring
homomorphism
\begin{align}
\mathbbm{k}[\lambda_0, \ldots, \lambda_n] &\longrightarrow \mathbbm{k}[t_1^\pm, \ldots ,t_d^\pm]
\otimes_\mathbbm{k} \mathbbm{k}[\lambda_1, \ldots, \lambda_n] \notag \\ \lambda_i
&\mapsto \begin{cases} -\sum_{i=1}^n t^{{\mathbf{b}}_i}\otimes \lambda_i &
\text{for}\; i= 0; \\ \lambda_i & \text{for}\;
i=1,\ldots,n,\end{cases}
\end{align}
which induces a family of Laurent polynomials
\begin{align}\label{eq:FamLaurent}
\varphi_B\colon T_B \times \AA^n &\longrightarrow \AA^{n+1} = \AA^{1} \times \AA^{n} \, .
\end{align}
and an isomorphism
\begin{align}
i_\varphi: T_B \times \AA^n &\longrightarrow \Gamma \subseteq T_B \times \AA \times
\AA^n \notag
\end{align}
onto the graph $\Gamma$. Hence $\varphi_B = q_\Gamma \circ
i_\varphi$ and therefore $\mathcal{H}^0(\varphi_{B+} \mathcal{O}_{T_B \times
\AA^n}) \simeq \mathcal{H}^0(q_{\Gamma +} \mathcal{O}_\Gamma)$.
This recovers a special case of a theorem of \cite{Reich2},
\emph{i.e.} there is an exact sequence
\begin{gather}\label{eq:seqIII}
\xymatrix{0 \ar[r] & \underline{H^{d-1}_{dR}(T_B;\mathbbm{k})} \ar[r] &
\mathcal{H}^0(\varphi_{B+} \mathcal{O}_{T_B \times \AA^n}) \ar[r] & \mathcal{M}^0_A
\ar[r] & \underline{H^d_{dR}(T_B;\mathbbm{k})} \ar[r] & 0}
\end{gather}
which is isomorphic to the sequence \eqref{eq:seqII}.
\subsection{Vanishing Gau\ss--Manin system and the extension class}
In this section we show that the $A$-hypergeometric system is an
extension of a trivial vector bundle of rank one
by the quotient of
a Gau\ss--Manin system modulo
its flat sections. We show
that this extension does not split.
As before,
$\beta=0$ (and $A$ is saturated, homogeneous, and pointed).
\begin{dfn}
The \emph{vanishing Gau\ss--Manin system} $\mathcal{V}$ with respect to
the map $\varphi_B$ is the cokernel of the map
$\underline{H^{d-1}_{dR}(T_B\mathbbm{k})} \longrightarrow \mathcal{H}^0(\varphi_{B+}
\mathcal{O}_{T_B \times \AA^n})$. In other words,
\begin{equation}\label{eq:Pshortexseq}
0 \longrightarrow \mathcal{V} \longrightarrow \mathcal{M}^0_A \longrightarrow \underline{H^d_{dR}(T_B;\mathbbm{k})} \longrightarrow 0
\end{equation}
is exact. We write $V_A=\Gamma(\AA^{n+1},\mathcal{V}_A)$ and note the
short exact sequence
\[
0\to V_A\to M^0_A\to O_A\to 0.
\]
\hfill\ensuremath{\diamond}\end{dfn}
The terminology is borrowed from the vanishing cohomology of a
hyperplane section $j:X \hookrightarrow Y$ of an $n$-dimensional
projective variety $Y$ which is a direct summand $H^{n-1}(X) =
H^{n-1}(X)_{van} \oplus j^*H^{n-1}(Y)$.
The sheaf $\mathcal{V}$ appears perhaps for the first time in Stienstra's
article \cite[Formula (61)]{Stienstra98}, essentially as a restriction
of \eqref{eq:Pshortexseq} to the smooth locus (where all sheaves in
\eqref{eq:seqII} become vector bundles). However, our situation is
more general even in Stienstra's set-up since in \cite{Stienstra98} the matrix
$B$ is assumed to be homogeneous while it is arbitrary for us.
A natural question is: what is the extension class of $\mathcal{M}^0_A$ inside the
sequence \eqref{eq:Pshortexseq}? Our next result answers this question,
confirming a prediction of Duco van Straten.
\begin{thm}\label{thm-calcExt}
Write $\mathcal{O}$ for $H^d_{dR}(T_B;\mathbbm{k})
\otimes \mathcal{O}_{\AA^{n+1}}$. There are natural (in $\mathbbm{k}$) isomorphisms
\[
\Ext^i_\mathcal{D}(\mathcal{O},\mathcal{V})\simeq \begin{cases}
\mathbbm{k} &
\text{for}\; i = 1 \\ 0 & \text{else.} \end{cases}
\]
The class of the sequence \eqref{eq:Pshortexseq} is nonzero and
induced by the identity on $\mathcal{O}$ under the connecting morphism.
\end{thm}
\begin{proof}
Since $\AA^{n+1}$ is affine it suffices to compute on the level of
global sections. By Corollary \ref{cor-TorExt-zero},
$\Ext^\bullet_{D_A}(O_A,M_A^\beta)$ vanishes for $\beta\in \mathbb{N} A$.
Hence, $\Ext^i_{D_A}(O_A,V_A)=\Ext^{i-1}_{D_A}(O_A,O_A)$ and so has
exactly the prescribed $\mathbbm{k}$-space structure. In particular,
\eqref{eq:Pshortexseq} does not split.
The class of \eqref{eq:Pshortexseq} inside $\Ext^1_{D}(O_A,V)\simeq
\Ext^{0}_{D}(O_A,O_A)$ is the image of the identity on $\mathcal{O}_A$
under the connecting morphism induced by \eqref{eq:Pshortexseq},
compare \cite[Sec.~3.4]{Weibel}. Since the connecting
morphism is an isomorphism, this element is nontrivial.
\end{proof}
\bibliographystyle{amsalpha}
|
\section*{Introduction
\setcounter{equation}{0}
We say that a vector field on a Riemannian manifold is {\medit geodesic\/} if
its integral curves are re\-pa\-ram\-e\-trized geodesics. The present paper
discusse
\begin{equation}\label{trp}
\begin{array}{l}
\mathrm{triples\ }\hskip.7pt(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\hskip.7pt\mathrm{\ consisting\ of\ a\ compact\
complex\ manifold\ }\hsM}%{{\mathrm{M}}\hskip-.7pt\mathrm{,\ a}\\
\mathrm{K}\ddot{\mathrm{a}}\mathrm{hler\ metric\ }\,g\,\mathrm{\ on\
}\,M}%{{\mathrm{M}}\hskip-1.5pt\mathrm{,\ and\ a\ nonconstant\ function\
}\,\vt:M}%{{\mathrm{M}}\hskip-.7pt\to\mathrm{I\!R}}%{\mathbf{R}\mathrm{,}\\
\mathrm{the\hskip.7pt\hs\ }\,g\hyp\mathrm{gra\-di\-ent\hskip.7pt\hs\ of\hskip.7pt\hs\ which\hskip.7pt\hs\
is\hskip.7pt\hs\ both\hskip.7pt\hs\ geodesic\hskip.7pt\hs\ and\hskip.7pt\hs\
real}\hyp\mathrm{hol\-o\-mor\-phic.}
\end{array}
\end{equation}
We observe (Remark~\ref{dppdm}) that for $\,m=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}\hskip.7pt$ and
$\,d_\pm^{\phantom i}\hskip-.7pt=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\varSigma^\pm\hskip-.7pt$, where $\,\varSigma^+$ and $\,\varSigma^-$ are the
maximum and minimum level sets of $\,\vt$, one then has
\begin{equation}\label{ddm}
d_+^{\phantom i}+\hskip.7pt\,d_-^{\phantom i}\hskip.7pt\,\ge\,\,m\,-\,1\,\,\ge\hskip.7pt\,\,d_\pm^{\phantom i}\hskip.7pt\,\ge\,\,0\hskip.4pt,
\end{equation}
and every $\,(d_+^{\phantom i},d_-^{\phantom i},m)\in\hbox{$\mathsf{Z}\hskip-4.5pt\mathsf{Z}$}^3$ satisfying (\ref{ddm}) is realized by
some $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ with (\ref{trp}).
One of our three main results, Theorem~\ref{cpbdl}, classifies the triples
(\ref{trp}) such that
\begin{equation}\label{spn}
\mathrm{Ker}\hskip2.3ptd\pi^+\mathrm{\ and\
}\,\hskip.7pt\mathrm{Ker}\hskip2.3ptd\pi^-\mathrm{\ span\ an\ in\-te\-gra\-ble\
distribution\ on\ }\,M}%{{\mathrm{M}}'\hskip-.7pt.
\end{equation}
Here $\,M}%{{\mathrm{M}}'\hskip-.7pt=M}%{{\mathrm{M}}\smallsetminus(\varSigma^+\hskip-1.5pt\cup\varSigma^-)$, while
$\,\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt\to\varSigma^\pm$ sends each
$\,x\inM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp$ to the unique point nearest $\,x\,$ in
$\,\varSigma^\pm\hskip-.7pt$. (In case (\ref{trp}) $\,\pi^\pm$ always are disk-bundle
projections, and their vertical distributions
$\,\mathrm{Ker}\hskip2.3ptd\pi^\pm$ span a vector sub\-bun\-dle of
$\,T\nnhM}%{{\mathrm{M}}'\hskip-.7pt$, cf.\ Section~\ref{cc}; however, (\ref{trp}) does not imply
(\ref{spn}) -- see Remark~\ref{nintg}.)
As a consequence of Theorem~\ref{cpbdl}, in every triple with (\ref{trp}) and
(\ref{spn}),
\begin{equation}\label{bdl}
\begin{array}{l}
M}%{{\mathrm{M}}\,\mathrm{\hskip-.7pt\ is\hskip-.7pt\ bi\-hol\-o\-mor\-phic\hskip-.7pt\ to\hskip-.7pt\ a\hskip-.7pt\ bundle\hskip-.7pt\
of\hskip-.7pt\ (pos\-i\-tive\hskip.7pt}\hyp\hskip.7pt\mathrm{di\-men\-sion\-al)\hskip-.7pt\ complex}\\
\mathrm{projective\hskip.7pt\ spaces\hskip.7pt\ over\hskip.7pt\ some\hskip.7pt\ base\hskip.7pt\ manifold\
}\,\,\,B\,\,\mathrm{\ with\ }\,\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5ptB\ge0.
\end{array}
\end{equation}
The remaining two main results of the paper, Theorems~\ref{tgimm}
and~\ref{dicho}, deal with the general case of (\ref{trp}), that is, do not
assume (\ref{spn}).
According to Theorem~\ref{tgimm}, whenever $\,\varPi^\pm$ is a leaf of either
(obviously in\-te\-gra\-ble) vertical distribution
$\,\mathrm{Ker}\hskip2.3ptd\pi^{\hskip.7pt\pm}\hskip-.7pt$, the {\medit other\/} projection
$\,\pi^\mp$ maps $\,\varPi^\pm\hskip-.7pt\capM}%{{\mathrm{M}}'$ onto the image $\,F(\bbC\mathrm{P}^k)\,$ of
some totally geodesic hol\-o\-mor\-phic immersion $\,F:\bbC\mathrm{P}^k\hskip-.7pt\to\varSigma^\mp$
inducing on $\,\bbC\mathrm{P}^k$ a multiple of the Fu\-bi\-ni-Stu\-dy metric, with
$\,k=k_\pm^{\phantom i}\hskip-.7pt\ge0\,$ given by $\,k_\pm^{\phantom i}\hskip-.7pt=m-1-\hskip.7pt d_\pm^{\phantom i}$.
Both $\,\varSigma^\pm$ are themselves (connected) totally geodesic
compact complex sub\-man\-i\-folds of $\,M}%{{\mathrm{M}}\hskip-.7pt$, cf.\ Remark~\ref{ascdt}(iii).
The third main result reveals a dichotomy involving the assignment
\begin{equation}\label{asg}
M}%{{\mathrm{M}}'\hskip-.7pt\ni x\,
\mapsto\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathrm{Ker}\hskip2.3ptd\pi\hskip-.7pt_x^\mp\hskip-.4pt)
\in\hskip.7pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)\hskip7pt\mathrm{for}\hskip7pt
y=\pi^\pm\hskip-.7pt(x)\hskip.4pt,
\end{equation}
$\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)\,$ being the complex Grass\-mann\-i\-an,
with $\,k=k_\pm^{\phantom i}$ defined as before.
Specifically, Theorem~\ref{dicho} states that one of the following two cases
has to occur. First, (\ref{asg}) may be {\medit constant\/} on every leaf of
$\,\mathrm{Ker}\hskip2.3ptd\pi\hskip-.7pt_x^{\hskip.7pt\pm}$ in $\,M}%{{\mathrm{M}}'\hskip-.7pt$, that is, on
every fibre $\,\varPi^\pm$ of the projection
$\,\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt\to\varSigma^\pm$ restricted to
$\,M}%{{\mathrm{M}}'\hskip-.7pt$, with either sign $\,\pm\hskip.7pt$. Otherwise, $\,l=k_\mp^{\phantom i}$ and
$\,k=k_\pm^{\phantom i}$ are positive for both signs $\,\pm\hskip.7pt$, while (\ref{asg})
restricted to any such leaf $\,\varPi^\pm$ must be a composite mapping
$\,\varPi^\pm\hskip-1.5pt\to\hskip.4pt\bbC\mathrm{P}\hskip.4pt^l\hskip-1.5pt
\to\hskip.4pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)\,$
formed by a hol\-o\-mor\-phic bundle projection
$\,\varPi^\pm\hskip-1.5pt\to\hskip.4pt\bbC\mathrm{P}\hskip.4pt^l\hskip-.7pt$, having the fibre
$\,\bbC\smallsetminus\{0\}$, and a {\medit nonconstant hol\-o\-mor\-phic
embedding\/} $\,\bbC\mathrm{P}\hskip.4pt^l\hskip-1.5pt\to\hskip.4pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)$.
The first case of Theorem~\ref{dicho} is equivalent to condition (\ref{spn}),
and the immersions $\,\bbC\mathrm{P}^k\hskip-.7pt\to\varSigma^\pm\hskip-.7pt$, mentioned in the above summary
of Theorem~\ref{tgimm}, are then embeddings, for both signs $\,\pm\hskip.7pt$, while
their images constitute foliations of $\,\varSigma^\pm\hskip-.7pt$, both with the same leaf
space $\,B\,$ appearing in (\ref{bdl}). See Remark~\ref{infty}.
In the second case (cf.\ Remark~\ref{nncst}) the images of these immersions,
rather than being pairwise disjoint, are totally geodesic,
hol\-o\-mor\-phic\-al\-ly immersed complex projective spaces, an uncountable
family of which passes through each point of $\,\varSigma^\pm\hskip-.7pt$.
Three special classes of the objects (\ref{trp}) have been studied before. One
is provided by the gradient K\"ah\-\hbox{ler\hskip.7pt-}\hskip0ptRic\-ci sol\-i\-tons
discovered by Koiso \cite{koiso} and, independently, Cao \cite{cao}, where
$\,\vt\,$ is the sol\-i\-ton function; two more -- by special
K\"ah\-\hbox{ler\hskip.7pt-}\hskip0ptRic\-ci potentials $\,\vt\,$ on compact
K\"ah\-ler manifolds \cite{derdzinski-maschler-06}, and by triples with
(\ref{trp}) such that $\,M}%{{\mathrm{M}}\,$ is a (compact) complex surface
\cite{derdzinski-kp}. Each of these three classes satisfies (\ref{spn}).
The papers \cite{derdzinski-maschler-06,derdzinski-kp} provide complete
explicit descriptions of the classes discussed in them. Our
Theorem~\ref{cpbdl} generalizes their classification results, namely,
\cite[Theorem 16.3]{derdzinski-maschler-06} and
\cite[Theorem 6.1]{derdzinski-kp}.
For more details on the preceding two paragraphs, see Remark~\ref{clone}.
Functions with geodesic gradients on arbitrary Riemannian manifolds, usually
called {\it trans\-nor\-mal}, have been studied extensively as well
\cite{wang,miyaoka,bolton}.
Both authors' research was supported in part by a FAPESP\hskip-.7pt--\hskip.7pt OSU 2015
Regular Research Award (FAPESP grant: 2015/50265-6). The authors wish to thank
Fangyang Zheng for helpful comments.
\section{Preliminaries}\label{pr}
\setcounter{equation}{0}
Manifolds, mappings and tensor fields, including Riemannian metrics and
functions, are by definition of class $\,C^\infty\hskip-1.5pt$. A (sub)manifold
is always assumed connected.
Our sign convention about the curvature tensor $\,R=\hskip-1.5pt R\hskip-.4pt^\nabla$ of a
connection $\,\nabla\,$ in a vector bundle $\,E}%{{\mathrm{B}}\,$ over a manifold $\,M}%{{\mathrm{M}}\,$
is
\begin{equation}\label{cur}
R\hskip.4pt(\hskip-.4pt v,w)\hskip.4pt\xi\,=\,\nabla\hskip-3pt_w^{\phantom i}\hskip-.7pt\nabla\hskip-3pt_v^{\phantom i}\hskip.7pt\xi\,
-\,\nabla\hskip-3pt_v^{\phantom i}\hskip-.7pt\nabla\hskip-3pt_w^{\phantom i}\hskip.7pt\xi\,
+\,\nabla\hskip-3pt_{[v,w]}^{\phantom i}\hskip.4pt\xi
\end{equation}
for any section $\,\xi\,$ of $\,E}%{{\mathrm{B}}\,$ and vector fields $\,v,w\,$ tangent to
$\,M}%{{\mathrm{M}}\hskip-.7pt$. One may treat $\,R\hskip.4pt(\hskip-.4pt v,w)$, the covariant derivative
$\,\nabla\xi$, and any function $\,f\,$ on $\,M}%{{\mathrm{M}}\,$ as bundle morphisms
\begin{equation}\label{nwt}
R\hskip.4pt(\hskip-.4pt v,w),\hskip.7pt f:E}%{{\mathrm{B}}\toE}%{{\mathrm{B}}\hskip.4pt,\hskip22pt
\nabla\xi:T\nnhM}%{{\mathrm{M}}\toE}%{{\mathrm{B}}
\end{equation}
sending $\,\xi\,$ or $\,v\,$ as above to $\,R\hskip.4pt(\hskip-.4pt u,v)\hskip.4pt\xi$, $\,f\xi\,$
or, respectively, $\,\nabla\hskip-3pt_v^{\phantom i}\xi$. Notation of (\ref{nwt}) is used
in the next three displayed relations.
In the case of a Riemannian manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$, the symbol $\,\nabla\,$
will always stand for the Le\-vi-Ci\-vi\-ta connection of $\,g\,$ as well as
the $\,g$-gra\-di\-ent. Given a function $\,\vt\,$ and vector fields
$\,w,w\hskip.4pt'$ on $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$, one has the Lie-de\-riv\-a\-tive relation
\begin{equation}\label{lvg}
[\hskip-.7pt\pounds\hskip-.5pt_v^{\phantom i} g](\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)=2\hskip.4pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt,\hskip8pt\mathrm{where}
\hskip5ptv=\nabla\hskip-.5pt\vt\hskip5pt\mathrm{and}\hskip5ptS
=\nabla\hskip-.7pt v:T\nnhM}%{{\mathrm{M}}\to T\nnhM}%{{\mathrm{M}}\hskip.4pt,
\end{equation}
due to the local-coordinate equalities
$\,[\hskip-.7pt\pounds\hskip-.5pt_v^{\phantom i} g]_{jk}^{\phantom i}=v_{j,k}^{\phantom i}+v_{k,j}^{\phantom i}=2v_{j,k}^{\phantom i}$. For
vector fields $\,v,u\,$ on a manifold $\,M}%{{\mathrm{M}}\,$ and a bundle morphism
$\,B}%{{\mathrm{B}}:T\nnhM}%{{\mathrm{M}}\to T\nnhM}%{{\mathrm{M}}\hskip-.7pt$, the Leib\-niz rule gives
$\,[\hskip-.7pt\pounds\hskip-.5pt_v^{\phantom i}B}%{{\mathrm{B}}]u=[v,B}%{{\mathrm{B}} u]-B}%{{\mathrm{B}}[v,u]
=[\nabla\hskip-3pt_u^{\phantom i}B}%{{\mathrm{B}}]u+B}%{{\mathrm{B}}\nabla\hskip-3pt_u^{\phantom i} v
-\nabla\hskip-3pt_{B}%{{\mathrm{B}} u}^{\phantom i} v$, and so
\begin{equation}\label{lie}
\pounds\hskip-.5pt_v^{\phantom i}B}%{{\mathrm{B}}\,=\,\nabla\hskip-3pt_u^{\phantom i}B}%{{\mathrm{B}}\,+\,[B}%{{\mathrm{B}},\nabla\hskip-.7pt v]\hskip.4pt.
\end{equation}
Next, let $\,u\,$ be a Kil\-ling vector field on a Riemannian manifold
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$. The Ric\-ci and Bian\-chi identities imply, as in
\cite[bottom of p.\ 572]{derdzinski-roter-07}, the well-known relation
\begin{equation}\label{scd}
\nabla\hskip-3pt_v^{\phantom i}A\,=\,R(\hskip-.4pt u,v)\hskip.4pt,\hskip12pt\mathrm{with\ }\,\,A
=\nabla\hskip-.7pt u\hskip.4pt.
\end{equation}
Since the flow of a Kil\-ling field preserves the Le\-vi-Ci\-vi\-ta
connection, (\ref{scd}) also follows from the classical Lie-de\-riv\-a\-tive
equality$\,[\hskip-.7pt\pounds\hskip-.5pt_u^{\phantom i}\hskip-1.5pt\nabla]\hskip-.4pt _v^{\phantom i} w
=[\nabla\hskip-3pt_v^{\phantom i}A]\hskip.7pt w-R\hskip.4pt(\hskip-.4pt u,v)\hskip.4pt w$,
with $\,A=\nabla\hskip-.7pt u$, cf.\ \cite[formula (1.8) on p. 337]{schouten}, valid
for any connection $\,\nabla\,$
in $\,T\nnhM}%{{\mathrm{M}}\hskip-.7pt$.
Whenever $\,\vt:M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}\,$ is a function on a Riemannian manifold
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$, we have
\begin{equation}\label{tnd}
\nabla Q\,=\,2\hskip.4pt\nabla\hskip-3pt_v^{\phantom i} v\hskip.4pt,\hskip12pt\mathrm{where}\hskip7ptv
=\nabla\hskip-.5pt\vt\hskip7pt\mathrm{and}\hskip7ptQ=g(\hskip-.4pt v,v)\hskip.4pt,
\end{equation}
as one sees noting that, in local coordinates,
$\,(\vt\hskip-1.5pt_{,\hskip.7pt k}^{\phantom i}\vt^{\hskip.4pt,\hskip.4pt k})\hskip-.7pt_{,\hskip.4pt j}^{\phantom i}
=2\vt\hskip-1.5pt_{,\hskip.7pt kj}^{\phantom i}\vt^{\hskip.4pt,\hskip.4pt k}$. Also, obviously
\begin{equation}\label{dvt}
d_v^{\phantom i}\vt\,=\,g(\hskip-.4pt v,\nabla\hskip-.5pt\vt)\,=\,Q\hskip12pt\mathrm{if}\hskip7ptv=\nabla\hskip-.5pt\vt\hskip7pt
\mathrm{and}\hskip7ptQ=g(\hskip-.4pt v,v)\hskip.4pt.
\end{equation}
\begin{remark}\label{dvwwp}Relation (\ref{lvg}) becomes
$\,d_v^{\phantom i}[\hskip.7pt g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)]=2\hskip.4pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\,$ if, in addition,
$\,v$ commutes with $\,w\,$ and $\,w\hskip.4pt'\hskip-.7pt$. Namely,
$\,d_v^{\phantom i}=\hskip-.7pt\pounds\hskip-.5pt_v^{\phantom i}$ on functions, so that we may evaluate
$\,d_v^{\phantom i}[\hskip.7pt g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)]\,$ using the Leib\-niz rule for the Lie
derivative with $\,\pounds\hskip-.5pt_v^{\phantom i} w=\pounds\hskip-.5pt_v^{\phantom i} w\hskip.4pt'\hskip-.7pt=0$.
\end{remark}
\begin{remark}\label{gpgrd}Whenever the $\,g$-gra\-di\-ent
$\,v=\nabla\hskip-.5pt\vt\,$ of a function $\,\vt\,$ on a Riemannian manifold
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g)\,$ is tangent to a sub\-man\-i\-fold $\,\varPi\,$ with the
sub\-man\-i\-fold metric $\,g'\hskip-1.5pt$, the restriction of $\,v\,$ to
$\,\varPi\,$ obviously equals the $\,g'\hskip-1.5pt$-gra\-di\-ent of
$\,\vt:\varPi\to\mathrm{I\!R}}%{\mathbf{R}$.
\end{remark}
\begin{remark}\label{cifot}Given a manifold $\,M}%{{\mathrm{M}}\,$ and
$\,\sigma,\vt:M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}$, we call $\,\sigma\,$ {\it a\/ $\,C^\infty$ function
of\/} $\,\vt\,$ if $\,\vt\,$ is nonconstant (so that its range $\,\vt(M}%{{\mathrm{M}})\,$
is an interval) and $\,\sigma=\chi\circ\tau\,$ for some $\,C^\infty$ function
$\,\chi:\vt(M}%{{\mathrm{M}})\to\mathrm{I\!R}}%{\mathbf{R}$. Note that $\,\chi\,$ is then uniquely determined by
$\,\sigma$ and $\,\vt$. We will denote by $\,\sigma\,$ both the original
function $\,M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}\,$ and the function $\,\chi:I\to\mathrm{I\!R}}%{\mathbf{R}\,$ of the variable
$\,\vt\in I\hskip-.7pt$.
\end{remark}
Let $\,(t,s)\mapsto x(t,s)\inM}%{{\mathrm{M}}\,$ be a fixed {\medit variation of curves\/}
in a manifold $\,M}%{{\mathrm{M}}\hskip-.7pt$, that is, a $\,C^\infty$ mapping in which the real
variables $\,t,s\,$ range independently over intervals. The {\medit partial
derivative\/} $\,x_t^{\phantom i}$ (or, $\,x_s^{\phantom i}$) then assigns to each
$\,(t_0^{\phantom i},s_0^{\phantom i})\,$ the velocity vector at $\,t_0^{\phantom i}$ (or, $\,s_0^{\phantom i}$) of the
curve $\,t\mapsto x(t,s_0^{\phantom i})\,$ or, respectively, $\,s\mapsto x(t_0^{\phantom i},s)$.
(Thus, $\,x_t^{\phantom i}$ and $\,x_s^{\phantom i}$ are sections of a specific pull\-back
bundle.) A connection $\,\nabla\,$ on $\,M}%{{\mathrm{M}}\,$ allows us to define
the {\medit mixed sec\-ond-or\-der partial derivatives\/} $\,x_{ts}^{\phantom i}$ and
$\,x_{st}^{\phantom i}$ of the variation, so that, for instance, the value of
$\,x_{ts}^{\phantom i}$ at $\,(t_0^{\phantom i},s_0^{\phantom i})\,$ is the $\,\nabla\hskip-.7pt$-co\-var\-i\-ant
derivative, at the parameter $\,s_0^{\phantom i}$, of the vector field
$\,s\mapsto x_t^{\phantom i}(t_0^{\phantom i},s)\,$ along the curve $\,s\mapsto x(t_0^{\phantom i},s)$, and
analogously for $\,x_{st}^{\phantom i}$. Obviously, $\,x\hskip-.7pt_{st}^{\phantom i}\hskip-.7pt=x_{ts}^{\phantom i}$ when
$\,\nabla\,$ is tor\-sion-free, cf.\ \cite[p.\ 101]{derdzinski-maschler-06}.
\begin{remark}\label{jcobi}For a tor\-sion-free connection $\,\nabla\,$ on a
manifold $\,M}%{{\mathrm{M}}\,$ and a smooth variation
$\,(t,s)\mapsto x(t,s)=\exp_{y(s)}^{\phantom i}\hskip.7pt(t-t_0^{\phantom i})\xi(s)\,$ of
$\,\nabla\hskip-.7pt$-ge\-o\-des\-ics, with $\,(t,s)\,$ near $\,(t_0^{\phantom i},s_0^{\phantom i})$ in
$\,\bbR^2$ and a vector field
$\,s\mapsto\xi(s)\in T\hskip-3pt_{y(s)}^{\phantom i}\hskip-.7ptM}%{{\mathrm{M}}\,$ along a curve
$\,s\mapsto y(s)\inM}%{{\mathrm{M}}\hskip-.7pt$, let $\,t\mapsto\hat w(t)\,$ be the Ja\-co\-bi
vector field along the geodesic $\,t\mapsto x(t)=x(t,s_0^{\phantom i})\,$ defined by
$\,\hat w(t)=x_s^{\phantom i}(t,s_0^{\phantom i})\,$ (notation of the last paragraph). Then
$\,[\nabla\hskip-3pt_{\dot x}^{\phantom i}\hat w](t_0^{\phantom i})
=[\nabla\hskip-3pt_{\dot y}^{\phantom i}\hskip.7pt\xi](s_0^{\phantom i})$,
which is nothing else than $\,x\hskip-.7pt_{st}^{\phantom i}\hskip-.7pt=x_{ts}^{\phantom i}$ (see above) at
$\,(t,s)=(t_0^{\phantom i},s_0^{\phantom i})$. Also, clearly, $\,\hat w(t_0^{\phantom i})=\dot y(s_0^{\phantom i})$. Note
that $\,\nabla\hskip-3pt_{\dot y}^{\phantom i}\xi\,$ may be nonzero even if $\,y(s)=y\,$
is a constant curve, as it then equals the ordinary derivative of
$\,s\mapsto\xi(s)\in T\hskip-3pt_y^{\phantom i}\hskip-.7ptM}%{{\mathrm{M}}\,$ with respect to $\,s$.
\end{remark}
\begin{remark}\label{tlspc}Let $\,N}%{{\mathrm{N}}\,$ be a vector bundle over a manifold
$\,\varSigma$. We use the same symbol $\,N}%{{\mathrm{N}}\hskip.7pt$ for its total space, which we
identify, as a set, with
$\,\{(y,\xi):y\in\varSigma\,\mathrm{\ and\ }\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\}$. Given
a connection $\,\mathrm{D}\,$ in $\,N}%{{\mathrm{N}}\hskip.7pt$ and vector fields $\,v,w\,$
tangent to $\,\varSigma$,
\begin{equation}\label{cmt}
\mathrm{the\ }\hskip.7pt\mathrm{D}\hskip.4pt\hyp\hskip-.4pt\mathrm{hor\-i\-zon\-tal\ lifts\ of\
}\hskip.7pt v\hskip.7pt\mathrm{\ and\ }\hskip.7pt w\hskip.7pt\mathrm{\ commute\ if\ so\ do\
}\hskip.7pt v,w\hskip.7pt\mathrm{\ and\ }\hskip.7pt R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt v,w)=0
\end{equation}
(notation of (\ref{nwt})). Namely, at any $\,x=(y,\xi)\inN}%{{\mathrm{N}}\hskip-1.5pt$, the vertical
(or, horizontal) component of the Lie bracket of the horizontal lifts of
$\,v\,$ and $\,w\,$ equals
$\,R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt v_y^{\phantom i},w_y^{\phantom i})\hskip.4pt\xi$ (an easy exercise) or,
respectively, the horizontal lift of $\,[v,w]_y^{\phantom i}$, cf.\
\cite[p.\ 10]{kobayashi-nomizu}.
\end{remark}
\begin{remark}\label{nexpm}Let $\,\mathrm{D}\,$ be the normal connection in
the normal bundle $\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma\,$ of a totally geodesic
sub\-man\-i\-fold $\,\varSigma\,$ in a Riemannian manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$. We denote by
$\,\mathrm{Exp}^\perp\hskip-1.5pt:U\hskip-.7pt\toM}%{{\mathrm{M}}\,$ the normal exponential mapping
of $\,\varSigma$, the domain of which is an open sub\-man\-i\-fold $\,\,U$ of the
total space $\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma\,$ such that, for every normal space
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma$, where $\,y\in\varSigma^\pm\hskip-.7pt$, the intersection
$\,\,U\hskip-.7pt\capN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma\,$ is nonempty and star-shap\-ed (in the
sense of being a union of line segments emanating from $\,0$).
Remark~\ref{jcobi} leads to the following well-known description of the
differential $\,d\hskip.7pt\mathrm{Exp}^\perp_{(y,\,\xi)}$ of $\,\mathrm{Exp}^\perp$
at any $\,(y,\xi)\in\,U\hskip-1.5pt$, cf.\ Remark~\ref{tlspc}. Specifically, we may
assume that $\,\xi\ne0\,$ since, clearly,
$\,d\hskip.7pt\mathrm{Exp}^\perp_{(y,\,\xi)}=\hskip.7pt\mathrm{Id}$ when $\,\xi=0$, under
the obvious isomorphic identification
$\,T_{(y,0)}^{\phantom i}[N}%{{\mathrm{N}}\hskip-2.3pt\varSigma]
=T\hskip-3pt_y^{\phantom i}\varSigma\oplusN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma
=T\hskip-3pt_y^{\phantom i}M}%{{\mathrm{M}}\hskip-.7pt$. The point $\,y\in\varSigma\,$ and the normal vector
$\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma$ thus have the property that the nontrivial
geodesic $\,r\mapsto x(\hskip-.4pt r\hskip-.7pt)=\exp_y^{\phantom i}\hskip-.4pt r\xi\,$ is defined for all
$\,r\in[\hs0,1\hskip.4pt]$. If $\,r>0$, a vector tangent to
$\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma\,$ at $\,(y,r\xi)\,$ can be uniquely written as
$\,r\eta+w_r^{\hskip.5pt\text{\rm hrz}}\hskip-.7pt$, where
$\,\eta\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma=T_{(y,\,r\xi)}^{\phantom i}[N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma]\,$
is vertical and $\,w_r^{\hskip.5pt\text{\rm hrz}}$ denotes the $\,\mathrm{D}\hskip.4pt$-\hskip-.4pt hor\-i\-zon\-tal
lift of some $\,w\in T\hskip-3pt_y^{\phantom i}\varSigma$. Then, for the Ja\-cobi field
$\,r\mapsto\hat w(\hskip-.4pt r\hskip-.7pt)\,$ along our geodesic $\,r\mapsto x(\hskip-.4pt r\hskip-.7pt)\,$
such that $\,\hat w(0)=w\,$ and
$\,[\nabla\hskip-3pt_{\dot x}^{\phantom i}\hat w](0)=\eta$,
\begin{equation}\label{dxp}
d\hskip.7pt\mathrm{Exp}^\perp_{(y,\,r\xi)}(r\eta+w_r^{\hskip.5pt\text{\rm hrz}})\,
=\,\hat w(\hskip-.4pt r\hskip-.7pt)\hskip13pt\mathrm{whenever}\hskip6ptr\in[\hs0,1\hskip.4pt]\hskip.4pt.
\end{equation}
In fact, linearity of both sides in $\,(\eta,w)\,$ allows us to consider two
separate cases, $\,w=0\,$ and $\,\eta=0$. For $\,s\,$ close to $\,0\,$
in $\,\mathrm{I\!R}}%{\mathbf{R}\,$ and $\,r\in[\hs0,1\hskip.4pt]$, let us set
$\,x(r,s)=\exp_{y(s)}^{\phantom i}\hskip.7pt r\xi(s)$, where in the former case
$\,(y(s),\xi(s))=(y,\xi+s\eta)$, and in the latter $\,s\mapsto\xi(s)\,$ is the
$\,\mathrm{D}\hskip.4pt$-\hskip-.7pt par\-al\-lel normal vector field with $\,\xi(0)=\xi\,$
along a fixed curve $\,s\mapsto y(s)\in\varSigma\,$ such that $\,y(0)=y\,$ and
$\,\dot y(0)=w$. Thus, in both cases, the curve $\,s\mapsto (y(s),r\xi(s))\,$
in $\,\,U\hskip.7pt$ has, at $\,s=0$, the velocity $\,r\eta+w_r^{\hskip.5pt\text{\rm hrz}}$. The
velocity at $\,s=0$ of its $\,\mathrm{Exp}^\perp\hskip-.7pt$-im\-age curve
$\,s\mapsto x(r,s)\,$ therefore equals the left-hand side of (\ref{dxp}). At
the same time this last velocity is $\,\hat w(\hskip-.4pt r\hskip-.7pt)=x_s^{\phantom i}(r,0)\,$ for
$\,\hat w\,$ defined as in Remark~\ref{jcobi} with the variable $\,t\,$ and
$\,(t_0^{\phantom i},s_0^{\phantom i})\,$ replaced by $\,r\,$ and $\,(0,0)$. Now (\ref{dxp}) follows
since the two definitions of $\,\hat w\,$ agree: according to
Remark~\ref{jcobi}, both Ja\-cobi fields denoted by $\,\hat w\,$ satisfy the
same initial conditions at $\,s=0$.
\end{remark}
\begin{remark}\label{kiljc}Every Kil\-ling vector field $\,u\,$ on a
Riemannian manifold is a Ja\-co\-bi field along any geodesic
$\,t\mapsto x(t)$. In fact, the local flow of $\,u$, applied to the geodesic,
yields a variation of geodesics. (Equivalently, one may note that (\ref{scd})
with $\,v=\dot x$, evaluated on $\,\dot x$, is precisely the Ja\-co\-bi
equation.)
\end{remark}
\begin{remark}\label{ttgim}Let $\,\varPsi:\varPi\toM}%{{\mathrm{M}}\,$ be a totally geodesic
immersion of a manifold $\,\varPi\hskip.7pt$ in a Riemannian manifold $\,(M,g)$. If
$\,\varPsi(\hskip-.7pt\varLambda)\subseteq\varSigma\,$ and
$\,\varPsi(\varPi\hskip-.7pt\smallsetminus\varLambda)\subseteqM}%{{\mathrm{M}}\smallsetminus\varSigma\,$ for
sub\-man\-i\-folds $\,\varLambda\,$ of $\,\varPi\hskip.7pt$ and $\,\varSigma\,$ of
$\,M}%{{\mathrm{M}}\hskip-.7pt$, such that $\,\varSigma\,$ is totally geodesic in $\,(M,g)$, then, for
$\,\varSigma$ endowed with the sub\-man\-i\-fold metric,
$\,\varPsi:\varLambda\to\varSigma\,$ is a totally geodesic immersion.
In fact, every point of $\,\varLambda\,$ has a neighborhood $\,\,U\,$ in
$\,\varPi\,$ on which $\,\varPsi\,$ is an embedding with a totally geodesic image
$\,\varPsi(\hskip-.4pt u)$. Our claim now follows since the sub\-man\-i\-fold
$\,\varPsi(\varLambda\cap U)\,$ of $\,\varSigma$, being the
intersection of the totally geodesic sub\-man\-i\-folds $\,\varPsi(\hskip-.4pt u)$ and
$\,\varSigma$, must itself be totally geodesic.
\end{remark}
\begin{remark}\label{crvcm}Let $\,R,R\hskip.4pt'$ and $\,\hat R\,$ be the curvature
tensors of connections $\,\nabla,\nabla'$ in vector bundles $\,E}%{{\mathrm{B}},E}%{{\mathrm{B}}'$ over
a fixed base manifold and, respectively, of the connection $\,\hat\nabla\,$
induced by them in the vector bundle $\,\mathrm{Hom}\hskip.7pt(E}%{{\mathrm{B}},E}%{{\mathrm{B}}')$. Then
$\,\hat R\,$ is given by the com\-mu\-ta\-tor-type formula
$\,\hat R(\hskip-.4pt v,w)\hskip.4pt\varTheta\hskip.7pt
=\hskip.7pt[R\hskip.4pt'\hskip-.7pt(\hskip-.4pt v,w)]\varTheta\hskip.7pt-\hskip.7pt\varTheta[R(\hskip-.4pt v,w)]$, cf.\
(\ref{nwt}), for any section $\,\varTheta\,$ of
$\,\mathrm{Hom}\hskip.7pt(E}%{{\mathrm{B}},E}%{{\mathrm{B}}')\,$ (that is, any
vec\-tor-bun\-dle morphism $\,\varTheta:E}%{{\mathrm{B}}\toE}%{{\mathrm{B}}'$) and vector fields
$\,v,w\,$ tangent to the base. This trivially follows from (\ref{cur}) and the
fact that $\,[\hat\nabla\hskip-3pt_v^{\phantom i}\varTheta]\hskip.4pt\xi
=\nabla'\hskip-6pt_v^{\phantom i}(\varTheta\hskip.4pt\xi)-\varTheta\nabla\hskip-3pt_v^{\phantom i}\xi\,$
whenever $\,\xi\,$ is a section of $\,E}%{{\mathrm{B}}$.
\end{remark}
\section{Projectability of distributions}\label{pd}
\setcounter{equation}{0}
As usual, whenever $\,\pi:M}%{{\mathrm{M}}\toB\,$ is a mapping between manifolds, we
say that a vector field $\,w\,$ (or, a distribution $\,\mathcal{E}$) on
$\,M}%{{\mathrm{M}}\,$ is $\,\pi${\medit-pro\-ject\-a\-ble\/} if
\begin{equation}\label{prj}
d\pi\hskip-1.5pt_x^{\phantom i} w_x^{\phantom i}=u_{\pi(x)}^{\phantom i}\hskip12pt\mathrm{or,\
respectively,}\hskip8ptd\pi\hskip-1.5pt_x^{\phantom i}(\mathcal{E}\hskip-1.5pt_x^{\phantom i})=\mathcal{H}_{\pi(x)}^{\phantom i}
\end{equation}
for some vector field $\,u\,$ (or, some distribution $\,\mathcal{H}$) on
$\,B\,$ and all $\,x\inM}%{{\mathrm{M}}\hskip-.7pt$.
\begin{remark}\label{liebr}Let $\,\pi:M}%{{\mathrm{M}}\toB\,$ be a bundle projection. A
vector field $\,w\,$ on $\,M}%{{\mathrm{M}}\,$ is $\,\pi$-pro\-ject\-a\-ble if and only if,
for every section $\,v\,$ of the vertical distribution
$\,\mathcal{V}=\,\mathrm{Ker}\hskip2.3ptd\pi$, the Lie bracket $\,[v,w]\,$ is
also a section of $\,\mathcal{V}\hskip-.7pt$. This is easily verified in local
coordinates for $\,M}%{{\mathrm{M}}\,$ that make $\,\pi\,$ appear as a
Car\-te\-sian-prod\-uct projection.
\end{remark}
\begin{remark}\label{cmmut}For $\,\pi,M}%{{\mathrm{M}},B,\mathcal{V}\,$ as in
Remark~\ref{liebr}, a $\,\pi$-pro\-ject\-a\-ble vector field $\,w\,$ on
$\,M}%{{\mathrm{M}}\hskip-.7pt$, and $\,x\inM}%{{\mathrm{M}}\,$ such that $\,w_x^{\phantom i}\hskip-.7pt\ne0$, every prescribed
value $\,u_x^{\phantom i}\hskip-.7pt\in\mathcal{V}_{\hskip-.7pt x}^{\phantom i}$ is realized by a local section
$\,u\,$ of $\,\mathcal{V}\,$ commuting with $\,w$. Namely, we may first
prescribe such $\,u\,$ along a fixed co\-di\-men\-sion-one sub\-man\-i\-fold
containing $\,x$, which is transverse to $\,w\,$ at $\,x$, and then use the
local flow of $\,w\,$ to spread $\,u\,$ over a neghborhood of $\,x$.
\end{remark}
\begin{remark}\label{prjct}Given a vector field $\,v\,$ and a distribution
$\,\mathcal{E}\,$ on a manifold, the local flow $\,t\mapsto e^{tv}$ of $\,v\,$
preserves $\,\mathcal{E}\,$ if and only if, whenever $\,w\,$ is a local
section of $\,\mathcal{E}\hskip-.7pt$, so is $\,[v,w]$. Namely,
$\,[v,w]=\pounds\hskip-.5pt_v^{\phantom i} w$, while, denoting by
$\,\varTheta\mapsto (de^{tv})\varTheta\,$ the push-for\-ward action of
$\,e^{tv}$ on tensor fields $\,\varTheta\,$ of any type, we have
\begin{equation}\label{ddt}
d\hskip.7pt[(de^{tv})\varTheta]\hskip.4pt/dt\,
=\,-\hskip.7pt(de^{tv})\pounds\hskip-.5pt_v^{\phantom i}\varTheta\hskip.4pt.
\end{equation}
In fact, when $\,t=0$, (\ref{ddt}) is just the definition of
$\,\pounds\hskip-.5pt_v^{\phantom i} \varTheta$, while, for arbitrary $\,t,$ it follows
from the group\hskip.4pt-homo\-mor\-phic property of $\,t\mapsto e^{tv}\hskip-.7pt$.
\end{remark}
\begin{remark}\label{pralg}We say that a vector field $\,w\,$ (or, a
distribution $\,\mathcal{H}$) on a manifold $\,M}%{{\mathrm{M}}$ is {\medit
pro\-ject\-able along an in\-te\-gra\-ble distribution\/} $\,\mathcal{V}\,$ on
$\,M}%{{\mathrm{M}}\hskip-.7pt$, or $\,\mathcal{V}\hskip-.7pt${\medit-pro\-ject\-a\-ble}, if it is
$\,\pi$-pro\-ject\-a\-ble, as in (\ref{prj}), when restricted to any open
sub\-man\-i\-fold of $\,M}%{{\mathrm{M}}\,$ on which $\,\mathcal{V}$ forms the vertical
distribution $\,\mathrm{Ker}\hskip2.3ptd\pi\,$ of a bundle projection $\,\pi$.
For $\,w\,$ this amounts to invariance of $\,\mathcal{V}\,$ under the local
flow of $\,w$, cf.\ Remarks~\ref{liebr} and~\ref{prjct}.
\end{remark}
\begin{remark}\label{tglvs}For an in\-te\-gra\-ble distribution $\,\mathcal{Z}\,$ on a
Riemannian manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$, the following two conditions are
equivalent.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(i)}] $d_w^{\phantom i}[\hskip.7pt g(\hskip-.4pt v,v)]=0\,$ for all local sections $\,v\,$ of
$\,\mathcal{Z}\,$ and $\,w\,$ of $\,\mathcal{Z}^\perp$ such that $\,w\,$ is nonzero,
$\,\mathcal{Z}$-pro\-ject\-a\-ble, and $\,[v,w]=0$.
\item[{\rm(ii)}] Every leaf (maximal integral manifold) of $\,\mathcal{Z}\,$ is
totally geodesic in $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$.
\end{enumerate}
In fact, let $\,b\,$ be the second fundamental form of the leaves of $\,\mathcal{Z}$,
with $\,b(\hskip-.4pt v,v)\,$ equal to the $\,\mathcal{Z}^\perp$ component of
$\,\nabla\hskip-3pt_{v}^{\phantom i} v$. If $\,v\,$ and $\,w\,$ commute,
$\,d_w^{\phantom i}[\hskip.7pt g(\hskip-.4pt v,v)]=2\hskip.4pt g(\nabla\hskip-3pt_{w}^{\phantom i} v,v)
=2\hskip.4pt g(\nabla\hskip-3pt_{v}^{\phantom i} w,v)=-\nnh2\hskip.4pt g(\nabla\hskip-3pt_{v}^{\phantom i} v,w)
=-\nnh2\hskip.4pt g(b(\hskip-.4pt v,v),w)$, while $\,b\,$ is symmetric and $\,v,w\,$ as in
(i) realize, at any $\,x\inM}%{{\mathrm{M}}\hskip-.7pt$, any given elements of $\,\mathcal{Z}_x^{\phantom i}$
and $\,\mathcal{Z}^\perp_x$ (see Remark~\ref{cmmut}).
\end{remark}
\begin{remark}\label{frthr}Clearly, (ii) in Remark~\ref{tglvs} also follows
when $\,d_w^{\phantom i}[\hskip.7pt g(\hskip-.4pt v,v)]=0\,$ for all $\,v,w\,$ satisfying specific
further conditions besides (i), as long as the last line of
Remark~\ref{tglvs} still applies.
\end{remark}
\begin{lemma}\label{intgr}{\medit
For two in\-te\-gra\-ble distributions\/ $\,\mathcal{E}^\pm$ on a manifold\/
$\,M}%{{\mathrm{M}}\,$ such that the span\/ $\,\mathcal{E}\,$ of\/
$\,\mathcal{E}^+\hskip-1.5pt$ and\/ $\,\mathcal{E}^-\hskip-.7pt$ has constant dimension, the
following conditions are all equivalent.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] $\mathcal{E}\,\,$ is in\-te\-gra\-ble.
\item[{\rm(b)}] $\mathcal{E}^+$ is pro\-ject\-able along\/ $\,\mathcal{E}^-\hskip-1.5pt$.
\item[{\rm(c)}] $\mathcal{E}^-$ is pro\-ject\-able along\/ $\,\mathcal{E}^+\hskip-1.5pt$.
\end{enumerate}
If\/ {\rm(a)} -- {\rm(c)} hold, the distributions that\/ $\,\mathcal{E}^\pm$
locally project onto are in\-te\-gra\-ble as well.
}
\end{lemma}
\begin{proof}We may assume that $\,\mathcal{E}^+$ is the vertical distribution
of a bundle projection $\,\pi:M}%{{\mathrm{M}}\toB\,$ with connected fibres. First, let
$\,\mathcal{E}\,\,$ be in\-te\-gra\-ble. Since $\,\mathcal{E}\,$ contains
$\,\mathcal{E}^+\hskip-1.5pt=\,\mathrm{Ker}\hskip2.3ptd\pi$, its leaves are unions of
fibres and so their $\,\pi$-im\-ages form a foliation of $\,B$, tangent to a
distribution $\,\mathcal{H}\,$ satisfying (\ref{prj}), which proves
pro\-ject\-abil\-i\-ty of $\,\mathcal{E}\hskip-.7pt$, and hence of
$\,\mathcal{E}^-\hskip-.7pt$, along $\,\mathcal{E}^+\hskip-.7pt$. In other words, (a) implies
(c). Conversely, assuming (c), we obtain
$\,d\pi\hskip-1.5pt_x^{\phantom i}(\mathcal{E}\hskip-1.5pt_x^{\phantom i})=d\pi\hskip-1.5pt_x^{\phantom i}(\mathcal{E}\hskip-1.5pt_x^{\hskip.7pt-})
=\mathcal{H}_{\pi(x)}^{\phantom i}$ for all $\,x\inM}%{{\mathrm{M}}\,$ and some distribution
$\,\mathcal{H}$ on $\,B$, which is necessarily in\-te\-gra\-ble: its leaves
are $\,\pi$-im\-ages of the leaves of $\,\mathcal{E}^-\hskip-.7pt$. In\-te\-gra\-bil\-i\-ty of
$\,\mathcal{E}\,$ now follows, as its leaves are the $\,\pi$-pre\-im\-ages of
those of $\,\mathcal{H}$.
Finally, as (a) involves $\,\mathcal{E}^+$ and $\,\mathcal{E}^-$ symmetrically, it is also
equivalent to (b).
\end{proof}
\section{K\"ah\-ler manifolds}\label{km}
\setcounter{equation}{0}
For K\"ah\-ler manifolds we use symbols such as $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$, where $\,M}%{{\mathrm{M}}\,$
stands for the underlying complex manifold. Generally, in complex manifolds,
\begin{equation}\label{jcs}
J\,\mathrm{\ always\ denotes\ the\ com\-plex}\hyp\mathrm{struc\-ture\ tensor.}
\end{equation}
Let $\,v\,$ be a vector field on a K\"ah\-ler manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$. Since
$\,\nabla\hskip-1.5pt J=0$, one has
\begin{equation}\label{ajs}
A\,=\,JS\,\,\mathrm{\ if\ one\ sets\ }\,S=\nabla\hskip-.7pt v\,\mathrm{\ and\
}\,A=\nabla\hskip-.7pt u\hskip.4pt,\mathrm{\ for\ }\,u=J\hskip-.7pt v\hskip.4pt,
\end{equation}
$J,S,A\,$ being viewed as bundle morphisms $\,T\nnhM}%{{\mathrm{M}}\to T\nnhM}%{{\mathrm{M}}\hskip-.7pt$, cf.\
(\ref{nwt}). For the curvature tensor $\,R\,$ of a K\"ah\-ler
manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)\,$ and any vector fields $\,u,v\,$ on $\,M}%{{\mathrm{M}}\hskip-.7pt$,
\begin{equation}\label{rcm}
R(\hskip-.4pt u,v)=R(Ju,J\hskip-.7pt v):T\nnhM}%{{\mathrm{M}}\to T\nnhM}%{{\mathrm{M}}\hskip8pt\mathrm{and}\hskip8ptJ:T\nnhM}%{{\mathrm{M}}
\to T\nnhM}%{{\mathrm{M}}\hskip8pt\mathrm{commute.}
\end{equation}
In fact, the condition $\,\nabla\hskip-1.5pt J=0\,$ turns $\,\nabla\,$ into a
connection in $\,T\nnhM}%{{\mathrm{M}}\,$ treated as a {\medit complex\/} vector bundle,
$\,g\,$ being the real part of a $\,\nabla\hskip-1.5pt$-par\-al\-lel Her\-mit\-i\-an
fibre metric, that is,
\begin{equation}\label{gjw}
g(J\hskip-.7pt w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)\,=\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)
\end{equation}
for all vector fields $\,w,w\hskip.4pt'$ on $\,M}%{{\mathrm{M}}\hskip-.7pt$, and so the curvature
operators $\,R(\hskip-.4pt u,v)\,$ are all com\-plex-lin\-e\-ar and
skew-Her\-mit\-i\-an. The former property now amounts to commutation in
(\ref{rcm}), the latter to the equality
$\,g(R(\hskip-.4pt u,v)\hskip.4pt w,w\hskip.4pt'\hskip.4pt)
=g(R(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt u,v)=g(R(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt Ju,J\hskip-.7pt v)
=g(R(Ju,J\hskip-.7pt v)\hskip.4pt w,w\hskip.4pt'\hskip.4pt)$, with any vector fields $\,w,w\hskip.4pt'\hskip-.7pt$.
Real-hol\-o\-mor\-phic vector fields $\,v\,$ on K\"ah\-ler manifolds will
always be briefly referred to as {\medit hol\-o\-mor\-phic}. Since they are
characterized by $\,\pounds\hskip-.5pt_v^{\phantom i} J=0$, formula (\ref{lie}) for
$\,B}%{{\mathrm{B}}=J\,$ implies that, given a vector field $\,v\,$ on a K\"ah\-ler
manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$,
\begin{equation}\label{hol}
v\,\mathrm{\ is\ holomorphic\ if\ and\ only\ if\
}\,\hsS=\nabla\hskip-.7pt v\,\mathrm{\ commutes\ with\ }\,J\hskip.4pt,
\end{equation}
where $\,J,S:T\nnhM}%{{\mathrm{M}}\to T\nnhM}%{{\mathrm{M}}\,$ as in (\ref{nwt}). For any
hol\-o\-mor\-phic vector field $\,v$
\begin{equation}\label{kil}
\begin{array}{l}
J\hskip-.7pt v\hskip.7pt\mathrm{\ must\ be\ hol\-o\-mor\-phic\ as\ well,\ while\
}\hskip.4pt v\hskip.4pt\mathrm{\ is\ locally}\\
\mathrm{a\ gradient\ if\ and\ only\ if\ }\,\,u\hskip.7pt=J\hskip-.7pt v\,\,\mathrm{\ is\ a\
Kil\-ling\ field.}
\end{array}
\end{equation}
In fact, for $\,S=\nabla\hskip-.7pt v\,$ and $\,A=\nabla\hskip-.7pt u$, (\ref{ajs}) --
(\ref{hol}) give $\,A=JS=SJ$, and so $\,A+A\hskip-.7pt^*\hskip-.7pt=J(S-S^*)$, while the
lo\-cal-gra\-di\-ent property of $\,v\,$ amounts to $\,S-S^*\hskip-.7pt=0$, and the
Kil\-ling condition for $\,u\,$ reads $\,A+\nnhA\hskip-.7pt^*\hskip-.7pt=0$.
\begin{remark}\label{kilxp}As shown by Kobayashi \cite{kobayashi-fp}, if
$\,u\,$ is a Kil\-ling vector field on a Riemannian manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$,
the connected components of the zero set of $\,u\,$ are mutually isolated
totally geodesic sub\-man\-i\-folds of even co\-di\-men\-sions.
\end{remark}
\begin{lemma}\label{nontr}{\medit
If a complex manifold\/ $\,M}%{{\mathrm{M}}\,$ admits a
K\"ah\-ler metric\/ $\,g$, with the K\"ah\-ler form\/
$\,\omega=g(J\,\cdot\,,\,\cdot\,)$, and\/ $\,\varepsilon:\bbC\mathrm{P}^k\hskip-.7pt\toM}%{{\mathrm{M}}\,$ is a
nonconstant hol\-o\-mor\-phic mapping, then\/ $\,\varepsilon\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt\omega$ represents a
nonzero de Rham co\-ho\-mol\-o\-gy class in\/
$\,H^{2\hskip-.7pt}(\hskip-.4pt\bbC\mathrm{P}^k\hskip-1.5pt,\mathrm{I\!R}}%{\mathbf{R})$.
Whether a hol\-o\-mor\-phic mapping\/ $\,\varepsilon:\bbC\mathrm{P}^k\hskip-.7pt\toM}%{{\mathrm{M}}\,$ is constant,
or not, the same is the case for all hol\-o\-mor\-phic mappings\/
$\,\bbC\mathrm{P}^k\hskip-.7pt\toM}%{{\mathrm{M}}\,$ sufficiently close to\/ $\,\varepsilon\,$ in the\/
$\,C^0$ topology.
}
\end{lemma}
\begin{proof}Clearly, $\,\varepsilon\,$ remains nonconstant (and hol\-o\-mor\-phic)
when restricted to a suitable projective line $\,\bbC\mathrm{P}^1\hskip-.7pt\subseteq\bbC\mathrm{P}^k$.
In addition to being positive sem\-i\-def\-i\-nite everywhere, the restriction
$\,h\,$ of $\,\varepsilon\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt g\,$ to $\,\bbC\mathrm{P}^1$ must also be positive definite
somewhere (or else $\,h$, being Her\-mit\-i\-an, would vanish identically,
making $\,\varepsilon\,$ constant on $\,\bbC\mathrm{P}^1$). The integral of $\,\varepsilon\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt\omega\,$
over $\,\bbC\mathrm{P}^1$ is thus positive, proving our first claim. The second one
follows since nearby continuous mappings are, obviously, hom\-o\-top\-ic to
$\,\varepsilon$.
\end{proof}
\begin{remark}\label{fibra}We need the following well-known fact, valid
both in the $\,C^\infty$ and complex (hol\-o\-mor\-phic) categories: any
in\-te\-gra\-ble distribution with compact simply connected leaves constitutes
the vertical distribution of a bundle projection.
The required local trivializations are provided by the -- necessarily trivial
-- holonomy of the underlying foliation; see, for instance,
\cite[p.\ 71]{camacho-lins-neto}.
\end{remark}
\begin{remark}\label{known}We need two more well-known facts; cf.\
\cite[Example 1 of Sect.\ 2.2]{shokurov}.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] A continuous function $\,\,U\to\bbC\,$ on an open set
$\,\,U\subseteq\bbC$, hol\-o\-mor\-phic on $\,\,U\hskip-.7pt\smallsetminus\varLambda$,
where $\,\varLambda\subseteq\hskip.7pt U\,$ is discrete, is necessarily
hol\-o\-mor\-phic everywhere in $\,\,U\hskip-.7pt$.
\item[{\rm(b)}] The only injective hol\-o\-mor\-phic mappings
$\,\bbC\mathrm{P}^1\hskip-1.5pt\to\bbC\mathrm{P}^1$ are bi\-hol\-o\-mor\-phisms.
\end{enumerate}
\end{remark}
\begin{remark}\label{hlext}If $\,\varPsi:\varPi\toM}%{{\mathrm{M}}\,$ is a continuous
mapping between complex manifolds, and a co\-di\-men\-sion-one complex
sub\-man\-i\-fold $\,\varLambda\,$ of $\,\varPi\hskip-.7pt$, closed as a subset of
$\,\varPi\hskip-.7pt$, has the property that the restrictions of $\,\varPsi\,$ to
$\,\varPi\,$ and to the complement $\,\varPi\hskip-.7pt\smallsetminus\varLambda\,$ are
both hol\-o\-mor\-phic, then $\,\varPsi\,$ is hol\-o\-mor\-phic on
$\,\varPi\hskip-.7pt$.
In fact, let $\,p=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\varPi\hskip-.7pt$. When $\,p=1$, our claim is obvious from
Remark~\ref{known}(a). Generally, in local hol\-o\-mor\-phic coordinates
$\,z^1\hskip-.7pt,\dots,z^p$ for $\,\varPi\,$ such that $\,z^2\hskip-.7pt=\ldots=z^p\hskip-.7pt=0\,$
on the intersection of $\,\varLambda\,$ with the coordinate domain, the
complex partial derivatives of the components of $\,\varPsi\,$ (relative to any
local hol\-o\-mor\-phic coordinates in $\,M}%{{\mathrm{M}}$) all clearly exist: for
$\,\partial/\partial z^1$ this follows from the case $\,p=1$.
\end{remark}
\begin{remark}\label{posit}As usual, we call a differential $\,2$-form
$\,\omega\hskip.4pt$ on a complex manifold {\medit positive\/} if it equals the
K\"ah\-ler form $\,g(J\,\cdot\,,\,\cdot\,)\,$ of some K\"ah\-ler metric
$\,g$. This amounts to requiring closedness of $\,\omega\hskip.4pt$ along with
symmetry and positive definiteness of the twice-co\-var\-i\-ant tensor field
$\,-\hskip.7pt\omega\hskip.4pt(J\,\cdot\,,\,\cdot\,)$.
In any complex manifold, $\,d\hskip.4pt\omega=0\,$ and
$\,\omega\hskip.4pt(J\,\cdot\,,\,\cdot\,)\,$ symmetric whenever
$\,\omega=i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf$ or, equivalently,
$\,2\hskip.7pt\omega=-\hskip.7pt d\hskip1pt[\hskip-.4pt J^*\hskip-1.5pt d\hskip-1.2ptf]\,$ for a
real-val\-ued function $\,f\hskip-.7pt$, with the $\,1$-form
$\,J^*\hskip-1.5pt d\hskip-1.2ptf\hskip-.7pt$, also denoted by $\,(d\hskip-1.2ptf)\hskip-.4pt J$,
which sends any tangent vector field $\,v\,$ to $\,d\hskip-1.5pt_{J\hskip-.7pt v}^{\phantom i} f\hskip-1.5pt$.
Clearly,
\begin{equation}\label{idd}
2\hskip.4pt i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf\,
=\,\hs2\hskip.4pt i\hskip-.7pt f'\hskip-.4pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5pt\chi\,
-\,f''\hskip-.7pt d\chi\hskip-.7pt\wedge\hskip-.7pt J^*\hskip-1.5pt d\chi\hskip.4pt,\hskip12pt\mathrm{with}
\hskip6ptf'\hskip.7pt=\,d\hskip-1.2ptf/d\chi\hskip.4pt,
\end{equation}
if $\,f\,$ is a $\,C^\infty$ function of a function $\,\chi\,$ on the same
manifold (cf.\ Remark~\ref{cifot}). The ex\-te\-ri\-or-de\-riv\-a\-tive and
ex\-te\-ri\-or-prod\-uct conventions used here, for any $\,1$-forms
$\,\iota,\kappa\,$ and vector fields $\,u,v$, are
$\,(d\hskip.4pt\kappa)(u,v)=d_u[\kappa(v)]-d_v[\kappa(u)]-\kappa([u,v])\,$ and
$\,(\iota\wedge\kappa)(u,v)=\iota(u)\kappa(v)-\iota(v)\kappa(u)$. When, in
addition, $\,v\,$ is real-hol\-o\-mor\-phic, one has
\begin{equation}\label{ojv}
2\hskip.4pt\omega\hskip.4pt(J\hskip-.7pt v,\,\cdot\,)\,=\,-\hskip.4pt d(d_v^{\phantom i} f)\,
-\,J^*\hskip-1.5pt[d(d\hskip-.7pt_{J\hskip-.4pt v}^{\phantom i} f)\hskip-.7pt]\hskip10pt\mathrm{for}\hskip7pt\omega
=i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf.
\end{equation}
See \cite[Lemma 2]{derdzinski-bc}; the K\"ah\-ler metric used in
\cite{derdzinski-bc} always exists locally.
\end{remark}
\begin{remark}\label{ddnsq}For the real part $\,\langle\,,\rangle\,$ of a Her\-mit\-i\-an
inner product in a fi\-\hbox{nite\hskip.4pt-}\hskip0ptdi\-men\-sion\-al complex
vector space $\,\mathcal{N}\hskip-1.5pt$, let $\,\rho:\mathcal{N}\to[\hs0,\infty)\,$ and
$\,\mathcal{V}\,$ be the {\medit norm function\/} and {\medit complex
radial distribution\/} on $\,\mathcal{N}\smallsetminus\{0\}$, so that
$\,\rho(\xi)=\langle \xi,\xi\rangle^{\hs1\hskip-.7pt/2}$ and
$\,\mathcal{V}\hskip-1.5pt_\xi^{\phantom i}\hskip-.7pt=\mathrm{Span}_\bbC^{\phantom i}(\xi)$.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] $d\hskip-.7pt\rho^2$ is obviously given by
$\,\xi\mapsto2\langle\xi,\,\cdot\,\rangle$.
\item[{\rm(b)}] $i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-.5pt\rho^2$ coincides with
twice the K\"ah\-ler form $\,\langle J\,\cdot\,,\,\cdot\,\rangle\,$ of the
constant metric $\,\langle\,,\rangle$.
\item[{\rm(c)}] $d\hskip-.7pt\rho^2\wedge J^*\hskip-1.5pt d\hskip-.7pt\rho^2\hskip-.7pt$, on
$\,\mathcal{N}\smallsetminus\{0\}$, equals $\,-4\rho^2$ times the restriction of
$\,\langle J\,\cdot\,,\,\cdot\,\rangle\,$ to $\,\mathcal{V}\hskip-.7pt$.
\end{enumerate}
In fact, (b) -- (c) are immediate from (a) and Remark~\ref{posit}.
\end{remark}
\begin{remark}\label{cptfb}Let $\,\pi:M}%{{\mathrm{M}}\toB\,$ be a surjective submersion
between manifolds.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] If the pre\-im\-a\-ges $\,\pi^{-\nnh1}(y)$, $\,y\inB$, are
all compact, then $\,\pi\,$ can be factored as $\,M}%{{\mathrm{M}}\to\varPi\toB$, with
a bundle projection $\,M}%{{\mathrm{M}}\to\varPi\,$ having compact (connected) fibres, and
a finite covering projection $\,\varPi\toB$.
\item[{\rm(b)}] In the case where $\,\dimB=\dim\varPi\,$ and $\,M}%{{\mathrm{M}}\,$ is
compact, $\,\pi\,$ must necessarily be a (finite) covering projection.
\end{enumerate}
Namely, (a) is a well-known fact, easily verified using parallel transports
corresponding to a fixed vector sub\-bun\-dle $\,\mathcal{H}\,$ of
$\,T\nnhM}%{{\mathrm{M}}\,$ for which $\,T\nnhM}%{{\mathrm{M}}=\mathcal{V}\hskip-.7pt\oplus\hskip-1.5pt\mathcal{H}\,$
cf.\ \cite[Remark 1.1]{derdzinski-roter-08}, or derived as in
Remark~\ref{fibra}, since the foliation with the leaves $\,\pi^{-\nnh1}(y)\,$
has trivial holonomy. Part (b) -- in which the dimension equality means
that $\,\pi\,$ is lo\-cal\-ly-dif\-feo\-mor\-phic -- easily follows from (a).
\end{remark}
\begin{remark}\label{ftcvr}For a K\"ah\-ler manifold $\,(\varPi\hskip-.7pt,h)\,$ with
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\varPi=l$, any hol\-o\-mor\-phic mapping
$\,F:\bbC\mathrm{P}\hskip.4pt^l\hskip-.7pt\to\varPi\,$ such that $\,F^*\hskip-.7pt h\,$ is a positive constant
multiple of the Fu\-bi\-ni-Stu\-dy metric on $\,\bbC\mathrm{P}\hskip.4pt^l$ (cf.\
Remark~\ref{fbstm}) must be a bi\-hol\-o\-mor\-phism.
In fact, $\,F\,$ is then a covering projection (Remark~\ref{cptfb}(b)) and our
claim follows since, due to a result of Kobayashi \cite{kobayashi-ck},
$\,\varPi\,$ has to be simply connected.
\end{remark}
\section{Ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler
triples}\label{gk}
\setcounter{equation}{0}
Given a manifold $\,M}%{{\mathrm{M}}\,$ endowed with a fixed connection $\,\nabla\hskip-1.5pt$, we
refer to a vector field $\,v\,$ on $\,M}%{{\mathrm{M}}\,$ as {\medit geodesic\/} if the
integral curves of $\,v\,$ are re\-pa\-ram\-e\-trized
$\,\nabla\hskip-.7pt$-ge\-o\-des\-ics. Equivalently, for some function $\,\psi\,$ on
the open set $\,M}%{{\mathrm{M}}'\hskip-.7pt\subseteqM}%{{\mathrm{M}}\,$ on which $\,v\ne0$,
\begin{equation}\label{nvv}
\nabla\hskip-3pt_{v}^{\phantom i} v\,=\,\psi\hskip.4ptv\qquad\mathrm{everywhere\
in}\hskip7ptM}%{{\mathrm{M}}'.
\end{equation}
A function $\,\vt\,$ on a Riemannian manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)\,$ is said to
{\medit have a geodesic gradient\/} if its gradient $\,v\,$ is a geodesic
vector field relative to the Le\-vi-Ci\-vi\-ta connection $\,\nabla\hskip-1.5pt$.
Functions with geodesic gradients are also called {\it trans\-nor\-mal\/}
\cite{wang,miyaoka,bolton}.
\begin{lemma}\label{ggqft}{\medit
For a function\/ $\,\vt\,$ on a Riemannian manifold\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$, the
gradient of\/ $\,\vt\,$ is a geodesic vector field if and only if\/
$\,Q=g(\nabla\hskip-.5pt\vt,\nabla\hskip-.5pt\vt)\,$ is, locally in\/ $\,M}%{{\mathrm{M}}'\hskip-1.5pt$, a function of\/ $\,\vt$.
}
\end{lemma}
\begin{proof}By (\ref{tnd}), condition (\ref{nvv}) is equivalent to
$\,dQ\wedge\hskip.7pt d\vt=0$.
\end{proof}
\begin{definition}\label{ggktr}A {\medit ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple\/} $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ consists of any K\"ah\-ler manifold
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g)\,$ and a nonconstant real-val\-ued function $\,\vt\,$ on $\,M}%{{\mathrm{M}}\,$
such that the $\,g$-gra\-di\-ent $\,v=\nabla\hskip-.5pt\vt\,$ is both geodesic and
real-hol\-o\-mor\-phic.
Speaking of {\medit compactness\/} of $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, or its {\medit
dimension}, we always mean those of the underlying complex manifold
$\,M}%{{\mathrm{M}}\hskip-.7pt$, and we call two such triples
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt),\,(\hatM}%{{\mathrm{M}},\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt)$ {\medit iso\-mor\-phic\/} if
$\,\vt=\hat\vt\circ\varPhi\,$ and $\,g=\varPhi^*\hskip-1.5pt\hat{g\hskip2pt}\hskip-1.3pt\,$ for some
bi\-hol\-o\-mor\-phism $\,\varPhi:M}%{{\mathrm{M}}\hskip-.4pt\to\hatM}%{{\mathrm{M}}\hskip-.7pt$.
\end{definition}
For $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ as above, whenever the extrema of $\,\vt\,$ exist, we
will also consider
\begin{equation}\label{tmm}
\mathrm{the\ }\,\vt\hyp\mathrm{pre\-im\-ages\ }\,\varSigma^+\hskip-.7pt\mathrm{\ and\
}\,\varSigma^-\hskip-.7pt\mathrm{\ of\ \ }\,\vt\hskip-1.8pt_+\w\hskip-.7pt=\hskip.7pt\mathrm{max}\hskip2.7pt\vt\,
\mathrm{\ \ and\ \ }\,\vt\hskip-1.7pt_-\w\hskip-.7pt=\hskip.7pt\mathrm{min}\hskip2.7pt\vt\hskip.4pt.
\end{equation}
\begin{remark}\label{trivl}A ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent
K\"ah\-ler triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ can be {\medit trivially modified\/} to
yield $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,p\vt+q)$, with any real constants $\,p\ne0\,$ and $\,q$.
(Clearly, $\,\varSigma^\pm$ in (\ref{tmm}) then become switched if $\,p<0$.) Any
such $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ and any complex sub\-man\-i\-fold $\,\varPi$ of
$\,M}%{{\mathrm{M}}\hskip-.7pt$, tangent to $\,v=\nabla\hskip-.5pt\vt\,$ (that is, forming a union of integral
curves of $\,v$), and not contained in a single level set of $\,\vt$,
give rise (cf.\ Remark~\ref{gpgrd}) to the new
ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triple
$\,(\varPi,g'\hskip-.7pt,\vt')$, where $\,g'\hskip-.7pt,\vt'$ are the restrictions of $\,g\,$
and $\,\vt\,$ to $\,\varPi\hskip-.7pt$.
\end{remark}
As shown next, ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triples
naturally arise from suitable co\-ho\-mo\-ge\-ne\-i\-ty-one isometry groups.
\begin{lemma}\label{chone}{\medit
Let a connected Lie group\/ $\,G\hskip.7pt$ acting by hol\-o\-mor\-phic isometries on
a K\"ah\-ler manifold\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)$, and having some orbits of real
co\-di\-men\-sion\/ $\,1$, preserve a nontrivial hol\-o\-mor\-phic Kil\-ling
field\/ $\,u\,$ with zeros. If\/ $\,H^1\hskip-.7pt(M}%{{\mathrm{M}},\mathrm{I\!R}}%{\mathbf{R})=\{0\}$, then\/
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is a ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple and\/
$\,u=J(\nabla\hskip-.5pt\vt)\,\,$ for some\/ $\,G$-in\-var\-i\-ant function\/ $\,\vt\,$ on\/
$\,M}%{{\mathrm{M}}\hskip-.7pt$.
}
\end{lemma}
\begin{proof}Since $\,H^1\hskip-.7pt(M}%{{\mathrm{M}},\mathrm{I\!R}}%{\mathbf{R})=\{0\}$, (\ref{kil}) implies both the
existence of a function $\,\vt$ with $\,u=J(\nabla\hskip-.5pt\vt)$, and the fact that its
gradient $\,v=\nabla\hskip-.5pt\vt=-\hskip-.7pt Ju\,$ is hol\-o\-mor\-phic. Thus, elements of $\,G\,$
preserve $\,\vt\,$ up to additive constants. Let $\,\varSigma\,$ now be a fixed
connected component of the zero set of $\,u$, so that $\,G$, being connected,
leaves $\,\varSigma\,$ invariant, while $\,\vt\,$ is constant on $\,\varSigma\,$ (cf.\
Remark~\ref{kilxp}). The additive constants just mentioned are therefore equal
to $\,0$. Due to their $\,G$-in\-var\-i\-ance, the functions $\,\vt\,$ and
$\,Q=g(\nabla\hskip-.5pt\vt,\nabla\hskip-.5pt\vt)$ are constant along co\-di\-men\-sion-one orbits of
$\,G\,$ and, consequently, functionally dependent. (Note that the union of
such orbits is dense in $\,M}%{{\mathrm{M}}\hskip-.7pt$.) Consequently, by Lemma~\ref{ggqft}, the
gradient $\,v=\nabla\hskip-.5pt\vt\,$ is a geodesic vector field.
\end{proof}
\begin{remark}\label{reddt}The assumptions about triviality of
$\,H^1\hskip-.7pt(M}%{{\mathrm{M}},\mathrm{I\!R}}%{\mathbf{R})\,$ and hol\-o\-mor\-phic\-i\-ty of $\,u\,$ in
Lemma~\ref{chone} are well-known to be redundant when $\,M}%{{\mathrm{M}}\,$ is compact
\cite[p.\ 95, Corollary 4.5]{kobayashi-tg}; see also \cite[formula (A.2c) and
Theorem A.1]{derdzinski-kp}.
\end{remark}
The following fact will be used in the proof of Theorem~\ref{jacob}.
\begin{lemma}\label{jcbfl}{\medit
If a vector field\/ $\,w\,$ on a Riemannian manifold\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)\,$ is
orthogonal to a geodesic gradient\/ $\,v\,$ and commutes with\/ $\,v$, then\/
$\,w\,$ is a Ja\-co\-bi field along every integral curve of\/ $\,v/|v|\,$ in
the set\/ $\,M}%{{\mathrm{M}}'$ where\/ $\,v\ne0$.
}
\end{lemma}
\begin{proof}Fix $\,\vt:M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}\,$ with $\,v=\nabla\hskip-.5pt\vt$. For
$\,Q=g(\hskip-.4pt v,v):M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}$, (\ref{tnd}) and (\ref{nvv}) give
$\,dQ\wedge\hskip.7pt d\vt=0$, so that $\,|v|\,$ is, locally in $\,M}%{{\mathrm{M}}'\hskip-1.5pt$, a
$\,C^\infty$ function of $\,\vt\,$ (cf.\ Remark~\ref{cifot}). As
$\,\pounds_w^{\phantom i}\vt=0\,$ due to the orthogonality assumption, and
$\,\pounds_w^{\phantom i} v=0$, we now have $\,\pounds_w^{\phantom i}(\hskip-.4pt v/|v|)=0\,$ on $\,M}%{{\mathrm{M}}'\hskip-1.5pt$. The
local flow of $\,w$, applied to any integral curve of $\,v/|v|$, thus yields a
variation of unit-speed geodesics, and our claim follows.
\end{proof}
\begin{remark}\label{dmone}A compact ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler
triple of complex dimension $\,1\,$ is essentially, up to isomorphisms,
nothing else than the sphere $\,S^2$ with a rotationally invariant metric. In
fact, necessity of rotational invariance is due to (\ref{kil}), while its
sufficiency follows from Lemma~\ref{chone} with $\,G=S^1\hskip-.7pt$, for the sphere
treated as $\,\bbC\mathrm{P}^1$ with a K\"ah\-ler metric. (Once the $\,S^1$ action is
chosen, the required function $\,\vt\,$ becomes unique up to trivial
modifications, cf.\ Remark~\ref{trivl}.)
\end{remark}
\section{Examples: Grass\-mann\-i\-an and CP triples}\label{eg}
\setcounter{equation}{0}
In this section {\medit vector spaces\/} are complex (except for
Remark~\ref{dfrtl}) and fi\-\hbox{nite\hskip.4pt-}\hskip0ptdi\-men\-sion\-al. By
$\,k${\it-planes\/} in a vector space $\,\mathsf{V}\,$ we mean
\hbox{$\,k$-}\hskip0ptdi\-men\-sion\-al vector sub\-spaces of $\,\mathsf{V}\hskip-.7pt$.
When $\,k=1$, they will also be called {\medit lines\/} in $\,\mathsf{V}\hskip-.7pt$.
Given a vector space $\,\mathsf{V}\,$ and $\,k\in\{0,1,\dots,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}\}$, the
Grass\-mann\-i\-an $\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ is the set of all
$\,k$-planes in $\,\mathsf{V}\hskip-.7pt$. Each $\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ naturally
forms a compact complex manifold (see Remark~\ref{grass}), and
$\,\mathrm{P}\mathsf{V}=\mathrm{G}\hskip-.4pt_1^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ is the projective space of
$\,\mathsf{V}\hskip-.7pt$, provided that $\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}>0$. We will use the standard
identification
\begin{equation}\label{inc}
\mathrm{P}(\hskip-.4pt\bbC\times\mathsf{V})\hskip.7pt\,=\,\hskip.7pt\mathsf{V}\,\,\cup\,\hskip.7pt\mathrm{P}\mathsf{V}\hskip.4pt,
\end{equation}
of $\,\mathrm{P}(\hskip-.4pt\bbC\times\mathsf{V})\,$ with the disjoint union of an open
subset bi\-hol\-o\-mor\-phic to $\,\mathsf{V}\,$ and a complex sub\-man\-i\-fold
bi\-hol\-o\-mor\-phic to $\,\mathrm{P}\mathsf{V}\,$ via the bi\-hol\-o\-mor\-phism
sending $\,v\in\mathsf{V}$, or the line $\,\bbC v\,$ spanned by
$\,v\in\mathsf{V}\smallsetminus\{0\}$, to the line $\,\bbC(1,v)\,$ or, respectively,
$\,\bbC(0,v)$. The {\medit projectivization\/} of a hol\-o\-mor\-phic vector
bundle $\,N}%{{\mathrm{N}}\,$ over a complex manifold $\,\varSigma$ is, as usual, the
hol\-o\-mor\-phic bundle $\,\mathrm{P}\nnhN}%{{\mathrm{N}}\,$ of complex projective spaces
over $\,\varSigma\,$ with
\begin{equation}\label{prz}
\mathrm{the\ fibres}\hskip7pt
[\mathrm{P}\nnhN}%{{\mathrm{N}}]_y^{\phantom i}\,=\,\hskip.7pt\mathrm{P}\mathsf{V}\hskip7pt\mathrm{for}\hskip7pt\mathsf{V}\,
=\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\hskip.4pt,\hskip7pt\mathrm{whenever}\hskip7pty\in\varSigma\hskip.4pt.
\end{equation}
For a sub\-space $\,\mathsf{L}\,$ of a vector space $\,\mathsf{V}\,$ such that
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}\ge2$, let $\,G\,$ be the group of all
com\-plex-lin\-e\-ar au\-to\-mor\-phisms of $\,\mathsf{V}\,$ preserving both
$\,\mathsf{L}\,$ and a fixed Her\-mit\-i\-an inner product in $\,\mathsf{V}\hskip-.7pt$. We now
define a compact complex manifold $\,M}%{{\mathrm{M}}\,$ by
\begin{equation}\label{dta}
\begin{array}{rl}
\mathrm{i)}&M}%{{\mathrm{M}}=\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\mathrm{,\ where\
}\,0<k<\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}\,\mathrm{\ and\ }\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\mathsf{L}=1\mathrm{,\ or}\\
\mathrm{ii)}&M}%{{\mathrm{M}}=\mathrm{P}\mathsf{V}\mathrm{,\ allowing\
}\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\mathsf{L}\in\{1,\dots,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}-1\}\,\mathrm{\ to\ be\ arbitrary.}
\end{array}
\end{equation}
Then the hypotheses, and hence conclusions, of Lemma~\ref{chone} are
satisfied by these $\,M}%{{\mathrm{M}}\hskip-.7pt,G$, any $\,G$-in\-var\-i\-ant
K\"ah\-ler metric $\,g\,$ on $\,M}%{{\mathrm{M}}\hskip-.7pt$, and some $\,u$. Specifically,
$\,u\,$ is a vector field arising from the central circle subgroup $\,S^1$ of
$\,G\,$ formed by all unimodular elements of $\,G\,$ acting in both
$\,\mathsf{L},\mathsf{L}\hskip-1.5pt^\perp$ as multiples of $\,\mathrm{Id}$. See the remarks below.
The triples $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ arising via Lemma~\ref{chone} in cases
(\ref{dta}.i) and (\ref{dta}.ii) will from now on be called {\medit
Grass\-mann\-i\-an triples\/} and, respectively, {\medit CP triples}.
Since $\,G\,$ as above contains all unit complex multiples of $\,\mathrm{Id}$,
its action on $\,M}%{{\mathrm{M}}\,$ is not effective. Lemma~\ref{chone}
does not require effectiveness of the action.
\begin{remark}\label{grcpt}The co\-ho\-mo\-ge\-ne\-i\-ty-one assumption of
Lemma~\ref{chone} follows here from the fact that the orbits of $\,G\,$
coincide with the levels of a nonconstant real-an\-a\-lyt\-ic function
$\,f:M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}$. Specifically, in case (\ref{dta}.i),
$\,f(\mathsf{W})=|\hskip.7pt\mathrm{pr}\hskip.4pt(X,\mathsf{W})|^2\hskip-.7pt$, where
\begin{equation}\label{prx}
\mathrm{pr}\hskip.4pt(X,\mathsf{W})\,\mathrm{\ denotes\ the\ orthogonal\ projection\ of\
}\,X\,\mathrm{\ onto\ }\,\mathsf{W},
\end{equation}
and $\,X\,$ is some\hskip.7pt$/$any unit vector spanning $\,\mathsf{L}$, which yields
$\,G$-in\-var\-i\-ance of $\,f\hskip-1.5pt$. Conversely, if
$\,\mathsf{W},\tilde\mathsf{W}\inM}%{{\mathrm{M}}\,$ and $\,f(\mathsf{W})=f(\tilde\mathsf{W})$, an element of $\,G\,$
sending $\,\mathsf{W}\,$ to $\,\tilde\mathsf{W}\,$ is provided by any linear isometry
mapping the quadruple
$\,\mathsf{W},\mathsf{W}^\perp\hskip-1.5pt,\mathrm{pr}\hskip.4pt(X,\mathsf{W}),\mathrm{pr}\hskip.4pt(X,\mathsf{W}^\perp)$
onto $\,\tilde\mathsf{W},\tilde\mathsf{W}^\perp\hskip-1.5pt,\mathrm{pr}\hskip.4pt(X,\tilde\mathsf{W}),
\mathrm{pr}\hskip.4pt(X,\tilde\mathsf{W}^\perp\hskip-1.5pt)$. (Such an isometry will preserve
$\,X=\mathrm{pr}\hskip.4pt(X,\mathsf{W})+\mathrm{pr}\hskip.4pt(X,\mathsf{W}^\perp\hskip-1.5pt)$.) In case
(\ref{dta}.ii) we may use $\,f\,$ given by
$\,f(\mathsf{W})=|\hskip.7pt\mathrm{pr}\hskip.4pt(Y\hskip-3pt_\mathsf{W}^{\phantom i},\mathsf{L})|^2\hskip-.7pt$, with
$\,Y\hskip-3pt_\mathsf{W}^{\phantom i}$ standing for some\hskip.7pt$/$\hskip-.4pt any unit vector that spans
the line $\,\mathsf{W}\hskip-.7pt$.
\end{remark}
\begin{remark}\label{zeros}For a Grass\-mann\-i\-an or CP triple
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, critical points of $\,\vt$, that is, the zeros of $\,u\,$
or, equivalently, the fixed points of the central circle subgroup $\,S^1$
mentioned above, form the disjoint union of two (connected) compact complex
sub\-man\-i\-folds, which -- since $\,\vt\,$ is clearly constant on either of
them -- must be the same as $\,\varSigma^\pm$ in (\ref{tmm}). With $\,\approx\,$
denoting bi\-hol\-o\-mor\-phic equivalence, these $\,\varSigma^\pm$ are
\begin{equation}\label{spm}
\begin{array}{rlll}
\mathrm{a)}&\{\mathsf{W}\inM}%{{\mathrm{M}}:\mathsf{L}\subseteq\mathsf{W}\}\hskip-.7pt
\approx\hskip-.7pt\mathrm{Gr}\hskip-.7pt_{k-\nh1}^{\phantom i}[\hskip.4pt\mathsf{V}\hskip-2pt/\mathsf{L}]\hskip.4pt,\hskip9pt
&\mathrm{b)}&\{\mathsf{W}\inM}%{{\mathrm{M}}:\mathsf{W}\subseteq\mathsf{L}\hskip-2.5pt^\perp\}\hskip-.7pt
\approx\hskip-.7pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt,\\
\mathrm{c)}&\{\mathsf{W}\inM}%{{\mathrm{M}}:\mathsf{W}\subseteq\mathsf{L}\}\hskip-.7pt
\approx\hskip-.7pt\mathrm{P}\mathsf{L}\hskip.4pt
&\mathrm{d)}&\{\mathsf{W}\inM}%{{\mathrm{M}}:\mathsf{W}\subseteq\mathsf{L}\hskip-2.5pt^\perp\}\hskip-.7pt
\approx\hskip-.7pt\mathrm{P}\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt,
\end{array}
\end{equation}
where (\ref{spm}.a) -- (\ref{spm}.b) correspond to (\ref{dta}.i), and
(\ref{spm}.c) -- (\ref{spm}.d) to (\ref{dta}.ii). In fact, each of the four
sets clearly consists of fixed points of $\,S^1\hskip-1.5pt$. Conversely, suppose that
$\,\mathsf{W}\inM}%{{\mathrm{M}}\,$ does not lie in the union of the sets (\ref{spm}.a) --
(\ref{spm}.b) (or, (\ref{spm}.c) -- (\ref{spm}.d)), and
$\,\varXi\in S^1\hskip-.7pt\smallsetminus\{\mathrm{Id},-\hskip-.7pt\mathrm{Id}\}$. Then
$\,\varXi(\mathsf{W})\ne\mathsf{W}\hskip-.7pt$. Namely, if $\,\varXi\,$ preserved $\,\mathsf{W}\hskip-.7pt$, the
equalities $\,\varXi(\mathsf{L})=\mathsf{L}\,$ and
$\,\varXi(\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt)=\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt$, along with
$\,\varXi(\mathsf{W})=\mathsf{W}\,$ and $\,\varXi(\mathsf{W}^\perp\hskip-1.5pt)=\mathsf{W}^\perp\hskip-1.5pt$, would imply
analogous equalities for the lines spanned by
$\,\mathrm{pr}\hskip.4pt(X,\mathsf{W}),\,\mathrm{pr}\hskip.4pt(X,\mathsf{W}^\perp)\,$ (case
(\ref{dta}.i)), or by
$\,\mathrm{pr}\hskip.4pt(X,\mathsf{L}),\,\mathrm{pr}\hskip.4pt(X,\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt)\,$
(case (\ref{dta}.ii)), cf.\ (\ref{prx}), $\,X\,$ being any unit vector in the
line $\,\mathsf{L}$ (or, respectively, in the line $\,\mathsf{W}$). In either case, the
plane $\,\mathsf{K}\,$ spanned by the two $\,\varXi\hskip-.7pt$-in\-var\-i\-ant lines would
contain a third such line, in the form of $\,\mathsf{L}\,$ (or, respectively,
$\,\mathsf{W}$). Thus, $\,\varXi\,$ restricted to $\,\mathsf{K}\,$ would be a multiple of
$\,\mathrm{Id}$, leading to a contradiction: in both cases, $\,\mathsf{K}\,$ contains
nonzero vectors from $\,\mathsf{L}\,$ and from $\,\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt$, which
are eigen\-vec\-tors of $\,\varXi\,$ for two distinct eigen\-values.
\end{remark}
\begin{remark}\label{cpone}All compact ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent
K\"ah\-ler triples of complex dimension $\,1$ are obviously isomorphic to CP
triples constructed from the data (\ref{dta}.ii) with $\,m=2$ and
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\mathsf{L}=1$. See Remark~\ref{dmone}.
\end{remark}
\begin{remark}\label{fbstm}Given the real part $\,\langle\,,\rangle\,$ of a Her\-mit\-i\-an
inner product in a vector space $\,\mathsf{V}\hskip-.7pt$, the {\medit Fu\-bi\-ni-Stu\-dy
metric\/} on $\,\mathrm{P}\mathsf{V}\,$ associated with $\,\langle\,,\rangle\,$ is, as usual,
uniquely characterized by requiring that the restriction of the projection
$\,\xi\mapsto\bbC\xi\,$ to the unit sphere of $\,\langle\,,\rangle\,$ be a Riemannian
submersion. Another such real part $\,\langle\,,\rangle\hskip-.4pt'$ yields the same
Fu\-bi\-ni-Stu\-dy metric as $\,\langle\,,\rangle\,$ only if $\,\langle\,,\rangle\hskip-.4pt'$ is a constant
multiple of $\,\langle\,,\rangle$. In fact, a $\,\bbC$-lin\-e\-ar auto\-mor\-phism of
$\,\mathsf{V}\,$ taking $\,\langle\,,\rangle\,$ to $\,\langle\,,\rangle\hskip-.4pt'$ descends to an isometry
$\,\mathrm{P}\mathsf{V}\to\mathrm{P}\mathsf{V}$ and hence equals a
$\,\langle\,,\rangle$-u\-ni\-tar\-y auto\-mor\-phism of $\,\mathsf{V}\,$ followed by
$\,r\hskip.4pt e^{i\theta}$ times $\,\mathrm{Id}\,$ for some $\,r,\theta\in\mathrm{I\!R}}%{\mathbf{R}$,
which gives $\,\langle\,,\rangle=r\langle\,,\rangle\hskip-.4pt'\hskip-.7pt$.
\end{remark}
\begin{remark}\label{grass}Given a vector space $\,\mathsf{V}\,$ and
$\,k\in\{1,\dots,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}\}$, we denote by $\,\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}$
and $\,\pi:\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\to\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$
the Stie\-fel manifold of all linearly independent ordered $\,k$-tuples of
vectors in $\,\mathsf{V}\,$ (forming an open sub\-man\-i\-folds of the
$\,k\hskip.7pt$th Cartesian power of $\,\mathsf{V}$) and, respectively, the projection
mapping sending each $\,\mathbf{e}\in\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ to
$\,\pi(\mathbf{e})=\mathrm{Span}\hskip.4pt(\mathbf{e})$. Then
$\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ has a unique structure of a compact complex
manifold of complex dimension $\,(n-k)k$, where $\,n=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}\hskip-.7pt$, such
that $\,\pi\,$ is a hol\-o\-mor\-phic submersion.
\end{remark}
\begin{remark}\label{tgvec}We will use the canonical
isomorphic identification
\begin{equation}\label{twg}
T\hskip-2.2pt_\mathsf{W}^{\phantom i}\hskip-.7pt[\hskip.7pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\hskip.7pt]\,\,
=\,\,\mathrm{Hom}\hskip.7pt(\mathsf{W},\mathsf{V}\hskip-1.8pt/\hskip.4pt\mathsf{W})\hskip12pt\mathrm{whenever}
\hskip7ptk\in\{0,1,\dots,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}\}\hskip.4pt,
\end{equation}
for the tangent space of the complex manifold $\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$
at any $\,k$-plane $\,\mathsf{W}\hskip-.7pt$, where $\,\mathrm{Hom}$ means `the space of
linear operators' and $\,\mathsf{V}\,$ is any vector space.
Namely, let $\,\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ and
$\,\pi:\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\to\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ be as in
Remark~\ref{grass}. Under the identification (\ref{twg}),
$\,H\in\mathrm{Hom}\hskip.7pt(\mathsf{W},\mathsf{V}\hskip-1.8pt/\hskip.4pt\mathsf{W})\,$ corresponds to
$\,d\pi_{\mathbf{e}}^{\phantom i}(\tildeH\hskip-.7pt\mathbf{e})
\in T\hskip-3pt_\mathsf{W}^{\phantom i}\hskip-.7pt[\hskip.7pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\hskip.7pt]\,$ for any linear
lift $\,\tildeH:\mathsf{W}\to\mathsf{V}\,$ of $\,H\,$ and any basis
$\,\mathbf{e}\,$ of $\,\mathsf{W}\,$ which, clearly, does not depend on how such
$\,\tildeH\,$ and $\,\mathbf{e}\,$ were chosen.
\end{remark}
\begin{remark}\label{dfrtl}Given a real (or, complex) manifold $\,\,U\hskip.7pt$ and
real (or, complex) vector spaces $\,\mathcal{T}\hskip-1.5pt,\mathcal{Y}\hskip-1.5pt$, let
$\,F:U\to\mathrm{Hom}\hskip.7pt(\mathcal{T}\hskip-1.5pt,\mathcal{Y})\,$ be a $\,C^\infty$ (or,
hol\-o\-mor\-phic) mapping giving rise to a {\medit constant\/} function
$\,U\ni\xi\mapsto\mathrm{rank}\hskip1.7ptF(\xi)\,$ or, equivalently,
leading to the same value of $\,k=\dim\hskip.7pt\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\xi)\,$
for all $\,\xi\in U\hskip-.7pt$. Then the mapping
\begin{equation}\label{ker}
U\ni\xi\,
\mapsto\,\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\xi)\in\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt\mathcal{T}
\end{equation}
is of class $\,C^\infty$ (or, hol\-o\-mor\-phic) and its differential
$\,T\hskip-2.3pt_\xi^{\phantom i} U\to\,\mathrm{Hom}\hskip.7pt(\mathsf{W},\mathcal{T}\hskip-1.8pt/\hskip.4pt\mathsf{W})\,$
at any $\,\xi\in U\hskip-.7pt$, with $\,\mathsf{W}=\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\xi)$, cf.\
(\ref{twg}), sends $\,\eta\in T\hskip-2.3pt_\xi^{\phantom i} U\,$ to the unique
$\,H:\mathsf{W}\to\mathcal{T}\hskip-1.8pt/\hskip.4pt\mathsf{W}$ having a linear lift
$\,\tildeH:\mathsf{W}\to\mathcal{T}\,$ such that
\begin{equation}\label{fxh}
F\hskip-.7pt(\xi)\circ\hskip-.7pt\tildeH\hskip7pt\mathrm{equals\ the\ restriction\
of}\hskip7pt-\hskip-3ptd\hskip-.8ptF\hskip-3pt_\xi^{\phantom i}\eta\hskip7pt\mathrm{to}
\hskip7pt\mathsf{W}.
\end{equation}
Here $\,d\hskip-.8ptF\hskip-3pt_\xi^{\phantom i}:T\hskip-2.3pt_\xi^{\phantom i} U
\to\,\mathrm{Hom}\hskip.7pt(\mathcal{T}\hskip-1.5pt,\mathcal{Y})$, so that $\,F\hskip-.7pt(\xi)\,$ and
$\,d\hskip-.8ptF\hskip-3pt_\xi^{\phantom i}\eta\,$ are linear operators
$\,\mathcal{T}\hskip-.7pt\to\mathcal{Y}\hskip-1.5pt$.
In fact, let `regular' mean $\,C^\infty$ or hol\-o\-mor\-phic. Given
$\,\xi\in U\hskip-.7pt$, we may select a subspace $\,\mathsf{V}\,$ of $\,\mathcal{T}\,$ so that
$\,\mathcal{T}\hskip-.7pt=\mathsf{V}\oplus\mathsf{W}\hskip-.7pt$, where
$\,\mathsf{W}\hskip-.7pt=\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\xi)$. For all $\,\eta\,$ near $\,\xi\,$
in $\,U\hskip-.7pt$, the restriction of $\,F(\eta)\,$ to $\,\mathsf{V}\,$ is clearly an
isomorphism onto the image of $\,F(\eta)$. Denoting by
$\,F^{-\nnh1}_{\hskip-2.2pt\eta}$ its inverse isomorphism, we see that
$\,F^{-\nnh1}_{\hskip-2.2pt\eta}\circ F(\eta)\,$ and $\,\mathrm{pr}\hskip-.4pt_\eta^{\phantom i}
=\mathrm{Id}-F^{-\nnh1}_{\hskip-2.2pt\eta}\circ F(\eta)\,$ coincide with the
di\-rect-sum projections of
$\,\mathcal{T}\hskip-.7pt=\mathsf{V}\oplus\hskip.7pt\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\eta)$ onto $\,\mathsf{V}\,$ and
$\,\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\eta)$. A fixed basis $\,e_1^{\phantom i},\dots,e_k^{\phantom i}$ of
$\,\mathsf{W}\,$ thus gives rise to the basis
$\,\mathrm{pr}\hskip-.4pt_\eta^{\phantom i}\hskip.4pt e_1^{\phantom i},\dots,\mathrm{pr}\hskip-.4pt_\eta^{\phantom i}\hskip.4pt e_k^{\phantom i}$ of any
such $\,\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\eta)$, depending regularly on $\,\eta$,
which constitutes a regular local lift of (\ref{ker}) valued in the Stie\-fel
manifold $\,\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathcal{T}\,$ (see Remark~\ref{grass}), proving
regularity of (\ref{ker}).
Replacing our $\,\xi\,$ with $\,\xi(0)\,$ for a curve
$\,t\mapsto\xi(t)\in U\hskip-.7pt$, and letting $\,e_1^{\phantom i}(0),\dots,e_k^{\phantom i}(0)$ be a fixed
basis of $\,\mathsf{W}\hskip-.7pt=\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\xi(0))$, we set
$\,e_j^{\phantom i}(t)=\mathrm{pr}\hskip-.4pt_{\xi(t)}^{\phantom i}\hskip.4pt[\hskip.4pt e_j^{\phantom i}(0)]\,$ and
$\,\eta(t)=\dot\xi(t)$. Suppressing from the notation the dependence on $\,t$,
one thus gets $\,[F(\xi)]\hskip.4pt e_j^{\phantom i}\hskip-.7pt=0\,$ and, by differentiation,
$\,[\hskip.7pt d\hskip-.8ptF\hskip-3pt_\xi^{\phantom i}\eta\hskip.4pt]\hskip.4pt e_j^{\phantom i}\hskip-.7pt
+[F(\xi)]\hskip.4pt\dot e_j^{\phantom i}\hskip-.7pt=0$. The operators
$\,\tildeP=\tildeP(t):\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\xi)\to\mathsf{V}$ defined by
$\,\tildeP\hskip-.7pt e_j^{\phantom i}\hskip-1.5pt=\dot e_j^{\phantom i}$, $\,j=1,\dots,k$, yield (\ref{fxh}) at
$\,t=0\,$ for $\,\tildeP\,$ instead of $\,\tildeH$. On the other hand,
$\,\mathbf{e}=\mathbf{e}(t)\,$ with
$\,\mathbf{e}=(e_1^{\phantom i},\dots,e_k^{\phantom i})\,$ is a regular lift of
$\,\mathrm{Ker}\hskip1.4ptF\hskip-.7pt(\xi(t))\,$ to $\,\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathcal{T}$,
showing that $\,d\pi_{\mathbf{e}}^{\phantom i}(\tildeP\mathbf{e})
=d\pi_{\mathbf{e}}^{\phantom i}\dot{\mathbf{e}}\,$ is the image of $\,\eta=\dot\xi\,$
under the differential of (\ref{ker}) at $\,\xi=\xi(t)$. This equality, at
$\,t=0$, uniquely determines $\,P:\mathsf{W}\to\mathcal{T}\hskip-1.8pt/\hskip.4pt\mathsf{W}\,$ for which
our $\,\tildeP:\mathsf{W}\to\mathcal{T}\,$ is a linear lift realizing
the image just mentioned as in the final paragraph of Remark~\ref{tgvec}),
so that $\,P\hskip.4pt=H$, and our claim follows.
\end{remark}
\section{Some relevant types of data}\label{td}
\setcounter{equation}{0}
We will repeatedly consider quadruples $\,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a,Q\,$ formed by
\begin{equation}\label{pbd}
\begin{array}{l}
\mathrm{a\ nontrivial\ closed\ interval\ }\,\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\hskip.7pt\mathrm{,\ a\
constant\ }\,\,a\in(0,\infty)\mathrm{,}\\
\mathrm{and\ a\ }\,C^\infty\mathrm{\ function\ }\,Q\,\mathrm{\ of\ the\
variable\ }\,\vt\in[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\mathrm{,\hskip-.7pt\ positive}\\
\mathrm{on\ }\,(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)\mathrm{,\hskip-.7pt\ such\ that\ }\hskip-1.5pt\,Q=0\,\mathrm{\ and\
}\hskip-.4pt\,dQ\hskip.4pt/\hskip-.7pt d\vt=\mp2a\hskip-.4pt\,\mathrm{\ at\ }\hskip-.4pt\,\vt=\vt\hskip-1.5pt_\pm\w\hskip.4pt,
\end{array}
\end{equation}
$\mp\,$ being the opposite sign of $\,\pm\hskip.4pt$. As explained below, we may then
also choose
\begin{equation}\label{sgn}
\begin{array}{l}
\mathrm{a\ sign\ }\hskip.7pt\pm\,(\mathrm{equal\ to\ }\hskip.7pt+\hskip.7pt\mathrm{\ or\
}\hskip.7pt-),\mathrm{\hskip-.7pt\ and\ a\ }C^\infty\hskip-2.5pt\mathrm{\
dif\-feo\-mor\-phism}\\
(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)\ni\vt\,\,\mapsto\,\,\rho\in(0,\infty)\,\,\mathrm{\ such\ that\
}\,\,d\hskip-.7pt\rho\hskip.7pt/d\vt\,=\,\mp\hskip.4pt a\rho/Q\hskip.4pt,
\end{array}
\end{equation}
a function $\,(0,\infty)\ni\rho\mapsto f\in\mathrm{I\!R}}%{\mathbf{R}$, unique up to an additive
constant, with
\begin{equation}\label{dfr}
a\rho\hskip.7pt\,d\hskip-1.2ptf\hskip-1.5pt/\hskip-.7pt d\hskip-.7pt\rho\,=\,2\hskip.7pt|\hskip.7pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|\hskip.4pt,
\end{equation}
and the unique increasing dif\-feo\-mor\-phism
$\,(0,\infty)\ni\rho\mapsto\sigma\in(0,\delta)\,$ such that
\begin{equation}\label{dsr}
a\rho\,d\sigma\hskip-.7pt/\hskip-.4pt d\hskip-.7pt\rho\,=\,Q^{\hs1\hskip-.7pt/2}\hskip10pt\mathrm{and}\hskip7pt
\sigma\hskip-.7pt\to0\hskip6pt\mathrm{as}\hskip6pt\rho\to0\hskip.4pt,
\end{equation}
the inverse of $\,\vt\mapsto\rho\,$ in (\ref{sgn}) being used to treat
$\,\vt,Q\,$ as functions of $\,\rho$, for
\begin{equation}\label{int}
\delta\in(0,\infty)\,\mathrm{\ equal\ to\ the\ integral\ of\
}\,Q^{-\nnh1\hskip-.7pt/2}\hskip.7pt d\vt\,\mathrm{\ over\ }\,(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)\hskip.4pt.
\end{equation}
In fact, one easily verifies that $\,\vt\mapsto\rho\,$ and
$\,\rho\mapsto\sigma\,$ with the stated properties exist, while $\,\delta\,$ is
finite. See \cite[Remark 5.1]{derdzinski-maschler-06},
\cite[Theorem~10.2(iii)]{derdzinski-kp}, \cite[p.\ 1661]{derdzinski-kp}.
\begin{lemma}\label{xtens}{\medit
Any\/ $\,f\,$ satisfying\/ {\rm(\ref{dfr})} has a\/ $\,C^\infty$ extension
to\/ $\,[\hs0,\infty)$, which is also a\/ $\,C^\infty$ function of\/
$\,\rho^2\hskip-.7pt\in[\hs0,\infty)\,$ in a sense analogous to
Remark\/~{\rm\ref{cifot}}. The first and second derivatives of
$\,f\,$ with respect to\/ $\,\rho^2$ obviously are\/
$\,|\hskip.7pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|/(a\rho^2)\,$ and\/
$\,(Q-2a\hskip.4pt|\hskip.7pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|)/(2\hskip.4pt a^2\hskip-2pt\rho^4)$.
}
\end{lemma}
\begin{proof}Since
$\,a\rho^2d\hskip-1.2ptf\hskip-1.5pt/\hskip-.4pt d\hskip-.7pt\rho^2\hskip-.7pt=|\hskip.7pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|$, it clearly
suffices to establish $\,C^\infty\hskip-1.5pt$-ex\-ten\-si\-bil\-i\-ty of
$\,\rho^2\hskip-.7pt\mapsto(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)/\rho^2$ to $\,[\hs0,\infty)$. To this end,
note that, according to
\cite[top line on p.\ 83 and Remark 4.3(ii)]{derdzinski-maschler-06},
$\,(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)/\rho^2$ (or, $\,Q/\rho^2$) may be extended to a $\,C^\infty$
function of $\,\vt\in[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\smallsetminus\{\vt\hskip-1pt_\mp\w\}\,$ (or, of
$\,\rho^2\hskip-.7pt\in[\hs0,\infty)$) with a nonzero value at $\,\vt\hskip-1.5pt_\pm\w$ (or,
respectively, at $\,0$). However, by (\ref{sgn}),
$\,2\hskip.4pt a\,d\vt/\hskip-.4pt d\hskip-.7pt\rho^2\hskip-.7pt=\mp\hskip.4pt Q/\rho^2\hskip-.7pt$, so that the variables
$\,\vt\,$ and $\,\rho^2$ just mentioned depend dif\-feo\-mor\-phic\-al\-ly on
each other.
\end{proof}
\begin{remark}\label{uniqe}It is the limit condition in (\ref{dsr}) that makes
$\,\sigma\hskip.7pt$ unique; by contrast, $\,\rho$ with (\ref{sgn}) is only unique up
to a positive constant factor.
\end{remark}
\begin{remark}\label{tends}In (\ref{sgn}), the increasing function
$\,\mp\rho\,$ of the variable $\,\vt\,$ clearly tends to $\,0\,$ as
$\,\vt\to\vt\hskip-1.5pt_\pm\w$, and to $\,\infty\,$ as $\,\vt\to\vt\hskip-1pt_\mp\w$.
\end{remark}
\begin{remark}\label{compo}The composite
$\,(0,\delta)\ni\sigma\mapsto\rho\mapsto\vt\in(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)\,$ of the inverses of
the above two dif\-feo\-mor\-phisms is the unique solution of the autonomous
equation $\,d\vt/d\sigma=\mp\hskip.4pt Q^{\hs1\hskip-.7pt/2}\hskip-.7pt$, with the sign $\,\pm\,$
fixed as in (\ref{sgn}), such that $\,\vt\to\vt\hskip-1.5pt_\pm\w$ as $\,\sigma\to0$. (We say
`autonomous' since (\ref{pbd}) makes $\,Q\,$ a function of $\,\vt$.) In fact,
any two solutions of the equivalent equation
$\,d\sigma/d\vt=\mp\hskip.4pt Q^{-\nnh1\hskip-.7pt/2}$ differ by a constant.
\end{remark}
\begin{remark}\label{iddbf}With (\ref{pbd}) -- (\ref{dfr}) fixed as above,
suppose that $\,\rho\,$ simultaneously denotes a positive function on a complex
manifold $\,M}%{{\mathrm{M}}'\hskip-.7pt$, which also turns $\,f\,$ into a function
$\,M}%{{\mathrm{M}}'\hskip-.7pt\to\mathrm{I\!R}}%{\mathbf{R}$. Now (\ref{idd}) for $\,\chi=\rho^2$ and the formulae in
Lemma~\ref{xtens} give
\begin{equation}\label{ddf}
4\hskip.4pt i\hskip.4pt a^2\hskip-2pt\rho^4\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf\,
=\,\hs4\hskip.4pt i\hskip.4pt a\hskip.4pt\rho^2|\hskip.7pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|\,
\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-.5pt\rho^2\hskip.7pt
+\,(2a\hskip.4pt|\hskip.7pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|-Q)\,d\hskip-.7pt\rho^2\wedge J^*\hskip-1.5pt d\hskip-.7pt\rho^2.
\end{equation}
\end{remark}
\begin{remark}\label{smdiv}If a $\,C^\infty$ function
$\,\vt\mapsto\zeta(\vt)\,$ on an interval $\,I\hskip.7pt$ having an endpoint $\,c$
vanishes at $\,c$, then $\,\zeta(\vt)=(\vt-c)\chi(\vt)\,$ for a $\,C^\infty$
function $\,\chi\,$ on $\,I\hskip.7pt$ equal, at $\,c$, to the derivative
$\,\zeta\hskip.4pt'$ of $\,\zeta$. This is the Taylor formula, with
$\,\chi(\vt)=\int_0^1\zeta\hskip.4pt'(c+s(\vt-c))\,ds$.
\end{remark}
\begin{remark}\label{adtet}Given a $\,C^\infty$ function
$\,t\mapsto\gamma(t)\,$ on an interval $\,I\hskip.7pt$ having the endpoint $\,0\,$
such that $\,(\gamma(0),\dot{\hskip-1.8pt\gamma\hskip0pt}(0))=(0,1)$, where
$\,(\hskip2.3pt)\hskip-1.5pt\dot{\phantom o}\hskip-.7pt=\,d/dt$, and $\,\gamma>0\,$ on
$\,t\in I\hskip-.7pt\smallsetminus\{0\}$, there exists a $\,C^\infty$
function $\,\theta:I\hskip-.7pt\to\mathrm{I\!R}}%{\mathbf{R}$, unique up to multiplications by positive
constants, for which $\,\dot\theta>0\,$ and
$\,\gamma\hskip.7pt\dot\theta=\hskip.4pt\theta\,$ everywhere in $\,I\hskip-.7pt$, while
$\,\theta(0)=0$. Namely, Remark~\ref{smdiv} implies that some $\,C^\infty$
functions $\,\beta\,$ (without zeros) and $\,\alpha\,$ on $\,I\hskip.7pt$ satisfy the
conditions $\,\gamma(t)=t/\beta(t)\,$ and $\,\beta(t)=1+t\dot\alpha(t)$. Thus,
$\,1/\gamma(t)=\dot\alpha(t)+1/t\,$ has the antiderivative
$\,t\mapsto\alpha(t)+\log|\hskip.4pt t\hskip-.2pt|\,$ on
$\,t\in I\hskip-.7pt\smallsetminus\{0\}$, and we may set
$\,\theta(t)=t\hskip.7pt e^{\alpha(t)}\hskip-.7pt$.
\end{remark}
\section{The Chern connection}\label{ck}
\setcounter{equation}{0}
Let $\,\langle\,,\rangle\,$ be the real part of a Her\-mit\-i\-an
fibre metric in a hol\-o\-mor\-phic complex vector bundle $\,N}%{{\mathrm{N}}\,$ over a
complex manifold $\,\varSigma$. The {\medit Chern connection\/} of $\,\langle\,,\rangle$, also
called its {\medit Her\-mit\-i\-an connection}, is the unique connection
$\,\mathrm{D}\,$ in $\,N}%{{\mathrm{N}}\,$ which makes $\,\langle\,,\rangle$ parallel and satisfies the
condition $\,\mathrm{D}^{0,1}\hskip-1.5pt=\,\overline{\hskip-2pt\partial}$, meaning that, for any
section $\,\xi\,$ of $\,N}%{{\mathrm{N}}\hskip-1.5pt$, the com\-plex-anti\-lin\-e\-ar part of the
real vec\-tor-bun\-dle morphism $\,\mathrm{D}\hskip.4pt\xi:T\hskip-1.5pt\varSigma\toN}%{{\mathrm{N}}$ equals
$\,\hskip.7pt\overline{\hskip-2pt\partial}\xi$, the image of $\,\xi\hskip.7pt$ under the Cau\-chy-Rie\-mann operator.
Cf.\ \cite[Sect.\ 1.4]{kobayashi-dg}.
The following five properties of the Chern connection $\,\mathrm{D}\,$ are
well known -- (e) is obvious; for (a) -- (d) see \cite[p.\ 32]{ballmann},
\cite[Propositions 1.3.5, 1.7.19 and 1.4.18]{kobayashi-dg}.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] $\mathrm{D}\,$ depends on $\,N}%{{\mathrm{N}}\,$ and $\,\langle\,,\rangle\,$ functorially
with respect to all natural operations, including $\,\mathrm{Hom}$, direct
sums, and pull\-backs under hol\-o\-mor\-phic mappings.
\item[{\rm(b)}] $R^{\mathrm{D}}\hskip-.7pt(J\hskip-.7pt w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)
=R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)$, with the notation of (\ref{nwt}) and
(\ref{jcs}), where $\,w,w\hskip.4pt'\hskip-.7pt,R^{\mathrm{D}}$ are any vector fields on
$\,\varSigma\,$ and, respectively, the curvature tensor of $\,\mathrm{D}$.
\item[{\rm(c)}] $\mathrm{D}\,$ is the Le\-vi-Ci\-vi\-ta connection of
$\,\langle\,,\rangle\,$ if $\,N}%{{\mathrm{N}}=T\hskip-1.5pt\varSigma\,$ and $\,\langle\,,\rangle\,$ is a K\"ah\-ler metric.
\item[{\rm(d)}] $\mathrm{D}\,$ coincides with the normal connection in the
normal bundle $\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma\,$ for any totally geodesic complex
sub\-man\-i\-fold $\,\varSigma\,$ a K\"ah\-ler manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)\,$ and the
Riemannian fibre metric $\,\langle\,,\rangle\,$ in $\,N}%{{\mathrm{N}}\,$ induced by $\,g$. (In addition,
it follows then that $\,N}%{{\mathrm{N}}\,$ must be a hol\-o\-mor\-phic sub\-bun\-dle of
$\,T\nnhM}%{{\mathrm{M}}\hskip-.7pt$.)
\item[{\rm(e)}] $\mathrm{D}\hskip.4pt$-par\-al\-lel sections of $\,N}%{{\mathrm{N}}\,$ are
hol\-o\-mor\-phic.
\end{enumerate}
Any given local hol\-o\-mor\-phic coordinates $\,z^\lambda$ in $\,\varSigma\,$ and
local hol\-o\-mor\-phic trivializing sections $\,e_b^{\phantom i}$ for $\,N}%{{\mathrm{N}}\hskip-1.5pt$,
on the same domain, associate with the Her\-mit\-i\-an fibre metric
$\,(\hskip2.3pt,\hskip1.3pt)$ having the real part $\,\langle\,,\rangle$, sections
$\,\xi\,$ of $\,N}%{{\mathrm{N}}\hskip-1.5pt$, and any connection $\,\mathrm{D}$, their component
functions $\,\gamma_{b\bar c}^{\phantom i}\hskip-.7pt
=(\hskip-.7pt e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)\,$ and
$\,\xi^b\hskip-.7pt,\varOmega_{\hskip-1.2pt\mu b}^{\,c},
\varOmega_{\hskip-1.2pt\bar\mu b}^{\,c}$, the latter characterized by
$\,\xi=\xi^be_b^{\phantom i}$, as well as $\,\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i}
=\varOmega_{\hskip-1.2pt\lambda b}^{\hskip.4pt c}e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-2.3pt$
and $\,\mathrm{D}\hskip-.7pt_{\bar\mu}^{\phantom i} e_b^{\phantom i}
=\varOmega_{\hskip-1.2pt\bar\mu b}^{\hskip.4pt c}e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-2.3pt$.
Here $\,\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i}$ and $\,\mathrm{D}\hskip-.7pt_{\bar\mu}^{\phantom i}$ denote the
$\,\mathrm{D}\hskip.4pt$-co\-var\-i\-ant differentiations in the direction od the
complexified coordinate vector fields
$\,\partial_\lambda^{\phantom i}=\partial/\partial z^\lambda$ and
$\,\partial_{\bar\mu}^{\phantom i}=\partial/\partial\bar z^{\bar\mu}\hskip-.7pt$, repeated
indices are summed over and, with
$\,\bar z^{\bar\mu}\hskip-.7pt,\bar\gamma_{\bar c\hskip.4pt b}^{\phantom i}$ and
$\,\bar\eta^{\bar c}$ standing for the complex conjugates of
$\,z^\mu\hskip-.7pt,\gamma_{c\bar b}^{\phantom i}$ and $\,\eta^c\hskip-.7pt$, Her\-mit\-i\-an symmetry of
$\,(\hskip2.3pt,\hskip1.3pt)$ amounts to
$\,\gamma_{b\bar c}^{\phantom i}\hskip-.7pt=\gamma_{\bar c\hskip.4pt b}^{\phantom i}$, while
$\,(\xi,\eta)=\gamma_{b\bar c}^{\phantom i}\,\xi^b\hskip-.7pt\eta^{\bar c}$ whenever
$\,\xi,\eta\,$ are sections of $\,N}%{{\mathrm{N}}\hskip-1.5pt$.
The real coordinate vector fields corresponding to the real coordinates
$\,\mathrm{Re}\,z^\mu,\,\mathrm{Im}\,z^\mu$ then are
$\,\partial_\mu^{\phantom i}\hskip-.7pt+\hskip.7pt\partial_{\bar\mu}^{\phantom{_{j_j}}}\hskip-3pt,
\,i\hskip.7pt(\partial_\mu^{\phantom i}\hskip-.7pt
-\hskip.7pt\partial_{\bar\mu}^{\phantom{_{j_j}}}\hskip-3pt)$, and so, given a
complexified vector field $\,v=v^\mu\partial_\mu^{\phantom i}\hskip-.7pt
+\hskip.4pt v^{\bar\mu}\partial_{\bar\mu}^{\phantom{_{j_j}}}\hskip-3pt$,
\begin{equation}\label{rvf}
v\,\mathrm{\ is\ real\ if\ and\ only\ if\ each\ } v^{\bar\mu}\mathrm{\ equals\
the\ complex\ conjugate\ of\ }\,v^\mu\hskip-.7pt.
\end{equation}
For a connection $\,\mathrm{D}\,$ to make $\,(\hskip2.3pt,\hskip1.3pt)\,$
parallel it is clearly necessary and sufficient that
$\,\partial_\mu^{\phantom i}\gamma_{b\bar c}^{\phantom i}
=\varOmega_{\hskip-1.2pt\mu b}^{\,e}\gamma_{e\bar c}^{\phantom i}
+\varOmega_{\hskip-1.2pt\mu\bar c}^{\,\bar e}\gamma_{b\bar e}^{\phantom i}$
and $\,\partial_{\bar\mu}^{\phantom i}\gamma_{b\bar c}^{\phantom i}
=\varOmega_{\hskip-1.2pt\bar\mu b}^{\,e}\gamma_{e\bar c}^{\phantom i}
+\varOmega_{\hskip-1.2pt\bar\mu\bar c}^{\,\bar e}\gamma_{b\bar e}^{\phantom i}$, where
$\,\varOmega_{\hskip-1.2pt\mu\bar c}^{\,\bar e}$ and
$\,\varOmega_{\hskip-1.2pt\bar\mu\bar c}^{\,\bar e}$ are defined to be the
complex conjugates of $\,\varOmega_{\hskip-1.2pt\bar\mu c}^{\,e}$ and
$\,\varOmega_{\hskip-1.2pt\mu c}^{\,e}$. On the other hand, the
condition $\,\mathrm{D}^{0,1}\hskip-1.5pt=\,\overline{\hskip-2pt\partial}\,$ is obviously equivalent to
$\,\mathrm{D}\hskip-.7pt_{\bar\mu}^{\phantom i} e_b^{\phantom i}=0$, that is,
$\,\varOmega_{\hskip-1.2pt\bar\mu b}^{\hskip.4pt c}=0$. Existence and uniqueness of
the Chern connection $\,\mathrm{D}$ now follow, its component functions being
\begin{equation}\label{chc}
\varOmega_{\hskip-1.2pt\bar\mu b}^{\hskip.4pt c}\hskip.7pt=\,0\hskip10pt
\mathrm{and}\hskip10pt\varOmega_{\hskip-1.2pt\lambda b}^{\,d}\hskip7pt
\mathrm{\ characterized\
by}\hskip7pt\varOmega_{\hskip-1.2pt\lambda b}^{\,d}\gamma_{d\bar c}^{\phantom i}\hskip.7pt
=\,\partial_\lambda^{\phantom i}\gamma_{b\bar c}^{\phantom i}\hskip.4pt.
\end{equation}
Consequently, the Chern connection $\,\mathrm{D}\,$ has the curvature
components
\begin{equation}\label{ccd}
R_{\lambda\bar\mu b\bar c}^{\phantom i}=-R_{\bar\mu\lambda b\bar c}^{\phantom i}
=\partial_\lambda^{\phantom i}\partial_{\bar\mu}^{\phantom i}\gamma_{b\bar c}^{\phantom i}
-\varOmega_{\hskip-1.2pt\lambda b}^{\,d}
\partial_{\bar\mu}^{\phantom i}\gamma_{d\bar c}^{\phantom i}\hskip.4pt,\hskip14pt
R_{\lambda\mu b\bar c}^{\phantom i}=R_{\bar\lambda\bar\mu b\bar c}^{\phantom i}=0
\end{equation}
(which implies (b) above). Here $\,R_{\lambda\bar\mu b\bar c}^{\phantom i}
=(R^{\mathrm{D}}\hskip-.7pt(\partial_\lambda^{\phantom i},\partial_{\bar\mu}^{\phantom i})\hskip.4pt
e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)$, and analogously for the other three
pairs $\,\bar\mu\lambda,\hskip.4pt\lambda\mu,\hskip.4pt\bar\lambda\bar\mu\,$ of indices. We
obtain (\ref{ccd}) from (\ref{cur}) via differentiation by parts, noting that
$\,[\hskip.4pt\partial_\lambda^{\phantom i},\partial_\mu^{\phantom i}]
=[\hskip.4pt\partial_\lambda^{\phantom i},\partial_{\bar\mu}^{\phantom i}]=0\,$ while, by (\ref{chc}),
$\,\mathrm{D}\hskip-.7pt_{\bar\mu}^{\phantom i} e_b^{\phantom i}=0\,$
and
$\,(\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)
=\partial_\lambda^{\phantom i}\gamma_{b\bar c}^{\phantom i}$. For instance,
$\,R_{\lambda\bar\mu b\bar c}^{\phantom i}
=(\mathrm{D}\hskip-.7pt_{\bar\mu}^{\phantom i}
\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)
=\partial_{\bar\mu}^{\phantom i}(
\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)
-(\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},
\mathrm{D}\hskip-1.5pt_\mu^{\phantom i} e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)
=\partial_{\bar\mu}^{\phantom i}\partial_\lambda^{\phantom i}\gamma_{b\bar c}^{\phantom i}
-\varOmega_{\hskip-1.2pt\lambda b}^{\,d}
\partial_{\bar\mu}^{\phantom i}\gamma_{\bar cd}^{\phantom i}$. Similarly,
$\,(\mathrm{D}\hskip-1.5pt_\mu^{\phantom i}
\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)
=\partial_\mu^{\phantom i}(
\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)
-(\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},
\mathrm{D}\hskip-.7pt_{\bar\mu}^{\phantom i} e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)\,$ equals
$\,\partial_\mu^{\phantom i}(
\mathrm{D}\hskip-1.5pt_\lambda^{\phantom i} e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)
=\partial_\mu^{\phantom i}\partial_\lambda^{\phantom i}\gamma_{b\bar c}^{\phantom i}$, which is symmetric in
$\,\lambda,\mu\,$ (and so $\,R_{\lambda\mu b\bar c}^{\phantom i}=0$).
Let $\,f:\varSigma\to\mathrm{I\!R}}%{\mathbf{R}$. With the notational conventions of Remark~\ref{posit},
\begin{equation}\label{ddb}
i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf\,\,
=\,\,[\hskip.4pt\partial_\lambda^{\phantom i}\partial_{\bar\mu}^{\phantom i} f\hskip.4pt]\,
dz^\lambda\hskip-.7pt\wedge\hskip.4pt d\bar z^{\bar\mu}\hskip.4pt,
\end{equation}
as $\,d\hskip-1.2ptf=[\hskip.4pt\partial_\lambda^{\phantom i} f\hskip.4pt]\,dz^\lambda\hskip-.7pt
+[\hskip.4pt\partial_{\bar\mu}^{\phantom i} f\hskip.4pt]\,d\bar z^{\bar\mu}$ and, in the case where
$\,f=z^\lambda$ (or, $\,f=\bar z^{\bar\mu}$), the com\-plex-val\-ued
$\,1$-form $\,J^*\hskip-1.5pt d\hskip-1.2ptf=(d\hskip-1.2ptf)\hskip-.4pt J\,$ equals
$\,i\hskip.7pt dz^\lambda$ or, respectively, $\,-i\hskip.7pt d\bar z^{\bar\mu}\hskip-.7pt$.
Equivalently,
\begin{equation}\label{lmz}
(i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf)
(\partial_\lambda^{\phantom i},\partial_{\bar\mu}^{\phantom i})\,\,
=\,\,\partial_\lambda^{\phantom i}\partial_{\bar\mu}^{\phantom i} f,\hskip12pt
(i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf)
(\partial_\lambda^{\phantom i},\partial_\mu^{\phantom i})\,\,
=\,\,(i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf)
(\partial\hskip.4pt_{\bar\lambda},\partial_{\bar\mu}^{\phantom i})\,\,=\,\,0\hskip.4pt.
\end{equation}
Thus, by (\ref{ddb}) and (\ref{rvf}), for
$\,\omega=i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5ptf\,$ and any real vector
fields $\,v,u\,$ on $\,\varSigma$,
\begin{equation}\label{ovu}
\omega\hskip.4pt(\hskip-.4pt v,u)\,\,=\,\,2\hskip.7pt\mathrm{Im}\hskip2pt
(v^\lambda u^{\bar\mu}\partial_\lambda^{\phantom i}\partial_{\bar\mu}^{\phantom i} f)\hskip.4pt.
\end{equation}
\begin{lemma}\label{ddrsq}{\medit
With\/ $\,\pi:N}%{{\mathrm{N}}\hskip-.7pt\to\varSigma\,$ and\/ $\,\rho:N}%{{\mathrm{N}}\hskip-.7pt\to[\hs0,\infty)\,$ denoting
the bundle projection and the norm function of\/ $\,\langle\,,\rangle$, \,for\/
$\,N}%{{\mathrm{N}}\hskip-.7pt,\varSigma,\langle\,,\rangle\,$ as above, the Chern connection\/ $\,\mathrm{D}\,$ of\/
$\,\langle\,,\rangle\,$ and the $\,2$-form\/
$\,\omega=i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-.5pt\rho^2$ satisfy the following
conditions.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] The horizontal distribution of\/ $\,\mathrm{D}\,$ constitutes
a complex vector sub\-bun\-dle of\/ $\,T\nnhN}%{{\mathrm{N}}\hskip-1.5pt$, and is\/
$\,\omega$-or\-thog\-o\-nal, in an obvious sense, to the vertical
distribution\/ $\,\mathrm{Ker}\hskip2.3ptd\pi$.
\item[{\rm(ii)}] Part\/ {\rm(b)} of Remark\/~{\rm\ref{ddnsq}} describes\/
$\,\omega\hskip.7pt$ restricted to any fibre\/ $\hsN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ \,of\/
$\hsN}%{{\mathrm{N}}\hskip-.7pt$, where\/ $\,y\in\varSigma$.
\item[{\rm(iii)}] Whenever\/ $\,x=(y,\xi)\inN}%{{\mathrm{N}}\hskip-1.5pt$, cf.\
Remark\/~{\rm\ref{tlspc}}, the restriction of\/ $\,\omega\hskip-.7pt_x^{\phantom i}$ to the
horizontal space of\/ $\,\mathrm{D}\,$ at\/ $\,x\,$ equals the\/
$\,d\pi\hskip-1.5pt_x^{\phantom i}$-pull\-back of the\/ $\,2$-form\/
$\,\langle\hskip-.7pt R_y^{\mathrm{D}}\hskip-.7pt(\,\cdot\,,\hskip.7pt\cdot\,)\hskip.4pt\xi,i\hskip.4pt\xi\rangle\,$ at\/
$\,y\in\varSigma$.
\item[{\rm(iv)}] The Chern connection\/ $\,\hat{\mathrm{D}}\,$ of\/
$\,e^\theta\hskip-1.5pt\langle\,,\rangle$, \,for any function\/ $\,\theta:\varSigma\to\mathrm{I\!R}}%{\mathbf{R}$, is related
to\/ $\,\mathrm{D}\,$ by\/ $\,\hat{\mathrm{D}}=\mathrm{D}
+(\partial\theta)\hskip-1.5pt\otimes\hskip-.7pt\mathrm{Id}\hskip.4pt$, so that\/
$\,\hat\varOmega_{\hskip-1.2pt\bar\mu b}^{\hskip.4pt c}\hskip-.7pt=0\,$ and\/
$\,\hat\varOmega_{\hskip-1.2pt\lambda b}^{\,c}\hskip-.7pt
=\varOmega_{\hskip-1.2pt\lambda b}^{\,c}\hskip-.7pt
+\hskip.7pt\delta_{\hskip-.7pt b}^c\partial\hskip-1.5pt_\lambda^{\phantom i}\hskip-.7pt\theta$. Also,
$\,\widehat v_x^{\phantom i}\hskip-.7pt\in\mathrm{Span}_\bbC^{\phantom i}(\widetilde v_x^{\phantom i},\xi)$,
where\/ $\,\widetilde v_x^{\phantom i}$ \,and\/ $\,\widehat v_x^{\phantom i}$ \,denote the\/
$\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal and\/
$\,\hat{\mathrm{D}}\hskip.4pt$-hor\-i\-zon\-tal lifts of any\/ $\,v\inT\hskip-2.3pt_y\w\hskip-.7pt\varSigma\,$ to\/
$\,x=(y,\xi)\inN}%{{\mathrm{N}}\hskip-.7pt$.\phantom{$1^{1^1}$}
\end{enumerate}
}
\end{lemma}
\begin{proof}In terms of complexified coordinate vector fields
$\,\partial_\lambda^{\phantom i},\,\partial_{\bar\mu}^{\phantom i}$ in $\,\varSigma\,$ and their analogs
$\,\partial_\lambda^{\phantom i},\,\partial_{\bar\mu}^{\phantom i},\,\partial_b^{\phantom i},
\,\partial_{\bar c}^{\phantom i}$ corresponding to the local hol\-o\-mor\-phic
coordinates $\,z^\lambda\hskip-.7pt,\xi^b$ in $\,N}%{{\mathrm{N}}\hskip-.7pt$, the
$\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal lifts
$\,\widetilde{\partial}_\lambda^{\phantom i},\,\widetilde{\partial}_{\bar\mu}^{\phantom i}$ of the
former are given by
\begin{equation}\label{hlo}
\widetilde{\partial}_\mu^{\phantom i}\hskip.7pt=\,\partial_\mu^{\phantom i}\hskip.7pt
-\,\varOmega_{\hskip-1.2pt\mu b}^{\hskip.4pt e}\xi^b\partial_e^{\phantom i}\hskip.7pt
-\,\varOmega_{\hskip-1.2pt\bar\mu\bar e}^{\hskip.4pt\bar c}\xi^{\bar e}
\partial_{\bar c}^{\phantom i}\hskip.4pt,\hskip16pt\widetilde{\partial}_{\bar\mu}^{\phantom i}\hskip.7pt
=\,\partial_{\bar\mu}^{\phantom i}\hskip.4pt.
\end{equation}
Since $\,J\partial_\lambda^{\phantom i}=i\hskip.7pt\partial_\lambda^{\phantom i}$ and
$\,J\partial_{\bar\mu}^{\phantom i}=-i\hskip.7pt\partial_{\bar\mu}^{\phantom i}$, (\ref{hlo}) implies
com\-plex-lin\-e\-ar\-i\-ty of the $\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal lift
operation relative to the com\-plex-stru\-cture tensors $\,J$, proving the
first part of (i). Assertion (ii) is in turn obvious from naturality of the
operator $\,i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}$.
Next, applying (\ref{lmz}) to $\,f=\rho^2$ and the coordinates
$\,z^\lambda\hskip-.7pt,\xi^b$ rather than just $\,z^\lambda\hskip-.7pt$, we obtain
$\,\omega\hskip.4pt(\partial_\lambda^{\phantom i},\partial_{\bar\mu}^{\phantom i})
=\xi^b\xi^{\bar c}\partial_\lambda^{\phantom i}\partial_{\bar\mu}^{\phantom i}\gamma_{b\bar c}^{\phantom i}$,
$\,\omega\hskip.4pt(\partial_e^{\phantom i},\partial_{\bar\mu}^{\phantom i})
=\xi^{\bar c}\partial_{\bar\mu}^{\phantom i}\gamma_{e\bar c}^{\phantom i}$,
$\,\omega\hskip.4pt(\partial_{\bar c}^{\phantom i},\partial_{\bar\mu}^{\phantom i})
=\omega\hskip.4pt(\partial_{\bar c}^{\phantom i},\partial_{\bar e}^{\phantom i})=0$,
$\,\omega\hskip.4pt(\partial_\lambda^{\phantom i},\partial_{\bar c}^{\phantom i})
=\xi^b\partial_\lambda^{\phantom i}\gamma_{b\bar c}^{\phantom i}$, and
$\,\omega\hskip.4pt(\partial_b^{\phantom i},\partial_{\bar c}^{\phantom i})=\gamma_{b\bar c}^{\phantom i}$. Thus,
$\,\omega\hskip.4pt(\widetilde{\partial}_\lambda^{\phantom i},\partial_{\bar c}^{\phantom i})=0\,$ by
(\ref{chc}), which amounts to the remaining claim in (i); similarly,
(\ref{hlo}) and (\ref{ccd}) yield
$\,\omega\hskip.4pt(\widetilde{\partial}_\lambda^{\phantom i},\widetilde{\partial}_{\bar\mu}^{\phantom i})
=\xi^b\xi^{\bar c}(\partial_\lambda^{\phantom i}\partial_{\bar\mu}^{\phantom i}\gamma_{b\bar c}^{\phantom i}
-\varOmega_{\hskip-1.2pt\lambda b}^{\,e}\partial_{\bar\mu}^{\phantom i}\gamma_{e\bar c}^{\phantom i})
=\xi^b\xi^{\bar c}R_{\lambda\bar\mu b\bar c}^{\phantom i}$. Now (iii) follows: for the
$\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal lifts $\,\widetilde v,\widetilde u\,$ of
any real vectors $\,v,u\inT\hskip-2.3pt_y\w\hskip-.7pt\varSigma$, (\ref{lmz}) -- (\ref{ovu}) and the last
equality give
$\,\omega\hskip-.7pt_x^{\phantom i}(\widetilde v,\widetilde u)=2\hskip.7pt\mathrm{Im}\hskip2pt
(v^\lambda u^{\bar\mu}R_{\lambda\bar\mu b\bar c}^{\phantom i}\xi^b\xi^{\bar c})$, while
at the same time $\,(R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt v,u)\hskip.4pt\xi,i\hskip.4pt\xi)\,$ is
imaginary, and so
$\,\langle\hskip-.7pt R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt v,u)\hskip.4pt\xi,i\hskip.4pt\xi\rangle
=\mathrm{Im}\hskip2pt(R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt v,u)\hskip.4pt\xi,i\hskip.4pt\xi)
=(R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt v,u)\hskip.4pt\xi,i\hskip.4pt\xi)
=\xi^b\xi^{\bar c}
(R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt v,u)\hskip.4pt e_b^{\phantom i},e\hskip-.7pt_{c\phantom{b}}^{\phantom i}\hskip-3.2pt)\,$
which, analogously, equals $\,2\hskip.7pt\mathrm{Im}\hskip2pt
(v^\lambda u^{\bar\mu}R_{\lambda\bar\mu b\bar c}^{\phantom i}\xi^b\xi^{\bar c})$.
Finally, (\ref{chc}) and (\ref{hlo}) imply (iv).
\end{proof}
\section{Examples: Vector bundles}\label{ev}
\setcounter{equation}{0}
The ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triples
constructed in this section are all noncompact. What makes them relevant is
the fact that some of them serve as universal building blocks for compact
ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triples. (See
Theorem~\ref{first}.)
We begin with data $\,\varSigma,h,N}%{{\mathrm{N}}\hskip-.7pt,\langle\,,\rangle,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a,Q,\pm\hskip.7pt,\vt\mapsto\rho\,$ and
$\,\rho\mapsto f\,$ consisting of
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] the real part $\hskip.7pt\langle\,,\rangle\hskip.7pt$ of a Her\-mit\-i\-an fibre metric
in a hol\-o\-mor\-phic complex vector bundle $\,N}%{{\mathrm{N}}\hskip.7pt$ of positive fibre
dimension over a K\"ah\-ler manifold $\,(\varSigma,h)$,
\item[{\rm(ii)}] some objects $\,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a,Q,\pm\hskip.7pt,\vt\mapsto\rho\,$ and
$\,\rho\mapsto f\,$ satisfying (\ref{pbd}) -- (\ref{dfr}).
\end{enumerate}
Letting $\,\pi:N}%{{\mathrm{N}}\to\varSigma\,$ stand for the bundle projection, $\,\mathrm{D}\,$
for the Chern connection of $\,\langle\,,\rangle\,$ (see Section~\ref{ck}), and $\,\rho\,$
both for the variable in (ii) and for the norm function
$\,N}%{{\mathrm{N}}\to[\hs0,\infty)$, we use the inverse mapping of $\,\vt\mapsto\rho$,
cf.\ (\ref{sgn}), to
\begin{equation}\label{trt}
\mathrm{treat\ }\,\vt,Q\,\mathrm{\ and\ }\,f\,\mathrm{\ as\ functions\
}\,N}%{{\mathrm{N}}\to\mathrm{I\!R}}%{\mathbf{R}\mathrm{,\ denoted\ here\ by\ }\hat\vt,\hat Q\,\mathrm{\ and\
}\,\hat f.
\end{equation}
Denoting by $\,\hat J\,$ (rather than $\,J$) the com\-plex-stru\-cture tensor
of $N}%{{\mathrm{N}}\hskip-1.5pt$, we define a K\"ah\-ler metric $\,\,\hat{g\hskip2pt}\hskip-1.3pt\,$ on $\,N}%{{\mathrm{N}}\,$ by
requiring the K\"ah\-ler forms $\,\hat\omega=\hat{g\hskip2pt}\hskip-1.3pt(\hat J\,\cdot\,,\,\cdot\,)\,$
and $\,\omega^h\hskip-.7pt=h(J\,\cdot\,,\,\cdot\,)$ to be related by
$\,\hat\omega\,=\,\hskip.7pt\pi\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt\omega^h\hskip.7pt
+\,i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5pt\hat f\hskip-.7pt$, which amounts to
\begin{equation}\label{hoh}
\hat{g\hskip2pt}\hskip-1.3pt\,\,=\,\,\pi\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt h\,\,-\,\,(i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5pt\hat f)
(\hat J\,\cdot\,,\,\cdot\,)\hskip.4pt.
\end{equation}
As $\,\hat\omega\,$ should be positive (Remark~\ref{posit}), it is necessary
to assume here that
\begin{equation}\label{thd}
\pi\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt h\,\,-\,\,(i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5pt\hat f)
(\hat J\,\cdot\,,\,\cdot\,)\,\,
\mathrm{\ is\ positive}\hyp\mathrm{definite\ at\ every\ point\ of\ }\,N}%{{\mathrm{N}}.
\end{equation}
The above construction uses the objects (i) -- (ii) with (\ref{thd}), and
leads to what is shown below (Theorem~\ref{ehggk}) to be a
ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triple
$\,(N}%{{\mathrm{N}}\hskip-.7pt,\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt)$.
It is convenient, however, to provide the following equivalent, though less
concise, description of $\,\hat{g\hskip2pt}\hskip-1.3pt\,$ and $\,\hat J\,$ restricted to the
complement $\,N}%{{\mathrm{N}}'\hskip-.7pt=N}%{{\mathrm{N}}\hskip-.7pt\smallsetminus\varSigma$ of the zero section in
$\,N}%{{\mathrm{N}}\hskip-1.5pt$. It uses the complex di\-rect-sum decomposition
\begin{equation}\label{tnp}
T\nnhN}%{{\mathrm{N}}'\,=\,\hskip.7pt\hat{\mathcal{V}}\hskip1.2pt\oplus\hskip.7pt\hat{\mathcal{H}}^\mp\hskip-1.5pt
\oplus\hskip.7pt\hat{\mathcal{H}}^\bullet\hskip-.7pt,
\end{equation}
in which $\,\hat{\mathcal{H}}^\bullet$ is the horizontal distribution of
$\,\mathrm{D}\,$ and
$\,\hat{\mathcal{V}}\hskip1.2pt\oplus\hskip.7pt\hat{\mathcal{H}}^\mp\hskip-.7pt
=\hskip.7pt\mathrm{Ker}\hskip2.3ptd\pi\,$ equals the vertical distribution, with the
summands $\,\hat{\mathcal{V}}\,$ and $\,\hat{\mathcal{H}}^\mp$ forming, on
each punctured fibre $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}$, the complex
radial distribution (Remark~\ref{ddnsq}) and, respectively, its
$\,\langle\,,\rangle$-or\-thog\-o\-nal complement in
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}$. (The word `complex' preceding
(\ref{tnp}) is justified by Lemma~\ref{ddrsq}(i).) To describe $\,\hat{g\hskip2pt}\hskip-1.3pt\,$
and $\,\hat J$, we declare that the three summands of (\ref{tnp}) are
$\,\hat J$-in\-var\-i\-ant and mutually $\,\hat{g\hskip2pt}\hskip-1.3pt$-or\-thog\-o\-nal, that
$\,\hat J\,$ restricted to $\,\hat{\mathcal{V}}\,$ agrees, along each
punctured fibre $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}$, with its standard
com\-plex-stru\-cture tensor of the complex vector space
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, that the differential of $\,\pi\,$ at every
$\,(y,\xi)\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}$, cf.\ Remark~\ref{tlspc},
maps $\,\mathcal{H}_{(y,\,\xi)}^{\phantom i}$ com\-plex-lin\-e\-ar\-ly onto $\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma\,$
and, with the constant $\,a\in(0,\infty)\,$ and function $\,\hat\vt\,$
appearing in (i) and (\ref{trt}),
\begin{equation}\label{hgv}
\begin{array}{rl}
\mathrm{a)}&
\hat{\mathcal{V}}\hskip-.7pt,\,\hat{\mathcal{H}}^\mp\mathrm{\ and\
}\,\hat{\mathcal{H}}^\bullet\mathrm{\ in\ (\ref{tnp})\ are\ mutually\
}\,\hat{g\hskip2pt}\hskip-1.3pt\hyp\mathrm{or\-thog\-o\-nal,}\phantom{\displaystyle\frac1{2_{1_1}}}\\
\mathrm{b)}&
a^2\hskip-.7pt\rho^2\hat{g\hskip2pt}\hskip-1.3pt\hskip.7pt=\,\hat Q\,\langle\,,\rangle\,\mathrm{\ on\ }\,\hat{\mathcal{V}}\hskip-.7pt,
\hskip16pta\hskip.4pt\rho^2\hat{g\hskip2pt}\hskip-1.3pt\hskip.7pt=\,2\hskip.7pt|\hskip.7pt\hat \vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|\,\langle\,,\rangle\,\mathrm{\ on\
}\,\hat{\mathcal{H}}^\pm\hskip-.7pt,\\
\mathrm{c)}&
\hat{g\hskip2pt}\hskip-1.3pt\hskip-1.5pt_x^{\phantom i}(\hskip-.7pt w_x^{\phantom i},w_x^{\phantom i}\hskip-5pt'\hskip1.7pt)=h_y^{\phantom i}(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)-
\displaystyle{\frac{|\hskip.4pt\hat\vt(x)-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|}{a\rho^2}}\,
\langle\hskip-.7pt R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,J\hskip-2pt_y^{\phantom i} w\hskip.4pt'\hskip.4pt)\hskip.4pt\xi,i\hskip.4pt\xi\rangle
\hskip8pt\mathrm{with}\hskip5pt\rho=|\hskip.4pt\xi|\hskip.4pt,
\end{array}
\end{equation}
at any $\,x=(y,\xi)\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}$, where
$\,w,w\hskip.4pt'$ are any two vectors in $\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma$, and $\,w_x^{\phantom i},w_x^{\phantom i}\hskip-5pt'$
denote their $\,\mathrm{D}\hskip.4pt$-\hskip-.4pt hor\-i\-zon\-tal lifts to $\,x$. The
vertical vector fields $\,\hat v,\hat u\,$ with
\begin{equation}\label{lvf}
\hat v_{(y,\,\xi)}^{\phantom i}\,=\,\mp\hskip.4pt a\hskip.4pt\xi\hskip.4pt,\hskip16pt
\hat u_{(y,\,\xi)}^{\phantom i}\,=\,\mp\hskip.4pt ai\hskip.4pt\xi\hskip.4pt,
\end{equation}
allow us to characterize the restrictions of $\,\hat{g\hskip2pt}\hskip-1.3pt\,$ and $\,\hat J\,$ to
$\,\hat{\mathcal{V}}=\mathrm{Span}\hskip.7pt(\hskip-.4pt\hat v,\hat u)\,$ by
\begin{equation}\label{ana}
\hat{g\hskip2pt}\hskip-1.3pt(\hat v,\hat v)\,=\,\hat{g\hskip2pt}\hskip-1.3pt(\hat u,\hat u)\,
=\,\hat Q\hskip.4pt,\hskip12pt\hat{g\hskip2pt}\hskip-1.3pt(\hat v,\hat u)\,=\,0\hskip.4pt,\hskip12pt\hat u\,
=\,\hat J\hskip-1.1pt\hat v\hskip.4pt.
\end{equation}
Note that the symmetry of
$\,\hat{g\hskip2pt}\hskip-1.3pt\hskip-1.5pt_x^{\phantom i}(\hskip-.7pt w_x^{\phantom i},w_x^{\phantom i}\hskip-5pt'\hskip1.7pt)\,$ in
$\,w_x^{\phantom i},w_x^{\phantom i}\hskip-5pt'$ reflects (b) in Section~\ref{ck}.
Lemma~\ref{ddrsq} easily implies that the definition (\ref{hgv}) of $\,\hat{g\hskip2pt}\hskip-1.3pt\,$
is actually equivalent to (\ref{hoh}), while condition (\ref{thd}) is nothing
else than positivity of the right-hand side in (\ref{hgv}.c) whenever
$\,w=w\hskip.4pt'\hskip-.7pt\ne0$.
\begin{theorem}\label{ehggk}{\medit
For any data\/ {\rm(i)} -- {\rm(ii)} with\/ {\rm(\ref{thd})}, let us define\/
$\,\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt\,$ by\/ {\rm(\ref{trt})} -- {\rm(\ref{hoh})}.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] $(N}%{{\mathrm{N}}\hskip-.7pt,g,\vt)\,$ is a ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler
triple.
\item[{\rm(b)}] The fibres $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}=\pi^{-\nnh1}(y)$,
$y\in\varSigma$, \,are totally geodesic complex \hbox{sub\-man\-i\-folds of
$\,(N}%{{\mathrm{N}}\hskip-.7pt,g)$.}
\item[{\rm(c)}] The zero section\/ $\,\varSigma\subseteqN}%{{\mathrm{N}}\,$ coincides with\/
$\,\varSigma^\pm\hskip-1.5pt$, \,the\/ $\,\vt\hskip-1.5pt_\pm\w\hskip.7pt$ level set of\/ $\,\vt$.
\item[{\rm(d)}] The\/ $\,\hat{g\hskip2pt}\hskip-1.3pt$-gra\-di\-ent\/
$\,\hat v=\hat\nabla\hskip-.5pt\hat\vt\,$ and\/
$\,\hat{S}=\hat\nabla\hskip-.7pt\hat v\,$ satisfy\/ {\rm(\ref{lvf})} --
{\rm(\ref{ana})} and the equality
\begin{equation}\label{gsw}
2\hskip.4pt\hat{g\hskip2pt}\hskip-1.3pt\hskip-1.5pt_x^{\phantom i}(\hat{S}\hskip-.7pt_x^{\phantom i} w_x^{\phantom i},w_x^{\phantom i}\hskip-5pt'\hskip1.7pt)\,=
\,\pm\,\frac{\hat Q(x)}{a\rho^2}\,
\langle\hskip-.7pt R_y^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,J\hskip-2pt_y^{\phantom i} w\hskip.4pt'\hskip.4pt)\hskip.4pt\xi,i\hskip.4pt\xi\rangle\hskip.4pt,
\hskip7pt\mathrm{where}\hskip5pt\rho=|\hskip.4pt\xi|>0 \hskip.4pt,
\end{equation}
the assumptions being the same as in\/ {\rm(\ref{hgv}.c)}.
\end{enumerate}
}
\end{theorem}
\begin{proof}The $\,\hat v$-di\-rec\-tion\-al derivatives of the norm squared
$\,\rho^2$ and of $\,\rho\,$ are, obviously, $\,\mp2\hskip.4pt a\rho^2$ and
$\,\mp\hskip.4pt a\rho$. As $\,d\vt/d\hskip-.7pt\rho=\mp\hskip.4pt Q/(a\rho)\,$ in (\ref{sgn}), we see using (\ref{trt}) and (\ref{ana}) that
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(e)}] $\hat Q=\hat{g\hskip2pt}\hskip-1.3pt(\hat v,\hat v)\,$ equals the
$\,\hat v$-di\-rec\-tion\-al derivative
$\,\hat{g\hskip2pt}\hskip-1.3pt(\hat v,\hat\nabla\hskip-.5pt\hat\vt)\,$ of $\,\hat\vt$.
\end{enumerate}
Furthermore, $\,\mathrm{D}\hskip.4pt$-par\-al\-lel transports preserve the real
fibre metric $\,\langle\,,\rangle$. Therefore,
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(f)}] $\rho,\hat\vt\,$ and $\,\hat Q\,$ are constant along every
$\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal curve in $\,N}%{{\mathrm{N}}\hskip-1.5pt$,
\end{enumerate}
due to (ii) and (\ref{trt}). The equality
$\,\hat v=\hat\nabla\hskip-.5pt\hat\vt\,$ now follows:
$\,\hat{\mathcal{V}}=\mathrm{Span}\hskip.7pt(\hskip-.4pt\hat v,\hat u)$, and $\,\hat\vt\,$
is a function of the norm $\,\rho$, so that
$\,\hat v-\hat\nabla\hskip-.5pt\hat\vt\,$ is $\,\hat{g\hskip2pt}\hskip-1.3pt$-or\-thog\-o\-nal to
$\,\hat{\mathcal{H}}^\bullet\hskip-.7pt,\,\hat{\mathcal{H}}^\pm\hskip-.7pt,\hat u\,$ and
$\,\hat v\,$ by (f), (\ref{hgv}.a), (\ref{ana}), and (e). Also, (\ref{lvf})
clearly gives hol\-o\-mor\-phic\-i\-ty of $\,\hat v$, while closedness and
positivity of the form
$\,\hat{g\hskip2pt}\hskip-1.3pt(\hat J\,\cdot\,,\,\cdot\,)=\hskip.7pt\pi\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt\omega^h\hskip-.7pt
+i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip-1.5pt\hat f\,$ imply that $\,\hat{g\hskip2pt}\hskip-1.3pt\,$ is a
K\"ah\-ler metric, and $\,\hat\vt\,$ has a geodesic $\,\hat{g\hskip2pt}\hskip-1.3pt$-gra\-di\-ent
$\,\hat v$, its $\,\hat{g\hskip2pt}\hskip-1.3pt$-norm squared $\,\hat Q\,$ being a function of
$\,\hat\vt$, cf.\ (\ref{trt}) and Lemma~\ref{ggqft}. We have thus proved (a).
Next, Remark~\ref{tends} yields (c).
The $\,\pi$-pro\-ject\-a\-ble local sections of
$\,\hat{\mathcal{H}}^\bullet$ are precisely the same as the
$\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal lifts of local vector fields tangent to
$\,\varSigma$, and their local flows act as $\,\mathrm{D}\hskip.4pt$-par\-al\-lel
transports between the fibres. As the the sub\-man\-i\-fold metrics of the
fibres are defined by (\ref{hgv}.a) -- (\ref{hgv}.b), this last action
consists -- by (f) -- of isometries which, being linear, also preserve the
vertical vector field $\,\hat v\,$ with (\ref{lvf}). Hence
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(g)}] $\hat v\,$ commutes with all local
$\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal lifts $\,w$,
\end{enumerate}
and, at the same time, applying Remark~\ref{tglvs} to any such $\,w\ne0\,$ we
obtain (b).
Finally, by (g) and Remark~\ref{dvwwp}, the left-hand side of (\ref{gsw})
equals the $\,\hat v$-di\-rec\-tion\-al derivative of the right-hand side in
(\ref{hgv}.c). To evaluate the latter, note that only the factor
$\,-\hskip.4pt|\hskip.4pt\hat\vt(x)-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|=\pm\hskip.4pt(\hskip.4pt\hat\vt(x)-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)\,$ in the second term
needs to be differentiated, as the first term and the remaining factor of the
second one are constant along $\,\hat v\,$ (due to constancy along
$\,\hat v\,$ of $\,\xi/\hskip-1.5pt\rho=\xi/\hskip-.4pt|\hskip.4pt\xi|$, obvious from (\ref{lvf})).
Now (e) implies (\ref{gsw}), completing the proof.
\end{proof}
A {\medit special K\"ah\-ler-Ric\-ci potential\/}
\cite{derdzinski-maschler-06} on a K\"ah\-ler manifold $\,(M}%{{\mathrm{M}}\hskip-.7pt,g)\,$ is any
nonconstant function $\,\vt:M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}\,$ such that $\,v=\nabla\hskip-.5pt\vt\,$ is
real-hol\-o\-mor\-phic, while, at points where $\,v\ne0$, all nonzero vectors
orthogonal to $\,v\,$ and $\,J\hskip-.7pt v\,$ are eigen\-vec\-tors of both
$\,\nabla\hskip-.7pt v$ and the Ric\-ci tensor, with
$\,\nabla\hskip-.7pt v:T\nnhM}%{{\mathrm{M}}\to T\nnhM}%{{\mathrm{M}}\,$ as in (\ref{nwt}). We then call
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ an {\medit SKRP triple}. All SKRP triples $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$
are ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triples, due to
their eas\-i\-ly-ver\-i\-fied property
\cite[Remark~7.1]{derdzinski-maschler-03} that $\,v=\nabla\hskip-.5pt\vt$, wherever nonzero,
is an eigenvector of $\,\nabla\hskip-.7pt v$. Cf.\ (\ref{nvv}).
Compact SKRP triples $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ have been classified in
\cite[Theorem 16.3]{derdzinski-maschler-06}. They are divided into Class 1, in
which $\,M}%{{\mathrm{M}}\,$ is the total space of a hol\-o\-mor\-phic $\,\bbC\mathrm{P}^1$ bundle,
and Class 2, with $\,M}%{{\mathrm{M}}\,$ bi\-hol\-o\-mor\-phic to $\,\bbC\mathrm{P}^m$ for
$\,m=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}\hskip-.7pt$.
\begin{lemma}\label{cltwo}{\medit
Up to isomorphisms, in the sense of Definition\/~{\rm\ref{ggktr}}, compact
SKRP triples of Class 2 are the same as CP triples constructed using\/
{\rm(\ref{dta}.ii)} with\/ $\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\mathsf{L}=1\hskip.4pt$.
}
\end{lemma}
\begin{proof}See \cite[Remark 6.2]{derdzinski-maschler-06}. (Note that the
case $\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\mathsf{L}=m-1\,$ in (\ref{dta}.ii) obviously leads to the same
isomorphism type.)
\end{proof}
In (i) above, $\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\varSigma\ge0$, which allows the possibility of a
one\hskip.7pt-\hskip-.7pt point base manifold $\,\varSigma=\{y\}$, so that, as a complex manifold,
$\,N}%{{\mathrm{N}}\,$ is a complex vector space, namely, the fibre
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$. According to
\cite[pp.\ 85-86]{derdzinski-maschler-06}, under the standard identification
(\ref{inc}) for $\,\mathsf{V}=N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, both $\,\hat{g\hskip2pt}\hskip-1.3pt\,$ and $\,\hat\vt\,$
then can be extended to the projective space
$\,\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})$, giving rise to a Class 2
SKRP triple $\,(M}%{{\mathrm{M}},\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt)$, where
$\,M}%{{\mathrm{M}}=\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})$.
\begin{lemma}\label{xplic}{\medit
The SKRP triples\/ $\,(M}%{{\mathrm{M}},g,\vt)\,$ just mentioned, with\/
$\,M}%{{\mathrm{M}}=\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})$, represent all
isomorphism types of compact SKRP triples of Class 2. Such types include all
compact ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triples of
complex dimension $\,1$.
}
\end{lemma}
\begin{proof}For the first part, see
\cite[Remark 6.2]{derdzinski-maschler-06}. The final clause is in turn
immediate from Remark~\ref{cpone} and Lemma~\ref{cltwo}.
\end{proof}
\begin{remark}\label{fbstu}As a consequence of the second part of
Remark~\ref{trivl}, for $\,(N}%{{\mathrm{N}}\hskip-.7pt,g,\vt)\,$ as in Theorem~\ref{ehggk}, every
fibre $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ is the underlying complex manifold of a
ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple, realizing a special case of
Theorem~\ref{ehggk}: that of a one\hskip.7pt-\hskip-.7pt point base manifold $\,\{y\}$. Its
projective compactification
$\,\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\,$ constitutes, for
reasons mentioned above, the underlying complex manifold of an SKRP triple of
Class 2. The resulting sub\-man\-i\-fold metric on the complement of
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ in
$\,\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\,$ (that is, on the
projective hyper\-plane at infinity, identified via (\ref{inc}) with
$\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$) equals
$\,2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)/\hskip-.4pt a$ times the Fu\-bi\-ni-Stu\-dy metric
associated -- as in Remark~\ref{fbstm} -- with $\,\langle\,,\rangle$.
Namely, let $\,\xi,\eta\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ have $\,\langle\xi,\xi\rangle=1\,$ and
$\,\langle\xi,\eta\rangle=\langle i\xi,\eta\rangle=0$. The curve $\,t\mapsto t\eta\,$ of
vectors $\,t\eta\,$ tangent to $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ at the points
$\,t\hskip.4pt\xi$, satisfies, in view of
(\ref{hgv}.b) and Remark~\ref{tends}, the limit relation
$\,\hat{g\hskip2pt}\hskip-1.3pt\hskip-1.5pt_{(y,\,t\hskip.4pt\xi)}^{\phantom i}(t\eta,t\eta)
\to2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)\langle\eta,\eta\rangle/\hskip-.4pt a\,$ as $\,t\to\infty$. At the
same time, $\,t\hskip.4pt\xi\,$ (or, the tangent vector $\,t\eta$) tends, as
$\,t\to\infty$, to the point $\,\bbC(0,\xi)\,$ of
$\,\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})
\smallsetminusN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$,
identified with $\,\bbC\xi\in\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ or,
respectively, to the vector tangent to $\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$
at $\,\bbC\xi\,$ which is the image of $\,\eta\,$ under
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm($*$)}] the differential of the projection $\,\xi\mapsto\bbC\xi\,$
restricted to the unit sphere of $\,\langle\,,\rangle$.
\end{enumerate}
The claim about the tangent vectors, which clearly implies our assertion, can
be justified as follows. The vector $\,t\eta\,$ equals $\,x_s^{\phantom i}(t,0)\,$
(notation preceding Remark~\ref{jcobi}) with
$\,x(t,s)=t\hskip.4pt(\xi+s\eta)\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, so that
$\,|\hskip.7pt x(t,s)|^2\hskip-.7pt=t^2(1+|s\eta|^2)\,$ and, setting
$\,\zeta(t,s)
=[1+t^2(1+|s\eta|^2)]^{-\nnh1\hskip-.7pt/2}(1,x(t,s))
\in\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, we get $\,|\hskip.4pt\zeta(t,s)|=1\,$ for the
di\-rect-sum Euclidean norm. Identifying $\,x(t,s)\,$ with
$\,\bbC(1,x(t,s))=\bbC\hskip.4pt\zeta(t,s)
\in\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})$, we see that $\,t\eta$,
treated as tangent to
$\,\mathrm{P}(\hskip-.4pt\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\,$ at $\,\bbC(1,t\hskip.4pt\xi)$,
is the image, under the analog of ($*$) for
$\,\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, of the vector
$\,\zeta_s^{\phantom i}(t,0)=(1+t^2)^{-\nnh1\hskip-.7pt/2}(0,x_s^{\phantom i}(t,0))
=(0,t\hskip.4pt(1+t^2)^{-\nnh1\hskip-.7pt/2}\eta)$ tangent to the unit sphere of
$\,\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ at the point
$\,\zeta(t,0)=(1+t^2)^{-\nnh1\hskip-.7pt/2}(1,t\hskip.4pt\xi)\,$ and having the required
limit $\,(0,\eta)\,$ as $\,t\to\infty$, which we identify with $\,\eta$.
\end{remark}
\begin{remark}\label{unnec}The construction summarized in Theorem~\ref{ehggk}
has an obvious generalization, arising when, in (\ref{pbd}),
$\,\vt\mapsto Q\,$ is assumed to be only defined on the half-open interval
$\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\smallsetminus\{\vt\hskip-1pt_\mp\w\}$, and $\,dQ\hskip.4pt/\hskip-.7pt d\vt=\mp2a\,$ at
$\,\vt=\vt\hskip-1.5pt_\pm\w$ just for one fixed sign $\,\pm\hskip.7pt$. Our discussion focuses on a
narrower case since this is the case relevant to the study of {\medit
compact\/} ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triples.
\end{remark}
\section{Local properties}\label{lp}
\setcounter{equation}{0}
Throughout this section $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is a fixed
ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triple (Definition~\ref{ggktr}). We
use the symbols
\begin{equation}\label{sym}
J,\,\,v,\,\,u,\,\,M}%{{\mathrm{M}}'\hskip-.7pt,\,\,\psi,\,\,Q,\,\,\mathcal{V},\,\,\mathcal{V}^\perp\hskip-.7pt,
\,\,S,\,\,A
\end{equation}
for the com\-plex-struc\-ture tensor $\,J:T\nnhM}%{{\mathrm{M}}\to T\nnhM}%{{\mathrm{M}}\,$ of the
underlying complex manifold $\,M}%{{\mathrm{M}}\hskip-.7pt$, the gradient $\,v=\nabla\hskip-.5pt\vt$, its
$\,J$-im\-age $\,u=J\hskip-.7pt v$, the open set $\,M}%{{\mathrm{M}}'$ where $\,v\ne0$, the
function $\,\psi\,$ on $\,M}%{{\mathrm{M}}'$ with (\ref{nvv}), the function
$\,Q=g(\hskip-.4pt v,v)\,$ on $\,M}%{{\mathrm{M}}\hskip-.7pt$, the distribution
$\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)\,$ on $\,M}%{{\mathrm{M}}'\hskip-1.5pt$, its orthogonal
complement, as well as the en\-do\-mor\-phisms $\,S=\nabla\hskip-.7pt v\,$ and
$\,A=\nabla\hskip-.7pt u\,$ of $\,T\nnhM}%{{\mathrm{M}}\hskip-.7pt$, cf.\ (\ref{nwt}). Under the above
hypotheses,
\begin{equation}\label{loc}
\begin{array}{rl}
\mathrm{a)}&v,u\,\mathrm{\ are\ both\ hol\-o\-mor\-phic,}\,|v|=|u|
=Q^{\hs1\hskip-.7pt/2}\hskip-.7pt,\mathrm{\ and\ }\,A\hskip.7pt=\hskip.7pt JS\hskip.7pt=\hsS\hskip-.7pt J\hskip.4pt,\\
\mathrm{b)}&u=J\hskip-.7pt v\,\mathrm{\ is\ a\ Kil\-ling\ field\ commuting\
with\ }\,v\hskip.4pt,\mathrm{\ and\ orthogonal\ to\ }\,v\hskip.4pt,\\
\mathrm{c)}&\nabla\hskip-3pt_{w}^{\phantom i}A\hskip.7pt=\hskip.7pt R(\hskip-.4pt u,w)\,\mathrm{\ and\
}\,\nabla\hskip-3pt_{w}^{\phantom i}S\hskip.7pt=\hskip.7pt-\hskip-.7pt J[R(\hskip-.4pt u,w)]\,\mathrm{\ for\ any\
vector\ field\ }\,w\hskip.4pt,\\
\mathrm{d)}&S\,\mathrm{\ is\ self}\hyp\mathrm{ad\-joint\ and\
}\,J,A\,\mathrm{\ are\ skew}\hyp\mathrm{ad\-joint\ at\ every\ point\ of\
}\,M}%{{\mathrm{M}}\hskip.4pt,\\
\mathrm{e)}
g([w,w\hskip.4pt'],u)=-\nh2\hskip.4pt g(A w,w\hskip.4pt'\hskip.4pt)\,\mathrm{\ for\ any\ local\
sections\ }\,w,w\hskip.4pt'\mathrm{\ of\ }\,\mathcal{V}^\perp\hskip.4pt,\\
\mathrm{f)}&\nabla\hskip-3pt_{v}^{\phantom i} v=\psi\hskip.4ptv
=-\nabla\hskip-3pt_u^{\phantom i} u\,\mathrm{\
and\ }\,\nabla\hskip-3pt_u^{\phantom i} v=\nabla\hskip-3pt_{v}^{\phantom i} u
=\psi\hskip.4ptu\,\mathrm{\ everywhere\ in\ }\,M}%{{\mathrm{M}}',\\
\mathrm{g)}&Q\,\mathrm{\ is,\ locally\ in\ }\,M}%{{\mathrm{M}}'\hskip-.7pt,\mathrm{\
a\ function\ of\ }\,\vt\hskip.4pt,\mathrm{\ and\ }\,2\psi=\hskip.7pt d\hskip.4pt Q/d\vt\hskip.4pt,\\
\mathrm{h)}&J,\hsS,\hsA\,\mathrm{\ and\ the\ local\ flows\ of\
}\,u\,\mathrm{\ and\ }\,v\,\mathrm{\ leave\ }\,\mathcal{V}\,\mathrm{\ and\
}\,\mathcal{V}^\perp\mathrm{\ invariant.}
\end{array}
\end{equation}
In (\ref{loc}.c), $\,R\,$ denotes the curvature tensor of $\,g$, and the
notation of (\ref{nwt}) is used.
In fact, hol\-o\-mor\-phic\-i\-ty of $\,v\,$ (cf.\ Definition~\ref{ggktr})
combined with (\ref{ajs}) -- (\ref{hol}) gives (\ref{loc}.a), $\,u\,$ being
hol\-o\-mor\-phic due to (\ref{hol}), as $\,A\hskip.7pt=\hskip.7pt JS\hskip.7pt=\hsS\hskip-.7pt J\,$
commutes with $\,J$. Next, (\ref{loc}.b) follows from (\ref{kil}) and the
Lie-brack\-et equality
$\,[u,v]=\nabla\hskip-3pt_u^{\phantom i} v-\nabla\hskip-3pt_{v}^{\phantom i} u=S\hskip-.7pt u-A v=S\hskip-.7pt u-S\hskip-.7pt J\hskip-.7pt v=0$,
obvious in view of (\ref{loc}.a), while (\ref{loc}.c) (or, (\ref{loc}.d)) is a
direct consequence of (\ref{scd}) and (\ref{loc}.a) or, respectively, of
of (\ref{loc}.b) combined with the fact that $\,v\,$ is a gradient. We now
obtain (\ref{loc}.e) from (\ref{loc}.d), noting that
$\,g(\nabla\hskip-3pt_{w}^{\phantom i} w\hskip.4pt'\hskip-.7pt,u)
=-\hskip.4pt g(\hskip-.4pt w\hskip.4pt'\hskip-.7pt,\nabla\hskip-3pt_{w}^{\phantom i} u)
=-\hskip.4pt g(\hskip-.4pt w\hskip.4pt'\hskip-.7pt,A w)$. On the other hand, (\ref{loc}.b), (\ref{loc}.a) and
(\ref{nvv}) yield
$\,\nabla\hskip-3pt_u^{\phantom i} v=\nabla\hskip-3pt_{v}^{\phantom i} u
=\nabla\hskip-3pt_{v}^{\phantom i}(J\hskip-.7pt v)=J\nabla\hskip-3pt_{v}^{\phantom i} v
=\psi\hskip.4ptJ\hskip-.7pt v=\psi\hskip.4ptu\,$ and so
$\,\nabla\hskip-3pt_u^{\phantom i} u=\nabla\hskip-3pt_u^{\phantom i}(J\hskip-.7pt v)=J\nabla\hskip-3pt_u^{\phantom i} v
=\psi\hskip.4ptJu=-\hskip.4pt\psi\hskip.4ptv$, establishing (\ref{loc}.f), while
Lemma~\ref{ggqft}, (\ref{tnd}) and (\ref{loc}.f) imply (\ref{loc}.g). That
$\,J,S,A\,$ all leave $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)\,$
invariant is
clear as $\,J\hskip-.7pt v=u$ and $\,Ju=-\hskip.4pt v\,$ while, by (\ref{loc}.f),
$\,S\hskip-.7pt v,S\hskip-.7pt u,A v,A u\,$ are sections of $\,\mathcal{V}\hskip-1.5pt$. The
same conclusion for $\,\mathcal{V}^\perp$ is now immediate from (\ref{loc}.d).
By (\ref{loc}.b), the local flows of $\,v\,$ and $\,u\,$ preserve $\,v,u\,$
and $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)$. The $\,u$-in\-var\-i\-ance of
$\,\mathcal{V}^\perp$ now follows from (\ref{loc}.b). Finally, let $\,w\,$ be
a section of $\,\mathcal{V}^\perp\hskip-1.5pt$. Writing $\,\langle\,,\rangle\,$ for $\,g$, we get
$\,\langle[v,w],v\rangle=\langle\nabla\hskip-3pt_{v}^{\phantom i} w-\nabla\hskip-3pt_{w}^{\phantom i} v,v\rangle
=-\hskip.4pt\langle w,\nabla\hskip-3pt_{v}^{\phantom i} v\rangle-\laS\hskip-.7pt w,v\rangle=-\laS\hskip-.7pt w,v\rangle
=-\langle w,S\hskip-.7pt v\rangle=0$, cf.\ (\ref{loc}.d) and (\ref{loc}.f). Similarly,
$\,\langle[v,w],u\rangle=\langle\nabla\hskip-3pt_{v}^{\phantom i} w-\nabla\hskip-3pt_{w}^{\phantom i} v,u\rangle
=-\hskip.4pt\langle w,\nabla\hskip-3pt_{v}^{\phantom i} u\rangle-\laS\hskip-.7pt w,u\rangle=-\laS\hskip-.7pt w,u\rangle
=-\langle w,S\hskip-.7pt u\rangle=0$. Thus, $\,[v,w]\,$ is a section of
$\,\mathcal{V}^\perp$ as well. In view of Remark~\ref{prjct}, this completes
the proof of (\ref{loc}.h). For easy reference, note that, by (\ref{loc}.a) --
(\ref{loc}.b),
\begin{equation}\label{gvv}
g(v,v)\,=\,g(u,u)\,
=\,Q\hskip.4pt,\hskip12ptg(v,u)\,=\,0\hskip.4pt,\hskip12ptu\,
=\,J\hskip-1.1ptv\hskip.4pt.
\end{equation}
\begin{lemma}\label{dvgww}{\medit
Under the assumptions preceding\/ {\rm(\ref{loc})}, on\/ $\,M}%{{\mathrm{M}}'\hskip-.7pt$,
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] the distribution\/ $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)\,$
is in\-te\-gra\-ble and has totally geodesic leaves,
\item[{\rm(b)}] a local section of\/ $\hskip.7pt\mathcal{V}^\perp\hskip-1.5pt$ is
pro\-ject\-able along\/ $\mathcal{V}\hskip.7pt$ if and only if it commutes with\/
$\hskip.7pt u\hskip.7pt$ and\/ $\hskip.7pt v$,
\item[{\rm(c)}] if local sections\/ $\,w\,$ and\/ $\,w\hskip.4pt'$ of\/
$\,\mathcal{V}^\perp\,$ commute with\/ $\,u\,$ and\/ $\,v$, then
\begin{equation}\label{dvg}
\begin{array}{rlrl}
\mathrm{i)}&d_v^{\phantom i}[\hskip.7pt g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)]\,=\,2\hskip.4pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt,
&\mathrm{ii)}&d_v^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]\,=\,2\psi\hskip.4pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt,\\
\mathrm{iii)}&d_v^{\phantom i}[Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]\,=\,0\hskip.4pt,&
\end{array}
\end{equation}
\item[{\rm(d)}] $d_v^{\phantom i} Q\hskip-1.5pt=\hskip-.7pt 2\psi Q\hskip.4pt\,$ and\/
$\hskip.7pt d_u^{\phantom i}[\hskip.7pt g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)]\hskip-.7pt=\hskip-.7pt d_u^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]\hskip-.7pt
=\hskip-.7pt d_u^{\phantom i} Q\hskip-1.5pt=0\,$ for any\/ $\,w,w\hskip.4pt'$ as in\/ {\rm(c)},
\item[{\rm(e)}] $[\nabla\hskip-3pt_{v}^{\phantom i}S]\hskip.4pt w\,=\,2(\psi\hskip.7pt-\hsS)S\hskip.4pt w\hskip.7pt\,$
whenever\/ $\,w\,$ is a local section of\/ $\,\mathcal{V}^\perp\hskip-.7pt$.
\end{enumerate}
}
\end{lemma}
\begin{proof}Assertions (a) -- (b) are obvious from (\ref{loc}.b) and,
respectively, Remark~\ref{liebr} combined with (\ref{loc}.h). Next, let
$\,\pounds\hskip-.5pt_v^{\phantom i} w=\pounds\hskip-.5pt_v^{\phantom i} w\hskip.4pt'\hskip-.7pt=\pounds\hskip-.5pt_u^{\phantom i} w
=\pounds\hskip-.5pt_v^{\phantom i} w\hskip.4pt'\hskip-.7pt=0$. Since $\,\pounds\hskip-.5pt_v^{\phantom i}$ and
$\,\pounds\hskip-.5pt_u^{\phantom i}$ act on functions as $\,d_v^{\phantom i}$ and $\,d_u^{\phantom i}$,
(\ref{lvg}) implies (\ref{dvg}.i), and $\,d_u^{\phantom i}[\hskip.7pt g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)]=0\,$ as
$\,\pounds\hskip-.5pt_u^{\phantom i} g=0\,$ by (\ref{loc}.b). For similar reasons,
$\,d_u^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]
=\pounds\hskip-.5pt_u^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]=0$. (Namely,
(\ref{loc}.c) gives $\,\nabla\hskip-3pt_u^{\phantom i}S=0$, so that (\ref{loc}.a) and
(\ref{lie}), with $\,u,S\,$ rather than $\,v,B}%{{\mathrm{B}}$, yield
$\,\pounds\hskip-.5pt_u^{\phantom i}S=0$.) On the other hand, by (\ref{gvv}),
$\,g(\hskip-.4pt v,v)=Q$. Now (\ref{tnd}), (\ref{loc}.f) and (\ref{loc}.b) imply that
$\,d_u^{\phantom i}\vt=d_u^{\phantom i} Q=0\,$ and $\,d_v^{\phantom i} Q=2\psi Q$, establishing (d).
Using (\ref{loc}.a) we get $\,g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)=g(JS\hskip-.7pt w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)
=g(A w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)\,$ which,
by (\ref{loc}.e), is nothing else than $\,-\hskip.4pt g([w,J\hskip-.7pt w\hskip.4pt'],u)/2$. Hence
$\,2\hskip.7pt d_v^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]=2\pounds\hskip-.5pt_v^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]
=-\pounds\hskip-.5pt_v^{\phantom i}[\hskip.7pt g(\hskip-.4pt u,[w,J\hskip-.7pt w\hskip.4pt']))]
=-\hskip.4pt[\hskip-.7pt\pounds\hskip-.5pt_v^{\phantom i} g](\hskip-.4pt u,[w,J\hskip-.7pt w\hskip.4pt']))$. (Our assumption that
$\,\pounds\hskip-.5pt_v^{\phantom i} w=\pounds\hskip-.5pt_v^{\phantom i} w\hskip.4pt'\hskip-.7pt=0$ gives
$\,\pounds\hskip-.5pt_v^{\phantom i}(J\hskip-.7pt w\hskip.4pt'\hskip.4pt)=0$, as $\,v\,$ is hol\-o\-mor\-phic,
which in turn yields $\,\pounds\hskip-.5pt_v^{\phantom i}[w,J\hskip-.7pt w\hskip.4pt']=0$, while
$\,\pounds\hskip-.5pt_v^{\phantom i} u=0$, cf.\ (\ref{loc}.b).) From (\ref{lvg}),
(\ref{loc}.f) and (\ref{loc}.a) we now obtain
$\,d_v^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]=-\hskip.4pt[\hskip-.7pt\pounds\hskip-.5pt_v^{\phantom i} g](\hskip-.4pt u,[w,J\hskip-.7pt w\hskip.4pt']))/2
=-\hskip.4pt g(S\hskip-.7pt u,[w,J\hskip-.7pt w\hskip.4pt'])=-\nh2g(\psi u,[w,J\hskip-.7pt w\hskip.4pt'])
=2\psi\hskip.4pt g(A w,\hskip-.7pt J\hskip-.7pt w\hskip.4pt'\hskip.4pt)\hskip-.7pt=\hskip-.7pt-\nnh2\psi\hskip.4pt g(J\nnhA w,w\hskip.4pt'\hskip.4pt)\hskip-.7pt
=2\psi\hskip.4pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)$, that is, (\ref{dvg}.ii), which, since
$\,d_v^{\phantom i} Q=2\psi Q\,$ by (d), also proves (\ref{dvg}.iii).
Finally, (\ref{loc}.h) and the equality
$\,\nabla\hskip-3pt_{v}^{\phantom i}S\hskip.7pt=\hskip.7pt-\hskip-.7pt J[R(\hskip-.4pt u,v)]$, cf.\ (\ref{loc}.c), combined with
(a), imply that $\,\nabla\hskip-3pt_{v}^{\phantom i}S-(2\psi\hskip.7pt-\hsS)S\,$ leaves
$\,\mathcal{V}^\perp$ invariant. To obtain (e), it now suffices to show that
$\,[\nabla\hskip-3pt_{v}^{\phantom i}S]\hskip.4pt w-(2\psi\hskip.7pt-\hsS)S\hskip.4pt w\,$ is orthogonal to
$\,w\hskip.4pt'$ for any local sections $\,w,w\hskip.4pt'$ of $\,\mathcal{V}^\perp\hskip-.7pt$. We are
free to assume here that $\,w=w\hskip.4pt'$ (due to self-ad\-joint\-ness of
$\,S=\nabla\hskip-.7pt v$) and that $\,w\,$ commutes with $\,u\,$ and $\,v\,$ (see
(b)). Differentiation by parts gives, by (\ref{dvg}.iii) and (\ref{loc}.d),
$\,g([\nabla\hskip-3pt_{v}^{\phantom i}S]\hskip.4pt w,w)=d_v^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w)]
-g(S\nabla\hskip-3pt_{v}^{\phantom i} w,w)-g(S\hskip-.7pt w,\nabla\hskip-3pt_{v}^{\phantom i} w)
=2\psi\hskip.4pt g(S\hskip-.7pt w,w)-2g(S\hskip-.7pt w,S\hskip-.7pt w)$, as required, with
$\,\nabla\hskip-3pt_{v}^{\phantom i} w=S\hskip-.7pt w\,$ since $\,[v,w]=0$.
\end{proof
\section{Horizontal Ja\-co\-bi fields}\label{hj}
\setcounter{equation}{0}
In addition to using the assumptions and notations of Section~\ref{lp}, we now
let $\,\varGamma$ stand for the underlying
\hbox{one\hskip.4pt-}\hskip0ptdi\-men\-sion\-al manifold of a fixed maximal integral
curve of $\,v\,$ in $\,M}%{{\mathrm{M}}'\hskip-.7pt$. We restrict the objects in (\ref{sym}) to
$\,\varGamma\,$ without changing the notation, and select a unit-speed
parametrization $\,t\mapsto x(t)\,$ of the geodesic $\,\varGamma$ such that
\begin{equation}\label{dxe}
\dot x\,=\,v/|v|\,=\,Q^{-\nnh1\hskip-.7pt/2}v\hskip10pt\mathrm{along}\hskip6pt\varGamma,
\hskip6pt\mathrm{where}\hskip6ptv=\nabla\hskip-.5pt\vt\hskip.4pt.
\end{equation}
As an obvious consequence of (\ref{dxe}), (\ref{dvt}) and Lemma~\ref{dvgww}(d),
\begin{equation}\label{dtq}
\dot\vt\,=\,Q^{\hs1\hskip-.7pt/2},\hskip20pt\dot Q\,=\,2\psi Q^{\hs1\hskip-.7pt/2},
\hskip20pt\mathrm{with}\hskip8pt(\hskip2.3pt)\hskip-1.5pt\dot{\phantom o}\hskip-.7pt=\,d/dt\,
=\,d_{\dot x}^{\phantom i}\hskip.4pt.
\end{equation}
Any constant $\,c\in[\mathrm{I\!R}}%{\mathbf{R}\smallsetminus\vt(\varGamma)]\cup\{\infty\}$, where
$\,\vt(\varGamma)\,$ is the range of $\,\vt\,$ on $\,\varGamma$, gives rise to the
function $\,\lambda_c^{\phantom i}:\varGamma\to\mathrm{I\!R}}%{\mathbf{R}\,$ defined by
\begin{equation}\label{lgm}
\lambda_c^{\phantom i}\,=\,\hskip.7pt Q/[2(\vt-c)]\hskip.4pt,
\end{equation}
the convention being that $\,\lambda_c^{\phantom i}$ is identically zero when
$\,c=\infty$. We denote by $\,\mathcal{W}\,$ the set of all
$\,\mathcal{V}^\perp\hskip-.7pt$-val\-ued vector fields
$\,t\mapsto w(t)\in\mathcal{V}^\perp_{\hskip-1ptx(t)}$ along $\,\varGamma\,$
satisfying the equation
\begin{equation}\label{nxw}
\nabla\hskip-3pt_{\dot x}^{\phantom i} w\,=\,Q^{-\nnh1\hskip-.7pt/2}S\hskip-.7pt w\hskip.4pt.
\end{equation}
Of particular interest to us are $\,c\,$ such that
\begin{equation}\label{wze}
\begin{array}{rl}
\mathrm{a)}&
c\in[\mathrm{I\!R}}%{\mathbf{R}\smallsetminus\vt(\varGamma)]\cup\{\infty\}\,\,\mathrm{\ and\
}\,\,\mathcal{W}[c]\ne\{0\}\hskip.4pt,\mathrm{\ \ where}\\
\mathrm{b)}&
\mathcal{W}[c]\,=\,\{w\in\mathcal{W}:S\hskip-.7pt w=\lambda_c^{\phantom i}\hskip-.7pt w\}\hskip.4pt.
\end{array}
\end{equation}
About pro\-ject\-abil\-i\-ty along $\,\mathcal{V}\,$ in (i) below, see
Remark~\ref{pralg} and Lemma~\ref{dvgww}(a).
\begin{theorem}\label{jacob}{\medit
Under the above hypotheses, the following conclusions hold.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] $\mathcal{V}^\perp\hskip-.7pt$-val\-ued solutions\/ $\,w\,$ to\/
{\rm(\ref{nxw})} are nothing else than restrictions to\/ $\,\varGamma\hskip.7pt$ of those
local sections of\/ $\hskip.7pt\mathcal{V}^\perp\hskip-1.5pt$ with domains containing\/
$\,\varGamma\hskip.7pt$ which are pro\-ject\-able along\/ $\,\mathcal{V}\hskip-.7pt$.
\item[{\rm(ii)}] All\/ $\,w\,$ as in\/ {\rm(i}), that is, all elements of\/
$\,\mathcal{W}\hskip-.7pt$, are Ja\-co\-bi fields along\/ $\,\varGamma\hskip-.7pt$.
\item[{\rm(iii)}] Every vector in\ $\,\mathcal{V}^\perp_{\hskip-1ptx(t)}$
equals\/ $\,w(t)\,$ for some unique\/ $\,w\in\mathcal{W}\hskip-.7pt$.
\item[{\rm(iv)}] $\mathcal{W}\,$ is a complex vector space of complex
dimension\/ $\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}-1$, and the direct sum of all\/
$\,\hskip.7pt\mathcal{W}[c]\,$ for\/ $\hsc\,$ in\/ {\rm(\ref{wze}.a)}, with\/
$\,w\mapsto J\hskip-.7pt w\,$ serving as the multiplication by\/ $\,i\in\bbC$.
\item[{\rm(v)}] A function\/ $\,t\mapsto\lambda(t)\,$ on the parameter
interval of\/ $\,t\mapsto x(t)\,$ satisfies the equation\/
$\,d\lambda/dt=2(\psi-\lambda)\lambda\hskip.4pt Q^{-\nnh1\hskip-.7pt/2}$, with\/
$\,\psi,Q\,$ evaluated at\/ $\,x(t)$, if and only if\/
$\,\lambda(t)=\lambda_c^{\phantom i}(x(t))$, cf.\ {\rm(\ref{lgm})}, for some\/
$\,c\in[\mathrm{I\!R}}%{\mathbf{R}\smallsetminus\vt(\varGamma)]\cup\{\infty\}\,$ and all\/ $\,t$.
\item[{\rm(vi)}] At any\/ $\,x=x(t)\in\varGamma\hskip-.7pt$, the eigen\-values of
$\,S\hskip-1.5pt_x^{\phantom i}:\mathcal{V}^\perp_{\hskip-1ptx}\hskip-.7pt
\to\mathcal{V}^\perp_{\hskip-1ptx}$, cf.\ {\rm(\ref{loc}.h)}, are precisely
the values\/ $\,\lambda_c^{\phantom i}(x)\,$ for all\/
$\,c\,$ in\/ {\rm(\ref{wze}.a)}. The eigen\-space of\/
$\,S\hskip-1.5pt_x^{\phantom i}:\mathcal{V}^\perp_{\hskip-1ptx}\hskip-.7pt
\to\mathcal{V}^\perp_{\hskip-1ptx}$ corresponding to\/
$\,\lambda_c^{\phantom i}(x)\,$ is\/ $\,\{w(t):w\in\mathcal{W}[c]\}$.
\item[{\rm(vii)}] $R(\hskip-.4pt w,u)\hskip.4pt u\hskip-1.5pt=\hskip-.7pt\hskip-1.5pt R(\hskip-.4pt w,v)\hskip.4pt v\hskip-1.5pt
=\hskip-1.5pt(\psi-S)S\hskip.4pt w\hskip-1.5pt=\hskip-.7pt\hskip-1.5pt R(\hskip-.4pt v,u)\hskip.4pt J\hskip-.7pt w\hskip-.7pt/2\,$ on\/ $\hsM}%{{\mathrm{M}}'\hskip-.7pt$
for sections\/ $\hskip.7pt w\hskip.4pt$ of\/ $\,\mathcal{V}^\perp\hskip-1.5pt$.
\item[{\rm(viii)}] If\/ $\,\vt(\varGamma)=(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)\,$ is bounded, then\/
$\,Q/(\vt-\hskip-.4pt\vt\hskip-1.8pt_+\w)\le\,2S\hskip.4pt\le Q/(\vt-\hskip-.4pt\vt\hskip-1.7pt_-\w)\,$ on\/
$\,\mathcal{V}^\perp\hskip-1.5pt$.
\end{enumerate}
}
\end{theorem}
\begin{proof}Any $\,w\,$ as in the second line of (i), restricted to
$\,\varGamma\hskip-.7pt$, becomes both a Ja\-co\-bi field (by Lemmas~\ref{jcbfl}
and~\ref{dvgww}(b)) and a $\,\mathcal{V}^\perp\hskip-.7pt$-val\-ued solution to
(\ref{nxw}) (since $\,S=\nabla\hskip-.7pt v$, so that (\ref{dxe}) and
Lemma~\ref{dvgww}(b) give
$\,\nabla\hskip-3pt_{\dot x}^{\phantom i} w=Q^{-\nnh1\hskip-.7pt/2}\nabla\hskip-3pt_{v}^{\phantom i} w
=Q^{-\nnh1\hskip-.7pt/2}\nabla\hskip-3pt_{w}^{\phantom i} v=Q^{-\nnh1\hskip-.7pt/2}S\hskip-.7pt w$). With
$\,\varGamma\,$ replaced by suitable shorter sub\-geodesics covering all points of
$\,\varGamma\hskip-.7pt$, the inclusion just established between the two vector spaces
appearing in (i) is actually an equality: in either class, the vector field in
question is uniquely determined by its initial value at any given point
$\,x\in\varGamma\hskip-.7pt$. This proves (i) -- (ii) as well as (iii) -- (iv), the latter
in view of the fact that $\,JS=S\hskip-.7pt J$, cf.\ (\ref{loc}.a).
For a $\,C^1$ function $\,\lambda\,$ defined on the parameter interval of
$\,t\mapsto x(t)$, one has
\begin{equation}\label{dlt}
\dot\lambda\,=\,2(\psi-\lambda)\lambda\hskip.4pt Q^{-\nnh1\hskip-.7pt/2}\hskip12pt
\mathrm{with}\hskip7pt(\hskip2.3pt)\dot{\phantom o}\hskip-.7pt=d/dt
\end{equation}
if and only if either $\,\lambda=0\,$ identically, or $\,\lambda\ne0\,$
everywhere and the function $\,c$ characterized by
$\,2c=2\vt-Q/\lambda\,$ is constant. (In fact, the ei\-ther-or claim about
vanishing of $\,\lambda\,$ is due to uniqueness of solutions of
in\-i\-tial-val\-ue problems, while (\ref{dtq}) yields
$\,2\dotc
=Q\lambda^{-\nnh2}[\dot\lambda-2(\psi-\lambda)\lambda\hskip.4pt Q^{-\nnh1\hskip-.7pt/2}\hskip.4pt]$.)
Now (v) easily follows, all nonzero initial conditions for (\ref{dlt}) at
fixed $\,t\,$ being realized by suitably chosen constants
$\,c\in\mathrm{I\!R}}%{\mathbf{R}\smallsetminus\vt(\varGamma)\,$ (and $\,\lambda=0\,$ satisfying (v)
with $\,c=\infty$).
On the other hand, from (\ref{dxe}) and Lemma~\ref{dvgww}(e),
\begin{equation}\label{nds}
[\nabla\hskip-3pt_{\dot x}^{\phantom i}S]\hskip.4pt w\,
=\,2Q^{-\nnh1\hskip-.7pt/2}(\psi\hskip.7pt-\hsS)S\hskip.4pt w\hskip.4pt,
\hskip9pt\mathrm{if\ }\,w\,\mathrm{\ is\ a\
}\,\mathcal{V}^\perp\hskip-.7pt\hyp\mathrm{val\-ued\ vector\ field\ along\ }\,\varGamma.
\end{equation}
Next, we fix $\,x=x(t)\in\varGamma\,$ and express any prescribed
ei\-gen\-val\-ue-ei\-gen\-vec\-tor pair for
$\,S\hskip-1.5pt_x^{\phantom i}:\mathcal{V}^\perp_{\hskip-1ptx}\hskip-.7pt
\to\mathcal{V}^\perp_{\hskip-1ptx}$ as $\,\lambda_c^{\phantom i}(x)\,$ and $\,w(t)$,
with some unique $\,c\in[\mathrm{I\!R}}%{\mathbf{R}\smallsetminus\vt(\varGamma)]\cup\{\infty\}$ and
$\,w\in\mathcal{W}\hskip-.7pt$. By (v), $\,\lambda=\lambda_c^{\phantom i}$ satisfies
(\ref{dlt}), so that, in view of (\ref{nxw}) and (\ref{nds}), the vector field
$\,\hat w=S\hskip-.7pt w-\lambda w\,$ is a solution of the linear homogeneous
differential equation
$\,\nabla\hskip-3pt_{\dot x}^{\phantom i} \hat w=Q^{-\nnh1\hskip-.7pt/2}(2\psi-2\lambda-S)\hskip.4pt\hat w$.
Since $\,\hat w\,$ vanishes at $\,x=x(t)$, it must vanish identically, which
establishes (vi).
Now let $\,w\in\mathcal{W}\hskip-.7pt$. As $\,\dot Q=2\psi Q^{\hs1\hskip-.7pt/2}$ (see
the lines following (\ref{dlt})), the Ja\-co\-bi equation and (\ref{nxw})
give, by (ii) and (\ref{nds}),
$\,R(\hskip-.4pt w,\dot x)\hskip.4pt\dot x=\nabla\hskip-3pt_{\dot x}^{\phantom i}
\nabla\hskip-3pt_{\dot x}^{\phantom i} w
=\nabla\hskip-3pt_{\dot x}^{\phantom i}[Q^{-\nnh1\hskip-.7pt/2}S\hskip-.7pt w]
=Q^{-\nnh1}(\psi\hskip.7pt-\hsS)S\hskip.4pt w$,
that is, $\,R(\hskip-.4pt w,v)\hskip.4pt v=(\psi\hskip.7pt-\hsS)S\hskip.4pt w$, the second equality in
(vii). Also, Lemma~\ref{dvgww}(e), (\ref{loc}.c) and (\ref{rcm}) yield
$\,2(\psi\hskip.7pt-\hsS)S\hskip.4pt w=[\nabla\hskip-3pt_{v}^{\phantom i}S]\hskip.4pt w
=-\hskip-.7pt J[R(\hskip-.4pt u,v)\hskip.4pt w]=-\hskip.4pt R(\hskip-.4pt u,v)\hskip.4pt J\hskip-.7pt w=R(\hskip-.4pt v,u)\hskip.4pt J\hskip-.7pt w=R(\hskip-.4pt v,J\hskip-.7pt v)\hskip.4pt J\hskip-.7pt w$, the last
equality in (vii). Combining the two relations, and repeatedly using
(\ref{rcm}), we get $\,2R(\hskip-.4pt w,v)\hskip.4pt v=R(\hskip-.4pt v,J\hskip-.7pt v)\hskip.4pt J\hskip-.7pt w$, that is,
$\,R(\hskip-.4pt w,v)\hskip.4pt v=R(\hskip-.4pt v,w)\hskip.4pt v+R(\hskip-.4pt v,J\hskip-.7pt v)\hskip.4pt J\hskip-.7pt w=R(J\hskip-.7pt v,J\hskip-.7pt w)\hskip.4pt v+R(\hskip-.4pt v,J\hskip-.7pt v)\hskip.4pt J\hskip-.7pt w$. Thus,
from the Bian\-chi identity,
$\,R(\hskip-.4pt w,v)\hskip.4pt v=R(\hskip-.4pt v,J\hskip-.7pt w)\hskip.4pt J\hskip-.7pt v=R(J\hskip-.7pt v,JJ\hskip-.7pt w)\hskip.4pt J\hskip-.7pt v=R(\hskip-.4pt w,u)\hskip.4pt u$, which proves (vii).
Finally, (viii) is an easy consequence of (vi) and (\ref{loc}.d).
\end{proof}
\section{Consequences of compactness}\label{cc}
\setcounter{equation}{0}
Let $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ be a fixed ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler
triple (Definition~\ref{ggktr}). We use the notation of (\ref{sym}),
(\ref{tmm}) and -- in (i) below -- the terminology of Remark~\ref{cifot}.
\begin{remark}\label{ascdt}According to
\cite[Lemmas 11.1, 11.2 and Remark 2.1]{derdzinski-kp}, the following holds
for any $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ as above with compact $\,M}%{{\mathrm{M}}\hskip-.7pt$, the objects
(\ref{tmm}), and $\,v=\nabla\hskip-.5pt\vt$.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] $Q=g(\hskip-.4pt v,v)\,$ is a $\,C^\infty$ function of $\,\vt$,
leading to data $\,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a,Q\,$ with (\ref{pbd}).
\item[{\rm(ii)}] The flow of the Kil\-ling vector field $\,u=J\hskip-.7pt v\,$
is periodic.
\item[{\rm(iii)}] $\varSigma^\pm$ are (connected) totally geodesic compact complex
sub\-man\-i\-folds of $\,M}%{{\mathrm{M}}\hskip-.7pt$.
\item[{\rm(iv)}] $\varSigma^+\hskip-1.5pt\cup\varSigma^-$ is the zero set of $\,v$, that is, the
set of critical points of $\,\vt$.
\end{enumerate}
(Conclusion (iv) is a special case of a result due to Wang
\cite[Lemma 3]{wang}.) Furthermore, restricting $\,\vt\mapsto Q\,$ in (i) to
the open interval $\,(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)\,$ we have
\begin{equation}\label{dqt}
dQ/d\vt\,=\,2\psi\hskip.4pt,\hskip12pt\mathrm{and\ so}\hskip7pt\psi\to\mp a\hskip7pt
\mathrm{as}\hskip7pt\vt\to\vt\hskip-1.5pt_\pm\w\hskip.4pt,
\end{equation}
$\psi\,$ being the function with (\ref{nvv}) on the open set
$\,M}%{{\mathrm{M}}'$ on which $\,v\ne0\,$ (so that $\,\psi\,$ is also a $\,C^\infty$
function of $\,\vt$). This is clear as (\ref{tnd}) and (\ref{nvv}) give
$\,dQ=2\psi\,d\vt\,$ on $\,M}%{{\mathrm{M}}\hskip-.7pt'\hskip-.7pt$. Finally, by \cite[Lemma 1]{wang} (see
also \cite[Example 8.1 and Lemma 8.4(iv)]{derdzinski-kp}),
\begin{equation}\label{vtg}
v\,\mathrm{\ is\ tangent\ to\ every\ geodesic\ normal\ to\ }\,\varSigma^\pm\hskip-.7pt.
\end{equation}
\end{remark}
\begin{remark}\label{prcrm}Under the assumptions of Remark~\ref{ascdt}, for
$\,a\,$ as in Remark~\ref{ascdt}(i), $\,\mp\hskip.4pt a\,$ is the unique nonzero
eigen\-val\-ue of the Hess\-i\-an of $\,\vt\,$ (that is, of
$\,S=\nabla\hskip-.7pt v$) at any critical point $\,y\in\varSigma^\pm\hskip-.7pt$. The
$\,\mp\hskip.4pt a$-eigen\-space of $\,S\hskip-1.5pt_y^{\phantom i}$ is the normal space
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt$, and
$\,\mathrm{Ker}\hskip2.7ptS\hskip-1.5pt_y^{\phantom i}=T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ (which thus constitutes the
$\,0$-eigen\-space of $\,S\hskip-1.5pt_y^{\phantom i}$ unless $\,\varSigma^\pm\hskip-.7pt=\{y\}$).
In fact, as $\,\vt\,$ is a Morse-Bott function
\cite[Example 8.1]{derdzinski-kp}, applying
\cite[Lemma 8.4(i)]{derdzinski-kp} we see that $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm$
is the eigen\-space of $\,S\hskip-1.5pt_y^{\phantom i}$ for its unique nonzero eigen\-val\-ue,
and so $\,\mathrm{Ker}\hskip2.7ptS\hskip-1.5pt_y^{\phantom i}=T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ in view of
self-ad\-joint\-ness of $\,S\hskip-1.5pt_y^{\phantom i}$. That the nonzero eigen\-val\-ue
equals $\,\mp\hskip.4pt a\,$ is obvious from (\ref{dqt}), since (\ref{nvv}) amounts
to $\,S\hskip-.7pt v=\psi\hskip.4ptv$. Cf.\ \cite[Theorem 1.3]{miyaoka}.
\end{remark}
Still assuming compactness of a ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler
triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, let $\,N}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm$ be the bundle of radius
$\,\delta\,$ normal open disks around the zero section in the normal bundle
$\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-.7pt$, with $\,\delta\,$ characterized by (\ref{int}).
According to \cite[Lemma 10.3]{derdzinski-kp}, $\,\delta\,$ is then the
distance between $\,\varSigma^+$ and $\,\varSigma^-\hskip-.7pt$, while, with
$\,\mathrm{Exp}^\perp$ as in Remark~\ref{nexpm},
\begin{equation}\label{nex}
\begin{array}{l}
\mathrm{the\ restriction\ to\ }N}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm\mathrm{\,of\ the\
normal\ exponential\ mapping}\\
\mathrm{Exp}^\perp\hskip-1.5pt:N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip.7pt\to\,M}%{{\mathrm{M}}\,\mathrm{\ is\ a\
dif\-feo\-mor\-phism\
}\,N}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm\to\,M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp.
\end{array}
\end{equation}
Cf.\ \cite{bolton}, \cite[Lemma 2]{wang}, \cite[Theorem 1.1]{miyaoka}. Its
inverse $\,M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt\toN}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm\hskip-.7pt$,
composed with the projection $\,N}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm\hskip-.7pt\to\varSigma^\pm\hskip-.7pt$, yields
a new disk-bundle projection
\begin{equation}\label{dbp}
\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt\to\varSigma^\pm.
\end{equation}
\begin{remark}\label{ptnrs}Clearly, $\,\pi^\pm\hskip-1.5pt\circ\mathrm{Exp}^\perp$
is the nor\-mal-bun\-dle projection
$\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt\to\varSigma^\pm\hskip-1.5pt$. Also, according to the lines
preceding (\ref{dbp}),
\begin{equation}\label{img}
\begin{array}{l}
\mathrm{the\ image\ }\,\,\pi^\pm(x)\,\,\mathrm{\ of\ any\
}\,\,x\inM}%{{\mathrm{M}}'\,\mathrm{\ is\ the\ unique\ }\,\,y\in\varSigma^\pm\,\mathrm{\ that\
can}\\
\mathrm{be\ joined\ to\ }\,\,x\,\,\mathrm{\ by\ a\ (necessarily\ unique)\
geodesic\ segment\ }\,\,\varGamma\hskip-3.3pt_x^{\phantom i}\,\mathrm{\ of}\\
\mathrm{length\ less\ than\ }\hskip.7pt\delta\hskip.7pt\mathrm{\ emanating\ from\
}\,y\,\mathrm{\ in\ a\ direction\ normal\ to\ }\hskip.7pt\varSigma^\pm\hskip-1.5pt,
\end{array}
\end{equation}
which implies \cite[Remark 4.6, Example 8.1 and
Theorem 10.2(iii)\hskip.7pt--\hskip.7pt(vi)]{derdzinski-kp} that $\,\pi^\pm$ sends every
$\,x\inM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp$ to the unique point nearest $\,x\,$ in
$\,\varSigma^\pm\hskip-.7pt$.
\end{remark}
In the next lemma, by a {\medit leaf\hskip.7pt} we mean -- as usual -- a maximal
integral manifold.
\begin{lemma}\label{vsbkr}{\medit
Under the above hypotheses,
$\,\mathcal{V}\hskip.7pt\subseteq\,\mathrm{Ker}\hskip2.3ptd\pi^\pm$ for the
in\-te\-gra\-ble distribution\/ $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)\,$
on\/ $\,M}%{{\mathrm{M}}'\hskip-.7pt=M}%{{\mathrm{M}}\smallsetminus(\varSigma^+\hskip-1.5pt\cup\varSigma^-)$, cf.\
Lemma\/~{\rm\ref{dvgww}(a)} and Remark\/~{\rm\ref{ascdt}(iv)}. If\/ $\,\xi$
is a unit vector normal to\/ $\,\varSigma^\pm$ at a point\/ $\,y$, then, with\/
$\,\delta\,$ as in\/ {\rm(\ref{nex})},
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] the punctured radius\/ $\,\delta\,$ disk\/
$\,\{z\xi:z\in\bbC\hskip6pt\mathrm{and}\hskip6pt0<|z|<\delta\}\,$ in\/
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm$ is mapped by\/ $\,\exp_y^{\phantom i}\,$
dif\-feo\-mor\-phic\-al\-ly onto a leaf\/ $\,\varLambda\,\,$ of\/ $\,\mathcal{V}\hskip-.7pt$.
\end{enumerate}
Furthermore, every leaf\/ $\,\varLambda\hskip.7pt\subseteqM}%{{\mathrm{M}}'$ of\/ $\,\mathcal{V}\,$ has
the following properties.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(b)}] The closure of\/ $\,\varLambda\,\,$ in\/ $\,M}%{{\mathrm{M}}\,$ is a totally
geodesic complex sub\-man\-i\-fold, bi\-hol\-o\-mor\-phic to\/ $\,\bbC\mathrm{P}^1$ and
equal to\/ $\,\varLambda\hskip.7pt\cup\{y_+^{\phantom i},y_-^{\phantom i}\}$, where\/ $\,y_\pm^{\phantom i}\hskip-.7pt\in\varSigma^\pm$
are such that\/ $\,\{y_\pm^{\phantom i}\}=\pi^\pm\hskip-.7pt(\varLambda)$.
\item[{\rm(c)}] The leaf\/ $\,\varLambda\,\,$ arises from\/ {\rm(a)} \,for\/ some\/
unit normal vector\/ $\,\xi\,$ at the point\/ $\,y=y_\pm^{\phantom i}$ corresponding to\/
$\,\varLambda\hskip.7pt\,$ as in\/ {\rm(b)}, and then
\begin{equation}\label{pex}
\pi^\mp\hskip-.7pt(\mathrm{Exp}^\perp\hskip-.7pt(y_\pm^{\phantom i},z\xi))\,
=\,y_\mp^{\phantom i}\hskip12pt\mathrm{whenever}\hskip6ptz\in\bbC\hskip6pt\mathrm{and}
\hskip6pt0<|z|\le\delta\hskip.4pt,
\end{equation}
$\mathrm{Exp}^\perp$ being the normal exponential mapping\/
$\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma\toM}%{{\mathrm{M}}\hskip-.7pt$.
\end{enumerate}
}
\end{lemma}
\begin{proof}Let us fix $\,x,y\,$ and $\,\varGamma\hskip-3.3pt_x^{\phantom i}$ as in
(\ref{img}). Due to Remark~\ref{ascdt}(iv), the Kil\-ling field
$\,u=J\hskip-.7pt v\,$ vanishes along $\,\varSigma^\pm\hskip-.7pt$, so that its
infinitesimal flow at $\,y\,$ preserves both $\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ and
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt$. The images of $\,\varGamma\hskip-3.3pt_x^{\phantom i}$ under
the flow transformations of $\,u\,$ thus are geodesic segments normal to
$\,\varSigma^\pm$ emanating from $\,y\,$ and, as a consequence of (\ref{img}),
$\,\pi^\pm$ maps them all onto $\,\{y\}$. In other words, the union of such
segments, with the point $\,y$ removed, is simultaneously a subset of the
$\,\pi^\pm\hskip-.7pt$-pre\-im\-age of $\,y\,$ as well as -- according to (\ref{vtg}),
(\ref{gvv}) and parts (ii), (iv) of Remark~\ref{ascdt} -- a surface embedded
in $\,M}%{{\mathrm{M}}'\hskip-1.5pt$. This surface is, due to its very definition and (\ref{vtg}),
tangent to both $\,u\,$ and $\,v\,$ which, in view of (\ref{nex}), yields (a);
note that, by (\ref{loc}.a) and Remark~\ref{prcrm}, the orbit of $\,\xi\,$
under the flow of $\,A=\nabla\hskip-.7pt u\,$ at $\,y\,$ consists of all unit
complex multiples of $\,\xi$.
What we just observed about the orbit of $\,\xi\,$ clearly ensures smoothness
of the closure of the leaf at $\,y$. By (\ref{nex}) and (\ref{vtg}), the
union of $\,\varGamma\hskip-3.3pt_x^{\phantom i}$ and its analog for the same point $\,x\,$
and the {\medit other\/} projection $\,\pi^\mp$ is a length $\,\delta\,$
geodesic segment joining $\,y\in\varSigma^\pm$ to its other endpoint
$\,y_\mp^{\phantom i}\in\varSigma^\mp\hskip-.7pt$. The above discussion of the images of such a
segment under the flow of $\,u\,$ applies equally well to $\,y_\mp^{\phantom i}$, so that
(b) -- (c) follow from Lemma~\ref{dvgww}(a) and the fact
that $\,x\inM}%{{\mathrm{M}}'$ was arbitrary.
\end{proof}
\begin{remark}\label{ppmgc}Let $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ be a Grass\-mann\-i\-an or CP
triple, constructed as in Section~\ref{eg} from some data (\ref{dta}.i) or
(\ref{dta}.ii). We use the notation of (\ref{sym}) and (\ref{dbp}).
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] We already know that the critical manifolds $\,\varSigma^\pm$ of
$\,\vt\,$ are given by (\ref{spm}).
\item[{\rm(b)}] In the case of (\ref{spm}.c) (or, (\ref{spm}.b) and
(\ref{spm}.d)), $\,\pi^\pm$ acts on $\,\mathsf{W}\,$ as the orthogonal projection
into $\,\mathsf{L}\,$ (or, respectively, into $\,\mathsf{L}\hskip-2.5pt^\perp$)
\item[{\rm(c)}] When $\,\varSigma^\pm$ has the form (\ref{spm}.a), $\,\pi^\pm$
sends $\,\mathsf{W}\,$ to $\,\mathsf{L}\oplus(\mathsf{W}\hskip-.7pt\cap\mathsf{L}\hskip-2.5pt^\perp)$.
\item[{\rm(d)}] The leaf of $\,\mathcal{V}\,$ through any $\,\mathsf{W}\inM}%{{\mathrm{M}}'$
consists
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(d1)}] for (\ref{dta}.i) -- of all
$\,\mathsf{L}\hskip-2.5pt'\oplus\mathsf{W}'\hskip-.7pt$, where
$\,\mathsf{W}'\hskip-.7pt=\mathsf{W}\hskip-.7pt\cap\mathsf{L}\hskip-2.5pt^\perp$ and $\,\mathsf{L}\hskip-2.5pt'$ is any
line in the plane $\,\mathsf{L}\oplus(\mathsf{W}'\hskip-.7pt\cap\mathsf{W}^\perp)\,$ other than the lines
$\,\mathsf{L}\,$ and $\,\mathsf{W}'\hskip-.7pt\cap\mathsf{W}^\perp$ themselves,
\item[{\rm(d2)}] for (\ref{dta}.ii) -- of all lines other than $\,\mathsf{W}'$ and
$\,\mathsf{W}''$ in the plane $\,\mathsf{W}'\hskip-.7pt\oplus\mathsf{W}''\hskip-1.5pt$, where $\,\mathsf{W}'$ and
$\,\mathsf{W}''$ denote the orthogonal projections of $\,\mathsf{W}\,$ into $\,\mathsf{L}\,$
and $\,\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt$.
\end{enumerate}
\end{enumerate}
Namely, in both cases, let $\,G'$ be the complex Lie group of all
com\-plex-lin\-e\-ar au\-to\-mor\-phisms of $\,\mathsf{V}\,$ preserving both
$\,\mathsf{L}\,$ and $\,\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt$. The obvious action of $\,G'$ on
the Stie\-fel manifold $\,\mathrm{St}\hskip-.4pt_k^{\phantom i}\hskip-1.5pt\mathcal{T}\,$ (see
Remark~\ref{grass}) descends to a hol\-o\-mor\-phic action on
$\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\hskip-.7pt$, which becomes one on
$\,\mathrm{P}\mathsf{V}=\mathrm{G}\hskip-.4pt_1^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ when $\,k=1$. The elements of
the center of $\,G'\hskip-.7pt$, restricted to both sub\-spaces $\,\mathsf{L}\,$ and
$\,\mathsf{L}\hskip-2.5pt^\perp\hskip-1.5pt$, are complex multiples of $\,\mathrm{Id}$, and
the action of the center on $\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ includes
the circle subgroup $\,S^1$ of $\,G\,$ generated by the Kil\-ling field $\,u$,
mentioned in the lines following (\ref{dta}). Hol\-o\-mor\-phic\-i\-ty of the
action implies that the flow of the gradient $\,v=\nabla\hskip-.5pt\vt$, related to $\,u\,$
via $\,u=J\hskip-.7pt v$, also consists of transformations of
$\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}\,$ arising from the action of the center, and --
for dimensional reasons -- the orbits of the center coincide with the leaves
of $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)$. This easily gives (d). Now (b)
-- (c) follow: by Lemma~\ref{vsbkr}(b), the two
$\,\pi^\mp\hskip-2pt$-im\-a\-ges of any leaf of $\,\mathcal{V}\,$ are the two
points that, added to the leaf, yield its closure.
\end{remark}
As $\,\mathcal{V}\hskip.7pt\subseteq\,\mathrm{Ker}\hskip2.3ptd\pi^\pm$
(Lemma~\ref{vsbkr}), we may define vector sub\-bun\-dles
$\,\mathcal{H}^\pm\hskip-1.5pt$ of $\,T\nnhM}%{{\mathrm{M}}'$ by
\begin{equation}\label{hpm}
\mathcal{H}^\pm\hskip.7pt
=\,\,\mathcal{V}^\perp\hskip-.7pt\cap\,\mathrm{Ker}\hskip2.3ptd\pi^\mp,\hskip22pt
\mathrm{so\ that}\hskip12pt\mathrm{Ker}\hskip2.3ptd\pi^\pm\hskip.7pt
=\,\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\mp.
\end{equation}
\begin{theorem}\label{dcomp}{\medit
Given a ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$
with compact\/ $\,M}%{{\mathrm{M}}\hskip-.7pt$,
$\,v,M}%{{\mathrm{M}}'\hskip-.7pt,Q,\mathcal{V}\hskip-.7pt,S,\varSigma^\pm\hskip-.7pt,\vt\hskip-1.5pt_\pm\w,\pi^\pm\hskip-.7pt,\mathcal{H}^\pm$
be defined as in\/ {\rm(\ref{sym})}, {\rm(\ref{tmm})} and\/ {\rm(\ref{hpm})}.
Then the bundle en\-do\-mor\-phism\/ $\,2(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)S\hskip.7pt-\hskip.7pt Q\,$ of\/
$\,T\nnhM}%{{\mathrm{M}}\hskip-.7pt$, restricted to\/ $\,\mathcal{V}^\perp\hskip-.7pt$, has constant rank
on\/ $\,M}%{{\mathrm{M}}'\hskip-1.5pt$, while
\begin{equation}\label{hvk}
\mathcal{H}^\mp\,
=\,\,\mathcal{V}^\perp\hskip-.7pt\cap\hskip.7pt\mathrm{Ker}\hskip2pt
[2(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)S\hskip.7pt-\hskip.7pt Q]\hskip.4pt,
\end{equation}
and some sub\-bun\-dle\/ $\,\mathcal{H}\,$ of\/ $\,T\nnhM}%{{\mathrm{M}}'$ yields an\/
$\,S$-in\-var\-i\-ant complex orthogonal decomposition
\begin{equation}\label{tme}
T\nnhM}%{{\mathrm{M}}'\hskip.7pt=\,\,\mathcal{V}\hskip1.2pt\oplus\hskip.7pt\mathcal{H}^+\hskip-1.5pt
\oplus\hskip.7pt\mathcal{H}^-\hskip-1.5pt\oplus\hskip.7pt\mathcal{H}\hskip.4pt.
\end{equation}
Furthermore, for any\/ $\,\varGamma\hskip-.7pt\subseteqM}%{{\mathrm{M}}'$ chosen as at the beginning of
Section\/~{\rm\ref{hj}},the closure of\/ $\,\varGamma$ in\/ $\,M}%{{\mathrm{M}}\,$ admits a
unit-speed\/ $\,C^\infty\hskip-1.5pt$ parametrization\/
$\,[\hskip.7pt t_-^{\phantom i},t_+^{\phantom i}]\ni t\mapsto x(t)\,$ which, restricted
to\/ $\,(t_-^{\phantom i},t_+^{\phantom i})$, is a parametrization of\/ $\,\varGamma\hskip.7pt$ satisfying\/
{\rm(\ref{dxe})} along with the following conditions.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] The endpoint\/ $\,y_\pm^{\phantom i}\hskip-.7pt=x(t_\pm^{\phantom i})\,$ lies in\/
$\,\varSigma^\pm\hskip-.7pt$, and\/ $\,\dot x(t_\pm^{\phantom i})\,$ is normal to\/ $\,\varSigma^\pm$ at\/
$\,y_\pm^{\phantom i}$.
\item[{\rm(b)}] Every solution\/
$\,(t_-^{\phantom i},t_+^{\phantom i})\ni t\mapsto w(t)\in\mathcal{V}^\perp_{\hskip-1ptx(t)}$ of\/
{\rm(\ref{nxw})} along\/ $\,\varGamma\hskip.7pt$ has a\/ $\,C^\infty\hskip-1.5pt$ extension to\/
$\,[\hskip.7pt t_-^{\phantom i},t_+^{\phantom i}]\,$ such that\/ $\,d\pi^\pm_{x(t)}[w(t)]=w(t_\pm^{\phantom i})\,$
whenever\/ $\,t\in(t_-^{\phantom i},t_+^{\phantom i})$.
\item[{\rm(c)}] The bundle projection\/
$\,\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt\to\varSigma^\pm$ is
hol\-o\-mor\-phic.
\item[{\rm(d)}] If\/ $\,w\in\mathcal{W}[\vt\hskip-1.5pt_\pm\w]$, cf.\ {\rm(\ref{wze})}, then,
in\/ {\rm(b)},
$\,w(t_\pm^{\phantom i})=0$, and\/ $\,[\nabla\hskip-3pt_{\dot x}^{\phantom i} w](t_\pm^{\phantom i})\,$
is normal to\/ $\,\varSigma^\pm$ at\/ $\,y_\pm^{\phantom i}\hskip-.7pt=x(t_\pm^{\phantom i})\,$ as well as
orthogonal to\/ $\,\dot x(t_\pm^{\phantom i})\,$ and\/ $\,J\dot x(t_\pm^{\phantom i})$.
\item[{\rm(e)}] If\/ $\,w\,$ lies in the direct sum of spaces\/
$\,\mathcal{W}[c]\ne\{0\}\,$ with\/ $\,c\ne\vt\hskip-1.5pt_\pm\w\hskip.4pt$, for a fixed sign\/
$\,\pm\hskip.7pt$, then\/ $\,w(t_\pm^{\phantom i})\,$ is tangent to\/ $\,\varSigma^\pm$ at\/
$\,y_\pm^{\phantom i}\hskip-.7pt=x(t_\pm^{\phantom i})$, and\/
$\,[\nabla\hskip-3pt_{\dot x}^{\phantom i} w](t_\pm^{\phantom i})=0$.
\item[{\rm(f)}] Whenever\/ $\,t\in(t_-^{\phantom i},t_+^{\phantom i})\,$ and\/ $\,x=x(t)$, the
assignment\/
$\,w(t)\mapsto(\hskip-.4pt w(t_\pm^{\phantom i}),\,[\nabla\hskip-3pt_{\dot x}^{\phantom i} w](t_\pm^{\phantom i}))$, with\/
$\,w\,$ as in\/ {\rm(b)}, is a\/ $\,\bbC$-lin\-e\-ar isomorphism\/
$\,\mathcal{V}^\perp_{\hskip-1ptx}\toT\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-2.3pt\times\nhN}%{{\mathrm{N}}'_y\hskip.7pt$,
where\/ $\,y=y_\pm^{\phantom i}$ and\/ $\,N}%{{\mathrm{N}}'_y\,$ denotes the orthogonal complement
of\/ $\,\mathrm{Span}\hskip.7pt(\dot x(t_\pm^{\phantom i}),\hskip.7pt J\dot x(t_\pm^{\phantom i}))\,$ in\/
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt$. At the same time, $\,w(t)\,$ then equals
the image, under the differential
of the normal exponential mapping\/
$\,\mathrm{Exp}^\perp\hskip-1.5pt:N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt\toM}%{{\mathrm{M}}\,$ at\/
$\,(y,\xi)\inN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm$ given by\/ $\,y=x(t_\pm^{\phantom i})\,$ and\/
$\,\xi=(t-t_\pm^{\phantom i})\hskip.4pt\dot x(t_\pm^{\phantom i})$, of the vector tangent to\/
$\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm$ at\/ $\,(y,\xi)\,$ which equals the sum of the
vertical vector\/
$\,\eta=(t-t_\pm^{\phantom i})[\nabla\hskip-3pt_{\dot x}^{\phantom i} w](t_\pm^{\phantom i})\,$ and
the\/ $\,\mathrm{D}\hskip-.4pt$-\hskip-.4pt hor\-i\-zon\-tal lift of\/ $\,w(t_\pm^{\phantom i})\,$ to\/
$\,(y,\xi)$, for the normal connection\/ $\,\hskip.7pt\mathrm{D}\,$ in\/
$\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt$. Similarly, $\,u_{x(t)}^{\phantom i}$, for
$\,u=J\hskip-.7pt v$, is the image, under the differential of\/
$\,\mathrm{Exp}^\perp$ at\/ $\,(y,\xi)$, of the vertical vector\/
$\,\eta=\mp\hskip.4pt ai\hskip.4pt\xi$.
\item[{\rm(g)}] For any\/ $\,w,w\hskip.4pt'\hskip-.7pt\in\mathcal{W}\hskip-.7pt$, the function\/
$\,Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\,$ is constant on\/ $\,\varGamma\hskip.7pt$ and the
restriction of\/ $\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ to\/ $\,\varGamma\hskip.7pt$ is an af\-fine
function of\/ $\,\vt:\varGamma\hskip-.7pt\to\mathrm{I\!R}}%{\mathbf{R}\,$ with the derivative\/
$\,d\hskip.7pt[\hskip.7pt g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)]/\hskip-.7pt d\vt=2\hskip.4pt Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)$.
\item[{\rm(h)}] Explicitly, in\/ {\rm(g)}, with\/ $\,a\,$ as in
Remark\/~{\rm\ref{ascdt}(i)}, either sign\/ $\,\pm\hskip.7pt$, and\/
$\,y=y_\pm^{\phantom i}\hskip-.7pt=x(t_\pm^{\phantom i})$,
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(h1)}] $g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)
=(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)^{-\nnh1}|\hskip.4pt\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w|\hskip.7pt g_y^{\phantom i}(\hskip-.4pt w_\pm^{\phantom i},w_\pm'\hskip.4pt)\,$ if\/
$\,w\in\mathcal{W}[\vt\hskip-1pt_\mp\w]\,$ and\/ $\,w\hskip.4pt'\hskip-.7pt\in\mathcal{W}$,
\item[{\rm(h2)}] $g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)=g_y^{\phantom i}(\hskip-.4pt w_\pm^{\phantom i},w_\pm'\hskip.4pt)
-a^{-\nnh1}|\hskip.4pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|\,g_y^{\phantom i}(R_y^{\phantom i}(\hskip-.4pt w_\pm^{\phantom i},J\hskip-2pt_y^{\phantom i} w_\pm'\hskip.4pt)\hskip.4pt
\dot x\hskip-.7pt_\pm^{\phantom i},J\hskip-2pt_y^{\phantom i}\dot x\hskip-.7pt_\pm^{\phantom i})\,$ if\/ $\,w,w\hskip.4pt'$ both satisfy the
assumption of\/ {\rm(e)},
\item[{\rm(h3)}] $g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)
=2\hskip.4pt a^{-\nnh1}|\hskip.4pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|\,g_y^{\phantom i}([\nabla\hskip-3pt_{\dot x}^{\phantom i} w]_\pm^{\phantom i},
[\nabla\hskip-3pt_{\dot x}^{\phantom i} w\hskip.4pt'\hskip.4pt]_\pm^{\phantom i})\,$ if\/
$\,w,w\hskip.4pt'\hskip-.7pt\in\mathcal{W}[\vt\hskip-1.5pt_\pm\w]$,
\end{enumerate}
\end{enumerate}
where the subscript\/ $\,\pm\,$ next to\/
$\,w,w\hskip.4pt'\hskip-.7pt,\nabla\hskip-3pt_{\dot x}^{\phantom i} w,\nabla\hskip-3pt_{\dot x}^{\phantom i} w\hskip.4pt'$
or\/ $\,\dot x\,$ represents their evaluation at\/ $\,t_\pm^{\phantom i}$.
}
\end{theorem}
\begin{remark}\label{bthex}Since $\,|\hskip.4pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|=\mp\hskip.4pt(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)\,$
and $\,\pm\hskip.4pt(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)=\vt\hskip-1.5pt_\pm\w\hskip-.7pt-\vt\hskip-1pt_\mp\w$, applying $\,d/\hskip-.7pt d\vt\,$ to the
right-hand side in (h1), or (h2), or (h3), we get the three values
\[
(\vt\hskip-1.5pt_\pm\w\hskip-.7pt-\vt\hskip-1pt_\mp\w)^{-\nnh1}\hskip-.7pt g_y^{\phantom i}(\hskip-.4pt w_\pm^{\phantom i},\hskip-.7pt w_\pm'\hskip.4pt),\hskip3.6pt
\pm\hskip.4pt a^{-\nnh1}\hskip-.7pt g_y^{\phantom i}(R_y^{\phantom i}(\hskip-.4pt w_\pm^{\phantom i},\hskip-.7pt J\hskip-2pt_y^{\phantom i} w_\pm'\hskip.4pt)\hskip.4pt
\dot x\hskip-.7pt_\pm^{\phantom i},\hskip-.7pt J\hskip-2pt_y^{\phantom i}\dot x\hskip-.7pt_\pm^{\phantom i})\hskip.4pt,\hskip3.5pt
\mp2\hskip.4pt a^{-\nnh1}\hskip-.7pt g_y^{\phantom i}([\nabla\hskip-3pt_{\dot x}^{\phantom i} w]_\pm^{\phantom i},
\hskip-.7pt[\nabla\hskip-3pt_{\dot x}^{\phantom i} w\hskip.4pt'\hskip.4pt]_\pm^{\phantom i}).
\]
As a consequence of parts (g) -- (h) of Theorem~\ref{dcomp}, this triple
provides the three expressions for
$\,2\hskip.4pt Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\,$ in the cases (h1), (h2) and (h3),
respectively.
Note that the three different formulae for $\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ in (h1),
(h2) and -- with the reversed sign -- in (h3), are all simultaneously valid
when $\,w,w\hskip.4pt'\hskip-.7pt\in\mathcal{W}[\vt\hskip-1pt_\mp\w]$.
\end{remark}
\begin{remark}\label{inequ}Under the assumptions of Theorem~\ref{dcomp},
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] the relation $\,\xi=(t-t_\pm^{\phantom i})\hskip.4pt\dot x(t_\pm^{\phantom i})\,$ in (f)
clearly gives $\,\dot x\hskip-.7pt_\pm^{\phantom i}=\mp\hskip.7pt\xi/\hskip-.4pt|\hskip.4pt\xi|\,$ in (h2),
\item[{\rm(ii)}] by (d) -- (f), the images under the differential of
$\,\mathrm{Exp}^\perp$ of vertical (or, horizontal) vectors
tangent to $\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm$ at the point $\,(y,\xi)\,$ appearing in
(f) have the form
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm($*$)}] $w(t)\,$ for $\,w\,$ satisfying the hypothesis of (d) (or,
respectively, of (e)),
\end{enumerate}
\item[{\rm(iii)}] the differential of $\,\pi^{\hskip.7pt\pm}$ at any $\,x\inM}%{{\mathrm{M}}'$
maps the summands $\,\mathcal{H}_x^\pm$ and $\,\mathcal{H}_x^{\phantom i}$ in (\ref{tme})
iso\-mor\-phic\-al\-ly onto the images
$\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)$ and
$\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^{\phantom i})$, orthogonal to each other in
$\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ for $\,y=\pi^\pm\hskip-.7pt(x)$,
\item[{\rm(iv)}] one has $\,(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)\hskip.7pt g_x^{\phantom i}(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)
=|\hskip.4pt\vt(x)-\hskip-.4pt\vt\hskip-1pt_\mp\w|\hskip.7pt
g_y^{\phantom i}(d\pi\hskip-.7pt_x^{\hskip.7pt\pm}w,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}w\hskip.4pt'\hskip.4pt)\,$ whenever
$\,w\in\mathcal{H}_x^\pm$ and $\,w\hskip.4pt'\hskip-.7pt\in\mathcal{V}^\perp_{\hskip-1ptx}$
at any $\,x\inM}%{{\mathrm{M}}'\hskip-.7pt$, while $\,y=\pi^\pm\hskip-.7pt(x)$,
\item[{\rm(v)}] (\ref{tmm}) and (a) imply the inequality of
Theorem~\ref{jacob}(viii) everywhere in $\,M}%{{\mathrm{M}}'\hskip-.7pt$.
\end{enumerate}
Only (iii) and (iv) require further explanations. For (iii),
$\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}$ is injective on the space
$\,\mathcal{H}_x^\pm\hskip-.7pt\oplus\mathcal{H}_x$, orthogonal, by (\ref{hpm}) and
(\ref{tme}), to its kernel
$\,\mathcal{V}_{\hskip-.7pt x}^{\phantom i}\oplus\mathcal{H}_x^\mp\hskip-.7pt$. Or\-thog\-o\-nal\-i\-ty
in (\ref{hpm}) also shows, via (\ref{hvk}), (\ref{wze}.b) and (\ref{loc}.d),
that vectors in $\,\mathcal{H}_x^\pm$ (or, in $\,\mathcal{H}_x$) have the form
($*$) with $\,x=x(t)$, cf.\ (f), and the former remain orhogonal to the latter
as $\,t\,$ varies, leading to (iii) as a consequence of the final clause of
(b). Assertion (iv) is nothing else than (h1) for $\,w=w(t)\,$ at $\,x=x(t)$,
cf.\ (f), where $\,y=\pi^\pm\hskip-.7pt(x)\,$ and $\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}w=w_\pm^{\phantom i}$ by
(\ref{img}) and (b).
\end{remark}
\begin{remark}\label{holom}As another immediate consequence of
Theorem~\ref{dcomp}, the assignment
$\,x\mapsto d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)
=d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{V}_{\hskip-.7pt x}^{\phantom i}\oplus\mathcal{H}_x^\pm\hskip-.7pt)\,$
defines a hol\-o\-mor\-phic section of the bundle over $\,M}%{{\mathrm{M}}'$ arising via
the pull\-back under $\,\pi^\pm$ from
$\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.4pt(T\hskip-1.5pt\varSigma^\pm\hskip-.7pt)$, for a suitable integer
$\,k=k_\pm^{\phantom i}$. Here
$\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.4pt(T\hskip-1.5pt\varSigma^\pm\hskip-.7pt)\,$ is the Grass\-mann\-i\-an bundle
with the fibres $\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)$, $\,y\in\varSigma^\pm$ (cf.\
Section~\ref{eg}), hol\-o\-mor\-phic\-i\-ty and the equality
$\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)
=d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{V}_{\hskip-.7pt x}^{\phantom i}\oplus\mathcal{H}_x^\pm\hskip-.7pt)\,$ are
clear from Theorem~\ref{dcomp}(c) (which also implies, due to (\ref{hpm}),
that $\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\pm$ is a hol\-o\-mor\-phic
sub\-bun\-dle of $\,T\nnhM}%{{\mathrm{M}}'$) and (\ref{tme}) (which, combined with
(\ref{hpm}), ensures constancy of the dimension $\,k=k_\pm^{\phantom i}$ of the spaces
$\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)$).
\end{remark}
\section{Proof of Theorem~\ref{dcomp}}\label{pt}
\setcounter{equation}{0}
We begin by establishing (a) - (f) under the stated assumptions about
$\,\varGamma\hskip-.7pt$.
Let $\,(t_-^{\phantom i},t_+^{\phantom i})\mapsto x(t)\,$ be a parametrization of $\,\varGamma\hskip.7pt$ with
(\ref{dxe}). As $\,\vt\,$ then is clearly an increasing function of $\,t$, it
has some limits $\,\hat\vt\hskip-1.6pt_\pm\w$ as $\,t\to t_\pm^{\phantom i}$, finite due to boundedness of
$\,\vt$. The length of $\,\varGamma\hskip.7pt$ obviously equals the integral of
$\,Q^{-\nnh1\hskip-.7pt/2}$ over $\,(\hat\vt\hskip-1.7pt_-\w,\hat\vt\hskip-1.8pt_+\w)\subseteq(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)$, and so it is finite in
view of (\ref{int}). This implies the existence of limits $\,x(t_\pm^{\phantom i})\,$ of
$\,x(t)\,$ as $\,t\to t_\pm^{\phantom i}$. Furthermore, each $\,x(t_\pm^{\phantom i})\,$ lies in
$\,\varSigma^\pm$ since, if one $\,x(t_\pm^{\phantom i})\,$ did not, Remark~\ref{ascdt}(iv)
would yield $\,v\ne0\,$ at $\,x(t_\pm^{\phantom i})$, contradicting maximality of
$\,\varGamma\hskip-.7pt$. Thus, $\,[\hskip.7pt t_-^{\phantom i},t_+^{\phantom i}]\mapsto x(t)\,$ parametrizes the closure
of $\,\varGamma\hskip.7pt$. Next, $\,M}%{{\mathrm{M}}\smallsetminus(\varSigma^+\hskip-1.5pt\cup\varSigma^-)\,$ is, by
(\ref{nex}) and (\ref{vtg}), a disjoint union of maximal integral curves of
$\,v$, each of which has two limit points, one in $\,\varSigma^-$ and one in
$\,\varSigma^+\hskip-.7pt$, and the corresponding limit directions of the curve are normal
to $\,\varSigma^-$ and $\,\varSigma^+\hskip-.7pt$. Since $\,\varGamma\hskip.7pt$ is one of these curves, (a)
follows.
In (b), a $\,C^\infty$ extension to $\,[\hskip.7pt t_-^{\phantom i},t_+^{\phantom i}]\,$ must exist as
$\,w\,$ is a Ja\-co\-bi field; see Theorem~\ref{jacob}(ii). To obtain (d) --
(e), we fix $\,w\in\mathcal{W}[c]$, so that, from (\ref{lgm}) -- (\ref{wze}),
\begin{equation}\label{swe}
\mathrm{i)}\hskip6ptS\hskip-.7pt w\,=\,Qw/[2(\vt-c)]\hskip.4pt,\hskip22pt
\mathrm{ii)}\hskip6pt\nabla\hskip-3pt_{\dot x}^{\phantom i} w\,=\,Q^{\hs1\hskip-.7pt/2}\hskip.7pt w/[2(\vt-c)]
\hskip.4pt.
\end{equation}
Let $\,y=x(t_\pm^{\phantom i})\,$ and $\,w_y^{\phantom i}=w(t_\pm^{\phantom i})$. By
Remark~\ref{ascdt}(iv), $\,Q=\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w=0\,$ on $\,\varSigma^\pm$ while, in view of
(a) and (\ref{dqt}), $\,Q/[2(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)]\,$ evaluated at $\,x(t)\,$ tends to
$\,\mp a\ne0\,$ as $\,t\to t_\pm^{\phantom i}$. If $\,c=\vt\hskip-1.5pt_\pm\w$, (\ref{swe}.ii)
multiplied by $\,Q^{\hs1\hskip-.7pt/2}$ thus yields $\,w_y^{\phantom i}=0$, and the relation
$\,S\hskip-.7pt w\hskip.4pt'\hskip-.7pt=Qw\hskip.4pt'\hskip-.7pt/[2(\vt-c)]\,$ for
$\,w\hskip.4pt'\hskip-.7pt=\nabla\hskip-3pt_{\dot x}^{\phantom i} w$, obvious from (\ref{swe}), implies
that $\,[\nabla\hskip-3pt_{\dot x}^{\phantom i} w](t_\pm^{\phantom i})$ lies in the
$\,\mp\hskip.4pt a$-eigen\-space of $\,S\hskip-1.5pt_y^{\phantom i}$. When $\,c\ne\vt\hskip-1.5pt_\pm\w$, (\ref{swe}.i)
and (\ref{swe}.ii) give, respectively, $\,S\hskip-1.5pt_y^{\phantom i} w_y^{\phantom i}=0\,$ and
$\,[\nabla\hskip-3pt_{\dot x}^{\phantom i} w](t_\pm^{\phantom i})=0$. Due to Remark~\ref{prcrm},
this proves (d) and (e): or\-thog\-o\-nal\-i\-ty in (d) follows since $\,w\,$
and $\,\nabla\hskip-3pt_{\dot x}^{\phantom i} w\,$ take values in
$\,\mathcal{V}^\perp\hskip-.7pt$, for $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)$ (so
that $\,g(\hskip-.4pt w,v)=g(\hskip-.4pt w,u)=0$), while $\,\dot x=v/|v|\,$ by (\ref{dxe}),
and $\,u=J\hskip-.7pt v$.
Furthermore, the assignment in (f) is well-defined, injective,
com\-plex-lin\-e\-ar and $\,(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-2.3pt\times\nhN}%{{\mathrm{N}}'_y)$-val\-ued
due to parts (iii), (ii), (iv) of Theorem~\ref{jacob} and, respectively, (d)
-- (e). The first claim of (f) thus follows since both spaces have the same
dimension. The second (or, third) one is in turn immediate from (\ref{dxp})
applied, at $\,r=1$, to any $\,w\in\mathcal{W}\hskip-.7pt$, cf.\
Theorem~\ref{jacob}(ii) (or, to $\,w=u$), with $\,y,\xi,\eta\,$ as in (f), and
$\,\hat w\,$ defined by $\,\hat w(\hskip-.4pt r\hskip-.7pt)=w(rt+(1-r)\hskip.4pt t_\pm^{\phantom i})$. (That
$\,r\mapsto\hat w(\hskip-.4pt r\hskip-.7pt)\,$ then is a Ja\-cobi field along the geodesic
$\,r\mapsto x(rt+(1-r)\hskip.4pt t_\pm^{\phantom i})\,$ follows from Theorem~\ref{jacob}(ii) or,
respectively, (\ref{loc}.b) and Remark~\ref{kiljc}, while, in the latter case,
due to (\ref{loc}.a) along with Remarks~\ref{ascdt}(iv) and~\ref{prcrm},
$\,w=u\,$ realizes the initial conditions
$\,(\hskip-.4pt u,\hskip.7pt\nabla\hskip-3pt_{dx\hskip-.4pt/\hskip-.4pt dr}^{\phantom i} u)=(0,\mp\hskip.4pt ai\hskip.4pt\xi)\,$ at
$\,r=0$.)
The remaining equality $\,d\pi^\pm_{x(t)}[w(t)]=w(t_\pm^{\phantom i})\,$ in (b) now
becomes an obvious consequence of the second part of (f) combined with the
first line of Remark~\ref{ptnrs}. This proves (b) and, combined with
Theorem~\ref{jacob}(iv), implies (c).
Next, for $\,t\mapsto x(t)\,$ as in (a) -- (f), any $\,t\in(t_-^{\phantom i},t_+^{\phantom i})$, a
fixed sign $\,\pm\hskip.4pt$, and $\,x=x(t)$, Theorem~\ref{jacob}(iii), (\ref{hpm})
and (b) give
$\,\mathcal{H}^\mp_x
=\{w(t):w\in\mathcal{W}\hskip.7pt\,\,\mathrm{and}\,\,w(t_\pm^{\phantom i})=0\}$. Writing any
$\,w\in\mathcal{W}\,$ as $\,w=w\hskip.4pt'\hskip-.7pt+w\hskip.4pt''\hskip-.7pt$, where
$\,w\hskip.4pt'\hskip-.7pt\in\mathcal{W}[\vt\hskip-1.5pt_\pm\w]\,$ and $\,w\hskip.4pt''$ lies in the direct sum of
the spaces $\,\mathcal{W}[c]\ne\{0\}\,$ with\/ $\,c\ne\vt\hskip-1.5pt_\pm\w\hskip.4pt$,
cf.\ Theorem~\ref{jacob}(iv), we see that, by (d) -- (e), the isomorphism in
(f) sends $\,w\hskip.4pt'\hskip-.7pt(t)\,$ and $\,w\hskip.4pt''\hskip-.7pt(t)$, respectively, to pairs of
the form $\,(0,\,\cdot\,)\,$ and $\,(\,\cdot\,,0)$. Thus,
$\,w(t)\in\mathcal{H}^\mp_x$ if and only if $\,w\hskip.4pt''\hskip-.7pt=0$, that is,
$\,w\in\mathcal{W}[\vt\hskip-1.5pt_\pm\w]$. Combining Theorem~\ref{jacob}(vi) with (\ref{lgm})
and (\ref{wze}.b), one now obtains (\ref{hvk}), so that (\ref{hpm}) implies
the con\-stant-rank assertion preceding (\ref{hvk}). On the other hand,
$\,\mathcal{H}^+_x$ and $\,\mathcal{H}^-_x$ are mutually orthogonal at every
$\,x\inM}%{{\mathrm{M}}'\hskip-.7pt$, being, by (\ref{hvk}), contained in eigen\-spaces
corresponding to different eigen\-values of the self-ad\-joint operator
$\,S\hskip-1.5pt_x^{\phantom i}$, cf.\ (\ref{loc}.d), so that (\ref{tme}) follows.
Let $\,w,w\hskip.4pt'\hskip-.7pt\in\mathcal{W}\hskip-.7pt$. Constancy of
$\,Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\,$ along $\,\varGamma\hskip.7pt$ trivially follows from
(\ref{dvg}.iii) and (\ref{vtg}), cf.\ Lemma~\ref{dvgww}(b) and parts (i) --
(ii) of Theorem~\ref{jacob}. The operators $\,d/\hskip-.7pt d\vt\,$ and $\,d_v^{\phantom i}$
acting on functions $\,\varGamma\hskip-.7pt\to\mathrm{I\!R}}%{\mathbf{R}\,$ are in turn related by
$\,d_v^{\phantom i}\hskip-.7pt=Q\,d/\hskip-.7pt d\vt$, since (\ref{dxe}) gives
$\,d_v^{\phantom i}\hskip-.7pt=Q^{\hs1\hskip-.7pt/2}d_{\dot x}^{\phantom i}\hskip-.7pt=Q^{\hs1\hskip-.7pt/2}d/\hskip-.7pt dt$, while
$\,d/\hskip-.7pt dt=Q^{\hs1\hskip-.7pt/2}d/\hskip-.7pt d\vt\,$ due to (\ref{dtq}). Now (g) is
immediate from (\ref{dvg}.ii).
In (h), all three right-hand sides are af\-fine functions of $\,\vt\,$ with
the correct values at $\,t=t_\pm^{\phantom i}$ (that is, limits at the endpoint
$\,y_\pm^{\phantom i}\hskip-.7pt=x(t_\pm^{\phantom i})$). Proving (h) is thus reduced by (g) to showing
that, in each case, $\,\chi=2\hskip.4pt Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\,$ coincides
with the derivative of the right-hand side provided by Remark~\ref{bthex},
which -- even though $\,\chi\,$ is constant on $\,\varGamma\hskip-.7pt$, cf.\ (g) -- will be
achieved via evaluating the limit of $\,\chi\,$ at $\,y_\pm^{\phantom i}\hskip-.7pt\in\varGamma\hskip.7pt$ or,
equivalently, at $\,t_\pm^{\phantom i}\hskip-.7pt\in[\hskip.7pt t_-^{\phantom i},t_+^{\phantom i}]$. When
$\,w\in\mathcal{W}[\vt\hskip-1pt_\mp\w]$, (\ref{wze}.b) and (\ref{lgm}) imply that
$\,2\hskip.4pt Q^{-\nnh1}S\hskip-.7pt w=(\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w)^{-\nnh1}w\,$ and, consequently,
$\,\chi=(\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w)^{-\nnh1}g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ has the value (and limit)
$\,\pm\hskip.4pt(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)^{-\nnh1}g_y^{\phantom i}(\hskip-.4pt w_\pm^{\phantom i},w_\pm'\hskip.4pt)\,$ at
$\,y=y_\pm^{\phantom i}$, as required in (h1).
Let $\,w,w\hskip.4pt'$ now satisfy the hypotheses of (e). Consequently, along
$\,\varGamma\hskip-.7pt\smallsetminus\{y_\pm^{\phantom i}\}$,
\begin{equation}\label{ttz}
Q\hskip.4pt,\hskip9ptS\hskip-.7pt w\hskip.4pt',\hskip9ptQ^{-\nnh1}g(S\hskip-.7pt w,S\hskip-.7pt w\hskip.4pt'\hskip.4pt)\hskip13pt
\mathrm{all\ tend\ to}\hskip6pt0\hskip6pt\mathrm{at}\hskip6pty\hskip.4pt,\hskip6pt
\mathrm{where}\hskip6pty=y_\pm^{\phantom i}\hskip.4pt.
\end{equation}
In fact, $\,Q(y)=0\,$ by (a). Next, $\,Q^{-\nnh1}S\hskip-.7pt w\,$ is bounded near the
endpoint $\,y\,$ of $\,\varGamma\hskip-.7pt\smallsetminus\{y\}$ (and similarly for
$\,w\hskip.4pt'\hskip.4pt$); to see this, we may assume that
$\,w\in\mathcal{W}[c]\,$ with $\,c\ne\vt\hskip-1.5pt_\pm\w$, cf.\ (e), and then
(\ref{wze}.b) and (\ref{lgm}) give
$\,2\hskip.4pt Q^{-\nnh1}S\hskip-.7pt w=(\vt-c)^{-\nnh1}w$, which is bounded as
$\,\vt\to\vt\hskip-1.5pt_\pm\w$ since, due to (b), $\,w\,$ has a limit at $\,t=t_\pm^{\phantom i}$. Now
(\ref{ttz}) follows.
In view of (\ref{ttz}) and (b), we may now evaluate the limit of
$\,\chi=2\hskip.4pt Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)$ as $\,t\to t_\pm^{\phantom i}$ using
l'H\^opital's rule: it coincides with the limit of
$\,2\hskip.7pt d_{\dot x}^{\phantom i}[\hskip.7pt g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)]/\dot Q$. By (\ref{loc}.c) for
$\,\dot x\,$ rather than of $\,w$, (\ref{nxw}), (\ref{loc}.d) and (\ref{dtq}),
this last expression is the sum of two terms,
$\,\psi^{-\nnh1}Q^{-\nnh1\hskip-.7pt/2}g(R(\hskip-.4pt u,\dot x)\hskip.4pt w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)\,$ and
$\,2\psi^{-\nnh1}Q^{-\nnh1}g(S\hskip-.7pt w,S\hskip-.7pt w\hskip.4pt'\hskip.4pt)$. According to
(\ref{ttz}) and (\ref{dqt}), only the first term contributes to the limit and,
as it equals $\,\psi^{-\nnh1}g(R(J\dot x,\dot x)\hskip.4pt w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)$, cf.\
(\ref{gvv}) and (\ref{vtg}), relation (\ref{dqt}) yields (h2).
Finally, suppose that $\,w,w\hskip.4pt'\hskip-.7pt\in\mathcal{W}[\vt\hskip-1.5pt_\pm\w]$. It follows that
\begin{equation}\label{qoh}
Q^{-\nnh1\hskip-.7pt/2}w\,
\to\,\mp\hskip.4pt a^{-\nnh1}[\nabla\hskip-3pt_{\dot x}^{\phantom i} w]_\pm^{\phantom i}\hskip13pt
\mathrm{at}\hskip6pty\hskip.4pt,\hskip6pt\mathrm{where}\hskip6pty=y_\pm^{\phantom i}\hskip.4pt,
\end{equation}
and analogously for $\,w\hskip.4pt'\hskip-.7pt$. Namely, $\,Q\,$ and $\,w\,$ vanish at
$\,y\,$ (see (a), (d)), while
$\,(Q^{\hs1\hskip-.7pt/2})\hskip-1.5pt\dot{\phantom o}\hskip-1.5pt=\psi\,$ by (\ref{dtq}), and so
$\,[\nabla\hskip-3pt_{\dot x}^{\phantom i} w]/(Q^{\hs1\hskip-.7pt/2})\hskip-1.5pt\dot{\phantom o}\hskip-1.5pt
=\psi^{-\nnh1}\nabla\hskip-3pt_{\dot x}^{\phantom i} w$. L'H\^opital's rule and (\ref{dqt}) now
imply (\ref{qoh}). Since
$\,S\hskip-1.5pt_y^{\phantom i}[\nabla\hskip-3pt_{\dot x}^{\phantom i} w]_\pm^{\phantom i}\hskip-.7pt
=\mp\hskip.4pt a\hskip.4pt[\nabla\hskip-3pt_{\dot x}^{\phantom i} w]_\pm^{\phantom i}$ by (d) and
Remark~\ref{prcrm}, assertion (h3) is obvious from (\ref{qoh}), completing the
proof of Theorem~\ref{dcomp}.
\begin{remark}\label{dppdm}With the same notations and assumptions as in
Theorem~\ref{dcomp}, denoting by $\,k_\pm^{\phantom i}$ and $\,q\,$ the complex fibre
dimensions of the sub\-bundles $\,\mathcal{H}^\pm$ and $\,\mathcal{H}\,$ of
$\,T\nnhM}%{{\mathrm{M}}'\hskip-.7pt$, we have, for $\,m=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}\,$ and
$\,d_\pm^{\phantom i}\hskip-.7pt=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\varSigma^\pm\hskip-.7pt$,
\begin{equation}\label{dmq}
d_+^{\phantom i}+\hskip.7pt\,d_-^{\phantom i}\hskip.7pt=\,\hskip.7pt m\,-\,1\,+\,q\hskip.4pt,
\end{equation}
as one sees adding the equalities $\,d_\pm^{\phantom i}\hskip-.7pt=m-1-k_\pm^{\phantom i}$ and
$\,m=1+k_+^{\phantom i}\hskip-1.5pt+k_-^{\phantom i}\hskip-1.5pt+q$ (the former due to (\ref{dbp}) and (\ref{hpm}),
the latter to (\ref{tme})). Consequently,
\begin{equation}\label{dpp}
d_+^{\phantom i}+\hskip.7pt\,d_-^{\phantom i}\hskip.7pt\,\ge\,\,m\,-\,1\hskip.4pt,
\end{equation}
with equality if and only if the distribution $\,\mathcal{H}\,$ in (\ref{tme})
is \hbox{$\,0$-}\hskip0ptdi\-men\-sion\-al, that is, if
\begin{equation}\label{tmp}
T\nnhM}%{{\mathrm{M}}'\,=\,\,\mathcal{V}\hskip1.2pt\oplus\hskip.7pt\mathcal{H}^+\hskip-1.5pt
\oplus\hskip.7pt\mathcal{H}^-\hskip-1.5pt.
\end{equation}
The explicit descriptions of $\,\varSigma^\pm$ in (\ref{spm}.c) -- (\ref{spm}.d)
clearly show that
\begin{equation}\label{dem}
d_+^{\phantom i}+\hskip.7pt\,d_-^{\phantom i}\hskip.7pt=\,\hskip.7pt m\,-\,1\,\,\mathrm{\ \ for\ every\ CP\ triple\
}\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\hskip.4pt.
\end{equation}
\end{remark}
\section{Examples: Nontrivial modifications}\label{nm}
\setcounter{equation}{0}
\begin{remark}\label{realz}For any two functions $\,\vt\mapsto Q\,$ and
$\,\hat\vt\mapsto\hat Q\,$ having the properties listed in (\ref{pbd}), with
the same $\,\vt\hskip-1.5pt_\pm\w$ and $\,a$, there must exist an increasing $\,C^\infty\hskip-1.5pt$
dif\-feo\-mor\-phism $\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\ni\vt\mapsto\hat\vt\in[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\,$
which realizes
\begin{equation}\label{qdt}
\begin{array}{l}
\mathrm{the\ equality\ }\hskip.7pt\hat Q\,d\hskip-.7pt/\hskip-1.5pt d\hat\vt\hskip.7pt
=\hskip.4pt Q\,d\hskip-.7pt/\hskip-1.5pt d\vt\,\mathrm{\ of\ vector\ fields\ on\
}\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\,\mathrm{\ expressed}\\
\mathrm{in\ terms\ of\ the\ two\
dif\-feo\-mor\-phic\-al\-ly}\hyp\mathrm{re\-lat\-ed\ coordinates\ }
\,\hat\vt\,\mathrm{\ and\ }\,\vt.
\end{array}
\end{equation}
Such a dif\-feo\-mor\-phism is unique up to compositions from the left (or,
right) with transformations forming the flow of the first (or, second) vector
field in (\ref{qdt}).
To see this, apply Remark~\ref{adtet} to $\,t=\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w$ and
$\,\gamma=\mp\hskip.4pt Q/(2a)$ (or, $\,t=\hat\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w$ and
$\,\gamma=\mp\hskip.4pt\hat Q/(2a)$), obtaining a function $\,\theta\,$ (or,
$\,\hat\theta$), unique up to a positive constant factor and vanishing at
$\,\vt=\hskip-.4pt\vt\hskip-1.5pt_\pm\w$ (or, $\,\hat\vt=\hskip-.4pt\vt\hskip-1.5pt_\pm\w$), with
$\,d\hskip.4pt\theta\hskip-.4pt/\hskip-.7pt d\vt=\theta/Q\,$ (or,
$\,d\hskip.4pt\hat\theta\hskip-.4pt/\hskip-.7pt d\hat\vt=\hat\theta/\hat Q$), the derivative being
positive everywhere in $\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]$. Adjusting the constant factor, we may
require $\,\vt\mapsto\theta\,$ and $\,\hat\vt\mapsto\hat\theta\,$ to be
increasing dif\-feo\-mor\-phisms of $\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]$ onto {\medit the
same\/} interval having the endpoint $\,0$, and then define
$\,\vt\mapsto\hat\vt\,$ by declaring $\,\theta$ ``equal to''
$\,\hat\vt\mapsto\hat\theta$, that is, letting $\,\vt\mapsto\hat\vt\,$ be
$\,\vt\mapsto\theta\,$ followed by the inverse of
$\,\hat\vt\mapsto\hat\theta$. Consequently,
$\,d\hat\vt\hskip-.4pt/\hskip-.7pt d\vt=\hat Q/Q\,$ on $\,(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)$, which amounts to
(\ref{qdt}).
The uniqueness clause is obvious: the only self-dif\-feo\-mor\-phisms
$\,\zeta\,$ of $\,(\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w)$ preserving a given vector field without zeros
are its flow transformations, since $\,\zeta$ acts on an integral curve as a
shift of the parameter.
\end{remark}
\begin{theorem}\label{ntrmd}{\medit
For the data\/ $\,\vt\hskip-1.5pt_\pm\w$ and\/ $\,\vt\mapsto Q\,$ related via
Remark\/~{\rm\ref{ascdt}(i)} to a given compact ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, and any increasing\/ $\,C^\infty\hskip-1.5pt$
dif\-feo\-mor\-phism\/ $\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\ni\vt\mapsto\hat\vt\in[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]$,
there exists a\/ $\,C^\infty\hskip-1.5pt$ function\/
$\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]\ni\vt\mapsto\phi\in\mathrm{I\!R}}%{\mathbf{R}$, unique up to additive constants,
such that\/ $\,\,\hat\vt\,=\,\vt\hskip.7pt+\hskip.4pt Q\,d\phi\hskip-.4pt/\hskip-.7pt d\vt$.
With\/ $\,\hat\vt,\phi\,$ treated, due to their dependence on\/ $\,\vt$, as
functions on the complex manifold\/ $\,M}%{{\mathrm{M}}\hskip-.7pt$, the formula\/ $\,\hat{g\hskip2pt}\hskip-1.3pt=g
-2(i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip0pt\phi)(J\,\cdot\,,\,\cdot\,)\,$
then defines another K\"ah\-ler metric on\/ $\,M}%{{\mathrm{M}}\hskip-.7pt$, and\/
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] $(M}%{{\mathrm{M}},\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt)\,$ is a new ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple.
\end{enumerate}
In addition, denoting by\/ $\,\hat\vt\mapsto\hat Q\,$ the analog of\/
$\,\vt\mapsto Q$ arising when Remark\/~{\rm\ref{ascdt}(i)} is applied to\/
$\,(M}%{{\mathrm{M}},\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt)$, and by\/ $\,\hat\nabla\hskip-.5pt\hat\vt\,$ the\/
$\,\hat{g\hskip2pt}\hskip-1.3pt$-gra\-di\-ent of\/ $\,\hat\vt$, \,one has\/ {\rm(\ref{qdt})} and\/
$\,\hat\nabla\hskip-.5pt\hat\vt=\hskip-.7pt\nabla\hskip-.5pt\vt$.
}
\end{theorem}
\begin{proof}As $\,\hat\vt=\vt\hskip.7pt+\hskip.4pt Q\hskip.4pt\phi'\hskip-.7pt$, where
$\,(\hskip2.3pt)'\hskip-.7pt=d\hskip-.7pt/\hskip-1.5pt d\vt$, our assumption about
$\,\vt\mapsto\hat\vt\,$ gives $\,\hat\vt'\hskip-.7pt>0$ and
$\,\vt\hskip-1.7pt_-\w\hskip-.7pt\le\hskip.7pt\hat\vt\le\vt\hskip-1.8pt_+\w$, leading to the inequalities
\begin{equation}\label{neq}
\vt\hskip-1.7pt_-\w\le\,\vt\hskip.7pt+\hskip.7pt Q\hskip.4pt\phi'\le\hskip.7pt\vt\hskip-1.8pt_+\w\hskip7pt\mathrm{(strict\ except\
at}\hskip5pt\vt=\vt\hskip-1.5pt_\pm\w\mathrm{)\ and}\hskip7pt1+Q'\phi'\hskip-.7pt+Q\hskip.4pt\phi''\hskip-.7pt>0\hskip.4pt.
\end{equation}
Note that $\,\phi\,$ exists since, by Remark~\ref{smdiv} and
(\ref{pbd}), $\,\hat\vt-\hskip-.7pt\vt\,$ and $\,Q\,$ are smoothly divisible by
$\,\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w$, their quotients being equal at $\,\vt\hskip-1.5pt_\pm\w$ to the value of
$\,\hat\vt'\hskip-.7pt-1\,$ and $\,\mp2a$, respectively, and so, as $\,\hat\vt'\hskip-.7pt>0$,
\begin{equation}\label{fgt}
\mp2a\hskip.4pt\phi'\hskip.7pt>\,-\nnh1\hskip8pt\mathrm{at}\hskip7pt\vt=\vt\hskip-1.5pt_\pm\w\hskip.4pt.
\end{equation}
For the self-ad\-joint bundle en\-do\-mor\-phism $\,K\,$ of $\,T\nnhM}%{{\mathrm{M}}\,$
with $\,\hat{g\hskip2pt}\hskip-1.3pt=g(K\,\cdot\,,\,\cdot\,)\,$ one has
\begin{equation}\label{cei}
K\,\hskip.7pt=\,\,\mathrm{Id}\,\hskip.7pt+\,2\phi'\nhS\,
+\,\phi''[\hskip.4pt g(\hskip-.4pt v,\,\cdot\,)\hskip.4pt v+g(\hskip-.4pt u,\,\cdot\,)\hskip.4pt u]\hskip.4pt,
\end{equation}
where $\,v,u,S\,$ are, as usual, given by $\,v=\nabla\hskip-.5pt\vt$, $\,u=J\hskip-.7pt v$, and
$\,S=\nabla\hskip-.7pt v$.
We proceed to prove positivity of $\,K\,$ at all points of $\,M}%{{\mathrm{M}}\hskip-.7pt$,
considering two separate cases: $\,y\in\varSigma^\pm$ and
$\,x\inM}%{{\mathrm{M}}'\hskip-.7pt=M}%{{\mathrm{M}}\smallsetminus(\varSigma^+\hskip-1.5pt\cup\varSigma^-)$, cf.\
Remark~\ref{ascdt}(iv).
If $\,y\,$ lies in either critical manifold $\,\varSigma^\pm\hskip-.7pt$, the relations
$\,v_y^{\phantom i}=u_y^{\phantom i}=0\,$ and $\,\vt(y)=\vt\hskip-1.5pt_\pm\w$ imply positivity of $\,K\hskip-1.5pt_y^{\phantom i}$ as a
consequence of (\ref{fgt}) since, by Remark~\ref{prcrm}, any eigen\-val\-ue of
$\,S\hskip-1.5pt_y^{\phantom i}$ must be equal to $\,0\,$ or $\,\mp\hskip.4pt a$.
On $\,M}%{{\mathrm{M}}'\hskip-.7pt$, we use the $\,S$-in\-var\-i\-ant decomposition
$\,T\nnhM}%{{\mathrm{M}}'\hskip-.7pt=\mathcal{V}\oplus\mathcal{V}^\perp\hskip-1.5pt$, cf.\ (\ref{tme}).
In view of (\ref{loc}.f) -- (\ref{loc}.g), the restriction of
$\,2S=2\nabla\hskip-.7pt v\,$ to $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)\,$ equals
$\,Q'$ times $\,\mathrm{Id}$. Using (\ref{cei}) and (\ref{gvv}) we now see
that $\,K\,$ acts in $\,\mathcal{V}\,$ via multiplication by the function
$\,1+Q'\phi'\hskip-.7pt+Q\hskip.4pt\phi''\hskip-.7pt$, which is positive according to (\ref{neq}).
Theorem~\ref{jacob}(vi) states in turn that the eigen\-values of
$\,S\hskip-1.5pt_x^{\phantom i}:\mathcal{V}^\perp_{\hskip-1ptx}\hskip-.7pt
\to\mathcal{V}^\perp_{\hskip-1ptx}$, for $\,x\inM}%{{\mathrm{M}}'\hskip-.7pt$, have the form
$\,\lambda_c^{\phantom i}(x)\,$ with (\ref{lgm}) and (\ref{wze}.a). Writing
$\,K,S,\vt,Q,\phi'$ instead of their values at $\,x$, we conclude from
(\ref{cei}) that the corresponding eigen\-values of $\,(\vt-c)\hskip.4pt K\,$ are
$\,\vt\hskip.7pt+\hskip.7pt Q\hskip.4pt\phi'\hskip-.7pt-c\,$ and so, due to the (strict) first inequality
of (\ref{neq}), they all lie in the interval $\,(\vt\hskip-1.7pt_-\w\hskip-.7pt-c,\vt\hskip-1.8pt_+\w\hskip-.7pt-c)$.
Positivity of $\,K\,$ on $\,\mathcal{V}\,$ thus easily follows both when
$\,c<\hskip.4pt\vt\hskip-1.7pt_-\w\hskip-1.5pt<\vt\,$ and when $\,\vt<\hskip.4pt\vt\hskip-1.8pt_+\w\hskip-1.5pt<c$.
Consequently, $\,\hat{g\hskip2pt}\hskip-1.3pt\,$ is a K\"ah\-ler metric on $\,M}%{{\mathrm{M}}\hskip-.7pt$, with
the K\"ah\-ler form $\,\hat\omega=\hat{g\hskip2pt}\hskip-1.3pt(\hat J\,\cdot\,,\,\cdot\,)$ related to
$\,\omega=g(\hat J\,\cdot\,,\,\cdot\,)\,$ by
$\,\hat\omega=\omega+2i\hskip.7pt\partial\hskip1.7pt\overline{\hskip-2pt\partial}\hskip0pt\phi$. Applying
(\ref{ojv}) to $\,v=\nabla\hskip-.5pt\vt\,$ and $\,\phi\,$ rather than $\,f\hskip-1.5pt$, we obtain
$\,\hat{g\hskip2pt}\hskip-1.3pt(\hskip-.4pt v,\,\cdot\,)=g(\hskip-.4pt v,\,\cdot\,)-2\hskip.4pt\omega\hskip.4pt(J\hskip-.7pt v,\,\cdot\,)
=d\vt+d(d_v^{\phantom i}\phi)$. (Note that $\,J\hskip-.7pt v=u$ and (\ref{gvv}) gives
$\,d_u^{\phantom i}\phi=d_u^{\phantom i}\vt=0$, since $\,\phi\,$ is a function of $\,\vt$.) As
$\,d_v^{\phantom i}\vt=Q$, cf.\ (\ref{dvt}), $\,v\,$ is thus the $\,\hat{g\hskip2pt}\hskip-1.3pt$-gra\-di\-ent
of $\,\vt+d_v^{\phantom i}\phi=\vt\hskip.7pt+\hskip.4pt Q\hskip.4pt\phi'\hskip-.7pt=\hskip.7pt\hat\vt$. On the other hand,
again from (\ref{dvt}), $\,\hat Q=\hat{g\hskip2pt}\hskip-1.3pt(v,v)\,$ equals
$\,d_v^{\phantom i}\hat\vt=\hat\vt'd_v^{\phantom i}\vt=\hat\vt'Q$, which is a function of $\,\vt$,
and of $\,\hat\vt$, proving both (a) (see Lemma~\ref{ggqft}) and (b).
\end{proof}
\begin{remark}\label{alltq}Let $\,G\,$ be the group of all auto\-mor\-phisms
(Definition~\ref{ggktr}) of a given compact ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$. Then every quadruple
$\,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a$, $\hat\vt\mapsto\hat Q\,$ satisfying the analog of (\ref{pbd})
arises when Remark~\ref{ascdt}(i) is applied to a suitably chosen
$\,G$-in\-var\-i\-ant ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple
$\,(M}%{{\mathrm{M}},\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt)\,$ with the same underlying complex manifold $\,M}%{{\mathrm{M}}\hskip-.7pt$.
In fact, a trivial modification (see Remark~\ref{trivl}) followed by rescaling
of the metric allows us to assume that $\,\vt\hskip-1.5pt_\pm\w$ and $\,a\,$ are the same as
those for $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$. Our claim is now obvious from Remark~\ref{realz}
and Theorem~\ref{ntrmd}.
\end{remark}
\begin{remark}\label{spcas}As a special case of Remark~\ref{alltq}, for the
first triple using the Fu\-bi\-ni-Stu\-dy metric $\,g\,$ and $\,G\,$ as in the
lines preceding (\ref{dta}), all quadruples $\,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a,\hskip.7pt\vt\mapsto Q$
with (\ref{pbd}) are realized, via Remark~\ref{ascdt}(i), by \,CP triples
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ having arbitrarily fixed values of $\,m=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}\,$
and $\,d_\pm^{\phantom i}\hskip-.7pt=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\varSigma^\pm$ that satisfy (\ref{dem}).
\end{remark}
\begin{remark}\label{cnvrs}Conversely, we can apply Remark~\ref{realz} and
Theorem~\ref{ntrmd} to canonically modify any given \,CP triple, obtaining one
with the Fu\-bi\-ni-Stu\-dy metric and the same group $\,G$.
We will not use the eas\-i\-ly-ver\-i\-fied fact that, for such a
Fu\-bi\-ni-Stu\-dy \,CP triple,
$\,(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)\hskip.4pt Q=2a\hskip.4pt(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)(\vt\hskip-1.8pt_+\w\hskip-2.3pt-\vt)\,$
and, in (\ref{dta}.ii), the value of $\,\vt\,$ at $\,\bbC(\xi+\eta)$, where
$\,\xi\in\mathsf{L}\,$ and $\,\eta\in\mathsf{L}\hskip-2.5pt^\perp\hskip-.7pt$, equals
$\,(\vt\hskip-1.5pt_\pm\w|\hskip.4pt\xi|^2\hskip-.7pt+\vt\hskip-1pt_\mp\w|\eta|^2)/(|\hskip.4pt\xi|^2\hskip-.7pt+|\eta|^2)\,$ for some sign
$\,\pm\hskip.7pt$.
\end{remark}
\section{The nor\-mal-ge\-o\-des\-ic bi\-hol\-o\-mor\-phisms}\label{ng}
\setcounter{equation}{0}
In this section $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is a fixed compact
ge\-o\-des\-\hbox{ic\hskip.7pt-}\hskip0ptgra\-di\-ent K\"ah\-ler triple
(Definition~\ref{ggktr}). We use the notation of (\ref{sym}), denote by
$\,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a,Q\,$ the data (\ref{pbd}) associated with $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$
(see Remark~\ref{ascdt}(i)), and choose for them the further data (\ref{sgn})
-- (\ref{int}), so that a sign $\,\pm\,$ is fixed as well. We also let
$\,\varSigma,N}%{{\mathrm{N}}\hskip-.7pt,h,\langle\,,\rangle\,$ and $\,\mathrm{D}\,$ stand for $\,\varSigma^\pm\hskip-1.5pt$, the
normal bundle $\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt$, the sub\-man\-i\-fold metric of
$\,\varSigma$, the Riemannian fibre metric in $\,N}%{{\mathrm{N}}\,$ induced by $\,g$, and
the Chern connection of $\,\langle\,,\rangle\,$ in $\,N}%{{\mathrm{N}}\hskip-1.5pt$, cf.\ (d) in
Section~\ref{ck}. We write $\,(y,\xi)\inN}%{{\mathrm{N}}\,$ when $\,y\in\varSigma\,$ and
$\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, as in Remark~\ref{tlspc}.
Using the normal exponential dif\-feo\-mor\-phism
$\,\mathrm{Exp}^\perp\hskip-1.5pt:N}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm\hskip-.7pt
\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp$ in (\ref{nex}), we define
$\,\varPhi=\hskip.7pt\varPhi^\pm\hskip-1.5pt:N}%{{\mathrm{N}}\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt$,
depending on the sign $\,\pm\hskip.7pt$, to be the composite
\begin{equation}\label{phe}
\varPhi\,\,=\,\,\mathrm{Exp}^\perp\hskip-1.5pt\circ\Delta\hskip.4pt,
\end{equation}
where $\,\Delta:N}%{{\mathrm{N}}\toN}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm$ is given by $\,\Delta(y,\xi)=y\,$ if
$\,\xi=0\,$ and, otherwise,
\begin{equation}\label{dyx}
\begin{array}{l}
\Delta(y,\xi)\hskip.7pt=\hskip.7pt(y,t\hskip.4pt\xi)\mathrm{,\ where\
}\,t\hskip.7pt=\hskip.7pt\sigma\hskip-1.5pt/\hskip-1.5pt\rho\,\mathrm{\ for\ }\,\rho\hskip.7pt=\hskip.4pt|\hskip.7pt\xi|\mathrm{,\
the\ function\ }\,\hskip.7pt\sigma\\\mathrm{of\ the\ variable\
}\,\rho\in[\hs0,\infty)\,\mathrm{\ being\ chosen\ as\ above,\ with\
(\ref{dsr}).}
\end{array}
\end{equation}
Note that $\,\Delta\,$ is a homeo\-mor\-phism and, restricted to the complement
$\,N}%{{\mathrm{N}}'\hskip-.7pt=N}%{{\mathrm{N}}\hskip-.7pt\smallsetminus\varSigma$ of the zero section, it becomes a
dif\-feo\-mor\-phism
$\,N}%{{\mathrm{N}}'\hskip-.7pt\toN}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm\hskip-1.5pt\smallsetminus\varSigma^\pm\hskip-.7pt$. In fact,
$\hskip.7pt t\hskip.4pt\xi\hskip.7pt$ with $\,t=\hskip.7pt\sigma\hskip-1.5pt/\hskip-1.5pt\rho\,$ determines $\,\xi\,$
(smoothly if $\,\xi\ne0$), since $\,|\hskip.4pt t\hskip.4pt\xi|=\sigma\,$ and $\,\sigma\,$
determines $\,\rho$ according to the line preceding (\ref{dsr}).
Consequently, $\,\varPhi:N}%{{\mathrm{N}}\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp$ is a
homeo\-mor\-phism, and the restriction $\,\varPhi:N}%{{\mathrm{N}}'\hskip-.7pt\toM}%{{\mathrm{M}}'$ a
dif\-feo\-mor\-phism. In addition,
\begin{equation}\label{pcf}
\begin{array}{l}
\pi^\pm\hskip-1.5pt\circ\varPhi^\pm\mathrm{\ equals\ the\
nor\-mal}\hyp\mathrm{bun\-dle\ projection\
}\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt\to\varSigma^\pm
\end{array}
\end{equation}
due to (\ref{phe}), the fibre\hskip.4pt-\hskip-.4pt preserving property of $\,\Delta$, and
the first line of Remark~\ref{ptnrs}.
\begin{remark}\label{sends}Suppose that a vector field $\,w\,$ on $\,N}%{{\mathrm{N}}'$ is
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] the $\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal lift of a vector
field on $\,\varSigma$, or
\item[{\rm(b)}] a vertical vector field of the form
$\,(y,\xi)\mapsto\varTheta\xi\,$ for some com\-plex-lin\-e\-ar
vec\-tor-bun\-dle morphism $\,\varTheta:N}%{{\mathrm{N}}\toN}%{{\mathrm{N}}\hskip-1.5pt$, skew-ad\-joint
relative to $\,\langle\,,\rangle\,$ at every point.
\end{enumerate}
Then $\,\Delta$, restricted to $\,N}%{{\mathrm{N}}'\hskip-.7pt$, sends $\,w\,$ onto its restriction to
$\,N}%{{\mathrm{N}}'\capN}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm\hskip-.7pt.$
In fact, let $\,r\mapsto(y(\hskip-.4pt r\hskip-.7pt),\xi(\hskip-.4pt r\hskip-.7pt))\,$ be an integral curve of
$\,w$. Then the function $\,r\mapsto|\hskip.4pt\xi(\hskip-.4pt r\hskip-.7pt)\hskip-.4pt|$ is constant, and
so, by (\ref{dyx}),
$\,\Delta(y(\hskip-.4pt r\hskip-.7pt),\xi(\hskip-.4pt r\hskip-.7pt))=(y(\hskip-.4pt r\hskip-.7pt),c\hskip.7pt\xi(\hskip-.4pt r\hskip-.7pt))\,$ with some
real constant $\,c$. This proves our claim since, in case (b), $\,w\,$
restricted to every fibre $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, being a linear vector field
on $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, is invariant under multiplications by scalars.
\end{remark}
\begin{theorem}\label{first}{\medit
For either critical manifold\/ $\,\varSigma^\mp$ of\/ $\,\vt\,$ in any compact
ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, the
triple\/ $\,(M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-1.5pt,g,\vt)\,$ is isomorphic to
one constructed in Section\/~{\rm\ref{ev}} \hskip.8ptfrom some data\/
{\rm(\ref{pbd})} -- {\rm(\ref{sgn})} and\/ $\,\varSigma,h,N}%{{\mathrm{N}}\hskip-.7pt,\langle\,,\rangle$.
The data consist of\/ {\rm(\ref{pbd})} associated with\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$
as in Remark\/~{\rm\ref{ascdt}(i)}, any choice of\/ $\,\vt\mapsto\rho\,$
with\/ {\rm(\ref{sgn})} for\/ {\rm(\ref{pbd})} and our fixed sign\/
$\,\pm\hskip.7pt$, the sub\-man\-i\-fold metric\/ $\,h\,$ and normal bundle\/
$\,N}%{{\mathrm{N}}=N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-.7pt$ of\/ $\,\varSigma=\varSigma^\pm\hskip-1.5pt$, and the
fibre metric\/ $\,\langle\,,\rangle\,$ in\/ $\,N}%{{\mathrm{N}}\hskip.7pt$ induced by\/ $\,g$. Furthermore,
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] the required isomorphism\/
$\,N}%{{\mathrm{N}}\hskip-.7pt\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp$ is provided by the mapping\/
$\,\varPhi=\hskip.7pt\varPhi^\pm$ with\/ {\rm(\ref{phe})}, which, in particular, must
be bi\-hol\-o\-mor\-phic,
\item[{\rm(ii)}] $\varPhi\,$ sends the horizontal distribution of the Chern
connection\/ $\,\mathrm{D}\,$ of\/ $\,\langle\,,\rangle\,$ in\/ $\,N}%{{\mathrm{N}}\hskip-1.5pt$, cf.\ {\rm(d)}
of Section\/~{\rm\ref{ck}}, onto the summand\/
$\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\pm$ in\/ {\rm(\ref{tme})},
\item[{\rm(iii)}] the leaves of\/ $\,\hskip.7pt\mathcal{V}\,$ are precisely the same
as the\/ $\,\varPhi\hskip.4pt$-im\-a\-ges of all punctured complex lines through\/
$\,0\,$ in the normal spaces of\/ $\,\varSigma$.
\end{enumerate}
In the special case where\/
$\,T\nnhM}%{{\mathrm{M}}'\,=\,\,\mathcal{V}\hskip1.2pt\oplus\hskip.7pt\mathcal{H}^+\hskip-1.5pt
\oplus\hskip.7pt\mathcal{H}^-\hskip-1.5pt$, that is, the summand distribution\/
$\,\mathcal{H}$ \hbox{in\/ {\rm(\ref{tme})}} is\/
\hbox{$\,0$-}\hskip0ptdi\-men\-sion\-al, formula\/ {\rm(\ref{hgv}.c)} used in
the construction of \,Section\/~{\rm\ref{ev}} may also be replaced by the
following equality, using the simplified notation of\/
{\rm(\ref{hgv}.c):}
\begin{equation}\label{alt}
\hat{g\hskip2pt}\hskip-1.3pt(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,=\,\frac{|\hskip.4pt\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w|}{\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w}\,h(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt.
\end{equation}
}
\end{theorem}
\begin{proof}It suffices to prove that the restriction of $\,\varPhi\,$ to
$\,N}%{{\mathrm{N}}'\hskip-.7pt=N}%{{\mathrm{N}}\hskip-.7pt\smallsetminus\varSigma\,$ is an isomorphism between the
ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triples $\,(N}%{{\mathrm{N}}'\hskip-1.5pt,\hat{g\hskip2pt}\hskip-1.3pt,\hat\vt)\,$
and $\,(M}%{{\mathrm{M}}'\hskip-1.5pt,g,\vt)$, since the analogous conclusion about $\,\varPhi\,$
itself then follows from \cite[Lemma 16.1]{derdzinski-maschler-06}.
We start by establishing the equality
\begin{equation}\label{vcf}
\vt\circ\varPhi\,=\,\hat\vt\hskip.4pt.
\end{equation}
Namely, $\,|\hskip.4pt\rho\hskip.4pt\xi|=\rho\,$ for any $\,\rho\in(0,\infty)\,$ and any
$\,(y,\xi)\inN}%{{\mathrm{N}}$ with $\,|\hskip.7pt\xi|=1$, so that
$\,\varPhi(y,\rho\hskip.4pt\xi)=x\hskip-.7pt_\sigma^{\phantom i}$, where
$\,x\hskip-.7pt_\sigma^{\phantom i}=\exp_y^{\phantom i}\sigma\xi\,$ and $\,\sigma\hskip.7pt$ depends on $\,\rho\,$
as in (\ref{dsr}). Since $\,\sigma\mapsto x\hskip-.7pt_\sigma^{\phantom i}$ is a unit-speed
geodesic, (\ref{vtg}) and (\ref{dtq}) give
$\,d\hskip.7pt[\vt(x\hskip-.7pt_\sigma^{\phantom i})]/d\sigma=\mp\hskip.7pt Q^{\hs1\hskip-.7pt/2}\hskip-.7pt$, the sign factor
being due to the relation $\,d\hskip.4pt(x\hskip-.7pt_\sigma^{\phantom i})/d\sigma=\mp\hskip.4pt v/|v|\,$
(immediate from (\ref{tmm}) with $\,v=\nabla\hskip-.5pt\vt$). Here $\,Q=g(\hskip-.4pt v,v)\,$ depends
on $\,\vt(x\hskip-.7pt_\sigma^{\phantom i})\,$ as in Remark~\ref{ascdt}(i). However, according to
Remark~\ref{compo} and the text preceding (\ref{hgv}.a) -- (\ref{hgv}.b), the
same autonomous equation
$\,d\hskip.7pt[\hat\vt(y,\rho\hskip.4pt\xi)]/d\sigma=\mp\hskip.4pt Q^{\hs1\hskip-.7pt/2}$ holds when
$\,\vt(x\hskip-.7pt_\sigma^{\phantom i})\,$ is replaced by $\,\hat\vt(y,\rho\hskip.4pt\xi)$, with {\medit
the same} dependence of $\,Q\,$ on the unknown function. The uniqueness clause
of Remark~\ref{compo} thus gives
$\,\vt(\varPhi(y,\rho\hskip.4pt\xi))=\vt(x\hskip-.7pt_\sigma^{\phantom i})=\hat\vt(y,\rho\hskip.4pt\xi)$, as
required.
One has two complex di\-rect-sum decompositions,
$\,T\nnhM}%{{\mathrm{M}}'=\,\mathcal{V}\hskip1.2pt\oplus\hskip.7pt\mathcal{H}^\mp\hskip-1.5pt
\oplus\hskip.7pt\mathcal{H}^\bullet$ and
$\,T\nnhN}%{{\mathrm{N}}'=\,\hat{\mathcal{V}}\hskip1.2pt\oplus\hskip.7pt\hat{\mathcal{H}}^\mp\hskip-1.5pt
\oplus\hskip.7pt\hat{\mathcal{H}}^\bullet\hskip-.7pt$, orthogonal relative to $\,g\,$ and,
respectively, $\,\hat{g\hskip2pt}\hskip-1.3pt$. The former arises from (\ref{tme}) if one sets
$\,\mathcal{H}^\bullet\hskip-.7pt=\,\mathcal{H}^\pm\hskip-1.5pt\oplus\hskip.7pt\mathcal{H}$.
In the latter $\,\hat{\mathcal{V}}\hskip-.7pt,\hat{\mathcal{H}}^\mp$ and
$\,\hat{\mathcal{H}}^\bullet$ are the distributions introduced in the lines
following (\ref{tnp}). First, for $\,\hat u\,$ as in (\ref{lvf}) and our
$\,u=J\hskip-.7pt v$, where $\,v=\nabla\hskip-.5pt\vt$, we show that
\begin{equation}\label{prs}
\begin{array}{rl}
\mathrm{i)}&\Delta\,\mathrm{\ preserves\
}\,\hat{\mathcal{V}}\hskip-.7pt,\,\hat{\mathcal{H}}^\mp\hskip-1.5pt,
\hat{\mathcal{H}}^\bullet\mathrm{\ and\ }\,\hat u\hskip.4pt,\\
\mathrm{ii)}&\mathrm{Exp}^\perp\mathrm{\ sends\
}\,\hat{\mathcal{V}}\hskip-.7pt,\,\hat{\mathcal{H}}^\mp\hskip-1.5pt,
\hat{\mathcal{H}}^\bullet\hskip-.7pt,\hskip.7pt\hat u\,\mathrm{\ to\
}\,\mathcal{V}\hskip-.7pt,\,\mathcal{H}^\mp\hskip-1.5pt,\mathcal{H}^\bullet\hskip-.7pt,\hskip.7pt u\hskip.4pt,\\
\mathrm{iii)}&
\mathrm{both\ }\,\Delta\,\mathrm{\ and\ }\,\mathrm{Exp}^\perp\mathrm{\ act\
com\-plex}\hyp\mathrm{lin\-e\-ar\-ly\ on\ }\,\hat{\mathcal{H}}^\mp\mathrm{\
and\ }\,\hat{\mathcal{H}}^\bullet\hskip-1.5pt.
\end{array}
\end{equation}
More precisely, $\,\Delta\,$ (or, $\,\mathrm{Exp}^\perp$) appearing in
(\ref{dyx}) (or, (\ref{nex})), restricted to $\,N}%{{\mathrm{N}}'$ (or,
$\,N}%{{\mathrm{N}}'\capN}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm$), sends
$\,\hat{\mathcal{V}}\hskip-.7pt,\,\hat{\mathcal{H}}^\mp\hskip-1.5pt,
\hat{\mathcal{H}}^\bullet\hskip-.7pt,\hskip.7pt\hat u\,$ onto their restrictions to
$\,N}%{{\mathrm{N}}'\capN}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm$ (or, respectively, onto
$\,\mathcal{V}\hskip-.7pt,\,\mathcal{H}^\mp\hskip-1.5pt,\mathcal{H}^\bullet\hskip-.7pt,\hskip.7pt u$). The
claims about $\,\hat{\mathcal{V}}\,$ in (\ref{prs}.i) -- (\ref{prs}.ii) follow
as $\,\Delta\,$ clearly preserves each leaf of $\,\hat{\mathcal{V}}\hskip-.7pt$, that is,
each punctured complex line through $\,0\,$ in the normal space
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma\,$ at any point $\,y\in\varSigma$, while, by
Lemma~\ref{vsbkr}(a), $\,\mathrm{Exp}^\perp$ maps the leaves of
$\,\hat{\mathcal{V}}\,$ intersected with $\,N}%{{\mathrm{N}}'\hskip-1.5pt\capN}%{{\mathrm{N}}^\delta\hskip-1.5pt\varSigma^\pm$
onto leaves of $\,\mathcal{V}\hskip-.7pt$. This also proves (ii). Next, the
class of vertical vector fields of Remark~\ref{sends}(b) obviously includes
$\,\hat u\,$ and, locally, some of them span $\,\hat{\mathcal{H}}^\mp\hskip-.7pt$.
Remark~\ref{sends} thus yields the remainder of (\ref{prs}.i), while
(\ref{prs}.iii) for $\,\Delta\,$ follows from com\-plex-lin\-e\-ar\-i\-ty of the
$\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal lift operation (due to
Lemma~\ref{ddrsq}(i)), and the fact that $\,\Delta\,$ acts on the vertical vector
fields in Remark~\ref{sends}(b) as the identity operator. On the other hand,
(\ref{prs}.ii) in the case of $\,\hat{\mathcal{H}}^\mp$ and
$\,\hat{\mathcal{H}}^\bullet$ (or, of $\,\hat u$) is an immediate consequence
of the second (or, third) claim in Theorem~\ref{dcomp}(f). (To be specific,
for $\,\hat{\mathcal{H}}^\mp$ and $\,\hat{\mathcal{H}}^\bullet$ this is clear
from Remark~\ref{inequ}(ii) combined with (\ref{hpm}) -- (\ref{tme}).)
Finally, the com\-plex-lin\-e\-ar\-i\-ty assertion of Theorem~\ref{dcomp}(f)
implies (\ref{prs}.iii).
By (\ref{prs}), the dif\-feo\-mor\-phism
$\,\varPhi=\mathrm{Exp}^\perp\hskip-1.5pt\circ\Delta:N}%{{\mathrm{N}}'\hskip-.7pt\toM}%{{\mathrm{M}}'$ maps
$\,\hat{\mathcal{V}}\hskip-.7pt,\hat{\mathcal{H}}^\mp$ and
$\,\hat{\mathcal{H}}^\bullet$ onto $\,\mathcal{V}\hskip-.7pt,\mathcal{H}^\mp$ and
$\,\mathcal{H}^\bullet\hskip-1.5pt$. Proving the theorem is thus reduced to showing
that
\begin{equation}\label{mps}
\begin{array}{l}
\hat J\,\mathrm{\ and\ }\,\hat{g\hskip2pt}\hskip-1.3pt\mathrm{,\ on\ each\ of\ the\ three\ summands\
}\,\,\hat{\mathcal{V}}\hskip-.7pt,\,\hat{\mathcal{H}}^\mp\mathrm{\ and\
}\,\hat{\mathcal{H}}^\bullet\hskip-1.5pt\mathrm{,\ correspond}\\
\mathrm{under\ the\ differential\ }\,d\varPhi\,\mathrm{\ to\ }\,J\,\mathrm{\
and\ }\,g\,\mathrm{\ on\
}\,\mathcal{V}\hskip-.7pt,\,\mathcal{H}^\mp\mathrm{\ and\
}\,\mathcal{H}^\bullet\hskip-1.5pt\mathrm{,\ respectively\hskip-.7pt.}
\end{array}
\end{equation}
To begin with, for $\,\hat Q\,$ as in Section~\ref{ev}, $\,\hat v\,$ given by
(\ref{lvf}), and our $\,v=\nabla\hskip-.5pt\vt$,
\begin{equation}\label{qav}
\varPhi\,\,\mathrm{\ pushes\ }\,\hat Q,\hat u\,\mathrm{\ and\
}\,\hat v\,\mathrm{\ forward\ onto\ }\,Q,u\,\mathrm{\ and\ }v\hskip.4pt.
\end{equation}
In the case of $\,\hat Q\,$ this amounts to $\,Q\circ\varPhi=\hat Q$, which is
a trivial consequence of (\ref{vcf}) and the fact that $\,\hat Q\,$ was
defined in Section~\ref{ev} to be the same function of $\,\hat\vt\,$ as
$\,Q\,$ is of $\,\vt$. For $\,\hat u$, (\ref{qav}) follows from (\ref{prs})
and (\ref{phe}). Next, any integral curve of $\,\hat v\,$ in
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}\,$ has, up to a shift of the
parameter, the form $\,r\mapsto(y,e^{\mp\hskip.4pt ar}\xi)$ with a unit vector
$\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, so that
$\,\Delta(y,e^{\mp\hskip.4pt ar}\xi)=(y,\sigma\hskip.7pt\xi)$, where in addition to the curve
parameter $\,r\hskip-.7pt$, two more real variables are used:
$\,\rho=e^{\mp\hskip.7pt ar}\hskip-1.5pt$,
and $\,\sigma\,$ related to $\,\rho\,$ via (\ref{dsr}). The chain rule thus
yields
$\,d\sigma/dr=\mp\hskip.4pt a\rho\,d\sigma\hskip-.7pt/\hskip-.4pt d\hskip-.7pt\rho=\mp\hskip.4pt Q^{\hs1\hskip-.7pt/2}\hskip-.7pt$,
while $\,\varPhi(y,e^{\mp\hskip.4pt ar}\xi)=x(\hskip-.7pt\sigma\hskip-.7pt)\,$ for
$\,x(\hskip-.7pt\sigma\hskip-.7pt)=\exp_y^{\phantom i}\sigma\xi$. Since
$\,\sigma\mapsto x(\hskip-.7pt\sigma\hskip-.7pt)\,$ is a unit-speed geodesic, (\ref{vtg}) and
(\ref{dxe}) give $\,d\hskip.7pt[x(\hskip-.7pt\sigma\hskip-.7pt)]/d\sigma=\mp\hskip.7pt Q^{\hs1\hskip-.7pt/2}\hskip-.7pt$,
with $\,Q\,$ evaluated at $\,x(\hskip-.7pt\sigma\hskip-.7pt)$, and the sign factor arising
from (\ref{tmm}), as $\,v=\nabla\hskip-.5pt\vt$. Applying the chain rule again, we obtain
$\,d\hskip.7pt[x(\hskip-.7pt\sigma\hskip-.7pt)]/dr=v\hskip-.7pt_{x(\hskip-.7pt\sigma\hskip-.7pt)}^{\phantom i}$ and, consequently,
(\ref{qav}).
The claim made in (\ref{mps}) about
$\,\hat{\mathcal{V}}=\mathrm{Span}\hskip.7pt(\hat v,\hat u)\,$ and
$\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)\,$ is now obvious from (\ref{qav})
and (\ref{gvv}) along with (\ref{ana}).
For the remaining two pairs of summands, (\ref{mps}) in the case of
$\,\hat J,J\,$ (or, $\,\hat{g\hskip2pt}\hskip-1.3pt,g$) is a direct consequence of (\ref{prs}) and
(\ref{phe}) (or, respectively, of (i) -- (ii) in Remark~\ref{inequ} along with
parts (h2) -- (h3) of Theorem~\ref{dcomp}, (\ref{vcf}) and (\ref{hgv})). Note
that, by (\ref{dyx}), $\,\Delta\,$ leaves $\,\xi/\hskip-.4pt|\hskip.4pt\xi|\,$ unchanged, while
$\,\rho=|\hskip.4pt\xi|\,$ in (h2).
Finally, if $\,T\nnhM}%{{\mathrm{M}}'=\,\mathcal{V}\hskip1.2pt\oplus\hskip.7pt\mathcal{H}^+\hskip-1.5pt
\oplus\hskip.7pt\mathcal{H}^-$ in (\ref{tme}), Remark~\ref{inequ}(ii) allows us to
use (h1) in Theorem~\ref{dcomp}, instead of (h2), obtaining (\ref{alt}).
\end{proof}
\begin{corollary}\label{totgd}{\medit
Suppose that\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is a compact ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple. Then, for $\,\mathcal{V}\,$ and\/ $\,\mathcal{H}^\pm$
appearing in\/ {\rm(\ref{tme})}, with either sign\/ $\,\pm\hskip.7pt$, the
distribution\/ $\,\mathcal{V}\oplus\mathcal{H}^\pm$ is in\-te\-gra\-ble and
its leaves are totally geodesic in\/ $\,(M}%{{\mathrm{M}}'\hskip-.7pt,g)$.
}
\end{corollary}
\begin{proof}Use Theorem~\ref{first} and Theorem~\ref{ehggk}(b) (or -- for
in\-te\-gra\-bil\-i\-ty -- (\ref{hpm})).
\end{proof}
\section{Immersions of complex projective spaces}\label{ic}
\setcounter{equation}{0}
In the next result the inclusions
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}
\subseteq\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\,$
and $\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}
\subseteq\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\,$
come from the standard identification (\ref{inc}) for
$\,\mathsf{V}=N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, where $\,y\in\varSigma^\pm\hskip-.7pt$. Let us also note
that, by (\ref{hpm}) and Corollary~\ref{totgd}, the restriction to the normal
space
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}=N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt
\subseteqN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm$
of the bi\-hol\-o\-mor\-phism
$\,\varPhi:N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp$ (see
Theorem~\ref{first}) constitutes
\begin{equation}\label{tgh}
\mathrm{a\ totally\ geodesic\ hol\-o\-mor\-phic\ embedding\
}\,\,\,\varPhi:N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt.
\end{equation}
\begin{theorem}\label{tgimm}{\medit
Given a compact ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ and a fixed sign\/ $\,\pm\hskip.7pt$, let\/ $\,y\,$ be a point
of the critical manifold $\,\varSigma^\pm\hskip-.7pt$. Then the following conclusions hold.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] The embedding\/
$\,\varPhi:N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\,$ with\/
{\rm(\ref{tgh})} has an extension to a totally geodesic hol\-o\-mor\-phic
immersion\/ $\,\varPsi:\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\toM}%{{\mathrm{M}}\hskip-.7pt$.
\item[{\rm(b)}] The mapping\/ $\,\varPsi\,$ in\/ {\rm(a)} restricted to the
projective hyper\-plane\/ $\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}
\subseteq\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})$ at infinity is a totally
geodesic hol\-o\-mor\-phic immersion\/
$\,F:\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\to\varSigma^\mp\hskip-.7pt$, and the metric that it
induces on\/ $\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ equals\/
$\,2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)/\hskip-.4pt a\,$ times the Fu\-bi\-ni-Stu\-dy metric,
cf.\ Remark\/~{\rm\ref{fbstm}}, arising from the inner product\/ $\,g_y^{\phantom i}$
in\/ $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\hskip.4pt$, \,for\/ $\,a,\vt\hskip-1.5pt_\pm\w$ as in
Remark\/~{\rm\ref{ascdt}(i)},
\item[{\rm(c)}] the images of the immersion\/
$\,F:\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\to\varSigma^\mp$ in\/ {\rm(b)} and of its
differential at any point\/ $\,\bbC\xi$, where\/
$\,(y,\xi)\inN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm$ and\/ $\,\xi\ne0$, coincide with the\/
$\,\pi^\mp\hskip-2pt$-im\-a\-ge of the leaf of\/
$\,\mathrm{Ker}\hskip2.3ptd\pi^{\hskip.7pt\pm}$ in\/ $\,M}%{{\mathrm{M}}'\hskip-.7pt$ passing through\/
$\hskip.7pt\,x=\varPhi(y,\xi)\,$ and, respectively, with the subspace\/
$\,d\pi\hskip-.7pt_x^\mp\hskip-.7pt(\mathcal{H}_x^\mp)
=d\pi\hskip-.7pt_x^\mp\hskip-.7pt(\mathcal{V}_{\hskip-.7pt x}^{\phantom i}\oplus\mathcal{H}_x^\mp)\,$ of\/
$\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\mp\hskip-.7pt$.
\end{enumerate}
}
\end{theorem}
\begin{proof}As a consequence of Theorems~\ref{first} and~\ref{dcomp}(c), the
composite $\,\pi^\mp\hskip-1.5pt\circ\varPhi\,$ maps
$\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\pm$ (the complement of the
zero section in $\,N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm$) hol\-o\-mor\-phic\-al\-ly into
$\,\varSigma^\mp\hskip-.7pt$. The restriction of $\,\pi^\mp\hskip-1.5pt\circ\varPhi\,$ to
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}
\subseteqN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\pm\hskip-.7pt$, being, by (\ref{pex})
and (\ref{phe}), constant on each punctured complex line through $\,0$, thus
descends to
\begin{equation}\label{dsc}
\mathrm{a\ hol\-o\-mor\-phic\ immersion\
}\,\,\,F:\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\to\varSigma^\mp,
\end{equation}
where the immersion property of $\,F\,$ is an immediate consequence of the
fact, established below, that both
$\,\pi^\mp\hskip-1.5pt:\varPhi\hskip.7pt(\nhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}\hskip-1.5pt)
\to\varSigma^\mp$ and
$\,\pi^\mp\hskip-1.5pt\circ\varPhi:N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}\to\varSigma^\mp$
have constant (complex) rank, equal to
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\hskip.7pt-\hs1$. As $\,\varPhi\,$ is a
bi\-hol\-o\-mor\-phism, it suffices to verify this last claim for the former
mapping; we do it noting that
$\,\varPi=\varPhi\hskip.7pt(\nhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\smallsetminus\{0\}\hskip-1.5pt)$
coincides with the $\,\pi^\pm\hskip-.7pt$-pre\-im\-age of $\,y\,$ (due to (\ref{pcf})
and Remark~\ref{ptnrs}), and hence forms a leaf of
$\,\mathrm{Ker}\hskip2.3ptd\pi^\pm\hskip-.7pt
=\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\mp$ restricted to $\,M}%{{\mathrm{M}}'\hskip-.7pt$, cf.\
(\ref{hpm}). That $\,\pi^\mp:\varPi\to\varSigma^\mp$ satisfies the required
rank condition is now clear: the kernel of its differential at any point
$\,x\,$ coincides, by (\ref{hpm}) and (\ref{tme}), with
$\,\mathcal{V}_{\hskip-.7pt x}^{\phantom i}$, while $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)$.
The mapping $\,\varPsi:\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\toM}%{{\mathrm{M}}\hskip-.7pt$,
equal to $\,\varPhi\,$ on $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ and to $\,F\,$ on
$\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, is continuous. Namely, if it were not,
we could pick a sequence
$\,\xi\hskip-.7pt_j^{\phantom i}\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, $\,j=1,2,\dots\hskip.7pt$, such that
$\,|\hskip.7pt\xi\hskip-.7pt_j^{\phantom i}|\to\infty\,$ and $\,\xi\hskip-.7pt_j^{\phantom i}/|\hskip.7pt\xi\hskip-.7pt_j^{\phantom i}|\to\xi\,$ as
$\,j\to\infty\,$ for some unit vector $\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, while no
sub\-se\-quence of the image sequence $\,\varPsi(\xi\hskip-.7pt_j^{\phantom i})\,$ tends to
$\,F(\bbC\xi)$. The resulting limit relation
$\,\sigma\hskip-3pt_j^{\phantom i}\to\delta$, where $\,\sigma\hskip-3pt_j^{\phantom i}$ corresponds
to $\,\rho\hskip-1.5pt_j^{\phantom i}=|\hskip.7pt\xi\hskip-.7pt_j^{\phantom i}|\,$ as in the line preceding (\ref{dsr}),
combined with (\ref{phe}), now gives
$\,\varPsi(\xi\hskip-.7pt_j^{\phantom i})=\varPhi(\xi\hskip-.7pt_j^{\phantom i})
=\mathrm{Exp}^\perp\hskip-.7pt(y,
\sigma\hskip-3pt_j^{\phantom i}\hskip.7pt\xi\hskip-.7pt_j^{\phantom i}\hskip-.7pt/\hskip-2.3pt\rho\hskip-1.5pt_j^{\phantom i}\hskip-1.5pt)$
which -- due to continuity of $\,\mathrm{Exp}^\perp$ and (\ref{pex}) --
converges to $\,\mathrm{Exp}^\perp\hskip-.7pt(y,\delta\xi)=y_\mp^{\phantom i}$, for a specific
point $\,y_\mp^{\phantom i}$. However, (a) -- (b) in Lemma~\ref{vsbkr} and the definition
of $\,F\,$ also give $\,y_\mp^{\phantom i}=F(\bbC\xi)$, which contradicts our choice of
$\,\xi\hskip-.7pt_j^{\phantom i}$, proving continuity of $\,\varPsi$.
Hol\-o\-mor\-phic\-i\-ty of $\,\varPsi\,$ is now obvious from Remark~\ref{hlext}
applied to $\,\varPi=\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})$ and its
co\-di\-men\-sion-one complex sub\-man\-i\-fold
$\,\varLambda=\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$. Furthermore,
\begin{equation}\label{imm}
\varPsi\,\,\,\mathrm{is\ an\ immersion}.
\end{equation}
To see this, first note that $\,\varPsi\,$ has two restrictions, $\,F\,$ to
$\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ and $\,\varPhi\,$ to the dense open
sub\-man\-i\-fold $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, already known to be immersions,
the former into $\,\varSigma^\mp\hskip-1.5pt$, cf.\ (\ref{tgh}) -- (\ref{dsc}). Next, for
any unit vector $\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$, if $\,\varLambda\hskip-.4pt'$ denotes
the projective line in
$\,\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\,$ joining $\,\bbC(1,0)\,$ to the
point $\,\bbC\xi\in\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ (identified via
(\ref{inc}) with $\,\bbC(0,\xi)$), then the restriction of $\,\varPsi\,$ to
$\,\varLambda\hskip-.4pt'$ is an embedding with the image
$\,\varLambda=\varPsi(\varLambda\hskip-.4pt'\hskip.4pt)\,$ forming a complex sub\-man\-i\-fold
of $\,M}%{{\mathrm{M}}\hskip-.7pt$, bi\-hol\-o\-mor\-phic to $\,\bbC\mathrm{P}^1\hskip-.7pt$, and intersecting each
of $\,\varSigma^+$ and $\,\varSigma^-$ orthogonally at a single point. In fact,
Lemma~\ref{vsbkr} yields all the claims just made except the `embedding'
property; we obtain the latter from Remark~\ref{known}(b), which we use to
conclude that the resulting hol\-o\-mor\-phic mapping
$\,\varPsi\hskip-1.5pt:\varLambda\hskip-.4pt'\hskip-.4pt\to\varLambda$, being injective (since so is
$\,\varPhi$), must be a bi\-hol\-o\-mor\-phism. Now (\ref{imm}) follows.
For obvious reasons of continuity, (\ref{tgh}) implies that the
hol\-o\-mor\-phic immersion
$\,\varPsi:\mathrm{P}(\bbC\times\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i})\toM}%{{\mathrm{M}}\,$ is totally
geodesic, which establishes (a). Finally, Remarks~\ref{fbstu},
\ref{ascdt}(iii), \ref{ttgim} and Theorem~\ref{first} give rise to (b),
completing the proof.
\end{proof}
\begin{remark}\label{codim}For $\,m,d_\pm^{\phantom i},k_\pm^{\phantom i},q\,$ as in
Remark~\ref{dppdm}, the co\-di\-men\-sion
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\varSigma^\mp\hskip-1.5pt-\hskip.4pt\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ of the immersion
$\,F\,$ in Theorem~\ref{tgimm}(b) equals $\,q$. In fact,
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}=m-d_\pm^{\phantom i}\hskip-.7pt-1$, and so, by (\ref{dmq}),
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\varSigma^\mp\hskip-.7pt-\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}=(m-d_\pm^{\phantom i}\hskip-.7pt-1)-d_\mp^{\phantom i}=q$.
\end{remark}
\begin{remark}\label{spmcp}Suppose that the distribution $\,\mathcal{H}\,$ in
(\ref{tme}) is \hbox{$\,0$-}\hskip0ptdi\-men\-sion\-al or, in other words,
$\,T\nnhM}%{{\mathrm{M}}'\,=\,\,\mathcal{V}\hskip1.2pt\oplus\hskip.7pt\mathcal{H}^+\hskip-1.5pt
\oplus\hskip.7pt\mathcal{H}^-\hskip-1.5pt$. Then, for either sign $\,\pm\hskip.7pt$, the critical
manifold $\,\varSigma^\pm\hskip-.7pt$, with its sub\-man\-i\-fold metric, must be
bi\-hol\-o\-mor\-phic\-al\-ly isometric to a complex projective space carrying
the Fu\-bi\-ni-Stu\-dy metric multiplied by
$\,2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)/\hskip-.4pt a\hskip.4pt$.
In fact, the isometric immersion $\,F\,$ of Theorem~\ref{tgimm}(b), having
co\-di\-men\-sion zero (cf.\ Remark~\ref{codim}), is necessarily a
bi\-hol\-o\-mor\-phism (Remark~\ref{ftcvr}).
\end{remark}
\section{Consequences of condition {\rm(\ref{spn})}}\label{ch}
\setcounter{equation}{0}
The results stated and proved below use Definition~\ref{ggktr}, the notations
of (\ref{sym}), (\ref{dbp}), (\ref{hpm}), and the notion of projectability
introduced in Section~\ref{pd}.
\begin{lemma}\label{intpr}{\medit
For a compact ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, the following three conditions are mutually equivalent.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(i)}] The distribution\/
$\,\hskip.7pt\mathcal{Z}=\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$ on\/
$\,M}%{{\mathrm{M}}'\hskip.4pt$ is in\-te\-gra\-ble.
\item[{\rm(ii)}] $\mathrm{Ker}\hskip2.3ptd\pi^-\hskip.7pt
=\,\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\,$ is\/
$\,\pi^+\hskip-2pt$-pro\-ject\-able.
\item[{\rm(iii)}] $\mathrm{Ker}\hskip2.3ptd\pi^+\hskip.7pt
=\,\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^-\,$ is\/
$\,\pi^-\hskip-2pt$-pro\-ject\-able.
\end{enumerate}
In\/ {\rm(ii)} -- {\rm(iii)} one may also replace\/
$\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\pm\,$ by\/ $\,\mathcal{H}^\pm$ or\/
$\,\mathcal{Z}$. If\/ {\rm(i)} -- {\rm(iii)} hold, then\/{\rm:}
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(iv)}] The immersions of Theorem\/~{\rm\ref{tgimm}(c)} are all
embeddings.
\item[{\rm(v)}] The\/ $\,\pi^\pm\hskip-1.2pt$-im\-a\-ges\/ $\,\mathcal{Z}^\pm\hskip-1.5pt$ of
the in\-te\-gra\-ble distribution\/ $\,\hskip.7pt\mathcal{Z}\,$ on\/ $\,M}%{{\mathrm{M}}'$ are
in\-te\-gra\-ble hol\-o\-mor\-phic distributions on\/ $\,\varSigma^\pm$ and have
totally geodesic leaves bi\-hol\-o\-mor\-phic\-al\-ly isometric to complex
projective spaces carrying\/ $\,2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)/\hskip-.4pt a\,$ times the
Fu\-bi\-ni-Stu\-dy metric, cf.\ Theorem\/~{\rm\ref{tgimm}(b)}. These leaves
coincide with the images of the embeddings in\/ {\rm(iv)}, and form the fibres
of hol\-o\-mor\-phic bundle projections\/
$\,\mathrm{pr}^\pm\hskip-1.5pt:\varSigma^\pm\hskip-1.5pt\toB^\pm\,$ for some compact complex base
manifolds\/ $\,B^\pm\hskip-1.5pt$.
\item[{\rm(vi)}] The summand\/ $\,\mathcal{H}\,$ in\/ {\rm(\ref{tme})} is\/
$\,\pi^\pm\hskip-1.2pt$-pro\-ject\-able and its\/
$\,\pi^\pm\hskip-1.2pt$-im\-a\-ge coincides with the orthogonal complement
of\/ $\,\mathcal{Z}^\pm$ in\/ $\,T\hskip-1.5pt\varSigma^\pm\hskip-.7pt$.
\item[{\rm(vii)}] The leaf space\/ $\,B=M}%{{\mathrm{M}}'\hskip-3pt/\hskip-1.8pt\mathcal{Z}\,$
admits a unique structure of a compact complex manifold such that the
quo\-tient projection\/ $\,M}%{{\mathrm{M}}'\hskip-.7pt\toM}%{{\mathrm{M}}'\hskip-3pt/\hskip-1.8pt\mathcal{Z}\,$
constitutes a hol\-o\-mor\-phic fibration while, for either sign\/ $\,\pm\,$
and\/ $\,\mathrm{pr}^\pm\hskip-1.5pt:\varSigma^\pm\hskip-1.5pt\toB^\pm\,$ as in\/ {\rm(iv)}, the
mapping\/ $\,B\toB^\pm\hskip-1.5pt$, sending each leaf
of\/ $\,\mathcal{Z}\,$ to its image under\/ $\,\mathrm{pr}^\pm\hskip-1.5pt\circ\pi^\pm\hskip-1.5pt$,
is a bi\-hol\-o\-mor\-phism.
\item[{\rm(viii)}] There exists a unique hol\-o\-mor\-phic bundle projection\/
$\,\pi:M}%{{\mathrm{M}}\toB\,$ with\/ $\,\mathrm{Ker}\hskip2.3ptd\pi=\mathcal{Z}$ on\/
$\,M}%{{\mathrm{M}}'$ such that, for both signs\/ $\,\pm\hskip.7pt$, the restriction of\/
$\,\pi\,$ to\/ $\,M}%{{\mathrm{M}}'$ equals\/
$\,\beta^\pm\hskip-1.5pt\circ\mathrm{pr}^\pm\hskip-1.5pt\circ\pi^\pm\hskip-.7pt$, where\/
$\,\beta^\pm$ is the inverse of the bi\-hol\-o\-mor\-phism\/
$\,B\toB^\pm$ in\/ {\rm(vii)}.
\item[{\rm(ix)}] $R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,
=\,-ia\hskip.4pt(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)^{-\nnh1}h(J\hskip-.7pt w,w\hskip.4pt'\hskip.4pt):N}%{{\mathrm{N}}\hskip-.7pt\toN}%{{\mathrm{N}}\hskip-1.5pt$,
with the notation of\/ {\rm(\ref{nwt})}, for the sub\-man\-i\-fold metric\/
$\,h\,$ of\/ $\,\varSigma^\pm\hskip-.7pt$, the normal connection\/ $\,\mathrm{D}\,$ in its
normal bundle\/ $\,N}%{{\mathrm{N}}=N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-.7pt$, any vector field\/
$\,w\hskip.4pt'$ on\/ $\,\varSigma^\pm\hskip-.7pt$, and any section\/ $\,w\,$ of\/ $\,\mathcal{Z}^\pm\hskip-.7pt$,
cf.\ {\rm(v)}.
\end{enumerate}
}
\end{lemma}
\begin{proof}Since $\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\pm$ are both
in\-te\-gra\-ble by (\ref{hpm}), the mutual equivalence of (i), (ii), (iii)
and the in\-te\-gra\-bil\-i\-ty claim in (v) are all immediate from
Lemma~\ref{intgr} applied to
$\,\mathcal{E}^\pm\hskip-1.5pt=\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\pm\hskip-1.5pt$, along with
(\ref{hpm}) and (\ref{tme}). The immersions mentioned in
Theorem~\ref{tgimm}(c) thus have nonsingular images, namely, the leaves
$\,\varPi\,$ of the distribution $\,\mathcal{Z}^\pm$ in (v), so that (iv)
follows from Remark~\ref{ftcvr} applied to
$\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ standing for $\,\bbC\mathrm{P}\hskip.4pt^l\hskip-.7pt$, with
$\,l=k_\mp^{\phantom i}$ defined in Remark~\ref{dppdm}, and such a leaf $\,\varPi\hskip-.7pt$.
The remaining part of (v) is a direct consequence of Theorem~\ref{tgimm}(b)
and Remark~\ref{fibra}.
At any $\,y\in\varSigma^\pm\hskip-.7pt$, the image
$\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)$ is now independent of the choice
of $\,x\inM}%{{\mathrm{M}}'$ with $\,\pi^\pm\hskip-.7pt(x)=y$, and hence so is its orthogonal
complement $\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^{\phantom i})$ in $\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ (see
Remark~\ref{inequ}(iii)), proving assertion (vi).
The mappings $\,B\toB^\pm$ in (vii) are obviously bijective, and lead to
an identification $\,B^+\hskip-1.5pt=\hskip.7ptB^-$ which is a bi\-hol\-o\-mor\-phism, as
one sees restricting $\,\pi^\pm$ to ``local'' complex sub\-man\-i\-folds of
$\,M}%{{\mathrm{M}}'$ which the composite bundle projections
$\,M}%{{\mathrm{M}}'\hskip-1.5pt\to\varSigma^\pm\hskip-1.5pt\toB^\pm$ (with fibres provided by the leaves of
$\,\mathcal{Z}$) send bi\-hol\-o\-mor\-phic\-al\-ly onto open sub\-man\-i\-folds of
$\,B^\pm\hskip-1.5pt$. This yields (vii). For (viii), it suffices to note that the
two composite bundle projections
$\,\mathrm{pr}^\pm\hskip-.7pt\circ\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-1.5pt
\toB\,$ agree, by (vii), on the intersection $\,M}%{{\mathrm{M}}'$ of their domains, cf.\
Remark~\ref{ascdt}(iv), while the union of their domains is $\,M}%{{\mathrm{M}}\hskip-.7pt$.
For (ix), Theorem~\ref{first} allows us to identify
$\,M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp$ with $\,N}%{{\mathrm{N}}\,$ so that (\ref{hgv}.c) and
(\ref{gsw}) hold under the assumptions following (\ref{hgv}). Since $\,w\,$
lies in the $\,\pi^\pm\hskip-1.2pt$-im\-a\-ge $\,\mathcal{Z}^\pm\hskip-1.5pt$ of
$\,\mathcal{H}^\pm\hskip-1.5pt$, cf.\ (ii), (iii), (v), formula (\ref{hvk}) gives
$\,2S\hskip-.7pt w=Qw/(\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w)\,$ for its $\,\mathrm{D}\hskip.4pt$-hor\-i\-zon\-tal
lift, also denoted by $\,w$. Replacing
$\,2S\hskip-.7pt w\,$ in (\ref{gsw}) with $\,Qw/(\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w)\,$ and multiplying the
result by $\,(\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w)\hskip.4pt Q^{-\nnh1}\hskip-.7pt$, we get an expression for
$\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ which, equated to (\ref{hgv}.c), yields
$\,\langle\hskip-.7pt R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)\hskip.4pt\xi,i\hskip.4pt\xi\rangle=
-\hskip.4pt a\hskip.4pt(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)^{-\nnh1}\langle\xi,\xi\rangle\hskip.7pt h(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)$,
since $\,\rho^2\hskip-.7pt=\langle\xi,\xi\rangle$ while, obviously,
$\,|\hskip.4pt\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w|=\mp\hskip.4pt(\vt-\hskip-.4pt\vt\hskip-1.5pt_\pm\w)$. Applying the last equality to
$\,J\hskip-.7pt w\,$ instead of $\,w$, and using (b) in Section~\ref{ck} along with
Her\-mit\-i\-an symmetry of
$\,\langle\hskip-.7pt R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt\xi,i\hskip.4pt\eta\rangle
=-\langle i\hskip.4pt R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\hskip.4pt\xi,\eta\rangle$ in $\,\xi,\eta$,
we obtain the required relation in (ix)
\end{proof}
Note that the above proof of (ix) in Lemma~\ref{intpr} actually uses the
assumptions (i) -- (iii): without them, the formula
$\,\langle\hskip-.7pt R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,J\hskip-.7pt w\hskip.4pt'\hskip.4pt)\hskip.4pt\xi,i\hskip.4pt\xi\rangle=
-\hskip.4pt a\hskip.4pt(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)^{-\nnh1}\langle\xi,\xi\rangle\hskip.7pt h(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)$,
rather than being valid for any given $\,w\in\mathcal{Z}_y^\pm\hskip-.7pt$,
$\,y\in\varSigma^\pm\hskip-.7pt$, and {\medit all\/} vectors $\,\xi\,$ normal to $\,\varSigma^\pm$
at $\,y$, would hold only when $\,w\,$ lies in some subspace of $\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$
depending on $\,\xi$.
Let us now fix a K\"ah\-ler manifold $\,(\hat\varSigma,\hat h)$, and consider
pairs $\,N}%{{\mathrm{N}}\hskip-.7pt,\langle\,,\rangle\,$ formed by a hol\-o\-mor\-phic complex vector bundle
$\,N}%{{\mathrm{N}}\,$ over $\,\hat\varSigma\,$ and the real part $\,\langle\,,\rangle\,$ of a Her\-mit\-i\-an
fibre metric in $\,N}%{{\mathrm{N}}\hskip-1.5pt$, the Chern connection of which -- see
Section~\ref{ck} -- satisfies the curvature condition
$\,R^{\mathrm{D}}\hskip-.7pt(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)
=2i\hskip.4pt\hat h(J\hskip-.7pt w,w\hskip.4pt'\hskip.4pt):N}%{{\mathrm{N}}\hskip-.7pt\toN}%{{\mathrm{N}}$
for any vector fields $\,w,w\hskip.4pt'$ tangent to $\,\hat\varSigma$, where the notation
of (\ref{nwt}) is used.
\begin{lemma}\label{rdehj}{\medit
Whenever\/ $\,\hat\varSigma\,$ is simply connected and such\/
$\,N}%{{\mathrm{N}}\hskip-.7pt,\langle\,,\rangle\,$ exist, they are essentially unique, in the sense that, given
another pair\/ $\,N}%{{\mathrm{N}}'\hskip-.7pt,\langle\,,\rangle\hskip-.4pt'$ with the same property, some
hol\-o\-mor\-phic vec\-tor-bun\-dle isomorphism\/ $\,N}%{{\mathrm{N}}\hskip-.7pt\toN}%{{\mathrm{N}}'$ takes\/
$\,\langle\,,\rangle\,$ to\/ $\,\langle\,,\rangle\hskip-.4pt'\hskip-.7pt$.
}
\end{lemma}
\begin{proof}Remark~\ref{crvcm} implies that the Chern connections
$\,\mathrm{D}\,$ and $\,\mathrm{D}'$ induce a flat metric connection in the
bundle $\,\mathrm{Hom}\hskip.7pt(N}%{{\mathrm{N}}\hskip-.7pt,N}%{{\mathrm{N}}')$. The required isomorphism is now
provided by a global parallel section of $\,\hskip.7pt\mathrm{Hom}\hskip.7pt(N}%{{\mathrm{N}}\hskip-.7pt,N}%{{\mathrm{N}}')\,$
chosen so as to transform $\,\langle\,,\rangle$ into $\,\langle\,,\rangle\hskip-.4pt'$ at one point, and its
hol\-o\-mor\-phic\-i\-ty follows from (e) in Section~\ref{ck}.
\end{proof}
\begin{theorem}\label{cptrp}{\medit
For a compact ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$,
the following two conditions are equivalent.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] $(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is iso\-mor\-phic to a \,CP\ triple,
defined as in Section\/~{\rm\ref{eg}}.
\item[{\rm(ii)}] $d_+^{\phantom i}+\hskip.7pt\,d_-^{\phantom i}\hskip.7pt=\,m\,-\,1$, where\/ $\,m=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}\,$
and\/ $\,d_\pm^{\phantom i}\hskip-.7pt=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\varSigma^\pm\hskip-.7pt$. In other words, cf.\
Remark\/~{\rm\ref{dppdm}},
$\,T\nnhM}%{{\mathrm{M}}'\hskip.7pt=\,\,\mathcal{V}\hskip1.2pt\oplus\hskip.7pt\mathcal{H}^+\hskip-1.5pt
\oplus\hskip.7pt\mathcal{H}^-\hskip-1.5pt$, that is, $\,\mathcal{H}\,$ in\/ {\rm(\ref{tme})}
is\/ \hbox{$\,0$-}\hskip0ptdi\-men\-sion\-al.
\end{enumerate}
In this case, the assertion of \,Theorem\/~{\rm\ref{first}}, including\/
{\rm(\ref{alt})}, is satisfied by\/
$\,(M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-1.5pt,g,\vt)$, with either fixed sign\/
$\,\pm\,$ and\/ $\,(\varSigma,h)\,$ bi\-hol\-o\-mor\-phic\-al\-ly isometric
to a complex projective space carrying\/
$\,2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)/\hskip-.4pt a\,$ times the Fu\-bi\-ni-Stu\-dy metric,
$\,N}%{{\mathrm{N}}\,$ and\/ $\,\langle\,,\rangle\,$ being, up to a hol\-o\-mor\-phic vec\-tor-bun\-dle
isomorphism, the normal bundle of the latter treated as a linear variety in\/
$\,\bbC\mathrm{P}^m$ and its Her\-mit\-i\-an fibre metric induced by the
Fu\-bi\-ni-Stu\-dy metric of\/ $\,\bbC\mathrm{P}^m\hskip-.7pt$.
Furthermore, the isomorphism types of \,CP\ triples\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ having
any given values of\/ $\,d_\pm^{\phantom i}$ and\/ $\,m\,$ in\/ {\rm(ii)} are in a
natural bijective correspondence, obtained by applying
Remark\/~{\rm\ref{ascdt}(i)}, with quadruples\/
$\,\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w,a,\hskip.7pt\vt\mapsto Q\,$ that satisfy\/ {\rm(\ref{pbd})}.
}
\end{theorem}
\begin{proof}First, (i) implies (ii) according to (\ref{dem}).
Assuming now (ii), let us use Remark~\ref{spcas} to select a CP triple
$\,(\bbC\mathrm{P}^m\hskip-.7pt,g'\hskip-.7pt,\vt')$ realizing the same data $\,d_\pm^{\phantom i},\vt\hskip-1.5pt_\pm\w,a\,$ and
$\,\vt\mapsto Q$, in (ii) above and Remark~\ref{ascdt}(i), as our
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ (which also establishes the surjectivity part of the final
clause). With either fixed sign $\,\pm\hskip.7pt$, denoting
$\,\varSigma^\mp\hskip-.7pt,\varSigma^\pm$ by $\,\varSigma,\varPi$, and their analogs for
$\hskip.7pt(\bbC\mathrm{P}^m\hskip-.7pt,g'\hskip-.7pt,\vt')\hskip.7pt$ by $\,\varSigma'\hskip-.7pt,\varPi'\hskip-.7pt$, we choose the
isomorphisms $\,N}%{{\mathrm{N}}\hskip-.7pt\toM}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varPi$ and
$\,N}%{{\mathrm{N}}'\hskip-.7pt\to\bbC\mathrm{P}^m\hskip-.7pt\hskip-1.5pt\smallsetminus\hskip-.7pt\varPi'$ by applying
Theorem~\ref{first}(i) to both triples. As (i) has already been shown to yield
(ii), we may now also apply Remark~\ref{spmcp} to both of them, identifying
the critical manifolds $\,\varSigma,\varSigma'$ (and their sub\-man\-i\-fold metrics) with
a complex projective space $\,\hat\varSigma\,$ (and, respectively, with the
Fu\-bi\-ni-Stu\-dy metric $\,\hat h\,$ multiplied by
$\,2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)/\hskip-.4pt a$). Next, (ix) in Lemma~\ref{intpr} holds
for both triples, so that the pairs $\,N}%{{\mathrm{N}}\hskip-.7pt,\langle\,,\rangle\,$ and $\,N}%{{\mathrm{N}}'\hskip-.7pt,\langle\,,\rangle\hskip-.4pt'$
associated with them via Theorem~\ref{first} satisfy, along with
$\,\hat\varSigma=\varSigma=\varSigma'$ and $\,\hat h$, the assumptions -- as well as the
conclusion -- of Lemma~\ref{rdehj}. Thus, some hol\-o\-mor\-phic
vec\-tor-bun\-dle isomorphism $\,N}%{{\mathrm{N}}\hskip-.7pt\toN}%{{\mathrm{N}}'$ takes $\,\langle\,,\rangle\,$ to
$\,\langle\,,\rangle\hskip-.4pt'$ and, since the metrics $\,\hat{g\hskip2pt}\hskip-1.3pt,\hat{g\hskip2pt}\hskip-1.3pt'$ on $\,N}%{{\mathrm{N}}\,$ and $\,N}%{{\mathrm{N}}'$
constructed in Section~\ref{ev} depend only on $\,\langle\,,\rangle,\langle\,,\rangle\hskip-.4pt'$ (aside from the
data fixed above and shared by both triples), this isomorphism is a
hol\-o\-mor\-phic isometry of $\,(N}%{{\mathrm{N}}\hskip-.7pt,\hat{g\hskip2pt}\hskip-1.3pt)$ onto $\,(N}%{{\mathrm{N}}'\hskip-.7pt,\hat{g\hskip2pt}\hskip-1.3pt')$,
sending $\,\vt\,$ to its analog on $\,N}%{{\mathrm{N}}'\hskip-.7pt$. In view of
\cite[Lemma 16.1]{derdzinski-maschler-06}, it can be extended to an
isomorphism between the triples $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ and
$\,(\bbC\mathrm{P}^m\hskip-.7pt,g'\hskip-.7pt,\vt')$. We thus obtain injectivity in the final clause
and the fact that (ii) yields (i).
\end{proof}
\section{Horizontal extensions of CP triples}\label{he}
\setcounter{equation}{0}
Once again, we use the notation of (\ref{sym}), (\ref{tmm}) and (\ref{nex}),
assuming $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ to be a compact ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple (Definition~\ref{ggktr}).
\begin{lemma}\label{ttgds}{\medit
Suppose that\/ conditions\/ {\rm(i)} -- {\rm(iii)} along with the other
assumptions of Lemma\/~{\rm\ref{intpr}} hold for a triple\/
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, and\/ $\,\pi,B\,$ are as in
Lemma\/~{\rm\ref{intpr}(viii)}.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] Given a\/ $\,\pi$-pro\-ject\-a\-ble nonzero local section\/
$\,w\,$ of the distribution\/ $\,\mathcal{H}\,$ in\/ {\rm(\ref{tme})},
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a1)}] $w\,$ commutes with the vector fields\/ $\,v=\nabla\hskip-.5pt\vt\,$ and\/
$\,u=J\hskip-.7pt v$,
\item[{\rm(a2)}] $w\,$ is $\,\pi^\pm\hskip-1.2pt$-pro\-ject\-able for both
signs\/ $\,\pm\hskip.7pt$,
\item[{\rm(a3)}] the local flow of\/ $\,w\,$ in\/ $\,M}%{{\mathrm{M}}'$ preserves the
distributions\/ $\,\mathcal{V}\hskip-.7pt,\mathcal{H}^+$ and\/ $\,\mathcal{H}^-\hskip-1.5pt$.
\end{enumerate}
\item[{\rm(b)}] The leaves of the in\-te\-gra\-ble distribution\/
$\,\hskip.7pt\mathcal{Z}=\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$ on\/
$\,M}%{{\mathrm{M}}'\hskip.4pt$ are totally geodesic complex sub\-man\-i\-folds of\/
$\,M}%{{\mathrm{M}}'\hskip.4pt$ and all the local flows mentioned in\/ {\rm(a3)} act between them
via local isometries.
\end{enumerate}
}
\end{lemma}
\begin{proof}Any $\,w\,$ in (a) is normal to the totally geodesic leaves of
the in\-te\-gra\-ble distributions $\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\pm$
(see Corollary~\ref{totgd}), while $\,v,u\,$ are both tangent to them, as
$\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)$. Therefore,
$\,\nabla\hskip-3pt_v^{\phantom i} w\,$ and $\,\nabla\hskip-3pt_u^{\phantom i} w$, being, as a
result, also normal to those leaves for both signs\/ $\,\pm\hskip.7pt$, are -- by
(\ref{tme}) -- sections of $\,\mathcal{H}$. The same is true of
$\,\nabla\hskip-3pt_w^{\phantom i} v,\nabla\hskip-3pt_w^{\phantom i} u$ (and hence of
$\,[v,w],[u,w]$) due to $\,S$-in\-var\-i\-ance in (\ref{tme}), with
$\,S=\nabla\hskip-.7pt v$ and $\,\nabla\hskip-.7pt u=A=JS=SJ$, cf.\ (\ref{loc}.a). At the
same time, $\,\pi$-pro\-ject\-a\-bil\-i\-ty of $\,w$ implies, via
Remark~\ref{liebr} and Lemma~\ref{intpr}(viii), that $\,[v,w]\,$ and
$\,[u,w]\,$ are sections of
$\,\mathcal{Z}=\mathrm{Ker}\hskip2.3ptd\pi=\mathcal{H}^\perp$. We thus obtain (a1)
along with (a3) for $\,\mathcal{V}\hskip-.7pt$. Next, (a2) follows: due to
$\,\pi$-pro\-ject\-a\-bil\-i\-ty of $\,w$, with $\,y\in\varSigma^\pm$ fixed,
$\,d\pi\hskip-1.5pt_x^{\phantom i} w_x^{\phantom i}$ is independent of the choice of $\,x\inM}%{{\mathrm{M}}'$ such that
$\,\pi^\pm\hskip-.7pt(x)=y$, and hence so must be $\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}w_x^{\phantom i}$, as the
differential at $\,y\,$ of the bundle projection
$\,\beta^\pm\hskip-1.5pt\circ\mathrm{pr}^\pm\hskip-1.5pt:\varSigma^\pm\hskip-1.5pt\toB\,$ (see (vii) --
(viii) in Lemma~\ref{intpr}) sends $\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}w_x^{\phantom i}$ to
$\,d\pi\hskip-1.5pt_x^{\phantom i} w_x^{\phantom i}$, which determines $\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}w_x^{\phantom i}$ uniquely
due to its being orthogonal, by Lemma~\ref{intpr}(v), to $\,\mathcal{Z}_y^\pm\hskip-.7pt$,
for the vertical distribution $\,\mathcal{Z}^\pm$ of
$\,\beta^\pm\hskip-1.5pt\circ\mathrm{pr}^\pm\hskip-1.5pt$.
We obtain the remainder of (a3) by noting that, for either fixed sign
$\,\pm\hskip.7pt$,
\begin{equation}\label{lft}
\begin{array}{l}
\mathrm{all\ vectors\ in\ }\,\,\mathcal{H}^\pm\mathrm{\ are\ realized\ by\
}\,\pi^\pm\hskip-1.2pt\hyp\mathrm{pro\-ject\-able\ local}\\
\mathrm{sections\ }\hskip.7pt w^\pm\hskip-1.5pt\mathrm{\ of\ }\hskip.7pt\mathcal{H}^\pm\hskip-1.5pt\mathrm{\
commuting\ with\ }\hskip.7pt w\hskip.4pt\mathrm{,\ the\
}\hskip.7pt\pi^\pm\hskip-1.2pt\hyp\mathrm{im\-a\-ges}\\
\hat w^\pm\hskip-1.5pt\mathrm{\ of\ which\ also\ commute\ with\ the\
}\pi^\pm\hskip-1.2pt\hyp\mathrm{im\-a\-ge\ \hskip-.7pt}\,\hat w\,\mathrm{\hskip-.7pt\ of\
\hskip-.7pt}\,w.
\end{array}
\end{equation}
Namely, since the $\,\pi^\pm\hskip-1.2pt$-im\-a\-ge $\,\hat w\,$ of $\,w\,$ is
obviously $\,\mathcal{Z}^\pm\hskip-1.2pt$-pro\-ject\-able, we may prescribe the
$\,\pi^\pm\hskip-1.2pt$-im\-a\-ge $\,\hat w^\pm$ of $\,w^\pm$ to be a local
section of $\,\mathcal{Z}^\pm$ commuting with $\,\hat w\,$ (see (v) -- (vi) in
Lemma~\ref{intpr} and Remark~\ref{cmmut}), and then lift $\,\hat w^\pm$ to
$\,\mathcal{H}^\pm\hskip-.7pt$, using Remark~\ref{inequ}(iii). For the resulting lift
$\,w^\pm\hskip-.7pt$, (\ref{cmt}) and parts (ix), (vi) of Lemma~\ref{intpr} give
$\,[w,w^\pm]=0$.
We now derive (b) from Remark~\ref{tglvs}. According to Remark~\ref{frthr}, it
suffices to establish (i) in Remark~\ref{tglvs} for local sections of
$\,\mathcal{Z}\,$ having the form $\,w\hskip.4pt'\hskip-.7pt=w^0\hskip-.7pt+\hskip.4pt w^+\hskip-1.5pt+\hskip.4pt w^-$ with
$\,w^\pm$ satisfying (\ref{lft}) and $\,w^0$ equal to a
con\-stant-co\-ef\-fi\-cient combination of $\,v$ and $\,u$. Orthogonality in
(\ref{tme}) combined with (\ref{gvv}) shows that
$\,g(\hskip-.4pt w\hskip.4pt'\hskip-.7pt,w\hskip.4pt'\hskip.4pt)\,$ equals a constant multiple of $\,Q\,$ plus the
sum of the terms $\,g(\hskip-.4pt w^\pm\hskip-.7pt,w^\pm\hskip-.7pt)$. Verifying part (i) of
Remark~\ref{tglvs} thus amounts to showing that
$\,d_w^{\phantom i} Q=d_w^{\phantom i}[\hskip.7pt g(\hskip-.4pt w^\pm\hskip-.7pt,w^\pm\hskip-.7pt)]=0$. In terms of the
$\,\pi^\pm\hskip-1.2pt$-im\-a\-ge $\,\hat w^\pm$ of $\,w^\pm\hskip-.7pt$,
Remark~\ref{inequ}(iv) gives
$\,(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\vt\hskip-1.7pt_-\w)\hskip.7pt g(\hskip-.4pt w^\pm\hskip-.7pt,w^\pm\hskip-.7pt)
=(\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w)\hskip.7pt h(\hat w^\pm\hskip-1.5pt,\hat w^\pm\hskip-.7pt)$, where $\,h\,$ is the
sub\-man\-i\-fold metric of $\,\varSigma^\pm\hskip-.7pt$. Since $\,\vt\hskip-1.5pt_\pm\w$ are constants, our
claim is thus reduced to two separate parts, $\,d_w^{\phantom i} Q=d_w^{\phantom i}\vt=0\,$ and
$\,d_w^{\phantom i}[h(\hat w^\pm\hskip-1.5pt,\hat w^\pm\hskip-.7pt)]=0$. The former part is immediate:
$\,Q\,$ is a function of $\,\vt$, cf.\ Remark~\ref{ascdt}(i), while $\,w\,$
and $\,v=\nabla\hskip-.5pt\vt\,$ are sections of the mutually orthogonal summands
$\,\mathcal{H}\,$ and $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)$ in
(\ref{tme}). For the latter part, (a2) allows us to replace $\,w\,$ by its
$\,\pi^\pm\hskip-1.2pt$-im\-a\-ge $\,\hat w$, noting that
$\,h(\hat w^\pm\hskip-1.5pt,\hat w^\pm\hskip-.7pt)$ is the
$\,\pi^\pm\hskip-1.2pt$-pull\-back of a function defined (locally) in
$\,\varSigma^\pm\hskip-.7pt$. Now $\,d_w^{\phantom i}[h(\hat w^\pm\hskip-1.5pt,\hat w^\pm\hskip-.7pt)]=0$ due to the
fact that (ii) implies (i) in Remark~\ref{tglvs}, and $\,\hat w$, or
$\,\hat w^\pm\hskip-.7pt$, is normal or, respectively, tangent to the totally geodesic
leaves of the in\-te\-gra\-ble distribution $\,\mathcal{Z}^\pm\hskip-.7pt$, while $\,\hat w$,
besides being -- as noted above -- projectable along $\,\mathcal{Z}^\pm\hskip-.7pt$, also
commutes with $\,\hat w^\pm$ (see (v) -- (vi) in Lemma~\ref{intpr} and
(\ref{lft})).
\end{proof}
We say that a (lo\-cal\-ly-triv\-i\-al) hol\-o\-mor\-phic fibre bundle
carries a specific lo\-cal-type {\medit fibre geometry\/} if such a geometric
structure is selected in each of its fibres and suitable local
$\,C^\infty$ trivializations make the structures appear the same in all
nearby fibres. For instance, hol\-o\-mor\-phic complex vectors bundle endowed
with Her\-mit\-i\-an fibre metrics may be referred to as
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\roman{enumi}}}
\item[{\rm(i)}] hol\-o\-mor\-phic bundles of Her\-mit\-i\-an vector spaces.
\end{enumerate}
The fact that (i) leads to the presence of the distinguished Chern connection
(Section~\ref{ck}) has obvious generalizations to two situations (ii) -- (iii)
discussed below.
By a {\medit horizontal distribution\/} for a hol\-o\-mor\-phic bundle
projection $\,\pi:M}%{{\mathrm{M}}\toB$ between complex manifolds, also called a
{\medit connection in the hol\-o\-mor\-phic bundle\/} $\hsM}%{{\mathrm{M}}$ \hbox{over
$\,B$,} we mean any $\,C^\infty$ real vector sub\-bun\-dle
$\,\mathcal{H}\,$ of $\,T\nnhM}%{{\mathrm{M}}\hskip-.7pt$, complementary to the vertical
distribution $\,\mathrm{Ker}\hskip2.3ptd\pi$, so that
$\,T\nnhM}%{{\mathrm{M}}\,$ is the direct sum of $\,\mathrm{Ker}\hskip2.3ptd\pi\,$ and
$\,\mathcal{H}$. {\medit Horizontal lifts\/} of vectors tangent to $\,B$,
and of piecewise $\,C^1$ curves in $\,B$, as well as {\medit parallel
transports\/} along such curves, are then defined in the usual fashion,
although the maximal domain of a lift of a curve (or, of a parallel transport)
may in general be a proper sub\-in\-ter\-val of the original domain interval.
This last possibility does not, however, occur in bundles with compact fibres,
or in vector bundles with linear connections, where horizontal lifts of curves
and parallel transports are all {\medit global}.
We proceed to describe the {\medit Chern connection\/} $\,\mathcal{H}\,$ in
the cases of
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(ii)}] hol\-o\-mor\-phic bundles of Fu\-bi\-ni-Stu\-dy complex
projective spaces, and
\item[{\rm(iii)}] hol\-o\-mor\-phic bundles of \,CP triples, over any complex
manifold $\,B$.
\end{enumerate}
Their fibre geometries consist of
Fu\-bi\-ni-Stu\-dy metrics (Remark~\ref{fbstm}) and, respectively, the
structures of a \,CP triple (Section~\ref{eg}).
For (ii), $\,\mathcal{H}\,$ arises since local $\,C^\infty$ trivializations
mentioned earlier may be chosen so as to share their domains with local
hol\-o\-mor\-phic trivializations; the former make the fibre geometry appear
constant, and the latter turn the bundle, locally, into the projectivization
(\ref{prz}) of a hol\-o\-mor\-phic vector bundle $\,E}%{{\mathrm{B}}\,$ endowed with a
Her\-mit\-i\-an fibre metric $\,(\hskip2.3pt,\hskip1.3pt)\,$ that induces the
Fu\-bi\-ni-Stu\-dy metrics of the original fibres. Since
$\,(\hskip2.3pt,\hskip1.3pt)$ is unique up to multiplications by positive
functions (Remark~\ref{fbstm}), Lemma~\ref{ddrsq}(iv) easily implies that its
choice does not affect the resulting parallel transports between the
projectivized fibres, thus giving rise to $\,\mathcal{H}$.
The Chern connection $\,\mathcal{H}\,$ now also arises in case (iii) since,
according to Remark~\ref{cnvrs}, (iii) is a sub\-case of (ii). The situation
is, however, more special: the critical manifolds -- analogs of (\ref{tmm}) --
in the fibres now constitute two hol\-o\-mor\-phic bundles $\,\varSigma^\pm$ of
Fu\-bi\-ni-Stu\-dy complex projective spaces over $\,B\,$ (with fibre
dimensions that need not be both positive; see Remark~\ref{spmcp}), contained
as sub\-bundles in the original bundle, and invariant under all
$\,\mathcal{H}$-par\-al\-lel transports. Also, the fibre-ge\-om\-e\-try
gradients and their $\,J$-im\-a\-ges (analogous to what we normally denote by
$\,v=\nabla\hskip-.5pt\vt\,$ and $\,u$) together form two hol\-o\-mor\-phic vertical vector
fields $\,v\,$ and $\,u=J\hskip-.7pt v$ on the total space. This is immediate from
the preceding paragraph, with the two sub\-bundles $\,\varSigma^\pm$ corresponding
to a $\,(\hskip2.3pt,\hskip1.3pt)$-or\-thog\-o\-nal hol\-o\-mor\-phic
decomposition $\,E}%{{\mathrm{B}}\hskip-.7pt=E}%{{\mathrm{B}}^+\hskip-1.5pt\oplusE}%{{\mathrm{B}}^-$ of the lo\-cal\-ly-de\-fined
vector bundle $\,E}%{{\mathrm{B}}\hskip-.7pt$, cf.\ (\ref{dta}.ii) and (\ref{spm}.c) --
(\ref{spm}.d), while the flow of $\,u$, described in the lines following
(\ref{dta}), acts in both $\,E}%{{\mathrm{B}}^\pm$ via multiplications by two (unrelated)
constant unit complex scalars. In case (ii), or (iii),
\begin{equation}\label{prt}
\begin{array}{l}
\mathrm{the\ }\hskip.7pt\mathcal{H}\hyp\mathrm{par\-al\-lel\ transports\ are\
hol\-o\-mor\-phic\ isometries\ or,}\\
\mathrm{respectively,\ CP}\hskip-.7pt\hyp\mathrm{triple\ isomorphisms\ between\ the\
fibres,}\\
\end{array}
\end{equation}
which holds for (ii) since it does for (i), cf.\ Section~\ref{ck} and,
consequently, also extends to the case of (iii) via the canonical
modifications in Remarks~\ref{alltq} and~\ref{cnvrs}.
The following assumptions and notations will now be used to construct compact
ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triples, each of which we call a
{\medit horizontal extension\/} of the \,CP triple provided by any fibre
$\,(\pi^{-\nnh1}(z),g^z\hskip-.7pt,\vt^z)$.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] $\pi\hskip-.7pt:\nnhM}%{{\mathrm{M}}\hskip-.7pt\to\hskip-.7ptB\,$ and $\,\mathcal{H}\,$ are the
bundle projection and the Chern connection of a hol\-o\-mor\-phic bundle of
\,CP triples with a compact base $\,B\,$ and the
\hbox{\,CP\hskip-.7pt-}\hskip0pttriple fibres $\,(\pi^{-\nnh1}(z),g^z\hskip-.7pt,\vt^z)$,
$\,z\inB$, while $\,\varSigma^\pm$ stand for the above sub\-bundles of
Fu\-bi\-ni-Stu\-dy complex projective spaces, invariant under
$\,\mathcal{H}$-par\-al\-lel transports.
\item[{\rm(b)}] We let $\,\vt\hskip-1.5pt_\pm\w,a\,$ be the data associated with
some\hskip.7pt$/$\hskip-.4pt any fibre $\,(\pi^{-\nnh1}(z),g^z\hskip-.7pt,\vt^z)$ as in
Remark~\ref{ascdt}(i), and $\,\vt:M}%{{\mathrm{M}}\to\mathrm{I\!R}}%{\mathbf{R}\,$ (or,
$\,\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt\to\varSigma^\pm$) be the
$\,C^\infty$ function (or, hol\-o\-mor\-phic bundle projection) which,
restricted to each $\,\pi^{-\nnh1}(z)$, equals $\,\vt^z$ or, respectively, the
version of (\ref{dbp}) corresponding to $\,(\pi^{-\nnh1}(z),g^z\hskip-.7pt,\vt^z)$. We
also set $\,M}%{{\mathrm{M}}'\hskip-.7pt=M}%{{\mathrm{M}}\smallsetminus(\varSigma^+\hskip-1.5pt\cup\varSigma^-)$.
\item[{\rm(c)}] One is given two K\"ah\-ler metrics $\,h\hskip-.7pt^\pm$ on the total
spaces $\,\varSigma^\pm$ of our hol\-o\-mor\-phic bundles of Fu\-bi\-ni-Stu\-dy
complex projective spaces such that either $\,h\hskip-.7pt^\pm$ makes the fibres
$\,\varSigma^\pm_{\hskip-.7pt z}\hskip-.7pt$, $\,z\inB$, or\-thog\-o\-nal to $\,\mathcal{H}\,$
along $\,\varSigma^\pm$ and, restricted to each fibre, $\,h\hskip-.7pt^\pm$ equals
$\,2(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)/\hskip-.4pt a\,$ times the Fu\-bi\-ni-Stu\-dy metric of
$\,\varSigma^\pm_{\hskip-.7pt z}\hskip-.7pt$.
\item[{\rm(d)}] We define a Riemannian metric $\,g\,$ on $\,M}%{{\mathrm{M}}'$ by
requiring that $\,\mathcal{H}\,$ be $\,g$-or\-thog\-o\-nal to the vertical
distribution $\,\mathrm{Ker}\hskip2.3ptd\pi$, that $\,g\,$ agree on the fibres
$\,\pi^{-\nnh1}(z)\,$ with the metrics $\,g^z\hskip-.7pt$, and that
$\,(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)\hskip.7pt g\hskip.7pt=(\vt-\hskip-.4pt\vt\hskip-1.7pt_-\w)\hsh\hskip-1.5pt^+\hskip-.7pt
+(\vt\hskip-1.8pt_+\w\hskip-2.3pt-\vt)\hsh\hskip-1.5pt^-$ on $\,\mathcal{H}$, the symbols $\,h\hskip-.7pt^\pm$ being
also used for the $\,\pi^\pm\hskip-1.2pt$-pull\-backs of $\,h\hskip-.7pt^\pm\hskip-.7pt$, cf.\ (b)
-- (c).
\item[{\rm(e)}] Our final assumption is that the Riemannian metric $\,g\,$ on
the dense open sub\-man\-i\-fold $\,M}%{{\mathrm{M}}'$ has an extension to a K\"ah\-ler
metric on $\,M}%{{\mathrm{M}}$ (still denoted by $\,g$).
\end{enumerate}
\begin{remark}\label{adggk}Under the hypotheses (a) -- (e), the resulting
horizontal extension $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is actually a
ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple. Namely, being a part of the
geometry of the fibres $\,(\pi^{-\nnh1}(z),g^z\hskip-.7pt,\vt^z)$, the functions
$\,\vt^z$ are preserved by $\,\mathcal{H}$-par\-al\-lel parallel transports,
that is, $\,\vt\,$ is constant along $\,\mathcal{H}$, and so its (vertical)
$\,g$-gra\-di\-ent must, by Remark~\ref{gpgrd}. coincide with the
hol\-o\-mor\-phic vertical vector field $\,v\,$ described in the lines
preceding (\ref{prt}). On the other hand, the function $\,Q=g(\hskip-.4pt v,v)$, equal
- consequently - to its fibre version, is a specific function of $\,\vt$.
Thus, by Lemma~\ref{ggqft}, $\,\vt\,$ has a hol\-o\-mor\-phic geodesic
$\,g$-gra\-di\-ent.
\end{remark}
\begin{remark}\label{hhext}Whenever a compact ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is a horizontal extension arising as in
Remark~\ref{adggk}, the distribution
$\,\mathcal{Z}=\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$ on
$\,M}%{{\mathrm{M}}'$ coming from the decomposition (\ref{tme}) for $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$
coincides, on $\,M}%{{\mathrm{M}}'\hskip-.7pt$, with the vertical distribution
$\,\mathrm{Ker}\hskip2.3ptd\pi\,$ of the bundle projection
$\,\pi\hskip-.7pt:\nnhM}%{{\mathrm{M}}\hskip-.7pt\to\hskip-.7ptB\,$ (see (a) above) and, consequently,
$\,\mathcal{Z}\,$ is integrable.
In fact, applying Remark~\ref{tglvs} to $\,\mathcal{H}$-hor\-i\-zon\-tal lifts
$\,w\,$ of local vector fields on $\,B$, we see that, by (\ref{prt}),
$\,\mathrm{Ker}\hskip2.3ptd\pi\,$ has totally geodesic leaves. Using
(\ref{img}) for both $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ and the fibres
$\,(\pi^{-\nnh1}(z),g^z\hskip-.7pt,\vt^z)$, we now conclude that the projections
$\,\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\mp\hskip-.7pt\to\varSigma^\pm$ defined in (b)
are the same as those in (\ref{dbp}). (Note that, due to the orthogonality
requirement in (c), the minimizing geodesic segment in $\,\pi^{-\nnh1}(z)$,
$\,z\inB$, joining a point $\,x\in\pi^{-\nnh1}(z)\,$ to $\,\varSigma^\pm_{\hskip-.7pt z}$
a normal to $\,\varSigma^\pm_{\hskip-.7pt z}\hskip-.7pt$, serves as the segment with the same
properties for $\,M}%{{\mathrm{M}}\,$ rather than $\,\pi^{-\nnh1}(z)$.) Now (\ref{hpm})
implies that the distribution
$\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$ is
contained in $\,\mathrm{Ker}\hskip2.3ptd\pi\,$ and, restricted to every fibre,
equals the analog of
$\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$ for the
fibre, that is, its tangent bundle (see Theorem~\ref{cptrp}(ii)). Thus,
$\,\mathcal{Z}\,$ coincides with the full vertical distribution
$\,\mathrm{Ker}\hskip2.3ptd\pi$.
\end{remark}
\begin{theorem}\label{cpbdl}{\medit
A ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, with
compact\/ $\,M}%{{\mathrm{M}}\hskip-.7pt$, satisfies one$/\hskip-.4pt$all of the
mu\-tu\-al\-ly-e\-quiv\-a\-lent conditions\/ {\rm(i)} -- {\rm(iii)} of
Lemma\/~{\rm\ref{intpr}}, if and only if it is iso\-mor\-phic to a horizontal
extension of a CP triple, defined as above using\/ {\rm(a)} -- {\rm(e)}.
}
\end{theorem}
\begin{proof}Remark~\ref{hhext} clearly yields the `if' part of our claim.
Conversely, let $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ satisfy (i) -- (iii) in
Lemma~\ref{intpr}. Lemma~\ref{intpr}(viii) states that
$\,\mathcal{Z}=\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$
coincides, on $\,M}%{{\mathrm{M}}'\hskip-.7pt$, with the vertical distribution
$\,\mathrm{Ker}\hskip2.3ptd\pi$ of the hol\-o\-mor\-phic bundle projection
$\,\pi:M}%{{\mathrm{M}}\toB$. Also, in view of Remark~\ref{trivl}, the leaves of
$\,\mathcal{Z}\,$ form ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triples, due to their
being complex sub\-man\-i\-folds of $\,M}%{{\mathrm{M}}\,$ tangent to $\,v=\nabla\hskip-.5pt\vt\,$ (since
$\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)$) and, as they are also totally
geodesic (see Lemma~\ref{ttgds}(b)), (\ref{hvk}) and the
$\,S$-in\-var\-i\-ance in (\ref{tme}), with $\,S=\nabla\hskip-.7pt v$, imply via
Theorem~\ref{cptrp} that they are all iso\-mor\-phic to \,CP\ triples. The
local isometries of Lemma~\ref{ttgds}(b) can obviously be made global due to
compactness (see the lines preceding (ii) above) which, consequently, turns
$\,M}%{{\mathrm{M}}\,$ into a hol\-o\-mor\-phic bundle of \,CP triples over $\,B$, in the
sense of (iii).
On the other hand, the $\,g$-or\-thog\-o\-nal complement of
$\,\mathcal{Z}=\mathrm{Ker}\hskip2.3ptd\pi\,$ is equal, on $\,M}%{{\mathrm{M}}'\hskip-.7pt$, to the summand
$\,\mathcal{H}\,$ in (\ref{tme}). Thus, $\,\mathcal{H}\,$ constitutes a
connection in the bundle $\,M}%{{\mathrm{M}}\,$ over $\,B$, as defined in the lines
following (i), and -- being the intersection of the horizontal distribution of
the Chern connections $\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\pm$ in the normal
bundles $\,N}%{{\mathrm{N}}=N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-.7pt$, cf.\ Theorem~\ref{first}(ii) --
$\,\mathcal{H}\,$ itself is, according to (a) in Section~\ref{ck}, the Chern
connection of the hol\-o\-mor\-phic bundle $\,M}%{{\mathrm{M}}\,$ of \,CP triples over
$\,B$.
This provides parts (a) -- (b) of the data (a) -- (e) required above, with
$\,\varSigma^\pm$ and $\,\vt\hskip-1.5pt_\pm\w,a\,$ given by (\ref{tmm}) and, respectively,
Remark~\ref{ascdt}(i). The sub\-man\-i\-fold metrics $\,h\hskip-.7pt^\pm$ of $\,\varSigma^\pm$
have, by (v) -- (vi) in Lemma~\ref{intpr} and the final clause of
Theorem~\ref{dcomp}(b), all the properties needed for (c).
To show that $\,g\,$ satisfies (d), consider two $\,\pi$-pro\-ject\-a\-ble
nonzero local sections $\,w,w\hskip.4pt'$ of the distribution
$\,\mathcal{H}=\mathcal{Z}^\perp\hskip-.7pt$, cf.\ (\ref{tme}). According to
Lemma~\ref{ttgds}(a) and the last line of Remark~\ref{pralg}, $\,w\,$ and
$\,w\hskip.4pt'$ are projectable along $\,\mathcal{V}\,$ and $\,\mathcal{H}^\pm\hskip-.7pt$,
as well as $\,\pi^\pm\hskip-1.2pt$-pro\-ject\-able, for either sign
$\,\pm\hskip.7pt$. Their restrictions to any fixed normal geodesic segment
$\,\varGamma\hskip.7pt$ emanating from $\,\varSigma^\pm$ thus lie in the space $\,\mathcal{W}\,$
(cf.\ (\ref{vtg}) and (i) -- (ii) in Theorem~\ref{jacob}) and, by
Theorem~\ref{dcomp}(g), $\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ restricted to $\,\varGamma\hskip.7pt$ is
a (possibly nonhomogeneous) linear function of $\,\vt$. The same linearity
condition obviously holds for $\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ when $\,g\,$ is defined
as in (d), rather than being the metric of our triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$. The
two definitions of $\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ must now agree, as the two linear
functions have -- in view of Remark~\ref{inequ}(iii) and the final clause of
Theorem~\ref{dcomp}(b) -- the same values
$\,h\hskip-.7pt^\pm\hskip-.7pt(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$ at either endpoint $\,\vt\hskip-1.5pt_\pm\w$ of the interval
$\,[\hskip.4pt\vt\hskip-1.7pt_-\w,\vt\hskip-1.8pt_+\w]$.
\end{proof}
\begin{remark}\label{clone}All compact SKRP triples of Class 1 (cf.\
Section~\ref{ev}) must be
\begin{equation}\label{cdo}
\mathrm{iso\-mor\-phic\ to\ horizontal\ extensions\ of\ CP\ triples\ of\
complex\ dimension\ }\,1\hskip.4pt,
\end{equation}
while those of Class 2 are themselves CP triples of a special type. The former
claim is easily verified using \cite[Theorem 16.3]{derdzinski-maschler-06};
for the latter, see Lemma~\ref{cltwo}.
The classification result of \cite[Theorem 6.1]{derdzinski-kp} may be
rephrased as the conclusion (\ref{cdo}) about all compact
ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triples $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ with
$\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}=2\,$ other than Class 2 SKRP triples are. Similarly,
(\ref{cdo}) is the case -- by their very construction -- for the gradient
K\"ah\-\hbox{ler\hskip.7pt-}\hskip0ptRic\-ci sol\-i\-tons of Koiso \cite{koiso} and
Cao \cite{cao}, mentioned in the Introduction.
\end{remark}
\section{Con\-stant-rank multiplications}\label{cr}
\setcounter{equation}{0}
In this section all vector spaces are
fi\-\hbox{nite\hskip.4pt-}\hskip0ptdi\-men\-sion\-al and complex. Bi\-lin\-e\-ar
mappings of the type discussed here arise in any compact
ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple (see Theorem~\ref{ctrkm}), which
leads to the dichotomy conclusion of Theorem~\ref{dicho}.
A {\medit con\-stant-rank multiplication\/} is any bi\-lin\-e\-ar mapping
$\,\mu:\mathcal{N}\hskip-1.5pt\times\hskip-.7pt\mathcal{T}\to\mathcal{Y}\hskip-.7pt$, where $\,\mathcal{N},\mathcal{T}\hskip-1.5pt,\mathcal{Y}\,$ are vector
spaces, such that the function $\,\mathcal{N}\smallsetminus\{0\}\ni\xi
\mapsto\mathrm{rank}\hskip1.7pt\mu(\xi,\,\cdot\,)\,$ is constant or,
equivalently, $\,\dim\hskip.7pt\mathrm{Ker}\hskip1.7pt\mu(\xi,\,\cdot\,)\,$ is the
same for all nonzero $\,\xi\in\mathcal{N}\hskip-.7pt$. When
$\,\dim\hskip.7pt\mathrm{Ker}\hskip1.7pt\mu(\xi,\,\cdot\,)=k\,$ for all
$\,\xi\in\mathcal{N}\smallsetminus\{0\}$, we also say that
$\,\mu:\mathcal{N}\hskip-1.5pt\times\hskip-.7pt\mathcal{T}\to\mathcal{Y}\,$ {\medit has the constant rank\/}
$\,\dim\mathcal{T}\hskip-2.3pt-k$. With the notations of Section~\ref{eg}, such
$\,\mu\,$ leads to a mapping
\begin{equation}\label{phi}
\varepsilon:\mathrm{P}\hskip-.7pt\mathcal{N}\to\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt\mathcal{T}\hskip10pt
\mathrm{given\ by}\hskip8pt\varepsilon(\hskip-.4pt\bbC\xi)\hskip.7pt
=\hskip.7pt\mathrm{Ker}\hskip1.7pt\mu(\xi,\,\cdot\,)\hskip6pt
\mathrm{for}\hskip5pt\xi\in\mathcal{N}\smallsetminus\{0\}\hskip.4pt.
\end{equation}
\begin{lemma}\label{dfphi}{\medit
For\/ $\,\mu\,$ and\/ $\,\varepsilon\,$ as above,
$\,\mathcal{N}\smallsetminus\{0\}\ni\xi
\mapsto\hskip.7pt\mathrm{Ker}\hskip1.7pt\mu(\xi,\,\cdot\,)
\in\hskip.7pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt\mathcal{T}\,$ and\/ $\,\varepsilon$ are both hol\-o\-mor\-phic.
In terms of the identification\/ {\rm(\ref{twg})}, the differential of the
former mapping at any\/ $\,\xi\in\mathcal{N}\smallsetminus\{0\}\,$ sends\/
$\,\eta\in\mathcal{N}\,$ to the unique\/
$\,H\in\mathrm{Hom}\hskip.7pt(\mathsf{W},\mathcal{T}\hskip-1.8pt/\hskip.4pt\mathsf{W})\,$ with\/
$\,\mu(\eta,w)=\mu(\xi,\hskip-1.5pt-\hskip-1.5pt\tilde H\hskip-.7pt w)\,$ for all\/
$\,w\in\mathsf{W}=\varepsilon(\hskip-.4pt\bbC\xi)$, where\/ $\,\tilde H:\mathsf{W}\to\mathcal{T}\,$ is any linear
lift of\/ $\hskip.7pt H$.
}
\end{lemma}
\begin{proof}This is obvious if one sets $\,F(\xi)=\mu(\xi,\,\cdot\,)\,$ in
Remark~\ref{dfrtl}.
\end{proof}
\begin{example}\label{cstrk}Any given con\-stant-rank multiplication
$\,\mu:\mathcal{N}\hskip-1.5pt\times\hskip-.7pt\mathcal{T}\to\mathcal{Y}\,$ leads to further such multiplications,
$\,\mu':\mathcal{N}\hskip-1.5pt\times\hskip-.7pt\mathcal{T}\hskip.4pt'\hskip-.7pt\to\mathcal{Y}'$ and
$\,\mu^*:\mathcal{N}\hskip-1.5pt\times\hskip-.7pt\mathcal{Y}^*\hskip-.7pt\to\mathcal{T}^*\hskip-.7pt$, obtained by setting
$\,\mu'(\xi,\,\cdot\,)=\gamma[\mu(\xi,\alpha\,\cdot\,)]\,$ and
$\,\mu^*(\xi,\,\cdot\,)=[\mu(\xi,\,\cdot\,)]^*\hskip-.7pt$. Here $\,\mathcal{T}\hskip.4pt'\hskip-.7pt,\mathcal{Y}'$
are vector spaces, $\,\alpha:\mathcal{T}\hskip.4pt'\hskip-.7pt\to\mathcal{T}\,$ (or, $\,\gamma:\mathcal{Y}\to\mathcal{Y}'$) is
surjective (or, injective) and linear, while $\,[\hskip3.2pt]^*$ stands for
the dual of a vector space or a linear operator.
\end{example}
\begin{lemma}\label{holem}{\medit
If\/ $\,\mu:\mathcal{N}\hskip-1.5pt\times\hskip-.7pt\mathcal{T}\to\mathcal{Y}\,$ has the constant
rank\/ $\,\dim\mathcal{T}\hskip-1.5pt-k\,$ and\/ $\,\varepsilon\,$ with\/ {\rm(\ref{phi})}
is nonconstant, then\/ $\,\varepsilon\,$ is a hol\-o\-mor\-phic embedding.
Whether\/ $\,\varepsilon\,$ is constant, or not, the same is the case for all
multiplications\/ $\,\mathcal{N}\hskip-1.5pt\times\hskip-.7pt\mathcal{T}\to\mathcal{Y}$ of the constant rank\/
$\,\dim\mathcal{T}\hskip-1.5pt-k$, sufficiently close to\/ $\,\mu$.
}
\end{lemma}
\begin{proof}Let $\,\mathsf{W}\in\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt\mathcal{T}$. The subset of $\,\mathcal{N}\,$
consisting of $\,0\,$ and all \hbox{$\xi\in\mathcal{N}\smallsetminus\{0\}\,$} with
$\,\varepsilon(\hskip-.4pt\bbC\xi)=\mathsf{W}\,$ is a vector sub\-space. In fact, if
$\,\xi,\eta\in\mathcal{N}\smallsetminus\{0\}\,$ and
$\,\mathsf{W}=\mathrm{Ker}\hskip1.7pt\mu(\xi,\,\cdot\,)
=\mathrm{Ker}\hskip1.7pt\mu(\eta,\,\cdot\,)$, then
$\,\mathsf{W}\subseteq\mathrm{Ker}\hskip1.7pt\mu(\zeta,\,\cdot\,)\,$ for any
$\,\zeta\in\mathrm{Span}\hskip.4pt(\xi,\eta)\,$ and, unless $\,\zeta=0$, this
inclusion is actually an equality due to the con\-stant-rank property of
$\,\mu$.
Therefore, $\,\varepsilon\hskip.7pt$-pre\-im\-ages of points of
$\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt\mathcal{T}\,$ are linear sub\-va\-ri\-e\-ties in
$\,\mathrm{P}\hskip-.7pt\mathcal{N}$. If $\,\varepsilon$ is nonconstant, all these
sub\-va\-ri\-e\-ties are ze\-\hbox{ro\hskip.4pt-}\hskip0ptdi\-men\-sion\-al, that is,
$\,\varepsilon\,$ has to be injective. Namely, by Lemma~\ref{nontr}, for the
K\"ah\-ler form $\,\omega\,$ of any K\"ah\-ler metric on
$\,\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt\mathcal{T}\hskip-1.5pt$, the integral of $\,\varepsilon\hskip-2pt\mathop{\phantom{\vrule width1pt height2.7pt depth0pt}}\limits^*\hskip-2.7pt\omega\,$
over any projective line $\,\mathcal{L}\,$ in $\,\mathrm{P}\hskip-.7pt\mathcal{N}\,$ is nonzero, and
so $\,\mathcal{L}$ cannot lie in the $\,\varepsilon\hskip.7pt$-pre\-im\-age of a point. Also,
Lemma~\ref{dfphi} guarantees hol\-o\-mor\-phic\-i\-ty of $\,\varepsilon$.
Let $\,\varepsilon\hskip.7pt$ now be nonconstant. Then $\,\varepsilon\,$ must be an embedding, that
is, $\,d\varepsilon_{\bbC\xi}^{\phantom i}$ is injective at any
$\,\bbC\xi\in\mathrm{P}\hskip-.7pt\mathcal{N}\,$ or, equivalently, the differential of
$\,\xi\mapsto\hskip.7pt\mathrm{Ker}\hskip1.7pt\mu(\xi,\,\cdot\,)\,$
at any $\,\xi\in\mathcal{N}\smallsetminus\{0\}\,$ has the kernel $\,\bbC\xi$.
Namely, in Lemma~\ref{dfphi} we may set $\,\tilde H=0$ when $\,H=0$, and so
$\,\eta\,$ lies in the kernel if and only if the inclusion
$\,\mathsf{W}\subseteq\mathrm{Ker}\hskip1.7pt\mu(\eta,\,\cdot\,)$ holds for
$\,\mathsf{W}\hskip-.7pt=\varepsilon(\hskip-.4pt\bbC\xi)$. Unless $\,\eta=0$, this inclusion is, as
before, an equality, and injectivity of $\,\varepsilon\,$ then yields
$\,\eta\in\bbC\xi$, which completes the proof, the final clause being
an immediate consequence of that in Lemma~\ref{nontr}.
\end{proof}
Given a compact ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$,
we use the notation of (\ref{sym}) and (\ref{tmm}) to set, for
$\,\xi,\eta\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm$ and $\,w\inT\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt$, with either
fixed sign $\,\pm\hskip.7pt$,
\begin{equation}\label{zxe}
Z_y^\pm\hskip-.7pt(\xi,\eta)\hskip.4pt w\,=\,a\hskip.4pt g_y^{\phantom i}(\xi,\eta)\hskip.4pt w\,
+\,(\vt\hskip-1.8pt_+\w\hskip-2.2pt-\hskip-.7pt\vt\hskip-1.7pt_-\w)\hskip.7pt R_y^{\phantom i}(\xi,J\hskip-2pt_y\eta)\hskip.4pt J\hskip-2pt_y^{\phantom i} w\hskip.4pt.
\end{equation}
Thus, $\,Z_y^\pm\hskip-.7pt(\xi,\eta)\hskip.4pt w\inT\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt$, as $\,\xi,\eta\,$ are
tangent, and $\,w\,$ normal, to the totally geodesic leaf through $\,y\,$ of
the $\,J$-in\-var\-i\-ant in\-te\-gra\-ble distribution
$\,\mathrm{Ker}\hskip2.3ptd\pi^\pm\hskip-.7pt
=\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^\mp\hskip-.7pt$, cf.\ (\ref{hpm}),
Theorem~\ref{dcomp}(c), Corollary~\ref{totgd}, and the first line of
Remark~\ref{ptnrs}. Also, denoting by $\,Z_y^\pm\hskip-.7pt(\xi,\eta)\,$ the
en\-do\-mor\-phism $\,w\mapsto Z_y^\pm\hskip-.7pt(\xi,\xi)\hskip.4pt w\,$ of $\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt$,
one has
\begin{equation}\label{zex}
Z_y^\pm\hskip-.7pt(\xi,\eta)=Z_y^\pm\hskip-.7pt(\eta,\xi)
=Z_y^\pm\hskip-.7pt(J\hskip-2pt_y^{\phantom i}\xi,J\hskip-2pt_y^{\phantom i}\eta)\hskip.4pt w\hskip.4pt,
\hskip16pt
J\hskip-2pt_y^{\phantom i}[Z_y^\pm\hskip-.7pt(\xi,\eta)]=[Z_y^\pm\hskip-.7pt(\xi,\eta)]J\hskip-2pt_y^{\phantom i}\hskip.4pt,
\end{equation}
as an obvious consequence of (\ref{rcm}) and (\ref{gjw}). Next, we define a
com\-plex-bi\-lin\-e\-ar mapping
$\,\mu_y^\pm\hskip-1.5pt:N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-1.5pt\timesT\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt\to
\overline{\mathrm{Hom}}_\bbC^{\phantom i}(N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)\,$
by
\begin{equation}\label{mxw}
\mu_y^\pm(\xi,w)\,=\,Z_y^\pm\hskip-.7pt(J\hskip-2pt_y^{\phantom i}\xi,\,\cdot\,)\hskip.4pt w\,
+\,Z_y^\pm\hskip-.7pt(\xi,\,\cdot\,)J\hskip-2pt_y^{\phantom i} w\hskip.4pt.
\end{equation}
By $\,\overline{\mathrm{Hom}}_\bbC^{\phantom i}$ we mean here `the space of
anti\-lin\-e\-ar operators' and
$\,\overline{\mathrm{Hom}}_\bbC^{\phantom i}(N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)$
is treated as a complex vector space in which the multiplication by
$\,i\,$
acts via composition with $\,J\hskip-2pt_y^{\phantom i}$ from the left. (The product thus equals
the given operator $\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt\toT\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ {\medit
followed\/} by $\,J\hskip-2pt_y^{\phantom i}$.) Anti\-lin\-e\-ar\-i\-ty of $\,\mu_y^\pm(\xi,w)\,$
and com\-plex-bi\-lin\-e\-ar\-i\-ty of $\,\mu_y^\pm$ are both obvious from
(\ref{zex}).
\begin{theorem}\label{ctrkm}{\medit
For a compact ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, a fixed sign\/ $\,\pm\hskip.7pt$, and any point\/
$\,y\in\varSigma^\pm\hskip-.7pt$, the mapping\/ $\,\mu_y^\pm$ with\/ {\rm(\ref{mxw})} is a
con\-stant-rank multiplication, cf.\ Section\/~{\rm\ref{cr}}. Furthermore,
if\/ $\,\varepsilon=\varepsilon_y^\pm$ corresponds to\/
$\,\mu=\mu_y^\pm$ as in\/ {\rm(\ref{phi})} and\/ $\,\xi\,$ is any nonzero
vector normal to\/ $\,\varSigma^\pm$ at\/ $\,y$, then
\begin{equation}\label{dph}
\begin{array}{rl}
\mathrm{i)}&\varepsilon_y^\pm(\hskip-.4pt\bbC\xi)\,=\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)\,
=\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{V}_{\hskip-.7pt x}^{\phantom i}\oplus\mathcal{H}_x^\pm\hskip-.7pt)\hskip.4pt,
\hskip12pt\mathrm{where}\hskip7ptx=\varPhi(y,\xi)\hskip.4pt,\phantom{_{j_j}}\\
\mathrm{ii)}&\varepsilon_y^\pm(\hskip-.4pt\bbC\xi)\,
=\,\mathrm{Ker}\hskip1.7ptZ_y^\pm\hskip-.7pt(\xi,\xi)\hskip.4pt,\hskip40pt\mathrm{for}
\hskip7ptZ_y^\pm\hskip-.7pt(\xi,\xi)\hskip7pt\mathrm{as\ \ in\ \ (\ref{zex}).}
\phantom{^{1^1}}
\end{array}
\end{equation}
}
\end{theorem}
\begin{proof}Whenever $\,x=\varPhi(y,\xi)\,$ and
$\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt\smallsetminus\hskip-.7pt\{0\}$, we have
\begin{equation}\label{krz}
\mathrm{Ker}\hskip1.7ptZ_y^\pm\hskip-.7pt(\xi,\xi)\,=\,
d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)\,
=\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{V}_{\hskip-.7pt x}^{\phantom i}\oplus\mathcal{H}_x^\pm\hskip-.7pt)\hskip.4pt.
\end{equation}
In fact, let $\,x=x(t)\in\varGamma\hskip.7pt$ as in Theorem~\ref{dcomp}, with some fixed
$\,t\in(t_-^{\phantom i},t_+^{\phantom i})$. According to (\ref{hvk}) and parts (iii), (iv), (vi) of
Theorem~\ref{jacob}, the vectors forming $\,\mathcal{H}_x^\pm$ are precisely
the values $\,w(t)\,$ for all $\,w\,$ as in Theorem~\ref{dcomp}(e) which also
have the property that
$\,2\hskip.4pt(\vt-\hskip-.4pt\vt\hskip-1pt_\mp\w)\hskip.4pt Q^{-\nnh1}g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)=g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$
whenever $\,w\hskip.4pt'$ satisfies the hypotheses of Theorem~\ref{dcomp}(e). Since
the values $\,w_\pm'$ in Theorem~\ref{dcomp}(h2) fill $\,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ (cf.\
assertions (d) -- (f) of Theorem~\ref{dcomp}), replacing $\,g(\hskip-.4pt w,w\hskip.4pt'\hskip.4pt)\,$
and $\,g(S\hskip-.7pt w,w\hskip.4pt'\hskip.4pt)\,$ in the last equality with the expressions
provided by Theorem~\ref{dcomp}(h2) and Remark~\ref{bthex}, we easily verify,
using (\ref{rcm}) and Remark~\ref{inequ}(i), that $\,w(t)\in\mathcal{H}_x^\pm$
if and only if $\,Z_y^\pm\hskip-.7pt(\xi,\eta)\hskip.4pt w_\pm^{\phantom i}\hskip-.7pt=0$. Now the final clause
of Theorem~\ref{dcomp}(b) (or, Remark~\ref{holom}) yields the first
\hbox{(or, second) equality in (\ref{krz}).}
To simplify notations, let us write $\,g,Z,J\,$ rather than
$\,g_y^{\phantom i},Z_y^\pm\hskip-.7pt,J\hskip-2pt_y^{\phantom i}$. Since $\,x=\varPhi(y,\xi)\,$ in (\ref{krz}) and
$\,\varPhi\,$ is hol\-o\-mor\-phic (Theorem~\ref{first}), (\ref{krz}) and
Remark~\ref{holom} clearly imply that, for a suitable integer $\,k=k_\pm^{\phantom i}$,
the resulting mapping
\begin{equation}\label{mpg}
N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt\smallsetminus\hskip-.7pt\{0\}\ni\xi\,\,
\mapsto\,\,\mathrm{Ker}\hskip1.7ptZ(\xi,\xi)
\in\hskip.7pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)\hskip12pt\mathrm{is\
hol\-o\-mor\-phic.}
\end{equation}
The $\,C^\infty$ version of the assumptions
listed in Remark~\ref{dfrtl} is thus satisfied if one chooses
$\,\,U\hskip-.7pt,\mathcal{T}\hskip-1.5pt,\mathcal{Y}\,$ to be
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt\smallsetminus\hskip-.7pt\{0\},T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt,T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$
and sets $\,F(\xi)=Z(\xi,\xi)$. By (\ref{fxh}), the differential
of (\ref{mpg}) at any nonzero $\,\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm$ sends
any $\,\eta\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm$ to the
unique $\,H:\mathsf{W}\to\mathcal{T}\hskip-1.8pt/\hskip.4pt\mathsf{W}\hskip-.7pt$, where
$\,\mathsf{W}=\hskip.7pt\mathrm{Ker}\hskip1.7ptZ(\xi,\xi)$, with a linear lift
$\,\tilde H:\mathsf{W}\to\mathcal{T}\hskip-.7pt=T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm$ such that
$\,Z(\xi,\xi)\circ\tilde H\,$ equals the restriction of
$\,-\nnh2Z(\xi,\eta)\,$ to $\,\mathsf{W}\hskip-.7pt$. (We have
$\,d\hskip-.8ptF\hskip-3pt_\xi^{\phantom i}=2Z(\xi,\,\cdot\,)$ since $\,Z(\xi,\eta)\,$ is
real-bi\-lin\-e\-ar and symmetric in $\,\xi,\eta$, cf.\ (\ref{zex}).)
Consequently,
\begin{equation}\label{cpl}
2Z(\xi,\eta)\hskip.4pt w\,=\,-Z(\xi,\xi)\hskip.4pt\tilde H\hskip-.7pt w\hskip12pt\mathrm{for\ all\
}\hskip5ptw\in\mathrm{Ker}\hskip1.7ptZ(\xi,\xi)\hskip.4pt.
\end{equation}
Com\-plex-lin\-e\-ar\-i\-ty of the differential, due to (\ref{mpg}), means
that (\ref{cpl}) will still hold if we replace $\,\eta\,$ with $\,J\hskip-.4pt\eta\,$
and $\,\tilde H\,$ with $\,J\hskip-.7pt\tilde H$. Then, from (\ref{zex}) and
(\ref{cpl}), $\,2Z(J\xi,\eta)\hskip.4pt w=-\nnh2Z(\xi,J\hskip-.4pt\eta)\hskip.4pt w
=Z(\xi,\xi)\hskip.4pt J\hskip-.7pt\tilde H\hskip-1.5pt w=J[Z(\xi,\xi)\hskip.4pt\tilde H\hskip-.7pt w]
=-\nnh2J[Z(\xi,\eta)\hskip.4pt w]=-\nnh2Z(\xi,\eta)\hskip.4pt J\hskip-.7pt w$. In other words,
$\,Z(J\xi,\eta)\hskip.4pt w+Z(\xi,\eta)\hskip.4pt J\hskip-.7pt w=0\,$ whenever
$\,w\in\mathrm{Ker}\hskip1.7ptZ(\xi,\xi)$ and
$\,\eta\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt$. Thus, by (\ref{mxw}),
$\,\mathrm{Ker}\hskip1.7ptZ(\xi,\xi)\subseteq\varepsilon_y^\pm(\hskip-.4pt\bbC\xi)
=\mathrm{Ker}\hskip1.7pt\mu_y^\pm(\xi,\,\cdot\,)$, while the opposite
inclusion is obvious since (\ref{zex}) gives $\,Z(\xi,J\xi)=0$, and so the
expression $\,Z(J\xi,\eta)\hskip.4pt w+Z(\xi,\eta)\hskip.4pt J\hskip-.7pt w=0\,$ for
$\,\eta=J\xi\,$ equals $\,Z(\xi,\xi)\hskip.4pt w$.
The equality $\,\mathrm{Ker}\hskip1.7ptZ(\xi,\xi)=\varepsilon_y^\pm(\hskip-.4pt\bbC\xi)\,$
and (\ref{krz}) -- (\ref{mpg}) complete the proof.
\end{proof}
The description of $\,\dot x\hskip-.7pt_\pm^{\phantom i}$ in the lines preceding (\ref{mpg}) also
gives
\begin{equation}\label{nng}
g_y^{\phantom i}(Z_y^\pm\hskip-.7pt(\xi,\xi)w,w)\,\ge\,0\hskip12pt\mathrm{for\ all}\hskip7pt
\xi\inN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip7pt\mathrm{and}\hskip7pt
w\inT\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt,
\end{equation}
which one sees taking the limit of the equality in Theorem~\ref{dcomp}(h2)
with $\,w\hskip.4pt'\hskip-.7pt=w\,$ as $\,t\in(t_-^{\phantom i},t_+^{\phantom i})\,$ approaches the other endpoint
$\,t_\mp^{\phantom i}$ (and so $\,\vt\to\vt\hskip-1pt_\mp\w$).
\section{The dichotomy theorem}\label{dt}
\setcounter{equation}{0}
This section uses the notations listed at the beginning of Section~\ref{cc}
and the symbols $\,k_\pm^{\phantom i}$ of Remark~\ref{dppdm}. Any $\,y\in\varSigma^\pm$ leads
to the assignment
\begin{equation}\label{xtd}
N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-1.5pt\smallsetminus\hskip-1.5pt\{0\}\ni\xi\hskip-.4pt
\mapsto\hskip-.7pt d\pi\hskip-.7pt_x^{\hskip.7pt\pm}\hskip-.7pt(\mathcal{H}_x^\pm\hskip-.7pt)
\in\hskip-.4pt\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)\hskip.4pt,\hskip3.5pt\mathrm{where}
\hskip5ptx\hskip-.7pt=\hskip-.4pt\varPhi(y,\xi)\hskip5pt\mathrm{and}\hskip5ptk\hskip-.4pt=\hskip-.7pt k_\pm^{\phantom i}\hskip.4pt,
\end{equation}
$\varPhi=\hskip.7pt\varPhi^\pm$ being defined by (\ref{phe}). (Due to (\ref{hpm}) and
(\ref{tme}), $\,d\pi\hskip-.7pt_x^{\hskip.7pt\pm}$ is injective on $\,\mathcal{H}_x^\pm$.)
\begin{theorem}\label{dicho}{\medit
Given any compact ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler triple\/
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, one and only one of the following two cases occurs.
\begin{enumerate}
\def{\rm\alph{enumi}}{{\rm\alph{enumi}}}
\item[{\rm(a)}] Either the mappings\/ {\rm(\ref{xtd})} are all constant, for
both signs\/ $\,\pm\hskip.7pt$, or
\item[{\rm(b)}] each of\/ {\rm(\ref{xtd})}, for both signs\/ $\,\pm\hskip.7pt$,
descends to a nonconstant hol\-o\-mor\-phic embedding\/
$\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\hskip-.7pt\to\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-.7pt(T\hskip-2.3pt_y\w\hskip-.7pt\varSigma^\pm\hskip-.7pt)$,
where\/ $\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ is the projective space of\/
$\,N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}=N}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt$.
\end{enumerate}
Condition\/ {\rm(a)} holds if and only if\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ satisfies\/
{\rm(i)} -- {\rm(iii)} in Lemma\/~{\rm\ref{intpr}}.
}
\end{theorem}
\begin{proof}In view of Theorem~\ref{ctrkm}, we may use Lemma~\ref{holem} for
$\,\varepsilon=\varepsilon_y^\pm$ corresponding to $\,\mu=\mu_y^\pm$ as in (\ref{phi}),
concluding (from an obvious continuity argument) that, with either fixed sign
$\,\pm\hskip.7pt$, all the mappings (\ref{xtd}) descend to hol\-o\-mor\-phic
embeddings of $\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}$ unless they are
all constant. Their constancy for one sign implies, however, the same for the
other, since it amounts to (ii) or (iii) in Lemma~\ref{intpr}, while (ii) and
(iii) are equivalent. This completes the proof.
\end{proof}
\begin{remark}\label{infty}Case (a) of Theorem~\ref{dicho} is equivalent to
(\ref{spn}), as one sees combining Lemma~\ref{intpr}(i) with (\ref{hpm}).
According to (iv) -- (vi) in Lemma~\ref{intpr}), the immersions of
Theorem~\ref{tgimm}(c) are then embeddings and their images form the leaves of
foliations on $\,\varSigma^\mp\hskip-.7pt$, both of which have the same leaf space $\,B$.
\end{remark}
\begin{remark}\label{nncst}When (b) holds in Theorem~\ref{dicho}, images of
the totally geodesic hol\-o\-mor\-phic immersions of Theorem~\ref{tgimm}(c)
pass through every point $\,y\in\varSigma^\pm\hskip-.7pt$, realizing an uncountable family
of tangent spaces: the image of the embedding (\ref{xtd}).
\end{remark}
\section{More on Grass\-mann\-i\-an triples}\label{mg}
\setcounter{equation}{0}
We continue using the asumptions and notation of Section~\ref{he}
\begin{lemma}\label{lfspc}{\medit
The leaf space\/ $\,M}%{{\mathrm{M}}'\hskip-2.8pt/\mathcal{V}\,$ of the in\-te\-gra\-ble
distribution\/ $\,\mathcal{V}=\mathrm{Span}\hskip.7pt(\hskip-.4pt v,u)\,$ on\/
$\,M}%{{\mathrm{M}}'\hskip-.7pt=M}%{{\mathrm{M}}\smallsetminus(\varSigma^+\hskip-1.5pt\cup\varSigma^-)$, cf.\
Lemma\/~{\rm\ref{dvgww}(a)}, carries a natural structure of a compact complex
manifold of complex dimension\/ $\,m-1$, with\/ $\,m=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nhM}%{{\mathrm{M}}\hskip-.7pt$, such
that the quo\-tient-space projection\/
$\,M}%{{\mathrm{M}}'\hskip-.7pt\toM}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\,$ forms a hol\-o\-mor\-phic fibration
and, for either sign\/ $\,\pm\hskip.7pt$, the projectivization\/
$\,\mathrm{P}\nnhN}%{{\mathrm{N}}\hskip.7pt$ of the normal bundle\/
$\,N}%{{\mathrm{N}}=N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-.7pt$, defined as in\/ {\rm(\ref{prz})}, is
bi\-hol\-o\-mor\-phic to\/ $\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\,$ via the
bi\-hol\-o\-mor\-phisms sending each complex line\/ $\,\mathcal{L}\,$ through\/
$\,0\,$ in the normal space of\/ $\,\varSigma^\pm$ at any point to the\/
$\,\mathrm{Exp}^\perp\hskip-1.5pt$-im\-age of the punctured radius\/ $\,\delta\,$
disk in\/ $\,\mathcal{L}\hskip.4pt$, the latter image being a leaf of\/ $\,\mathcal{V}\,$
according to Lemma\/~{\rm\ref{vsbkr}(a)}.
The mappings\/ {\rm(\ref{dbp})}, restricted to\/ $\,M}%{{\mathrm{M}}'\hskip-.7pt$, descend to
hol\-o\-mor\-phic bundle projections
\begin{equation}\label{hbp}
\pi^\pm\hskip-1.5pt:M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\to\varSigma^\pm,
\end{equation}
also denoted by\/ $\,\pi^\pm\hskip-.7pt$, which, under the bi\-hol\-o\-mor\-phic
identifications\/
$\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}=\mathrm{P}(\nnhN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt)$ of
the preceding paragraph, coincide with the bundle projections\/
$\,\mathrm{P}(\nnhN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt)\to\varSigma^\pm\hskip-.7pt$.
}
\end{lemma}
\begin{proof}The restrictions
$\,\varPhi^\pm\hskip-1.5pt=\varPhi:N}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt\smallsetminus\hskip-.7pt\varSigma^\pm
\hskip-.7pt\toM}%{{\mathrm{M}}'$ given by (\ref{phe}) with the two possible signs $\,\pm\,$ are
bi\-hol\-o\-mor\-phisms (Theorem~\ref{first}), and hence so is the composite
of one of them followed by the inverse of the other. At the same time, by
Theorem~\ref{first}(iii), either of them descends to a bijection
$\,\mathrm{P}(\nnhN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt)
\toM}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\hskip-.7pt$, and the composite just mentioned yields a
bi\-hol\-o\-mor\-phism between
$\,\mathrm{P}(\nnhN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\pm\hskip-1.5pt)\,$ and
$\,\mathrm{P}(\nnhN}%{{\mathrm{N}}\hskip-2.3pt\varSigma^\mp\hskip-1.5pt)$. This turns
$\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\,$ into a compact complex manifold in a manner
independent of the bijection used. Our assertion is now immediate from
(\ref{pcf}).
\end{proof}
\begin{remark}\label{trint}The direct sum of the two vertical distributions
$\,\mathrm{Ker}\hskip2.3ptd\pi^\pm$ of the projections (\ref{hbp}) is a
distribution on $\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\hskip-.7pt$, since, at every point
$\,\varLambda\inM}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\hskip-.7pt$, they intersect trivially:
$\,\mathrm{Ker}\hskip2.3ptd\pi_{\hskip-1.5pt\varLambda}^+\hskip.7pt
\cap\,\mathrm{Ker}\hskip2.3ptd\pi_{\hskip-1.5pt\varLambda}^-=\hskip.7pt\{0\}$. In fact, as a
consequence of (\ref{tme}), the original vertical distributions on
$\,M}%{{\mathrm{M}}'\hskip-.7pt$, given by (\ref{hpm}), intersect along $\,\mathcal{V}\hskip-.7pt$.
\end{remark}
For a Grass\-mann\-i\-an triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ obtained as in
Section~\ref{eg} from some data (\ref{dta}.i), the descriptions of $\,\varSigma^\pm$
provided by (\ref{spm}.a), and
\begin{equation}\label{mpv}
\begin{array}{l}
M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\,\,\,
=\,\,\hskip.4pt\{(\mathsf{W},\mathsf{W}')\in\mathrm{Gr}\hskip-.7pt_k^{\phantom i}\hskip-1.5pt\mathsf{V}
\times\mathrm{Gr}\hskip-.7pt_{k-\nnh1}^{\phantom i}\hskip-2.3pt\mathsf{V}:\mathsf{W}'\hskip-.7pt\subseteq\mathsf{W}\hskip.7pt\}\hskip.4pt,
\mathrm{\ under\ which}\\
\pi^\pm\mathrm{\ in\ (\ref{hbp})\ correspond\ to\
}\,(\mathsf{W},\mathsf{W}')\mapsto\mathsf{W}\,\mathrm{\ and\ }\,(\mathsf{W},\mathsf{W}')\mapsto\mathsf{W}'\hskip-.7pt.
\end{array}
\end{equation}
the equality meaning a natural bi\-hol\-o\-mor\-phic identification. If
$\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ is in turn a CP triple, arising from (\ref{dta}.ii),
$\,\varSigma^\pm$ must be as in (\ref{spm}.b), and (\ref{mpv}) is replaced by
$\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}=\hskip.7pt\varSigma^+\hskip-1.5pt\times\varSigma^-\hskip-.7pt$, while
$\,\pi^\pm$ in (\ref{hbp}) then become the factor projections.
All these claims are immediate consequences of Remark~\ref{ppmgc}(d).
\begin{lemma}\label{ttlds}{\medit
For a fi\-\hbox{nite\hskip.4pt-}\hskip0ptdi\-men\-sion\-al complex vector space\/
$\,\mathsf{V}\hskip-.7pt$, any\/ $\,k\in\{1,\dots,\dim\mathsf{V}\}$, and\/
$\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\,$ given by\/ {\rm(\ref{mpv})}, let\/
$\,(\mathsf{W}_0^{\phantom i},\mathsf{W}_0'),(\mathsf{W},\mathsf{W}')\inM}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\hskip-.7pt$. Then there
exist an integer $\,p}%{q^+\ge1\,$ and\/
$\,(\mathsf{W}_j^{\phantom i},\mathsf{W}_j')\inM}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\hskip-.7pt$, $\,j=0,1,\dots,p}%{q^+$,
with\/ $\,(\mathsf{W}_{\!p}%{q^+}^{\phantom i},\mathsf{W}_{\!p}%{q^+}')=(\mathsf{W},\mathsf{W}')\,$ and\phantom{\hskip.7pt}
\hbox{$(\mathsf{W}_{\!j-\nh1}^{\phantom i},\mathsf{W}_{\!j-1}')\sim(\mathsf{W}_j^{\phantom i},\mathsf{W}_j')\,$} whenever\/
$\,j=1,\dots,p}%{q^+$, the notation\/ $\,(\tilde\mathsf{W},\tilde\mathsf{W}')\sim(\mathsf{W},\mathsf{W}')$
meaning that\/ $\,\mathsf{W}=\tilde\mathsf{W}\,$ or\/ $\,\mathsf{W}'=\tilde\mathsf{W}'\hskip-.7pt$.
}
\end{lemma}
\begin{proof}If $\,\mathsf{W}_0^{\phantom i}=\mathsf{W}$, our claim is obvious as
$\,(\mathsf{W}_0^{\phantom i},\mathsf{W}_0')\sim(\mathsf{W},\mathsf{W}')$. Otherwise we may first choose
$\,\mathsf{W}_1^{\phantom i}=\mathsf{W}_0^{\phantom i}$ and $\,\mathsf{W}_1'$ such that
$\,\mathsf{W}_0^{\phantom i}\cap\mathsf{W}\subseteq\mathsf{W}_1'\hskip-.7pt\subseteq\mathsf{W}_0^{\phantom i}$, and then select
$\,\mathsf{W}_2'=\mathsf{W}_1'$ along with $\,\mathsf{W}_2^{\phantom i}$ spanned by $\,\mathsf{W}_1'$ and a vector in
$\,\mathsf{W}\smallsetminus\mathsf{W}_0^{\phantom i}$. Now
$\,(\mathsf{W}_0^{\phantom i},\mathsf{W}_0')\sim(\mathsf{W}_1^{\phantom i},\mathsf{W}_1')\sim(\mathsf{W}_2^{\phantom i},\mathsf{W}_2')\,$ and
$\,\dim(\mathsf{W}_2^{\phantom i}\cap\mathsf{W})>\dim(\mathsf{W}_0^{\phantom i}\cap\mathsf{W})$. This step may be repeated for
$\,(\mathsf{W}_2^{\phantom i},\mathsf{W}_2')\,$ instead of $\,(\mathsf{W}_0^{\phantom i},\mathsf{W}_0')$, as long as
$\,\mathsf{W}_2^{\phantom i}\ne\mathsf{W}$.
\end{proof}
\begin{corollary}\label{strbg}{\medit
Let\/ $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ be any Grass\-mann\-i\-an triple, arising from some
data\/ {\rm(\ref{dta}.i)} as in Section\/~{\rm\ref{eg}}. Then the direct
sum\/ $\,\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$
appearing in Lemma\/~{\rm\ref{intpr}(i)} is a strongly
brack\-et-gen\-er\-at\-ing distribution on\/ $\,M}%{{\mathrm{M}}'\hskip-.7pt$, in the sense that
any two points of\/ $\,M}%{{\mathrm{M}}'$ can be joined by a piecewise\/ $\,C^\infty$ curve
tangent to\/
$\,\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-\hskip-.7pt$.
}
\end{corollary}
\begin{proof}According to (\ref{mpv}), whenever
$\,(\tilde\mathsf{W},\tilde\mathsf{W}')\sim(\mathsf{W},\mathsf{W}')\,$ in Lemma~\ref{ttlds}, both
$\,(\tilde\mathsf{W},\tilde\mathsf{W}')\,$ and $\,(\mathsf{W},\mathsf{W}')\,$ must lie in the same fibre
of one of the bundle projections (\ref{hbp}). As the fibres of either
projection (\ref{hbp}), being complex projective spaces (see the last line in
Lemma~\ref{lfspc}), are connected, the strong brack\-et-gen\-er\-at\-ing
property thus follows for the di\-rect-sum distribution of Remark~\ref{trint}.
Our claim is now immediate since
$\,\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$ projects
onto that latter distribution under the quo\-tient-space projection
$\,M}%{{\mathrm{M}}'\hskip-.7pt\toM}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\hskip-.7pt$, which also has connected fibres
(bi\-hol\-o\-mor\-phic to twice-punc\-tur\-ed complex projective lines, cf.\
Lemma~\ref{vsbkr}(b)).
\end{proof}
\begin{remark}\label{nintg}A compact ge\-o\-des\-ic-gra\-di\-ent K\"ah\-ler
triple need not, in general, satisfy conditions (i) -- (iii) of
Lemma~\ref{intpr}, that is, (\ref{spn}). Examples are provided by
all Grass\-mann\-i\-an triples $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)\,$ arising via Lemma~\ref{chone}
from data (\ref{dta}.i) such that $\,2\le k\le n-2$, where
$\,n=\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\hskip-1.5pt\mathsf{V}\hskip-.7pt$.
Namely, in (\ref{dmq}), $\,q=(k-1)(n-1-k)\,$ as $\,m=(n-k)k\,$ (see
Remark~\ref{grass}) and, similarly,
$\,\{d_+^{\phantom i},\hskip.7pt d_-^{\phantom i}\}=\{(n-k)(k-1),(n-1-k)k\}\,$ from (\ref{spm}.a) --
(\ref{spm}.b), where $\,\dim_{\hskip.4pt\bbC}\w\hskip-.7pt\nh\mathsf{L}=1\,$ by (\ref{dta}.i). Thus, $\,q>0\,$ and
$\,\hskip.7pt\mathcal{V}\hskip.7pt\oplus\mathcal{H}^+\hskip-3pt\oplus\mathcal{H}^-$ in
(\ref{tme}) is a proper sub\-bun\-dle of $\,T\nnhM}%{{\mathrm{M}}'\hskip-.7pt$. Consequently, due
to Corollary~\ref{strbg}, it cannot be in\-te\-gra\-ble.
\end{remark}
\begin{remark}\label{twweb}For any compact ge\-o\-des\-ic-gra\-di\-ent
K\"ah\-ler triple $\,(M}%{{\mathrm{M}}\hskip-.7pt,g,\vt)$, the leaf space
$\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\,$ carries what might be called a {\medit
hol\-o\-mor\-phic $\,2$-web of complex projective spaces}, formed by the two
hol\-o\-mor\-phic fibrations (\ref{hbp}) with fibres bi\-hol\-o\-mor\-phic to
(pos\-i\-\hbox{tive\hskip.4pt-}\hskip0ptdi\-men\-sion\-al) complex projective spaces,
having the triv\-i\-al-in\-ter\-sec\-tion property of Remark~\ref{trint}.
There is also a natural hol\-o\-mor\-phic complex line bundle over
$\,M}%{{\mathrm{M}}'\hskip-3pt/\mathcal{V}\hskip-.7pt$, the restriction of which to every fibre of
$\,\pi^+$ (or, $\,\pi^-$), with (\ref{hbp}), is bi\-hol\-o\-mor\-phic\-al\-ly
isomorphic to the tautological (or, respectively, dual tautological) bundle of
the fibre. Specifically, the complex line attached to a leaf
$\,\varLambda\hskip.7pt\subseteqM}%{{\mathrm{M}}'$ of $\,\mathcal{V}\,$ is
$\,\{0\}\cup\varPhi^{-\nnh1}\hskip-1.5pt(\varLambda)\subseteqN}%{{\mathrm{N}}\hskip-2.4pt_y^{\phantom i}\varSigma^\pm\hskip-.7pt$,
cf.\ Theorem~\ref{first}(iii); that changing the sign $\,\pm\,$ to $\,\mp\,$
leads to its dual complex line follows from
\cite[Remark 4.1]{derdzinski-maschler-06} and (\ref{hgv}.a) -- (\ref{hgv}.b).
\end{remark}
|
\section{Introduction}
Network data occurs in a range of fields, and its analysis has become a highly interdisciplinary effort
\citep{Easley:Kleinberg:2010:1,Handcock:Gile:2010:1,Kolaczyk:2009:1,Newman:2009}.
In statistical network analysis, two classes of models have recently received particular attention: Graphon models
\citep{Borgs:Chayes:Lovasz:Sos:Vesztergombi:2008,Hoff:Raftery:Handcock:2002:1,Orbanz:Roy:2014},
and the subclass of stochastic block models (SBMs) \citep{allmanetal11,bickel13,Goldenberg:Zheng:Fienberg:Airoldi:2009}.
These models parametrize a random graph by a
symmetric measurable function $w$, which can be interpreted as representing an adjacency matrix in the limit
of infinite graph size \citep{Borgs:Chayes:Lovasz:Sos:Vesztergombi:2008}. In a SBM, the function is in particular piece-wise constant.
Examples of statistical problems arising in this field include \emph{estimation problems} (see below),
\emph{class label recovery}
\citep{bickelchen09,Lei:Rinaldo:2015:1,Lei:Zhu:2017:1,vvvp,zhangzhou}, and \emph{signal detection}, which refers to testing for the presence of a signal in settings where observed data constitutes a network or array \citep{accd11,ariascv14,butuceaingster,verzelenac15}.
The work described here concerns estimation problems. SBMs are parametrized by the sizes of communities, which subdivide the domain of $w$ into regions of constant value, and by the values of the function $w$ on these regions. The estimation of these parameters is typically complicated by the fact that the specific subdivision in regions, or `labelling', is not observed.
Remarkably, the pointwise asymptotic results obtained by \citet{bickel13} imply SBM parameters can be recovered at the same rate regardless of whether or not labels are observed, at least asymptotically and on the interior of the parameter set. Earlier results include \citep{allmanetal11,celisseetal,channarondetal}. Other functionals can also be of interest; estimation of moments, for instance, is considered in \citet{ambroisematias12} and \citet{bickeletal11}.
In more general graphon models, one can estimate the complete function $w$
\citep{chaogr15,kvt16,wolfeohlede}.
Of particular relevance for our purposes is an idea highlighted by \citet*{chaogr15}: If one estimates
the entire parameter function $w$ in an uniform way, not
observing labels slows the rate---which is not the case in the setting studied by \citet{bickel13}.
In this paper, we consider uniform, non-asymptotic rates of convergence for estimation of
certain parameters of graphon models. We begin with one of the simplest graphon models conceivable, a two-class SBM specified by a single, scalar parameter. Even under this simple model, uniform and pointwise rates differ significantly.
We then derive results for SBMs with $k$ classes, where $k$ is not necessarily fixed and may grow to infinity with $n$. The phenomenon observed for $k=2$ generalises to a general $k$, with $k$ possibly growing with $n$, with a rate depending on $n,k$ and the possible sparsity of the graph. Finally, we derive results for certain functionals under a smooth graphon model.
Before we describe the results more precisely, we define the models and introduce the notation used in the sequel.
\vspace{.5cm}
{\em Stochastic blockmodels and graphon models.} Consider sampling at random
an undirected, simple graph $G$
on the vertex set $\EM{\mathcal{V}}=\{1,\ldots,n\}$ as follows. Fix some $k\in\{1,\ldots,n\}$.
Let ${\pi=(\pi_1,\ldots,\pi_k)}$ be a probability distribution
on the set ${\lbrace 1,\ldots,k\rbrace}$, with $\pi$ identified as a line vector of size $k$.
Let ${M}:=(M_{lm})$ be a symmetric ${k\times k}$ matrix with elements ${M_{lm}\in[0,1]}$.
To sample a graph $G$, we generate its adjacency matrix $X=(X_{ij})_{i,j\in\EM{\mathcal{V}}}$. Since $G$ is undirected, it suffices
to sample entries with ${i<j}$, and writing
Be$(p)$ as a shorthand for Bernoulli$(p)$,
\begin{enumerate}
\item For each vertex $i\in \EM{\mathcal{V}}$, independently generate a label ${\varphi(i)\,\sim\,\pi}$.
\item For each pair ${i<j}$ in $\EM{\mathcal{V}}$, independently sample ${X_{ij}\,|\, \varphi(i),\varphi(j)\,\sim\, \text{Be}(M_{\varphi(i)\varphi(j)})}$.
\end{enumerate}
In this notation, $\varphi$ is a (random) mapping $\varphi:\{1,\ldots,n\}\to\{1,\ldots,k\}$ that attributes a label to each node of the graph. It is random because labels are by definition randomly sampled.
The distribution $P_{\pi,M}$ so defined on the set of undirected, simple graphs
is called a \emph{stochastic blockmodel} of order $k$ with parameters $\pi$ and $M$.
One can also write
\begin{align}
\begin{aligned} \label{modsbm}
(\varphi(1),\ldots,\varphi(n))\ & \sim\ \pi^{\otimes n} \\
(X_{ij})_{i<j}\, \,|\,\, \varphi\ & \sim\ \bigotimes_{i<j} \text{Be}(M_{\varphi(i)\varphi(j)}),
\end{aligned}
\end{align}
where $\otimes$ denotes a tensor product of distributions, $\pi^{\otimes n}=\pi\otimes\cdots\otimes \pi$, and here and in the sequel $i<j$ refers to all pairs of indices
$(i,j)\in\EM{\mathcal{V}}^2$ with $i<j$. Any given $\varphi$ partitions the vertex set ${\lbrace 1,\ldots,n\rbrace}$ into
$k$ distinct classes.
We call $\pi$ the {\em proportions} vector and $M$ the matrix of {\em connectivity} parameters.
These models can be regarded as a special case of a more general class of random graphs, parametrized by
the set
of all measurable functions ${w:[0,1]^2\rightarrow[0,1]}$ that are symmetric,
i.e.\ ${w(x,y)=w(y,x)}$. Any such $w$ defines a random graph $G$: denoting by $\text{Unif}[0,1]$ the uniform distribution on $[0,1]$, and $(U_i)_i=(U_i)_{1\le i\le n}$, set
\begin{align}
\begin{aligned} \label{modgr}
(U_i)_{i} \ & \sim \ \text{Unif}[0,1]^{\otimes n} \\
(X_{ij})_{i<j}\, \,|\,\, (U_i)_i \ & \sim \
\bigotimes_{i<j} \text{Be}(w(U_i,U_j)).
\end{aligned}
\end{align}
The law $P_w$ of the graph $G$ defined by the random matrix $X$ in \eqref{modgr} is called a \emph{graphon model}
\citep{Borgs:Chayes:Lovasz:Sos:Vesztergombi:2008}.
{\em Mixture interpretation.}
The $n$-tuple $(U_i)$ in a graphon model, or, equivalenly, the mapping $\varphi$ in a SBM, are in general
not observed, and can hence be interpreted as latent variables. In other words, the distribution of
the data $(X_{ij})_{i<j}$ is a {\em mixture}. Also, either class of models generalises to directed graphs,
by dropping the symmetry constraints on $\pi$ and $w$, and requiring only ${i\neq j}$ rather than ${i<j}$;
in the following, we consider only the undirected case.
The data distribution in \eqref{modgr} remains invariant if $w$ is replaced by $w\circ g$, for any measure-preserving transformation $g$ of $[0,1]$: $P_w=P_{\tilde{w}}$ for ${\tilde{w}(x,y)=w(g(x),g(y))}$. More generally, two graphons $w$ and $w'$ are considered equivalent if ${P_w=P_{w'}}$. The equivalence class $\left<w\right>$ of $w$ is called a graph limit.
Similarly in \eqref{modsbm}, if $\sigma$ is a fixed arbitrary permutation of $\{1,\ldots,k\}$, with permutation matrix $\Sigma$, then $P_{\pi,M}=P_{\pi\Sigma,\Sigma M\Sigma^T}$. The parameters of the SBM can only be recovered up to label switching. We refer to \cite{allmanetal11} and \cite{celisseetal} for detailed identifiability statements.
Stochastic blockmodels are special cases of graphon models,
obtained by choosing $w$ as a histogram:
subdivide the unit interval into $k$ intervals ${I_s:=[\,\sum_{i<s}\pi_i,\sum_{i\leq s}\pi_i)}$
of respective lengths $\pi_s$, and set
\begin{equation}
\label{SBM:graphon}
w(x,y):=M_{ij} \quad\text{ for }\quad x\in I_i, y\in I_j\;.
\end{equation}
Then $P_w=P_{\pi,M}$. In a graphon model, the continuous vertex labels $U_i$
are almost surely distinct; in a stochastic block model, labels coincide whenever two
vertices belong to the same class. Thus, the SBM labels
can be regarded as discretization of graphon labels. Conversely, any graphon can be approximated
by a sequence of stochastic blockmodels of increasing order $k$; indeed, the set of stochastic
blockmodels---that is, of graphons of the form \eqref{SBM:graphon} for all $k$, $\pi$ and $M$---is
dense in the set of functions $w$ endowed with its natural topology \citep[see e.g.][for details]{Janson:2013:1}. This idea
can be used to construct SBM-valued estimators for graphons \citep{wolfeohlede,chaogr15}.
{\em Fixed and random design.} In models \eqref{modsbm}-\eqref{modgr}, the latent variables, respectively $\varphi$ and $U$, are random. Sometimes, a slightly different version of the model is considered, where $\varphi$ and $U$ are still unobserved, but fixed, non-random quantities. For instance, under this setting \eqref{modsbm} becomes
\[(X_{ij})_{i<j}\, \sim\ \bigotimes_{i<j} \text{Be}(M_{\varphi(i)\varphi(j)}),\]
for a given, unknown, $\varphi:\{1,\ldots,n\}\to\{1,\ldots,k\}$, and the data distribution is denoted $P_{\varphi,M}$.
Such models will be referred to as {\em fixed design} SBM and {\em random design} SBM respectively.
Some theoretical arguments simplify in the fixed design case, for which the data distribution is a product measure, rather than a mixture of products measures. Most results below are obtained in both fixed and random design case.\\
{\em Estimation in SBMs.} We briefly review some existing results on SBMs. For parameters that can be expressed as moments, such as the edge density, central limit theorems can be derived for empirical moments \citep{bickeletal11,ambroisematias12}.
Consider the estimation of the parameters $\pi$ and $M$ of the SBM itself, for a fixed (and known) number of classes $k$. The labels, given by the mapping $\varphi$, are not observed. If they were observed, Lemma 2 of \cite{bickel13} shows that the model is locally asymptotically normal, with rescaling rates respectively $1/\sqrt{n}$ and $1/n$ at any `interior' point $(\pi_0,M_0)$ (see below). For known labels, this suggests a difference in the speed of estimation for the proportions vector $\pi$ and connectivity matrix $M$. Remarkably, Theorems 1-2 of \cite{bickel13} show that in the original model \eqref{modsbm}---that is, for unobserved labels---the parameters $\pi$ and $M$ can be estimated at rate respectively $1/\sqrt{n}$ and $1/n$.
Denote by $\hat\pi$, $\hat M$ the profile maximum likelihood estimators
of $\pi, M$. Then, as $n\to\infty$,
\begin{align}
\begin{aligned} \label{asyn}
\sqrt{n}({{\hat \pi}}{}^* - \pi_0)\ & \to \ \EM{\mathcal{N}}(0,T_1) \\
n({{\hat M}}{}^* - M_0)\ & \to \ \EM{\mathcal{N}}(0,T_2)
\end{aligned}
\end{align}
where $( {{\hat \pi}}{}^*,{{\hat M}}{}^* )$ denote some label switched-version of $(\hat \pi,\hat M)$, which may be random and depend on $n$, and the convergence is in distribution, under $P_{\pi_0,M_0}$, to normal limits with covariance matrices $T_1$ and $T_2$. Here, $T_1$ is a $k\times k$ matrix, whereas $T_2$ is of size ${k^2\times k^2}$, and the matrices $\hat{M}$ and $M_0$ are treated as vectors of size $k^2$. It can also be of interest to consider sparser versions of the model, and \cite{bickel13} more generally obtain results when the connectivity matrix is normalised by a factor $\rho$ that may go to $0$ with $n$.
The asymptotic result \eqref{asyn} holds under the assumption that the number of classes $k$ is known and fixed as $n\to\infty$, say $k=k_0$, and that the true parameter $(\pi_0,M_0)$ is such that all coordinates of $\pi_0$ are nonzero, and no two lines of the matrix $M_0$ coincide. This can be thought of as saying that the true parameter should be in the `interior' of the parameter set. In the sequel, as we shall work with the {\em quadratic} risk, the rates in \eqref{asyn} will be squared, and we refer to a rate as {\em fast} if it is of order $1/n^2$, that is scaling as the number of edges, and as {\em slow} if it is of order $1/n$, that is scaling as the number of nodes.
{\em Our contribution.} The convergence $\eqref{asyn}$ is asymptotic and pointwise at the interior point $\pi_0, M_0$. One main reason behind the possibility to recover, under some conditions, the fast rates as in \eqref{asyn} even though labels are not observed is the possibility to estimate reasonably well the unobserved labels. However, as our results below will show, there are regions of the parameter set, that depend both on $n$ and the number of classes $k$, over which rates drop significantly. This is connects to the possibility of obtaining {\em non-asymptotic and uniform} rates result over the parameter set. To fix ideas, consider the following simple example. Take $n=30$, $k=5$, let $\pi$ be fixed for simplicity to equiproportions $\pi=[1/5,1/5,1/5,1/5,1/5]$, and let the true connectivity matrix $M=M_0$ be given below. One also defines a matrix $\tilde M$ as follows
\[
M_0 = \begin{bmatrix}
.55 & .45 & .4 & .1 & .7 \\
.45 & .55 & .4 & .1 & .7\\
.4 & .4 & .6 & .2 & .4 \\
.1 & .1 & .2 & .1& .4 \\
.7 & .7 & .4 & .4 & .2
\end{bmatrix},\qquad
\tilde{M} = \begin{bmatrix}
.5 & .5 & .4 & .1 & .7 \\
.5 & .5 & .4 & .1 & .7\\
.4 & .4 & .6 & .2 & .4 \\
.1 & .1 & .2 & .1& .4 \\
.7 & .7 & .4 & .4 & .2
\end{bmatrix}.
\]
One notes that the upper-left corner of $M_0$ is close to $\begin{bmatrix}
.5 & .5 \\
.5 & .5
\end{bmatrix}$. Replacing the upper-left corner of $M_0$ by this matrix leads to the matrix $\tilde M$. Note that the first two lines of $\tilde M$ are equal and that it in fact it corresponds to a SBM with $k=4$ classes instead of $5$. Does the fact that $M_0$ is close to $\tilde M$ change something in terms of estimation rates? To address this question, we consider the simplest possible statistical setting of SBMs where everything is known except one parameter, and we derive estimation rates for it. In this paper, we will focus on the {\em connectivity} parameters, that is on the elements of $M$, so the proportions vector $\pi$ will typically be kept fixed. Section \ref{sec2} considers the toy example where $k=2$. It will be seen that the estimation rate drops if one requires uniformity.
In Section \ref{sec3}, we consider SBMs with $k$ classes. We obtain a minimax lower bound for estimation of a parameter of the connectivity matrix that generalises the case $k=2$. The number $k$ of classes is allowed to depend on $n$, and the result is {\em local} around {\em any} possible SBM. For instance, it can be applied to the example with $M$ and $\tilde M$ above. We also show that the rate is sharp for `most' connectivity matrices $M$, at least if $k$ does not grow too fast with $n$. In Section \ref{sec4}, we derive some results for certain functionals of the graphon model \eqref{modgr} and conclude with a short discussion. Appendices \ref{secpr}-\ref{sec:lemmas} collect proofs and a number of useful lemmas. \\
{\em Notation.} Denote by ${[k]^n}$ the set
of all mappings $\{1,\ldots,n\}\to \{1,\ldots,k\}$.
A SBM is entirely specified through three parameters: the number of classes $k$, a vector of proportions $\pi$ that belongs to the $k$-dimensional simplex, and the connectivity matrix $M$ of size $k\times k$. The representation of the resulting distribution $P_{\pi,M}$ as a mixture is useful to make the link between the standard random design case and the fixed design case. We have
\begin{equation}\label{distrib}
P_{\pi,M} = \sum_{\varphi\in [k]^n} \mu_{\pi}[\varphi] \,
\bigotimes_{i<j} \text{Be}(M_{\varphi(i)\varphi(j)}),
\end{equation}
where, for $\EM{\mathcal{M}}(n;\pi_1,\ldots,\pi_k)$ the multinomial distribution
with parameters $n,\pi_1,\ldots,\pi_k$ and $N_j(\varphi)=\sum_{i=1}^n 1\!{\rm l}_{\varphi(i)=j}$ the number of times the label $j$ is present, the multinomial probabilities are given by
\[ \mu_{\pi}[\varphi] =
Pr\left[\,\EM{\mathcal{M}}(n;\pi_1,\ldots,\pi_k)=(N_1(\varphi),\ldots,N_k(\varphi))\,\right].\]
In the fixed design model, the labels given through $\varphi$ are fixed and unobserved, so that the distribution is $P_{\varphi,M}$ given by
\[ P_{\varphi,M} = \bigotimes_{i<j} \text{Be}(M_{\varphi(i)\varphi(j)}).\]
For $A$ a subset of the integers, let $|A|$ denote its cardinality. For $M$ a square matrix, let $\|M\|_F$ denote its Frobenius norm
and $\|M\|_{Sp}$ its spectral norm. Let $ER(p)$ denote the Erd\"os-Renyi distribution with parameter $p$ over $n$ nodes, that is ${ER(p)=ER(p,n)=\otimes_{i<j} \text{Be}(p)}$.
\section{Toy example: the case $k=2$} \label{sec2}
In this section, we consider the case $k=2$. Let, for $\theta\in [-1/2,1/2]$,
\begin{align}
e\ \ = & \quad \left[\frac12\, , \, \frac12\right] \label{prop2} \\
Q^\theta \ = &
\begin{bmatrix}
\frac12 + \theta & \frac12 - \theta\\
\frac12 - \theta & \frac12 + \theta
\end{bmatrix}. \label{simat}
\end{align}
Recalling that $P_{\pi,M}$ denotes the distribution of the data $X$ in the SBM \eqref{modsbm} with parameters $\pi,M$, consider the submodel
\begin{equation}
\EM{\mathcal{M}} = \left\{ P_\theta:= P_{e,Q^\theta},\quad \theta\in[-1/2,1/2] \right\}.
\end{equation}
The set $\EM{\mathcal{M}}$ is a $1$--dimensional submodel through the set of all SBMs with at most two classes. For $\theta=0$ the matrix $Q^0$ is degenerate and the model is simply an Erd\"os-Reyni graph model with parameter $1/2$, that is all edges are independent and have a probability $1/2$ of being present. In this notation, the case $\theta=0$ is somewhat overspecified and could also more simply be given by $1$-dimensional vectors $\tilde e=[\,1\,]$, $\tilde q=[\, 1/2\,]$ (note that the fact that there are various parameterisations for the case $\theta=0$ causes no identifiability issue, as the probability measure $P_0$ is well identified and distinct from all $P_\theta$, $\theta\neq 0$). Connectivity matrices with two parameters, one for intra-group and one for between-group connections (so-called affiliation models), are common in the literature on random graphs models \citep[e.g.][and references therein]{ambroisematias12,ariascv14,Mossel2016}.
In the fixed design case, the model is $\EM{\mathcal{M}}_F=\{P_{\theta,\varphi}:=P_{\varphi,Q^\theta},\quad
\theta\in[-1/2,1/2],\ \varphi\in [2]^n\}$. Expectations with respect to the measures $P_\theta$ and $P_{\theta,\varphi}$ are denoted respectively $E_\theta$ and $E_{\theta,\varphi}$.
\begin{thm} \label{thm-twocl}
Consider a stochastic blockmodel \eqref{modsbm} with $k=2$ specified by $\EM{\mathcal{M}}$, that is $P_\theta=P_{e,Q^\theta}$ with $e,Q^\theta$ given by \eqref{prop2}-\eqref{simat}. There exists a constant $c_1>0$ such that for all $n\ge 2$,
\[ \inf_{T} \sup_{\theta\in[-1/2,1/2]} E_{\theta} \left[ T(X) - \theta \right]^2
\ge \frac{c_1}{n}, \]
where the infimum is taken over all estimators $T$ of $\theta$ in the model $\EM{\mathcal{M}}$.
\end{thm}
\vspace{.5cm}
Theorem \ref{thm-twocl} states that, even in a very simple SBM with ${k=2}$ classes and only one unknown parameter in its connectivity matrix, the minimax estimation rate is no faster than $1/n$.
This is no contradiction to the fast rate obtained by \citet{bickel13} (meaning a $1/n$ rate for the convergence in distribution but a $1/n^2$ rate for the quadratic risk): the latter is a pointwise asymptotic result, and assumes that no two lines of the connectivity matrix are the same, whereas Theorem \ref{thm-twocl} is nonasymptotic and uniform. It shows that the rate in a two-class model changes for distributions close to an Erd\H{o}s-Renyi model ($k=1$); informally, models close to the `boundary' are harder to estimate. We note the result does not require the sub-model $\EM{\mathcal{M}}$ to include the Erd\H{o}s-Renyi model; see the remark below.
The phenomenon is reminiscent of effects familiar from community detection, where matrices similar to \eqref{simat} naturally arise as most difficult submodels. Community detection is a {\em testing} problem, though, as opposed to the {\em estimation} problem considered here.
For a different but related result in the very sparse case, see \citep{Mossel2016}.\\
\noindent {\em Remarks.} (a)
In Theorem \ref{thm-twocl}, one can take $c_1=1/107$; additionally, the supremum can be restricted to $(-\theta_n,\theta_n)$ for
\[ \theta_n=\frac{c_0}{\sqrt{n}} \qquad \text{and}\quad c_0=\frac{1}{3\cdot2^{3/4}}\approx 0.56.\]
Moreover, the proof implies that one can restrict the supremum to a set not actually containing $\theta=0$, but rather two points close enough to $\theta=0$, namely $\theta_1=c_1/\sqrt{n}$, $\theta_2=c_2/\sqrt{n}$ for suitably chosen, fixed constants $c_1, c_2>0$.\\
\noindent (b) A similar result as in Theorem \ref{thm-twocl} holds in the fixed design model, to wit
\[ \inf_{T_f} \sup_{\theta\in[-1/2,1/2],\,\varphi\in[2]^n}
E_{\theta,\varphi} \left[ T_f(X) - \theta \right]^2
\ge \frac{c_1}{n}, \]
where the infimum is taken over all estimators $T_f$ of $\theta$ in the fixed design model. \\
Theorem \ref{thm-twocl} provides a
lower bound. There is a corresponding, matching upper-bound, which we obtain next, by defining an estimator of $\theta$ whose maximum quadratic risk matches the lower bound of Theorem \ref{thm-twocl}. To do so, let, for any $\sigma$ an element of $[2]^n$, that is a mapping $\{1,\ldots,n\}\to\{1,2\}$,
\begin{align}
2Z_n(\sigma,X) & := - \sum_{i<j,\ \sigma(i)=\sigma(j)} (1-2X_{ij}) +
\sum_{i<j,\ \sigma(i)\neq \sigma(j)} (1-2X_{ij}). \label{zedns}
\end{align}
Maximising \eqref{zedns} in $\sigma$ leads to set
\[ \hat \sigma = \underset{\sigma\in [2]^n}{\text{argmax }} |Z_n(\sigma,X)| \]
which leads to the profile maximum likelihood estimate
\begin{align}
\hat \theta & = \frac{Z_n(\hat\sigma,X)}{b_n}\qquad\text{ and }\qquad b_n ={n\choose 2} = \frac{n(n-1)}{2} \label{pseml}.
\end{align}
This estimator
can be seen as a (pseudo)-maximum likelihood estimate, see Appendix \ref{sec-ub}.
\begin{thm} \label{thmub}
Consider a stochastic blockmodel \eqref{modsbm} with $k=2$ specified by $\EM{\mathcal{M}}$, that is, $P_\theta=P_{e,Q^\theta}$ with $e,Q^\theta$ given by \eqref{prop2}-\eqref{simat}. Let $\hat\theta$ be the estimator defined by \eqref{pseml}. There exists a constant $C_1>0$ such that for all $n\ge 2$,
\[ \sup_{\theta\in [-1/2,1/2]} E_{\theta}\left[ \hat\theta - \theta \right]^2
\le \frac{C_1}{n}.
\]
The same risk bound holds for $\hat\theta$ in the fixed design model, uniformly over $\theta$ and $\varphi\in [2]^n$.
\end{thm}
\vspace{.5cm}
{\it{Remark} (different parameter choices)}. It is not hard to check that the results of Theorems \ref{thm-twocl}-\ref{thmub} are unchanged if instead of $1/2$ in the matrix $Q^\theta$ in \eqref{simat}, another number $a_0\in(0,1)$ is used. If $a_0$ is bounded away from $0$ and $1$, assuming $\min(a_0,1-a_0)\ge \rho>0$, then the results are only modified by constants. Also, if the proportions vector $\pi$ is of the form
$[b\,,\,1-b]$ with $b>0$, similar results continue to hold, provided the matrix $Q^\theta$ is replaced by
\[
Q^\theta_b \ =
\begin{bmatrix}
\frac12 + c_b\theta & \frac12 - d_b\theta\\
\frac12 - d_b\theta & \frac12 + c_b\theta
\end{bmatrix}
\]
for suitable constants $c_b, d_b$ that depend on $b$ (one can take e.g. $c_b=1-b$ and $d_b=b$).
\vspace{.5cm}
Since the maximum likelihood estimator \eqref{ub-pre}
has to optimize over the set $[k]^n$, it is not easily computable.
A simple alternative is to use a spectral method, see e.g. \cite{Lei:Rinaldo:2015:1}. The proof of the following result (see Appendix \ref{sec:ub_spec}) is based on variation of tools used in \cite{Lei:Rinaldo:2015:1}.
With the convention that $X_{ii}=1/2$ and $X_{ji}=X_{ij}$, define the $n\times n$ matrix $\Delta$ by
\[
\Delta := X - \frac12 J, \quad\text{ where }\quad
J := \bigl(1\bigr)_{i,j\leq n}.
\]
{{\em Spectral algorithm }$\EM{\mathcal{S}}_2$}. Let $\lambda_{1}^a(\Delta)$ denote the largest eigenvalue in absolute value of $\Delta$ and set
\begin{equation} \label{te_spec}
\tilde\theta := \frac{\lambda_{1}^a(\Delta)}{n-1}\;.
\end{equation}
We refer to this procedure as spectral algorithm for $k=2$ and denote it $\EM{\mathcal{S}}_2$.
The intuition behind this estimator in the fixed design setting is the following. For $i\neq j$, we have
\[ E[X_{ij} - \frac12] = M_{\varphi(i)\varphi(j)}^\theta - \frac12
=(-1)^{1\!{\rm l}_{\varphi(i)\neq \varphi(j)}}\theta. \]
Set ${v=((-1)^{\mathds{1}\lbrace \varphi(i)=1\rbrace})_{i\leq n}}$ and $V := vv^t=\bigl((-1)^{\mathds{1}\lbrace \varphi(i)\neq\varphi(j)\rbrace}\bigr)_{i,j\leq n}$.
Then for non-random $\varphi$,
\[ E[\Delta] = \theta(V-I_n),\]
where $I_n$ is the identity matrix of size $n$. As $E[\Delta]$ is a rank $1$ matrix whose non-zero eigenvalue equals $(n-1)\theta$ (with $v$ the corresponding eigenvector), this leads us to introduce $\tilde\theta$ as in \eqref{te_spec}.
\begin{thm} \label{thm_ubspec}
In the same setting as in Theorem \ref{thmub}, let $\tilde\theta$ be the estimator defined by \eqref{te_spec}. There exists a constant $C>0$ such that for all $n\ge 2$,
\[ \sup_{\theta\in [-1/2,1/2]} E_{\theta}\left[ \tilde\theta - \theta \right]^2
\le \frac{C}{n}.
\]
The same risk bound holds for $\hat\theta$ in the fixed design model, uniformly over $\theta$ and $\varphi\in [2]^n$.
\end{thm}
A small simulation study in Section \ref{sec:simulation} illustrates the behaviour of the estimator.
The main takeaway of the results above is that the {\em uniform} quadratic rate of estimation along the submodel $\EM{\mathcal{M}}$ for estimating a connectivity parameter is exactly of order $n^{-1}$, up to constants, as follows from combining
Theorems \ref{thm-twocl} and \ref{thmub} (or \ref{thm_ubspec}). This `slow' rate (as compared to the asymptotic pointwise quadratic rate $n^{-2}$ of \eqref{simat}) arises even if all other parameters---here, the vector of proportions $\pi$---are assumed known.
The submodel built for ${k=2}$ can be regarded as a local perturbation of an Erd\"os-Renyi graph model with connection probability $1/2$. The drop in the rate is already noteworthy, as the rate of estimation of $p$ for a ER$(p)$ model is of the order $n^{-2}$. More generally, however, one may consider perturbations around a SBM with ${k-1}$ classes instead of the ER$(p)$ model, and ask whether similar results hold for other choices of $k$, and whether the rate changes with $k$. More precisely: Can one describe convergence rates along simple submodels, in one unknown connectivity parameter, that still exhibit slower rates of convergence than the expected pointwise rate? This question is considered in the next section.\\
\section{Main result and local minimax bounds for a SBM with $k$ classes} \label{sec3}
The connectivity matrix of a SBM with at most ${k-1}$ classes
is of the form
\begin{equation} \label{coM}
M =
\begin{bmatrix}
a_0 & a_1 & \cdots & a_{k-2} \\
a_1 & b_{11} & \cdots & b_{1 k-2} \\
\vdots & \vdots & & \vdots \\
a_{k-2} & b_{1 k-2} & \cdots & b_{k-2 k-2}
\end{bmatrix}
\qquad\text{ where }a_0,a_i,b_{ij}\in[0,1]\text{ for }i,j\in[k-2]\;.
\end{equation}
For simplicity of notation, and comparability to Section \ref{sec2}, we assume $a_0=1/2$ throughout. Results are easily adapted to the case $a_0\in (0,1)$, requiring only that $a_0$ be bounded away from $0$ and $1$, as in the remark above. If needed, one can ensure the number of classes is exactly ${k-1}$ by requiring no two rows of $M$ coincide, which will be (only) used for spectral estimators in Theorems \ref{thmspec} and \ref{thm-twocl-spa} below.
We consider $1$-dimensional submodels in the parameter space of connectivity matrices: Set
\begin{equation}
e_k = \left[\frac1k, \cdots, \frac1k \right], \label{propk}
\end{equation}
and, for coefficients $\{a_i\}, \{b_{ij}\}$ as above, define
\begin{equation}
M^\theta
= \begin{bmatrix}
\frac12 + \theta & \frac12 - \theta & a_1 & \cdots & a_{k-2} \\
\frac12 - \theta & \frac12 + \theta & a_1 & \cdots & a_{k-2} \\
a_1 & a_1 & b_{11} & \cdots & b_{1 k-2} \\
\vdots & \vdots & \vdots & & \vdots \\
a_{k-2} & a_{k-2} & b_{1 k-2} & \cdots & b_{k-2 k-2}
\end{bmatrix}
=
\begin{bmatrix}
Q^\theta & A\\
A^T & B
\end{bmatrix}
,
\label{zte}
\end{equation}
where
\[ Q^\theta \ =
\begin{bmatrix}
\frac12 + \theta & \frac12 - \theta\\
\frac12 - \theta & \frac12 + \theta
\end{bmatrix}, \quad
A =
\begin{bmatrix}
a_1 & a_2 & \ldots & a_{k-2} \\
a_1 & a_2 & \ldots & a_{k-2}
\end{bmatrix}, \quad
B=
\begin{bmatrix}
b_{11} & \cdots & b_{1 k-2} \\
\vdots & & \vdots \\
b_{1 k-2} & \cdots & b_{k-2 k-2}
\end{bmatrix}.
\]
Thus, $M^{\theta}$ is a symmetric $k\times k$ matrix, obtained from $M$ by replacing the scalar coefficient $a_0$ by the $2\times 2$ matrix $Q^{\theta}$, and repeating the vector $(a_i)_{1\le i\le k-2}$.
The number of nodes in a given class will be specified as follows.
In the random design model for simplicity below we choose the proportions vector $\pi$ in \eqref{modsbm} equiproportional and equal to $e_k$ in \eqref{propk}, although (as in the case $k=2$) analogous results can be obtained if the proportions are of similar sizes. In the fixed design model, we shall consider classes, given though the mapping $\varphi$, that are balanced in the following sense. Let $\Sigma_e$ denote the set of maps $\sigma\in [k]^n$ such that, for some constants $c_1,c_2$, for any $1\le j\le k$,
\[ c_1\frac{n}{k} \le |\sigma^{-1}(j)| \le c_2\frac{n}{k}.\]
The set $\Sigma_e$ thus consists of those maps $\sigma$ that produce $k$ classes all of size of order $n/k$.
Now consider the model defined by, for $e_k, M^\theta$ as in \eqref{propk}-\eqref{zte},
\begin{equation} \label{modk}
\EM{\mathcal{M}}_k = \left\{ P_\theta:= P_{e_k,M^\theta},\quad \theta\in[-1/2,1/2] \right\}.
\end{equation}
The set $\EM{\mathcal{M}}_k$ is a $1$--dimensional submodel through the set of all SBMs with at most $k$ classes. As before for $\theta=0$ the matrix $M^0$ has two identical rows, and the model becomes a SBM with at most $k-1$ classes, with connectivity matrix given by $M$ defined above. Even though the model is defined through only one unknown connectivity-type parameter $\theta$, the (uniform) rate of estimation turns out to be fairly slow, as Theorem \ref{thm-kcl} below shows.
\subsection{Lower bound result}
As before let us denote by $E_\theta$ the expectation under $P_\theta$ in the model $\EM{\mathcal{M}}_k$ given by \eqref{modk}.
\begin{thm} \label{thm-kcl}
Consider a stochastic blockmodel \eqref{modsbm} with $k\ge 2$ classes specified by $\EM{\mathcal{M}}_k$ in \eqref{modk}, that is $P_\theta=P_{e_k,M^\theta}$ with $e_k,M^\theta$ given by \eqref{propk}-\eqref{zte}, for fixed matrices $A, B$ with arbitrary coefficients.
There exists a constant $c_3=c_3(\rho)>0$, independent of $A,B$, such that, for all $n\ge 12k$,
\[ \inf_{T} \sup_{\theta\in[-1/2,1/2]} E_{\theta} \left[ T(X) - \theta \right]^2
\ge c_3\frac{k}{n}, \]
where the infimum is taken over all estimators $T$ of $\theta$ in the model $\EM{\mathcal{M}}_k$.
\end{thm}
Theorem \ref{thm-kcl} establishes that the minimax estimation rate of $\theta$ in the model \eqref{modk} is at best of the order $k/n$, uniformly over $k$ and $n$. An intuitive explanation for this particularly slow rate is as follows: the phenomenon observed for $k=2$ is still present but this time the part of the matrix $Z^\theta$ containing information about $\theta$ is smaller, as only of the order $2/k$ of the nodes will be assigned to classes $1$ or $2$, which are the elements of the connectivity matrix that depend on $\theta$.
An important point is that this lower bound is minimax {\em local} (as opposed to more commonly proved minimax global results) that is, not only does this slow rate occur around one specific least-favorable point in the parameter space, it does occur around {\em any} point. More precisely, for any $k\ge 2$ and the proportions vector being fixed, whatever connectivity matrix $M$ with $k-1$ classes as in \eqref{coM} we start from, there exists at least one submodel around $M$, namely $\EM{\mathcal{M}}_k$ in \eqref{modk}, such that estimation of a connectivity parameter in $M$ cannot be faster than $k/n$. One could have wondered from Theorem \ref{thm-twocl} if the slow rate there was not due to the fact that the model for $\theta=0$ was quite special, namely of Erd\"os-Reyni type. This is not the case. Proving such a local bound makes the proof of Theorem \ref{thm-kcl} more involved in the random design case, as one has to quantify the $L^1$-distance between two mixtures of probability measures, instead of between one fixed measure and a mixture as is often the case in proving minimax global bounds.
It is interesting to compare the rate in Theorem \ref{thm-kcl} to the one that would be obtained if the labels were observed. If $k$ is fixed, Lemma 2 in \citet{bickel13}
gives a quadratic rate of order $1/n^2$ for connectivity parameters when labels are observed. This result can be easily adapted to the case where $k$ possibly grows with $n$, say in an asymptotic setting with $n\to\infty$ and $k/n\to 0$, leading to a quadratic rate of order $(k/n)^2$. The rate in Theorem \ref{thm-kcl} is the square-root of this rate and thus much slower.
\subsection{Upper bound results}
We shall first focus on a computable estimator, that generalises the simple estimator $\tilde\theta$ for $k=2$ in \eqref{te_spec}, and that is obtained following recent ideas for spectral methods introduced by \citet{Lei:Rinaldo:2015:1} and \citet{Lei:Zhu:2017:1}. For the sake of comparison, we then consider a $k$-classes counterpart of the maximum likelihood-type estimator \eqref{pseml}. This will be done under some fairly mild assumptions of the matrix $M$. These conditions are for simplicity of presentation and could, in some cases, be improved. Our main purpose here is to show that, for `typical' matrices $A$ and $B$ in \eqref{zte}, the rate of estimation of $\theta$ in \eqref{zte} is indeed exactly of the order $k/n$. In section \ref{subsec:nec} below, we show that at least {\em some} conditions on possible matrices $A,B$ are necessary: for certain unfavourable matrices, the rate can drop below $k/n$.
\subsubsection{Estimation via a spectral algorithm} \label{sec-spec}
In the frame below, we define an algorithm {\tt Spec-$\theta$} that builds upon the spectral clustering method of \citet{Lei:Rinaldo:2015:1}, and on its refinement {\tt V-Clust} recently introduced by \citet{Lei:Zhu:2017:1}. The latter is based on a sample splitting idea, which under appropriate conditions on the connectivity matrix enables one to recover the labels {\em exactly}, with high probability.
Recall the assumed form of the connectivity matrix $M^\theta$ in \eqref{zte}. The conditions of the next results are in terms of an `aggregated' $(k-1)\times (k-1)$ matrix $N$ obtained from $M^\theta$ by merging the first and second row/columns when $\theta=0$, that is
\[ N = \begin{bmatrix}
1/2 & a_1 & \cdots & a_{k-2} \\
a_1 & b_{11} & \cdots & b_{1 k-2} \\
\vdots & \vdots & & \vdots \\
a_{k-2} & b_{1 k-2} & \cdots & b_{k-2 k-2}
\end{bmatrix}.\]
Recall that $\varphi$ denotes the true labelling map.
Define a labelling $\psi:[n]\to[k-1]$ by $\psi(v)=1$ if $\varphi(v)\in\{1,2\}$ and
$\psi(v)=\varphi(v)-1$ if $v\in \{3,\ldots,k\}$. That is, we `aggregate' nodes of label $1$ or $2$ in one class and renumber the remaining labels so that the label set is, now, $[k-1]$. For easy reference in the next frame and proof of Theorem \ref{thmspec}, we adopt the notation of \cite{Lei:Zhu:2017:1}: we set $g_v=\psi(v)$ for the true (aggregated) label of node $v\in[n]$ and $\EM{\mathcal{I}}^{(l)}=\{v\in[n]:\ g_v=l\}$.
The algorithm {\tt Spec-$\theta$} specified in the frame below has three steps. First, one runs the exact label recovery algorithm {\tt V-Clust} of \citet{Lei:Zhu:2017:1} for $K=k-1$ classes.
Under some conditions on the matrix $N$, see (A1)--(A2) below, this finds the `aggregate' labels $\psi$ above {\em up to label permutation} with high probability. Then the aim is to recover the aggregated class with original labels $1$ and $2$. Due to the label switching issue, this requires some extra condition on $N$. For simplicity (see also comments below) we assume in (A3) that the diagonal terms $b_{ii}$ are separated from $1/2$, which enables to estimate the aggregated class label $1$ by comparing diagonal empirical connectivities to $1/2$. Finally, in a third step one can run the spectral algorithm $\EM{\mathcal{S}}_2$ from Section \ref{sec2} on the nodes found at the previous step.
\vspace{.5cm}
\fbox{\begin{minipage}{0.9\textwidth}
\begin{center}
{\tt Algorithm: Spectral method for estimation of $\theta$ (Spec-$\theta$)}
\end{center}
{\bf Input:} adjacency matrix $X$ (where we set $X_{ii}=0$), number of classes $k$\\
{\bf Subroutines:} {\tt V-Clust} (Lei-Zhu), Initial community recovery $\EM{\mathcal{S}}$ (Lei-Rinaldo), Spectral algorithm $\EM{\mathcal{S}}_2$ for $k=2$ (Section \ref{sec2})
\begin{enumerate}
\item Apply {\tt V-Clust} on adjacency matrix $X$ using $k-1$ classes, $\EM{\mathcal{S}}$ and $V=2$
\[ \hat g = \text{{\tt V-Clust}}(X,k-1,V,\EM{\mathcal{S}}). \]
\item Set
$\hat\EM{\mathcal{I}}^{(1)}=\{v\in[n]:\ \hat g_v=\hat \ell\,\}$, where
\[ \hat \ell \, = \,\underset{l\in[k-1]}{\text{argmin}}\
\Bigg|\,
\frac{1}{{{\hat{g}^{-1}(l)|} \choose {2}}}\sum_{i<j,\, i,j\in \hat{g}^{-1}(l)} X_{ij}\,
-\,\frac12
\,\Bigg|
\]
\item Run spectral algorithm $\EM{\mathcal{S}}_2$ for $k=2$ on corresponding nodes and set
\[ \hat\theta = \EM{\mathcal{S}}_2(X^{\hat \EM{\mathcal{I}}^{(1)}}),\]
where $X^{\hat \EM{\mathcal{I}}^{(1)}}$ is the induced adjacency matrix over nodes in $\hat \EM{\mathcal{I}}^{(1)}$.
\end{enumerate}
\end{minipage}}
\vspace{.5cm}
We set $K=k-1$ and make the following assumptions, where $C$ is a large enough universal constant,
\begin{enumerate}
\item[(A1)] $N$ is full rank
and any two rows of $N$ are separated by at least $\gamma=\gamma(K)>0$ in $\ell_2$-norm.
\item[(A2)] For $\lambda=\lambda(K)$ the smallest absolute eigenvalue of $N$,
\begin{equation*}
n \lambda(K) \gamma(K)\ge CK^{4.5},\quad n\gamma(K)^2\ge CK^3\log{n},\quad
n\ge CK^3.
\end{equation*}
\item[(A3)] For all $i\in\{1,\ldots,k-2\}$,
\[ |b_{ii}-1/2|\ge \kappa, \]
where $\kappa=\kappa(K)\ge C\sqrt{K(\log{n})/n}$.
\end{enumerate}
Comments on (A1)--(A3) follow below. For a version for sparse graphs, see Section \ref{sec_spa}.
\begin{thm} \label{thmspec}
In the fixed design SBM model with $k$ classes, under the assumptions (A1)--(A3),
let us set, for $c$ a small enough universal constant and $K=k-1$,
\begin{equation} \label{def_tk}
T_K := c \frac{\lambda(K) \gamma(K)^{1/2}}{K^{5/4}} \wedge \frac{\kappa}{4}.
\end{equation}
Then the obtained $\hat \theta$ from algorithm {\tt Spec-$\theta$} satisfies, for $C_3$ a large enough constant,
\[ \sup_{|\theta|\le T_K,\, \varphi\in\Sigma_e} E_{\theta,\varphi}\left[ \hat\theta-\theta\right]^2 \le C_3\frac{k}{n}. \]
\end{thm}
The algorithm {\tt Spec-$\theta$}, unlike the likelihood method considered below, only uses the fact that the connectivity matrix is of the form $M^\theta$, but does not use specific knowledge of the vector $a$ and matrix $B$ to compute $\theta$.
Conditions (A1) and (A2) are typical for spectral methods; their specific form is that assumed by \citet{Lei:Zhu:2017:1}, with the initial recovery algorithm being that of \citet{Lei:Rinaldo:2015:1}. If $K$ is fixed independently of $n$, then (A2) follows from (A1) if $n$ is large enough.
Condition (A3) is specific to our problem, and assumed in this form only for simplicity of exposition: To identify the special cluster arising from the $1/2$ coefficient in the matrix $N$ (step 2. in {\tt Spec-$\theta$}), some identifiability condition is needed, because even the refined spectral clustering algorithm of \cite{Lei:Zhu:2017:1} can only recover the original labels up to a permutation. Condition (A3) can be replaced with any other condition that ensures cluster $1$ can be identified from a noisy, permuted version of $N$ (with noise amplitude going to zero fast, as $k/n$). Note that, if $k$ is fixed and $n$ large enough, (A3) simply requires the diagonal terms of $B$ to differ from $1/2$.
Finally, a comment on $T_K$ in \eqref{def_tk}. The label recovery in Steps 1--2 is run with $k-1$ classes, and hence joins two of the $k$ classes in the sample. The restriction on the range of $\theta$ ensures the classes joined are the first two, with high probability.
Indeed, here we are interested in the situation where $\theta$ may be small, which makes identification of labels difficult, and the rate slow; if $\theta$ is large, the problem becomes easier. Again, note that if $k$ is fixed, the condition simply requires that $|\theta|$ is smaller than a given constant.
\subsubsection{Estimation via a maximum likelihood approach} \label{subsec-mle}
For a given subset $S\subset\{1,2,\ldots,n\}$, define
\begin{equation} \label{znew}
2Z_n(\sigma,S,X)=- \sum_{i<j,\ i,j\in S,\, \sigma(i)=\sigma(j)} (1-2X_{ij}) +
\sum_{i<j,\ i,j\in S,\, \sigma(i)\neq \sigma(j)} (1-2X_{ij}).
\end{equation}
This criterion function is similar to the one in \eqref{zedns}, but restricts to a subset $S$ of indices.
We can simplify the proof of Theorem \ref{ubk} below, without loss of generality, by maximizing over a grid.
Let $\Theta_n=\{i/(2n^2),\ i= -n^2,\ldots,n^2\}$ be a regular grid within $\Theta=[-1/2,1/2]$, and define
\begin{align}
(\tilde\sigma,\tilde \theta)\ & :=\ \underset{\sigma\in\Sigma_e,\, \theta\in\Theta_n}{\text{argmin }}
\ \sum_{i<j} (X_{ij} - Z^\theta_{\sigma(i)\sigma(j)})^2 \label{ub-pre} \\
\tilde{S}_I\ & := \ \tilde\sigma^{-1}(\{1,2\}). \label{sun-hat}
\end{align}
Equation \eqref{ub-pre} defines a global maximum-likelihood type estimator, which is then used to obtain an estimate $\tilde S_I$ of the set of nodes labelled $1$ or $2$. Given this estimate, one can apply the profile-type method already used in the case ${k=2}$:
For $\tilde S_I$ as in \eqref{sun-hat}, $\tilde n_k= {|\tilde S_I| \choose 2}$, and $Z_n$ as in \eqref{znew}, set
\begin{align}
\hat \sigma_I & = \underset{\sigma\in\Sigma_e}{\text{argmax }} |Z_n(\sigma,\tilde S_I,X)| \label{sigi} \\
\hat \theta & = \frac{Z_n(\hat\sigma_I, \tilde S_I,X)}{\tilde n_k}. \label{thgen}
\end{align}
We require the coefficient $a_0$ of the matrix $M$ in \eqref{coM} to be sufficiently distinct from the remaining entries: Let $\EM{\mathcal{C}}=\{a_{i}, b_{ij}, 1\le i,j\le k-2\}$ be the set of coefficients of the matrices $A$ and $B$ in \eqref{zte}, with $a_0=1/2$,
\begin{equation} \label{techm}
\min_{c\in\EM{\mathcal{C}}} \left\{ |c-a_0| \right\} \ge 2\kappa >0.
\end{equation}
\vspace{-.1cm}
\begin{thm} \label{ubk}
Consider a stochastic blockmodel \eqref{modsbm} with $k\ge 2$ classes specified by $P_\theta=P_{e_k,M^\theta}$ with $e_k,M^\theta$ given by \eqref{propk}-\eqref{zte}, for fixed matrices $A, B$. Define $\hat\theta=\hat\theta(X)$ as in \eqref{thgen}. Suppose \eqref{techm} holds and that, for some small enough $d$ and $\kappa$ as in \eqref{techm},
\begin{equation} \label{techc}
k^3 \log{k}\le d \kappa^4 n.
\end{equation}
Then there exists a universal constant $C_1>0$ such that for $n\ge 5$,
\[ \sup_{|\theta|\le \kappa} E_{\theta}\left[\hat\theta - \theta\right]^2
\le C_1\frac{k}{n}.\]
The same risk bound holds for $\hat\theta$ in the fixed design model, uniformly over $|\theta|\le \kappa, \varphi\in \Sigma_e$.
\end{thm}
Note $\kappa$ in \eqref{techc} may depend on $k$ and $n$, and may go to zero in a framework where $k,n$ go to infinity. Two examples for the behaviour of $\kappa$ are given below. Condition \eqref{techm} is similar in spirit to Condition (A3) for Theorem \ref{thmspec} but stronger. Similar comments can be made, and again the specific assumed form of the condition is for simplicity and could be improved.
On the other hand, for Theorem \ref{ubk} no other condition on $A, B$ in \eqref{zte} are required, unlike for the spectral algorithm, that requires (A1)--(A2).\\
\noindent {\em Example 1 (well-separated block).} If $\kappa$ is a fixed positive constant e.g. $1/4$, then the submatrix $Q^\theta$ is well separated from the other coefficients of the matrix $M$. The procedure above then correctly picks up a sensible approximation of the true set $\sigma^{-1}(\{1,2\})$ via $\tilde S_I$ and the rate $k/n$ is achieved, as long as $k$ does not grow faster than $n^{1/3}/\log{n}$, an already fairly important number of classes.
\\
\noindent {\em Example 2 (randomly sampled matrix $M$).} Suppose that the symmetric matrix ${M=:(c_{ij})}$ in \eqref{coM} is a random matrix whose upper triangular entries are drawn i.i.d. with uniform distribution $\EM{\mathcal{U}}[0,1]$, except $c_{11}=1/2$. The distribution of $|c_{ij}-1/2|$ except for $i=j=1$ is then $\EM{\mathcal{U}}[0,1/2]$, and it is a standard fact that the first order statistic of an uniformly distributed sample of size $N$ is Beta$(1,N)$ distributed. That implies the random variable $2\min_{c_{ij}\in\EM{\mathcal{C}}} |c_{ij}-1/2|$ has law $\text{Beta}(1,k(k-1)/2-1)$. Therefore, $\kappa$ in \eqref{techm} is of order no less than $1/k^2$ with high probability. From \eqref{techc} one deduces that for $k$ of the form $n^\delta$ with $\delta<1/11$ and $n$ large enough, the rate $k/n$ is achieved uniformly and locally, for typical matrices $M$.
Inspection of the proof of Theorem \ref{ubk} reveals that $k=o(n^\delta)$ with $\delta<1/7$ in fact suffices for the rate $k/n$ to be attained with high probability when $M$ is random: this is achieved by distinguishing $c_{ij}$ of the types $a_i$ or $b_{ij}$ in the proof and noting that the minimum of $|a_{i}-1/2|$ over $i$ will be of larger order $k^{-1}$, instead of $k^{-2}$ for the minimum over $i,j$ of $|b_{ij}-1/2|$.\\
\subsection{Necessity of conditions on $M$} \label{subsec:nec}
What precedes shows that the rate $k/n$ is achieved under conditions on $M$ in \eqref{coM} and/or $k$. In general, we expect the rate to depend on the matrices $M$. Although we do not investigate this point in full here, we discuss it briefly.
The estimation methods investigated in Sections \ref{sec-spec} and \ref{subsec-mle} require the upper-left ${2\times 2}$ block of $M^\theta$ to be sufficiently separated from at least part of the other entries of $M^\theta$. Among those matrices $M^\theta$ whose upper-left corner equals $Q^\theta$, a worst case scenario should correspond to a matrix whose coefficients in $A$ and $B$ all equal $1/2$. This leads to the matrix
\begin{equation} \label{special}
\check M^\theta
= \begin{bmatrix}
\frac12 + \theta & \frac12 - \theta & \frac12 & \cdots & \frac12\\
\frac12 - \theta & \frac12 + \theta & \frac12 & \cdots & \frac12\\
\frac12 & \frac12 & \frac12 & \cdots & \frac12 \\
\vdots & \vdots & \vdots & & \vdots \\
\frac12 & \frac12 & \frac12 & \cdots & \frac12
\end{bmatrix},
\end{equation}
which is of course heavily over-specified from the SBM perspective.
Consider the SBM in a fixed design case, where $\varphi:\{1,\ldots,n\}\to \{1,\ldots,k \}$ is unobserved. Suppose all classes $\sigma^{-1}(i)$ are of cardinality of order $n/k$, and the connectivity matrix is given by \eqref{special}. This specific model can be regarded as a special case of the setting considered from a testing perspective by \citet{butuceaingster} and \citet{ariascv14}. From Theorem 4.3 of \cite{butuceaingster}, one can deduce that the minimax rate for the quadratic risk when estimating $\theta$ is no better than $\rho_n=\min\left(\frac{k^2}{n},\sqrt{\frac{k\log k}{n}}\right)$, for $k,n\to \infty$ and $\rho_n=o(1)$. The rate is therefore no better than $k^2/n$ for $k\le n^{1/3}$, and remains much slower than $k/n$ even for $k>n^{1/3}$.
\subsection{Extension to sparse graphs} \label{sec_spa}
So far for simplicity we have considered {\em dense} graphs in the sense that at least some elements of the connectivity matrix (e.g. $1/2+\theta$ or $1/2-\theta$) are bounded away from zero.
An $\alpha_n$--sparse SBM model is generally defined as one in which the connectivity matrix $M$ can be written, for $\alpha_n$ a sequence going to $0$ with $n$, as $M = \alpha_n M_0$,
for $M_0$ a nonnegative symmetric matrix with maximum entry $1$ \citep[e.g.][]{bickeletal11,Lei:Rinaldo:2015:1}. Here, we assume that the connectivity matrix is $M^\theta(\alpha_n)$ with
\begin{equation} \label{zte_spa}
M^\theta(\alpha_n) = \alpha_n M^\theta,
\end{equation}
and $M^\theta$ as in \eqref{zte}. Then the largest coefficient of $M^{\theta}$ is between $\alpha_n/2$ and $\alpha_n$, as the coefficients of the upper $2\times 2$ block are $\alpha_n(1/2\pm\theta)$. We also set, for $\theta\in[-1/2,1/2]$,
\begin{equation}
Q^\theta(\alpha_n) = \alpha_n Q^\theta, \qquad
Q^\theta \ =
\begin{bmatrix}
\frac12 + \theta & \frac12 - \theta\\
\frac12 - \theta & \frac12 + \theta
\end{bmatrix}. \label{simatspa}
\end{equation}
In constructing upper bounds below, we assume that for $C_s$ a large enough constant,
\[ (B0)\quad\qquad \alpha_n\ge C_s \frac{\log{n}}{n}, \]
as up to a constant $\log{n}/n$ is the typical boundary between the moderately sparse and very sparse situations, the later requiring different tools, see \cite{Lei:Rinaldo:2015:1}. For simplicity we also assume that $\alpha_n$ is known for the upper-bound results.
\begin{thm} \label{thm-twocl-spa}
Consider a stochastic blockmodel \eqref{modsbm} with $k=2$ specified by $P_\theta=P_{e,Q^\theta(\alpha_n)}$ with $e,Q^\theta(\alpha_n)$ given by \eqref{prop2}-\eqref{simatspa}. There exists a constant $c_1>0$ such that for all $n\ge 2$,
\[ \inf_{T} \sup_{\theta\in[-1/2,1/2]} E_{\theta} \left[ T(X) - \theta \right]^2
\ge c_1\left(1\wedge \frac{1}{n\alpha_n}\right), \]
where the infimum is taken over all estimators $T$ of $\theta$ in the model $\EM{\mathcal{M}}$.
Furthermore, if $\Delta_n=X-\alpha_nJ/2$, and $\lambda_1^a(\Delta_n)$ the largest absolute eigenvalue of $\Delta_n$, set $\tilde\theta:=\lambda_1^a(\Delta_n)/\{(n-1)\alpha_n\}$. Then, under (B0), for some constant $C>0$ and $n\ge 2$,
\[ \sup_{\theta\in[-1/2,1/2]} E_\theta[(\tilde\theta-\theta)^2] \le \frac{C}{n\alpha_n}. \]
\end{thm}
The case of $k$ classes carries over to the sparse situation as follows. The lower bound result is only modified by a scaling factor $1/\alpha_n$. For upper bounds, considering the more easily computable spectral algorithm {\tt Spec-$\theta$} only, Assumption (A2) is replaced by (B2) below, where $N$ has the same definition as in Section \ref{sec-spec}.
\begin{enumerate}
\item[(B2)] For $\lambda=\lambda(K)$ the smallest absolute eigenvalue of $N$,
there exists $C>0$ such that
\begin{equation*}
n \alpha_n\lambda(K) \gamma(K)\ge CK^{4.5},\quad n\alpha_n\gamma(K)^2\ge CK^3\log{n},\quad
n\ge CK^3.
\end{equation*}
\end{enumerate}
\begin{thm} \label{thm-kcl-spa}
Consider a stochastic blockmodel \eqref{modsbm} with $k\ge 2$ classes specified by $\EM{\mathcal{M}}_k$ in \eqref{modk}, that is $P_\theta=P_{e_k,M^\theta}$ with $e_k,M^\theta$ given by \eqref{propk}--\eqref{zte_spa}, for fixed matrices $A, B$ with arbitrary coefficients.
There exists a constant $c_3=c_3(\rho)>0$, independent of $A,B$, such that, for all $n\ge 12k$,
\[ \inf_{T} \sup_{\theta\in[-1/2,1/2]} E_{\theta} \left[ T(X) - \theta \right]^2
\ge c_3\left(1 \wedge \frac{k}{n\alpha_n}\right), \]
where the infimum is taken over all estimators $T$ of $\theta$ in the model $\EM{\mathcal{M}}_k$.
Let $\EM{\mathcal{S}}_{2,\alpha_n}$ be the algorithm for $k=2$ classes in the sparse case described in Theorem \ref{thm-twocl-spa}. Consider the fixed-design setting and suppose $(B0), (B2), (A1)$ and $(A3)$ are satisfied. Then the algorithm {\tt Spec}-$\theta$ used with subroutine $\EM{\mathcal{S}}_{2,\alpha_n}$ outputs an estimator $\hat\theta$ that satisfies, for $T_K$ as in \eqref{def_tk},
\[ \sup_{|\theta|\le T_K,\, \varphi\in \Sigma_e} E_{\theta,\varphi}\left(\hat\theta-\theta\right)^2
\le C \frac{k}{n\alpha_n}.\]
\end{thm}
Similar comments as for Theorems \ref{thm-kcl}--\ref{thmspec} can be made. Also, in the case that $k$ does not grow with $n$, then (B2) follows from (B0) for $n$ larger than a fixed constant. The proof of the lower bound in Theorem \ref{thm-kcl-spa} is similar to that of Theorem \ref{thm-kcl} using the normalisation as in the proof of Theorem \ref{thm-twocl-spa} and is omitted. The upper bound result includes that of Theorem \ref{thmspec} and is proved in Appendix \ref{sec:ub_spec}.
\subsection{Simulation study}
\label{sec:simulation}
\begin{figure}
\makebox[\textwidth][c]{
\begin{tikzpicture}
\begin{scope}[scale=3.5]
\draw (0,0)--(.4,0)--(.4,-.4)--(0,-.4)--cycle;
\draw (.2,0)--(.2,-.4);
\draw (0,-.2)--(.4,-.2);
\node at (.1,-.1) {\scriptsize $\frac{1}{2}\!+\!\theta$};
\node at (.3,-.1) {\scriptsize $\frac{1}{2}\!-\!\theta$};
\node at (.3,-.3) {\scriptsize $\frac{1}{2}\!+\!\theta$};
\node at (.1,-.3) {\scriptsize $\frac{1}{2}\!-\!\theta$};
\end{scope}
\begin{scope}[xshift=7cm,yshift=-.7cm]
\node at (0,0) {\includegraphics[height=3.2cm]{risk_2classes_spectral_M1000.png}};
\node at (4,0) {\includegraphics[height=1.36cm]{legend.png}};
\end{scope}
\end{tikzpicture}
}
\caption{Estimation of $\theta$ in the two-class case, using the spectral estimator \eqref{te_spec}.
Graphs of size ${n=50,100,500,1000,1500}$ are generated from the graphon
on the right, for ${\theta=0.01,0.025,0.05,0.1}$. Shown is the empirical risk (computed over
1000 experiments) as a function of the sample size $n$.}
\label{fig:k:2}
\end{figure}
Computationally feasible estimators---the spectral estimator
for ${k=2}$, and the Spec-$\theta$ estimator for ${k>2}$---can be tested in simulation: Draw $n$ vertices from
a stochastic block model as in \eqref{modk} with a given value of $\theta$, compute the respective estimate, and report the
empirical quadratic risk. Figure \ref{fig:k:2} shows how the risk develops as a function of sample size
for different values of $\theta$, for the two-community model \eqref{simat}.
For $k>2$ communities, the model is given by the connectivity matrix \eqref{zte}.
Simulation results for $k=5$, with ${a_1=\frac{1}{12}}$, ${a_2=\frac{11}{12}}$ and ${a_3=1}$, are shown in
Figure \ref{fig:k:5}.
Small values of $\theta$ give larger risk, while for larger values of $\theta$, one falls into the `fast rate' regime, where the rate becomes essentially as fast as if labels were observed-- the empirical risk then behaves as
$\hat{R}_n\approx C/n^2$.
\section{Results for a class of functionals of smooth graphons}\label{sec4}
Stochastic block models can be identified with piece-wise constant graphons; we now consider the case where
the graphon is a smooth function instead.
Let $w:[0,1]^2\to [0,1]$
a measurable function, let $\left< w\right>$ be its graphon equivalence class, and denote by
${P_{\left< w\right>}=P_w}$ the distribution of data $X$ generated
by the graphon model \eqref{modgr}.
Consider the problem of estimating the functional
\begin{equation}\label{sig}
\tau(\left< w\right>)=\left[\int_{[0,1]^2} \Big(w(x,y)-\int_{[0,1]^2}w\Big)^2 dxdy\right]^{1/2},
\end{equation}
for any representer $w$ of $\left< w\right>$.
This is well defined in terms of the graphon, as the integral is invariant under any simultaneous (Lebesgue-)measure-preserving transformation of $x$ and $y$.
The statistic \eqref{sig} can be interpreted as a `graphon-standard deviation'.
Its estimation under a smooth graphon model is, in a sense, analogous to the problem of estimating the functional $\theta$ in the simple SBM with two classes discussed in Section \ref{sec2}: Let $w_\theta$ be the piece-wise constant graphon characterizing the SBM defined by \eqref{prop2}-\eqref{simat}. Since $\tau({\langle} w_\theta {\rangle})=|\theta|$, estimating $\theta$ is then indeed equivalent to estimation of $\tau(\left< w\right>)$ (if one considers only positive values of $\theta$).
Under a 2-class SBM, the results of Section \ref{sec2} show $\theta$ in \eqref{simat} cannot be estimated faster than $c/n$. It is natural to ask whether the same still holds if one works with `smoother' graphons instead of histograms (where we refer to ${\langle} w {\rangle}$ as smooth if at least one of its representers is a smooth function). The following result answers this question, both for $\tau$ and for a larger class of functionals containing $\tau$.
Let $\EM{\mathcal{P}}_B$ be the collection of all graphons that admit a representer which is a polynomial in $x, y$, with degree bounded by some integer $D\ge 2$ and coefficients bounded by an arbitrary constant $M>0$ (this boundedness restriction is only to ensure a --nearly, up to a log term-- matching upper-bound in the next result). For any $0\le \theta\le 1$, let us denote by $w_\theta$ the function from $[0,1]^2$ to $[0,1]$ given by
\begin{equation} \label{polyw}
w_\theta(x,y)=\frac12-\theta(x-\frac12)(y-\frac12)
\end{equation}
and let $w_0$ denote the constant function equal to $1/2$.
\begin{thm} \label{thm-quad}
Let $X$ be data from the graphon model \eqref{modgr}.
There exist constants $c_1,c_2>0$ such that, for $\tau(\cdot)$ defined by \eqref{sig},
\[ \frac{c_1}{n}\le \inf_{\hat \tau} \sup_{w \in \EM{\mathcal{P}}_B} E_{P_{w}}\left[\hat{\tau}(X) - \tau(\left< w\right>)\right]^2 \le \frac{c_2\log{n}}{n}, \]
where the infimum is taken over all possible estimators of $\tau({\langle} w{\rangle})$ in model \eqref{modgr}.
Let $\psi$ be an arbitrary functional defined on graphon equivalence classes such that for some $c>0$ and $w_\theta$ the function in \eqref{polyw},
\[ |\psi({\langle} w_\theta {\rangle})-\psi({\langle} w_0{\rangle})| \ge c|\theta|. \] Then for some $d>0$,
\[ \inf_{\hat \psi} \sup_{w \in \EM{\mathcal{P}}_B} E_{P_{w}}\left[\hat{\psi}(X) - \psi(\left< w\right>)\right]^2 \ge \frac{d}{n}.\]
\end{thm}
Thus, the minimax rate cannot be faster than $c/n$, even if one estimates a functional as elementary as \eqref{sig}, and restricts the model to a simple, small class of smooth graphons $w$---here, graphons with a polynomial representer. The functional
\[ \psi(P_w) = \int_{[0,1]^2} \Big|f(x,y) - \int_{[0,1]^2} |f(x,y)|dxdy \Big|dxdy\]
is another example satisfying the assumptions of Theorem \ref{thm-quad}.
\begin{figure}
\makebox[\textwidth][c]{
\begin{tikzpicture}
\begin{scope}[scale=0.8, every node/.style={scale=0.8}]
\begin{scope}[scale=3.5]
\draw (0,0)--(1,0)--(1,-1)--(0,-1)--cycle;
\draw (.2,0)--(.2,-1);
\draw (.4,0)--(.4,-1);
\draw (.6,0)--(.6,-1);
\draw (.8,0)--(.8,-1);
\draw (0,-.2)--(1,-.2);
\draw (0,-.4)--(1,-.4);
\draw (0,-.6)--(1,-.6);
\draw (0,-.8)--(1,-.8);
\node at (.1,-.1) {\scriptsize $\frac{1}{2}\!+\!\theta$};
\node at (.3,-.1) {\scriptsize $\frac{1}{2}\!-\!\theta$};
\node at (.3,-.3) {\scriptsize $\frac{1}{2}\!+\!\theta$};
\node at (.1,-.3) {\scriptsize $\frac{1}{2}\!-\!\theta$};
\node at (.1,-.5) {\scriptsize $\frac{1}{12}$};
\node at (.3,-.5) {\scriptsize $\frac{1}{12}$};
\node at (.5,-.5) {\scriptsize $\frac{1}{12}$};
\node at (.5,-.3) {\scriptsize $\frac{1}{12}$};
\node at (.5,-.1) {\scriptsize $\frac{1}{12}$};
\node at (.7,-.1) {\scriptsize $\frac{11}{12}$};
\node at (.7,-.3) {\scriptsize $\frac{11}{12}$};
\node at (.7,-.5) {\scriptsize $\frac{11}{12}$};
\node at (.7,-.7) {\scriptsize $\frac{11}{12}$};
\node at (.5,-.7) {\scriptsize $\frac{11}{12}$};
\node at (.3,-.7) {\scriptsize $\frac{11}{12}$};
\node at (.1,-.7) {\scriptsize $\frac{11}{12}$};
\node at (.9,-.1) {\scriptsize $1$};
\node at (.9,-.3) {\scriptsize $1$};
\node at (.9,-.5) {\scriptsize $1$};
\node at (.9,-.7) {\scriptsize $1$};
\node at (.9,-.9) {\scriptsize $1$};
\node at (.7,-.9) {\scriptsize $1$};
\node at (.5,-.9) {\scriptsize $1$};
\node at (.3,-.9) {\scriptsize $1$};
\node at (.1,-.9) {\scriptsize $1$};
\end{scope}
\end{scope}
\begin{scope}[xshift=8.5cm,yshift=-1.4cm]
\node at (0,0) {\includegraphics[height=3.2cm]{risk_5classes_spectheta_M1000.png}};
\end{scope}
\end{tikzpicture}
}
\caption{Estimation of $\theta$ in the five-class case, using the Spec-$\theta$ algorithm.
Graphs of size ${n=50,100,500,1000,1500}$ are generated from the graphon
on the right, for ${\theta=0.01,0.025,0.05,0.1}$. Shown is the empirical risk (computed over
1000 experiments) as a function of the sample size $n$. For small values of $\theta$, convergence
slows visibly.}
\label{fig:k:5}
\end{figure}
\section{Discussion}
\citet{chaogr15} show that, if one estimates the parameter function $w$ of a graphon model, not observing the vertex labels---in this case, the variables $U_i$ in
\eqref{modgr}---does (in general) impact on the optimal rate.
In the present paper, we have considered uniform estimation of certain functionals of graphon models (in particular, the loss function is quite different from theirs).
For estimation of certain graph functionals---including the connectivity parameters considered by \citet{bickel13}---we have shown that
the uniform, minimax rate does depend on whether the labels are observed, i.e.\ the phenomenon described by
\cite{chaogr15} persists even if one does not try to recover the entire function $w$. The fast quadratic rate $1/n^2$ is not achievable uniformly. If the number $k$ of classes is known and fixed, the quadratic rate becomes $1/n$. If the number of classes $k$ grows with $n$, the rate drops to $k/n$. We have used some mild assumptions on the part of the connectivity matrix other than the ``difficult'' ${2\times 2}$ submodel. If those assumptions are not satisfied, the rate may even drop further. Similar results also hold for sparse graphs.
Interestingly, for the functionals considered here, the uniform rate is always, regardless of the number of classes $k$, much below the rate in the case where labels would be observed. This is in contrast with the problem of recovery of the mean adjacency matrix considered in \cite{chaogr15}, where for $k$ is larger than $\sqrt{n\log{n}}$, the (non--normalised) rate $k^2+n\log{k}$ is dominated by the `parametric' rate $k^2$, the rate if labels are observed.
Aspects of our proofs reflect the fact that graphon models constitute a specific type of mixture model,
and estimation in mixtures can be difficult if mixture components are hard to distinguish; although no general theory of these phenomena seems to exist, we refer to the early work on estimation in finite mixture models by \cite{Hartigan83} and \cite{bickelchernoff}, and e.g.$\,$ to \cite{kahn} and \cite{gmc16} for more recent results.
|
\section{Introduction}
Anticipated synchronization (AS), as proposed originally by H. Voss~\cite{Voss00,Voss01a}, is a particular kind of lag synchronization that can occur in two unidirectionally-coupled dynamical systems (Sender-Receiver) when the receiver is subject to a self-inhibitory feedback loop.
In the counterintuitive AS regime, the receiver system can predict the future dynamics of the sender for certain parameter values.
Anticipated synchronization has been found both experimentally and numerically in different fields including optics~\cite{Masoller01b,Liu02,Tang03}, electronic circuits~\cite{Voss02}, neuronal systems~\cite{Ciszak03,Toral03b,Matias11,Pyragiene13,Matias14,Matias15,Matias16} and more~\cite{Kostur05,Pyragas08,Ciszak15,Voss16}.
In neuronal systems, AS was originally studied numerically by Ciszak and coworkers~\cite{Ciszak03,Toral03b} using a Fitzhugh-Nagumo model with diffusive coupling.
Chemical synapses in a three-neuron sender-receiver-interneuron (SRI) motif were included by Matias et al.~\cite{Matias11}.
Using the Hodgkin-Huxley model, a transition from the more intuitive delayed-synchronization (DS) regime to the AS regime was found when changing the inhibitory conductance impinging on the receiver neuron.
Recently, the ideas introduced in~\cite{Matias11} were extended to neuronal populations ~\cite{Matias14} to explain the observations of a positive Granger Causality, with well-defined directional influence, accompanied by either a positive or negative phase lag in the recordings of the motor cortex activity of monkeys while doing a visual task ~\cite{Brovelli04,Salazar12}.
Despite the interest attracted by AS in neuronal circuits, a thorough understanding of how this particular state is stablished is missing.
In this paper we use the well-known phase resetting curve (PRC) approach to gain insight into the AS regime and in particular into the DS/AS transition that occurs in the SRI motif of model neurons shown in Fig.~\ref{fig:intervals}a.
In its pulsatile version, PRCs describe how the spiking time of an oscillating neuron is altered by synaptic inputs.
PRCs have a long history in the analysis of coupled oscillators~\cite{PRCbook}.
More than fifty years ago, for instance, this technique was used to understand how excitatory and inhibitory pulses could decrease or increase firing rates of pacemaker neurons~\cite{Perkel64}.
Despite its generality, the technique is particularly suitable when a neuron receives one input per cycle.
As reviewed by Goel and Ermentrout~\cite{Goel02}, as well as by Canavier and Achuthan~\cite{Canavier10pulse,CanavierBookChapter}, pulsatile PRC was applied in some particular cases, namely, models of two uni- and bi-directionally coupled neurons, neurons arranged in a ring configuration, two-dimensional and all-to-all networks.
PRCs were also measured experimentally in different biological systems, from neurons~\cite{Galan05} to circadian rhythms~\cite{Czeisler86, Strogatz90light}.
We aim at comparing the predictions from the PRC technique with numerical simulations of the SRI motif whose nodes are described by Hodgkin-Huxley neuronal models with chemical synapses.
The manuscript is organized as follows.
The SRI motif and neuronal models are described in section~\ref{HH}.
In section~\ref{results} we develop the PRC map for this system and compare the results with the numerical integration of the full model.
Finally, we summarize our results in section~\ref{conclusions}.
\section{The Sender-Receiver-Interneuron motif\label{HH}}
In the SRI motif, the sender node S projects an excitatory synapse onto the receiver node R, which also receives an inhibitory projection from the node I.
Moreover, the node R projects an excitatory synapse onto node I, closing an excitatory-inhibitory loop (see Fig.~\ref{fig:intervals}a).
Each node of the motif is described by the Hodgkin-Huxley model~\cite{HH52}, which consists of four differential equations describing the evolution of the membrane potential and the currents flowing across a patch of an axonal membrane~\cite{Koch}:
\begin{eqnarray}
C_m \frac{dV}{dt} &=& \overline{G}_{Na} m^3 h (E_{Na}-V) + \overline{G}_{K} n^4 (E_{K}-V) \nonumber \\
&& + G_m (V_{rest}-V) + I_c + \sum I_{syn} \label{eq:dvdt}\\
\frac{dx}{dt} &=& \alpha_x(V)(1-x) -\beta_x(V) x \label{eq:alphaHH} \;.
\end{eqnarray}
$V$ is the membrane potential, $x\in\{h,m,n\}$ are the gating variables for sodium ($h$ and $m$) and
potassium ($n$). The capacitance of a $30\times 30\times \pi$~$\mu$m$^2$ equipotential patch of
membrane is $C_m = 9\pi$~pF~\cite{Koch}.
$E_{Na}=115$~mV, $E_{K}=-12$~mV and $V_{rest}=10.6$~mV are the reversal potentials of the
Na$^+$, K$^+$ and leakage currents, respectively.
The maximal conductances are $\overline{G}_{Na}=
1080\pi$~nS, $\overline{G}_{K} =324\pi$~nS and $G_m=2.7\pi$~nS,
respectively.
$I_{syn}$ accounts for the chemical synapses arriving from other neurons and $I_c$ accounts for an external constant current.
In the absence of synapses $I_{syn}=0$ and for $I_c=280$~pA the neuron spikes with a period equals to $T=14.68$~ms.
The voltage-dependent rate variables in the Hodgkin-Huxley model have the form:
\begin{eqnarray}%
\alpha_n(V) & = & \frac{10-V}{100(e^{(10-V)/10}-1)}, \\
\beta_n(V) & = & 0.125e^{-V/80}, \\
\alpha_m(V) & = & \frac{25-V}{10(e^{(25-V)/10}-1)}, \\
\beta_m(V) & = & 4e^{-V/18},\\
\alpha_h(V) & = & 0.07e^{-V/20}, \\
\beta_h(V) & = & \frac{1}{(e^{(30-V)/10}+1)}, \label{eq:betah}
\end{eqnarray}%
where all voltages are measured in mV and the resting potential is shifted to zero mV.
For the synapses we assumed a current-based model given by:
\begin{equation}
I_{syn}(t) = g_{syn} V_{syn} \sum_{spikes} \alpha(t-t_{spike}) \;.
\label{eq:HHcurrent}
\end{equation}
$V_{syn}$ is taken, without loss of generality, equal to $1$~mV. $g_{syn}$ represents the
maximal conductances which are different for AMPA ($g_{exc}$) and GABA$_A$ ($g_{inh}$) mediated synapses.
The internal sum is extended over all the presynaptic spikes occurring at $t_{spike}$.
The $\alpha(t)$ function, that models the postsynaptic conductance, is described by the following equation:
\begin{equation}
\label{eq:alphafunction}
\alpha(t)= \pm \frac{1}{\tau_{-}-\tau_{+}} ( \exp{(-t/\tau_{-})} - \exp{(-t/\tau_{+})} ).
\end{equation}
The positive signal accounts for excitatory synpases whereas the negative for inhibitory ones. The parameters $\tau_{-}$ and $\tau_{+}$ stand respectively for the decay and rise time of the function and determine the duration of the synaptic response.
In the simulations we fix the maximum excitatory conductance $g_{exc}=1000$~nS, $\tau_{-}=6.0$~ms and $\tau_{+}=0.1$~ms.
\section{Results \label{results}}
\subsection{Phase map}
\begin{figure}
\begin{center}
\includegraphics[width=1.1\columnwidth,clip]{Fig1}
\end{center}
\caption{
\label{fig:intervals}
a) Three coupled neurons in a SRI configuration.
Each spike of the receiver (R) is perturbed by the synaptic
current from the sender (S) and the interneuron (I),
whereas each spike of the interneuron is only perturbed
by the synaptic current from the receiver.
b) The Poincar\'e map of this configuration provides
the time differences between the three neurons
in the phase-locking regime.
}
\end{figure}
In order to apply the PRC approach to the SRI configuration shown in Fig.~\ref{fig:intervals}, one has to consider that the central neuron receives two inputs per cycle when locked in the 1:1 solution: one excitatory (from the sender) and another inhibitory (from the interneuron).
Following the approach initially developed in Ref.~\cite{MatiasTese}, we define $t_{R}[n]$ as the spiking time at the $n$-th cycle of the receiver, which we take as the reference to measure time differences.
Let $t_{S}[n]$ and $t_{I}[n]$ be, respectively, the spiking times of the sender and interneuron immediately after $t_{R}[n]$ and $T_S$, $T_R$ and $T_I$ the free-running periods of the neurons, as shown in Fig.~\ref{fig:intervals}b.
To construct the return map, we introduce the variables (depicted in Fig.~\ref{fig:intervals}b)
\begin{eqnarray}
\beta_n & \equiv &t_{S}[n] - t_{R}[n] \; , \nonumber \\
\gamma_n & \equiv &t_{R}[n+1] - t_{I}[n] \; , \label{eq:variables}
\\
\alpha_n & \equiv &t_{I}[n] - t_{R}[n] \; . \nonumber
\end{eqnarray}
From the above definitions, $\beta_n$ and $\alpha_n$ measure, respectively, the timing of the excitatory and inhibitory inputs relative to the receiver cycle.
$\gamma_n$ measures the timing of the excitatory input relative to the interneuron cycle.
The $PRC_x$ of a given neuron $x$ is defined as the difference between its free-running period and the period after a perturbation is applied (so that positive PRCs imply period shortening).
We use the synaptic funtion in Eq.~\ref{eq:alphafunction} as the appiled perturbation in such a way that $PRC_x(\delta)$ is the response due to an input $\alpha[(t-\delta)$mod$(T_x)$].
We start by analyzing the simplest case of the interneuron, whose $PRC_I$, shown in Fig.~\ref{fig:singlePRC}, depends only on the excitatory input from the receiver.
From Fig.~\ref{fig:intervals}b we can start building the return map.
The interval between two consecutive spikes of the I neuron satisfies
\begin{equation}
\label{eq:alpha}
T_{I} + (-PRC_I(\gamma_n)) = \gamma_n + \alpha_{n+1}\; .
\end{equation}
The analysis of the R neuron is more complicated, because it receives two inputs from different neurons at different times within one period.
In the most general form, therefore, $PRC_R$ depends on the two variables $\alpha$ and $\beta$ that, as can be seen in Fig.~\ref{fig:intervals}b, satisfy the condition:
\begin{equation}
T_{R} + (-PRC_R(\beta_{n},\alpha_{n})) = \alpha_{n} + \gamma_{n} \; .
\end{equation}
Isolating $\gamma_n$ we get
\begin{equation}
\label{eq:gamma}
\gamma_{n} = T_{R} - PRC_R(\beta_{n},\alpha_{n})) - \alpha_{n} \equiv \gamma_n(\beta_{n},\alpha_{n}) \; .
\end{equation}
This indicates that Eq.~\ref{eq:alpha} can be written in terms of $\alpha_n$, $\alpha_{n+1}$ and $\beta_n$.
It is usually assumed that $PRC_R(\beta_{n},\alpha_{n})$ can be decomposed as the sum of two single-variable PRCs~\cite{Canavier10pulse}.
We will show later that this approximation fails precisely in the region of parameter space where AS occurs.
The interval between the $n$-th spike of the receiver and the $(n+1)$-th spike of the sender satisfies, as shown in Fig.~\ref{fig:intervals}b,
\begin{equation}
\label{eq:beta}
\beta_n + T_S = T_R + (-PRC_R(\beta_n,\alpha_n)) + \beta_{n+1} \; .
\end{equation}
Finally, we obtain, using Eq.~\ref{eq:beta} and combining Eqs.~\ref{eq:alpha} and \ref{eq:gamma} the following two-dimensional map:
\begin{eqnarray}
\beta_{n+1} &=& \beta_{n} + PRC_R(\beta_{n},\alpha_{n}) + T_{S} - T_{R}
\; ,\label{eq:prcmapbeta} \\
\alpha_{n+1}&=& \alpha_{n} +PRC_R(\beta_{n},\alpha_{n}) - PRC_I(\gamma_{n}(\beta_n,\alpha_n)) \nonumber \\
&& + T_{I} - T_{R} \; \label{eq:prcmapalpha}.
\end{eqnarray}
Two important assumptions were made here~\cite{Netoff05}.
First, we assumed that the inputs affect only the following spike of each neuron, meaning that second order effects of the PRC are neglected~\cite{Canavier10pulse,CanavierBookChapter}.
Second, we considered that the three neurons fire once in each cycle (which we know to be true from numerical integration of the equations~\cite{Matias11}).
\begin{figure}
\begin{center}
\includegraphics[width=0.9\columnwidth,clip]{Fig2}
\end{center}
\caption{Phase Resseting Curve for the interneuron as a function of $\gamma$ (see Eq.~\ref{eq:variables}).
$\gamma^*$ is the stable fixed point solution obtained with the condition given by Eqs.~\ref{eq:fixedpoints}.
}
\label{fig:singlePRC}
\end{figure}
\subsection{Phase-locked solutions and stability}
To gain insight into the transition from anticipated to delayed synchronization (AS/DS), we look for the fixed point solutions of Eqs.~\ref{eq:prcmapbeta} and \ref{eq:prcmapalpha}.
We start with the case where the three neurons have the same periods.
\subsubsection{Identical free-running periods}
Assuming that the free-running periods of all three neurons are identical, $T_S = T_R = T_I = T$, the fixed point solutions $(\alpha^*,\beta^*)$ are given by
\begin{eqnarray}
PRC_R(\beta^{*},\alpha^{*}) &=& 0 \nonumber \\
PRC_R(\beta^{*},\alpha^{*}) - PRC_I(\gamma^{*}) &=& 0 \; , \label{eq:fixedpoints}
\end{eqnarray}
where $\gamma^* = \gamma_n(\alpha^*,\beta^*)$ as defined in Eq.~\ref{eq:gamma}.
In the phase-locking regime one therefore has $PRC_I(\gamma^{*})=0$.
The analysis of the system of equations \ref{eq:fixedpoints} can be done in two steps.
First we find the stable fixed point solution for the one dimensional $PRC_I(\gamma^*)$ (note in Fig.~\ref{fig:singlePRC} that the curve has two fixed points, the one with negative slope being the stable one~\cite{Canavier10pulse,CanavierBookChapter}%
).
Since Eq.~\ref{eq:alpha} implies $\alpha^* = T - \gamma^*$, the search of the zero of $PRC_R(\alpha^*,\beta^*)$ only requires the line with constant $\alpha^*$ to be scanned.
In Fig.~\ref{fig:PRC2D} we show $PRC_R(\beta,\alpha)$ as a function of its two arguments.
This function is obtained by numerically integrating the HH equations for the R neuron subject to one excitatory and one inhibitory inputs at different times of the R neuron period.
These two inputs are given by Eq.~\ref{eq:alphafunction}, with their appropriate parameters.
For simplicity, and without loss of generality, we keep the excitatory conductance $g_{exc}$ fixed, while we change the values of $g_{inh}$.
The points of interest in the figure are those that satisfy $PRC_R=0$.
In order to find the stable fixed point solutions of Eq.~\ref{eq:fixedpoints} we should scan the values of the PRC in a vertical line in Figs.~\ref{fig:PRC2D} corresponding to the stable $\alpha^*=T-\gamma^*$ value shown in Fig.~\ref{fig:singlePRC}.
The stable solution is the one that crosses zero with negative slope when increasing $\beta$ (filled circles in Figs.~\ref{fig:PRC2D}a,b).
The line $\beta=\alpha$ is of particular importance, corresponding to the excitatory and inhibitory inputs arriving simultaneously at the receiver.
It can be clearly seen, when comparing panels b and d in Fig.~\ref{fig:PRC2D}, that the combined effect of the two pulses is very different in the full PRC function than when we just add the effect of them independently.
It is worth noting that different values of the inhibitory conductance lead to severe changes in the $PRC_R(\beta,\alpha)$ landscape, particularly impacting the position of the fixed points relative to the pulsating period (compare Figs.~\ref{fig:PRC2D}a and b).
This change corresponds to a transition in the synchronization regime (from DS to AS).
Let us define the spike timing difference $\tau_{SR}$ between sender and receiver (in the phase-locking regime) as the difference between their closest spikes, i.e. $\tau_{SR}=t_{R}-t_{S}$.
Consequently, if $\beta^* < T/2$ (see Eqs.~\ref{eq:variables} and Fig.~\ref{fig:intervals}b), then the system is in a AS regime characterized by $\tau=-\beta^*$.
On the contrary, if $\beta^* > T/2$, the system operates in the DS regime which is characterized by $\tau_{SR}=T-\beta^*$.
\begin{widetext}
\begin{figure}[!th]%
\includegraphics[width=\linewidth]{Fig3-1200}
\caption{
{\bf (Color online) Phase Response Curve of the receiver neuron due to two inputs per cycle}.
$PRC$ is color-coded (in ms) as a function of its two variables $\alpha$ and $\beta$.
In (a) and (b) we plot the full function $PRC_R(\beta,\alpha)$, whereas in panels (c) and (d) we plot the approximation
$PRC_R(\alpha)+PRC_R(\beta)$.
In the upper (lower) panels, $g_{inh}=200$ nS ($g_{inh}=1000$ nS).
The filled circles correspond to the stable fixed point, predicting delayed synchronization in panels a, c and d, and anticipated synchronization in panel b.
The prediction of panel d is incorrect (see text for details).
}
\label{fig:PRC2D}
\end{figure}%
\end{widetext}
We now check the accuracy of the PRC prediction and compare it with the numerical simulations of the full HH model.
We show in Figs.~\ref{fig:PRC2D}a and b two examples of $PRC_R{(\beta,\alpha)}$ for two different values of the inhibitory conductance $g_{inh}$.
For the parameters of Fig.~\ref{fig:PRC2D}a (small inhibitory conductance), the stable fixed point $\beta^*$ is clearly $>T/2$ for any value of $\alpha^*$.
The system therefore operates in the DS regime, as long as a stable $\gamma^*$ (and consequently $\alpha^*$) exists.
In addition, in Fig.~\ref{fig:PRC2D}c we show the results when one employs the decomposition $PRC_R{(\beta,\alpha)} \approx PRC_R(\alpha)+PRC_R(\beta)$.
This is the usual and simplest approximation when an oscillator receives two inputs per cycle~\cite{Canavier10pulse,CanavierBookChapter}.
In this case, $PRC_R(\beta)$ and $PRC_R(\alpha)$ represent the phase-resetting curves of the receiver when it is subject to \textit{either} an excitatory \textit{or} an inhibitory input, respectively.
Note that the general qualitative results of Fig.~\ref{fig:PRC2D}c are remarkably similar to those of Fig.~\ref{fig:PRC2D}a.
Moreover, the fixed points in both figures are almost identical.
Indeed, these results predict well the stationary spiking time difference $\tau_{SR}$ directly measured in the simulations of the full Hodgkin-Huxley motif.
In Fig.~\ref{fig:tau_gi}a we plot the time difference $\tau_{SR}$ versus the inhibitory conductance $g_{inh}$.
In the numerical simulations of the full SRI motif (full circles), delayed synchronization ($\tau_{SR} > 0$, see time traces in Fig.~\ref{fig:tau_gi}b) is obtained for $g_{inh} \lesssim 800$~nS, whereas beyond this value an anticipated synchronization regime takes over ($\tau_{SR} < 0$, see time traces in Fig.~\ref{fig:tau_gi}c).
For $g_{inh} \gtrsim 1020$~nS, a phase drift regime is reached.
When compared with those of the PRC approach (filled squares in Fig.~\ref{fig:tau_gi}a), namely, the fixed point solutions of Eqs.~\ref{eq:fixedpoints}, results agree very well.
The agreement extends for the whole $g_{inh}$ range, including the second transition from AS to the phase drift regime.
Interestingly, when we approximate $PRC_R(\beta,\alpha)$ by $PRC_R(\beta)+PRC_R(\alpha)$, a good agreement is obtained only for relatively small $g_{inh}$ (filled triangles in Fig.~\ref{fig:tau_gi}a).
Why does the approximation break down?
By examining Fig.~\ref{fig:PRC2D} it can be seen from panels b and d that the PRC landscapes are drastically different.
In Fig.~\ref{fig:PRC2D}d, the approximation keeps the fixed point solution in the high-$\beta^*$ range, therefore predicting delayed synchronization.
This can also be seen in the projection along a constant $\alpha^*$ value shown in Fig.~\ref{fig:tau_gi}e.
With the full $PRC_R(\beta,\alpha)$ (Fig.~\ref{fig:PRC2D}b), an increasing inhibitory conductance causes the fixed point to cross the zero-lag solution and a small value of $\beta^*$ is obtained, predicting an anticipated synchronization regime.
This transition is also illustrated in Fig.~\ref{fig:tau_gi}d.
For even larger values of $g_{inh}$, a transition to the phase drift regime is obtained.
Interestingly, the shape of $PRC_R(\beta,\alpha^*)$ shown in Fig.~\ref{fig:tau_gi}d for $g_{inh}=1200$~nS is reminiscent of that of a type-I excitable neuron, which is consistent with the absence of a locked regime.
\begin{figure}
\centering
\begin{minipage}{0.9\linewidth}
\includegraphics[width=0.9\linewidth,clip]{Fig4a}
\end{minipage}
\begin{minipage}{0.9\linewidth}
\includegraphics[width=0.9\linewidth,clip]{Fig4b}
\end{minipage}
\begin{minipage}{0.9\linewidth}
\centerline{\includegraphics[width=0.9\linewidth,clip]{Fig4c}}
\end{minipage}
\begin{minipage}{0.9\linewidth}
\centerline{\includegraphics[width=0.9\linewidth,clip]{Fig4d}}
\end{minipage}
\caption{
{\bf (Color online) Comparing numerical simulation with PRC prediction. }
a) Time delay between sender and receiver as a function of the inhibitory conductance $g_{inh}$.
b) and c) time traces of the membrane potential given by Eq.~\ref{eq:dvdt} for two different $g_{inh}$ values: in b) $g_{inh}=200$ nS and in c) $g_{inh}=1000$ nS.
In panel b the system is locked in a delayed-synchronization regime while in panel c it is locked in the anticipated-synchronization regime.
d) The one dimensional $PRC_R(\beta,\alpha^*)$ is plotted for a fixed value of $\alpha^*=T-\gamma^*$, where $\gamma^*$ is obtained from Fig.~\ref{fig:singlePRC}, for different values of $g_{inh}$.
e) The one dimensional $PRC_R(\beta)+ PRC_R(\alpha^*)$ is plotted for the same value of $\alpha^*$ as in d.
The difference in panels d and e reflect the discrepancies in the calculations of the fixed-point solutions for $\beta$.
}
\label{fig:tau_gi}
\end{figure}
\subsubsection{Different free-running periods}
Up to now, we focussed on the DS/AS transition assuming identical free-running periods for all the neurons~\cite{Matias11,Matias14,Matias16}.
The results presented in the previous section also assumed all periods to be identical.
However, the PRC approach, as presented in Eqs.~\ref{eq:prcmapbeta}-\ref{eq:prcmapalpha}, allows an extension to the case of different free-running periods.
Moreover, if $T_R$ does not change, the same $PRC_R(\beta,\alpha)$ shown in Fig.~\ref{fig:PRC2D}a and b is, in principle, still valid.
This, therefore, strengthens the predictive power of the PRC approach.
To probe it, we analyze a particularly relevant scenario where the interneuron has a different period than the others.
In neuroscience, it is often the case that inhibitory neurons spike faster than excitatory ones~\cite{Izhikevich07}.
We therefore focus on examining the dependence of the synchronization regimes on the free-running period $T_I$ of the interneuron.
From Eqs.~\ref{eq:prcmapbeta} and~\ref{eq:prcmapalpha} the fixed point solutions for $T_S \neq T_R \neq T_I$ become:
\begin{eqnarray}
PRC_R(\beta,\alpha) & = & \Delta T_{RS} \label{eq:PRCRDT} \\
PRC_I(\gamma; T_I) & = & \Delta T_{IS} \label{eq:PRCIDT}\; ,
\end{eqnarray}
where $\Delta T_{RS}=T_R-T_S$ and $\Delta T_{IS}=T_I-T_S$.
Note that we have now included an explicit dependence of $PRC_I$ on $T_I$.
To avoid recalculating $PRC_I(\gamma; T_I)$ for every $T_I$, we use instead an approximation that assumes that changes in the period amounts to a simple rescaling of the corresponding phase-resetting curve as:
\begin{equation}
\label{eq:gammanew}
PRC_I(\gamma; T_I) = \frac{T_I}{T} PRC_I\left(\gamma \frac{T_I}{T}; T\right)\; ,
\end{equation}
where $PRC_I\left(\gamma ; T\right)$ is the function shown in Fig.~\ref{fig:singlePRC} for the case of the three neurons having identical period $T$.
The analysis of Eq.~\ref{eq:PRCIDT} with Eq.~\ref{eq:gammanew} allows discriminating two possibilities in terms of the $PRC_I$:
$\it i)$ if $T_S>T_I$ (or equivalently the sender frequency is smaller than the interneuron frequency) the fixed point solutions exist until $\Delta T_{IS}$ reaches the minimum value of $PRC_I(\gamma; T_I)$. At this value the two fixed points collide and disappear and the system enters into a phase-drift regime.
$\it ii)$ if $T_I>T_S$ (or equivalently the sender frequency is higher than the interneuron frequency) the fixed point solutions exist until $\Delta T_{IS}$ reaches the maximum of $PRC_I(\gamma; T_I)$. At this value the two fixed points collide and disappear and the system enters into a second phase-drift regime.
\begin{figure}
\begin{center}
\includegraphics[width=0.9\columnwidth,clip]{Fig5}
\end{center}
\caption{
\label{fig:differentT}
{\bf (Color online) Comparison between the PRC prediction and SRI simulation for different free-running periods of the Interneuron.}
Time delay between Sender and Receiver as a function of the Interneuron period for the full simulation of the SRI motif (circles) and the PRC prediction (squares).
The vertical dashed line corresponds to the free-running period of both the Sender and the Receiver.
The inhibitory conductance is $g_{inh} = 1000$~nS, so that the function $PRC(\alpha,\beta)$ corresponds to that of Fig.~\ref{fig:PRC2D}b. }
\end{figure}
In Fig.~\ref{fig:differentT} we plot the time difference between sender and receiver $\tau_{SR}$ as a function of the free-running period of the interneuron $T_I$.
$\tau_{SR}$ is calculated with the fixed point solutions $\gamma^*$ obtained from Eq.~\ref{eq:gammanew} in combination with Fig.~\ref{fig:PRC2D}b.
A very good agreement can be seen when comparing the PRC's prediction with the numerical simulation of the full HH model.
For values of $T_I \lesssim 14.5$~ms and $T_I \gtrsim 18.7$~ms the phase-locked solution is lost and the system enters into a phase-drift regime.
\section{Concluding remarks \label{conclusions}}
In this paper we have used a phase-reseting curve (PRC) approach to gain insight into the transition from delayed to anticipated synchronization and anticipated synchronization to phase drift regime in a sender-receiver-interneuron motif.
Initially we assumed identical parameters and operating conditions for the three neurons.
The PRC of the receiver neuron was computed as a function of two inputs per cycle: one arriving from the sender and another from the interneuron.
We found that the description of the PRC in terms of two variables is essential to correctly match numerical results obtained from the full neuronal and synaptic model.
In particular our PRC approach correctly predicts the transition from the anticipated-synchronization to the phase-drift regime.
On the contrary, if the typical approximation is used, considering the sum of two PRCs from independent stimulus, the results significantly depart from the numerical solutions, with the largest discrepancies at intermediate to large values of the inhibitory conductance.
Moreover, this approximation does not account either for the AS/DS transition nor the AS/phase-drift regime transition observed both numerically in the full neuronal model and with the two-variable PRC.
We have also explored the PRC prediction when the neurons had different free-running periods.
Under this condition, the PRC calculation is easily extended, in particular when only the period of the interneuron element is varied.
Assuming that the PRC of the interneuron modifies according to a simple rescaling factor when its period changes,
we also obtain a very good agreement with the numerical simulations, highlighting the strength of the method.
Further investigations including different types of synapses and neuronal models (type-I vs. type-II excitability) as well as different pulsating regimes will be reported in a forthcoming publication.
\begin{acknowledgments}
We thank CNPq grant 310712/2014-9, FACEPE grant APQ-0826-1.05/15, CAPES grant PVE 88881.068077/2014-01 for financial support.
C.R.M. acknowledges support from the Spanish Ministerio de Econom\'{\i}a y Competitividad (MINECO) and Fondo Europeo de Desarrollo Regional (FEDER) through project TEC2016-80063-C3-3-R (AEI/FEDER, UE).
This article was produced as part of the activities of FAPESP Research, Innovation and Dissemination Center for Neuromathematics (grant 2013/07699-0, S. Paulo Research Foundation).
\end{acknowledgments}
|
\section{hyperfine interaction}
In this supplemental section we estimate the order of magnitudes of the hyperfine interaction arising from the Fermi contact, dipolar, and orbital contributions per nucleus. We start by writing the vector potential generated by a nuclear spin ${\bf I}$,
\begin{eqnarray}
{\bf A}_n &=& \frac{\mu_0} {4\pi} \frac{ \hbar \gamma_{n} {\bf I} \times {\bf r}} {r^3},
\end{eqnarray}
with the magnetic moment $\hbar \gamma_{n} {\bf I}$ carried by the nuclear spin and the gyromagnetic ratio $\gamma_n$. Here $r = |{\bf r}|$ is the distance between the electron and nucleus with ${\bf r} = (x,y,z)$. The magnetic field $\triangledown \times {\bf A}_n $ is then felt by an electron with spin ${\bf S}$, leading to the Fermi contact and dipolar contributions~\citesupp{Okvatovity:2016_S,Lunde:2013_S},
\begin{eqnarray}
H_{\textrm{hf}}^{\textrm{Fc}} &=& -\frac{\mu_0} {4\pi} \frac{8\pi} {3} \hbar \gamma_{n} g_{e} \mu_{B} {\bf S} \cdot {\bf I} \delta({\bf r}), \\
H_{\textrm{hf}}^{\textrm{dip}} &=& \frac{\mu_0} {4\pi} \hbar \gamma_{n} g_{e} \mu_{B} {\bf I} \cdot
\left[ \frac{ {\bf S} - 3 \hat{r} ({\bf S} \cdot \hat{r} ) } {r^3} \right],
\end{eqnarray}
with $g_{e}$ and $\mu_{B}$ being the electron gyromagnetic constant and Bohr magneton, respectively.
Following Ref.~\citesupp{Okvatovity:2016_S}, we obtain the orbital contribution with the Peierls substitution of ${\bf A}_n$ into the electron kinetic energy $H_{\textrm{kin}}$. For the edge states of two-dimensional topological insulators (2DTIs), we have $ H_{\textrm{kin}}({\bf p}) = v_F \sigma^z p^x$ with the Pauli matrix $\sigma^z$ and momentum $p^x = -i\hbar \partial_x $, from which we get
\begin{eqnarray}
H_{\textrm{hf}}^{\textrm{orb}} &=& H_{\textrm{kin}} ({\bf p} - e {\bf A}_n) - H_{\textrm{kin}} ({\bf p}) \nonumber\\
&=& \frac{\mu_0} {4\pi} \hbar \gamma_{n} e v_F \sigma^z \frac{ ( y I^z - z I^y ) } {r^3}.
\label{Eq:H_orb_2DTI}
\end{eqnarray}
This term, being proportional to $\sigma^z$, does not lead to the electron spin flip, and is not relevant to our analysis in the disordered phase, i.e., it does not enter the backscattering strength $D_{\textrm{hf}}$ in the main text. It does not contribute to the RKKY coupling, either. Nonetheless, we estimate its magnitude below, as well as the Fermi contact and dipolar contributions.
To proceed, we estimate the energy scales of the following matrix elements~\citesupp{Okvatovity:2016_S},
\begin{subequations}
\label{Eq:M_hf}
\begin{eqnarray}
\left < \Psi \left| H_{\textrm{hf}}^{\textrm{Fc}} \right| \Psi'\right> & = & -\frac{2\mu_0} {3} \hbar \gamma_{n} g_{e} \mu_{B} \left < \Psi \left| {\bf S} \cdot {\bf I} \delta({\bf r}) \right| \Psi'\right> , \label{Eq:M_fc} \\
\left < \Psi \left| H_{\textrm{hf}}^{\textrm{dip}} \right| \Psi'\right> &=& \frac{\mu_0} {4\pi} \hbar \gamma_{n} g_{e} \mu_{B}
\left < \Psi \left| {\bf I} \cdot \left[ \frac{ {\bf S} - 3 \hat{r} ({\bf S} \cdot \hat{r} ) } {r^3} \right] \right| \Psi'\right> ,\nonumber \\ \label{Eq:M_dip} \\
\left < \Psi \left| H_{\textrm{hf}}^{\textrm{orb}} \right| \Psi'\right>
&=& \frac{\mu_0} {4\pi} \hbar \gamma_{n} e v_F \left< \Psi \left| \sigma^z \frac{ ( y I^z - z I^y ) } {r^3} \right| \Psi'\right>, \label{Eq:M_orb}
\end{eqnarray}
\end{subequations}
where we choose the initial $\left| \Psi' \right>$ and final $\left| \Psi \right>$ states to be the edge states with the opposite (same) velocities for the spin-flip $S^{x,y}$ (spin-conserving $S^{z}$) terms.
To proceed, we express the edge states as the product of the Bloch amplitude $u_{\textrm{B}}$ and the envelope function ($\left| \Psi_{R/L} ({\bf r})\right>$ for the right/left state),
\begin{eqnarray}
\left| \Psi \right>,~\left|\Psi'\right> &=& u_{\textrm{B}}({\bf r}) \left| \Psi_{R/L} ({\bf r}) \right>,
\end{eqnarray}
where the Bloch amplitude $u_{\textrm{B}} ({\bf r})$ satisfies $\int d{\bf r} \; |u_{\textrm{B}} ({\bf r})|^2 f({\bf r}) \approx \int d{\bf r} \; f({\bf r})$ for functions $f({\bf r})$ that are smooth over the atomic scale $a_0$. We assume that $\left| \Psi_{R/L} ({\bf r})\right>$ can be factorized into the longitudinal and transverse parts,
\begin{eqnarray}
\left| \Psi_{R/L} ( {\bf r}) \right> &=& C_{\parallel} e^{\pm i k_{F} x} \left| \downarrow/\uparrow \right> \otimes \left| \Psi_{\perp} (y,z) \right>,
\end{eqnarray}
where the longitudinal part is written as the product of the spatial part (with the normalization factor $C_{\parallel} =1/\sqrt{L} $) and the spin state $\left| \sigma \right>$, and the transverse part fulfills $\left< \Psi_{\perp} (y,z)\right. \left| \Psi_{\perp} (y,z) \right> =1$. We now estimate the matrix elements in Eqs.~(\ref{Eq:M_hf}).
\subsection{I. Fermi contact contribution}
We start with the Fermi contact contribution. Here we compute the matrix element of the spin-flip $S^{+}\equiv S^{x} + i S^{y}$ term (the spin-conserving $S^{z}$ term should give similar results),
\begin{eqnarray}
E_{\textrm{Fc}} &\equiv& \frac{2\mu_0}{3} \hbar \gamma_{n} g_{e} \mu_{B} \left < \Psi \left| S^{+} \delta({\bf r}) \right| \Psi'\right> \nonumber \\
&=& \frac{2\mu_0}{3} \hbar \gamma_{n} g_{e} \mu_{B} \frac{1}{L w^2} \left|u_{\textrm{B}}( {\bf r}=0) \right|^2,
\label{Eq:E_fc}
\end{eqnarray}
where $w$ is the transverse length scale, and the values of $\eta \equiv \left|u_{\textrm{B}}( {\bf r}=0)\right|^2 $ for the relevant nuclei were estimated in semiconductor systems~\citesupp{Gueron:1964_S,Paget:1977_S,Schliemann:2003_S}. Although Refs.~\citesupp{Gueron:1964_S,Paget:1977_S,Schliemann:2003_S} are not about 2DTI materials (e.g. InAs and GaSb compounds), we expect that the $\eta$ value here to be of the same order of the magnitudes. For the Hg and Te nuclei in HgTe, however, the value of $\eta$ may be smaller due to the larger principal quantum number of the outermost electrons, as discussed in Ref.~\citesupp{Lunde:2013_S}.
\subsection{II. Dipolar contribution}
We now turn to the dipolar contribution to the hyperfine interaction. Because the spatial dependence in Eq.~(\ref{Eq:M_dip}) is smooth over the atomic scale, the details of $u_{\textrm{B}}({\bf r})$ do not play a role. Since we are only interested in the overall scale, we simplify the spatial dependence as $r^{-3}$, and estimate the spin-flip $S^{+}$ term in Eq.~(\ref{Eq:M_dip}),
\begin{eqnarray}
E_{\textrm{dip}}
&\equiv& \frac{\mu_0} {4\pi} \hbar \gamma_{n} g_{e} \mu_{B} C_{\parallel}^2 \int_{-\infty}^{\infty} dx \; e^{ 2i k_{F} x} \left < \Psi_{\perp} (y,z) \left| \frac{ 1 } {r^3} \right| \Psi_{\perp} (y,z) \right>. \nonumber\\
\end{eqnarray}
Assuming that the integrals over $y$ and $z$ coordinates lead to
\begin{eqnarray}
\left < \Psi_{\perp} (y,z) \left| \frac{ 1 }{r^3} \right| \Psi_{\perp} (y,z) \right> &=& \frac{1} { (x^2 + w^2)^{3/2}},
\end{eqnarray}
and performing the remaining $x$ integral, we get
\begin{eqnarray}
E_{\textrm{dip}} &\sim& \frac{\mu_0} {4\pi} \hbar \gamma_{n} g_{e} \mu_{B} \frac{0.1}{L w^2},
\label{Eq:E_dip}
\end{eqnarray}
where the factor $0.1$ comes from the suppression due to the oscillatory integrand. Comparing with the Fermi contact contribution Eq.~(\ref{Eq:E_fc}) gives
\begin{eqnarray}
\frac{E_{\textrm{dip}} } {E_{\textrm{Fc}}} &\sim& \frac{3}{80\pi\eta} \sim O(10^{-2}) \times \frac{1}{\eta},
\end{eqnarray}
which means the dipolar contribution is at least two orders smaller than the Fermi contact contribution as long as $\eta>1$, even though we do not know the exact $\eta$ value in InAs/GaSb.
Therefore, the dipolar contribution to the hyperfine interaction is much weaker than the Fermi contact contribution.
\subsection{III. Orbital contribution}
We now turn to the orbital contribution. Even though the orbital contribution does not cause spin flip, we still estimate its magnitude for the sake of completeness. Again, we use the fact that $y/r^3$ and $z/r^3$ are smooth on the atomic scale, so the details of $u_{\textrm{B}}( {\bf r})$ can be neglected. The matrix element Eq.~(\ref{Eq:M_orb}) can be written as
\begin{eqnarray}
\frac{\mu_0} {4\pi} \hbar \gamma_{n} e v_F C_{\parallel}^2 \int_{-\infty}^{\infty} dx \; \left < \Psi_{\perp} (y,z) \left| \frac{ ( y I^z - z I^y ) } {r^3} \right| \Psi_{\perp} (y,z) \right>. \nonumber \\
\end{eqnarray}
To proceed, we assume the integral over the transverse part can be written as,
\begin{subequations}
\begin{eqnarray}
\left < \Psi_{\perp} (y,z) \left| \frac{ y } {r^3} \right| \Psi_{\perp} (y,z) \right> &=& \frac{w_{y}} { (x^2 + w^2)^{3/2}} ,\\
\left < \Psi_{\perp} (y,z) \left| \frac{ z } {r^3} \right| \Psi_{\perp} (y,z) \right> &=& \frac{w_{z}} { (x^2 + w^2)^{3/2}},
\end{eqnarray}
\end{subequations}
with $w^2 =w_y^2 + w_z^2 $. Performing the remaining $x$ integral, we obtain the magnitude of Eq.~(\ref{Eq:M_orb}),
\begin{eqnarray}
E_{\textrm{orb}} &\sim& \frac{\mu_0} {4\pi} \hbar \gamma_{n} e v_F \frac{1}{L w}.
\end{eqnarray}
Comparing with the Fermi contact contribution Eq.~(\ref{Eq:E_fc}) gives
\begin{eqnarray}
\frac{E_{\textrm{orb}} }{E_{\textrm{Fc}}} &\sim& \frac{3}{8\pi} \frac{e v_F w}{g_e \mu_B \eta} \sim O(10^{-1}) \times \frac{1}{\eta},
\end{eqnarray}
where we have used the parameters, $\mu_0=4\pi \times 10^{-7}$~Vs/Am, $g_e=2$, $\mu_B= 5.8 \times 10^{-5}$~eV/T, $v_{F}=4.6\times 10^4~$m/s, and $w = 10~$nm. As mentioned above, even though we do not have the exact $\eta$ value, we find the above ratio to be much smaller than 1 for any $\eta>1$.
As a summary, both the dipolar and orbital contributions to the hyperfine interaction are much weaker than the Fermi contact contribution.
\subsection{IV. Hyperfine coupling}
We note that the energy scale $E_{\textrm{Fc}}$ in Eq.~(\ref{Eq:E_fc}) is not the hyperfine coupling $A_0$ in the main text. Since we define $A_0$ such that $H_{\textrm{hf}} \sim (A_0/\rho_{\textrm{nuc}}) \rho_{e} {\bf S} \cdot {\bf I} $ with the nuclear and electron densities, $\rho_{\textrm{nuc}}=8/a_0^3$ and $\rho_{e} \sim 1/(Lw^2)$, respectively, we have
\begin{eqnarray}
A_0 &\sim & \frac{16\mu_0}{3} \hbar \gamma_{n} g_{e} \mu_{B} \frac{\eta}{a_{0}^3}.
\label{Eq:A_0}
\end{eqnarray}
In Refs.~\citesupp{Gueron:1964_S,Paget:1977_S,Schliemann:2003_S}, the $\eta$ values for the relevant nuclei are given by $\eta_{\textrm{In}} = 6.3 \times 10^3$, $\eta_{\textrm{Sb}} = 1.1 \times 10^4$, $\eta_{\textrm{Ga}} = 2.7 \times 10^3$, and $\eta_{\textrm{As}} = 4.5 \times 10^3$. If we take the average value, $\eta=6.1 \times 10^3$, for our estimation, along with the parameters $\hbar \gamma_{n}=6 \times 10^{-8}~$eV/T, $a_0=6.1~${\AA}, $g_e=2$, and $\mu_B= 5.8 \times 10^{-5}$~eV/T, we obtain $A_{0}\sim O(100~\mu\textrm{eV})$. Since, however, the edge states are mixtures of $s$- and $p$-orbital states, the actual contribution from the $s$-orbital state may be somewhat smaller. We note that Ref.~\citesupp{Lunde:2013_S} investigated the hyperfine interaction in HgTe 2DTIs, and found that, for the spin-flip terms, the contribution from the Fermi contact dominates over the other contributions.
We also note that generalizing our model with an anisotropic hyperfine coupling ($A_x \neq A_y$) may modify the backscattering strength, e.g. by replacing $A_0^2 \rightarrow (A_x^2+A_y^2)/2$, but does not lead to a qualitative difference. As a result, we keep only the Fermi contact contribution in the main text for our analysis.
\section{Dipole-dipole interaction of nuclei}
In this supplemental section we discuss the effects of the dipole-dipole interaction between the nuclear spins on the electron backscattering. The nuclear dipolar interaction is much weaker than the electron-nuclear hyperfine interaction~\citesupp{Paget:1977_S}. It is nevertheless an important ingredient for the dissipation of the nuclear spin polarization.
The (dynamical) polarization would otherwise accumulate during spin-flip backscattering processes~\citesupp{Lunde:2012_S,Kornich:2015_S} and therefore prevent subsequent backscattering events. Even though the nuclear dipolar interaction is not explicitly written in our Hamiltonian Eqs.~(\ref{eq:H_LL})--(\ref{Eq:H_hf}), we implicitly include it in our analysis by considering that the nuclear subsystem is in its (thermal) ground state, whether ordered or not. The dipole-dipole nuclear spin diffusion is the mechanism for the nuclear subsystem to return to this ground state upon excitations by current.
We note that this makes our theory different from, e.g. backscattering on a single magnetic impurity or a spin bath where such a dissipation channel is absent, and the backscattering is trivially shut down once these magnetic impurities become polarized.
\section{RKKY interaction}
In this supplemental section we discuss the RKKY interaction mediated by a helical Tomonaga-Luttinger liquid. Similar to nonhelical systems~\citesupp{Simon:2008_S,Braunecker:2009a_S,Braunecker:2009b_S,Meng:2014a_S,Hsu:2015_S}, here we integrate out the electronic degrees of freedom in the hyperfine interaction to obtain the RKKY interaction between the localized nuclear spins. The interaction strength is proportional to the electronic spin susceptibility, and can be calculated along the line of Ref.~\citesupp{Giamarchi:2003_S}. The RKKY strength develops a dip at $q=2k_{F}$, with the magnitude,
\begin{eqnarray}
J^{x}_{2k_{F}} &\approx& - \frac{ \sin (\pi K) }{8\pi^2} \frac{K A_{0}^2}{ \Delta }
\left(\frac{\lambda_{T}}{2\pi a} \right)^{2-2K}
\left| \frac{ \Gamma\left( 1-K\right) \Gamma\left( \frac{K}{2} \right) }
{\Gamma\left(\frac{2-K}{2} \right)}\right|^{2}, \nonumber \\
\label{Eq:JRKKY}
\end{eqnarray}
with the Gamma function $\Gamma(x)$. Equation~(\ref{Eq:JRKKY}) is then used in the calculation of the magnon energy $\hbar \omega_{\textrm{mag}}$ and the resistance due to the magnon emission $R_{\textrm{mag}}^{\textrm{em}}$ [Eq.~(\ref{Eq:R_mag})] in the main text.
\section{Interplay of the electronic and nuclear subsystems}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\linewidth]{Overhauser.eps}
\caption{Backscattering due to the Overhauser field arising from the nuclear spin order with the ground state value $\left< \tilde{I}\right>_{-}$. A gap $\Delta_{\textrm{m}}$ opens below the Fermi surface.}
\label{Fig:backscattering}
\end{figure}
In this supplemental section we comment on two important features of the interplay of the electronic and nuclear subsystems.
First, in the ordered phase, the Overhauser field arising from the ordered nuclear spins induces an electronic gap $\Delta_{\textrm{m}}$ below the Fermi surface, as shown in Fig.~\ref{Fig:backscattering}. This gap may provide for additional experimental signatures of the nuclear spin order. Due to the separation of the time scales of the electron and nuclear subsystems, one can rapidly change the chemical potential via a gate voltage, while the spatial modulation of the ordered nuclear spins (thus the Overhauser field and the position of the gap) remains intact. Therefore, by sweeping the chemical potential across the gap, experimental signatures can be sought via techniques that were employed to detect the effects of the nuclear spins in semiconducting systems, such as transport, optical, and NMR measurements~\citesupp{Dobers:1988_S,Smet:2002_S,Chekhovich:2013_S,Tiemann:2014_S}.
Second, a finite gap $\Delta_{\textrm{hx}}$ due to the nuclear order-assisted backscattering on impurities reduces the RKKY coupling, which we did not take into account in the main text. This mechanism imparts a negative feedback onto the effects of the nuclear spin order on the resistance.
With the exact solution being beyond the scope of this work, we only remark that we expect our results to remain qualitatively valid as long as $\Delta_{\textrm{hx}} \ll \epsilon_{F}$, which preserves the sharp RKKY dip around momentum $2k_F$, albeit with a reduced height~\citesupp{Klinovaja:2013b_S,Meng:2014a_S,Hsu:2015_S}.
\bibliographystylesupp{apsrev4-1}
\bibliographysupp{HLLsupp}
\end{document}
|
\section{The nearly-free-particle approximation of 1+1 D space-time crystal}
\begin{figure}
\includegraphics[height=0.45\columnwidth, width=0.45\columnwidth]
{zone_folding_a.eps}
\includegraphics[height=0.45\columnwidth, width=0.45\columnwidth]
{zone_folding_b.eps}
\includegraphics[height=0.5\columnwidth, width=0.6\columnwidth]
{zone_folding_c.eps}
\caption{(a). The nearly-free-particle dispersion without the
zone-folding.
When two momentum-energy vectors are differed by a reciprocal lattice
vector $B$, whose Fourier component of the space-time lattice
potential $V_B\neq 0$, they are marked in pairs.
$V_B$ results in level repulsion and open a gap at the value of $2|V_B|$.
(b) The dispersion duplicated in extended MEBZs.
The relevant MEBZs are sketched in the background for clarity.
The first MEBZ is centered on the origin.
(c) We smoothly vary the loop structure so that the winding
number structure becomes transparent.
}
\label{fig:zonefolding}
\end{figure}
In this section, we expand the discussion on the nearly-free-particle
band structure of the space-time crystal and explicitly demonstrate the
momentum-energy Brillouin zone (MEBZ) folding procedure.
We consider a weak space-time lattice potential
$V(\mathbf{r}, t)$ and work in the framework of the generalized
Floquet-Bloch theorem based on Equations (2) and (3) in the main text.
Consider a Floquet-Bloch state $\psi_{\kappa,m}(\mathbf r, t)$ with
the good quantum number $\kappa=(\mathbf{k},\omega_m)$ and the band
index $m$.
Its wavefunction component in terms of each reciprocal lattice vector
$B=(\mathbf G, \Omega)$ is denoted as $c_{m,B}$.
By construction, if $c_{m, B}$ is a solution, then $c_{m,B-B\prime}$
with $B^\prime=(\mathbf{G^\prime}, \Omega^\prime)$ corresponds to the
solution with an equivalent quasi-momentum and energy
$(\mathbf{k}+\mathbf{G}', \omega_m+\Omega')$.
Nevertheless, this remains the same state as before.
In order to remove this redundancy, $\kappa$ can be constrained
in the first momentum-energy Brillouin zone (FMEBZ), i.e., the unit cell
in the reciprocal space centered at the origin.
Dispersions residing in high order MEBZs can be folded into
the FMEBZ.
The general process is as follows:
First fold the free spectrum $\varepsilon_0(\mathbf k)$ into the FMEBZ.
When it crosses, the corresponding $\kappa$'s of the unfolded spectrum
are differed by a momentum-energy reciprocal lattice vector $B$.
These two plane-wave states are hybridized in the presence of the
nonzero Fourier component $V_B$, leading to the level repulsion at
the crossing point and yielding the Floquet-Bloch wavefunctions.
Consequently, the quadratic spectrum breaks into loops winding
around in the FBZ.
Fig. \ref{fig:zonefolding} demonstrates an example of zone-folding
of the dispersion corresponding to the Fig 1. (a) in the main text.
Fig. \ref{fig:zonefolding} (a) presents an unfolded dispersion
in the presence of a weak space-time potential.
The gap opening points are marked in pairs on the unfolded dispersion:
The dashed arrows linking two points represent the non-vanishing Fourier
components $V_B$ of the external potential.
In Fig. \ref{fig:zonefolding} (b), the extended MEBZ representation
is used, i.e., the dispersion is duplicated being shifted by all
the momentum-energy reciprocal lattice vectors.
The FMEBZ boundary is marked by the solid line.
The Floquet-Bloch dispersion forming two loops after folded into
the FMEBZ as shown in Fig. \ref{fig:zonefolding} (c).
The parts of the spectrum from $2\rightarrow 3$ and from $3^\prime \to 4^\prime$
does not participate in forming these loops, and are omitted.
We can smoothly vary the loop structure in the FMEBZ such that
the winding numbers of each loop become transparent: The red and blue
loops have the winding numbers $(0, 1)$ and $(1, 0)$, respectively.
These two loops actually cross, however, they do not open the gap
due to the lack of non-zero Fourer component $V_B$ with $B$ connecting
two momentum-energy vectors at the crossing point as shown in Fig.
\ref{fig:zonefolding} (a).
Otherwise, we will arrive at the situation shown in Fig 1. (b) in
the main text, where the two loops merge into one with the winding
number $(1,1)$.
The above example demonstrates the connection between the winding
numbers and the non-vanishing Fourier components of the external
potential.
The crossing of bands can be protected from splitting by space-time
symmetries of the system.
For example, as the example shown in the main text, the glide
time-reversal $g_t$ protects the spectral double degeneracy.
\section{Construction of the space-time group via group cohomology}
Each space group is constructed from a static Bravais lattice.
Similarly, each space-time group is constructed from a Bravais lattice $M$
constituted of the $D+1$ dimensional space-time mixed discrete
translations, and a magnetic point group in $d$ dimensions
$G_m(D)$ that leaves $M$ invariant.
Here ``magnetic" refers to the reflection with respect to time,
{\it i.e.}, time-reversal operation.
Below we use denote the space-time group by the symbol $ST$.
\subsection{The general procedure}
The hierarchical classification scheme of the space-time group starts with
the {\it crystal systems}.
Each crystal system is labeled by a set of {\it Bravais lattices}
$\{M\}$ and these lattices share the same magnetic point group symmetry.
Below we use the same symbol $M$ to represent the lattice translation
group associated to the Bravais lattice $M$.
The magnetic point group symmetry of a crystal is often smaller than
the that of the underlying Bravais lattice.
As a result, each crystal system can be further divided into different
{\it geometry} crystal classes (GCC) according to different magnetic
point group symmetries.
Each GCC can be further classified into different arithmetic crystal
classes (ACC) based on a particular Bravais lattice and a particular
magnetic point group.
It is worth noting that, the same Bravais lattice and magnetic point
group can give rise to different ACCs depending on the realizations
of magnetic point group operations.
For each ACC, based on its unique Bravais lattice and magnetic point
group, different space-time groups can be constructed.
Such hierarchy is in parallel to the
space group classification scheme of static lattices.\cite{Prince2004}.
\begin{table}[hptb]
\setlength\extrarowheight{4pt}
\centering
\begin{tabular}{|c|c|c|}
\hline
Point Group &$G_m(2)$ & Generators \\ \hline
\multirow{2}{*}{$C_1$} &$1$ & \\ \cline{2-3}
&$11'$ &$m_t$ \\ \hline
\multirow{3}{*}{$D_1$} &$m$ &$m_x$ \\ \cline{2-3}
&$m1'$ &$m_x, m_t$ \\ \cline{2-3}
&$m'$ &$m_xm_t$ \\ \hline
\multirow{3}{*}{$C_2$} &$2$ &$R_\pi$ \\ \cline{2-3}
&$21'$ &$R_\pi, m_t$ \\ \cline{2-3}
&$2'$ &$R_\pi m_t$ \\ \hline
\multirow{4}{*}{$D_2$} &$mm2$ &$m_x, m_y$ \\ \cline{2-3}
&$mm21'$ &$m_x, m_y, m_t$ \\ \cline{2-3}
&$m'm2'$ &$m_xm_t, m_y$ \\ \cline{2-3}
&$m'm'2$ &$m_xm_t, m_ym_t$ \\ \hline
\multirow{2}{*}{$C_3$} &$3$ &$R_{2\pi/3}$ \\ \cline{2-3}
&$31'$ &$R_{2\pi/3}, m_t$ \\ \hline
\multirow{3}{*}{$D_3$} &$3m$ &$R_{2\pi/3}, m_x$ \\ \cline{2-3}
&$3m1'$ &$R_{2\pi/3}, m_x, m_t$ \\ \cline{2-3}
&$3m'$ &$R_{2\pi/3}, m_xm_t$ \\ \hline
\multirow{3}{*}{$C_4$} &$4$ &$R_{\pi/2}$ \\ \cline{2-3}
&$41'$ &$R_{\pi/2}, m_t$ \\ \cline{2-3}
&$4'$ &$R_{\pi/2}m_t$ \\ \hline
\multirow{4}{*}{$D_4$} &$4mm$ &$R_{\pi/2}, m_x$ \\ \cline{2-3}
&$4mm1'$ &$R_{\pi/2}, m_x, m_t$ \\ \cline{2-3}
&$4'm'm$ &$R_{\pi/2}m_t, m_x$ \\ \cline{2-3}
&$4m'm'$ &$R_{\pi/2}, m_xm_t$ \\ \hline
\multirow{3}{*}{$C_6$} &$6$ &$R_{\pi/3}$ \\ \cline{2-3}
&$61'$ &$R_{\pi/3}, m_t$ \\ \cline{2-3}
&$6'$ &$R_{\pi/3}m_t$ \\ \hline
\multirow{4}{*}{$D_6$} &$6mm$ &$R_{\pi/3},m_x$ \\ \cline{2-3}
&$6mm1'$ &$R_{\pi/3},m_x, m_t$ \\ \cline{2-3}
&$6'm'm$ &$R_{\pi/3}m_t,m_x$ \\ \cline{2-3}
&$6m'm'$ &$R_{\pi/3},m_xm_t$ \\ \hline
\end{tabular}
\caption{The 2D magnetic point groups and their relations to the usual
ten 2D point groups.
The magnetic groups are denoted by using the international notation,
where $'$ means the time reversal operation.
In the third column, the symmetry generators for each magnetic point
group are listed.
The symmetry operation $m_t$ stands for the time reversal, $R_\theta$
represents rotation in the $x$-$y$ plane at the angle of $\theta$,
$m_{x}$ and $m_y$ are the reflections with respect to the $x$- and
$y$-directions, respectively.
The symbols of the magnetic group with $'$
}
\label{tb:mp2D}
\end{table}
\begin{table}[hptb]
\setlength\extrarowheight{5.5 pt}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Crystal System &Bravais Lattice &MP Group &ACC & $ST(2,1)$ \\ \hline
Triclinic &Primitive &$1, 2'$ &$2$ &$2$ \\ \hline
\multirow{2}{*}{T-Monoclinic} &Primitive &\multirow{2}{*}{$11', 2, 21'$} &$3$ &$8$ \\ \cline{2-2} \cline{4-5}
&Centered & &$3$ &$5$ \\ \hline
\multirow{2}{*}{R-Monoclinic} &Primitive &\multirow{2}{*}{$m, m', m'm2'$} &$3$ &$8$ \\ \cline{2-2} \cline{4-5}
&Centered & &$3$ &$5$ \\ \hline
\multirow{5}{*}{Orthorhombic} &Primitive &\multirow{5}{*}{\pbox{20cm}{$mm2, m'm'2$\\$mm21', m1'$}} &$4$ &$68$ \\ \cline{2-2} \cline{4-5}
&T-Base-Centered & &$4$ &$15$ \\ \cline{2-2} \cline{4-5}
&R-Base-Centered & &$5$ &$22$ \\ \cline{2-2} \cline{4-5}
&Face-Centered & &$4$ &$7$ \\ \cline{2-2} \cline{4-5}
&Body-Centered & &$4$ &$15$ \\ \hline
\multirow{2}{*}{Tetragonal} &Primitive &\multirow{2}{*}{\pbox{20cm}{$4,41', 4'$\\$4mm, 4mm1'$\\$4'm'm, 4m'm' $}}&$8$ &$49$ \\ \cline{2-2} \cline{4-5}
&Body-Centered & &$8$ &$19$ \\ \hline
\multirow{2}{*}{Trigonal} &Primitive &\multirow{2}{*}{\pbox{20cm}{$3, 6', 3m$ \\$3m' , 6'm'm$}} &$8$ &$18$ \\ \cline{2-2} \cline{4-5}
&Rhombohedral & &$5$ &$7$ \\ \hline
Hexagonal &Primitive & \pbox{3cm}{$6, 61', 31'$\\$6mm, 6m'm'$\\$6mm1', 3m1'$} &$8$ &$27$ \\ \hline
\end{tabular}
\caption{Classification of the 2+1D space-time groups.
There are 7 space-time crystal systems, 14 space-time
Bravais lattices (The primitive trigonal lattice is
the same as the primitive hexagonal lattice).
The 31 magnetic point groups are uniquely assigned to
the 7 crystal systems, as listed in the third column.
In the 4th and 5th columns, the numbers of the ACC
and the space-time groups for each Bravais lattice
are listed.
}
\label{tb:sptgrps2D}
\end{table}
\subsection{The method of group cohomology}
Given an ACC labeled by $M$ and $G_m$, where $M$ is a Bravais lattice and
$G_m$ is a magnetic point group symmetry, all its space-time groups can
be classified via the method of group cohomology theory, similar to the
space group classifications \cite{Hiller1986}.
A space-time group $ST$ is the group extension of the lattice
translation group of $M$ by the magnetic point group $G_m$,
described by the following exact sequence.
\begin{eqnarray}
1\rightarrow M \rightarrow ST \rightarrow G_m\rightarrow 1,
\end{eqnarray}
where $\rightarrow$ means a mapping.
For two consecutive mappings, the image of the first mapping is the
same as the kernel of the second one.
Such a group extension can be constructed by associating each magnetic
point group element $g$ in $G_m$ by a $D+1$ dimensional fractional
translation $c(g)$ not belonging to the lattice translation group.
$c(g)$ can be viewed as a mapping from $G_m$ to $T_{D+1}/M$
where $T_{D+1}$ is the continuous translation group in $D+1$ dimensions.
$c(g)$ satisfies
\begin{eqnarray}
c(1)=0, \ \ c(g_1g_2)=c(g_1)+g_1c(g_2).
\label{eq:cocycle}
\end{eqnarray}
We denote $(c(g),g)$ as the combined operation of first applying $g$ followed
by $c(g)$, then the set of these elements form a group following the rule
of product as
\begin{eqnarray}
(c(g_1), g_1) (c(g_2),g_2)=(c(g_1g_2), g_1g_2).
\end{eqnarray}
To classify all the space-time groups for the same ACC, a key observation is
that all the different ways of mapping themselves form an Abelian group.
Given two distinct mapping $c$ and $d$, we define their product $c\cdot d$
as
\begin{eqnarray}
c\cdot d (g)=c(g)+d(g),
\end{eqnarray}
which satisfies Eq. \ref{eq:cocycle} as well.
This group is denoted as $Z^1(G_m, T_{D+1}/M)$.
However, not all the elements in $Z^1(G_m, T_{D+1}/M)$ correspond to distinct
types of space-time groups.
Without specifying the equivalence relations, $Z^1(G_m, T_{D+1}/M)$
is not finite.
An obvious condition is that space-time groups that related by shifting the
origin are of the same type.
For example, the map of
\begin{equation}} \def\ee{\end{equation}
c_u(g)=gu-u,
\label{eq:coboundary}
\ee
simply shifts the magnetic point group operation origin to $u \in T_{D+1}/M$,
which should be identified with the trivial map $c(g)=0$.
The maps with the structure in Eq. \ref{eq:coboundary} form a group
$B^1(G_m, T_{D+1}/M)$.
The first cohomology group of $G_m$ with coefficients in $T_{D+1}/M$
is defined as the quotient group of
\begin{eqnarray}
H^1(G_m, T_{D+1}/M)=Z^1(G_m, T_{D+1}/M)/B^1(G_m, T_{D+1}/M),
\nonumber \\
\end{eqnarray}
Each element of $H^1(G_m, T_{D+1}/M)$ corresponds to one space-time group
of a particular ACC characterized by the magnetic point group $G_m$ and
the Bravais lattice $M$.
The identity of $H^1(G_m, T_{D+1}/M)$ is the trivial map that $c(g)=0$
for all $g$ in the magnetic point group $G_m$.
The corresponding space-time group is the semi-direct product of the
lattice translation group $M$ and the magnetic point group $G_m$, which
is called the ``symmorphic" space-time group.
By definition, each ACC only contains one symmorphic space-time group.
Other elements in $H^1(G_m, T/M)$ correspond to ``nonsymmorphic" space-time
groups.
For a crystal with ``nonsymmorphic" space-time group symmetries, it is
typically not invariant under the magnetic point group operations in $G_m$.
However, not all the elements of the cohomology group $H^1(G_m, T_{D+1}/M)$
lead to distinct space-time groups either.
For example, two elements in $H^1(G_m, T_{D+1}/M)$ related by a global
rotation should be identified.
Hence, we invoke the second equivalence relation that all elements
in $H^1(G_m, T_{D+1}/M)$ related by linear transformations $\rho$ are
identified, where $\rho$ leaves \textit{all} the magnetic point groups
of the crystal system unchanged.
\section{Classification of the 2+1D space-time group}
\begin{figure}
\includegraphics[height=0.35\columnwidth,width=0.4\columnwidth]{t_monoclinic.eps}
\includegraphics[height=0.35\columnwidth,width=0.35\columnwidth]{r_monoclinic.eps}
\caption{The different types of monoclinic space-time crystal systems.
(a) The $t$-monoclinic crystal system.
The $c$-axis is along the temporal direction and perpendicular to
the $xy$-plane.
(b) The $r$-monoclinic crystal system.
The $c$-axis is along one spatial direction, say, the $y$-direction,
and is perpendicular to the $xt$-plane.
The maximal magnetic point groups for (a) and (b)
are $21'$ and $m'm2$, respectively.
}
\label{fig:mono}
\end{figure}
\begin{figure}
\includegraphics[height=0.35\columnwidth, width=0.35\columnwidth]
{tcentered.eps}
\includegraphics[height=0.35\columnwidth, width=0.35\columnwidth]
{rcentered.eps}
\caption{The two types of base-centered Bravais lattices in the
space-time orthorhombic crystal system.
(a) The $t$-base-centered lattice, and (b) the $r$-base-centered
lattice.
The centered bases lie in the $xy$ and $xt$ planes, respectively.
}
\label{fig:base_orthg}
\end{figure}
In the main text, we list the classification of the space-time groups
in 1+1 dimensions as a proof of concept.
In this section, we focus on the case of 2+1 dimensions.
The case of 3+1 dimensions is left for future study.
In 2 dimensions, there exist 10 crystallographic point groups, which give
rise to 31 magnetic point groups as listed in Table. \ref{tb:mp2D}.
By combining the magnetic point groups with the discrete translation
symmetries in $T_{2+1}$, we obtain 7 space-time crystal systems, 14
Bravais lattices, and 275 space-time groups.
The relation among space-time crystal systems, Bravais lattices,
magnetic point groups, and space-time crystals is summarized in
Table \ref{tb:sptgrps2D}.
We adopt the terminology mostly from crystallography for 3D static
systems \cite{Aroyo2016}.
The detailed structure of each space-time group structure will be
published somewhere else.
The difference between the 2+1D space-time group and the 3D static space
groups is outlined below.
The main point is the specialty of the temporal direction: The
time reversal operation is anti-unitary, and space-time mixed
rotations are not allowed in the space-time group.
The second factor significantly affects the classification.
As shown in the first column of Table. \ref{tb:sptgrps2D}, the
difference already appears on the level of crystal systems.
Since the temporal and spatial directions are non-equivalent,
the cubic crystal system is absent in 2+1 D, and in addition,
there are two kinds of monoclinic crystal systems, $r$-monoclinic
and $t$-monoclinic, depending on whether the $c$-axis is along
the temporal or spacial directions.
The primitive unit cells for both $r$- and $t$-monoclinic space-time
crystal systems in Fig. \ref{fig:mono}.
The maximal magnetic point groups of these two crystal systems
are nonequivalent.
In the first case, it is $21'$ generated by the spatial rotation
$R_\pi$ and time reversal $m_t$, while it is $m'm2$ in the second
case generated by $m_ym_t, m_x$.
Furthermore, within the orthorhombic space-time crystal system,
there are two different base-centered Bravais lattices illustrated
in Fig. \ref{fig:base_orthg}, depending on the centered bases lying
in the $xy$ and $x(y)t$-planes, respectively.
The former is denoted as $r$-base-centered lattice, and the latter
is $t$-base-centered lattice.
As a result, there exist 5 Bravais lattices in the orthorhombic space-time
crystal system rather than 4 in its 3D static counterpart.
In these 2+1D space-time crystal systems, the number of the space-time groups
within each specific ACC is also in general different from its 3D
static counterpart.
On the other hand, when it comes to the tetragonal, the trigonal
and the hexagonal crystal systems, there is a one-to-one
correspondence between the 2+1 D space-time groups and the 3D
space groups, since the rotation plane must be purely spatial.
\section{Protected degeneracies due to non-commutative symmetry operators}
In the section, we discuss the protected band structure degeneracy
occurring at high symmetric points in MEBZ due to the non-symmorphic
nature of the space-time group symmetry.
A generic group element $g$ in the space-time group takes the form,
\begin{equation}} \def\ee{\end{equation}
g=T_\mathbf{r}(\mathbf{u})T_t(\tau) R m_t^s
\ee
where $T_\mathbf{r}(\mathbf{u})$ is the spatial translation, $T_t(\tau)$
is the temporal translation; $R$ is a point group operator
acting only on the spatial dimensions; $m_t$ is the anti-unitary
time-reversal operation;
$s=1$ or 0 determines whether $g$ includes time-reversal and is anti-unitary.
Consider two operations $g_1$ and $g_2$ in the little group of a
high symmetric point $\kappa=(\mathbf{k},\omega)$.
The degeneracy condition at $\kappa$ can be obtained by commuting
these two operators.
After some algebra, we arrive at
\begin{eqnarray}
g_1g_2&=&T_\mathbf{r}(\tilde{\mathbf{u}})T_t(\tilde{t})g_2g_1\tilde{R}
\end{eqnarray}
with
\begin{eqnarray}
\tilde{R}&=&R_1^{-1}R_2^{-1}R_1R_2, \nonumber \\
\mathbf{\tilde{u}}&=&(\mathbf{I}-R_2)\mathbf{u_1}
-(\mathbf{I}-R_1)\mathbf{u_2}, \nonumber \\
\tilde{t}&=&2 s_2 t_1-2 s_1 t_2.
\label{SM:degeneracy}
\end{eqnarray}
We assume $R_1$ and $R_2$ commute, i.e., $\tilde{R}=\mathbf{I}$, then
their operations on Floquet-Bloch wavefunctions with $\kappa$ satisfy
\begin{eqnarray}
M_{g_1}M_{g_2}=
e^{i \mathbf{k}\cdot\tilde{\mathbf{u}}-i \omega \tilde{t}}M_{g_2} M_{g_1}.
\label{eq:commute}
\end{eqnarray}
Depending on whether $g_1$ and $g_2$ are unitary or anti-unitary,
there are three different cases.
First, if neither of $g_1$ and $g_2$ flips the direction of time,
i.e., both $M_{g_1}$ and $M_{g_2}$ are unitary,
then $\tilde{t}=0$ and the phase factor in Eq. (\ref{eq:commute})
is independent of $\omega$.
This is the situation which has been studied in the main text.
Second, one of the $g$ operators, without loss of generality,
say, $g_1$, flips the direction of time and the other does not.
In this case, $M_{g_1}$ is anti-unitary while $M_{g_2}$ is unitary.
Now $\tilde{t}=2t_1$, and the prefactor in Eq. \ref{eq:commute}
does have frequency dependence.
However, due to the involving of the anti-unitary operator, the degeneracy
condition is more subtle than the previous case and shall
be discussed more carefully.
Consider a Floquet-Bloch state $\psi_\kappa$ as an eigenstate of
the unitary operator $g_2$,
\begin{equation}} \def\ee{\end{equation}
M_{g_2} \psi_\kappa=e^{i \mathbf{k}\cdot \mathbf{u_2}-i \omega t_2}
e^{i\theta} \psi_\kappa,
\ee
in which we explicitly separate the phase dependence on $\kappa$
, and $\theta$ \textit{only} depends on the point group
operation $R_1$.
Based on Eq. \ref{eq:commute}, one can show that $M_{g_1} \psi_\kappa$
is also an eigenstate of $g_2$ with, in principle, a different eigenvalue,
\begin{equation}} \def\ee{\end{equation}
M_{g_2} M_{g_1} \psi_\kappa=e^{-i\mathbf{k}\cdot (\mathbf{u_2}+\tilde{\mathbf{u}})-i \omega t_2}e^{-i\theta} M_{g_1} \psi_\kappa.
\ee
Therefore, if the two phases do not equal, i.e.
$e^{i \mathbf{k}\cdot (2\mathbf{u}_2+\tilde{\mathbf{u}})+2i\theta}\neq1$,
the space-time symmetry-protected degeneracy occurs.
Similar to the previous case, the degeneracy condition does not depend
on the frequency component of $\kappa$.
The last case is when both $g_1$ and $g_2$ flip the time-direction,
i.e., both $M_{g_1}$ and $M_{g_2}$ are anti-unitary.
This can be reduced to the 2nd case by defining $g_1^\prime=g_1g_2$,
whose $M_{g_1^\prime}$ is unitary again.
Then $g_1^\prime$ and $g_2$ satisfy
\begin{eqnarray}
g_1^\prime g_2=T_{\mathbf r} (\mathbf{u}) T_t (\tau) g_2 g_1^\prime,
\end{eqnarray}
where $\mathbf{u}$ and $\tau$ are defined according in
Eq. \ref{SM:degeneracy}.
In short, the degeneracy condition arising from the space-time symmetries
considered here does not depend on the frequency.
This is expected since one can always shift the frequency of the spectrum
by adding a constant to the time-dependent Hamiltonian.
In principle, one could also study the degeneracy condition resulting
from the interplay between the space-time symmetry and the chiral symmetry.
|
\section{Introduction}
\label{sec:intro}
The standard model of cosmology - Lambda Cold Dark Matter (LCDM) - is one of the biggest accomplishments
of modern physics of the last three decades. This model describes how the Universe cooled down and expanded
from the initial fireball of the Big Bang and formed the large-scale structure (LSS) observed presently in
wide and deep galaxy spectroscopic surveys. Astonishingly, LCDM, being a very simple model characterised by only six
free parameters, passes a tremendous number of robust observational tests. It explains the features and correlations
observed in the Cosmic Microwave Background (CMB) \citep[{\frenchspacing\it e.g.}][]{WMAP9,Planck1}, the primordial nucleosynthesis and light element abundance
\citep{Yang1984,Walker1991},
the growth of tiny primordial density perturbations into LSS \citep{Percival2001,Tegmark2004,Beutler2017} and
the late-time observed accelerated expansion \citep{acceleration1,acceleration2,Percival2010,Weinberg2013}.
However, this spectacular success comes with a high price, since LCDM is mostly phenomenological in its nature.
This is because in the model, the main contributors to the cosmic energy budget are dark matter (DM) and dark energy (DE).
The physical nature of both of these constituents is far from being understood and assessed.
While the observational evidence for dark matter's existence is overwhelming, with the general consensus being that
the last piece of this puzzle is still missing due to difficulties related to hunting for the DM particle
in Earth-based laboratories like CERN and Fermilab \citep{Ellis2000,Ahmed2009,Klasen2015}, the conceptual and theoretical problems related to DE
are of a much more profound nature. One of the core assumptions of LCDM is that General Relativity (GR) is
an adequate description of gravity on all cosmic scales, from the size of the Solar System ($\sim$AU) to the scale
of particle horizon ($\sim$ Gpcs). Within this picture the only possibility to accommodate the late-time accelerated
expansion of the Universe is
via a perfect fluid of an exotic equation of state parameter $w=-1$, dubbed Dark Energy. The simplest and most natural
candidate for DE is Einstein's cosmological constant (CC) -- $\Lambda$ -- reflecting here the zero-point energy
associated with quantum vacuum fluctuations. Alas, identification of $\Lambda$ with DE is spoiled by the gargantuan
discrepancy between the observed tiny value of CC and the theoretical prediction from quantum field theory.
This precipice has at least 50 orders of magnitude \citep{Carroll2001}. One of the commonly adopted solutions here consists of an arbitrary
assumption that by some symmetry of nature the true Einstein's constant is set to exactly zero, and the observed
accelerated expansion is due to an exotic scalar-field or other similar phenomena dominating the cosmic energy budget
at late time \citep{Copeland2006}. The main problem with this approach is that most theories experience
only weak coupling of the scalar field to matter, predicting growth of structure that is the same as in GR. This
makes these models very hard to falsify.
Another way to approach the DE problem is rooted in the observation that GR has been rigorously tested only on small
scales of the order of hundreds Astronomical Units at most \citep{Bertotti2003,Will2014,Baker2015,Brax2014,Raveri2015,Berti2015,Abbott2016}.
Thus, using GR to describe intergalactic, cosmological
and ultimately horizon scales, is an extrapolation by a spectacular 15 orders of magnitude. Therefore, one could argue for
a different interpretation of the late-time Universe's acceleration, which would not be due to a mysterious DE, but
could be a manifestation of the breakdown of GR on cosmological scales. Such scenarios have received large attention
in the literature over the past two decades and are commonly described as Modified Gravity (MG) theories. Here, viable theories
are usually built as infra-red modifications to GR that can fuel the low redshift acceleration of the space-time via virtue
of modifications to the Friedman equations stemming from an altered Einstein-Hilbert action \cite{cddett2005,cfps2011}. Most of such non-trivial modifications
of GR exhibit additional degrees of freedom, propagation of which will locally manifests as an additional fifth-force acting
on test matter particles. Such a fifth force can be usually described in terms of a local effective Newtonian gravitational constant
($G_{eff}$). Notwithstanding, as mentioned earlier, we have stringent precise tests of GR on small scales, thus any prospective
MG theory needs to employ a theoretical mechanism that would allow it to recover standard GR-behaviour on Solar System scales
and around dense bodies (like black holes and neutron stars). Such an appliance is called a {\it screening mechanism} in MG parlance.
While the screening mechanism allows a given theory to pass the small-scale and strong field regime observational tests,
its introduction comes with a high price, as theories with screening exhibit much stronger non-linear behaviour \citep{bdlw2012}.
The standard cosmological model, as any other successful physical theory, is placed under continuous scrutiny.
The fact that GR has not been rigorously tested on cosmological scales, put together with the fact that on a theoretical side
we are provided with a plethora of interesting modified theories, indicates that we are in an urgent need of precise observational
tests of gravity applied to intergalactic and cosmological scales. This worthy goal was set as the aim of many currently undergoing
and approaching observational endeavours, such as Euclid and DESI \citep{Laureijs2011,Levi2013,DESI2016} to name a few. These programmes aspire to chart
the Universe's large scale structure on vast spatial and time scales. The hope is that by analysing properties of LSS one can
measure the growth rate of structure at different epochs, which when determined with enough accuracy would constitute a strong
null-test for GR on cosmic scales \cite{Jain2010,Bean2010,Berti2015}. The validity of this test relies on the observation that within the standard paradigm
the observed LSS structures arose from tiny primordial density perturbations via a mechanism called {\it the gravitational instability}.
This mechanism explains how due to self-gravity, in an expanding background, small initial overdense regions
($\rho(\vec{x})>\langle\rho\rangle$) collapse into bound, and finally virialised, structures that host luminous
galaxies today. Here, the theory aspires to describe the growth on tiny irregularities in the course of the cosmic expansion history
from initial density perturbations of the order of $\sim10^{-5}$, as observed in the CMB \citep{Kogut1993}, to present day DM haloes
characterised by central densities of $\stackrel{>}{{}_\sim} 10^6$ \citep[see {\frenchspacing\it e.g.}][]{Frenk1985,Makino1998,FrenkWhite2012}.
Thus, to connect the theoretical predictions with observations we need an accurate description of the growth
of structures spanning more then 10 orders of magnitude in density.
On sufficiently large scales ({\frenchspacing\it i.e.}{} $\stackrel{>}{{}_\sim} 100h^{-1}\,{\rm Mpc}$) or interdependently at sufficiently early times ($z\simgt1$)
the growth of structures can be described by linear and weakly non-linear perturbation theory (PT). This picture has been shown
to be accurate and has been tested rigorously in the past, both for GR and non-GR theories \citep[see {\frenchspacing\it e.g.}][]{BCGS_book}. However, the bulk amount
of cosmological observations concerns the regime where the cosmic structure and its dynamics are deep in the non-linear regime.
The only way to probe and study the non-linear regime of gravitational instability is to use sophisticated and expensive
computer simulations. Use of simulations to study the growth of LSS has become a standard practice over the past three decades
and is now considered a mature field \citep{Davis1985,Springel2006}.
Indeed, computer cosmological simulations have undeniably become a powerful tool of modern cosmology, but they have also introduced
a major obstacle that makes their use for model testing difficult. Namely, for each set of initial conditions described by some
chosen values of cosmological parameters and for each specific GR/MG model, one needs to run a separate dedicated computer simulation.
In addition, due to intrinsic non-linearity of the screening mechanism employed by MG, dedicated numerical codes are also needed
for running non-standard gravity simulations. Covering the non-linear regime of MG structure formation with sufficient resolution
is also much more computationally expensive than standard GR simulations. These difficulties made the study of MG
theories very non-trivial and challenging. Nonetheless, we need to stress out that it is absolutely necessary to use dedicated
MG simulations for rendering theoretical predictions for growth rate of structures. This is because both the complicated nature
of galaxy formation physics and the non-linear character of screening mechanisms introduce various degeneracies and biases \citep{mog_gadget}.
In order to formulate self-consistent predictions for MG signatures in an observable one needs to assess, understand and disentangle
various systematic effects from the ones that are purely a result of altered dynamics stemming from additional MG degrees of freedom \citep{Bull2016}.
The situation is not as hopeless as one might expect when considering the very rich phenomenology of many MG theories.
Many different screening mechanisms can be categorised as either being screened by gravitational potential or density.
The first is a broad category, where the screening suppresses the fifth force by either making locally the scalar field very
massive ({\it the chameleon mechanism})\cite{kw2004}, forcing a small value of scalar field (the {\it symmetron } fields)\cite{dlmw2012},
or suppressing the strength of coupling of the scalar field ({\it the dilaton fields}) to matter \cite{bbdls2011} in high density regions.
The flagship example of the latter screening category is the {\it Vainsthein mechanism}\cite{Vainshtein:1972sx}, where due to higher-order
derivative interactions in the vicinity of massive objects the scalar field fluctuations attain significant kinetic terms and thus decouple
from the matter fields.
In the past decade the non-linear gravitational clustering in MG theories has received much attention
\citep[see {\frenchspacing\it e.g.} reviews in][]{cfps2011,Koyama2016,Bull2016}.
With the chameleon screening employed by the scalar-tensor theory of $f(R)$ being by far the most thoroughly inquired theory \cite{hs2007,Sotiriou:2008rp}.
It has been shown that this class of theories exhibit very interesting behaviour at late times predicting strong observational
features in anisotropic galaxy
redshift space clustering \cite{jblkz2012}, matter and velocity power spectra \cite{LiHell2013}, halo and galaxy dynamics \cite{Hellwing2014PhRvL},
real and redshift space halo clustering \cite{halosvoids_fr,Arnalte-Mur2017} and higher-order clustering statistics \cite{Hellwing2013}.
On the other hand the gravitational instability mechanism with Vainshtein
screening was probed to a much lower extent, as most of the studies concerned either only two-point matter clustering statistics
\cite{schmidt2009,Li2013,Barreira2013},
some basic morphological LSS features \cite{Falck2014, Falck2015}, screening of dark matter halos \cite{Falck2015} or simple halo dynamics \cite{Hellwing2014PhRvL}, and further those studies
were mostly based on simulations
with both limited resolution and volume. The results presented in this paper are aimed to amend this situation and provide comprehensive
insights into nature of gravitational instability within Vainshtein mechanism type of screened theories.
As our test-bed we chose to take the so-called {\it normal branch of Dvali-Gabadadze-Poratti} model (nDGP), a higher-dimensional gravity
brane-world model that employs Vainshtein mechanism \cite{DGP2000}. The choice of nDGP as our test case for Vainshtein is motivated by the fact
that this model, in contrast to so-called self-accelerating branch (sDGP) \cite{sDGP} or other models
like covariant Galileons \cite{Nicolis:2008in} that employ Vainsthein,
is fully consistent with LCDM's expansion history as precisely determined by modern observations. However, we ought to mention
that there is a major drawback here related to nDGP. In order achieve its compatibility with the data, it still requires some
small amount of DE \cite{schmidt2009}. Hence, its attractiveness as an alternative explanation of accelerated expansion is largely diminished.
This being said, we want to stress that our main intention here is not to present and test a new fundamental theory of gravity
as an alternative explanation to cosmic acceleration, but rather to study phenomenology of a large-class of models by assessing
the impact on cosmic structure formation of a fifth-force moderated by Vainshtein screening.
In addition to the gravitational instability, the second conventional assumption of the standard model for the formation
of structures is that the primordial density fluctuations were described by Gaussian random field statistics.
The structure formation theory is bound by providing an appropriate description of the initial power spectrum of the density fluctuations.
Now, by taking the main ingredient to be cold DM, one obtains a hierarchical model of structure formation,
where the clustering and gravitational collapse proceeds from small to large scales. For power law perturbation spectra, $P\propto k^{n_{\rm s}}$,
this is always true, provided $n_{\rm s}>-3$. In the nDGP gravity we assume that all elements of the structure formation model are the same as in LCDM,
except for the non-linear modifications to gravity surfacing through the Vainshtein screened fifth force. As the enhanced fifth-force dynamics
become important at different cosmic scales and epochs, it will lead to departures from the standard well established and tested hierarchical clustering
paradigm of GR. This changes in turn should be imprinted in hierarchical clustering statistics of matter and DM haloes.
For the case of MG models with different screening mechanisms it was shown that their modified dynamics in most cases
leave strong imprints on the matter clustering hierarchy, especially in the higher order moments \cite{Hellwing2010,Hellwing2013}.
Since the hierarchical clustering as a main prediction of the gravitational instability scenario was so thoroughly tested in
the case of GR and some other MG models, it is now imperative to conduct such studies also for the Vainshtein class of fifth-force cosmologies.
This defines the main goal of the analysis presented in this paper.
This paper is organised as follows: in \S\ref{sec:nDGP_desc} we provide a general description of the physical properties of the nDGP model,
\S\ref{sec:simulation} covers the details of N-body simulations used in this study. In Sec. \S\ref{sec:hierarchical_theory} we present
the main features of hierarchical clustering theory, this is followed by the main results of our analysis presented in \S\ref{sec:results}.
Finally, we conclude in \S\ref{sec:conclusions}.
\section{nDGP gravity model and simulations }
\label{sec:nDGP_desc}
\subsection{Model}
We consider the normal branch Dvali Gabadadze Porrati (nDGP) braneworld model \cite{DGP2000} that has the same expansion
history as LCDM. Under the quasi-static perturbations, the Poisson equation is given by~\cite{Koyama2007}
\begin{eqnarray}
\nabla^2 \Psi & =& 4 \pi G a^2 \rho \delta + \frac{1}{2} \nabla^2 \varphi, \nonumber\\
\nabla^2 \varphi & + & \frac{r_c^2}{3 \mathcal{B}(a) a^2}
[(\nabla^2 \varphi)^2 -(\nabla_i \nabla_j \varphi)(\nabla^i \nabla^j \varphi)]
= \frac{8 \pi G a^2}{3 \mathcal{B}(a)} \rho \delta, \nonumber\\
\label{eq:phievo1}
\ea
where $\Psi$ is the Newtonian potential, $G$ is the Newton constant, $a$ is the scale factor and $\varphi$ is an additional scalar field in the model.
The function $\mathcal{B}(a)$ is given by
\begin{equation}
\mathcal{B}(a) = 1 + 2 H r_c \left(1+ \frac{\dot{H}}{3 H^2} \right),
\label{eq:beta}
\end{equation}
where $r_c$ is the cross-over scale above which gravity becomes five-dimensional. Now by defining $\Omega_{r_c} \equiv1/ (2 r_c H_0)^2$
we obtain a single parameter that defines any given nDGP model. Here, $H_0$ is the present-day value of the Hubble parameter and we
have followed a usual convention with $c=1$.
On linear scales, where we can ignore the non-linearities of the scalar field, the fifth-force enhances the Newtonian gravity.
This can be quantified by means of an effective Newton constant, which will be given by $G_{eff} = G \{1+ 1/[3 \mathcal{B}(a)]\}$.
Note that $\mathcal{B}(a)$ is positive and decreasing in time, so
the growth of structure formation is always enhanced in this model and the enhancement becomes larger at late times. For
a larger $\Omega_{r_c}$, $\mathcal{B}(a)$ is smaller so the enhancement of gravity is stronger. On small scales, the non-linearity
of the scalar fields by virtues of the Vainshtein screening suppresses the coupling between the scalar field and matter.
Thus the effective gravity approaches the GR case, $G_{eff}\rightarrow G$.
This class of models experience rich phenomenology, as the Vainshtein mechanism is intrinsically non-linear.
\subsection{Simulations}
\label{sec:simulation}
\begin{table}[h!]
\caption{\label{tab:halo_samples} The number densities and the corresponding cut-off halo masses
for our halo sample at $z=0$.}
\medskip
\begin{tabular}{cccc}
\hline
\hline
number density & GR $M_{min}$ & nDGPa $M_{min}$ & nDGPb $M_{min}$\\
$h^{-3}\,{\rm Mpc}^3$ & $h^{-1}{\rm M}_{\odot}$ & $h^{-1}{\rm M}_{\odot}$ & $h^{-1}{\rm M}_{\odot}$ \\
\hline
\hline
$1.4\times10^{-3}$ & $1.6\times10^{12}$ & $1.6\times10^{12}$ & $1.72\times10^{12}$\\
\hline
$9\times10^{-4}$ & $1.93\times10^{12}$ & $1.94\times10^{12}$ & $1.96\times10^{12}$\\
\hline
$1\times10^{-4}$ & $2.54\times10^{13}$ & $2.57\times10^{13}$ & $2.7\times10^{13}$\\
\hline
\hline
\end{tabular}
\end{table}
In our analysis we will use a set of N-body simulations run for LCDM and two nDGP models.
These simulations were conducted using the AMR code ECOSMOG~\citep{ECOSMOG}.
The background cosmology is taken from WMAP9~\citep{WMAP9}: $\Omega_m = 0.281$, $h=0.697$, and $n_s=0.971$.
The box length is $1024$Mpc/h with $1024^3$ dark matter particles used and a starting redshift of $49$.
The initial conditions were generated using {\tt MPGrafic}\footnote{Available at \url{http://www2.iap.fr/users/pichon/mpgrafic.html}}.
Using $z_{ini}=49$ assures that the system will be evolved for time long enough in order to wipe-out any
transients effects that are affecting higher moments of initial particle distribution displaced
by Lagrangian methods \cite{Scoccimarro1998,transients1,transients2}
This design sets the resulting mass resolution at $m_p\cong7.8\times10^{10}M_{\odot}h^{-1}$ and the Nyquist fluid approximation
limit of $k_{Nyq}\cong\pi$ h/Mpc.The most refined AMR grid were at the level 16, setting a maximal force
resolution at $\epsilon=1024/2^{16}=0.015$Mpc/h.
The LCDM run will constitute our fiducial GR-reference point, in addition we simulate two nDGP models implementing
two values of the cross-over scale parameter: $\Omega_{rc} = 0.0124, 0.438$. The first model (nDGPa)
is characterised by only a mildly enhances growth of structure history and we will treat is as a borderline case.
The latter model (nDGPb) with the large value of $\Omega_{rc}$ parameter should experience
sizeably larger differences from the GR case, fostering a more realistic detection prognosis.
We evolve the dark matter density and velocity field from the initial redshift to the present epoch, selecting and saving
for the analysis snapshots taken at three specific epochs: $z=0,0.5$ and $1$. To identify DM haloes in our snapshots
we resort to the {\tt ROCKSTAR} FOF phase-space halo finder \cite{Behroozi2013}. We keep all the haloes and subhaloes with 20 or more
particles for further analysis. As a main proxy for halo mass we settle down for a commonly used virial mass
$M_{200}\equiv 4/3\pi R_{200}^3 200\times\rho_{c}$. Where $R_{200}$ is a boundary radius at which the spherically
averaged matter density enclosed inside is equal to 200 times the Universe critical density $\rho_{c}$.
At $z=0$ for all three models we find a very similar number of $'\sim1.5\times10^6$ distinct haloes and subhaloes,
with a very similar satellite fractions of $\sim6\%$ for GR and mild nDGP models and $7\%$ for the strong nDGP model.
Once we constructed our halo catalogues we split them into samples of a fixed number density. This allows us to
approximate in a very simplistic way mock galaxy samples in the spirit of abundance matching \cite{Vale2004}. For our richest
and most complete sample we pick centrals+satellite with a number density of $\av{n}=1.4\times10^{-3}(h^{-1}\,{\rm Mpc})^{-3}$.
We also consider sparser centrals-only samples with an effective number densities of $9$ and $1\times10^{-4}(h^{-1}\,{\rm Mpc})^{-3}$.
The specific details like the minimum halo mass cut-off for each sample are given in Tab.~\ref{tab:halo_samples}.
\begin{figure}
\includegraphics[angle=0,width=0.48\textwidth]{fig/errors_compare.pdf}
\caption{The contributions of the cosmic variance and the shot noise to the correlation functions
estimators we use in this work.}
\label{fig:errors}
\end{figure}
\section{Hierarchical clustering}
\label{sec:hierarchical_theory}
Here, our main focus will be on the matter and halo density fields, which we will describe in terms of the density contrast,
a quantity that measures local departure from a background uniform density. Thus we define
\begin{equation}
\label{eqn:de_contrast}
\rho(\vec{x},t) = \langle\rho(t)\rangle\left[1+\delta(\vec{x},t)\right]\,,
\end{equation}
where $\langle\rho(t)\rangle$ is the average background density of given tracers (matter or haloes),
and $\delta(\vec{x},t)$ (the density contrast) characterises local deviations from the background.
Now, the full statistical information about the density field and all its correlation's properties is encoded
in the density probability distribution function $p(\delta)$.
Cosmologies employing cold dark matter spectra for initial Gaussian fluctuations
exhibit so-called {\it hierarchical clustering}. Here, the first structures to emerge
from expanding smooth background are tiny haloes corresponding to the smallest density peaks with sizes just above the dark matter
streaming scale \cite{1980Peebles,Davis1985}. As the Universe expands, larger and larger density perturbations reach the turn-around
radius and start to collapse to form bigger structures. Simultaneously, some small haloes that formed earlier
become satellites of bigger and younger structures and eventually sink and merge into them. Thus structure formation proceeds from
small to large scales. In this picture the gravitational interactions during the evolution drive away the density probability
distribution function from its initial Gaussian characteristics. This is reflected by the growth of higher-order
moments of the density field, which measure departure from Gaussianity \cite{Juszkiewicz1993,skew_cur_pt,grav_inst_pt}.
The first and most basic statistics that characterise clustering is the two-point correlation function: $\xi(r)$.
This is defined as the excess probability (with respect to a Poisson process) of finding
two matter particles (or haloes) contained in two volume elements d$V_1$ and d$V_2$ at a distance $r$ (see {\frenchspacing\it e.g.} \citep{1980Peebles}):
\begin{equation}
\label{eqn:2pcf-definition}
\textrm{d}P_{12}(r) \equiv \bar{n}^2[1+\xi(r)]\textrm{d}V_1\textrm{d}V_2\,,
\end{equation}
where $\bar{n}$ is the mean matter (halo) number density. In general $\xi(r)$ characterises the strength
of matter (halo) clustering across cosmic scales and epochs. However, as mentioned above the gravitationally induced
evolution gives rise to significant departures of matter and halo density distribution functions from a normal one .
For a non-Gaussian $p(\delta)$, the knowledge of only the second moment is no longer enough to fully characterise the field,
as Wick's theorem no longer holds.
In this context, the so-called reduced moments or cumulants of the distribution function $p(\delta)$ are especially useful.
The $n$-th cumulant of the distribution function $p(\delta)$ is defined by recursive relation to the $n$-th
moments. This relation can be expressed by the cumulant generating function \citep[eg.][]{Lokas_kurt}
\begin{equation}
\label{eqn:gen_func}
\ad{n}_{\rm c}\equiv M_n = {\partial^n {\rm ln}\,\av{e^{t\delta}}\over \partial t^n}\bigg|_{t=0}\,.
\end{equation}
The cumulants now can be expressed in terms of the central moments, specifically, for the first five cumulants we have \citep{ber1994,GaztanagaAPM94}
\begin{eqnarray}
\label{eqn:cumulants}
\ad{}_{\rm c} &=& 0,\,\,\textrm{(the mean)}\nonumber\\
\ad{2}_{\rm c} &=& \ad{2}\equiv\sigma^2,\,\,\textrm{(the variance)}\nonumber\\
\ad{3}_{\rm c} &=& \ad{3},\,\,\textrm{(the skewness)}\nonumber\\
\ad{4}_{\rm c} &=& \ad{4} - 3\ad{2}_{\rm c}^2,\,\,\textrm{(the kurtosis)}\nonumber\\
\ad{5}_{\rm c} &=& \ad{5} - 10\ad{3}_{\rm c}\ad{2}_{\rm c}\,.
\ea
Here the $n$-th cumulant is obtained by taking the value of the $n$-th moment of the distribution $p(\delta)$
and removing from it the contributions from all the decompositions of a set of $n$ points in its subsets
multiplied (for each decomposition) by the cumulants corresponding to each subset \citep{ber1994}.
For a field described by a normal distribution with a zero mean all cumulants, but the variance $\sigma^2$, vanish. In the standard
random field theory, the first two non-disappearing connected moments above variance have special meaning as they describe specific
shape departures of the distribution function from a Gaussian. The skewness describes the asymmetry of the
distribution function and the kurtosis details the flattening of tails with respect to a Gaussian. Higher-order
moments characterise even more convoluted aberrations of the distribution function shape.
Various studies of the higher order cumulants of the cosmic density fields have revealed that for the case of initial adiabatic Gaussian
density perturbations described by a power-law spectrum, the gravitational instability mechanism produces a quasi-Gaussian clustering hierarchy
of connected moments \cite{ber1992}. Moreover, in the linear and weakly non-linear regime this hierarchy is preserved
by gravitational time evolution \cite{Fry1984a,Fry1984b}. Commonly, this hierarchy is described in terms of the so-called {\it hierarchical scaling} relations:
\begin{equation}
\label{eqn:hier_amplitudes}
S_n\equiv {\ad{n}_c\over\ad{2}_c^{n-1}}\,.
\end{equation}
Here $S_n$ denote {\it reduced cumulants} (also called hierarchical amplitudes), and for unsmoothed fields these are constant. In reality however,
one always deals with density fields that are smoothed at some given scale. So what can be actually measured from observations and compared
to predictions from N-body simulations are actually the {\it volume averaged} $n-point$ moments (correlation functions) of the underlying
smoothed density field $\delta_R$. Now, if we define a smoothed field as
\begin{equation}
\label{eqn:smoothed_de}
\delta_R(\vec{x})=\int\delta(\vec{x'})W(|\vec{x}-\vec{x'}|/R_w)\textrm{d}^3x'\,,
\end{equation}
with $W(x/R_w)$ being a spherically symmetric smoothing window. For practical reasons it is convenient to limit our analysis to only
one window function that is easily applied to numerical simulations.
Namely, we will deal with density fields smoothed over a ball of radius
$R=R_w$ with a window function normalised to unity
\begin{eqnarray}
\int W(y)\textrm{d}^3y = 1\quad \textrm{and}\\\nonumber
\int W(y)y^2\textrm{d}^3y = R_{\rm w}^2\,.
\ea
This filter function describes what is commonly known as the {\it top-hat} filtering.
Now the volume averaged n-point connected moments can be defined as
\begin{flalign}
\label{eqn:xi_volume_av}
&\overline{\xi}_n(R)\equiv\av{\delta_R^n}_{\rm c} =&&\\\nonumber
&= \int \textrm{d}^3 x_1\ldots\textrm{d}^3 x_n\xi(\mathbf{x_1}\ldots\mathbf{x_n})W(x_1/R_{\rm TH})\ldots W(x_n/R_{\rm TH})&&.
\end{flalign}
Classical PT for GR cosmology predicts that the hierarchical amplitudes computed for smoothed density fields should be weakly monotonic
decreasing functions of the smoothing scale $R$ \cite{BCGS_book}. This also has been also widely confirmed by
comparisons with N-body simulations \cite{npoint_omega_cdm}.
The scale-dependence of the reduced cumulants is a function of increasingly complicated combinations of growing order derivatives
of the matter variance taken effectively at a given smoothing scale. For example for the reduced skewness and kurtosis this was shown to be \citep{ber1994}
\begin{eqnarray}
\label{eqn:s3s4_gamma}
S_3 = {34\over 7} + \gamma_1\,,\nonumber\\
S_4 = {60712\over 1323} + {62\over 3}\gamma_1 + {7\over 3}\gamma_1^2+{2\over 3}\gamma_2\,,
\ea
with $\gamma$ factors enumerating the variance derivatives
\begin{equation}
\label{eqn:pt_gamma}
\gamma_n(R)\equiv {\textrm{d}^n\textrm{log}\sigma^2(R)\over \textrm{d}\textrm{log}^n R}\,.
\end{equation}
In the Einstien-De Sitter Universe ({\frenchspacing\it i.e.}{} with $\Omega_m=1$) the skewness and kurtosis are constants, as
effectively on all scales $\gamma_n\rightarrow 0$. In LCDM however,
the $\gamma$ factors represent corrections to the reduced cumulants due to varying with scale shape of the matter power spectrum
\citep[see {\frenchspacing\it e.g.}][]{Juszkiewicz1993,BCGS_book}.
In the nDGP model the power spectrum is modified at non-linear scales by the fifth-force dynamics \cite{schmidt2009}. Thus we can expect that this
effect should also be reflected in higher-order moments. This effect was found to be significant for other MG models like $f(R)$ \cite{Hellwing2013}
and ReBEL \cite{Hellwing2010}, where the degree of deviation from the GR prediction increased with the order of the cumulant.
We can expect that this well understood picture can get complicated, if we consider haloes rather then the matter density fields. DM haloes
are biased tracers of the underlying smooth density fields. Therefore, for the case of reduced cumulants the hierarchical scaling
relations will not only be simple functions of the power spectrum derivatives, but also will be described by
higher-order hierarchical biasing \cite{FG1993,npoint-halos}.
In general, we can expect that the higher-order bias can be a complicated function of scale, time and halo mass. However, as in nDGP there are
no environmental effects, the higher-order bias should take roughly the same time and scale dependence as in GR, with only the halo mass
being the major difference driver. A detailed analysis of the hierarchical biasing in nDGP (and in MG theories in general)
is well beyond the scope of this paper, as it would merit a whole dedicated study. Thus, we leave it for future work, and will not discuss further
the bias issues in the current paper.
\subsection{PT Variance and Skewness estimators}
\label{subsec:PT_formalism}
To test our numerical results we will benchmark them against the estimators available in the context of the Eulerian perturbation theory.
Our approach will be to use a perturbative formalism to calculate the variance and reduced skewness of the matter fields
\cite{1980Peebles,Goroff:1986ep,Juszkiewicz1993}.
Here, the additional non-linear evolution of the density contrast is modelled by also tracing the contribution from the peculiar velocity ($v_p(\vec{x})$)
divergence field:
\begin{equation}
\label{eqn:expansion_scalar}
\theta(\vec{x})={\nabla\cdot v_p(\vec{x})\over aH(a)f}\,\,,
\end{equation}
where $H(a)$ is the Hubble function and $f\equiv \textrm{d}{\rm ln}\,{D_+}/\textrm{d}{\rm ln}\,{a}$ is the logarithmic growth rate. Here, the scale factor $a$ is used
as a mean cosmic time variable and $D_+$ is the growing mode of the linear perturbation theory solution.
For the sake of brevity and simplicity it is very convenient to express all relevant quantities and work in Fourier space.
Thus, we further define
\begin{equation}
\label{eqn:delK}
\delta(\vec{k}) \, \equiv \, (2\pi)^{-2/3}\,\int \delta(\vec{x})\,e^{-i{\vec{k} \cdot \vec{x}}}\,d^3{\vec{x}}\,,
\end{equation}
and
\begin{equation}
\label{eqn:delK}
\theta(\vec{k}) \, \equiv \, (2\pi)^{-2/3}\,\int \theta(\vec{x})\,e^{-i{\vec{k} \cdot \vec{x}}}\,d^3{\vec{x}}\,,
\end{equation}
respectively for the density contrast and the velocity divergence fields.
The classical approach is to solve the continuity and Euler equations for $\delta(\vec{k})$ order by order
(for a comprehensive review see {\frenchspacing\it e.g.}~\cite{BCGS_book}). The evolution equations are expressed below as follows:
\begin{widetext}
\begin{eqnarray}
\label{eqn:pt_evolution1}
a \frac{\partial \delta(\vec{k})}{\partial a}+\theta(\vec{k}) =
-\int\frac{d^3\vec{k}_1d^3\vec{k}_2}{(2\pi)^3}\delta_{\rm D}(\vec{k}-\vec{k}_1-\vec{k}_2)
\alpha(\vec{k}_1,\vec{k}_2)\,\theta(\vec{k}_1)\delta(\vec{k}_2),
\label{eq:Perturb1}
\ea
\begin{eqnarray}
\label{eqn:pt_evolution2}
a \frac{\partial \theta(\vec{k})}{\partial a}+
\left(2+\frac{a H'}{H}\right)\theta(\vec{k})
-\left(\frac{k}{a\,H}\right)^2\,\Phi(\vec{k})=
-\frac{1}{2}\int\frac{d^3\vec{k}_1d^3\vec{k}_2}{(2\pi)^3}
\delta_{\rm D}(\vec{k}-\vec{k}_{12})
\beta(\vec{k}_1,\vec{k}_2)\,\theta(\vec{k}_1)\theta(\vec{k}_2)
\label{eq:Perturb2}
\ea
\end{widetext}
where the prime denotes a derivative w.r.t the scale factor $a$, $\vec{k}_{12} = \vec{k}_1+\vec{k}_2$, $\delta_D$
is the Dirac delta function and the mode coupling kernels, $\alpha$ and $\beta$, are given by
\begin{eqnarray}
\label{eqn:pt_kernels}
\alpha(\vec{k}_1,\vec{k}_2)=1+\frac{\vec{k}_1\cdot\vec{k}_2}{|\vec{k}_1|^2},
\quad \beta(\vec{k}_1,\vec{k}_2)=
\frac{(\vec{k}_1\cdot\vec{k}_2)\left|\vec{k}_1+\vec{k}_2\right|^2}{|\vec{k}_1|^2|\vec{k}_2|^2}.
\label{alphabeta}
\end{eqnarray}
The linearised equations are given by setting $\alpha=\beta=0$ and expanding the Poisson term, $k^2\Phi$, to linear order
in the density perturbations. Gravity effects the evolution of the perturbations through the Newtonian potential $\Phi$
and one can encode any modifications to gravity there. The general order solutions for the density contrast can then be expressed as
\begin{align}
\delta_n(\vec{k}; a) =& \int d^3\vec{k}_1...d^3 \vec{k}_n \delta_D(\vec{k}-\vec{k}_{1...n}) \nonumber \\
& \times F_n(\vec{k}_1,...,\vec{k}_n; a) \delta_0(\vec{k}_1)...\delta_0(\vec{k}_n) \label{nth1}
\end{align}
where the $n^{th}$ order kernel $F_n$ is obtained by recursively solving Eqn.(\ref{eq:Perturb1}) and Eqn.(\ref{eq:Perturb2})
at $n^{th}$ order in the density contrast and velocity divergence. $\delta_0$ is the initial density contrast
which we assume is Gaussian. We solve for these kernels numerically at linear and second
order within nDGP gravity and GR. This is done using the tool described in \cite{Bose:2016qun}.
To get the smoothed fields we simply perform a Fourier transformation on Eqn.(\ref{eqn:smoothed_de})
\begin{equation}
\label{eqn:smothed_de_k}
\delta(R) = \int\frac{\textrm{d}^3k'}{(2\pi)^3}\delta(\vec{k'})W(k'R)\,,
\end{equation}
where $W(k'R)$ is the Fourier transform of the smoothing function. For top-hat smoothing this is simply
\begin{equation}
\label{eqn:fourier_top_hat}
W(k'R) = \frac{3}{(k'R)^3}(\sin(k'R)-k'R\cos(k'R))
\end{equation}
The variance of the smoothed fields is then
\begin{equation}
\label{eqn:sig2_pt_estimator}
\sigma^2(R) = \int \frac{\textrm{d}^3k}{(2\pi)^3} W(Rk)^2 F_1(\vec{k};a)^2 P_L(k)
\end{equation}
where the initial linear power spectrum, $P_L(k)$ is defined as
$\delta_D(\vec{k}+\vec{k}')P_L(k) = \langle \delta_0(\vec{k}) \delta_0(\vec{k}') \rangle$.
As the nDGP gravity effects at very early times are negligible, for both models we can use
the LCDM initial linear power spectrum. This can be calculated for a given cosmology using a Boltzmann
solver code such as CLASS \cite{Blas:2011rf}.
On top of the variance we also consider the third moment, expressed as reduced skewness.
By using the properties of Gaussian random fields we have \citep{1980Peebles} $\ad{3}_c = 3 \langle (\delta^{(1)})^2 \delta^{(2)}\rangle$.
In terms of the generalised density contrast kernels and linear power spectrum this can be expressed as
\begin{align}
\ad{3}_c = 3&\int \frac{\textrm{d}^3k_1}{(2\pi)^3} \frac{\textrm{d}^3k_2}{(2\pi)^3} P_L(k_1)P_L(k_2) F_1(k_1;a) W(Rk_1) \nonumber \\
\times & \big[ 2W(Rk_2)W(Rk_{12})F_1(k_2;a) F_2(\vec{k_1},\vec{k_2};a) \nonumber \\ &+ W(0)W(Rk_1)F_1(k_1;a)F_2(\vec{k_2},-\vec{k_2};a) \big] \label{eq:3point}
\end{align}
Note that in GR and nDGP the last term (the $3^{rd}$ line) of Eqn.(\ref{eq:3point}) vanishes. Finally, performing some of the integrals we
obtain the following expression for the reduced skewness
\begin{widetext}
\begin{eqnarray}
\label{eqn:s3_pt_estimator}
S_3 = 3\int dk_1dk_2du k_1^2 k_2^2 P(k_1)P(k_2)F_1(k_1;a) F_1(k_2;a) F_2(k_1,k_2,u;a)W(Rk_1)W(Rk_2)W(Rk_{12}) \Bigg/ \nonumber \\
\bigg[ \int dk k^2 P(k)F_1(k;a)^2W(kR)^2\bigg]^2
\ea
\end{widetext}
We will use the estimators of Eqn.(\ref{eqn:sig2_pt_estimator}) and Eqn.(\ref{eqn:s3_pt_estimator}) computed for the initial power spectrum
sued in our N-body simulations, together with $F_1$ and $F_2$ evolution kernels expressed specifically for nDGP and GR
(see \cite{Bose2017} for details) as our PT prediction for the skewness and variance of the matter density field.
\subsection{Numerical methods}
\label{subsec:numerical_methpds}
We will specifically consider matter and halo auto-correlation function and also volume averaged moments. For the 2-point correlation
function (2PCF) we will use a simple estimator
\begin{eqnarray}
\label{eq:2p_estimator}
\xi(r) = {DD(r) \over N \bar{n} v(r)} -1 \, ,
\ea
where $DD(r)$ is the number of pairs of tracers with separation in the range $[r, r+\Delta r]$, $N$ is the total
number of tracers in the sample, $\bar{n}$ is their number density, and $v(r)$ is the volume of a spherical shell
of radius $r$ and width $\Delta r$. As we will deal only with N-body simulation data the sample selection functions in all cases
are complete, isotropic, and homogeneous. In addition, because we have periodic boundary conditions, there are no edge effects.
This makes the estimations of $\xi(r)$ for matter and halo samples straightforward. We rely on a numerically efficient publicly
available package {\tt CUTE}\footnote{Available here \url{http://members.ift.uam-csic.es/dmonge/CUTE.html}} \cite{cute}.
For the case of volume averaged moments we adopt the fast method presented in \cite{Hellwing2013}. Here the density field for a given sample
is estimated using {\it the Delaunay Tesselation Field Estimator} method (DTFE)\citep{sv2000,vs2009} implemented in the publicly available
software {\tt DTFE} provided by \citep{cv2011}. Once the density field on a regular grid with $N_g$ cells have been constructed
we proceed to apply a range of top-hat smoothing (using a FFT method) and calculate the central moments for smoothed fields:
\begin{eqnarray}
\label{eqn:central_moments_numerical}
\av{\delta_R^n} = {1\over N_{\rm g}}\sum_{i}^{N_{\rm g}}\left(\delta_R^i - \av{\delta_R}\right)^n\,.
\ea
Then we proceed to obtain connected moments and reduced cumulant using the relations of Eqn. (\ref{eqn:cumulants})-(\ref{eqn:xi_volume_av}).
In the general case our correlation function estimators will be affected by two sources of errors : the shot noise due to sparse sampling
and classical cosmic variance. Since we have only one realisation for the initial perturbation phases, we adopt a standard Poisson estimator
for the cosmic variance error, which here is a function of the number of independent density modes present in the simulation box at a given scale.
To estimate the shot noise we use the same estimator with the number counts per bin as the sampling indicator. In the case of our DM samples
the shot noise for nearly all scales ({\frenchspacing\it i.e.}{} $R>2h^{-1}\,{\rm Mpc}$) is sub-dominant to the cosmic variance contribution.
For haloes we find that the cosmic variance is dominant at large scales of $\stackrel{>}{{}_\sim} 10-30h^{-1}\,{\rm Mpc}$
(depending on the halo sample number density), while at smaller scales
the shot noise dominates. We illustrate our error budget contributions in the Fig.~\ref{fig:errors}
\section{Results}
\label{sec:results}
\begin{figure}
\includegraphics[angle=0,width=0.48\textwidth]{fig/variance_compare.pdf}
\caption{Comparison of matter density variance for $z=0$ estimated from N-body simulations (points) with
the PT predictions of Eqn.(\ref{eqn:sig2_pt_estimator}). The lower panel shows the fractional difference of both
nDGP models taken w.r.t the GR case. The shaded region illustrate the cosmic variance error for the GR fiducial case.}
\label{fig:variance_PT}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=0.48\textwidth]{fig/skewness_compare.pdf}
\caption{Same as the previous figure, but for the reduced density skewness $S_3$ compared against the estimator
of Eqn.~(\ref{eqn:s3_pt_estimator}).}
\label{fig:skewness_PT}
\end{figure}
In the following we present the results of our analysis, where we use two-point correlations functions
alongside the higher-order reduced cumulants as well as the density distribution functions themselves,
to study the gravitational instability mechanism and the hierarchical clustering it induces for the matter and halo
density fields estimated for a range of scales at three distinctive cosmic epochs $z=0, 0.5$ and $1$.
\subsection{Dark Matter}
\label{subsec:dm}
We start by focusing on the dark matter density field. First, we want to test the implementation of our
numerical moments estimators. We do this by comparing our N-body measurements for the variance, $\sigma^2$,
and the reduced skewness, $S_3$, with the PT predictions. The results of this procedure are shown in
Figs.~\ref{fig:variance_PT} and ~\ref{fig:skewness_PT}. Both figures illustrate a strikingly good match
of PT and N-body predictions, which agree nearly perfectly down to $R=10h^{-1}\,{\rm Mpc}$ for variance and
down to $\sim15h^{-1}\,{\rm Mpc}$ for the reduced skewness. This is the case for all three models studied here.
Below these scales the non-linearity in the density fields become severe and PT starts to significantly
underpredict the moments. However, the good agreement at scales that are already mildly non-linear
reassures that our numerical estimators are accurate and unbiased.
Interestingly, we can notice that
when we focus only on the deviations of nDGP models from the fiducial GR predictions, as quantified by
fractional differences ratios ($\Delta X= X_{nDGP}/X_{GR}-1$) shown in the lower panels, the PT predictions
remain surprisingly accurate down to the smallest scales we consider ($1h^{-1}\,{\rm Mpc}$).
While the PT fails to capture all the effects of the non-linear gravitational evolution at small-scales
and underpredicts the absolute amplitudes of the cumulants, this failure has a universal character
for all our models. This behaviour most likely reflect the fact that both of our nDGP models exhibit only
very mild deviations from GR gravitational dynamics.
\subsubsection{1-point statistics}
\label{subsec:dm:pdfs}
\begin{figure*}
\includegraphics[angle=0,width=0.96\textwidth]{fig/pdf_compare.pdf}
\caption{The distribution functions for our models computed for density fields smoothed at $R=0.5, 25$ and $50h^{-1}\,{\rm Mpc}$ (panels from left to right).
For each panel the GR case is marked by solid blue line, nDGP ${\Omega}_{rc}=0.0124$ is dashed orange and nDGP ${\Omega}_{rc}=0.438$
is marked by dot-dashed line. In each upper panel three groups of lines correspond to three redshifts $z=0,0.5$ and $1$, where the data for latter two
was scaled down for clarity. For each case the smaller bottom panel illustrates the $p(\delta)$ ratio taken w.r.t the GR case only for $z=0$ case.
Mark the change to linear scaling for the $\delta+1$ axis in the most left panel.}
\label{fig:pdf_compare}
\end{figure*}
\begin{figure}
\includegraphics[angle=-90,width=0.48\textwidth]{fig/all_pk_compare_nDGP.pdf}
\caption{The matter density power spectrum computed at $z=0$ for our fiducial GR model (solid line)
and two nDGP flavours (dotted and dashed-dotted lines). The shaded region illustrate the cosmic variance error.
The bottom panel illustrates the fractional difference of both MG models w.r.t. the GR case.}
\label{fig:matter_pk}
\end{figure}
It is very illustrative to begin our analysis by taking a close look at 1-point statistics of the cosmic density field. This is readily characterised
by 1D density probability distribution functions (PDFs), $p(\delta+1)/(\Delta\delta)$. For a primordial density perturbation field described by adiabatic
fluctuations these functions are Gaussian on all scales. For evolved fields the distribution functions develop exponential tails and at sufficiently
large scales can be described by log-normal distributions \cite{Bernardeau1995}. In all cases the linear and non-linear gravitational evolution
driving the structure formation
is encoded in the shape departures of the density $p(\delta)$ from a Gaussian.
To get a first intuition of the differences caused by modified nDGP dynamics we show the 1D density PDF's in Fig.~\ref{fig:pdf_compare}.
Here, the three panels from left to right correspond to three different top-hat smoothing lengths applied to our reconstructed density fields:
$R=0.5, 25$ and $50h^{-1}\,{\rm Mpc}$ respectively.
For all scales we show the PDFs obtained at three different epochs for $z=0, 0.5$ and $1$, where the lines corresponding to $z>0$
are tiered down for clarity. First, as expected by following the previous results from PT and N-body simulations
\cite{schmidt2009,Koyama2007}, we notice that the difference from the fiducial GR case is maximised at the present time.
Another interesting feature illustrated by Fig.~\ref{fig:pdf_compare} is the broadening of the PDFs towards low density ($\delta<0$) values. This feature
is present at all smoothing scales we consider. It clearly indicates that we can expect deeper density profiles of cosmic voids, which also should get larger,
as matter is more effectively evacuated towards surrounding higher density areas. While this is a very promising feature, and potentially could leave
an observable imprint in {\frenchspacing\it e.g.} void lensing potentials, we will not study this here further, leaving this interesting opportunity for a future work.
Additionally, we can observe that at larger smoothing scales the high-density tails of the nDGP distribution functions get significantly enhanced.
This feature reflects a simple conservation of mass, as the matter that was evacuated from the voids has to be deposited somewhere.
This is reflected in an increased virial mass of big haloes ({\frenchspacing\it i.e.}{} cluster and supercluster scales),
as was noted in \cite{halosvoids_fr,Hellwing2010boosting,Hellwing2013Halos}. Hence, this compensation
effect is much better appreciated at large smoothing scales, as $R_{th}=0.5h^{-1}\,{\rm Mpc}$ is way too small to encompass cluster and super-cluster density
perturbations. The general impression we can obtain by this analysis is that reduced cumulants of higher-order, like for example kurtosis,
should bear the nDGP signature at both small and large scales, while for the case of the PDF asymmetry measure, the skewness, we can expect that
the MG-induced differences should be much more pronounced at smaller scales. We will get back to this observation later, when we discuss
the higher-order density field statistics.
\subsubsection{2-point statistics}
\label{subsec:haloes:2pt_stat}
\begin{figure*}
\includegraphics[angle=0,width=0.48\textwidth]{fig/all_matter_xiN_compare.pdf}
\includegraphics[angle=0,width=0.48\textwidth]{fig/all_matter_xiN_compare_z1.pdf}
\caption{Hierarchy of n-point volume averaged correlation functions $\overline{\xi}_n(R)$ computed
for smoothed dark matter density fields at $z=0$ (left panel) and $z=1$ (right panel).
Shaded regions (hardly visible) mark a combination of sampling and cosmic variance errors for the GR cases.}
\label{fig:matter_Xins}
\end{figure*}
Studying the 1-point density field statistics was useful to obtain general insights into the differences between GR and nDGP
at various scales and density thresholds. However, in the observational context, the 2-point statistics used commonly to study
clustering in cosmological and extragalactic context should be much more illuminative. Fig.~\ref{fig:variance_PT} already hints
at nDGP's effects on the matter field variance. There we can observe that the second moment is enhanced by an approximately
constant factor on all scales. The amplitude of this increase takes $\sim4\%$ ($\sim20\%$) for $\Omega_{rc}=0.0124$ ($\Omega_{rc}=0.438$) models.
This is of course readily characterised by increased effective $\sigma_8\equiv\sqrt{\sigma^2}(8h^{-1}\,{\rm Mpc})$ values of the nDGP simulations,
which are higher by $2\%$ and $10\%$ for both nDGP models respectively. The variance of a smoothed density field is a position space
counterpart of a matter density power spectrum $P(k)=\langle \delta_k\delta_k^*\rangle$, which is a 2-point variance statistic for Fourier modes.
For completeness we show the power spectra alongside their fractional deviations in Fig.~\ref{fig:matter_pk}. The picture fostered here is
consistent with the behaviour exhibited by real-space variance, $\sigma^2$. The degree of the deviation from a GR fiducial value, as well as
the similar scale independence of this enhancement agree well with what we have observed earlier. However, thanks to the decomposition of
the density field into an orthogonal basis of density fluctuation modes, we can easier depict characteristic features of the Vainshtein screening
mechanism. Namely, at small scales, where the density field variance is dominated by the interiors of virialised haloes, we can start
to observe the suppression of the nDGP enhanced clustering due to effective screening of the fifth-force. This was studied in much greater
detail by other authors (see {\frenchspacing\it e.g.} \cite{MOG_comp}), but we can confirm that our simulations, up to their rather limited mass resolution, abide
to those findings (see also \cite{Falck2015}).
An important observation here consist of noting that as we have predicted, the simple 2-point statistics is not capable of fully capturing and
describing the otherwise complicated departures of the nDGP cosmic density field from the fiducial GR case, as was hinted by the PDFs shown in
Fig.~\ref{fig:pdf_compare}. This indicates a need to look at higher-order moments in order to get further insights into the gravitational
evolution of clustering in MG theories with the Vainshtein screening.
\subsubsection{Higher moments and hierarchical cumulants}
\label{subsec:dm:higher_moments}
\begin{figure*}
\includegraphics[angle=0,width=0.48\textwidth]{fig/all_matter_SN_compare.pdf}
\includegraphics[angle=0,width=0.48\textwidth]{fig/all_matter_SN_compare_z1.pdf}
\caption{Hierarchical clustering amplitudes $S_n(R)$ computed for smoothed dark matter density fields at $z=0$ (left panel) and $z=1$ (right panel).
Shaded regions (hardly visible) mark a combination of sampling and cosmic variance errors for the GR cases.}
\label{fig:matter_Sns}
\end{figure*}
In this section we look into the behaviour of higher-order moments of the density fields on a range of cosmic scales and epochs. We first take a look
at the averaged correlation functions themselves before focusing on the reduced cumulants. In Fig.~\ref{fig:matter_Xins} we plot the beautiful
correlation hierarchy formed by all the central moments from $\overline{\xi}_2$ up to $\overline{\xi}_9$ at $z=0$ (left panel) and $z=1$ (right panel).
For all our models the consecutive higher-order correlation functions exhibit tiered behaviour with the same monotonic scale-dependence.
Here, the picture broadly agrees with what we have observed already for the variance alone. Namely, the main effect of the nDGP
fifth-force is reflected in the $\overline{\xi}_n$'s amplitudes, which are increased by a nearly constant factor over the range of the scales.
We can also comment on a clear emergence of the effects induced by the finite-volume of our simulations. In agreement with well
establish theory\cite{Colombi1994}, the higher-order moments are more severely affected by the limited size of our simulations, which
is indicated by the fact that the moments from $\overline{\xi}_6$ and above could only be reliably estimated up to scales of $\sim 10-20h^{-1}\,{\rm Mpc}$.
The situation gets much more interesting when we contemplate the reduced cumulants, $S_n$'s, which are shown for the same two redshifts
in Fig.~\ref{fig:matter_Sns}. We can observe here the sensitivity of these statistics to the non-linear gravitational evolution,
as was established by many other authors \cite{Juszkiewicz1993,Lokas_kurt,grav_inst_pt,grav_inst_pt2,White1999}.
In contrast to the connected moments alone, the nDGP-induced deviations of the matter hierarchical
amplitudes are characterised by a strong scale-dependence. Here, at $z=0$, the deviations take significant values at small scales $R\simlt10h^{-1}\,{\rm Mpc}$, while
going back to the fiducial GR predictions at larger separations. This behaviour is nearly not present at $z=1$, indicating that
the features we observe here in $S_n$'s are induced by non-linear gravitational dynamics, which at late times
induce the change of the power spectrum shape at small scales, where the Vainsthein mechanism is ta play
(see {\frenchspacing\it e.g.}{} the high-$k$ tail in the lower panel of Fig.~\ref{fig:matter_pk}).
The ratios for reduced cumulants as described in Eqns.(\ref{eqn:hier_amplitudes}) and (\ref{eqn:pt_gamma}),
are sensitive by this shape change induced by modified structure formation. Hence we can confirm
the potential of the density field hierarchical amplitudes as potential probes of modified gravitational dynamics, a finding already
emphasised in clustering studies for a different class of MG theories \cite{Hellwing2010,Hellwing2013}.
\begin{figure*}
\includegraphics[angle=-90,width=0.98\textwidth]{fig/matter_s345_zevo.pdf}
\caption{Fractional differences from the GR taken across three epochs $z=0, 0.5, 1$ (rows of panels from top to bottom)
of the first three matter density reduced cumulants $S_3, S_4$ and $S_5$ (columns of panels from left to right).
Shaded regions illustrate the total error budget (cosmic variance + shot noise) on the ratios.}
\label{fig:matter_s3s4s5}
\end{figure*}
\begin{figure*}
\includegraphics[angle=0,width=0.99\textwidth]{fig/all_haloes_xi2_compare_1.pdf}
\caption{Two-point correlation functions computed as a pair separation function R. {\it Left panel:} Two-point statistics
for subsampled dark matter particles three our models: GR (solid blue), nDGP $\Omega_{rc}=0.0124$ (dashed orange) and
nDGP $\Omega_{rc}=0.438$ (dashed-dotted red). The shaded regions mark illustrative cosmic variance error for the GR case.
Three distinctive epochs are shown: $z=0,0.5$ and $1$. Consecutive values of $r^2\xi_2(r)$ were scaled down to allow better
presentation.
{\it Middle panel:} Same as the left panel but for all DM haloes (centrals and satellites) identified in our simulations.
{\it Right panel:} Same as the middle panel but only for main haloes.
In all three cases the lower subpanels illustrate the fractional departure from the fiducial GR case ($\xi_2/\xi_2^{GR}-1$)
at $z=0$. Here the error bars mark the Poison sampling errors for the pair number counts in each $r+\Delta r$ separation bin.
In contrast the shaded region (shown only for $z=0$ for clarity) reflect the cosmic variance error contribution.
}
\label{fig:mat_halo_2PCRF}
\end{figure*}
We now want to obtain a more quantitative measures of the general $S_n$'s behaviour observed above. Since our $N>5$ moments are severely
affected by finite-volume effects at even intermediate scales, in the following we will limit our analysis to only first three
reduced cumulants: skewness ($S_3$), kurtosis ($S_4$) and $S_5$. We are interested in the fractional differences ($\Delta S$) of the nDGP hierarchical
amplitudes taken as always w.r.t. the fiducial GR case assessed for a range of scales and epochs. We plot the results for the three
cumulants taken at three redshifts: $z=0,0.5$ and $1$ in the nine panels of Fig.~\ref{fig:matter_s3s4s5}. As already observed in the previous
plots, the non-GR signatures decrease with redshift and increase with the reduced cumulant order. For all the cases except the reduced
skewness, we can observe some erratic behaviour at large scales. This is driven by the relatively limited volume of our simulation,
and we can attribute this behaviour to noise, as indicated clearly by the cosmic variance shaded areas engulfing all the lines
at $R\stackrel{>}{{}_\sim} 10h^{-1}\,{\rm Mpc}$. However, at large scales there could be potentially interesting features appearing around the BAO wiggle,
which is a common feature predicted to exist at a fraction of the acoustic horizon scale in the higher order moments (see the details
in \citep{SkewBAO}). Our current limited simulations prevents us from charting this territory, and we will leave the studies of the BAO-related
scales for the future.
Focusing on small scales, where cosmic variance is sub-dominant, we note that there is a strong signal at a level of
significance $>2-3\sigma$ present for both our nDGP models in $S_4$ and $S_5$. The predicted deviations of the large cross-over scale
model of $\Omega_{rc}=0.0124$ for $S_3$ are too small to be significant in the presence of our sampling noise. However, for the case
of the stronger nDGP model its signature is prominent enough to constitute a $>2\sigma$ strong signal for all redshifts even in the skewness alone.
\subsection{Haloes}
\label{subsec:haloes}
It is always very insightful to trace and study the correlation hierarchy induced by gravitational clustering on the matter density field.
In the previous section we have seen that the fifth-force induced by the scalar-field propagation in the nDGP cosmology affect
the higher-order moments of the matter density field only at relatively small scales. This was consistent with the behaviour we could
trace in the full 1D pdf's of the density field. However, in reality the matter density field is not accessible to us directly
from observations. This is why we will now analyse the clustering hierarchy exhibited by dark matter halo populations as found
in our simulations. As described in Sec.\S\ref{sec:simulation} we split our DM haloes into various populations, based on their spatial
averaged abundances. In this way, we can study the clustering of different density tracers, and in principle such halo populations
can be (after making some simplifying assumptions) related to various populations of observed luminous galaxies. We shall not attempt
to model here any more complicated effects that are related to a given galaxy survey selection, nor any geometry effects.
This would require construction of a dedicated survey-specific mock galaxy catalogue and our current simulations are too limited in
volume and resolution to allow for such a procedure. However, in the future, once more advanced and bigger MG simulations become
available, it will be worthwhile to apply our analysis to survey-dedicated mocks.
\subsubsection{2-point statistics}
\label{subsec:haloes:2pt_stat}
\begin{figure}
\includegraphics[angle=0,width=0.46\textwidth]{fig/haloes_z_SN_compare.pdf}\\
\includegraphics[angle=0,width=0.46\textwidth]{fig/haloes_centrals_z_SN_compare.pdf}
\caption{The hierarchical amplitudes tiered from the reduced skewness, $S_3$, (the most bottom group of lines in each panel)
to $S_9$ (the top group of lines in each panel) calculated for the DM halo overdensity field at $z=0$. In the top panel
we show the data for all haloes found in our simulations (centrals+sattellites), while in the bottom panel we use
main haloes only (centrals).
}
\label{fig:haloes_Sns}
\end{figure}
We begin with taking a look at the 2-point clustering statistics expressed in terms of the configuration space 2-point correlation
functions $\xi(r)$. In Fig.~\ref{fig:mat_halo_2PCRF} we plot in three panels, the scaled $r^2\xi(R)$ for the dark matter sub-sample
with $\overline{n}=1.8\times10^{-3}h^3/$Mpc$^3$ (left panel), the central+satellites halo sample with $\overline{n}=1.4\times10^{-3}h^3/$Mpc$^3$ at $z=0$
(middle panel) and the centrals only sample with $\overline{n}=1.3\times10^{-3}h^3/$Mpc$^3$ (right panel). We consider the three epochs
$z=0,0.5$ and $1$, with the higher redshifts lines downscaled for brevity. As usual we show the relative ratios to the fiducial GR case
in the bottom panels, here displayed only for the $z=0$ case, where the deviations from GR are the strongest.
Since now we have individual pair-number counts for each model in each separation bin $r+\Delta r$, we decide
to show the Poisson error bars reflecting given number-counts separately from the cosmic variance error (shaded region). For the matter $\xi_2$
we observe that the nDGP deviations appear on all scales probed, however due to large cosmic variance plaguing our simulations,
the $1\sigma$ significance is limited to pair-separation scales of $r\stackrel{<}{{}_\sim} 15(40)h^{-1}\,{\rm Mpc}$ for nDGPa (nNDGb) models.
The situation for both our most-abundant halo populations is in stark contrast. Firstly, we need to note that the relative differences are much smaller.
While for the DM 2PCF the model with the smallest cross-over scale attained a difference of a $25\%$ magnitude, for haloes the maximum departure
from the GR 2PCF amplitude is around $5\%$, except for the smallest separations $r\sim1-3h^{-1}\,{\rm Mpc}$ where it reaches $10\%$ strength.
We have also check the 2PCF amplitudes for our more diluted halo samples. There, the error due to sparse-sampling shot noise is much more
severe rendering all the differences from the GR at small scales to be statistically insignificant.
This was a very important exercise. The study of the $\xi_2(r)$ amplitudes has revealed that in the case of nDGP gravity the strong signal
present in the matter density field clustering gets strongly suppressed and nearly diminishes when one look at the clustering of DM haloes.
There still might be a possibility to identify and extract a more robust MG signal in the 2PCF alone. For example one might try to use a specific
combinations of $\xi_2$ amplitudes taken at different scales and for different galaxy/halo samples such as the clustering ratios advocated
in \cite{Arnalte-Mur2017}. However, a detailed study of this kind would require simulations with more realisations, so the cosmic variance contribution would
be minimised. Thus we postpone it for the future, when such data sets will become available.
\subsubsection{Higher moments and hierarchical cumulants}
\label{subsec:healoes:higher_moments}
\begin{figure*}
\includegraphics[angle=-90,width=0.96\textwidth]{fig/haloes_cent_s345_zevo.pdf}
\caption{Fractional differences from the GR taken at $z=0$ for three halo samples:
$\av{n}=1.4\times10^{-3}h^3/$Mpc$^3$(centrals+satellites), $\av{n}=9\times10^{-4}h^3/$Mpc$^3$
and $\av{n}=1\times10^{-4}h^3/$Mpc$^3$ (rows of panels from top to bottom)
of the first three reduced cumulants $S_3, S_4$ and $S_5$ (columns of panels from left to right).
Shaded regions illustrate the total error budget (cosmic variance + shot noise for each halo sample) on the ratios.}
\label{fig:haloes_S345}
\end{figure*}
As we have mentioned a couple of times already the two-point statistics are not enough to tell the full story. As in the case of the DM clustering,
we can expect that MG effects on the halo clustering statistics might be better visible in the higher-order moments. The general picture for the case
of plain central averaged moments $\overline{\xi}_n$ is very similar to the one we have observed for the matter clustering moments shown in
Fig.~\ref{fig:matter_Xins}, thus we will skip this and jump straight away to the reduced cumulants $S_n$'s. We plot them at $z=0$ for two
halo samples (centrals only and centrals+satellites) in Fig.~\ref{fig:haloes_Sns}. First, we would like to focus on the impact made by removing
the satellites from the halo sample. This effect is clearly visible in the lower panel for $R\stackrel{<}{{}_\sim} 3h^{-1}\,{\rm Mpc}$, where especially the higher-order
cumulants starting from $S_6$ experience a dramatic suppression of amplitude at those scales. This effect is caused by halo exclusion. By removing
the subhaloes from our sample we are unable to probe the 1-halo regime any more. Thus at scales that are comparable with halo sizes we are left
with only two limiting cases, one or zero halo count in a given region of space. This is a well known effect \cite{Bardeen1986,Baldauf2013}.
However, an important observation
can be made here. By studying the scale and the magnitude of the halo exclusion effect we have found that they are in quantitative agreement
for all three models we study here. This is consistent with the very similar satellites fractions we have found for all the three simulations
(see Sec.\S\ref{sec:simulation}).
The second important feature fostered by the results shown in Fig.~\ref{fig:haloes_Sns} is that the differences between models are much larger
and also extend to larger scales than what we have observed for the matter field. The additional contrast here is the fact that now the differences
from GR can take both negative and positive sign, while they only were negative in the case of DM. As we will see later, especially for
the $\Omega_{rc}=0.438$ case, for most of the scales considered the reduced cumulants take actually larger amplitudes than in the GR case. A behaviour which
is exactly opposite to the one we have witnessed for the matter density hierarchical amplitudes. Since the reduced cumulants are scaled by
the averaged variance this effect ought to be driven mostly by higher-order biasing of haloes.
To obtain a more quantitative insight into the nDGP effects we will focus on the first three cumulants ($S_3-S_5$) estimated at $z=0$ for which
we find the strongest differences. We show the relative deviations from GR of these hierarchical amplitudes for three halo samples in Fig.\ref{fig:haloes_S345}.
Since the ratios get too noisy to allow for a more robust analysis we take an average 3-bin centred value and use cubic splines to smooth
the data.
The general trend that the non-GR signature gets stronger as we consider higher moments is also confirmed. Interestingly, we observe that
the relative differences here can take much larger values than in the case of DM cumulants and they are carried up to much larger scales.
While for the matter density all cumulants started to converge on the GR values for $R\stackrel{>}{{}_\sim} 10h^{-1}\,{\rm Mpc}$, here for all our halo samples
the differences can still be as large as $10-40\%$ up to $R\stackrel{<}{{}_\sim} 50h^{-1}\,{\rm Mpc}$.
For the case of the reduced skewness the difference is only significant for the stronger nDGPb model and only for the richest
one of our halo samples (centrals+satellites), where the magnitude of non-GR signal is varying from $5\%$ to $10\%$ at the scales we consider.
Despite the fact that $\Delta S_3$ is relatively small, there is a significant departure from the GR mean,
reaching for $4\leq R(h^{-1}\,{\rm Mpc})^{-1}\leq 9$ a $3\sigma$ statistical strength.
By considering two main halo samples only ($\av{n}=9\times 10^{-4}h^3/$Mpc$^{3}$ and
$\av{n}=1\times 10^{-4}h^3/$Mpc$^{3}$) we observe, that
the relative differences from the GR case are larger for the sparser of the two samples. We need to conclude however that the errors connected
with the shot-noise and sparse-sampling are so severe that the non-GR signature is enveloped by the noise for the case
the two central halo samples. We can expect that going to higher spatial abundances (hence lower-mass haloes) would
improve the situation. However, a new class of higher-resolution simulations are needed to perform such a study. It is also clear that for
all three halo samples the $\Omega_{rc}=0.0124$ signature is too weak to be significant. It emerges above the sampling-noise error envelop
only at the smallest probed scale ($R=1h^{-1}\,{\rm Mpc}$) for centrals+satellites sample, alas with a statistical signal less than $2\sigma$.
Now we want to focus on the reduced kurtosis, since it appears that this cumulant can be potentially the most promising observable to look
for a significant non-GR signature. Here all halo samples contain a significant nDGP signal at different scales .
For the large cross-over scale of nDGPa model we find that at $R\stackrel{<}{{}_\sim} 3h^{-1}\,{\rm Mpc}$ the signal reaches a few $\sigma$ significance.
For the stronger nDGPb model, the signal is significant and very prominent for all scales when we consider the centrals+satellites
sample and up to $R\simlt20h^{-1}\,{\rm Mpc}$ for centrals with $\av{n}=9\times 10^{-4}h^3/$Mpc$^{3}$. For the lowest density sample the departure from
the GR case is so strong that it is still significant in the regime $2\stackrel{<}{{}_\sim} R/(h^{-1}\,{\rm Mpc})\simlt20$. This is a very promising result, as
such a number density of tracers can be already attained by for example the Luminous Red Galaxy (LRG) sample in the SDSS/BOSS survey \cite{SDSS_LRGs}.
Thus, our finding open an exciting possibility to use the reduced kurtosis of LRG clustering as a discriminatory test for non-GR theories
involving the Vainshtein screening. For the large cross-over scale model we observe that it departures from the GR case in a significant way only
at small scales. For the main halo sample at $R\leq 3h^{-1}\,{\rm Mpc}$ the signal becomes marginally ($2-3\sigma$) strong.
Finally, we move to $S_5$, which fosters, as predicted, the largest relative deviations from the GR case. Alas, at larger separations ($R\simgt20h^{-1}\,{\rm Mpc}$)
and for sparser samples the ratio of nDGP to GR becomes very erratic and starts to fluctuate around zero. This reflects both the fact that
our simulations are characterised by relatively low volume coverage and the fact that higher-order moments are very vulnerable to sparse sampling noise.
Nonetheless, for our richest sample the non-GR signature reaches $\sim40\%$($\sim10\%$) for nDGPb(nDGPa) at $R\sim8-10h^{-1}\,{\rm Mpc}$.
The corresponding statistical significances are $8\sigma$ and $3\sigma$ respectively. For the centrals only with $\av{n}=9\times 10^{-4}h^3/$Mpc$^{3}$
we find that the strong nDGP model experiences a significant departure from the GR for $R\simlt20h^{-1}\,{\rm Mpc}$, where at the smallest scales
($R\simlt5h^{-1}\,{\rm Mpc}$) the statistical significance for the $S_5$ reaches up to $6-7\sigma$ from the GR mean. Here, for the same sample,
the mild nDGP model at larger separations fluctuate around the GR case, but again at the small scales $<3h^{-1}\,{\rm Mpc}$ the deviation become strong, reaching
magnitude of $5\sigma$. For both models and at all scales the ratio of $S_5$ experiences a large scatter with the $\av{n}=1\times 10^{-4}h^3/$Mpc$^{3}$
sample. Here, the signal could in principle be strong enough for a prospective detection, but our simulations are not good enough to allow us
to make that claim.
We can conclude that our analysis indicates that the reduced kurtosis, $S_4$, appears to offer a most beneficial combination of the non-GR signature
amplitude and noise properties to maximise the signal strength. While $S_5$ fosters usually stronger relative differences, it appears to be
really noisy for our halo samples as well. However in principle, one could imagine that with a data sample covering a much larger volume, the
impact of the finite-volume effect would be suppressed, increasing the quality of the signal in $S_5$. We also see that for most of the cases,
if one considers only a very mild MG model, one needs to go to really small scales ($R\simlt5h^{-1}\,{\rm Mpc}$) to look for a prominent signal.
At such small scales, deep into the non-linear regime, the modelling of galaxy clustering becomes very complicated due to various
degeneracies connected with baryonic effects and non-linear galaxy bias \cite{skewness_barions,mog_gadget,vanDaalen2014,Hellwing2016}.
\section{Discussion and Conclusions}
\label{sec:conclusions}
In this paper we have studied the gravitational instability mechanism of the GR and two nDGP models as our chosen test-case examples
of MG with the Vainshtein screening mechanism. To follow the gravitational evolution into the non-linear regime
we have conducted and analysed a series of N-body simulations implementing WMAP9 cosmogony and the non-linear dynamics
of GR and nDGP with $\Omega_{rc}=0.0124$ and $0.438$. Specifically, we have focused on using both low and high-order clustering statistics
to study the matter density field and halo clustering across the cosmic epochs (from $z=1$ to $z=0$) and scales (from $1$ to $100h^{-1}\,{\rm Mpc}$).
For the haloes we have considered a few different populations characterised by a different spatial abundances
(from $\av{n}=1.9\times10^{-3}h^3/$Mpc$^{3}$ to $1\times10^{-4}h^3/$Mpc$^{3}$). Thus, in the spirit of the abundance matching, we have made
a first attempt towards modelling of observables that can be easier associated with the different observational samples of many galaxy surveys.
Here we provide a list of our findings constructed to emphasise the most important ones:
\begin{itemize}
\item For all considered models and density field statistics we have found that the maximal relative difference from the GR case is always attained at
$z=0$ and is generally a monotonic function of time.
\item The investigation of one dimensional $p(\delta+1)$ functions has revealed that the pdfs of all models take maximums at the same averaged density values,
but are characterised by different shapes of low and high density tails. The most prominent was the increased MG pdf width
(reflected in measured later larger variance). We also noticed a significantly larger asymmetry (especially for the smallest smoothing scale $R=0.5h^{-1}\,{\rm Mpc}$).
In all cases and for all epochs the 1-point density statistics indicated that the nDGP matter density fields are characterised by enhancements at both
lower and higher density tails.
\item Both the matter density field variance and the power spectrum fostered consistent picture. Here, the non-GR models are characterised by a nearly
constant and scale-independent enhancement factors of $4$ and $20\%$ for our mild and strong nDGP models respectively.
\item We found a remarkable agreement of the PT-based estimators for the variance and the skewness with our N-body results. The absolute amplitudes
are in percent-level agreement down to $10(15)h^{-1}\,{\rm Mpc}$ for the variance (skewness). At the same time the relative deviations form GR retained the accuracy
even down to smaller scales of $\sim5h^{-1}\,{\rm Mpc}$.
\item The high-order moments of the matter field, $\overline{\xi}_2-\overline{\xi}_9$ were only mildly affected by the MG dynamics. However,
the corresponding hierarchical amplitudes $S_3-S_8$ are marked with significant deviations from the GR case at small scales $R\stackrel{<}{{}_\sim} 10h^{-1}\,{\rm Mpc}$
for the $z=0$. At higher redshifts the differences are typically sizeable smaller. The main feature is that nDGP models are characterised
by smaller amplitudes $S_n$ w.r.t. the GR fiducial case at small scales, while they converge to the GR values at larger separations ({\frenchspacing\it i.e.} $R\stackrel{>}{{}_\sim} 10h^{-1}\,{\rm Mpc}$).
This reflect the fact that at larger scales the matter power spectrum ($P(k)$) and variance ($\sigma^2(R)$) are enhanced by a constant factor, but
the nDGP gravity does not change their shape.
\item In the 2PCF the relative deviations from the GR were observed to be much stronger for the DM case than in the halo sample. With the deviations
in the matter clustering stretched out to even large scales $\sim 10h^{-1}\,{\rm Mpc}$. This was not found to be the case for the halo 2PCF, where the significant
differences were contained only to the smallest pair separations $R\simlt10h^{-1}\,{\rm Mpc}$.
\item The non-GR signal was much stronger in higher-order halo clustering statistics. Reduced moments are characterised by departures from GR-fiducial
values up to much larger scales ($R\stackrel{>}{{}_\sim} 50h^{-1}\,{\rm Mpc}$) than in the case of DM clustering. A side remark would be that the halo exclusion effect was
marked prominently in higher $S_n$'s for the centrals-only sample.
\item Qualitatively we have found that the skewness offers a good chance constrain only strong nDGP models (like our nDGPb) and only for
our halo sample with the highest spatial density. On the other hand, our study has pinpointed the reduced kurtosis as an excellent candidate for
a very promising cosmological MG probe. Here the nDGPb signal is prominent on all nearly scales and for all halo samples.
The model with large cross-over scale nDGPa is also harbouring strong signal, alas contained to only small scales of $R\simlt3h^{-1}\,{\rm Mpc}$. In general,
we were able to identify the non-GR signal with a good statistical significance, varying from 3 to $8\sigma$ for a given model.
\item Lastly, we have observed even stronger relative signal in $S_5$, however this was accompanied by a much larger sampling noise effectively reducing
the statistical significance of it. We advertise however, that with a better and larger-volume sample it might become feasible to extract
MG signal from $S_5$ with a large statistical significance.
\end{itemize}
In the literature we find abundant studies pointing toward a very similar picture as fostered by our analysis. In general, we see a trend
in dark matter clustering properties in models exhibiting a non-negligible fifth-force on intergalactic scales, and this
is shared among models despite a specific screening mechanism they employ. Namely, we observe that the fifth-force acting on DM
in MG theories enhances the gravitational instability process. This is reflected in a more efficient transfer of matter from underdense
to ovedense regions, which results in cosmically depressed regions, like cosmic voids, to be characterised
by lower densities \cite[see {\frenchspacing\it e.g.}][]{Hellwing2009,halosvoids_fr,Clampitt2013}.
Due to mass conservation the average density remains the same as in the GR case, hence the first affected clustering characteristic
is of the second-order, thus the variance of the density field. We saw that both in the configuration space variance as well as in the
power spectrum. The gravitational instability also naturally leads to more complicated shape deviations of the density field
distribution function. This is reflected in growing amplitude of the asymmetry rank -- the skewness,
or tails squashing as measured by the kurtosis. To understand why the fifth force dynamics, while enhancing the density variance w.r.t.
to the fiducial GR case, actually leads to lower skewness and kurtosis (and higher-order cumulants as well), we need to recall that
there is an intrinsic asymmetry imprinted in the density field described by the density contrast $\delta$. While on the positive side of
the distribution the density can grow to arbitrarily large values, reaching presently for example orders of $\sim10^{6}$ in the very centres
of cluster sized DM haloes, on the negative size there is a fundamental barrier of $\delta=-1$, which reflect the limiting case of a zero density
in an absolutely empty region of space. Thus in MG scenarios, and in the nDGP in particular, more empty voids lead to reduced asymmetry
of the PDF thanks to the enhanced $\delta<0$ tail of the distribution. One could argue that in general the more emptier voids are compensated
by increased density inside haloes and filaments. But since the latter are occupying much smaller fraction of the Universe, the overall
effects of voids prevails in the higher-order cumulants.
The DM density field is not accessible to observations, but one is able to measure the distribution of the total matter by using weak lensing
techniques. It has been already proven that weak lensing statistics such as the lensing convergence $\kappa$ and its power spectrum $C_{\kappa\kappa}$
are useful cosmological probes.
Our results for the matter density moments indicate that the non-GR signature we found should also be present
in weak lensing statistics \cite{Schmidt2008,Beynon2010}.
However, it would be very hard to assess the feasibility of the weak lensing statistics as probe of nDGP-kind of MG. This is because,
as our studies have found, most of the MG induced effects are confined to relatively small scales $R\stackrel{<}{{}_\sim} 10h^{-1}\,{\rm Mpc}$. There, the baroynic effects
related to energetic processes of galaxy formation are a source of a major influence on the total matter distribution
\cite{skewness_barions,Hellwing2016}. A separate dedicated study, involving self-consistent hydrodynamical simulations with galaxy
formation physics are needed to test the robustness and usability using the distortions of matter clustering as a MG probe.
The second part of the analysis presented in this work yields however a much more optimistic picture. Our studies of the higher-order
halo clustering statistics have shown that there is a very appealing opportunity to use the reduced cumulants as cosmological probes of non-GR
models and simultaneously consistency checks for GR. Here, especially the reduced kurtosis appears as the most promising statistics.
Since, we have not attempted any realistic galaxy modelling beyond some simple abundance matching approach, we should caution that
there are still a number of potential systematics effects related to both galaxy formation and survey selection that can potentially
weaken the MG signal we have found. However, the fact that the strong signal was present in all three halo samples we considered,
is an optimistic indicator that the MG signature persists over a range of halo masses. Thus, we can expect that it should be also present
in galaxy samples constructed using various selection criteria.
Another source of potential worry consists of a notion that the clustering statistics that are measured from spectroscopic
galaxy surveys concerns the position of galaxies in the redshift space. In the present study we have only focused on position space
clustering neglecting the effects of halo peculiar velocities. In MG theories we can expect in general that the fifth-force is affecting
the comic velocity field much stronger than the density alone \cite{LiHell2013,Hellwing2014PhRvL}, this may lead to a signal degeneracy at scales where
the both clustering and halo velocities are strongly affected. Thus, redshift space distortions contrary to classical approach, where they
are a source of rich cosmological information, in the case of the reduced cumulants are a source of another systematic.
Withal, it has been shown that in the case of the hierarchical amplitudes, the redshift space distortion appears to affect the variance
and the higher order moments to a similar degree, and since these two quantities appears in the denominator and numerator of $S_n$ formula
respectively, the overall redshift space effect is suppressed \cite{Hivon1995}. Thus, one can hope that the signal we have identified in this work
would persists also in the redshfts space cumulants of galaxy clustering.
Therefore, we strongly advocate here the urgent need for a good quality galaxy mock catalogues for MG theories that would allow to test
for the robustness of MG signal in galaxy $S_n$'s. Once such data sets will become available, we will be given the opportunity to
exploit already rich galaxy clustering data from the surveys like
SDSS/BOSS\cite{Ross2006}, 2dF\cite{Baugh2004,Croton2004_2dF} or VIPERES\cite{Cappi2015} and even more powerful data that will
come from future grand-scale surveys like Euclid and DESI, to perform an independent and robust tests of GR and non-GR models on
intergalactic scales.
\section*{Acknowledgements}
The authors are very grateful to Baojiu Li for inspiring discussions and for providing the {\tt ECOSMOG} code
that was used to run the simulations in this paper.
WAH and KK are supported from European Research Council (grant number 646702 ``CosTesGrav'').
WAH also acknowledges support from the Polish National Science Center under contract \#UMO-2012/07/D/ST9/02785.
KK is also supported by the UK Science and Technologies Facilities Council
grants ST/N000668/1. BB was supported by University of Portsmouth.
GBZ is supported by NSFC Grant No. 11673025, and by a Royal Society-Newton Advanced Fellowship.
The simulation used in this work were carried on the Sciama HPC cluster on the University of Portsmouth.
This work used the DiRAC Data Centric system at Durham University,
operated by the Institute for Computational Cosmology on behalf of
the STFC DiRAC HPC Facility (\url{www.dirac.ac.uk}). This equipment
was funded by BIS National E-infrastructure capital grant ST/K00042X/1,
STFC capital grant ST/H008519/1, and STFC DiRAC Operations grant
ST/K003267/1 and Durham University. DiRAC is part of the National
E-Infrastructure. This research was carried out with the support of the
HPC Infrastructure for Grand Challenges of Science and Engineering
Project, co-financed by the European Regional Development Fund under
the Innovative Economy Operational Programme.
\renewcommand{\bibname}{References}
\bibliographystyle{h-physrev}
|
\section{Introduction}
\label{sec:introduction}
Deep learning spearheaded by convolutional neural networks (CNN) and recurrent neural
networks (RNN) has been pushing the frontiers in many computer vision
applications~\cite{lecun2015deep}. Deep learning algorithms have achieved or surpassed
human-levels of perception in some applications~\cite{he2016deep,xiong2016microsoft},
enabling them to be deployed in the real-world applications.
The key challenges of deep learning algorithms are the computational complexity and
model size, which impede their deployment on end-user client devices,
thus limiting them to cloud-based high performance servers.
For example, AlexNet~\cite{krizhevsky2012imagenet}, a popular CNN model, requires 1.45 billion
operations per input image with 240MB weights~\cite{suda2016throughput}.
Many researchers have explored hardware accelerators for CNNs to enable deployment
of CNNs in embedded devices and demonstrated good performance at low power
consumption~\cite{qiu2016going,shin2017dnpu}. Reducing the data precision is a
commonly used technique for improving the energy efficiency of CNNs.
Typically, CNNs are trained on high performance CPU/GPU with 32-bit floating-point data.
Fixed-point representation with shorter bit-width for CNN weights and
activations has been widely
explored~\cite{judd2015reduced,gupta2015deep,gysel2016hardware,lin2015fixed},
which significantly reduces the storage requirements, memory
bandwidth and power consumption without sacrificing accuracy.
This work focuses on the number representation schemes for implementing
CNN inference. The representation scheme has the following requirements:
\begin{itemize}
\item Accuracy: the representation should achieve the desired network accuracy
with limited bit-width.
\item Efficiency: the representation can be implemented in hardware efficiently.
\item Consistency: the representation should be consistent across different CNNs.
\end{itemize}
Based on these requirements, we propose using floating-point numbers for CNN weights
and fixed-point numbers for activations. We justify this choice from both
algorithmic and hardware implementation perspectives.
From the algorithmic perspective, using popular large-scale CNNs
such as AlexNet~\cite{krizhevsky2012imagenet},
SqueezeNet~\cite{iandola2016squeezenet}, GoogLeNet~\cite{szegedy2015going}
and VGG-16~\cite{simonyan2014very},
we show that the representation range of the weights is the main factor that
determines the inference accuracy, which can be better represented
in floating-point format.
From the hardware perspective, we show that multiplication
can be implemented more efficiently with one floating-point operand and one
fixed-point operand generating a fixed-point product.
The rest of the paper is organized as follows, Section~\ref{sec:related_work}
discusses related work. Section~\ref{sec:background} gives some background
about different number representation formats and typical hardware implementation
of CNNs. Section~\ref{sec:scheme} describes the proposed number representation
scheme. Section~\ref{sec:algorithmic} and Section~\ref{sec:implementation}
discuss the scheme from the algorithmic and implementation perspective.
Section~\ref{sec:results} presents the experimental results, and
Section~\ref{sec:conclusion} concludes the paper.
\section{Related Work}
\label{sec:related_work}
Precision of the neural network weights and activations plays a major role in
determining the efficiency of the CNN hardware or software implementations.
A lot of research focuses on replacing the standard 32-bit floating-point
data with reduced precision data for CNN inference.
For example, Gysel et al.~\cite{gysel2016ristretto} propose representing
both CNN weights and activations using minifloat, i.e., floating-point number
with shorter bit-width.
Since fixed-point arithmetic is more hardware efficient than floating-point
arithmetic, most research focuses on fixed-point quantization.
Gupta et al.~\cite{gupta2015deep} present the impacts of different fixed-point
rounding schemes on the
accuracy. Judd et al.~\cite{judd2015reduced} demonstrate that the minimum
required data precision not only
varies across different networks, but also across different layers of the
same network. Lin et al.~\cite{lin2015fixed}
present a fixed-point quantization methodology to identify the optimal data
precision for all layers of a network.
Gysel et al.~\cite{gysel2016hardware} present a framework
{\em{Ristretto}} for fixed-point quantization and re-training of CNNs based on
{\em{Caffe}}~\cite{jia2014caffe}.
Researchers have also explored training neural networks directly
with fixed-point weights. In~\cite{hammerstrom1990vlsi}, the author
presents a hardware architecture for on-chip learning with
fixed-point operations. More recently, in~\cite{courbariaux2014training},
the authors train neural networks with floating-point, fixed-point and
dynamic fixed-point formats and demonstrate that fixed-point
weights are sufficient for training. Gupta et al.~\cite{gupta2015deep} demonstrate network
training with 16-bit fixed-point weights using stochastic rounding scheme.
Many other approaches for memory reduction of neural networks have been explored.
Han et al.~\cite{han2015deep}
propose a combination of network pruning, weight quantization during training
and compression based on Huffman
coding to reduce the VGG-16 network size by 49X.
In~\cite{deng2015reduced}, the authors propose to store both 8-bit quantized
floating-point weights and 32-bit full precision weights. At runtime, quantized weights
or full-precision weights are randomly fetched in order to reduce memory bandwidth.
The continuous research effort to reduce the data precision
has led to many interesting demonstrations with 2-bit weights~\cite{intel2bitnn}
and even binary weights/activations~\cite{binarynet,xnornet}.
Zhou et al.~\cite{dorefanet} demonstrate AlexNet
training with 1-bit weights, 2-bit activations and 6-bit gradients.
These techniques require additional re-training and can result in sub-optimal
accuracies.
In contrast to prior works, this work proposes quantization of a pre-trained
neural network weights into floating-point numbers and implementation of
activations in fixed-point format both for
memory reduction and hardware
efficiency. It further shows that floating-point representation of weights
achieves better range/accuracy trade-off
compared for the fixed-point representation of same number of bits and
we empirically demonstrate it on state of the art
CNNs such as AlexNet~\cite{krizhevsky2012imagenet}, VGG-16~\cite{simonyan2014very},
GoogLeNet~\cite{szegedy2015going}
and SqueezeNet~\cite{iandola2016squeezenet}.
Although this work is based on quantization only without
the need for retraining the network, retraining may also be applied to
reclaim part of the accuracy loss due to quantization.
\section{Background}
\label{sec:background}
\subsection{Fixed-Point Number Representation}
\label{subsec:background_fixed_point}
Fixed-point representation is very similar to integer representation. The
difference is that integer has a scaling factor of 1 and fixed-point can have a
pre-defined scaling factor as power of 2.
Some examples of fixed-point numbers are shown in Fig.~\ref{fig:fixed-point}.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./fixed_point_demo.png}
\caption{
Examples of fixed-point representation. The arrow indicates the position of
radix point. So the first number is the corresponding integer value scaled by
$2^{-5}$ and the second one is scaled by $2^{-7}$.
}
\label{fig:fixed-point}
\end{figure}
Usually, all fixed-point representation is assumed to share the same scaling
factor during the entire computation. In some scenarios, the computation can
be classified into different sets, e.g., for different CNN layers, with
fixed-point numbers of different scaling factors. This is also referred as
dynamic fixed-point
representation~\cite{courbariaux2014training}.
\subsection{Floating-point Number Representation}
\label{subsec:background_floating_point}
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./floating_point_demo.png}
\caption{
Example of floating-point representation.
With the implicit bit, the range for the significand part is limited to be $[1, 2)$.
Without the implicit bit, i.e., leading bit as 0, the range becomes
$[0, 0.5)$.
}
\label{fig:floating-point}
\end{figure}
One example of floating-point number representation is shown in
Fig.~\ref{fig:floating-point}.
For a floating-point representation, there are typically three parts: sign,
mantissa and exponent. The sign bit determines whether the number is a positive
or negative number. The mantissa determines the significand part and the exponent
determine the scale of the value. Usually, there are some special encodings used
for representing some special numbers (e.g., 0, NaN and +/- infinity),
For binary floating-point numbers, the mantissa can assume an implicit bit, which
is also adopted by IEEE floating-point standard. This ensures that the value of
mantissa is always between 1 and 2, so the leading bit 1 can be omitted to save
storage space. However, such an implicit bit places a limit on the smallest
representable number for the significand part.`
The exponent is typically represented as an unsigned integer number with a bias.
For example, for an 8-bit exponent with a bias of 127, it can represent numbers from
-127 to 128, i.e, 0-127 to 255-127.
\subsection{Hardware Implementation of CNNs}
\label{subsec:background_cnn_hw}
CNNs typically consist of multiple convolution layers interspersed by pooling, ReLU and
normalization layers followed by fully-connected layers. Convolution and
fully-connected layers are the most compute and data intensive layers
respectively~\cite{qiu2016going}.
The computation in these layers consist of multiply-and-accumulate (MAC) operations.
The data path is illustrated in Fig.~\ref{fig:data_flow},
where the input features are multiplied with the weights to get the intermediate data
(i.e., partial sums). These partial sums are accumulated to generate the output features.
Since fixed-point arithmetic is typically more efficient for hardware implementation,
most hardware accelerators implement the MAC operations using fixed-point representation.
The power/area breakdown of the CNN hardware accelerator mostly depends on the data flow
architecture. For example, in Eyeriss~\cite{chen2016eyeriss}, in each processing element (PE),
MAC and memory account for about 9\% and 52\% area respectively.
For Origami~\cite{cavigelli2015origami}, MAC and memory account for about
32\% and 34\% area respectively.
\section{Proposed Number Representation Scheme}
\label{sec:scheme}
The overview of our proposed number representation scheme is shown in
Fig.~\ref{fig:data_flow}. Different from most existing CNN implementations,
we propose using a combination of floating-point and fixed-point
representations.
The network weights are represented as floating-point numbers while the
input/output features are represented as fixed-point numbers.
The multiplier is implemented to take one floating-point number and one
fixed-point number and produces output, i.e., intermediate data, in
fixed-point format.
The intermediate data is in fixed-point format and can have wider bit-width
than the input/output features.
The accumulation is the same as fixed-point adder, which can have higher
bit-width.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./data_flow.png}
\caption{
Overview of the data path for convolutional and fully-connected layer
operations with the proposed number representation scheme.
The weights are represented using floating-point numbers, while the
input/output features and intermediate data are represented using
fixed-point numbers.
}
\label{fig:data_flow}
\end{figure}
From hardware perspective, it is more efficient to implement multiplication using
floating-point number and addition using fixed-point number.
The multiplication operations have one fixed-point number input, one floating-point
number input and fixed-point number output.
This can be implemented with a multiplier and a shifter.
The multiplier will multiply the fixed-point number with the mantissa part of the
floating-point number, and the shifter will shift the results according to the
exponent value. Therefore, we propose the hardware architecture illustrated in
Fig.~\ref{fig:data_flow}.
The accumulation/addition works with fixed-point numbers, which can be wider
(i.e., with larger bit-width) than either of the inputs. This part is similar to
most fixed-point number based implementation of CNN accelerators.
\section{Algorithmic Perspective}
\label{sec:algorithmic}
In this section, we investigate and explain why the proposed number representation
scheme is better from the algorithmic perspective.
Section~\ref{subsec:fixed_point_accuracies} demonstrates that different CNNs can have
inconsistent fixed-point bit-width requirements for representing the weights.
Section~\ref{subsec:weight_distribution} investigates this inconsistency by
analyzing CNN weight distribution and properties.
Section~\ref{subsec:range_vs_precision} shows that the representation range is the
main factor that determines the inference accuracy.
Section~\ref{subsec:floating_point_accuracies} shows that floating-point representation
is more efficient and consistent representation for CNN weights.
\subsection{CNN Accuracy with Fixed-Point Weights}
\label{subsec:fixed_point_accuracies}
To evaluate different number representation schemes, we implement
weight quantization based on {\em{Caffe}}~\cite{jia2014caffe} framework.
To make a fair comparison, we assume that there is always a sign bit for representing
negative values for both fixed-point and floating-point numbers.
We will represent the bit-width of floating point representation as $m+e$, where
$m$ is the number of mantissa bits and $e$ is the number of exponent bits.
We apply the weight quantization on four popular CNN networks:
AlexNet~\cite{krizhevsky2012imagenet},
SqueezeNet~\cite{iandola2016squeezenet}, GoogLeNet~\cite{szegedy2015going}
and VGG-16~\cite{simonyan2014very}.
We evaluate the network accuracy by doing quantization for all convolutional and
fully-connected layer weights. The activation is quantized to 16-bit fixed-point.
For each layer, we normalize the weights so that the maximum absolute value equals 1.
This is similar to the dynamic fixed-point
quantization~\cite{moons2016energy,gysel2016hardware,courbariaux2014training}.
All accuracy results are top-1 accuracy and normalized with the top-1 accuracy
using 32-bit floating-point representation.
The results of top-5 accuracy correlate with that of top-1 accuracy.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./network_accuracies_fixed.png}
\caption{Normalized accuracy of the networks using fixed-point representation with
different bit-width.}
\label{fig:acc_fixed}
\end{figure}
The network accuracy results using fixed-point representation are shown in
Fig.~\ref{fig:acc_fixed}. Two observations can be made here:
\begin{itemize}
\item For all networks, the accuracy starts increasing sharply after a certain
threshold. For all networks, the normalized accuracy increases
from close to 0 to close to 1 within 2 or 3 bits difference. This suggests that
there is something dramatically different with these additional bits.
\item Among different networks, the required number of bits is very inconsistent.
With 7-bit fixed-point number, AlexNet and SqueezeNet can achieve close to
full accuracy, while GoogLeNet and VGG-16 have very low accuracy.
GoogLeNet and VGG-16 need 10 to 11 bits to achieve full accuracy.
\end{itemize}
This inconsistency in bit-width requirements across different CNNs
poses challenges for hardware implementation.
For the design to be general-purpose and future-proof, the designer has to use
margined bit-width or use runtime adaptation~\cite{moons2016energy},
both of which incur significant overhead.
\subsection{Weight Distribution}
\label{subsec:weight_distribution}
There can be several reasons that cause the inconsistency in Fig.~\ref{fig:acc_fixed}.
The network depth is one of the possible reasons. Similar to the idea
in~\cite{lin2015fixed}, the fixed-point quantization can be modeled as quantization
noise for each layer. The network accuracy may drop further, i.e., accumulate
more noise, as the network depth increases.
The other possible reason is the number of MAC operations in each layer.
Small quantization error can accumulate over a large amount of MAC operations.
For example, the total number of MAC operations to calculate one output for convolutional and
fully-connected layers in AlexNet are (363, 1200, 2304, 1728, 1728, 9216,
4096, 4096).
However, none of the above reasons can explain the first observation earlier about
the sharp, instead of gradual, change in accuracy.
To further investigate this, we plot the weight distribution of four different layers
in AlexNet in Fig.~\ref{fig:alex_lin_hist}.
Most weights have small values even after the per-layer normalization. The distribution
is concentrated at the center, which is also the motivation for Huffman encoding
of the weights proposed in~\cite{han2015deep}.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./alexnet_conv_lin_hist.png}
\caption{
The weight distribution of four different layers in AlexNet.
{\em{conv1}} and {\em{conv3}} are convolutional layers, and fc7 and fc8 are fully-connected
layers.
}
\label{fig:alex_lin_hist}
\end{figure}
To better visualize the difference, we plot the same weight distribution in log-scale
in Fig.~\ref{fig:alex_hist}.
This plot is easier to spot the weight distribution difference and explains the accuracy
behavior under fixed-point quantization observed in Fig.~\ref{fig:acc_fixed}.
Under fixed-point quantization, the layer with the most small-valued weights, {\em{conv3}},
will be the most susceptible to quantization errors.
With 4-bit fixed-point representation, more than 90\%
weights in {\em{conv3}} are unrepresentable, i.e., quantized to 0.
This also explains why stochastic rounding works better than round-to-nearest reported
in~\cite{gupta2015deep}.
Since the layers are cascaded, the accuracy of the entire network is limited
by the weakest layer. This is why the network produces close to 0 accuracy with small
bit-width as shown in Fig.~\ref{fig:acc_fixed}.
With higher fixed-point bit-width, a larger number of weights in {\em{conv3}} become
representable, which results in the quick increase of the network accuracy.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./alexnet_conv_hist.png}
\caption{
The weight distribution in log-scale of four different layers in AlexNet.
The weights are first converted into binary floating-point format. The plot
is based on the exponent value of the weights.
}
\label{fig:alex_hist}
\end{figure}
The inconsistency in bit-width requirements observed in
Fig.~\ref{fig:acc_fixed} can also be explained with
the weight distribution. We pick the two layers with largest and smallest
weights from AlexNet and VGG-16 and plot the weight distribution in
Fig.~\ref{fig:hist_comparison}.
The layer {\em{conv3$\_$1}} in VGG-16 has more weights with smaller exponent values.
This explains why VGG-16 requires more bit-width when using fixed-point representation.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./weight_dist_comparison.png}
\caption{
The weight distribution in log-scale for different layers in AlexNet and VGG-16.
For each network, we pick the layers with the largest and smallest mean in
the exponent value of the weights.
}
\label{fig:hist_comparison}
\end{figure}
\subsection{Range vs. Precision}
\label{subsec:range_vs_precision}
The results in Fig.~\ref{fig:acc_fixed} and the weight distributions in
Fig.~\ref{fig:alex_hist} show that the network can achieve almost
full accuracy, e.g., with 7-bit fixed-point for AlexNet, even when most
weights are barely representable, i.e., only with 1 or 2 significant bits.
This means that representation range, i.e, the ability to represent
larger/smaller values, is more important than representation precision,
i.e., differentiation between nearby values.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./network_accuracies_float.png}
\caption{Normalized accuracy of the networks using floating-point representation
with different exponent bit-width. The floating-point number has 3-bit mantissa with
the implicit bit.}
\label{fig:acc_float}
\end{figure}
Since the representation range and representation precision is hard to decompose
in fixed-point representation, we investigate this using floating-point
representation.
For floating-point representation, the mantissa bit-width controls the precision
and the exponent bit-width controls the range.
Fig.~\ref{fig:acc_float} highlights some of the results of network accuracy
using floating-point representation with varying exponent bit-width (i.e.,
representation range).
The floating-point has 3-bit mantissa with the implicit bit. The implicit
bit limits the value of the significant part, so that the representation
range is controlled by the exponent part.
With floating-point representation, the networks show consistent trend,
with almost 0 accuracy with 2-bit exponent and quickly increase to almost
full accuracy with 4-bit exponent.
Unlike the behavior seen in Fig.~\ref{fig:acc_fixed},
this is expected as 2 additional bits
in exponent offers 4X increase in the representation range.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./network_accuracies_exp_2.png}
\caption{Normalized accuracy of the networks using floating-point representation
with limited exponent range.
Range of 4 is equivalent to 2-bit exponent, and
range of 8 is equivalent to 3-bit exponent.
The floating-point number has 2-bit mantissa with
the implicit bit.}
\label{fig:acc_float_exp2}
\end{figure}
To get more insight into the impact of representation range, we also run
experiments with floating-point like representation where we limit the
exponent range rather than the exponent bits.
For example, exponent range of 4 is equivalent to
2-bit exponent and exponent range of 8 is equivalent to 3-bit exponent.
The results of exponent range experiments are shown in Fig.~\ref{fig:acc_float_exp2}.
The floating-point number has 2-bit mantissa with the implicit bit.
The behavior of the accuracy is similar to the results with fixed-point
representation, i.e., accuracy increases rapidly from almost 0 to almost 1.
The relative ordering exponent range requirements also matches the bit-width
requirements for fixed-point representation. This suggests that representation range is
the main impacting factor for network accuracy.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./network_accuracies_exp_6.png}
\caption{Normalized accuracy of the networks using floating-point representation
with limited exponent range.
Range of 4 is equivalent to 2-bit exponent, and
range of 8 is equivalent to 3-bit exponent.
The floating-point number has 6-bit mantissa with
the implicit bit.}
\label{fig:acc_float_exp6}
\end{figure}
With 2-bit mantissa, some networks saturate with normalized accuracy less
than 1, as shown in Fig.~\ref{fig:acc_float_exp2}.
To compare with the effect of precision, the experiments are repeated
with 6-bit mantissa as shown in Fig.~\ref{fig:acc_float_exp6}.
Comparing Fig.~\ref{fig:acc_float_exp6} with Fig.~\ref{fig:acc_float_exp2},
the additional 4 bits in mantissa does not have significant impact on
the network accuracy. This also validates the initial hypothesis that
representation range is more important than representation accuracy.
\subsection{CNN Accuracy with Floating-Point Weights}
\label{subsec:floating_point_accuracies}
Comparing to fixed-point representation, floating-point is better for
representation range, which increases exponentially with the exponent bit-width.
The results shown in Fig.~\ref{fig:acc_float} show that 4-bit exponent is adequate
and consistent across different networks.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./network_accuracies_float_mantissa.png}
\caption{Normalized accuracy of the networks using floating-point representation
with different mantissa bit-width. The floating-point number has 4-bit exponent
and is with the implicit bit.}
\label{fig:acc_float_mantissa}
\end{figure}
The next question for floating-point representation is how much precision,
i.e., how many mantissa bits are needed.
Fig.~\ref{fig:acc_float_mantissa} highlights some of the results of network accuracy
using floating-point representation with varying mantissa bit-width.
The floating-point number has 4-bit exponent, and is with the implicit bit.
Most networks have very high accuracy even with 1-bit mantissa and achieve full
accuracy with 3-bit mantissa. This is also consistent across different networks,
which further proves that the inconsistency with fixed-point representation seen
in Fig.~\ref{fig:acc_fixed}
is mainly from the inconsistent requirements for representation range rather
than from the representation precision.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./network_accuracies_float_no_implicit.png}
\caption{Normalized accuracy of the networks using floating-point representation
with different exponent bit-width. The floating-point number has 4-bit mantissa without
the implicit bit.}
\label{fig:acc_float_no}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./network_accuracies_float_mantissa_no_implicit.png}
\caption{Normalized accuracy of the networks using floating-point representation
with different mantissa bit-width. The floating-point number has 3-bit exponent.
The mantissa is without the implicit bit.}
\label{fig:acc_float_mantissa_no}
\end{figure}
We also repeat the experiments with floating-point representation without the
implicit bit in mantissa. The results are highlighted in Fig.~\ref{fig:acc_float_no}
and Fig.~\ref{fig:acc_float_mantissa_no}.
The implicit bit in mantissa limits the range of the significand part to
$[0.5, 2)$. Removing it helps further extend the range of the representation,
especially for representing small numbers. That is why the network
accuracy saturates with 3-bit exponent instead of 4-bit.
The implicit bit also improves the precision of the significand part. Hence,
we need 4-bit mantissa instead of 3-bit to achieve full accuracy.
\section{Implementation Perspective}
\label{sec:implementation}
This section motivates the proposed number representation scheme from
the hardware implementation perspective. The implementation considerations
are discussed in Section~\ref{subsec:implementation_consideration}.
Hardware trade-off results are presented in Section~\ref{subsec:hw_results}.
\subsection{Hardware Implementation Considerations}
\label{subsec:implementation_consideration}
As discussed in Section~\ref{subsec:background_cnn_hw}, computations in CNNs
are typically implemented as MAC operations.
For the same 32-bit wide operations, hardware implementation of fixed-point
arithmetic can be more efficient than floating-point arithmetic.
This is one of the reasons why most of previous work focuses on the optimization
for fixed-point representation.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./multiplier_demo.png}
\caption{
Illustration of hardware implementation of multiplication with two
fixed-point numbers and mixed fixed-point and floating-point numbers.
}
\label{fig:multiplier}
\end{figure}
The comparison becomes less obvious when the bit-width is smaller, especially
when the number of exponent bits in floating-point representation is small.
For example, as shown in Fig.~\ref{fig:multiplier},
multiplier with a floating-point number can be implemented with
a multiplier (of the bit-width of mantissa) and a barrel shifter (of the bit-width of
exponent). This can be more efficient than multiplying two fixed-point numbers,
as the multiplier becomes smaller and the shifter is simpler.
\subsection{Hardware Trade-off Results}
\label{subsec:hw_results}
To validate the claim in Section~\ref{subsec:implementation_consideration},
we implement the multiplier with different operand
configurations using a commercial 16nm process technology and libraries.
The results are highlighted in Fig.~\ref{fig:hw_tradeoff}.
The floating-point configuration is represented
as $m$+$e$, i.e., mantissa bit-width + exponent bit-width.
Here we assume the baseline is a multiplier with two 8-bit fixed-point operands,
i.e., 8$\times$8.
The area and power numbers are all normalized with respect to the
baseline.
With the same bit-width, the proposed scheme of combining fixed-point and
floating-operands can reduce both area and power.
The reduction increases with less mantissa and more exponent bits, as a
shifter is more efficient than a multiplier.
As an example, the floating-point operand with 4-bit mantissa and 4-bit
exponent, i.e. 8x4+4, can reduce the power and area by more than 20\%,
compared to 8x8 fixed-point multiplication.
\begin{figure}[t]
\centering
\includegraphics[width = 0.95\columnwidth]{./hardware_tradeoff.png}
\caption{
Hardware implementation trade-off results for different multiplier
configurations. The area and power are normalized
with respect to the case with two 8-bit fixed-point operands.
The floating-point operands are represented as mantissa bit-width +
exponent bit-width.
}
\label{fig:hw_tradeoff}
\end{figure}
\section{Experimental Results}
\label{sec:results}
As discussed in Section~\ref{subsec:range_vs_precision}, we perform the weight
quantization based on {\em{Caffe}}~\cite{jia2014caffe}. We denote the representation as
($m$, $e$), where $m$ is the mantissa bit-width and $e$ is the exponent
bit-width. $e$=0 means fixed-point representation.
\begin{table}
\small
\caption{AlexNet accuracy with different representations.
$e$=0 means fixed-point representation. The floating-point
number is with the implicit bit.
}
\label{tbl:alex_accuracy}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& $e$=0 & $e$=1 & $e$=2 & $e$=3 & $e$=4 & $e$=5 \\
\hline
$m$=1 & 0.002 & 0.002 & 0.003 & 0.920 & 0.918 & 0.918 \\
\hline
$m$=2 & 0.003 & 0.003 & 0.003 & 0.983 & 0.982 & 0.982 \\
\hline
$m$=3 & 0.002 & 0.003 & 0.003 & 0.995 & 0.995 & 0.995 \\
\hline
$m$=4 & 0.016 & 0.003 & 0.003 & 0.997 & 1.001 & 1.001 \\
\hline
$m$=5 & 0.775 & 0.002 & 0.003 & 0.994 & 0.999 & 0.998 \\
\hline
$m$=6 & 0.979 & 0.002 & 0.003 & 0.995 & 0.999 & 0.999 \\
\hline
$m$=7 & 0.996 & 0.002 & 0.003 & 0.995 & 0.999 & 0.999 \\
\hline
$m$=8 & 0.998 & 0.002 & 0.003 & 0.995 & 1.000 & 1.000 \\
\hline
$m$=9 & 1.001 & 0.002 & 0.003 & 0.995 & 1.000 & 1.000 \\
\hline
$m$=10& 0.999 & 0.002 & 0.003 & 0.995 & 1.001 & 1.001 \\
\hline
\end{tabular}
\end{table}
We evaluate the network accuracy with different bit-width setting.
Table~\ref{tbl:alex_accuracy} highlights some results for AlexNet.
The network is non-functioning (i.e., with close to 0 accuracy)
when $e$ is small, i.e., with limited representation range.
To achieve full accuracy, AlexNet requires 8 bits for fixed-point
representation, i.e., (7, 0) with 1 sign bit, or 7 bits for floating-point
representation with
(3, 3) configuration.
If the implementation only targets AlexNet, the proposed
number representation can achieve 12.5\% weight storage reduction and
8\% power reduction in multiplication.
The benefit will increase for CNNs that require more fixed-point
bit-width.
As discussed earlier, one of the requirement for number representation
scheme is the consistency across different networks.
This is especially important for the hardware implementation
to be future-proof and viable for different CNN models.
Some results of the normalized accuracy of different network are
highlighted in Table~\ref{tbl:network_accuracy}.
The 7-bit fixed-point configuration used for AlexNet also works
for SqueezeNet, but is not adequate for GoogLeNet and VGG-16.
10-bit fixed-point representation is required to get consistent
accuracy across all networks used in this study.
\begin{table}
\small
\caption{Normalized Accuracy for different networks.}
\label{tbl:network_accuracy}
\begin{tabular}{|c|c|c|c|c|}
\hline
($m$, $e$) & AlexNet & SqueezeNet & GoogLeNet & VGG-16 \\
\hline
(7, 0) & 1.00 & 1.00 & 0.85 & 0.02 \\
\hline
(10, 0) & 1.00 & 0.99 & 0.99 & 1.00 \\
\hline
(3, 4) & 0.99 & 0.99 & 0.99 & 1.00 \\
\hline
\end{tabular}
\end{table}
By using proposed number representation scheme, we only need 7-bit
floating-point, i.e, (3, 4) configuration. Therefore, we can replace
11-bit weights (10-bit fixed-point number plus sign bit) with 8-bit weights
(3-bit mantissa, 4-bit exponent and 1 sign bit).
This results in 36\% storage reduction for weights and 50\% power reduction
in multiplication.
\section{Conclusion}
\label{sec:conclusion}
In this work, we propose CNN inference implementation with
floating-point weights and fixed-point activations. We give
the motivation for the proposed number representation scheme from
both algorithmic and hardware implementation perspectives.
The proposed scheme can reduce the weight storage by up to 36\%
and the multiplier power by up to 50\%.
Future work will investigate the impacts of network topology
and training on the number representation requirements.
|
\section{Introduction}\label{intro}
Jon Borwein discussing generalisations in the area of nonsmooth optimisation
\cite[p.\ 4]{Bor16} writes:
{\small
\begin{quote}
In his thesis Francis Clarke extended Moreau-Rockafellar max formula
to all locally Lipschitz functions. Clarke replaced
the right Dini directional derivative $D_hf(x)$ by
$$
D^c_hf(x) = \limsup_{0<t\to 0,y\to x}
\frac{f(y+th)-f(y)}{t}.
$$
Somewhat miraculously the mapping $p$ sending $h \to D^c_hf(x)$
is always continuous and sublinear in $h$, and so if we define
$\partial^Cf(x)=\partial p(0)=\{y\in X^*: \langle y, h\rangle\le D^c_hf(x), \forall h\in X\}$, Moreau-Rockafellar max formula leads directly to:
\medbreak\noindent
\textbf{Theorem} (Clarke).
{\it Let $f: E\to{\mathbb R}$ be a locally Lipschitz function. Then
\begin{equation}\label{intro1}
D^c_hf(x)=\sup_{y\in \partial^Cf(x)}\langle y, h\rangle
\end{equation}
for all $h\in E$. [...]
\if{In particular $\partial^Cf(x)$ is nonempty. Moreover, $\partial^Cf(x)$
is a singleton if and only if $f$ is strictly differentiable at $x$.}\fi}
\medbreak\noindent
In truth Clarke, appealing to Rademacher’s theorem, originally defined $\partial^Cf(x)$
as the closed convex hull of limits of nearby points of differentiability.
This makes (\ref{intro1}) seem even more remarkable.
There is, however, a dark side to the situation \cite[Cor.\ 9]{BMW01}.
Recall that a set in a Banach space is generic if it contains intersection of countably many dense open sets. The complement is thus very
small topologically.
\medbreak\noindent
\textbf{Theorem} (Generic triviality \cite{BMW01}).
{\it Let $A$ be an open subset
of a Banach space $X$. Then the
set of non-expansive functions on $A$ with $\partial^Cf(x)\equiv B_{X^*}$
for all $x$ in $A$ is generic in the uniform norm on $A$.}
\medbreak\noindent
In other words, in the sense of Baire category the Clarke subdifferential
(likewise the limiting subdifferential in the separable case) of almost
all functions contains no information at any point
of $A$. [...]
\if{Of course none of these functions are convex since then the function is generically strictly Fr\'echet differentiable.
This is the cost of abstraction --- if a construction always works for a
very broad class it usually works only passingly well.}\fi
So, for most Lipschitz functions the Clarke calculus is vacuous.
That is why serious researchers
work with well structured subclasses such \textit{semi-algebraic, partially smooth}
or \textit{essentially smooth} functions.
\end{quote}
}
The fact that subdifferentials cannot discriminate between functions
is a serious drawback according to
Terry Rockafellar \cite{Roc82} who argues:
{\small
\begin{quote}
In subgradient optimization, interest centers on methods for
minimizing $f$ that are based on being able to generate for each $x$
at least one (but not necessarily every) $y\in \partial^Cf(x)$, or perhaps just
an approximation of such a vector $y$. One of the main hopes is
that by generating a number of subgradients at various points in
some neighborhood of $x$, the behavior of $f$ around $x$ can roughly be
assessed. In the case of a convex function $f$ this is not just
wishful thinking, and a number of algorithms, especially those of
bundle type (e.g., Lemarechal 1975 and Wolfe 1975) rely on such an
approach. In the nonconvex case, however, there is the possibility,
without further assumptions on $f$ than local Lipschitz continuity,
that the multifunction $\partial^Cf: x\mapsto \partial^Cf(x)$
may be rather bizarrely disassociated from $f$.
An example given at the end of this section has
$f$ locally Lipschitzian, yet such that there exist many other locally
Lipschitzian functions $g$, not merely differing from $f$ by an additive
constant, for which $\partial^Cg(x) = \partial^Cf(x)$ for all $x$.
Subgradients alone cannot discriminate between the properties of
these different functions
and therefore cannot be effective in determining their local minima.
\if{
[...] The key seems to lie in postulating the existence of the ordinary
directional derivatives and some sort of relationship between them and $\partial^Cf$.
Mifflin (1977a and 1977b), most notably has worked in this direction.}\fi
\end{quote}
}
In this paper, we consider
lower semicontinuous functions and arbitrary subdifferentials.
We identify a large subclass of lower semicontinuous functions
whose radial subderivative at a given point of their domain
can be fully expressed in terms of the
subdifferential of the function at neighbouring points.
We show that these functions are precisely the functions
whose radial subderivative satisfies a mild directional upper
semicontinuity property, independently from any subdifferential. Such functions are said to be
upper semismooth. This class includes the proper lower semicontinuous
(directionally, approximately) convex functions,
the regular directionally Lipschitz functions,
the Mifflin semismooth functions,
the Thibault-Zagrodny directionally stable functions.
We show that, as expected, the class of upper semismooth functions
satisfies the subdifferential determination property,
that is, if two functions have the same subdifferential
and one of the functions is upper semismooth, then the two functions
are equal up to an additive constant.
\if{A class of functions $\mathcal{F}$ is said to have the
\textit{subdifferential determination property on an
open subset $\Omega \subset X$} if for any
two lsc functions $f, g:X\to \xR$ with $g$ in the class $\mathcal{F}$,
the following holds:
\begin{equation}\label{suddiffdet}
\partial f(x)=\partial g(x) \text{ for all } x\in \Omega
\Longrightarrow f=g + Const \text{ on }\Omega.
\end{equation}
}\fi
J.-J. Moreau \cite{Mor65} was the first to consider this property
for the class of proper lower semicontinuous convex functions defined on
Hilbert spaces. His result was later extended by Rockafellar \cite{Roc70}
to the same class of functions defined
on general Banach spaces. Since then
this property has been the object of intensive research and
various classes of non convex functions have been considered in this context.
We refer to the papers by L. Thibault and D. Zagrodny \cite{TZ05,TZ10}
for a detailed account of the history of this property.
The technique we use to prove the subdifferential
determination property is simple: from the subdifferential assumption on
the two functions, one of them being upper semismooth,
we derive an inequality between the radial subderivatives of
the functions; we then conclude by invoking a mean value theorem with
Dini subderivatives. The structure of the paper is as follows.
In Section \ref{Dinisect}, we revisit the mean value theorems with Dini
subderivatives. In Section \ref{subsubsect}, we discuss subderivatives
and subdifferentials and recall the duality formula linking them.
In Section \ref{subsubsect}, we define the class of upper semismooth functions
and give the main examples of functions in this class.
In Section \ref{determinationsect}, we prove our main theorems on the
subdifferential determination property.
\section{Mean value theorems with Dini subderivatives}\label{Dinisect}
The lower right-hand Dini derivative,
or lower radial subderivative from the direction $d=+1$,
of a function $\psi:{\mathbb R}\to{]}{-\infty},+\infty]$
at a point $t_0\in{\mathbb R}$ where $\psi$ is finite is denoted by
$$
\psi^r(t_0;+1):=\liminf_{t\searrow 0}\frac{\psi(t_0+t)-\psi(t_0)}{t}.
$$
Its upper version is denoted by
$$
\psi^r_+(t_0;+1):=\limsup_{t\searrow 0}\frac{\psi(t_0+t)-\psi(t_0)}{t}.
$$
\begin{proposition}[Mean value inequality]\label{mvi}
Let $\psi:[0,1]\to{]}{-\infty},+\infty]$ be lower semicontinuous on $[0,1]$ and finite at $0$.
Then, for every real number $\lambda\le \psi(1)-\psi(0)$, there
exists $t_0\in [0,1{[}$ such that $\psi(t_0)\le \psi(0)+t_0\lambda$ and
$\lambda\le \psi^r(t_0;+1).$
\end{proposition}
\begin{proof}
For completeness, we recall the elementary argument,
as given, e.g., in \cite[Lemma 4.1]{JL13} or in \cite[Lemma 3.1]{JL14}.
For $t\in [0,1]$, let $g(t):=\psi(t)-t \lambda$.
The function $g:[0,1]\to {]}{-\infty},+\infty]$ is lsc
on the compact $[0,1]$ and $g(0)=\psi(0)\le \psi(1)-\lambda=g(1)$.
Hence $g$ attains its
minimum on $[0,1]$ at a point $t_0\ne 1$.
Consequently, $\psi(t_0)-t_0\lambda=g(t_0)\le g(0)=\psi(0)$ and
since $g(t_0+t)\ge g(t_0)$ for every $t\in {]}0,1-t_0]$,
it follows that
\[
\forall t\in {]}0,1-t_0],\quad \frac{\psi(t_0+t)-\psi(t_0)}{t} \ge \lambda.
\]
Passing to the limit inferior as $t\searrow 0$, we get
$\psi^r(t_0;+1)\ge \lambda$ as claimed.
\end{proof}
\begin{proposition}[Mean value theorem: semicontinuous version]
\label{recov-subdiv-basic}
Let $\varphi:[0,1]\to{]}{-\infty},+\infty]$ be lower semicontinuous on $[0,1]$ and finite at $0$
and let
$\gamma:[0,1]\to [-\infty,+\infty{[}$ be upper semicontinuous on $[0,1]$
and finite at $0$.
Assume that for every $t\in [0,1[$,
there exists a real number $\rho(t)$ such that
$\varphi^r(t;+1)\le \rho(t)\le \gamma^r(t;+1)$.
Then
\begin{equation*}
\varphi(1)-\varphi(0)\le \gamma(1)-\gamma(0).
\end{equation*}
\end{proposition}
\begin{proof}
Let $\psi:=\varphi-\gamma$. Then $\psi:[0,1]\to{]}{-\infty},+\infty]$ is lower semicontinuous
on $[0,1]$ and finite at $0$.
So, according to Proposition \ref{mvi}, for every real number
$r\le \psi(1)-\psi(0)$, there
exists $t_0\in [0,1[$ such that $r\le \psi^r(t_0;+1)$.
There exists a real number $\rho(t_0)$ such that
\begin{equation}\label{recov-subdiv11}
\varphi^r(t_0;+1)\le \rho(t_0)\le \gamma^r(t_0;+1).
\end{equation}
Since
\begin{align*}
r\le \psi^r(t_0;+1)=
\liminf_{t\searrow 0}\left(\frac{\varphi(t_0+t)-\varphi(t_0)}{t}
-\frac{\gamma(t_0+t)-\gamma(t_0)}{t}\right)
\end{align*}
and since, by (\ref{recov-subdiv11}),
\begin{align*}
\rho(t_0)\le \liminf_{t\searrow 0}\frac{\gamma(t_0+t)-\gamma(t_0)}{t},
\end{align*}
for every $\varepsilon>0$ one can find $t_\varepsilon>0$ such that for all $t\in ]0,t_\varepsilon]$
\begin{align*}
r-\varepsilon< \frac{\varphi(t_0+t)-\varphi(t_0)}{t}
-\frac{\gamma(t_0+t)-\gamma(t_0)}{t}
\quad\text{ and }\quad
\rho(t_0)-\varepsilon\le \frac{\gamma(t_0+t)-\gamma(t_0)}{t},
\end{align*}
hence, for all $t\in {]}0,t_\varepsilon]$,
\begin{align*}
r+\rho(t_0)-2\varepsilon< \frac{\varphi(t_0+t)-\varphi(t_0)}{t}.
\end{align*}
Passing to the limit inferior, we get
\begin{align*}
r+\rho(t_0)-2\varepsilon\le \varphi^r(t_0;+1).
\end{align*}
Since $\varphi^r(t_0;+1)\le \rho(t_0)$ by (\ref{recov-subdiv11}),
it follows that $r-2\varepsilon\le 0$.
As $\varepsilon>0$ was arbitrary, we conclude that $r\le 0$.
Therefore, for every real number $r\le \psi(1)-\psi(0)$ one has $r\le 0$,
proving that
$\psi(1)-\psi(0)\le 0$. This amounts to
$\varphi(1)-\varphi(0)\le \gamma(1)-\gamma(0)$, as claimed.
\end{proof}
\begin{remark}
(a) In Proposition \ref{recov-subdiv-basic}, the semicontinuity conditions
on $\varphi$ and $\gamma$ cannot be relaxed. Indeed, consider
the functions $\varphi,\gamma:[0,1]\to{\mathbb R}$
defined by
$$
\varphi(t):= 0 \quad\mbox{for all } t\in [0,1]
\quad \mbox{and}\quad
\gamma(t):=\left\{
\begin{array}{ll}
1 & \mbox{if~~} t\in {[}0,1/2{[}\\
0 & \mbox{if~~} t\in {[}1/2,1{]}
\end{array}
\right..
$$
Then, $\varphi:[0,1]\to{\mathbb R}$ is continuous,
$\gamma:[0,1]\to{\mathbb R}$ is continuous except at point $1/2$ where it is
merely right-continuous, and
$\varphi^r(t;+1)= \gamma^r(t;+1)=0$ for all $t\in[0,1[$.
Yet the conclusion in Proposition \ref{recov-subdiv-basic} is false:
$\varphi(1)-\varphi(0)=0> \gamma(1)-\gamma(0)=-1$.
\smallbreak
(b) In Proposition \ref{recov-subdiv-basic}, the finiteness conditions
$\varphi^r(t;+1)<+\infty$ and $\gamma^r(t;+1)>-\infty$ for every $t\in [0,1[$,
cannot be relaxed.
Indeed, consider the functions $\varphi,\gamma:[0,1]\to{\mathbb R}$
defined by
$$
\varphi(t):=
\left\{
\begin{array}{ll}
2 & \mbox{if~~} t\in {]}0,1]\\
0 & \mbox{if~~} t= 0
\end{array}
\right.
\quad \mbox{and}\quad \gamma(t)=\sqrt{t}\quad\mbox{for all } t\in [0,1].
$$
Then, $\varphi:[0,1]\to{]}{-\infty},+\infty]$ is lower semicontinuous on $[0,1]$ with finite values,
$\gamma:[0,1]\to{\mathbb R}$ is continuous and
$\varphi^r(t;+1)\le \gamma^r(t;+1)$ for every $t\in [0,1[$.
Moreover, $\varphi^r(t;+1)$ and $\gamma^r(t;+1)$ are finite for every $t\in ]0,1[$,
but $\varphi^r(0;+1)= \gamma^r(0;+1)=+\infty$, and indeed
the conclusion of Proposition \ref{recov-subdiv-basic} is false:
$\varphi(1)-\varphi(0)=2> \gamma(1)-\gamma(0)=1$.
\end{remark}
\begin{proposition}[Mean value theorem: continuous version]
\label{recov-subdiv-basic-continuous}
Let $\varphi,\gamma:[0,1]\to{\mathbb R}$ be continuous.
Assume there is a countable subset $C\subset [0,1]$ such
that for every $t\in [0,1]\setminus C$,
there exists a real number $\rho(t)$ such that
$\varphi^r_+(t;+1)\le \rho(t)\le \gamma^r(t;+1)$.
Then
\begin{equation*}
\varphi(1)-\varphi(0)\le \gamma(1)-\gamma(0).
\end{equation*}
\end{proposition}
\begin{proof}
We follow the pattern of the proof of \cite[(8.5.1)]{Die69}.
For any $\varepsilon>0$, we will show that
$$
\varphi(1)-\varphi(0)\le \gamma(1)-\gamma(0) + 3\varepsilon;
$$
the left hand side being independent of $\varepsilon$, this will complete the proof.
Let $C:=\{c_n: n\in {\mathbb N}\}$ be the given countable subset of $[0,1]$.
Consider the set
$$
A:=\{ t\in [0,1]: \forall t'\in [0,t],\
\varphi(t')-\varphi(0)\le \gamma(t')-\gamma(0) + \varepsilon t'+\varepsilon \sum_{c_n<t'} 2^{-n}\}.
$$
It is clear that $0\in A$ and that if $t\in A$, then $[0,t]\subset A$.
Let $t_0=\sup A$. From the continuity of $\varphi$ and $\gamma$ it follows that
$t_0\in A$, so $[0,t_0]=A$. Therefore we need only prove that $t_0=1$.
\smallbreak
Suppose $t_0<1$.
If $t_0\not\in C$, there is a real number $\rho(t_0)$ such that
$$\varphi^r_+(t_0;+1)\le \rho(t_0)\le \gamma^r(t_0;+1),$$
so, by definition of the subderivatives, we can find $\eta>0$ such that
for every $s\in {]}0,\eta]$,
\begin{gather*}
\varphi(t_0+s)-\varphi(t_0)\le (\rho(t_0)+\varepsilon/2)s\quad\text{and}\quad
(\rho(t_0)-\varepsilon/2)s\le \gamma(t_0+s)-\gamma(t_0),
\end{gather*}
hence
$$
\varphi(t_0+s)-\varphi(t_0)\le \gamma(t_0+s)-\gamma(t_0)+\varepsilon s,
$$
and since $t_0\in A$, we deduce
\begin{align*}
\varphi(t_0+s)-\varphi(0) &\le \gamma(t_0+s)-\gamma(0) +
\varepsilon (t_0+s)+\varepsilon \sum_{c_n<t_0} 2^{-n}\\
&\le \gamma(t_0+s)-\gamma(0) +
\varepsilon (t_0+s)+\varepsilon \sum_{c_n<t_0+s} 2^{-n},
\end{align*}
hence $t_0+\eta\in A$ contrary to the definition of $t_0$.
If $t_0\in C$, the set $\{n\in{\mathbb N}: t_0=c_n\}$ is not empty; by continuity
of $\varphi$ and $\gamma$, we can find $\eta>0$ such that
for every $s\in {]}0,\eta]$,
\begin{gather*}
\varphi(t_0+s)-\varphi(t_0)\le (\varepsilon/2) \sum_{c_n=t_0} 2^{-n}
\quad\text{and}\quad
0\le \gamma(t_0+s)-\gamma(t_0)+(\varepsilon/2) \sum_{c_n=t_0} 2^{-n},
\end{gather*}
hence from $t_0\in A$ we deduce again
\begin{align*}
\varphi(t_0+s)-\varphi(0) &\le \gamma(t_0+s)-\gamma(0) +
\varepsilon t_0+\varepsilon \sum_{c_n<t_0+s} 2^{-n}\\
&\le \gamma(t_0+s)-\gamma(0) +
\varepsilon (t_0+s)+\varepsilon \sum_{c_n<t_0+s} 2^{-n},
\end{align*}
which is a contradiction.
\end{proof}
\section{Subderivatives and subdifferentials}\label{subsubsect}
In the sequel, $X$ is a real Banach space,
$X^*$ is its topological dual,
and $\langle .,. \rangle$ is the duality pairing.
For $x, y \in X$, we let $[x,y]:=\{ x+t(y-x) : t\in[0,1]\}$;
the sets $]x,y[$ and $[x,y[$ are defined accordingly.
Set-valued operators $T:X\rightrightarrows X^*$
are identified with their graph $T\subset X\times X^*$.
All extended-real-valued functions $f : X\to{]}{-\infty},+\infty]$ are assumed to be
lower semicontinuous (lsc)
and \textit{proper}, which means that
the set $\dom f:=\{x\in X: f(x)<\infty\}$ is non-empty.
\medbreak
The framework, terminology and notation are the same as in our work
\cite{Las16}. For the reader's convenience, we briefly recall
the main definitions and facts.
For a lsc function $f:X\to{]}{-\infty},+\infty]$, a point $\bar{x}\in\dom f$ and
a direction $u\in X$,
we consider the following basic subderivatives:
- the (lower right Dini) \textit{radial subderivative}:
\begin{equation*}
f^r(\bar{x};u):=\liminf_{t\searrow 0}\,\frac{f(\bar{x}+tu)-f(\bar{x})}{t},
\end{equation*}
its upper version:
\begin{equation*}
f^r_+(\bar{x};u):=\limsup_{t\searrow 0}\,\frac{f(\bar{x}+tu)-f(\bar{x})}{t},
\end{equation*}
and its upper strict version (the \textit{Clarke subderivative}):
\begin{equation*}
f^0(\bar{x};u):=
\limsup_{t \searrow 0 \atop{(x,f(x)) \to (\bar{x},f(\bar{x}))}}\frac{f(x+tu) -f(x)}{t};
\end{equation*}
- the (lower right Dini-Hadamard) \textit{directional subderivative}:
\begin{equation*}
f^d(\bar{x};u):=
\liminf_{t \searrow 0 \atop{u' \to u}}\frac{f(\bar{x}+tu')-f(\bar{x})}{t},
\end{equation*}
and its upper strict version (the Clarke-Rockafellar subderivative):
\begin{equation*}
f^\uparrow(\bar{x};u):= \sup_{\delta>0}
\limsup_{t \searrow 0 \atop{(x,f(x)) \to (\bar{x},f(\bar{x}))}}
\inf_{u' \in B_{\delta}(u)}\frac{f(x+tu') -f(x)}{t}.
\end{equation*}
It is immediate from these definitions that the following inequalities hold
($\rightarrow$ means $\le$):
\begin{align*}
f^r(\bar{x};u) & \rightarrow f^r_+(\bar{x};u)\rightarrow f^0(\bar{x};u)\\
\uparrow \quad & \qquad\qquad\qquad\quad\uparrow\\
f^d(\bar{x};u) & \qquad\longrightarrow \quad\quad f^\uparrow(\bar{x};u)
\end{align*}
For $f$ locally Lipschitz at $\bar{x}$,
one has $f^r(\bar{x};u)=f^d(\bar{x};u)$ and $f^0(\bar{x};u)=f^\uparrow(\bar{x};u)$.
For $f$ lsc convex, one has $f^d(\bar{x};u)=f^\uparrow(\bar{x};u)$.
A function $f$ satisfying such an equality is called \textit{regular}.
However, in general, $f^d(\bar{x};u)<f^\uparrow(\bar{x};u)$.
\medbreak
Next, given a lsc function $f:X\to{]}{-\infty},+\infty]$ and a point $\bar{x}\in\dom f$,
we consider the following two basic subsets of the dual space $X^*$:
- the \textit{Moreau-Rockafellar subdifferential}
(the subdifferential of convex analysis):
\begin{equation*}\label{convex-sdiff}
\partial_{MR} f (\bar{x}) :=
\{ x^* \in X^* : \langle x^*,y-\bar{x}\rangle + f(\bar{x}) \leq f(y),\, \forall y \in X \};
\end{equation*}
- the \textit{Clarke subdifferential}, associated to the Clarke-Rockafellar subderivative:
\begin{eqnarray*}\label{Csub}
\partial_{C} f(\bar{x}) := \{x^* \in X^* : \langle x^*,u\rangle \leq
f^\uparrow(\bar{x};u), \, \forall u \in X\}.
\end{eqnarray*}
All the classical subdifferentials (proximal,
Fr\'echet, Hadamard, Ioffe, Michel-Penot, \ldots)
lie between these two subsets,
and for a lsc convex $f$, all these subdifferentials coincide.
\smallbreak
In the sequel, we call \textit{subdifferential} any operator $\partial$ that associates
a set-valued mapping $\partial f: X \rightrightarrows X^\ast$
to each function $f$ on $X$ so that
\begin{equation*}\label{inclusdansClarke}
\partial_{MR} f\subset \partial f\subset \partial_{C} f
\end{equation*}
and the following \textit{Separation Principle}
is satisfied on $X$:
\medbreak
(SP)
\textit{For any lsc $f,\varphi$ with $\varphi$ convex Lipschitz
near $\bar{x}\in\dom f $,
if $f+\varphi$ admits a local minimum at $\bar{x}$, then
$0\in \widehat{\del} f(\bar{x})+ \partial \varphi(\bar{x}),$ where
\begin{multline*
\widehat{\del} f(\bar{x}):= \{\, \bar{x}^*\in X^*: \mbox{there is a net }((x_\nu,x^*_\nu))_\nu\subset \partial f \mbox{ with }\\
(x_\nu,f(x_\nu))\to (\bar{x},f(\bar{x})),\ x^*_\nu{\stackrel{w^*}{\longrightarrow}}\; \bar{x}^*,\ \limsup_\nu\,\langle x^*_\nu,x_\nu-\bar{x}\rangle\le 0\,\}.
\end{multline*}
}
The Clarke subdifferential, the Michel-Penot (moderate) subdifferential and
the Ioffe subdifferential satisfy the Separation Principle in any Banach space.
The elementary subdifferentials (proximal, Fr\'echet, Hadamard, \ldots),
as well as their viscosity and limiting versions,
satisfy the Separation Principle in appropriate Banach spaces.
See, e.g. \cite{Iof12,JL13,Pen13} and the references therein.
\smallbreak
Subdifferentials satisfying the Separation Principle
are densely defined:
\begin{theorem}[Density of subdifferentials]\label{dense}
Let $X$ be a Banach space,
$f:X\to{]}{-\infty},+\infty]$ be lsc and $\bar{x}\in\dom f$.
Then, there exists a sequence $((x_n,x_n^*))_n\subset\partial f$ such that
$x_n\to \bar{x}$, $f(x_n)\to f(\bar{x})$ and $\limsup_n \langle x_n^*, x_n-\bar{x}\rangle\le 0$.
\end{theorem}
\begin{proof}
See \cite[Theorem 2.1]{JL14}.
\end{proof}
A sequence $(x_n)\subset X$ is said to be
\textit{directionally convergent to $\bar{x}$ in the direction $v\in X$},
written $x_n\to_v \bar{x}$,
if there are two sequences $t_n\searrow 0$ (that is, $t_n\to 0$ with $t_n>0$)
and $v_n\to v$ such that
$x_n=\bar{x} + t_n v_n$ for all $n$.
Observe that for $v=0$, $x_n\to_v\bar{x}$ simply means $x_n\to\bar{x}$.
\smallbreak
We call \textit{subderivative associated to a subdifferential $\partial f$}
at a point $(\bar{x},u)\in \dom f\times X$, the \textit{support function} of
the set $\partial f(\bar{x})$ in the direction $u$,
which we denote by
$$
f^\partial(\bar{x};u):=
\sup \,\{\langle \bar{x}^*,u \rangle : \bar{x}^*\in\partial f(\bar{x})\}.
$$
\smallbreak
Subderivatives and subdifferentials are linked by the following formula
where, given $f:X\to\xR$, we have denoted by $f':\dom f\times X\to \overline{{\mathbb R}}$
any function lying between the subderivatives $f^d$ and $f^\uparrow$, that is:
$f^d\le f'\le f^\uparrow$:
\begin{theorem}[Subderivative-subdifferential duality formula]\label{formula}
Let $X$ be a Banach space,
$f:X\to{]}{-\infty},+\infty]$ be lsc, $\bar{x}\in\dom f$ and $u\in X$.
Then, for any direction $v\in X$ and any real number $\alpha\ge 0$, one has
\begin{subequations}\label{formula0}
\begin{align}
\limsup_{x\to_v\bar{x}} f^r(x;u+\alpha (\bar{x}-x))&=
\limsup_{x\to_v\bar{x}} f'(x;u+\alpha (\bar{x}-x)) \label{formula0a}\\
&=\limsup_{x\to_v\bar{x}}\,f^\partial(x;u+\alpha (\bar{x}-x)).
\label{formula0b}
\end{align}
\end{subequations}
\end{theorem}
\begin{proof}
See \cite[Theorem 3]{Las16}.
\end{proof}
\begin{remark}
(a) For $f$ locally Lipschitz at $\bar{x}$, the formulas \eqref{formula0}
do not depend on $\alpha\ge 0$ since
\begin{equation*}\label{formula0lip}
\limsup_{x\to_v\bar{x}} f^r(x;u+\alpha (\bar{x}-x))=\limsup_{x\to_v\bar{x}} f^r(x;u),
\end{equation*}
but they may depend on the direction $v\in X$:
for $f:x\in {\mathbb R}\mapsto f(x):=-|x|$ and $u\ne 0$, one has
$$\limsup_{x\to_u 0} f^r(x;u)=-|u|<\limsup_{x\to 0} f^r(x;u)=|u|.$$
\smallbreak
(b) For arbitrary lsc $f$, the value of the expressions
in \eqref{formula0}
may depend on $\alpha\ge 0$ even for convex $f$.
Indeed, in general for a lsc convex $f$ one has
$$f^r(\bar{x};u)<\limsup_{x\to\bar{x}} f^r(x;u),$$
while always (see \cite[Proposition 4]{Las16})
\begin{equation}\label{convex-uss}
f^r(\bar{x};u)=\inf_{\alpha\ge 0}\limsup_{x\to\bar{x}} f^r(x;u+\alpha (\bar{x}-x)).
\end{equation}
\end{remark}
\section{Upper semismooth functions}\label{usssect}
A lsc function $f:X\to{]}{-\infty},+\infty]$ is said to be \textit{radially accessible} at a point $\bar{x}\in\dom f$ from a direction $u\in X$ provided
$$f(\bar{x})=\liminf_{t\searrow 0}f(\bar{x}+tu),$$
or equivalently, provided
there exists a sequence $t_n\searrow 0$ such that
$f(\bar{x}+t_n u)\to f(\bar{x})$.
Every convex lsc function $f:X\to{]}{-\infty},+\infty]$
is radially accessible at $\bar{x}$ from any $u\in X$ such that
$\bar{x}+u\in \dom f$. On the other hand,
it is easily seen that if a lsc $f$ satisfies
$f^r(\bar{x};u)<+\infty$, then $f$ is radially accessible at $\bar{x}$ from $u$.
The converse is not true: the function
$f:t\in{\mathbb R}\mapsto \sqrt{|t|}$ is continuous, yet $f^r(0;u)=+\infty$.
For more examples and properties, see \cite{Las16}.
\begin{proposition}[Radial stability of the upper radial subderivative]\label{devdir}
Let $X$ be a Hausdorff locally convex space, $f:X\to{]}{-\infty},+\infty]$ be lsc,
$\bar{x}\in\dom f$ and $u\in X$
such that $f$ is radially accessible at $\bar{x}$ from $u$.
Then, there is a sequence
$\mu_n\searrow 0$ such that $f(\bar{x}+\mu_n u)\to f(\bar{x})$ and
\begin{equation}\label{applimvi1}
f^r_+(\bar{x};u)\le \liminf_{n\to +\infty}f^r(\bar{x}+\mu_n u;u).
\end{equation}
In particular,
\begin{equation}\label{below00}
f^r_+(\bar{x};u)\le \inf_{\alpha\ge 0}
\limsup_{x\to_u\bar{x}} f^r(x;u+\alpha (\bar{x}-x)).
\end{equation}
\end{proposition}
\begin{proof}
See \cite[Proposition 7]{Las16}.
\end{proof}
Theorem \ref{formula}, Formula \eqref{convex-uss}
for convex lsc functions and Proposition \ref{devdir}
suggest to consider the following class of lsc functions.
\begin{defn}\label{defdsemismooth}
A lsc function $f:X\to{]}{-\infty},+\infty]$ is said to be \textit{upper semismooth}
at a point $\bar{x}\in\dom f$ in the direction $u\in X$
provided
\begin{align}
\inf_{\alpha\ge 0}\limsup_{x\to_u \bar{x}} f^r(x; u +\alpha (\bar{x}-x))\le f^r(\bar{x};u).
\label{dsemismooth}
\end{align}
\end{defn}
\begin{remark}\label{remdefdsemismooth}
(a) In case $f$ is locally Lipschitz at $\bar{x}$, \eqref{dsemismooth}
boils down to
\begin{align}\label{dsemismoothLip}
\limsup_{x\to_u \bar{x}} f^r(x; u)\le f^r(\bar{x};u). \tag{\ref{dsemismooth}Lip}
\end{align}
We then essentially recover the class of locally Lipschitz upper semismooth
functions considered by Borwein-Moors \cite[p.\ 305]{BM98a}
(with the slight difference that
the upper radial subderivative $f^r_+$ is used there
instead of the lower radial subderivative $f^r$ used here).
As observed in \cite[p.\ 305]{BM98a}, the terminology is
justified by the characterization of semismooth functions given by
Correa-Jofr\'e \cite[Corollary 6.3]{CJ89}
(see also Proposition \ref{exsemismooth}\,(d) below).
\smallbreak
(b) In case $f$ is radially accessible at $\bar{x}$ from the direction $u$,
the inequality \eqref{dsemismooth} becomes:
\begin{align}\label{dsemismoothequal}
f^r(\bar{x};u)=
\inf_{\alpha\ge 0}\limsup_{x\to_u \bar{x}} f^r(x; u +\alpha (\bar{x}-x)),
\tag{\ref{dsemismooth}bis}
\end{align}
and in addition, $f^r(\bar{x};u)=f^r_+(\bar{x};u)$,
i.e.\ the lower radial derivative and its upper version
coincide at $\bar{x}$ in the direction $u$.
This follows by combining \eqref{dsemismooth} with the inequality
\eqref{below00} in Proposition \ref{devdir}.
\smallbreak
(c) One can have $f^r(\bar{x};u)=\pm\infty$ in \eqref{dsemismooth}.
For example, the continuous function
$x\mapsto \sqrt{|x|}$ is upper semismooth
at $\bar{x}=0$ in the direction $u=1$ with $f^r(0;1)=+\infty$,
while the continuous function
$x\mapsto -\sqrt{|x|}$ is upper semismooth
at $\bar{x}=0$ in the direction $u=1$ with $f^r(0;1)=-\infty$.
\end{remark}
\medbreak
We shall also consider a strict variant of the above notion:
\begin{defn}\label{defsemismooth}
A lsc function $f:X\to{]}{-\infty},+\infty]$ is said to be \textit{strictly upper semismooth}
at a point $\bar{x}\in\dom f$ in the direction $u\in X$
provided
\begin{align}
\inf_{\alpha\ge 0}\limsup_{x\to \bar{x}} f^r(x; u +\alpha (\bar{x}-x))\le f^r(\bar{x};u).
\label{semismooth}
\end{align}
In fact, equality holds in \eqref{semismooth}
since the reverse inequality is always true.
\end{defn}
\begin{remark}\label{remdefsemismooth}
(a) As above, in case $f$ is locally Lipschitz at $\bar{x}$,
\eqref{semismooth} boils down to
\begin{align}\label{semismoothLip}
\limsup_{x\to \bar{x}} f^r(x; u)\le f^r(\bar{x};u).
\tag{\ref{semismooth}Lip}
\end{align}
Since $\limsup_{x\to \bar{x}} f^r(x; u)=f^0(\bar{x};u)$ according to
Borwein-Str{\'o}jwas \cite[Theorem 2.1]{BS89},
the inequality \eqref{semismoothLip} is actually equivalent to the equality
$f^r(\bar{x};u)=f^0(\bar{x};u)$.
The terminology is therefore justified since the latter equality
means that the lower radial subderivative and its upper strict version
coincide at $\bar{x}$ in the direction $u$.
Locally Lipschitz functions satisfying such an equality in every direction
$u\in X$ are called (Clarke) \textit{regular at $\bar{x}$} (see also below the extension
of this concept to the general case of lsc functions).
\smallbreak
(b) Evidently, \eqref{semismooth} is more demanding than \eqref{dsemismooth},
so every strictly upper semismooth function is upper semismooth.
The converse is not true:
the locally Lipschitz function $f:x\in{\mathbb R}\mapsto -|x|$ satisfies
$$\limsup_{x\to_u 0} f^r(x;u)=-|u|=f^r(0;u)<\limsup_{x\to 0} f^r(x;u)=|u|,$$
so $f$ is upper semismooth at $\bar{x}=0$ from any $u\ne 0$ but not strictly
upper semismooth.
\end{remark}
We now proceed to give examples of strictly and non-strictly upper semismooth
functions.
Let us recall the definition of the concepts we shall consider.
Let $X$ be a Banach space,
$f:X\to{]}{-\infty},+\infty]$ be lsc, $\bar{x}\in\dom f$ and $u\in X$, $u\ne 0$.
Then $f$ is said to be:
$\bullet$\
\textit{semismooth} (Mifflin \cite{Mif77})
at $\bar{x}$ provided $f$ is locally Lipschitz and for all $u\in X$,
$$((x_n,x^*_n))\subset \partial_C f \mbox{ with }
x_n\to_u \bar{x} \Rightarrow \langle x^*_n,u\rangle \to f^r(\bar{x};u).$$
\smallbreak
$\bullet$\
\textit{directionally Lipschitz at $\bar{x}$ with respect to $u$}
(Rockafellar \cite{Roc80}) if $f$ is lsc and
\begin{equation}\label{dirlip}
\limsup_{\stackrel{(x,f(x))\to (\bar{x},f(\bar{x}))}{t\searrow 0, v\to u}} \frac{f(x+tv)-f(x)}{t}<\infty.
\end{equation}
\smallbreak
$\bullet$\
\textit{regular at $\bar{x}$} (Rockafellar \cite{Roc79})
if $f^d(\bar{x};v)=f^\uparrow(\bar{x};v)$ for every $v\in X$.
\smallbreak
$\bullet$\
\textit{directionally approximately convex at $\bar{x}$}
(see Daniilidis-Georgiev \cite{DG04},
Daniilidis-Jules-Lassonde \cite{DJL09} and the references therein), if for every
$u\in S_X$ and $\varepsilon>0$ there exists $\delta>0$
such that for all $x,y\in B(\bar{x},\delta)$, with $x\not=y$ and
$(x-y)/\|x-y\|\in B(u,\delta)$, and all $t\in [0,1]$
\begin{equation}\label{dac}
f(tx+(1-t)y)\leq tf(x)+(1-t)f(y)+\varepsilon t(1-t)\Vert x-y\Vert.
\end{equation}
(In finite-dimensional spaces, a
locally Lipschitz function is (directionally) approximately convex
if and only if it is lower-$C^1$,
cf.\ \cite{Spi81,DG04}.)
\begin{proposition}[Examples of (strictly) upper semismooth functions]
\label{exsemismooth}
Let $X$ be a Banach space,
$f:X\to{]}{-\infty},+\infty]$ be lsc, $\bar{x}\in\dom f$ and $u\in X$ with $u\ne 0$
and $\bar{x}+u\in\dom f$.
Each of the following $f$ is strictly upper semismooth at $\bar{x}$ in the direction $u$:
\smallbreak
{\rm(a)}
$f$ is directionally Lipschitz at $\bar{x}$ with respect to $u$ and regular at $\bar{x}$;
\smallbreak
{\rm(b)}
$f$ is convex;
\smallbreak
{\rm(c)}
$f$ is directionally approximately convex.
\smallbreak\noindent
The following $f$ is upper semismooth at $\bar{x}$ in the direction $u$:
\smallbreak
{\rm(d)}
$f$ is locally Lipschitz and semismooth at $\bar{x}$.
\end{proposition}
\begin{proof}
(a)
For $f$ directionally Lipschitz at $\bar{x}$ with respect to $u$, one has
$f^\uparrow(\bar{x};u)=f^0(\bar{x};u)$
(cf.\ \cite[Theorem 3]{Roc80}). If moreover $f$ is regular at $\bar{x}$,
that is $f^d(\bar{x};v)=f^\uparrow(\bar{x};v)$ for every $v\in X$,
we derive that $f^d(\bar{x};u)=f^0(\bar{x};u)$.
Then,
\begin{equation*}
f^r(\bar{x};u)\ge f^d(\bar{x};u)=f^0(\bar{x};u)=\limsup_{x\to \bar{x}} f^r(x; u)\ge
\inf_{\alpha\ge 0}\limsup_{x\to \bar{x}} f^r(x; u +\alpha (\bar{x}-x)).
\end{equation*}
Hence \eqref{semismooth} is satisfied.
\smallbreak
(b) We know from \cite[Proposition 4]{Las16} that,
if $f$ is lsc\ convex, then
$$f^r(\bar{x};u)=\inf_{\alpha\ge 0}\limsup_{x\to\bar{x}} f^r(x;u+\alpha (\bar{x}-x)).$$
Hence, \eqref{semismooth} holds.
\smallbreak
(c)
A directionally approximately convex function satisfies the following property
(cf.\ \cite[Proposition 1]{DJL09}):
for every $\epsilon>0$
and $u\in S_X$ there exists $\delta>0$ such that for all
$x\in B(\bar{x},\delta)$ and all $v\ne 0$ so that
$x+v\in B(\bar{x},\delta)$ and $v/\|v\|\in B(u,\delta)$, one has
\begin{equation}\label{technic2}
f^\uparrow(x;v)\le f(x+v)-f(x)+\varepsilon\|v\|.
\end{equation}
Without loss of generality, we may assume that the given $u$ belongs to $S_X$.
Let $\epsilon>0$ and $\delta>0$ such that \eqref{technic2} holds.
Fix $0<t<\delta$ and consider any $x\in B(\bar{x},t\delta/2)$.
Let $v=tu+\bar{x}-x$. Then $x+v\in B(\bar{x},\delta)$
and $v=t(u+(\bar{x}-x)/t\in {]}0,\delta B(u,\delta){[}$.
It follows from \eqref{technic2} that for any such $x$,
\begin{equation*}
f^\uparrow(x;tu+\bar{x}-x)\le f(\bar{x}+tu)-f(x)+\varepsilon\|tu+\bar{x}-x\|.
\end{equation*}
Hence, for every $0<t<\delta$, since $f$ is lsc\ at $\bar{x}$,
\begin{equation*}
\limsup_{x\to\bar{x}}f^\uparrow(x;tu+\bar{x}-x)\le f(\bar{x}+tu)-f(\bar{x})+t\varepsilon.
\end{equation*}
So, for every $\varepsilon>0$,
\begin{equation*}
\liminf_{t\searrow 0}
\limsup_{x\to\bar{x}}f^\uparrow(x;\frac{tu+\bar{x}-x}{t})\le
\liminf_{t\searrow 0}\frac{f(\bar{x}+tu)-f(\bar{x})}{t}+\varepsilon.
\end{equation*}
Consequently,
\begin{equation*}
\inf_{\alpha\ge 0}
\limsup_{x\to\bar{x}}f^\uparrow(x;u+\alpha(\bar{x}-x))\le
f^r(\bar{x};u).
\end{equation*}
Hence \eqref{semismooth} holds since we can replace $f^\uparrow$
by $f^r$ in the left hand side according to Theorem \ref{formula}.
\smallbreak
(d)
By \cite[Corollary 6.3]{CJ89},
a locally Lipschitz function $f$ is semismooth if and only if
$$f^r(\bar{x};u)=\lim_{x\to_u \bar{x}} f^r(x; u).$$
Hence (\ref{dsemismooth}Lip) holds.
\end{proof}
Besides the examples given in Proposition \ref{exsemismooth},
more elaborated classes of functions have been considered
to deal with the subdifferential determination property.
Classes of functions and results based on measure and integration theories
(e.g.\ Borwein-Moors \cite{BM98a} or Thibault-Zagrodny \cite{TZ10})
are discussed elsewhere.
Here, we discuss further the class of functions introduced by
L. Thibault and D. Zagrodny in \cite{TZ05}:
given a subdifferential $\partial$, a lsc function
$g:X\to{]}{-\infty},+\infty]$ is called \textit{$\partial$-subdifferentially and directionally stable}
(\textit{sds} for short) on $\Omega$
provided that for every $u \in \Omega \cap\dom \partial g$ and $v\in \Omega \cap\dom g$
with $v\ne u$, the following properties hold:
\begin{itemize}\itemsep-1ex\topsep0pt
\item[(i)] the function $t\mapsto \gamma (t):= g(u+t(v-u))$ is finite and
continuous on $[0, 1]$;
\item[(ii)] for any $t\in [0, 1{[}$, the right derivative
$\gamma_+'(t)$ exists and is less than $+\infty$;
\item[(iii)] for each fixed $y\in [u, v{[}$ and for each real number $\varepsilon > 0$, there exists some
$r_0 \in {]}0, 1{[}$ such that for any $w=y+r(v-u)$ with
$r \in {]}0, r_0 ]$ and for every
$(x_n,x^*_n) \in \partial g$ with $x_n \to x_0 \in [y, w{[}$ one has
\begin{equation}\label{sds}
\limsup_{n\to\infty}\, \langle x^*_n,w-x_n\rangle \le g^r(y;w-x_0)+\varepsilon\|w-x_0\|.
\end{equation}
\end{itemize}
\begin{proposition}[sds implies strictly upper semi-smooth]\label{sds-uss}
Let $\Omega\subset X$ a nonempty open convex subset of a Banach space $X$
and let $f:X\to\xR$ be a lsc function with $\Omega\cap \dom f\ne\emptyset$.
If the function $f:X\to\xR$ is sds on $\Omega$, then
for every $\bar{x}\in \Omega\cap \dom \partial f$, every $u\in X$ with
$\bar{x}+u\in \Omega\cap \dom f$ and every $t\in [0,1{[}$,
the function $f$ is strictly upper semismooth at $\bar{x}+tu$ in the direction $u$
and its radial subderivative $f^r(\bar{x}+tu;u)$ is finite .
\end{proposition}
\begin{proof}
Let $f:X\to\xR$ be sds on $\Omega$.
Let $\bar{x}\in\Omega\cap\dom \partial f$, $u\in X$ with $u\ne 0$,
such that $\bar{x}+u\in \Omega\cap\dom f$.
We apply the above definition of a sds function with $u$ and $v$
respectively replaced by $\bar{x}$ and $\bar{x}+u$.
Let $\varepsilon>0$ and let $\bar{x}_t:=\bar{x}+tu\in [\bar{x},\bar{x}+u{[}$.
By condition (iii) of the definition, there exists
$r_0 \in {]}0, 1{[}$ such that for any $w=\bar{x}_t+ru$ with
$r \in {]}0, r_0 ]$ and for every $(x_n,x^*_n) \in \partial f$
with $x_n \to x_0 \in [\bar{x}_t, w{[}$ one has
$$
\limsup_{n\to\infty}\, \langle x^*_n,w-x_n\rangle \le f^r(\bar{x}_t;w-x_0)+
\varepsilon\|w-x_0\|.
$$
Since subdifferentials are densely defined (Theorem \ref{dense}),
for every $x_0 \in [\bar{x}_t, w{[}$ there does exist a sequence
$(x_n,x^*_n) \in \partial f$ with $x_n \to x_0$. Given $x_0 \in [\bar{x}_t, w{[}$,
write $w$ as $w=x_0+r_1 u$ with $r_1>0$ so that the above relation
can be written as
$$
\limsup_{n\to\infty}\, \langle x^*_n,r_1 u+x_0-x_n\rangle \le f^r(\bar{x}_t;r_1u)+\varepsilon\|r_1u\|.
$$
Dividing by $r_1$ we get that for every $x_0\in [\bar{x}_t, w{[}$
and every $(x_n,x^*_n) \in \partial f$ with $x_n \to x_0$,
\begin{equation*}
\limsup_{n\to\infty}\, \langle x^*_n,u+\frac{x_0-x_n}{r_1 }\rangle \le
f^r(\bar{x}_t;u)+\varepsilon\|u\|,
\end{equation*}
hence, for every $x_0\in [\bar{x}_t, w{[}$,
\begin{equation*}
\inf_{\alpha\ge 0}\limsup_{x\to x_0} f^\partial (x; u +\alpha (x_0-x)\rangle\le f^r(\bar{x}_t;u)+\varepsilon\|u\|.
\end{equation*}
Invoking Theorem \ref{formula}, we conclude that for every $x_0\in [\bar{x}_t, w{[}$, it holds
\begin{equation}\label{fin}
f^r(x_0;u)\le \inf_{\alpha\ge 0}\limsup_{x\to x_0} f^r (x; u +\alpha (x_0-x)\rangle
\le f^r(\bar{x}_t;u)+\varepsilon\|u\|.
\end{equation}
The mean value inequality (Proposition \ref{mvi}) produces a point
$x_0\in [\bar{x}_t, w{[}$ such that $f^r(x_0;u)>-\infty$, hence also
$f^r(\bar{x}_t;u)>-\infty$ in view of \eqref{fin}.
Combining this with condition (ii) of the definition of sds, we see that
$f^r(\bar{x}_t;u)$ is finite.
Next, considering \eqref{fin} with $x_0=\bar{x}_t$ and noting that $\varepsilon>0$
was arbitrary, we see that \eqref{semismooth} holds with $\bar{x}=\bar{x}_t$,
that is, $f$ is strictly upper semismooth at $\bar{x}_t$ in the direction $u$.
\end{proof}
The next theorem states a key property of (strictly)
upper semismooth functions: roughly, a lsc function is (strictly)
upper semismooth at some point if and only if
its radial subderivative at this point can be recovered
from the values of the subdifferential
of the function at directional limiting points (at limiting points).
\begin{theorem}[Recovering the radial subderivative from a
subdifferential]\label{semismoothsubdiff}
Let $X$ be a Banach space,
$f:X\to{]}{-\infty},+\infty]$ be lsc, $\bar{x}\in\dom f$ and $u\in X$, $u\ne 0$.
Let also $\partial$ be an arbitrary subdifferential.
{\rm (a)}
Assume $f$ is radially accessible at $\bar{x}$ from $u$.
Then $f$ is upper semismooth at $\bar{x}$ in the direction $u$ if and only if
\begin{align}\label{semismoothsubdiff1}
f^r(\bar{x};u)=\inf_{\alpha\ge 0}\limsup_{x\to_u \bar{x}}
f^\partial (x:u+\alpha (\bar{x}-x)).
\end{align}
{\rm (b)} $f$ is strictly
upper semismooth at $\bar{x}$ in the direction $u$ if and only if
\begin{align}\label{semismoothsubdiff2}
f^r(\bar{x};u)=\inf_{\alpha\ge 0}\limsup_{x\to \bar{x}}
f^\partial (x:u+\alpha (\bar{x}-x)).
\end{align}
\end{theorem}
\begin{proof}
(a) By Theorem \ref{formula} with $v=u$,
the formulas (\ref{semismoothsubdiff1}) and (\ref{dsemismooth}bis) are the same.
\smallbreak
(b) We observed that equality holds in (\ref{semismooth});
this equality and \eqref{semismoothsubdiff2} are the same according to
Theorem \ref{formula} with $v=0$.
\end{proof}
\section{Subdifferential determination property}\label{determinationsect}
The two theorems of this section assert, with slightly different
assumptions, that the upper semismooth functions have the
subdifferential determination property.
\begin{theorem} [Subdifferential determination property]
\label{determination}
Let $X$ be a Banach space and $\Omega\subset X$ be a nonempty open convex subset.
Let $f,g:X\to{]}{-\infty},+\infty]$ be lsc with $\Omega\cap \dom f\ne\emptyset$.
Assume that for every $\bar{x}\in \Omega\cap \dom \partial f\cap \dom \partial g$ and
every $u\in X$, $u\ne 0$, with $\bar{x}+u\in \Omega\cap \dom f\cap \dom g$,
the points $\bar{x}_t:=\bar{x}+tu$ satisfy the following properties:
\smallbreak
{\rm(\ref{determination}.1)}
$t\mapsto f(\bar{x}_t)$ and $t\mapsto g(\bar{x}_t)$ are continuous on $[0,1]$;
\smallbreak
{\rm(\ref{determination}.2)} there is a countable subset $C\subset [0,1]$ such
that for every $t\in [0,1]\setminus C$,
either $f^r_+(\bar{x}_t;u)$ or $g^r(\bar{x}_t;u)$ is finite, and
$g$ is upper semismooth at $\bar{x}_t$ in the direction $u$.
\smallbreak\noindent
Then,
\begin{equation}\label{impli}
\partial f(x)\subset \partial g(x) \mbox{ for all } x\in\Omega\cap \dom f
\Longrightarrow f=g + {\rm const} \text{ on }\Omega\cap \dom f.
\end{equation}
\end{theorem}
\begin{proof}
Assume
\begin{equation}\label{final0}
\partial f(x)\subset \partial g(x) \quad \mbox{for all } x\in\Omega\cap \dom \partial f.
\end{equation}
The beginning of the proof is the same as in
Thibault-Zagrodny \cite[Theorem 3.21]{TZ10}.
We may suppose that $\Omega\cap \dom f$ is not a singleton because otherwise the result is obvious.
Then, if $\Omega\cap \dom f$ contains two distinct points $x,y$,
the set $\Omega\cap \dom \partial f$ also contains two distinct points $\bar{x},\bar{y}$ by the
density of $\dom\partial f$ into $\dom f$ (Theorem \ref{dense}).
From \eqref{final0} it follows that $\bar{x}$ and $\bar{y}$ also belong to
$\dom \partial g\subset \dom g$.
\smallbreak
\textit{First step.}
Fix $\bar{x}\in\Omega\cap \dom \partial f=\Omega\cap \dom \partial f\cap \dom \partial g$
and $\bar{y}\in \Omega\cap \dom f\cap\dom g$, with $\bar{x}\ne\bar{y}$.
We claim that
\begin{equation}\label{fin1}
f(\bar{y})-f(\bar{x})\le g(\bar{y})-g(\bar{x}).
\end{equation}
Let $u:=\bar{y}-\bar{x}$, hence $\bar{x}_t=\bar{x}+tu=\bar{x}+t(\bar{y}-\bar{x})$,
and let $t\in [0,1]\setminus C$.
By Assumption (\ref{determination}.1),
$f$ is radially accessible at $\bar{x}_t$ from $u$, so in view of
Proposition \ref{devdir} and Theorem \ref{formula}
\begin{equation}\label{fin2}
f^r_+(\bar{x}_t;u)\le \inf_{\alpha\ge 0}\limsup_{x\to_u \bar{x}_t}\,
f^\partial(x;u+\alpha(\bar{x}_t-x)).
\end{equation}
By Assumption (\ref{determination}.2),
$g$ is upper semismooth at $\bar{x}_t$ in the direction $u$,
hence according to Definition \ref{defdsemismooth} and Theorem \ref{formula}
\begin{equation}\label{fin3}
\inf_{\alpha\ge 0}\limsup_{x\to_u \bar{x}_t}\,g^\partial (x;u+\alpha(\bar{x}_t-x))\le g^r(\bar{x}_t;u).
\end{equation}
From \eqref{final0}, the right-hand side of \eqref{fin2} is less than
or equal to the left-hand side of \eqref{fin3}. We therefore conclude that
\begin{equation}\label{fin4}
f^r_+(\bar{x}_t;u)\le g^r(\bar{x}_t;u) \text{ for all } t\in [0,1]\setminus C.
\end{equation}
Now, consider the functions $\varphi:t\in [0,1]\mapsto f(\bar{x}_t)$ and
$\gamma:t\in [0,1]\mapsto g(\bar{x}_t)$. By Assumption (\ref{determination}.1),
they are finite and continuous on $[0,1]$
and, by Assumption (\ref{determination}.2), for every $t\in [0,1]\setminus C$,
either $\varphi^r_+(t;+1)$ or $\gamma^r(t;+1)$ is finite.
On the other hand, \eqref{fin4} can be reformulated as
\begin{equation}\label{fin4b}
\varphi^r_+(t;+1)\le \gamma^r(t;+1) \text{ for all } t\in [0,1]\setminus C.
\end{equation}
So we may invoke
Proposition \ref{recov-subdiv-basic-continuous} to derive
that $\varphi(1)-\varphi(0)\le \gamma(1)-\gamma(0)$, that is,
\eqref{fin1} holds.
\smallbreak
\textit{Second step.}
In the first step, we have shown that \eqref{fin1}
holds for every point $\bar{x}\in\Omega\cap \dom \partial f$
and $\bar{y}\in \Omega\cap \dom f\cap\dom g$.
Now, let $\bar{x}\in\Omega\cap\dom f$ and $\bar{y}\in \Omega\cap \dom f\cap\dom g$.
Applying the density Theorem \ref{dense}, we find
a sequence $(\bar{x}_n)_n$ in $\dom \partial f$ such that $\bar{x}_n\to \bar{x}$ and $f(\bar{x}_n)\to f(\bar{x})$.
By \eqref{fin1}, for every $n\in {\mathbb N}$ and $\bar{y}\in \Omega\cap \dom f\cap\dom g$ it holds
$$f(\bar{y})-f(\bar{x}_n)\le g(\bar{y})-g(\bar{x}_n).$$
Because $f(\bar{x}_n)\to f(\bar{x})$ and $g$ is lower semicontinuous at $\bar{x}$,
passing to the limit we get
$$f(\bar{y})-f(\bar{x})\le g(\bar{y})-g(\bar{x}).$$
This inequality shows that $g(\bar{x})$ is finite whenever $\bar{x}\in\Omega\cap\dom f$,
that is, $\Omega\cap\dom f= \Omega\cap\dom f\cap\dom g$.
So finally one has
$$
f(\bar{y})-f(\bar{x})\le g(\bar{y})-g(\bar{x}) \text{ for all } \bar{x},\bar{y} \in \Omega\cap\dom f.
$$
Then, interchanging the role of $\bar{x}$ and $\bar{y}$, we derive that in fact
$$
f(\bar{y})-f(\bar{x})= g(\bar{y})-g(\bar{x}) \text{ for all } \bar{x},\bar{y} \in \Omega\cap\dom f,
$$
which means that $f=g + {\rm const}$ on $\Omega\cap \dom f$.
The proof is complete.
\end{proof}
If instead of the continuous version of the mean value theorem
(Proposition \ref{recov-subdiv-basic-continuous})
we use the semicontinuous version
(Proposition \ref{recov-subdiv-basic}),
we obtain a variant of the above theorem with
weaker assumptions on $f$ but stronger on $g$. The proof being
similar will not be repeated.
\begin{theorem} [Subdifferential determination property: semicontinuous variant]
\label{determination2}
Let $X$ be a Banach space and $\Omega\subset X$ be a nonempty open convex subset.
Let $f,g:X\to{]}{-\infty},+\infty]$ be lsc with $\Omega\cap \dom f\ne\emptyset$.
Assume that for every $\bar{x}\in \Omega\cap \dom \partial f\cap \dom \partial g$ and
every $u\in X$, $u\ne 0$, with $\bar{x}+u\in \Omega\cap \dom f\cap \dom g$,
the points $\bar{x}_t:=\bar{x}+tu$ satisfy the following properties:
\smallbreak
{\rm(\ref{determination2}.1)}
$t\mapsto g(\bar{x}_t)$ is continuous on $[0,1]$;
\smallbreak
{\rm(\ref{determination2}.2)}
for every $t\in [0,1{[}$, either $f^r(\bar{x}_t;u)$ or $g^r(\bar{x}_t;u)$ is finite, and
either {\rm(a)} or {\rm(b)} holds:
\\\hspace*{0.7cm}
{\rm(a)} $f$ is radially accessible at $\bar{x}_t$ from $u$
and $g$ is upper semismooth at $\bar{x}_t$ in direction $u$,
\\\hspace*{0.7cm}
{\rm(b)} $g$ is strictly upper semismooth at $\bar{x}_t$ in direction $u$.
\smallbreak\noindent
Then,
\begin{equation*
\partial f(x)\subset \partial g(x) \mbox{ for all } x\in\Omega\cap \dom f
\Longrightarrow f=g + {\rm const} \text{ on }\Omega\cap \dom f.
\end{equation*}
\end{theorem}
\begin{remark}
(a) Theorem \ref{determination} is new in the context of mean-valued based theorems.
Its assumption on $g$ is much weaker than the one in Theorem \ref{determination2}.
It should rather be compared and contrasted with integration-based results
such as those in \cite{BM98a,TZ10}. Since the technique and concepts are
totally different, this will be done in a separate paper.
\smallbreak
(b) Theorem \ref{determination2} unifies several results.
As an illustration, we mention three of them, which cannot
be derived from each other but which are all special cases
of Theorem \ref{determination2}, since the functions they involve
are either upper semismooth (case b1) or strictly upper semismooth
(cases b2 and b3).
\\\hspace*{0.7cm}
(b1) Correa-Jofr\'e \cite[Proposition 5.4]{CJ89}, where
$g$ is locally Lipschitz, semismooth and whose Clarke subdifferential
is single-valued at any point of a dense subset of $X$;
\\\hspace*{0.7cm}
(b2) Thibault-Zlateva \cite[Theorem 3.3]{TZl05}, where
lsc regular functions $g$ which are
continuous relative to their domains and strictly directionally Lipschitz
are shown to have a ''local subdifferential determination property'';
\\\hspace*{0.7cm}
(b3) Thibault-Zagrodny \cite[Theorem 4.1]{TZ05}
(in the special case $\gamma=0$), where $g$ is sds.
\end{remark}
{\small
|
\section{Introduction}
Turbulent thermal convection plays very important role in a range of phenomena in astrophysical and geophysical systems, where the underlying surfaces over which convective turbulence occurs are always rough. As an idealised model for the study of turbulent thermal convection in general, turbulent Rayleigh-B\'enard convection (RBC), where a fluid layer confined between two horizontally parallel plates heated from below and cooled from above, has been studied extensively in the past two decades \citep*[for reviews, see e.g.][]{Siggia1994ARFM,Ahlers2009RMP,Xia2010ARFM,SchumacherChilla2012,Xia2013TAML}. In terms of equations of motion, there are two control parameters in RBC, i.e. the Rayleigh number $Ra=\alpha g\Delta T H^3/(\nu\kappa)$ and the Prandtl number $Pr=\nu/\kappa$, where $ \Delta T$ is the applied temperature difference across a fluid layer of height $H$, $g$ is the gravitational acceleration constant and $\alpha$, $\kappa$ and $\nu$ are respectively the thermal expansion coefficient, the thermal diffusivity and the kinematic viscosity of the working fluid. Any laboratory convection experiments are confined, thus the aspect ratio $\Gamma$ and the container shape come into the problem. One of the central issues in the study of turbulent RBC is to determine how the response parameters, i.e. the Nusselt number $Nu$ and the Reynolds number $Re$, depend on the control parameters, where $Nu$ is a measurement of the system's heat transport efficiency and $Re$ quantifies the turbulence intensity. The functional forms of $Nu(Ra, Pr)$ and $Re(Ra, Pr)$ are usually expressed in terms of power laws:
\begin{equation}
Nu\sim Ra^{\alpha}Pr^{\zeta} \qquad Re \sim Ra^{\beta}Pr^{\epsilon}
\end{equation}
A comprehensive review on this issue can be found in \citet{Ahlers2009RMP}. It is now generally accepted that there are multiple scalings in the $Ra-Pr$ phase space where the bulk flow is in a turbulent state according to the Grossmann and Lohse model for turbulent thermal convection \citep{GL2000JFM,GL2001PRL,GL2002PRE,GL2004PoF}.
Besides its fundamental importance in geophysical and astrophysical problems, turbulent thermal convection over rough surfaces is found to be a more effective way to transport heat, which is also useful in practical applications. Heat transport enhancement is observed when the top and bottom plates are rough and the thermal boundary layer (BL) is thinner than the roughness height \citep*{Shen1997PRL}. The enhancement is later found to be due to more thermal plumes being emitted from the tip of the roughness elements \citep*{Du1998PRL}.
Although the heat transport is enhanced, the scaling relation between $Nu$ and $Ra$ is found to be insensitive to the plate morphology \citep{Shen1997PRL,Du1998PRL}. Several later experiments show that the scaling exponents also change in convection cells with rough plates (rough cells hereafter) when compared with observations in convection cells with smooth plates (smooth cells hereafter). Using glass spheres adhered to the bottom plate of the convection cell, the scaling exponent is found to depend on the distribution of the roughness height \citep{Ciliberto1999PRL}. For $Ra>10^{12}$, a scaling exponent of $0.5$ is reported when grooved roughness is distributed on both the plates and sidewall of the convection cell \citep{Roche2001PRE}. A change of the scaling exponent from 0.28 in a smooth cell to 0.35 in a rough cell is observed \citep*{Qiu2005JoT}. The rough cell used in the experiments of \citep{Qiu2005JoT} had similar design with the ones used in \citet{Shen1997PRL} and \citet{Du1998PRL}.These authors have suggested that the discrepancy between their experiments and those of \citet{Shen1997PRL} and \citet{Du1998PRL} may be attributed to the different constructions and materials of the rough plates. In the earlier experiments, rough plates made of brass were attached on the surfaces the top and bottom smooth plates. While in the later experiment, the roughness elements were directly machined on a piece of copper. In addition to studying the global transport properties of the convection cell, the heat transport properties of individual plates with rough/smooth surfaces were also investigated \citep{Wei2014JFM}. It is found that heat transport property of individual plate differs significantly and depends on the nature of the opposite plate of the same convection cell and the pertaining temperature boundary conditions. By determining the $Nu$ separately for the bottom rough plate and the top smooth plate, a scaling exponent of 1/2 for the rough plate and of 1/3 for the smooth plate are reported \citep{Tisserand2011PoF}.
The effects of roughness elements on the local dynamics of turbulent RBC are also studied. Using flow visualization and near-wall temperature measurements, it is shown that the interaction of the large-scale circulation (LSC) and the secondary flow in the grooved region of the rough plate enhances plume emission which leads to the enhancement of heat transport \citep{DU2000JFM}. The study of local temperature fluctuations suggest that the enhancement of heat transport is determined by the dynamics in the near plate region \citep{Du2001PRE}. The effects of rough plates on the velocity BL have been investigated very recently \citep{Liot2016JFM}, where it is found that the flow field is very different before and after the heat transport enhancement transition.
Direct numerical simulations (DNSs) of turbulent RBC in rough cells also show that the scaling exponent increases \citep*{Verzicco2006JFM,Olga2011JFM,Wanger2015JFM,Wettlaufer2015EPL,Wettlaufer2017PRL}. The scaling exponent $\alpha$ increases to 0.37 in a rough cell with grooved walls at the top and bottom plates when compared with the smooth cell \citep{Verzicco2006JFM}. The scaling exponent was also found to depend on the geometry of the roughness elements or ``obstacles" in a quasi-2-D DNS \citep{Wanger2015JFM}. But the number of roughness elements in this study is very limited (only 4 ``obstacles" are used). In a 2-D study, using sinusoidal rough elements, the scaling exponent $\alpha$ is also found to depend on the wavelength of the roughness elements \citep{Wettlaufer2015EPL,Wettlaufer2017PRL}. In a DNS study of turbulent Taylor-Couette flow the scaling exponent is also found to increase with wall-roughness \citep{Zhu2016JFM}. These numerical studies imply the importance of roughness geometry on turbulent transport. However, the limited number of roughness element in a 3-D or a (quasi-) 2-D computation domain makes the comparison between the DNS studies and experimental ones less straightforward.
Although there are quite a number of previous experiments studying the effect of roughness on turbulent RBC, to the best of our knowledge no experimental study has been made to systematically varying roughness geometry. This motivates the present study in which the effects of roughness geometry on turbulent RBC have been investigated by systematically varying the roughness geometry. A parameter $\lambda$ which is defined as the height of a single roughness element over its base width is used to characterise the geometry of roughness elements. In addition to heat transport enhancement as found by previous studies, we find that the heat transport scaling in a rough cell with fixed roughness height $h$ may be classified into three regimes. The experiments further show that $\lambda$ can be used as a tuning parameter to manipulate heat transport scaling in rough cells in the heat transport enhanced regime. The Reynolds number and local temperature fluctuation scalings with $Ra$ also change with $\lambda$, suggesting that the geometry of roughness elements not only alters the global heat transport of the system but also influences the bulk turbulence and the dynamics of the LSC.
The remainder of the paper is organised as follows: The experimental setup and measurement techniques are introduced in \S \ref{Sec_Setup} which is divided into four subsections. The construction of the convection cell is described in \S \ref{SubSec_cell}. The properties of the working fluids used and the explored $Ra-Pr$ phase space are introduced in \S\ref{SubSec_fluid}. The temperature measurement techniques are documented in \S \ref{SubSec_temp}, and \S \ref{SubSec_PIV} introduces the measurement of the viscous BL using particle image velocimetry (PIV). \S \ref{Sec_Results} presents the experimental findings. The heat transport measurements for different values of $\lambda$ are reported in \S \ref{SubSec:Nu}. The transitions between different heat transport regimes are discussed in \S \ref{SubSec_Transition}. The Reynolds number and local temperature fluctuation measurements are presented in \S \ref{SubSec_Re} and \S \ref{SubSec_Temp_fluc}, respectively. A discussion of the proper parameter that characterises the system is reported in \S \ref{SubSec_con_para} before we summarise and conclude our findings in \S \ref{Sec_con}.
\section{Experimental setup and measurement techniques\label{Sec_Setup}}
\subsection{The convection cell\label{SubSec_cell}}
\begin{figure}
\includegraphics[width=\textwidth]{figure1-eps-converted-to.pdf}
\caption{\label{fig_pic}(Colour online) Photos of the bottom plate of a rough cell with $\lambda=1.0$. Left: top view; right: side view. The red heater connection wires can be seen on top of the left panel. Plates with $\lambda$ values of 0.5, 1.9 and 4.0 have the same design.}
\end{figure}
The experiments were carried out in four cylindrical convection cells. Each of the convection cells consisted of top and bottom plates and a Plexiglas sidewall. Except for the geometry of roughness elements on the top and bottom plates, the constructions of the convection cells were otherwise identical.
The sidewall was made of Plexiglas tube with a wall-thickness of 4 mm. The convection cells had a diameter $D$ of 19.2 cm and a height $H$ of 20.3 cm measured from the base of the roughness elements. Thus the aspect ratios $\Gamma=D/H$ of the convection cells were approximately unity. The top and bottom plates were made of 12 mm-thick copper except for the ones for rough cell with $\lambda=1.9$ which were made of aluminium. The thermal conductivity of aluminium is 205 W/(m$^o$C) at temperature 25 $^o$C, which is roughly half that of copper. The differences in heat transport properties in convection cells with plates made of copper and aluminium, i.e. the effects of finite conductivity of plates \citep{Verzicco2004PoF,Brown2005PoF}, have been shown to be minimal in the $Ra$ and $Pr$ range reported in the present studies \citep*{Wei2012PRE}. The pyramid-shaped roughness elements were directly machined on the top and bottom plates that were then plated with nickel to prevent oxidation. The roughness elements were arranged in a square lattice form (see figure \ref{fig_pic}), which was the same as those used previously \citep{Wei2014JFM, Xie2015JFM}. By keeping the height of roughness elements $h$ a constant at 8 mm and changing the base width $w$ of individual roughness elements from 16 mm to 2 mm, rough plates with roughness parameter $\lambda$, defined as the ratio of $h$ and $w$, at 0.5, 1.0, 1.9 and 4.0 were used in the experiments. With the increase of $\lambda$, the number density $n$ (in unit of number of roughness element per cm$^2$) also increases, i.e. $n$ = 0.38 ($\lambda = 0.5$), 1.5 ($\lambda = 1.0$), 6.1 ($\lambda = 1.9$), and 24.5 ($\lambda = 4.0$). To clarify whether the number density or the $\lambda$ is the key parameter that characterizes the system, another sets of rough plates with $\lambda$ = 4.0 and a roughness height of 4 mm were used. The present study implies that $\lambda$ is a more suitable parameter to characterise the geometry of the roughness elements and the response of the system, rather than the number density of the roughness elements.
The top plate was cooled by passing temperature-controlled water from a water bath (Poly Science, 9702) through a chamber fitted on its top. The cooling water entered the chamber through two inlets and left the chamber through two outlets. The arrangements of the inlets and outlets were such that the two inlets were along one diameter of the plate and the two outlets were along another diameter perpendicular to the inlets. Temperature stability of the cooling water was better than 0.01 $^o$C. A stirring blade driven by the inlet and outlet flows was installed inside the cooling chamber to ensure temperature uniformity across the top plate. The bottom plate was electrically heated by two rubber heaters connected in series and sandwiched in between the bottom plate and a copper plate with a thickness of 10 mm (see, e.g. figure \ref{fig_pic}). The heaters were connected to a DC power supply with a long term voltage stability of 99.99\% (Sorensen, XFR300-4). To ensure good thermal contact, a thin layer of thermally conductive paste was used between the heaters and the plates. The largest temperature difference within the same plate was about $1.5 \% \Delta T$. It should be noted that under this arrangement, the temperature boundary condition for the top plate was constant temperature and that for the bottom plate was constant heat flux.
To prevent heat leakage and to minimise the influence of the environmental temperature fluctuation, the convection cells were wrapped with Styrofoam with a thickness of 6 cm. A temperature regulated copper basin was placed beneath the convection cell to prevent heat leakage from the bottom plate of the convection cell by maintaining the same temperature at the top and bottom of the basin. The whole set up was put into a thermostat where the temperature was regulated to match the mean temperature $T_m$ of the top and bottom plate. The temperature stability of the thermostat was better than 0.1 $^o$C.
\subsection{The working fluids \label{SubSec_fluid}}
Only about one and a half decades of $Ra$ can be achieved in a single experimental setup using classical working fluid like water. To extend the $Ra$ range, two kinds of working fluids, namely deionized water and Flourinert FC770 (FC770 hereafter, 3M Inc.) were used. Water has been widely used in the study of turbulent convection, we will mention here only the properties of FC770. The data listed below are given by the manufacturer (3M Inc.) and are all obtained at the temperature $25^{o}$C. The FC770 has $\kappa$ = 0.063 W/(m$^o$C), $\alpha$ = 0.00148/$^o$C, $\nu = 7.9\times 10^{-7} $ m$^2$/s, the specific heat $c_p = $ 1038 J/(kg$^o$C), and the density $\rho$ = 1793 kg/m$^3$.
The working fluid was filled into the convection cell through a hole attached with a stainless steel tube ($\sim$ 8 mm in diameter) located at the centre of the top plate. The working fluids were degassed in the following way to ensure there was no bubble forming during a single run of the experiment, which typically lasts for three weeks. Deionized water was boiled for a hour and then cooled down to room temperature before filling into the convection cell. For FC770, it was first filled into the convection cell and degassed by keeping the temperature of the top and bottom plate, and thus the whole convection cell, at $50 ^o$C for 24 hours. Then the temperature of the convection cell was lowed to $25 ^o$C which was the $T_m$ for all the experiments using FC770, and the bubbles formed were totally removed.
The explored $Ra$ and $Pr$ phase space at different values of $\lambda$ is shown in figure \ref{fig_PhaseSpace}. Except for $\lambda=0.5$ which was investigated for a single $Pr$ (23.34) in the present studies, all the other values of $\lambda$ were studied for the four values of $Pr$, i.e. $Pr$ = 3.67 (water, $T_m = 50.00$ $^o$C), $Pr$ = 4.34 (water, $T_m = 40.00$ $^o$C), $Pr$ = 6.14 (water, $T_m = 25.00$ $^o$C), and $Pr$ = 23.34 (FC770, $T_m = 25.00 $ $^o$C). The experiments covered $7.50\times 10^7\leq Ra \leq 1.31\times 10^{11}$. The temperature difference applied across the convection cell using water as the working fluid for different values of $\lambda$ were: 0.98 $^o$C to 26.99 $^o$C ($\lambda = 4.0$); 0.47 $^o$C to 20.16 $^o$C ($\lambda = 1.9$) and 0.89 $^o$C to 24.75 $^o$C ($\lambda = 1.0$). The temperature difference applied using FC770 as working fluid at different values of $\lambda$ were: 1.40 $^o$C to 22.14 $^o$C ($\lambda = 4.0$); 0.99 $^o$C to 17.25 $^o$C ($\lambda = 1.9$); 1.11 $^o$C to 27.63 $^o$C ($\lambda = 1.0$) and 0.96 $^o$C to 29.66 $^o$C ($\lambda = 0.5$).
\begin{figure}
\begin{center}
\includegraphics[width=0.65\textwidth]{figure2-eps-converted-to.pdf}
\caption{\label{fig_PhaseSpace}(Colour online) $Ra-Pr$ phase space explored for different values of $\lambda$. The case for $\lambda$ = 0.5 is only studied for $Pr=$23.34. All experiments were conducted in convection cells with roughness height $h$ = 8 mm, except for the ones corresponding to hexagons that were conducted in convection cell with $h$ = 4 mm.}
\end{center}
\end{figure}
\subsection{The temperature measurements \label{SubSec_temp}}
Temperature of the top plate was measured using four thermistors (Omega, Model 44031) embedded at a distance of $D/4$ from the centre of the plate and uniformly in the azimuthal direction. Four thermistors at the mirror location of the top plate embedded in the bottom plate with one additional thermistor at the centre were used to measure its temperature. The thermistors were calibrated individually and separately in a temperature controlled water bath. The temperature range of calibration was from 5 $^o$C to 70 $^o$C with a temperature accuracy better than 0.01 $^o$C. The resistances of the thermistors were measured using a $6\frac{1}{2}$ digital multimeter (Keithley, 2700) and were converted to temperature using the calibration curve. The sampling rate of the temperature measurement was 0.7 Hz.
Local temperatures at the cell centre and $\sim$ 1 cm away from the sidewall were measured using two small thermistors with a bead diameter of $0.38$ mm and a time constant of 30 ms in liquid (Measurement Specialities, Model: G22K7MCD419). These small thermistors had a temperature accuracy of 0.01 $^o$C. They were guided into the convection cell using a stainless steel tube with an outer diameter of 1.1 mm and wall thickness of 0.2 mm (Goodfellow, Stainless Steel - AISI 304 Tube). Each of the small thermistors had an independent temperature measurement system specially designed to measure the temperature fluctuation. Detailed design of the measurement system can be found in \citet{Zhou2001PRL}. Briefly, the thermistor acted as one arm of a AC Wheatstone electrical bridge. The resistance fluctuations of the thermistor induced by the temperature fluctuations in the convection cell were converted to the voltage fluctuations of the electrical bridge which were magnified using a lock-in amplifier (Stanford Research System, SR830). The output signal from the lock-in amplifier was then digitised and stored by a dynamic signal analyser (HP, 35670A). The measured voltage was converted to resistance and then to temperature using the calibration curve of the thermistors. The sampling rate of the local temperature varied from 16 Hz to 128 Hz depending on $Ra$ and $Pr$ to ensure that the small scale temporal temperature fluctuations were fully resolved.
From the temperature measurements, the heat transport across the convection cell, i.e. the Nusselt number $Nu=qH/(\chi\Delta T \pi(D/2)^2)$, local temperatures, and the Reynolds number $Re=\frac{UH}{\nu}$ were measured as a function of $Ra$, $Pr$ and $\lambda$, where $q$ was the input heating power at the bottom plate, $\chi$ was the thermal conductivity of the working fluid and $U$ was the characteristic velocity of the LSC. Each measurement lasted at least 12 hours after the system had reached a stable state to ensure that the system had explored all flow states.
\subsection{The viscous boundary layer measurements \label{SubSec_PIV}}
The dynamics of viscous boundary layer in the rough cell with $\lambda=1.0$ was measured using a commercial particle image velocimetry (PIV) system that had been documented elsewhere \citep*{Xia2003PRE}. The flow was seeded with particles with a diameter of 2 $\mu$m and density matched with water. A 2D flow region with a size of 20 mm$\times$ 28 mm was measured within the LSC's circulation plane with $x$ denoting the direction along the circulation path of the LSC and $z$ denoting the direction pointing upwards at the centre of bottom plate (see, e.g. figure \ref{fig_PIV}). The origin of the coordinate system was located at the centre of the plate. To minimise optical distortion of the particle images induced by the curved surface of the cell, a jacket filled with water with flat window was adhered to the convection cell. The optical arrangement was such that the measurement plane was within one groove of the rough plate and the images were acquired perpendicular to it. A mask with the shape exactly matching that of the plate geometry when viewing from the front was applied to the measured images to exclude the parts blocked by the roughness elements in the captured images.
The velocity field was obtained by cross-correlating two consecutive images taken with a time separation that matches the velocity of the flow at different $Ra$. The velocity map consisted of 63$\times$ 79 vectors in the $x-$ and $z-$ directions respectively. Totally 26136 vector maps were acquired at a sampling rate of 2.2 Hz, which lasts for $\sim 3.3$ hours. The velocity fluctuation map was obtained by subtracting the mean velocity map from each of the instantaneous velocity map. The measurements were done at $Pr=4.34$ and $3.55\times10^8\leq Ra \leq 9.62\times10^9$. To lock the direction of the LSC, the convection cell was tilted with respect to the vertical direction by $\sim 0.5^o$.
\section{Results and discussion \label{Sec_Results}}
\subsection{Heat transport measurements}\label{SubSec:Nu}
We show in this section that the heat transport in turbulent RBC over rough plates may be classified into three regimes based on the scaling law of $Nu$ vs. $Ra$. In addition, the heat transport scaling in the enhanced regime (Regimes II and III) can be manipulated by changing the roughness geometry. The transition from Regime I to Regime II in a rough cell with $h$ = 4 mm and $\lambda=4.0$ is discussed in section \S \ref{Regime_I_II}. The heat transport measurement with changing the geometry of roughness elements, i.e. for different values of $\lambda$, is presented in \S \ref{Regime_II_III} together with the observation of Regime II and Regime III.
\subsubsection{Transition from Regime I to Regime II \label{Regime_I_II}}
\begin{figure}
\includegraphics[width=\textwidth]{figure_new-eps-converted-to.pdf}
\caption{\label{fig_Regime_I_II}(Colour online) ($a$) $Nu$ as a function of $Ra$ in a rough cell with a roughness elements height of 4 mm. The down-pointing triangles are taken from \citet{Denis2005JFM}($\Gamma=1.0, Pr=4.38$), the hexagons are taken from \citet{Sun2005JFM}($\Gamma=2.0, Pr=4.34$), the stars are taken from \citet{Wei2014JFM}. ($b$) The compensated plot of the same data in ($a$) .}
\end{figure}
Previously, \citet{Shen1997PRL} find that the heat transport cannot be altered in a rough cell if thermal BL thickness $\delta_{th}$ is larger than the height of the roughness elements. We call this Regime I of convective turbulence in a rough cell. As $\delta_{th}$ decreases with increasing $Ra$, it will eventually become smaller than the roughness height, i.e. $\delta_{th} < h$. When this happens turbulent convection enters a new, heat transport enhanced regime, termed Regime II here. This regime has been observed by a number of previous experiments \citep{Shen1997PRL,Du1998PRL,Qiu2005JoT,Tisserand2011PoF,Wei2014JFM}.
We plot in figure \ref{fig_Regime_I_II}($a$) $Nu$ as a function of $Ra$ in a rough cells with $h=4$ mm and $\lambda=4.0$. For comparison, the data from \citet{Denis2005JFM} in a smooth cell with $\Gamma=1, Pr=4.38$, from \citet{Wei2014JFM} in a smooth cell with $\Gamma=1, Pr=4.34$ and from \citet{Sun2005JFM} in a smooth cell with $\Gamma=2, Pr=4.34$ are plotted respectively as down-pointing-triangles, hexagons and stars. Figure \ref{fig_Regime_I_II}($b$) plots the corresponding data in a compensated form. It is seen that $Nu$ in the rough cells first overlaps nicely with those from the smooth cells for $Ra < 4\times 10^8$. With increasing $Ra$, there is a gradual transition, where the heat transport enhances. Within the transitional $Ra$ range $ 4\times 10^8 < Ra < 1\times 10^{10}$, there is no well defined scaling law between $Nu$ and $Ra$. For $Ra>1\times 10^{10}$, the data shows a power-law relation, i.e. $Nu\sim Ra^{0.60}$. These data clearly demonstrate the transition from Regime I to Regime II. Such a transition is also observed by \cite{Wei2014JFM} in a rough cell with $h=3$ mm. Using the relation $\delta_{th}=H/(2Nu)$ (which was valid only for the plate-wise averaged thermal BL thickness and for a smooth cell), we had $\delta_{th}=2.1$ mm at $Ra=3.4\times 10^8$, which might be considered to be close to the roughness height of 4 mm in an order of magnitude sense and broadly consistent with the proposed interpretation by \citet{Shen1997PRL}.
\subsubsection{Effects of roughness geometry on the heat transport enhanced regime\label{Regime_II_III}}
We present measurements of $Nu$ as a function of $Ra$ in convection cells with a roughness height $h$ of 8 mm and different roughness geometries in this subsection. The roughness height was much larger than $\delta_{th}$ for the explored range of $Ra$. Therefore, Regime I behaviour was not observed here. Instead, we focus on Regime II and Regime III (a new regime to be defined below).
\begin{figure}
\includegraphics[width=\textwidth]{figure3-eps-converted-to.pdf}
\caption{\label{fig1}(Colour online) Left panel: $Nu$ as a function of $Ra$ in rough cells for ($a$) $\lambda=4.0$; ($d$) $\lambda=1.9$; ($g$) $\lambda=1.0$ and ($j$) $\lambda=0.5$. The down-pointing triangles are from \citet{Denis2005JFM}($\Gamma=1.0, Pr=4.38$), the hexagons are from \citet{Sun2005JFM}($\Gamma=2.0, Pr=4.34$) and the stars are from \cite{Wei2014JFM}($\Gamma=1.0, Pr=4.38$). Data at $Pr=4.34$ in ($j)$ are from \citet{Wei2014JFM}; $Nu$ at $Pr=23.34$ at $\lambda = 4.0$ and 1.9 are larger than those at $Pr = 3.57$ at the same $Ra$. For a better comparison, they are shifted downwards respectively by 1.5 and 1.34 and displayed using squares with cross. The solid lines are power law fits to the data larger and smaller than $Ra^*\sim 4\times 10^9$. Right panel: The compensated plot of $Nu$ with the fitted $Ra$ scaling for $Ra$ larger and smaller than $Ra\sim 4\times 10^9$ from the left panel.}
\end{figure}
Figures \ref{fig1}($a, d, g, j$) show $Nu$ as a function of $Ra$ and $Pr$ in rough cells with $\lambda=4.0, 1.9, 1.0$ and 0.5, respectively. The solid lines in the figure are power law fit to each data set for $Ra$ larger and smaller than a critical Rayleigh number $Ra^*=4\times 10^9$. The same data from smooth cells as shown in figure \ref{fig_Regime_I_II} are shown in these figures. The general conclusion that one can draw from these figures is that the heat transport is enhanced in rough cells.
We first discuss the case with $\lambda=1.0$ (figure \ref{fig1}($g$)). Within the precision of the experiments, data at different values of $Pr$ but the same $Ra$ collapse onto each other, suggesting that the $Pr$-dependence in the rough cell with $\lambda=1.0$ is very weak in the present parameter range, which is consistent with observation in a smooth cell \citep*{Xia2002PRL}. Detailed examination of the data shows that there is a transition of the $Nu-Ra$ scaling at $Ra^*\sim 4\times 10^{9}$. Below and above $Ra^*$, the data can be fitted by different scaling laws, i.e. $Nu\sim Ra^{0.43\pm 0.01}$ and $Nu\sim Ra^{0.35\pm 0.01}$. The different scaling laws suggest that the heat transport enhancement regime in a rough cell can be further classified into two regimes for the $Pr$ range covered in the present experiments. We term the two regimes before and after $Ra^*$ as Regime II and Regime III. The data are compensated by the scaling exponents in Regimes II and III and shown in figure \ref{fig1}($h,i$). A gradual transition around $Ra^*\sim 4\times 10^9$ is clearly seen.
Data at $\lambda=4.0, 1.9$ and 0.5 are plotted in figure \ref{fig1} ($a, d, j$). One interesting feature revealed in figure \ref{fig1}($a$) and ($d$) is that the $Nu$ at $Pr=23.34$ is larger than that at $Pr=3.57$ when $Ra$ is the same. We note that this $Nu$-enhancement for FC770 ($Pr=23.34$), as compared to water ($Pr=3.57$), only occurs for the large $\lambda$ cases. In fact, for the smallest value of $\lambda$ ($=0.5$) there is a slight decrease in $Nu$. This is quite surprising and unexpected, as in the case of a smooth cells, $Nu$ is peaked around $Pr=3 \sim 4$ \citep{Xia2002PRL}. It is not clear why the heat transport enhancement in rough cells has this peculiar $Pr$-dependence, and furthermore this behaviour has a strong $\lambda$ dependence, i.e. its 150\% for $\lambda=4.0$, 134\% for $\lambda=1.9$, 4\% for $\lambda=1.0$, and -8\% for $\lambda=0.5$. One possible reason leading to such a $Pr$-abnormality is that the lifetime of the thermal plumes at larger $Pr$ is larger compared to the small $Pr$ case. Assuming that the thermal plumes have typical length scale comparable to the thermal BL thickness, the estimated thermal diffusion time scale, i.e. $\delta_{th}^2/\kappa$, for FC770 is 14.5 sec and that for water is 2.6 sec at $Ra\approx 10^{10}$. When normalised by the LSC turn-over time, the dimensionless plume lifetime for FC770 is 0.25 and that for water is 0.11 for $\lambda=4.0$. Another likely reason is the stronger plume clustering effect in rough cells with larger $\lambda$ as shown by the shadowgraph images in figure \ref{fig_shadow}, which also has the effect of increasing plume lifetime or slowing the decay of plumes. While it is difficult to quantify the combined effect of these two factors, it appears that larger values of $\lambda$ together with a larger $Pr$ result in a more efficient enhancement in the heat transport. For small $\lambda$ and/or small $Pr$, thermal plumes lose more heat to the surrounding fluid and therefore $Nu$ won't be enhanced as much.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\textwidth]{figure4-eps-converted-to.pdf}
\caption{\label{fig3}(Colour online) The scaling exponent $\alpha$ from power law fits to $Nu$ vs $Ra$, i.e. $Nu\sim Ra^{\alpha}$, in Regime II (circles) and Regime III (triangles) as a function of $\lambda$. Also plotted in the figure is the scaling exponent $\beta$ from power law fits to $Re$ vs $Ra$ (squares), i.e. $Re\sim Ra^{\beta}$, as a function of $\lambda$. The solids lines are empirical power law fits to the respective data sets: $\alpha_{II}=0.43\lambda^{0.24\pm 0.02}$, $\alpha_{III}=0.37\lambda^{0.23\pm 0.02}$ and $\beta=0.49\lambda^{0.08\pm0.01}$.}
\end{center}
\end{figure}
To determine more accurately the scaling law of $Nu$ vs. $Ra$, the data at $Pr=23.34$ in figure \ref{fig1}($a$) and ($d$) are shifted downwards by a constant, i.e. 1.50 in ($a$) and 1.34 in ($d$), to collapse data at different $Pr$ but the same $Ra$. The transition of the heat transport scaling around $Ra^*$ is observed for all $\lambda$ studied. The power law fits of $Nu=ARa^{\alpha}$ in Regimes II and III for different values of $\lambda$ are summarised in table \ref{tab:table1}. It should be noted that the shifted high $Pr$ data were included in the fitting. Since $Ra$ obtained using water as working fluid at $Pr=3.57$ and that using FC770 as working fluid at $Pr=23.34$ had an overlapping range of less than a half decade, i.e. from $6\times 10^9$ to $9\times 10^9$, a determination of the $Pr$ dependence in Regime III was not possible in the present experiments. It is seen that the heat transport law is distinctively different in Regime II and Regime III. This can be seem more clearly from the compensated plot in the right panel next to figure \ref{fig1}($a,d,j)$.
The experimentally determined scaling exponent $\alpha$, i.e. $Nu\sim Ra^{\alpha}$, in both Regime II and Regime III increases significantly with $\lambda$. In Regime II, the $\alpha$ increases from 0.36 to 0.59 when $\lambda$ is increased from 0.5 to 4.0; and in Regime III, it increases from 0.30 to 0.50. These data are plotted in figure \ref{fig3}. If a power law fit is attempted to these data within the limited range of $\lambda$, we obtain $\alpha_{II}=0.42\lambda^{0.23\pm 0.02}$ and $\alpha_{III}=0.35\lambda^{0.24\pm 0.02}$ which are shown as solid lines in the figure. Within the error bar, the $\lambda$-dependences of $\alpha_{II}$ and $\alpha_{III}$ are the same except that the pre-factor of $\alpha_{II}$ is larger than $\alpha_{III}$. The results suggest that the scaling law of the heat transport in turbulent RBC over rough plates can be manipulated using $\lambda$. In this sense, the present experiments demonstrate clearly the importance of roughness geometry on turbulent heat transport. Any realistic model for turbulent thermal convection in nature should take into account the effects of wall roughness.
Very recently, it is found that there is an upper bound for the heat transport in turbulent convection over rough plates, i.e. $Nu$ cannot grow faster than $Nu\sim Ra^{1/2}$ in the limit of $Ra\rightarrow\infty$ \citep{Goluskin2016JFM}. With increasing $Ra$, Regime II is just a transient, and Regime III is the dominated one. Our experiments showed that at $\lambda=4.0$ the scaling exponent already reached 0.50, which was consistent with the upper bound derived \citep{Goluskin2016JFM}. One interesting question is whether the scaling exponent will saturate to 0.50 with further increasing $\lambda$ which is very challenging to achieve experimentally due to technical difficulties in machining the plates. DNS studies may provide helpful insights into this aspect of the problem.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\textwidth]{figure5-eps-converted-to.pdf}
\caption{\label{fig_enhancement}(Colour online) The heat transport enhancements in Regime II (open symbols) and Regime III (solid symbols) as a function of $\lambda$.}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\def~{\hphantom{0}}
\begin{tabular}{cccc}
\textrm{$\lambda$}&\tabincell{c}{$Nu=ARa^{\alpha}$\\Regime II \hspace{0.5 cm} Regime III}&$Re=BRa^{\beta}Pr^{\epsilon}$&\tabincell{c}{$\sigma_T/\Delta T=C\times Ra^{\gamma}$\\center \hspace{0.5 cm} sidewall} \\
4.0 & $0.00052Ra^{0.59}$\quad$ 0.004Ra^{0.50 }$ &$0.088Ra^{0.551}Pr^{-0.66}$& $0.08Ra^{-0.09}$\quad$0.138Ra^{-0.08}$ \\
1.9 & $0.005Ra^{0.50}$\qquad$0.016Ra^{0.43}$ &$0.169Ra^{0.525}Pr^{-0.71}$& $0.19Ra^{-0.13}$\quad$0.135Ra^{-0.09}$ \\
1.0 &$0.015Ra^{0.43}$ \quad$0.10Ra^{0.35}$ &$0.395Ra^{0.486}Pr^{-0.69}$& $0.49Ra^{-0.18}$\quad$0.92Ra^{-0.18}$ \\
0.5 & $0.74Ra^{0.36}$ \qquad$0.15Ra^{0.30}$ &$0.481Ra^{0.472}Pr^{-0.72}$&$0.24Ra^{-0.16}$\quad$3.60Ra^{-0.23}$ \\
\end{tabular}
\caption{\label{tab:table1} Measured scaling law of the Nusselt number $Nu$ and normalized temperature fluctuations $\sigma_T/\Delta T$ in Regime III for different values of $\lambda$.}
\label{tab:kd}
\end{center}
\end{table}
We next study the $\lambda$-dependence of the heat transport enhancement behaviour in Regime II and Regime III. Figure \ref{fig_enhancement} shows the heat transport enhancement in rough cells compared with that in a smooth cell. It is seen that the heat transport enhancement in Regime II (open symbols in figure \ref{fig_enhancement}) first increases and then decreases with increasing $\lambda$. The situation is very different in Regime III (solid symbols in figure \ref{fig_enhancement}) where it is found that the heat transport enhancement increases monotonically with increasing $\lambda$. A maximum heat transport enhancement of 637\% is observed at $\lambda=4.0$.
We note that in a DNS study, \citet{Wanger2015JFM} show that the heat transport is enhanced more for slender roughness elements, i.e. larger $\lambda$, using only four roughness elements with a roughness height of 0.125H (``obstales" in \cite{Wanger2015JFM}) in a quasi-2-D convection cell. The observation in \cite{Wanger2015JFM} is qualitatively in agreement with what we find here, e.g. heat transport enhancement generally increases with $\lambda$ (see e.g. figure \ref{fig_enhancement}).
\subsection{Transition from Regime II to Regime III: The viscous boundary layer cross-over \label{SubSec_Transition}}
To understand what may have happened at $Ra^*$, i.e. the transition from Regime II to Regime III, we recall that there are two BLs in turbulent RBC, i.e. the thermal BL and the viscous BL. The viscous BL is thicker than the thermal BL for $Pr>1$. According to the Grossmann-Lohse model for turbulent RBC, BLs play crucial role in determine the global heat transport \citep{GL2000JFM}. The transition seen in figure \ref{fig1} may be understood in terms of the viscous BL cross-over, i.e. the roughness elements not only strongly perturb the thermal BL, but also alter the viscous BL in Regime III. To verify this, we directly measured the dynamics of the viscous boundary layer in a rough cell with $\lambda=1.0$.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{figure6-eps-converted-to.pdf}
\caption{\label{fig_PIV}(Colour online) The instantaneous ($a$) and time-averaged ($b$) velocity fields measured near the bottom plate within the large-scale circulation plane ($Ra=4.22\times 10^9$, $Pr=4.34$ and $\lambda=1.0$). The velocity magnitude (in units of mm/s) is colour coded with the colour bar nearby each figure, and also represented by the length of the vectors. The velocity field is not accessible experimentally below the dashed lines.}
\end{center}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure7-eps-converted-to.pdf}
\caption{\label{fig2} (Colour online)($a$) Typical time-averaged horizontal velocity $U$ and root-mean-square horizontal velocity $\sigma_u$ profiles ($Ra= 4.22\times10^9$, $\lambda=1.0$ and $Pr=4.34$). The vertical dashed line indicates the location of the viscous boundary layer. ($b$) Normalized viscous boundary layer thickness $\delta_v/H$ vs $Ra$. The dashed line marks the height of the roughness elements. The datum at the highest $Ra$ (symbols with cross, $\Delta T= 29.83K$) shows different behaviours with other data points which we believe is an artefact due to the strong optical distortion of the images used for PIV analysis.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure8-eps-converted-to.pdf}
\caption{\label{fig_Nu_Re} (Colour online) $Nu$ as a function of $Re$ measured in rough cells with different values of $\lambda$ at ($a$) $Pr=23.34$ ($b$) $Pr=6.14$ ($c$) $Pr=4.34$ and ($d$) $Pr=3.57$.}
\end{figure}
An example of the instantaneous velocity map measured at $Ra=4.22\times10^{9}$ and $Pr=4.34$ is shown in figure \ref{fig_PIV}($a$). The triangles outline the space occupied by the roughness elements schematically. The velocity magnitude is coded by both the colour and the length of the vectors. The instant velocity field shows that the mean flow is going from the right to the left and there are intermittent bursts of the flow velocity. Figure \ref{fig_PIV}($b$) depicts the time averaged velocity field with the velocity magnitude coded also in both colour and the length of the vectors. One may notice that the time-averaged maximum velocity magnitude is significantly reduced when compared with the instantaneous one as revealed by the scale bars. After averaging over time one sees that the mean velocity is uniform with respect to $x$ except very close to the plate, e.g. in between the roughness elements. A tiny counter-rotating vortex may be identified in the region -8 mm $< x < $ -5 mm and 0 mm $ < z < $ 4 mm. The mean velocity field varied weakly with $x$ (figure \ref{fig_PIV}($b$)). We thus averaged the mean velocity field along $x-$direction to obtain the mean velocity and the root-mean-square (rms) velocity profiles.
Figure \ref{fig2}($a$) shows an example of the horizontal mean velocity profile $U(z)$ and the profile of the rms velocity $\sigma_U(z)$ measured at $Ra=4.22\times 10^9$ and $Pr=4.34$. A remarkable feature shown in figure \ref{fig2}($a$) is that the rms velocity is comparable to the mean velocity very close to the plate, and it's even larger than the mean velocity within the viscous BL. The origin of the enhanced velocity fluctuation is not clear but speculatively this should largely correlate with the intermittent bursts of thermal plumes. In contrast, the maximum rms velocity in a smooth cell is only about 30\% of the maximum mean velocity and it is always smaller than the mean velocity within the boundary layer region \citep*{Sun2008JFM, Wei2013JFM}.
The viscous boundary layer thickness was determined using the so-called slope method: The intersection between the extrapolation of the linear part of the velocity profile very close to the plate and the maximum mean horizontal velocity was defined as the thickness of the viscous boundary layer $\delta_v$ (indicated by the vertical dashed line in figure \ref{fig2}($a$)) \citep{XinXiaTong1996PRL,XinXia1997PRE,QiuXia1998PRE,Sun2008JFM, Wei2013JFM}. The $Ra$-dependence of the viscous BL thickness normalised by the height of the convection cell is plotted in figure \ref{fig2}($b$). With increasing $Ra$, $\delta_u/H$ becomes thinner and thinner. At $Ra \sim 3\times 10^9$, $\delta_u/H$ approximately equals to $h/H$. This roughly corresponds to $Ra^*$ seen in the measured $Nu$ ($Ra^*\sim 4\times 10^9)$. This observation indicates that the transition of the heat transport law between Regimes II and III may be caused by the viscous BL becoming smaller than the height of the roughness elements. After this viscous BL cross-over, the roughness elements strongly perturb both the thermal and the viscous BLs. The thermal plumes emitted from the tip of pyramids can then be directly ejected into the bulk. For $Pr=23.34$, the estimated viscous BL thickness at the smallest $Ra$ explored in the experiments ($Ra=4.39\times 10^9$) was 0.04H \citep{Lam2002PRE}, which was very close to the height of roughness element, implying that the measurements with $Pr=23.34$ were all in Regime III.
In the asymptotically large $Ra$ limit, the thickness of the BLs is so small such that the roughness elements perturb both the thermal and the viscous BLs. The system is in Regime III in this case wherein $Nu$ is very sensitive to the LSC, or equivalently to the Reynolds number $Re$ associated with the LSC (to be defined in section \ref{SubSec_Re}). We plot in figure \ref{fig_Nu_Re} $Nu$ as a function of $Re$ for ($a$) $Pr$ = 23.34, ($b$) $Pr$ = 6.14, ($c$) $Pr$ = 4.34, and ($d$) $Pr$ = 3.57. On one hand, figure \ref{fig_Nu_Re} ($a$) shows that $Nu$ increases with $\lambda$ for all $Re$ measured at $Pr=23.34$, which are in the Regime III. On the other hand, $Nu$ at other $Pr$ where mostly within the Regime II shows weak dependence on $Re$. This is another evidence supporting that the viscous BL cross-over may also play important role in turbulent thermal convection with the presence of roughness elements.
\subsection{Reynolds number measurements\label{SubSec_Re}}
The self-organised large-scale circulation (LSC, or ``wind of turbulence'') is one of the fascinating features of turbulent RBC. The flow dynamics of the LSC is characterised by the associated Reynolds number $Re$. We study in this section how the roughness geometry alters the dynamics of the LSC by examining the $Re$ dependence on both $Ra$ and $Pr$. It has been well-known that there exits a well-defined low-frequency oscillation in turbulent RBC. The time scale of this oscillation is interpreted as the turnover time $T_f$ of the LSC (see, e.g. \citet*{Xie2013JFM} and references therein). By choosing a typical length scale of the LSC, e.g. 4H, the Reynolds number is calculated as:
\begin{equation}
Re=UH/\nu=4H^2/(T_f\nu)
\end{equation}
Time-resolved temperature signals measured $\sim 1$cm away from the sidewall were used to determine $T_f$ and thus the $Re$ as a function of $Ra$ and $Pr$ when changing the roughness geometry, i.e. at different values of $\lambda$. The turnover time of the LSC was determined by locating the time position of the first peak of the temperature auto-correlation function $C(\tau)$, which is defined as:
\begin{equation}
C(\tau)=\frac{\langle[T(t+\tau)-\langle T(t) \rangle_t ][T(t)-\langle T(t)\rangle_t]\rangle_t}{\sigma_T^2}
\end{equation}
where $\langle\cdots\rangle_t$ denoted time-averaging and $\sigma_T$ is the rms temperature. An example of the autocorrelation function measured in convection cell with $\lambda=4.0$ at $Ra=1.01\times 10^{11}$ and $Pr=23.34$ is depicted in figure \ref{fig_Oscillation}. Clear oscillations can be seen from $C(\tau)$. Fitting the data points nearby the first peak of $C(\tau)$ with a six order polynomial, the time location of the first peak was determined and denoted as $T_f$. The uncertainty of $T_f$ was from $0.06$ s to $0.008$ s with increasing $Ra$.
\begin{figure}
\centering
\includegraphics[width=0.65\textwidth]{figure9-eps-converted-to.pdf}
\caption{\label{fig_Oscillation} Autocorrelation function of the temperature measured $\sim 1$ cm away from sidewall in a rough cell with $\lambda=4.0$ at $Ra=1.01\times 10^{11}$ and $Pr=23.34$. Here $T_f$ is the turnover time of the LSC.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure10-eps-converted-to.pdf}
\caption{\label{fig_Re} (Colour online) Reynolds number $Re$ as a function of $Ra$ and $Pr$ in rough cells for ($a$) $\lambda=4.0$, ($c$) $\lambda=1.9$, ($e$) $\lambda=1.0$ and ($g$) $\lambda=0.5$. To determine more precisely the scaling exponent of $Re$ with respect to $Ra$, data measured at $Pr=3.53, 4.34$ and $6.14$ in ($a, c, e, g$) are multiplied by a $Pr$-dependent constant $f(Pr)$ to collapse with data at $Pr=23.34$. The shifted data are shown in the main plot in the right panel ($b, d, f, g$) with power law fit $Re f(Pr)\sim Ra^{\beta}$ shown as solid lines. The insets in ($b, d, f, g$) show the compensated plot $ReRa^{-\beta}$ with respect to $Pr$ with the solid lines representing power law fits to the data. See table \ref{tab:table1} for the fitted power laws.}
\end{figure}
Figure \ref{fig_Re} displays $Re$ as a function of $Ra$ and $Pr$ in rough cells with $\lambda=4.0$ ($a,b$), $\lambda=1.9$ ($c,d$), $\lambda=1.0$ ($e,f$) and $ \lambda=0.5$ $(g,h)$. Let's first discuss the case for $\lambda=4.0$. The $Re$ as a function of $Ra$ for $Pr=3.57, 4.34, 6.14$ and $23.34$ are plotted in figure \ref{fig_Re}($a$) on a log-log scale. The $Re$ exhibits power law dependence on $Ra$ at different $Pr$, i.e. $Re\sim Ra^{\beta}$. To determine more precisely $\beta$, data at $Pr$ = 3.57, 4.34 and 6.14 were multiplied by a $Pr$-dependent constant to collapse them with data at $Pr=23.34$ onto a single line. The shifted data are shown in the main plot of figure \ref{fig_Re}{($b$)}. A power law fits to the data over almost three decades of $Ra$ yielded: $Re\sim Ra^{0.551\pm0.001}$. The $Ra$-dependence of $Re$ was then divided out by compensating the data in figure \ref{fig_Re}($a$) with the so-determined power law, e.g. the inset of figure \ref{fig_Re}($b$). Within a relatively limited range of $Pr$, the data was also described by a power law, i.e. $Re/Ra^{0.551}\sim Pr^{-0.66\pm 0.02}$. The so-determined $Re(Ra, Pr)$ was:
\begin{equation}
Re=0.088Ra^{0.551}Pr^{-0.66}\quad (\lambda=4.0)
\end{equation}
Applying the same procedure to data obtained in rough cells with $\lambda=1.9$ (figure \ref{fig_Re}($c$)), $\lambda=1.0$ (figure \ref{fig_Re}($e$)), and $\lambda=0.5$ (figure \ref{fig_Re}($g$)), we determined the $Ra$ and $Pr$ dependence of $Re$ for each values of $\lambda$, respectively. The scaling laws of $Re=BRe^{\beta}Pr^{\epsilon}$ for different values of $\lambda$ are summarised in table \ref{tab:table1}.
On one hand, with $\lambda$ increasing from 0.5 to 4.0 the scaling exponent $\beta$, i.e. $Re\sim Ra^{\beta}$, increases from 0.472 to 0.551, which is plotted as a function of $\lambda$ in figure \ref{fig3}. The data may be described by an empirical power law within very limited range of $\lambda$, i.e. $\beta=0.49\lambda^{0.08\pm0.01}$. On the other hand, the scaling exponent $\epsilon$, i.e. $Re\sim Pr^{\epsilon}$, remains at the same value around $-0.70$ within experimental uncertainty except for $\lambda=4.0$. Whether there is a change of the $Pr$-dependence of $Re$ in the rough cell with $\lambda=4.0$ requires future investigation with a wider range of $Pr$. The change of $\beta$ with $\lambda$ suggests a change of the LSC dynamics when the geometry of the roughness elements is changed. It is thus interesting to study torsional and sloshing oscillations of the LSC dynamics in rough cells with different values of $\lambda$ as those done in a smooth cell \citep{Funfschilling2004PRL,Xi2009PRL,Xie2013JFM}. This is beyond the scope of the present studies and will be investigated in the future. Interestingly, the transition of the heat transport scaling law from Regime II to Regime III is not observed in the Reynolds number measurement.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure11-eps-converted-to.pdf}
\caption{\label{fig_ReRaNu} (Colour online) Reynolds number $Re$ as a function of $RaNu$ for different values of $\lambda$ with the solid line representing a power law fit to the data irrespective of $\lambda$. ($a$) $Pr=23.34$ and $Re=0.076\times (RaNu)^{0.36\pm0.01}$; ($b$) $Pr= 6.14$ and $0.266\times (RaNu)^{0.35\pm0.01}$; ($c$) $Pr= 4.34$ and $0.281\times (RaNu)^{0.36\pm0.01}$; ($d$) $Pr=3.57$ and Re=$0.340\times (RaNu)^{0.36\pm0.01}$.}
\end{figure}
It is found very recently that $Re$ is insensitive to the plate morphology for a given input heat flux for $\lambda=0$ and 0.5 \citep{Wei2014JFM}, i.e. when $Re$ is plotted against the flux Rayleigh number $RaNu$, data at different values of $\lambda$ collapse. We plot $Re$ as a function of $RaNu$ for different values of $\lambda$ and $Pr$ in figure \ref{fig_ReRaNu}. The data at the same $Pr$ but different $\lambda$ indeed overlap onto each other which is in agreement with the observation in \citet{Wei2014JFM}, but now for a wider range of $\lambda$ and $Pr$. In addition, data at the same $Pr$ can be fitted by a power law, i.e. $Re=0.076\times (RaNu)^{0.36\pm0.01}$ $(Pr=23.34)$, $Re=0.266\times (RaNu)^{0.35\pm0.01}$ $(Pr = 6.14)$, $Re=0.281\times (RaNu)^{0.36\pm0.01} (Pr=4.34) $, and $Re=0.340\times (RaNu)^{0.36\pm0.01}$ $(Pr = 3.57)$. The scaling exponents of these power laws fits are the same within experimental uncertainty but the magnitudes decrease with $Pr$.
\subsection{Local temperature fluctuations\label{SubSec_Temp_fluc}}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure12-eps-converted-to.pdf}
\caption{\label{fig_PDF_localTemp} (Colour online) Probability density functions (PDFs) of the standardised temperature measured ($a$) at the cell centre and ($b$) near sidewall for different values of $\lambda$ ($Ra=2.1\times 10^{10}$ and $Pr=23.34$). The root-mean-square temperatures as a function of $\lambda$ at the cell centre and sidewall are plotted in ($c$) and ($d$), respectively.}
\end{figure}
Since Regime II is a transitional state, we focus on temperature fluctuations measured in Regime III only. To simplify the analysis, we don't take into account the effects of $Pr$ and limit the analysis to measurements done in FC770 only ($Pr$=23.34). The results may shed light upon the mechanism leading to the change of the heat transport scaling in rough cells with different roughness geometries in Regime III.
The probability density function (PDF) of the standardised temperature $(T-\langle T \rangle)/\sigma_T$ measured at $Ra=2.1\times10^{11}$ at four different values of $\lambda$ at the cell centre and sidewall are displayed in figure \ref{fig_PDF_localTemp} ($a$) and ($b$) respectively, where $\sigma_T$ is the rms temperature that is shown as a function of $\lambda$ in figure \ref{fig_PDF_localTemp} ($c$) and ($d$) at the cell centre and sidewall. It is seen that the PDFs at different $\lambda$ overlap onto each other very nicely at the cell centre, suggesting that the bulk turbulence shares the similar dynamics at different values of $\lambda$. A notable feature seen in figure \ref{fig_PDF_localTemp} ($a$) is that the PDFs are significantly skewed toward the negative side, i.e. more cold plumes are detected in the cell centre than hot ones. This feature hasn't been observed either by us or reported by others in smooth cells with the same boundary conditions, i.e. constant temperature at the top boundary and constant heat flux at the bottom boundary. Thus, the reason for this phenomenon remains unknown to us at present. At sidewall, the shapes of the PDFs remain approximately invariant with $\lambda$ except the case with $\lambda=0.5$. For measurements with $\lambda=0.5$ the thermistor was so-positioned such that the upwelling hot plumes of the LSC were detected, while for the other cases the downwelling cold plumes of the LSC were detected. The positive tails of the PDFs overlap well onto each other while the negative tails, especially those larger than $3\sigma_T$, change with $\lambda$. The rms temperature at the cell centre increases considerably with $\lambda$ as can be seen from figure \ref{fig_PDF_localTemp}($c$), and that at the sidewall first remains almost the same for $\lambda$=0.5 and 1.0 and then increases with $\lambda$ as well. As the large temperature fluctuations are caused by thermal plumes, this observation suggests a change of the plume dynamics with increasing $\lambda$.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure13-eps-converted-to.pdf}
\caption{\label{fig_rms}(Colour online) $\lambda$-dependence of the normalised root-mean-square temperature as a function of $Ra$ at ($a$) the cell centre and ($b$) the sidewall. The solid lines are power law fits to the individual data set with the scaling exponents displayed nearby each data set. The data are measured at $Pr=23.34$.}
\end{figure}
The Rayleigh number dependence of normalised rms temperatures $\sigma_{T}/\Delta T$ for different values of $\lambda$ measured at the cell centre and near the sidewall are shown in figure \ref{fig_rms}($a$) and ($b$), respectively. With increasing $\lambda$, $\sigma_T/\Delta T$ enhances considerably. The changes lie not only in the magnitudes but also the scaling exponents with $Ra$. The scaling exponents of the power law fits, e.g. $\sigma_T/\Delta T=C\times Ra^{\gamma}$, are displayed nearby each data set in figure \ref{fig_rms}($a$) and ($b$) and the scaling laws of each data set are summarised in table \ref{tab:table1}. The scaling exponent increases from -0.16 to -0.09 with $\lambda$ increased from 0.5 to 4.0 in the cell centre. For comparison, \citet{Wei2016JFM} obtained an exponent of -0.17 in the centre of a smooth cell for a comparable $Pr$ ($Pr$ = 12.3). With a possible $Pr$-dependence of the temperature fluctuation unknown at the present, the experimental results suggest that the $Ra$-dependence of the temperature fluctuation at the cell centre becomes stronger with increasing $\lambda$. Near the sidewall, the scaling exponent increases from -0.23 to -0.08 in the $\lambda$ range covered in the experiments, suggesting a weakened dependence of temperature fluctuation on $\lambda$ at the sidewall. To the best of our knowledge, there has been no report on the smooth-cell value for this exponent in the similar $Pr$ range near sidewall.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure14-eps-converted-to.pdf}
\caption{\label{fig_sigma_T} (Colour online) Probability density function of the temperature deviation from the mean $T-\langle T\rangle$ normalised by the applied temperature between the top and bottom plate $\Delta T$ at ($a$) the cell centre and ($b$) the sidewall ($Ra=2.1\times 10^{10}$, $Pr=23.34$). }
\end{figure}
To gain further insight into the dynamics of temperature fluctuations with increasing $\lambda$, we examine the PDFs of the normalised temperature deviation from the mean $(T-\langle T \rangle)/ \Delta T$ for different values of $\lambda$ at $Ra=2.1\times 10^{10}$ at the cell centre and the sidewall respectively, as displayed in figure \ref{fig_sigma_T}($a$) and ($b$). The figure reveals that large events are becoming more probable with increasing $\lambda$ at the same $Ra$, i.e. hotter and colder plumes, suggesting that the effects of changing roughness geometry are to modify the dynamics of thermal plumes and make them more coherent and energetic (therefore decay slower). This is also supported by the shadowgraph visualisation of the flow field, such as those presented in figure \ref{fig_shadow}. These images are taken at $Ra=1.22\times 10^{11}$ and $Pr=$23.34. It is seen that, with increasing $\lambda$, both the density and the morphology of the thermal plumes change drastically. In the smooth cell there is a well-defined LSC with few thermal plumes passing through the cell centre (figure \ref{fig_shadow}($a$)). In the rough cell with $\lambda=1.0$, more thermal plumes are passing through the cell centre (figure \ref{fig_shadow}($b$)). In the rough cell with $\lambda=4.0$, thermal plumes occupy almost all the space in the convection cell (figure \ref{fig_shadow}($c$)). Furthermore, the plumes are now packed (or clustered) much more densely. This will result in their slower decay through thermal diffusion, which explains why it is now much more probable to find large temperature excursions in the cell centre (figure \ref{fig_sigma_T}($a$)).
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure15-eps-converted-to.pdf}
\caption{\label{fig_shadow} Shadowgraph images of the instantaneous flow fields in a rough cell with ($a$) $\lambda$=0.0 ($b$) $\lambda=1.0$ and ($c$) $\lambda=4.0$. The images were taken at $Ra=1.22\times 10^{11}$ and $Pr=23.34$.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure16-eps-converted-to.pdf}
\caption{\label{fig_plume} (Colour online)
($a$) The time averaged plume amplitude $\langle A \rangle$ normalised by $\Delta T$ and ($b$) the time averaged plume width $\langle w\rangle$ normalised by the large-scale circulation turnover time $T_f$ as a function of $Ra$ extracted from the temperature time series measured at the sidewall at $Pr=23.34$. The scaling of the power law fits are displayed nearby each data set.}
\end{figure}
To quantitatively understand the change of the dynamics of thermal plume, we extract time series of thermal plume amplitude $A$ and width $w$ using the time-resolved temperature data measured near sidewall. The detailed plume extraction method can be found in \citet{Zhou2016PRF}. The normalised time-averaged plume amplitude $\langle A\rangle/\Delta T$ and normalised time-averaged plume width $\langle w\rangle /\tau_0$ are shown in figure \ref{fig_plume}($a$) and ($b$) respectively. Figure \ref{fig_plume}($a$) reveals that with the increase of $\lambda$, the plume amplitude $A/\Delta T$ increases. The solid lines are power law fits to the data at different values of $\lambda$ with the scaling exponents listed nearby each data set. The scaling exponent of plume amplitude increases from -0.386 to -0.266 with $\lambda$ increasing from 0.5 to 4.0. In contrast, the $Ra-$dependence of the plume width remains almost the same with a scaling exponent around 0.25 for $\lambda$ smaller than 4.0. It should be noted that the scaling exponent of plume width for $\lambda=4.0$ increases considerably (0.341) when compared with other values of $\lambda$, which may be correlated with the drastically changed plume dynamics and morphology in rough cell with $\lambda=4.0$. As the thermal plumes are main heat carriers in turbulent RBC, the experiments suggest that the change of plume dynamics may lead to the change of the scaling law of heat transport for different values of $\lambda$.
\subsection{A proper parameter characterising the system: $\lambda$ vs number density of roughness elements \label{SubSec_con_para}}
When varying $\lambda$ the number density $n$ of roughness elements also varies, i.e. $n\simeq\pi/4\Gamma^2\lambda^2(H/h)^2$. In fact, the two parameters have one-to-one correspondence provided that the ratio of the roughness element height $h$ to the cell height $H$ and $\Gamma$ are fixed, which is the case for the present study. To figure out whether $\lambda$ is the most relevant parameter that characterises the system, we carried out another experiment, in which $\lambda$ was kept constant at 4.0 and the height of the roughness elements was halved. Thus the number density of the roughness elements is increased by a factor of 4.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{figure17-eps-converted-to.pdf}
\caption{\label{fig_Nu4} (Colour online) ($a$) $Nu$ as a function of $Ra$ and $Pr$ at fixed $\lambda=4.0$ measured in rough cells with roughness height $h$ of 4 mm and 8 mm. The down-pointing triangles are from \citet{Denis2005JFM} and the hexagons are from \citet{Sun2005JFM} in smooth cells. The solid lines are power law fits to the respective data set. The power law fitting in cell with $h=$ 8 mm is extrapolated to $Ra=2\times 10^{11}$. ($b$) Compensated plot of the same data in ($a$).}
\end{center}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.65\textwidth]{figure18-eps-converted-to.pdf}
\caption{\label{fig_Re4}(Colour online) $Re$ as a function of $Ra$ in rough cells with $h=$ 8 mm (solid symbols) and 4 mm (open symbols) at $Pr=$ 4.34 (triangles) and 23.34 (squares), respectively. In both cells $\lambda=4.0$. The solid lines are power fits to each data set. For $Pr=4.34$, $Re=0.028R^{0.56}$ ($h=8$ mm) $Re=0.014R^{0.58}$ ($h=$4 mm); and for $Pr=23.34$ $Re=0.013R^{0.55}$ ($h=8$ mm) $Re=0.0035R^{0.59}$ ($h=$4 mm). }
\end{figure}
We show in figure \ref{fig_Nu4}($a$) $Nu$ as a function of $Ra$ for measurements in $\lambda$ = 4.0 rough cells with $h=8$ mm and $h=4$ mm together with the same smooth cell data which have be shown in figure \ref{fig_Regime_I_II} and figure \ref{fig1}. The estimated viscous boundary layer thickness at the highest $Ra$ reached in the experiment is 4.8 mm. Thus, the measurements at $Pr=23.34$ and $h=4$ mm are in Regime II. A scaling exponent $\alpha$ = 0.60 is observed in rough cell with $h=4$ mm in Regime II, which is consistent with 0.59 observed in Regime II of rough cell with $h =$ 8 mm, which is indicated by the two parallel lines in figure \ref{fig_Nu4}. The similar $Ra$-dependence of $Nu$ in Regime II in rough cells with $h$ = 4 mm and 8 mm can be seen more clearly from the compensated plot in figure \ref{fig_Nu4}($b$). These observations suggest that $\lambda$ is not only a suitable parameter describing the geometry of the pyramid-shaped roughness elements, but also a suitable parameter to characterise the response of turbulent RBC over rough plates with different roughness geometries.
A noticeable feature from figure \ref{fig_Nu4}($a$) is that the measured $Nu$ in the rough cell with $h =$ 8 mm is larger than that in the rough cell with $h =$ 4 mm in Regime II despite the fact that they have similar $Ra$-dependence. This may be explained by the measured Reynolds number based on the circulation time of the LSC as plotted in figure \ref{fig_Re4}. It is seen that $Re$ in the rough cell with $h= 8$ mm is larger than that in the rough cell with $h=$4 mm, which is true for both $Pr =$ 4.34 and $Pr=$ 23.34. The result suggests that a faster LSC can transport heat more efficiently cross the convection cell.
\section{Conclusion and outlook\label{Sec_con}}
We studied systematically in this paper the effects of roughness geometry on the global heat transport, the flow dynamics and the local temperature fluctuations in turbulent Rayleigh-B\'enard convection over rough plates. The experiments were carried out in rough cells with pyramid-shaped roughness elements directly machined on the top and bottom plates. We defined a parameter $\lambda$ (the height of the roughness element over its base width) to characterise the geometry of the roughness element and to analyse the observed scaling behaviour of $Nu(Ra)$ and $Re(Ra, Pr)$, where $Ra$ and $Pr$ are respectively the Rayleigh number and Prandtl number, $Nu$ and $Re$ are respectively the Nusselt number and Reynolds number. We varied $\lambda$ from 0.5 to 4.0 by keeping the roughness height at a constant of 8 mm and changing the base width from 16 mm to 2 mm. In addition, a rough cell with $h =$4 mm was used. The experiments covered Rayleigh number $7.5\times 10^{5}\leq Ra \leq 1.31\times 10^{11}$ and $Pr$ from 3.57 to 23.34.
It is found that the heat transport is enhanced significantly by roughness elements, and the enhancement depends strongly on the roughness geometry. It is also found that the heat transport scaling, i.e. $Nu\sim Ra^{\alpha}$, may be classified in to three regimes in turbulent RBC over rough plates. In Regime I, the system is in a dynamically smooth state. The heat transport scaling in a rough cell is the same as that in a smooth cell in this regime. In Regimes II and III heat transport enhances. The scaling exponents $\alpha$ of $Nu$ vs. $Ra$ increase from 0.36 to 0.59 with $\lambda$ increased from 0.5 to 4.0 in regime II. It increases from 0.30 to 0.50 in Regime III with $\lambda$ increased from 0.5 to 4.0. The experiment thus clearly demonstrated that the heat transport scaling in turbulent RBC could be manipulated using $\lambda$ in Regimes II and III. The transition from Regime I to Regime II may be understood in terms of the thermal boundary layer thickness becoming smaller than the roughness height with increasing $Ra$ as suggested by a number of previous studies \citep{Shen1997PRL, Du1998PRL,Ciliberto1999PRL,Roche2001PRE,Tisserand2011PoF}. The transition from Regime II to Regime III may be understood in terms of viscous boundary layer crossing, i.e. the viscous BL thickness becomes thinner than the height of the roughness elements, as shown by direct measurement of the viscous boundary layer profiles in the present study. When this happens, the thermal plumes emitted from the tip of the roughness can be directly ejected into the bulk flow. An unexpected $Pr$ effect in turbulent RBC over rough plates was also observed: Larger heat transport enhancement was observed for larger values of $\lambda$ and larger $Pr$. This may be correlated with the strong clustering of thermal plumes as clearly seen from the shadowgraph visualization of the flow field (figure \ref{fig_shadow}).
The Reynolds number $Re$ is measured based on the turnover time of the large-scale circulation evaluated near the sidewall of the cells. Both the $Ra$ and $Pr$ dependence of $Re$ were studied for different roughness geometries, i.e. $Re=B \times Ra^{\beta} Pr ^{\epsilon}$. The scaling exponent $\beta$ increased from 0.472 to 0.551 with $\lambda$ increasing from 0.5 to 4.0. and $\epsilon$ remained at the same value around $-0.7$ except for the case of $\lambda=4.0$ ($\epsilon=-0.66$). With a limited range of $Pr$, a truly increased $\epsilon$ for $\lambda=4.0$ was not for assertion. Consistent with the observation in \citet{Wei2014JFM}, it is found that $Re$ is insensitive to the plate morphology for a fixed flux Rayleigh number, i.e $RaNu$, but now with a wider range of $Pr$ and $\lambda$ as compared with the experiment of \citet{Wei2014JFM}.
Time-resolved local temperatures at the cell centre and sidewall were measured using two small thermistors. The probability density functions (PDFs) of the standardised temperature ($T-\langle T\rangle)/\sigma_T$ in the cell centre remain almost invariant with respect to $\lambda$. The measured local temperature fluctuations at both the cell centre and sidewall increase considerably with $\lambda$. The scaling exponent of the normalised rms temperature ($\sigma_T$/$\Delta T$) with $Ra$ increases from $-0.16$ to $-0.09$ at the cell centre and from $-0.23$ to $-0.08$ at the sidewall with $\lambda$ increased from 0.5 to 4.0. By extracting thermal plume amplitude and width using a method described in \citet{Zhou2016PRF}, it is found that both the plume amplitude and width increase with $\lambda$. The scaling exponent of plume amplitude increases from -0.386 to -0.266 with $\lambda$ increasing from 0.5 to 4.0. The scaling exponent for the plume width remains at the same value around 0.25 except for the case of $\lambda=4.0$ which is -0.34. As thermal plumes are the main heat carriers in turbulent RBC, the experiments suggested that that the change of thermal plume dynamics might be a possible reason leading to the change of the scaling laws between $Nu$ and $Ra$ in Regime III.
The experiments demonstrated that the heat transport behaviour in turbulent thermal convection can be manipulated using a roughness parameter $\lambda$ that characterises the geometry of the roughness elements. It further showed that the flow dynamics and local temperature fluctuations are also altered by roughness geometry. There are several open questions: What is the effects of $\lambda$ on the dynamics of the large-scale circulation, e.g. twisting/sloshing oscillation, flow cessation/reversal? How will the vertical mean and variance temperature profiles change with $\lambda$?
Finally, we remark that the 0.5 scaling exponent observed for $\lambda$ = 4.0 in Regime III reminds one of the Kraichnan ultimate regime of turbulent convection \citep{Kraichnan1962PoF}. Whether Regime III for $\lambda=$ 4.0 really corresponds to the Kraichnan regime needs further investigations. The place to look would be the boundary layer, i.e. a truly turbulent boundary layer should be the ultimate test for the Kraichnan regime. A very recent DNS study suggests roughness may be used as a route to trigger the transition to the ultimate state of turbulent convection \citep{Wettlaufer2017PRL}. If this is indeed the case, then the dynamics of turbulent convection in the Kraichnan regime may be studied in a Rayleigh number range that is accessible using classical fluids like water in a table top experiment.
\section*{Acknowledgments}
We thank Yu-Hao He for help with the experiment. This work is supported by the Hong Kong Research Grants Council under grant Nos. CUHK1430115 and CUHK404513.
\bibliographystyle{jfm}
|
\section{Discussion and conclusion}
All the major WiFi chip families (Atheros, Intel, and Marvell) expose CSI~\cite{csitool, cupid}. Hence, we believe WiCapture\xspace can be added to any commodity WiFi infrastructure.
To enable high-speed communication, WiFi chips transmit reference signals on multiple frequencies and receivers, typically, have multiple antennas. Algorithm~\ref{algo:mt} can be extended to use these additional signals to improve AoD estimation. Also, Algorithm~\ref{algo:mt} tracks the target when it moves in a 2D plane. 3D tracking can be enabled by extending Algorithm~\ref{algo:mt} to find the direction of paths in a 3D space.
The paper focuses on positioning alone since that is the hardest to provide. Orientation can be relatively easily obtained via IMUs already available on many devices~\cite{gearvr}.
In conclusion, we developed and implemented a WiFi-based motion tracking system and demonstrated its performance in different deployment scenarios. A commodity WiFi-based position tracking system would potentially enable VR on mobile devices and revolutionize number of applications that can be enabled on top of motion tracking.
\section*{Acknowledgments}
\vspace{-2mm}
Manikanta Kotaru was supported by Thomas and Sarah Kailath Stanford Graduate Fellowship. The authors thank Prof. Gordon Wetzstein and Stanford Computational Imaging Group for valuable feedback on the paper and the anonymous reviewers for their insightful comments.
\vspace{-2mm}
\section*{Glossary}
\begin{Glo}
\item \textbf{AoD (Angle of Departure): }The angle of a path with respect to a line joining two antennas at the source.
\item \textbf{AP: }A WiFi access point or router.
\item \noindent \textbf{CSI (Channel State Information): }The received signal when the source transmits the reference signal.
\item \textbf{Frequency offset: }The difference in frequency of the reference signal used at two WiFi chips.
\item \textbf{Packet: }WiFi chunks data into units of efficient size for communication and each unit is called as a packet.
\end{Glo}
\section{Design}
\begin{figure*}
\begin{center}
\includegraphics[width=1\linewidth]{./figDesign/overall}
\end{center}
\caption{Target moves by $\vec{\delta}$. The new position of the target is shown with solid target icon and the old position with shaded icon. In the first step, for each AP, WiCapture\xspace obtains unit vectors along the direction of departure of all the paths. Here, there are two paths, green and red paths, with directions $\vec{r}_{\theta_1}$ and $\vec{r}_{\theta_2}$. The paths for first packet are shown using solid lines and the paths for the second packet using dashed lines. Next, the phase of complex attenuation along both the paths for both the packets is calculated. Finally, the effect of frequency offset on the phase values, $\nu_2 - \nu_1$, is removed to obtain equations dependent on just $\vec{r}_{\theta_1}$, $\vec{r}_{\theta_2}$ and $\vec{\delta}$, which can be solved to estimate $\vec{\delta}$.}
\label{fig:arch}
\end{figure*}
Fig.~\ref{fig:arch} shows how WiCapture\xspace would be deployed and summarizes the solution procedure. The target transmits normal WiFi signals. Each WiFi AP/receiver calculates CSI (or image) and sends it to a central server. The server uses CSI from all the access points (APs) to determine the position
of the target by performing the following two steps:
\vspace{-4mm}
\begin{Enumerate}
\item Estimate AoD and complex attenuation of all the paths.
\item Use the attenuation from consecutive CSIs along with AoD of the paths to estimate the displacement of the transmitter during the time between consecutive CSIs.
\end{Enumerate}
\subsection{Estimating AoD of all the paths}\label{sec:pathEst}
Let us rewrite the CSI for the first packet from Eq.~\ref{eq:H2_hat} as $\widehat{\textbf{H}}_1 = \textbf{A} (\textbf{F} \mathrm{e}^{j \nu_1})$. We can observe that CSI is nothing but a linear combination of steering vectors in $\textbf{A}$. If one obtains the steering vectors, then finding the AoD of all the paths is trivial using Eq.~\ref{eq:stvecaoa}. However, the problem of obtaining steering vectors from a single linear combination of theirs is ill-posed. Prior work~\cite{vanderveenJointSmoothingProof} has theoretically investigated that the more number of linear combinations of steering vectors with independent weights are provided, the higher the accuracy of the estimated steering vectors.
Note that the AoDs of paths are relatively stationary and do not change rapidly from packet to packet. For example, if the receiver is $3$ m away from the transmitter, a tiny motion of $1$ mm changes the AoD of any path by atmost $0.02$ degrees, which does not cause any measurable changes in the elements of the steering matrix. However, the weights in the linear combination change significantly as even a small motion creates significant change in the phase of complex attenuation along a path.
Using this insight, CSI from $P$ packets where AoD has not changed much are concatenated using Eq.~\ref{eq:csiCon}. CSI for any of these packets is linear combination of the same steering vectors. So,
\begin{equation}\label{eq:csiCon}
\textbf{X} = [\widehat{\textbf{H}}_1 \; \widehat{\textbf{H}}_2 \; \ldots \; \widehat{\textbf{H}}_P]
= \textbf{A} \textbf{G},
\end{equation}
where $\textbf{X} \in \mathbb{C}^{3 \times P}$, $\textbf{A} \in \mathbb{C}^{3 \times L}$ is steering matrix which is common for all the $P$ packets, $\textbf{G} \in \mathbb{C}^{L \times P }$ is a matrix of steering vector weights. This is standard form for applying well-known MUSIC algorithm to compute AoD~\cite{music, spotfi, xie2016xd}. The goal of the MUSIC algorithm is to find all the steering vectors given multiple linear combinations of the steering vectors and it provides an accurate estimate of $\textbf{A}$ due to diverse linear combinations.
\subsection{Estimating the displacement of the transmitter between consecutive packets}\label{sec:disp}
We will now show how the displacement of the WiFi transmitter between two consecutive packets can be estimated. Consider the CSI from the first two packets.
\noindent\textbf{Estimate the steering vectors' weights: }
CSI is a linear combination of steering vectors estimated using procedure in Sec.~\ref{sec:pathEst}. For $p^{\mathrm{th}}$ packet, these weights in the linear combination, $\widehat{\textbf{F}}_p \in \mathbb{C}^{L \times 1}$, can be obtained using
\begin{equation}\label{eq:AhatH}
\widehat{\textbf{F}}_p = \textbf{A}^\dagger \widehat{\textbf{H}}_p. \footnote{$ \textbf{B}^\dagger $ is the pseudo-inverse of $ \textbf{B} $.}
\end{equation}
Substituting Eq.~\ref{eq:H2_hat} into Eq.~\ref{eq:AhatH}, we can observe that
\begin{equation}\label{eq:Fhat}
\widehat{\textbf{F}}_1 = \textbf{F} \mathrm{e}^{j \nu_1}, \widehat{\textbf{F}}_2 = \textbf{D} \textbf{F} \mathrm{e}^{j \nu_2}
\end{equation}
\noindent\textbf{Estimate the change in the complex attenuation for each of the paths: }
Note that $\textbf{D} \in \mathbb{C}^{L \times L}$ and it's entries contain information about target's displacement (see Eq.~\ref{eq:H2}). We obtain an estimate ${\textbf{D}}$ by solving the convex optimization problem \ref{eq:con1} using standard procedures~\cite{ConvexBoyd}.
\begin{equation}\label{eq:con1}
\begin{aligned}
& \underset{{\textbf{D}}}{\text{minimize}}
& & \|\widehat{\textbf{F}}_2 - {\textbf{D}} \widehat{\textbf{F}}_1 \| \\
& \text{subject to}
& & {\textbf{D}} \text{ is diagonal}.
\end{aligned}
\end{equation}
From Equations~\ref{eq:Fhat},~\ref{eq:con1} and~\ref{eq:H2}, the $k^\mathrm{th}$ diagonal element of ${\textbf{D}}$, ${\textbf{D}}_{k,k}$, is an estimate of $\mathrm{e}^{-j2\pi (\vec{r}_{\theta_k}^\top \cdot \vec{\delta})/\lambda + j \nu_2 - j\nu_1}$. Note that unit vector in the direction of the departure can be obtained using $\vec{r}_{{\theta}_k} = [\cos({\theta}_k) \; \sin({\theta}_k)]^\top$; here ${\theta}_k$ is the AoD for the particular path estimated using procedure in Sec.~\ref{sec:pathEst}. So, if the term $(\nu_2 - \nu_1)$ is removed, one can estimate $\vec{\delta}$ from the phase of elements of $\textbf{D}$.
However, as discussed in Sec.~\ref{pre:cfo}, the term $\nu_2 - \nu_1$ due to frequency offset is orders of magnitude larger than the term $-2\pi (\vec{r}_{{\theta}_k}^\top \cdot \vec{\delta})/\lambda$ due to displacement. This change in phase due to frequency offset is not only unavailable but also extremely hard to predict~\cite{brown2012fundamental} making it hard to estimate and remove the term $\nu_2 - \nu_1$. This is precisely the phenomenon that led to long-held notion that phase information across packets is uncorrelated in commodity WiFi systems~\cite{nandakumar2014wi} and is unusable for tracking.
\noindent\textbf{Use attenuation-change between paths:
Our unique insight is that one can get rid of the effect of frequency offset by using the phase of the signal from multiple paths. This is surprising as multipath is traditionally viewed as a complicating distortion that needs to be compensated in many ToF depth sensor and WiFi localization systems~\cite{spotfi, tof_sparse_kadambi2013coded}. Notice that the effect of frequency offset is the same for all the paths, \textit{i.e.}, the term, $\nu_2 - \nu_1$, is present in the change of phase for all the paths. So, effect of the clock differences is removed by considering the change in the phase of signal along a path with respect to change in the phase of signal from another path.
Specifically, consider the phase of the complex number ${\textbf{D}}_{k+1,k+1}/{\textbf{D}}_{1,1}$. It is an estimate of $(-2\pi \vec{r}_{{\theta}_{k+1}}^\top \cdot \vec{\delta}/\lambda + \nu_2 - \nu_1) - (-2\pi \vec{r}_{{\theta}_1}^\top \cdot \vec{\delta}/\lambda + \nu_2 - \nu_1) = (-2\pi (\vec{r}_{{\theta}_{k+1}} - \vec{r}_{{\theta}_1})^\top \cdot \vec{\delta}/\lambda )$. Notice that the term, $\nu_2 - \nu_1$, canceled out by performing this operation. So, a $(L-1)$-dimensional vector $\vec{s}$ is calculated whose $k^{\mathrm{th}}$ element is
\begin{equation}\label{eq:s}
\vec{s}_k = \text{phase of } ({\textbf{D}}_{k+1,k+1}/{\textbf{D}}_{1,1}).
\end{equation}
Then $\textbf{R} \in \mathbb{R}^{(L-1) \times 2}$ is calculated whose $k^\mathrm{th}$ row is
\begin{equation}\label{eq:R}
\frac{-2\pi}{\lambda} [(\textrm{cos}({{{\theta}}_{k+1}})-\textrm{cos}({{{\theta}}_{1}})) \;\; (\textrm{sin}({{{\theta}}_{k+1}})-\textrm{sin}({{{\theta}}_{1}}))].
\end{equation}
One can then obtain an estimate of the displacement by solving the simple linear least squares problem \ref{eq:con2}.
\begin{equation}\label{eq:con2}
\vec{{\delta}} = \arg\!\min \| \textbf{R} \vec{{\delta}} - \vec{s} \|.
\end{equation}
If there are multiple APs, matrix $\textbf{R}$ and vector $\vec{s}$ obtained from multiple access points are concatenated vertically. If there are $L$ paths from the target to each of the $U$ APs, then the concatenated $\textbf{R} \in \mathbb{R}^{U*(L-1) \times 2}$ and the concatenated $\vec{s} \in \mathbb{R}^{U*(L-1)}$. Since the target displacement is same irrespective of the AP, one can estimate of the displacement by solving Eq.~\ref{eq:con2} using these concatenated matrices. We summarize the overall algorithm in Algorithm \ref{algo:mt}.
\vspace{-3mm}
\begin{algorithm}
\caption{WiCapture\xspace 's motion tracking algorithm}
\label{algo:mt}
\KwData{CSI of packets from target to each of the $U$ APs}
\KwResult{Trajectory traced by the target }
Initiate the trajectory at origin \;
\For{\text{each packet $p$ received at APs} }{
Consider packets received within the last $V$ seconds. Let the number of such packets be $P$ \;
Form $\textbf{X}$ from CSI of $P$ packets using Eq.~\ref{eq:csiCon} \;
Apply MUSIC~\cite{music} on $\textbf{X}$ to find AoD of $L$ paths \;
Obtain the steering vector weights using Eq.~\ref{eq:AhatH} \;
Obtain change in complex attenuation between $p^{\mathrm{th}}$ and $(p+1)^{\mathrm{th}}$ packets by solving~\ref{eq:con1} \;
Form $\textbf{R}$ using Eq.~\ref{eq:R} and $\vec{s}$ using Eq.~\ref{eq:s} \;
Update the trajectory by adding displacement obtained by solving~\ref{eq:con2} \;
}
\end{algorithm}
\vspace{-4mm}
\section{Evaluation}
We implemented WiCapture\xspace using off-the-shelf Intel 5300
WiFi cards which support three antennas. We employed Linux CSI tool~\cite{csitool} to obtain the
CSI. The WiFi cards operate in 5 GHz WiFi spectrum ($f$ in Sec.~\ref{sec:prel}). Also, the CSI information is quantized,
i.e., each of real and imaginary parts of CSI is represented using 8 bits.
The system used for evaluation consists of APs and a target device equipped with WiFi cards. The target/transmitter has a $3-$antenna circular array with distance between any two antennas equal to $2.6$ cm. The APs operate in monitor mode. CSI is calculated once every $6$ ms. So, the trajectory is estimated at an update rate of $167$ Hz for our evaluation experiments. An implementation is provided in~\cite{kotaru_wic_code}. We use $U=4$ APs, set $V=10$ s and $L=2$ paths in Algorithm~\ref{algo:mt}.
\subsection{Stationary experiments}
We start by examining the jitter/precision of WiCapture\xspace, \ie, how stationary the estimated position is when the target is stationary. This is important for VR because the scene displayed on the VR headset is not expected to change when the user is not moving. In our experiment, the target remains stationary and transmits $1000$ packets. The access points are placed at the corners of $5$ m $\times$ $6$ m space. The standard deviation of the trajectory obtained by using Algorithm~\ref{algo:mt} is used as the measure of jitter. The median jitter observed over 21 experiments at different target positions is $0.25$ mm.
We measured the jitter of Oculus DK2 system by placing it at the same positions. The Oculus camera is placed at $0.75$ m away from the headset. The median jitter observed for Oculus DK2 is $0.10$ mm. So, WiCapture\xspace's jitter in position estimation is comparable to that of a commercial position tracking system which requires dedicated infrastructure.
\subsection{Controlled tracking experiments}\label{sec:controlled}
Next we evaluate the resolution of WiCapture\xspace. We mount the target on a mechanical stage which has a least count of $0.005$ cm. The target is moved in increments of $0.1$ cm and then in decrements of $0.1$ cm so that it reaches the initial position as shown in Fig.~\ref{fig:mmStage}(a). The target transmits a WiFi packet at each position. Trajectory estimated by WiCapture\xspace is translated so that the initial position is origin. The maximum error in estimation of position of any point in the trajectory is $0.11$ cm.
We conducted another experiment where the target is moved mechanically to different positions on a trajectory shown in Fig.~\ref{fig:mmStage}(b). The maximum error in estimation of position of any point in the trajectory is $0.27$ cm. Note that WiCapture\xspace resolves even millimeter-level target motion.
\begin{figure}[t!]
\hfill
\subfigure[Trajectory 1]{\includegraphics[width = 0.49\linewidth]{./figEval/mmStage}}
\hfill
\subfigure[Trajectory 2]{\includegraphics[width = 0.49\linewidth]{./figEval/mmStage_8}}
\caption{Controlled trajectories where the target is positioned at different locations in a trajectory mechanically.}
\label{fig:mmStage}
\end{figure}
\subsection{Extensive tracking experiments}\label{sec:extensive}
Motion tracking accuracy of WiCapture\xspace is dependent on the multipath environment,
the material used in walls, the presence of metallic
objects, the density of WiFi AP deployment and many other
factors. In this evaluation, we test WiCapture\xspace's accuracy in different deployment scenarios.
\begin{figure*}[!htb]
\hfill
\subfigure[Indoor office deployment]{\includegraphics[width = 0.32\linewidth]{./figEval/setupLos}}
\hfill
\subfigure[Occlusion deployment]{\includegraphics[width = 0.32\linewidth]{./figEval/setupNlos}}
\hfill
\subfigure[Outdoor deployment]{\includegraphics[width = 0.32\linewidth]{./figEval/setupOutside}}
\caption{Experiment setups for the three deployments}
\label{setup}
\end{figure*}
\begin{figure*}[!htb]
\hfill
\subfigure[Indoor office deployment]{\includegraphics[width = 0.32\linewidth]{./figEval/losError}}
\hfill
\subfigure[Occlusion deployment]{\includegraphics[width = 0.32\linewidth]{./figEval/nlosError}}
\hfill
\subfigure[Outdoor deployment]{\includegraphics[width = 0.32\linewidth]{./figEval/outsideError}}
\caption{Cumulative Distribution Function (CDF) of trajectory error for the three deployments}
\label{err}
\end{figure*}
\noindent\textbf{Compared Approaches:} To the best of our knowledge, there exists no WiFi-based system that is geared towards \textit{tracking} the motion of a commodity WiFi chip. We faithfully implemented SpotFi, a state-of-the-art WiFi localization system and compared WiCapture\xspace against it. We did not compare with other non-vision based systems like ultrasound~\cite{liu2013guoguo} because of limitations detailed in section~\ref{sec:relatedWork} such as extensive installation of dedicated infrastructure.
\noindent\textbf{Ground truth: }We use an Oculus DK2~\cite{dk2Video} headset which is rigidly attached to the target to obtain ground truth trajectory. We placed the headset $0.9$ m away from the Oculus camera. Under these conditions, we calculated the accuracy of Oculus position tracking to be at sub-millimeter level using mechanical stage experiments (see Sec.~\ref{sec:controlled}).
\noindent\textbf{Metric (Trajectory error): }To focus on the shape of the trajectory alone, we translate the trajectory reported by each of the systems (Oculus DK2, SpotFi and WiCapture\xspace) by the initial position of the trajectory so that the initial position of all the trajectories is origin. Similarly, since the reference coordinate axes of different systems are not aligned, we rotate the trajectory reported by WiCapture\xspace (and SpotFi) so that the Root Mean Squared Error (RMSE) between the points of the rotated trajectory of WiCapture\xspace (and SpotFi) and the points on the Oculus trajectory is minimized. Note that we just shift and rotate the trajectories of WiCapture\xspace and SpotFi but do not scale the trajectories. As in other motion tracking systems~\cite{rfidraw}, absolute point-by-point position difference between this shifted and rotated trajectories and the ground truth trajectory is reported as the trajectory error.
\vspace{-4mm}
\subsubsection{Indoor office deployment}\label{sec:los}
\vspace{-2mm}
\textbf{Method:} We deployed WiCapture\xspace in a $5$ m $\times$ $6$ m room with access points placed at the four corners and the target is in the same room. This is the typical access point deployment density used in state-of-the-art WiFi localization systems~\cite{spotfi}. We traced $97$ trajectories with the target device. Fig.~\ref{fig:trace} shows a sample trajectory. The target was moved in a continuous manner on a table rather than point by point as done in Sec.~\ref{sec:controlled}. Fig.~\ref{setup}(a) shows the experimental setup.
\noindent\textbf{Analysis: }From Fig.~\ref{err}(a), WiCapture\xspace achieves a median trajectory error of $0.88$ cm. SpotFi's median trajectory error, $132$ cm, is more than two orders of magnitude larger than that of WiCapture\xspace. WiCapture\xspace thus achieves sub-centimeter-level motion tracking using commodity WiFi.
\vspace{-4mm}
\subsubsection{Occlusion deployment}
\vspace{-2mm}
\textbf{Method:} We evaluate WiCapture\xspace under challenging conditions where the target is one room and all the APs are occluded from the target either by furniture or by walls. Fig.~\ref{setup}(b) shows the experimental setup which shows couple of APs where one AP is placed outside the room where the target is placed and another AP is separated from the target through a cubicle. We traced $64$ different trajectories.
\noindent\textbf{Analysis: }Fig.~\ref{err}(b) plots the CDF of the trajectory error of WiCapture\xspace and SpotFi when either one or two access points are outside the room where the target is placed. However, note that all the access points are occluded whether they are inside the target's room or not. Under these conditions, WiCapture\xspace achieves a median trajectory error of $1.51$ cm. Thus, even in the challenging conditions where the target is occluded from the positioning system, WiCapture\xspace achieves accuracy acceptable for many applications.
We further stress-tested by placing three or more APs outside the room at which point the reconstructed trajectories accumulated large errors. So, WiCapture\xspace can provide accurate tracking only when atleast two access points are present which are not separated from the target by walls.
\vspace{-4mm}
\subsubsection{Outdoor deployment}
\vspace{-2mm}
\textbf{Method:} We evaluate WiCapture\xspace in a $5$ m $\times$ $6$ m outdoor space. The experiments are conducted in shade for accurate Oculus measurements. Fig.~\ref{setup}(c) shows the experimental setup. We traced $89$ trajectories.
\noindent\textbf{Analysis: }From Fig.~\ref{err}(c), WiCapture\xspace achieves a median trajectory error of $0.85$ cm compared to SpotFi's $57$ cm trajectory error. Thus, WiCapture\xspace can ubiquitously track a commodity WiFi device as long as there is WiFi infrastructure irrespective of whether the target is indoors or outdoors.
\subsection{Deep dive into WiCapture\xspace}
WiCapture\xspace achieves accurate motion tracking due to two novel techniques. First, WiCapture\xspace accurately resolves multipath by using multiple packets for estimation. Second, WiCapture\xspace accurately removes the phase distortion due to frequency offset between the transmitter and the receiver. We now test the significance of each of these factors individually. For the following experiments, we consider the data from the indoor office deployment scenario in Sec.~\ref{sec:los}.
\begin{figure}[t!]
\hfill
\subfigure[Improved AoD estimation accuracy]{\includegraphics[width = 0.49\linewidth]{./figEval/aodError}}
\hfill
\subfigure[Removing phase distortion due to different clocks]{\includegraphics[width = 0.49\linewidth]{./figEval/cfoError}}
\caption{(a) plots the CDF of AoD estimation error when using multiple packets (WiCapture\xspace) and when using one packet. (b) plots the CDF of trajectory error when using WiCapture\xspace and when using \textit{assume same clock} method which ignores frequency offset.}
\label{deepdive}
\end{figure}
\vspace{-4mm}
\subsubsection{Improved AoD estimation}
\vspace{-2mm}
\textbf{Method:} We compare WiCapture\xspace against an alternate method which uses single packet for AoD estimation (set $P$ equal to $1$ in Eq.~\ref{eq:csiCon}). Error of a particular method (single packet or multiple packets) is measured by using the absolute difference between ground truth direct path AoD (which is measured manually) and AoD estimated using the particular method that is closest to
this ground truth.
\noindent\textbf{Analysis:} From Fig.~\ref{deepdive}(a), $80^{\mathrm{th}}$ percentile AoD estimation error by using multiple packets is $11$ degrees and is $3\times$ smaller than the error obtained by using a single packet.
\vspace{-4mm}
\subsubsection{Removing phase distortion from frequency offset}
\vspace{-2mm}
\textbf{Method:} We compare WiCapture\xspace against an alternate method which solves a system of equations where each equation equates the change in the phase of complex attenuation of a path directly as a linear function of displacement without considering the offset from the clock differences. We call this alternate method as \textit{assume same clock}.
\noindent\textbf{Analysis:} From Fig.~\ref{deepdive}(b), the trajectory estimation errors are $6 \times$ worse when frequency offset is ignored.
\section{Introduction}
Immersive experiences like virtual reality (VR) require accurate tracking of the headset and other accessories like hand-motion controllers. Current commercial tracking systems like Oculus Rift~\cite{rift} and HTC Vive~\cite{vive} are outside-in where the tracking is performed using infrastructure external to the VR accessories. The external infrastructure is specialized and typically uses infrared (IR) cameras along with sensors on the headset to perform the tracking. These systems are very accurate but have the following limitations:
\vspace{-3mm}
\begin{Itemize}
\item{They require installing specialized hardware and dedicated infrastructure wherever user wants to experience VR. So if a user wishes to use VR headsets anywhere in her home, one would need IR cameras everywhere.}
\item{These systems are not occlusion resistant. For example, if the camera is blocked by furniture or if the user turns away from the camera, then the tracking fails.}
\item{These systems have limited range, typically around $2$ m in front of the camera~\cite{dk2Video}.}
\end{Itemize}
A competing technology to provide position tracking is inside-out position tracking found in systems like the Microsoft Hololens~\cite{holo}. These systems use cameras (both RGB and depth sensing) and implement vision based tracking algorithms on the headset. These systems are both accurate and infrastructure-free, however they come with certain limitations. Specifically, they significantly increase the complexity of the headset since they need to have several cameras as well as complex algorithms running on the headset to provide tracking. Further they are not robust, tracking fails in environments with transparent or texture-less objects (\eg a white wall)~\cite{patent:20160131761}. Finally and most importantly, these systems cannot be used for tracking peripherals such as hand-motion controllers; the complexity of inside-out tracking is too high to be implemented on such peripherals which are meant to be lightweight and cheap.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{./figEval/mainTraj}
\end{center}
\caption{Solid blue path estimated by WiCapture\xspace is compared against the dotted red ground truth trajectory.}
\label{fig:trace}
\end{figure}
In this paper, we present WiCapture, a novel VR position tracking system which addresses the above limitations of existing systems. WiCapture is a WiFi based position tracking system. Headsets transmit standard WiFi packets which are received by standard WiFi access points (APs). The WiFi APs receive metadata from each WiFi packet reception called Channel State Information (CSI) which encodes the transformation the environment has induced upon the transmitted WiFi signals. WiCapture invents novel algorithms that mine the CSI metadata to recover the position of the headset accurately. It has the following properties:
\begin{Itemize}
\item{WiCapture does not require special hardware, it uses commodity APs that can be bought in retail stores.}
\item{WiCapture is occlusion resistant, it continues to work even when the APs and headsets are occluded due to furniture or other objects in between.}
\item{WiCapture has larger range and operates across rooms.
\item{It is insensitive to room illumination or texture and it can work in the dark. Further, headset complexity is minimal, all the headset needs is a standard WiFi chip.}
\end{Itemize}
At a high level, as illustrated in Fig.~\ref{fig:rabbit}, WiCapture\xspace obtains the change in the position of the transmitter by using the change in the phase of the CSI between packets. As with the ToF (Time of Flight) cameras, the phase is distorted due to the signal received from reflectors; this phenomenon is called multipath propagation. However, unlike ToF cameras where the light transmitter and camera are time-synchronized, the phase of WiFi signal is also distorted due to the lack of synchronization of clocks at the WiFi transmitter and receiver. WiCapture\xspace tackles these challenges using novel algorithms that compensate for these distortions and provides accurate phase measurement which in turn enables accurate position tracking.
\subsection{Contributions}
\begin{Itemize}
\item WiCapture\xspace is the \textit{first} commodity WiFi-based sub-centimeter level accurate tracking system.
\item We \textit{developed} a novel technique to overcome the distortion due to clock differences by exploiting the multipath. This is surprising as multipath is traditionally viewed as a complication in localization systems~\cite{spotfi}.
\item WiCapture\xspace is the \textit{first} system that accurately disentangles signal from different paths by using CSI from multiple packets. The key observation is that the direction of the paths remain stationary over small intervals of time and CSI of all the packets obtained within this time can be used to resolve multipath accurately.
\item We \textit{built} WiCapture\xspace using commodity Intel 5300 WiFi chips ~\cite{csitool} which demonstrated a precision of $0.25$ mm and a position tracking error of $0.88$ cm.
\end{Itemize}
\subsection{Limitations}
WiCapture's current prototype however has two limitations compared to existing systems. It has higher latency since the tracking is computed in the network and then provided as an update to the headset. Second, it is less accurate than current outside-in position tracking systems. We believe that WiCapture\xspace's accuracy is acceptable for VR given the significant benefits that WiCapture provides around deployment, coverage and occlusion resistance.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{./figDesign/rabbit.pdf}
\end{center}
\caption{The change in the phase of CSI can be modeled in terms of the displacement of the target. WiFi waves in $5$ GHz spectrum have $6$ cm wavelength. So, even millimeter-level motion creates measurable phase shift.}
\label{fig:rabbit}
\end{figure}
\section{Preliminaries}\label{sec:prel}
In WiFi systems, data is chunked into units of efficient size for communication and each unit is called as a packet~\cite{11n}. At the start of each packet, a WiFi transmitter sends a reference signal. A WiFi receiver correlates the received signal with the reference signal and samples at time zero to calculate Channel State Information (CSI). This is similar to how image is formed in a ToF camera by correlating the received signal at each pixel with a reference signal~\cite{tof_f_bhandari2014resolving}. So, each antenna is a sensor element which performs job similar to a pixel in a ToF camera and CSI calculated at WiFi receiver is similar to an image obtained at a ToF camera. We now provide a brief primer on CSI (or image) formation model. We included a glossary at the end of the paper for WiFi-specific terms.
\subsection{CSI calculation}
Consider the reference complex sinusoid signal $\mathrm{e}^{j 2 \pi f t}$ of frequency $f$ emitted by the $q^\mathrm{th}$ antenna on the transmitter; here $j$ is the complex root of $-1$ and $t$ is time. The signal passes through wireless channel $h_{q}$ resulting in the received signal $h_{q} \mathrm{e}^{j 2 \pi f t}$. CSI corresponding to the $q^\mathrm{th}$ transmit antenna, $\hat{h}_{q}$, is obtained as
\vspace{-1mm}
\begin{equation}\label{eq:csiEst}
{\hat{h}}_{q} = \frac{1}{T} \int_{0}^{T} h_{q} \mathrm{e}^{j 2 \pi f t} \mathrm{e}^{-j 2 \pi f t + j \nu} \, dt = h_{q} \mathrm{e}^{j \nu},
\end{equation}
where $T$ is time for which sinusoid is transmitted and $\nu$ is the phase of the receiver's local sinusoidal signal relative to the phase of the transmitter's sinusoidal signal (see Fig.~\ref{fig:clk}).
In indoor environments, similar to ToF camera systems, transmitted signal travels along multiple paths and the signals from all the paths superpose to form the received signal. Each path is associated with an AoD (Angle of Departure) from the transmitter and attenuation of the signal along the path. We now describe the relation between the CSI and these path parameters. For this description, we consider a linear $3$-antenna array (see Fig.~\ref{fig:aoa}) although the model can be extended to any arbitrary antenna array geometry.
\subsection{Wireless channel modeling}
Consider an environment with $L$ paths; \eg, the setup in Fig.~\ref{fig:arch}(b) has $2$ paths. Let the spacing between two consecutive antennas at the transmitter be $d$ (see Fig.~\ref{fig:aoa}). Let the AP/receiver be in the same plane as the transmitter antenna array. Let $\theta_k$ denote the angle of departure (AoD) from the transmitter for the $k^\mathrm{th}$
path using the convention shown in Fig.~\ref{fig:aoa}. Let $\gamma_{k}$ denote the signal received along $k^{\mathrm{th}}$ path from the first transmitter antenna to the receiver.
The signal along $k^\mathrm{th}$ path travels different distances from different transmit antennas to the receiver. These differences result in different phases for CSI corresponding to different antennas. As described in ~\cite{arraytrack} and illustrated in Fig.~\ref{fig:aoa}, the vector of signals received from the transmit antennas along $k^\mathrm{th}$ path can be written as $\vec{a}(\theta_k) \gamma_{k} $, where
\begin{equation}\label{eq:stvecaoa}
\vec{a}(\theta_k) = [1 \; \mathrm{e}^{-j 2\pi d \cos (\theta_k) /\lambda} \; \mathrm{e}^{-j 4\pi d \cos (\theta_k) /\lambda} ]^\top .
\end{equation}
This vector $\vec{a}(\theta_k)$ is also known as the steering vector. We have as many steering
vectors as the number of paths.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.8\linewidth]{./figDesign/aod}
\end{center}
\caption{Uniform linear array consisting of $3$ antennas: For AoD of $\theta_k$, the target's signal travels an additional distance of $d \cos(\theta_k)$ from the second antenna in the array compared to the signal from the first antenna. This results in an additional phase of $-2\pi d \cos(\theta_k)/\lambda$ for the signal received from the second antenna compared to that from the first antenna.}
\label{fig:aoa}
\end{figure}
The signals from different paths superpose and result in the wireless channel $h_{q}$. So, wireless channel,
\begin{equation}\label{eq:H1}
\textbf{H} = [\vec{a}(\theta_1) \; \ldots \; \vec{a}(\theta_L)]\textbf{F} = \textbf{A} \textbf{F},
\end{equation}
where $\textbf{H} \in \mathbb{C}^{3 \times 1}$ is the wireless channel whose $q^\mathrm{th}$ element is $h_{q}$, $\textbf{F} \in \mathbb{C}^{L \times 1}$ is the matrix of complex attenuations whose $k^\mathrm{th}$ element is $\gamma_{k}$, $\textbf{A} \in \mathbb{C}^{3 \times L}$ is the steering matrix whose $k^\mathrm{th}$ column is the steering vector for $k^\mathrm{th}$ path.
Let $\widehat{\textbf{H}} \in \mathbb{C}^{3 \times 1}$ be the observed CSI (or image) whose $q^\mathrm{th}$ element is $\hat{h}_{q}$. Wireless channel $\textbf{H}$ would be related to the observed CSI matrix using the relation, $\widehat{\textbf{H}} = \textbf{H} \mathrm{e}^{j \nu}$.
\subsection{Change in channel due to displacement}\label{pre:disp}
We will now describe how the transmitter's motion affects the wireless channel. For the rest of the paper, we will index the CSI with corresponding packet, \ie, $\widehat{\textbf{H}}_p \in \mathbb{C}^{3 \times 1}$ is CSI for $p^\mathrm{th}$ packet and ${\textbf{H}_p} \in \mathbb{C}^{3 \times 1}$ represents the corresponding wireless channel. Let's say that the transmitter moved by a small displacement $\vec{\delta}$ between the times two consecutive reference signals are transmitted. Then the path-length for the $k^\mathrm{th}$ path changes by $(\vec{r}_{\theta_k} ^\top \cdot \vec{\delta})$ where $\vec{r}_{\theta_k}$ is the unit vector along the direction of departure for the particular path with AoD $\theta_k$. This induces a phase shift, $2\pi (\vec{r}_{\theta_k}^\top \cdot \vec{\delta})/\lambda$. So, attenuation of the $k^\mathrm{th}$ path gets multiplied by $\mathrm{e}^{-j2\pi (\vec{r}_{\theta_{k}}^\top \cdot \vec{\delta})/\lambda}$. Mathematically, if $\textbf{H}_1 = \textbf{A} \textbf{F}$ as in Eq.~\ref{eq:H1}, then the wireless channel for the second packet is
\begin{equation}\label{eq:H2}
\textbf{H}_2 = \textbf{A} \textbf{D} \textbf{F},
\end{equation}
where ${\textbf{H}_2} \in \mathbb{C}^{3 \times 1}$, $\textbf{A} \in \mathbb{C}^{3 \times L}$, ${\textbf{F}} \in \mathbb{C}^{L \times 1}$, and $\textbf{D} \in \mathbb{C}^{L \times L}$ is a diagonal matrix with entries $\mathrm{e}^{-j2\pi (\vec{r}_{\theta_1}^\top \cdot \vec{\delta})/\lambda}$, $\ldots$, $\mathrm{e}^{-j2\pi (\vec{r}_{\theta_{L-1}}^\top \cdot \vec{\delta})/\lambda}$, and $\mathrm{e}^{-j2\pi (\vec{r}_{\theta_L}^\top \cdot \vec{\delta})/\lambda}$.
\subsection{Phase distortion due to frequency offset}\label{pre:cfo}
\begin{figure}[t!]
\includegraphics[width=1\linewidth]{./figDesign/clk}
\caption{$T$ in Eq.~\ref{eq:csiEst} is on the order of few microseconds which is very small compared to the time interval between successive WiFi transmissions which is on the order of milliseconds. So, the relative phase can be assumed to be constant for $T$ period but changes on the order of few hundred radians within few milliseconds.}
\label{fig:clk}
\end{figure}
The transmitter and the receiver WiFi chips have different clocks and this creates an offset between the frequencies of the local reference sinusoids used at the transmitter and the receiver. Moreover, the offset is not constant with time as the clocks drift~\cite{zucca2005clock, brown2012fundamental}. As illustrated in Fig.~\ref{fig:clk}, the frequency offset results in a change in the relative phase between the sinudoids at the transmitter and the receiver (see Equation \ref{eq:csiEst}) from packet to packet. Let $\nu_p$ be the relative phase between the two sinusoids for $p^\mathrm{th}$ packet. Then,
\begin{equation}\label{eq:Hp}
\widehat{\textbf{H}}_p = {\textbf{H}_p} \mathrm{e}^{j \nu_p}.
\end{equation}
So, if the transmitter moved by $\vec{\delta}$ between consecutive packets, then using Equations~\ref{eq:H1},~\ref{eq:H2} and ~\ref{eq:Hp}, the CSI reported by the WiFi chip for the two packets can be written as
\begin{equation}\label{eq:H2_hat}
\widehat{\textbf{H}}_1 = \textbf{A} \textbf{F} \mathrm{e}^{j \nu_1}, \widehat{\textbf{H}}_2 = \textbf{A} \textbf{D} \textbf{F} \mathrm{e}^{j \nu_2}.
\end{equation}
Eq.~\ref{eq:H2_hat} relates the observed CSI to the transmitter's displacement. It is important to note that frequency offset of $20$ kHz is typical and WiFi standard allows it to be as high as $200$ kHz~\cite{11n}. This implies that the distortion in CSI due to frequency offset is orders of magnitude greater than the change caused by transmitter's motion. For example, if two consecutive WiFi packets are sent with a time-gap of $10$ ms (which is typical), then the relative phase between the transmitter and receiver sinusoids can change by $630$ radians and is observed in practice~\cite{nandakumar2014wi}. To compare, if the transmitter moved by $5$ mm during the same $10$ ms, then the change in the phase of any path is less than $0.6$ radians.
\section{Related work}\label{sec:relatedWork}
\textbf{Motion tracking} has been of great interest with applications in 3D object reconstruction, virtual/augmented reality, motion capture and motion control~\cite{welch2002motion, meyer1992survey}. Systems requiring infrared LEDs or photodiodes on the tracked object have short-range, limited field of view, difficulty in tracking multiple objects, and require line of sight between tracked object and sensor~\cite{kumar2016spatial, rift, patent:20160131761, laser, prakash}. Pulsed laser light based systems, in addition, require time synchronization of multiple laser light sources~\cite{laser}. WiCapture\xspace does not share these limitations as it has long range and typical obstructions like walls and humans are transparent to WiFi signals.
Magnetic signal-based systems~\cite{raab1979magnetic,polhemus} are occlusion-resistant but have a small range and are affected by distortions due to ferromagnetic materials ~\cite{wang1990tracking}. Radio frequency signal based techniques using RFIDs (Radio Frequency IDentification)~\cite{rfidraw, yang2014tagoram} and ultrawideband~\cite{gezici2005localization} signals demonstrated centimeter-level accuracy but have limited range and require specialized infrastructure that is not as ubiquitous as WiFi. Tracking systems using other modalities like ultrasound~\cite{foxlin1998constellation, shin2016application} and IR~\cite{vicon, dorfmuller1999robust} achieve high accuracy but require infrastructure dedicated for tracking.
\textbf{Infrastructure-free} approaches using inertial measurement units (IMUs) can only track the orientation of a device but not the position~\cite{esser2009imu}. Visual-inertial navigation systems~\cite{vio_engel2016direct, nerurkar2014c, cadena2016past} which use cameras and IMUs track the motion of a camera using natural image features unlike earlier systems~\cite{welch1999hiball} which required instrumentation of the environment with markers. However, visual systems have problems in environments with transparent or textureless objects~\cite{patent:20160131761} and are not applicable when the camera is occluded; for example, a phone cannot track itself when placed in a user's pocket. However, when they are applicable, we view WiCapture\xspace and infrastructure-free systems as complementary solutions that together can potentially form a robust and easy-to-deploy motion tracking system.
\textbf{Disentangling multipath} is a widely studied problem in Time of Flight (ToF) cameras and wireless literature as it enables several important applications like transient light imaging ~\cite{tof_tr_o2014temporal, tof_f_lin2014fourier, tof_tr_naik2015light}, wireless imaging~\cite{rfcapture_adib2015capturing, adib20143d} and localization ~\cite{spotfi, vasisht2016decimeter}. ~\cite{tof_f_rs4010021, tof_f_freedman2014sra, tof_f_kirmani2013spumic, tof_f_droeschel2010multi, tof_f_heide2013low, tof_f_dorrington2011separating, tof_f_bhandari2014resolving, tof_f_lin2014fourier, tof_f_kadambi2016macroscopic} explored transmitting signals at multiple frequencies to resolve multipath in ToF cameras. Similarly, wireless localization systems ~\cite{spotfi, xie2016xd} explored using multiple frequencies to resolve multipath. Unlike all the previous systems, we use signal received from multiple packets to improve the accuracy of estimated multipath parameters. This hinges on the fact that the multipath parameters like direction of the propagation paths are stationary over small time periods.
\textbf{WiFi-based localization systems} can be broadly classified into RSSI (Received Signal Strength Indicator) based, ToF based, and AoA (Angle of Arrival at the receiver) based systems. Signal strength (RSSI) based systems have 2-4 m localization error~\cite{radar,ez, Goswami2011} because of distortions in RSSI due to multipath. RSSI fingerprint-based approaches~\cite{horus, centaur} achieve $0.6$ m error but require an expensive, recursive fingerprinting operation. ToF based approaches achieve meter-level localization error ~\cite{mariakakis2014sail, xiong2015tonetrack, vasisht2016decimeter}. AoA based approaches achieved state-of-the-art decimeter-level localization error~\cite{spotfi,arraytrack,ubicarse,cupid,pinpoint}. WiCapture\xspace on the other hand is targeted towards estimating relative trajectory rather than the absolute position. WiFi-based tracking systems ~\cite{zeng2014your} can only classify the trajectory of the target WiFi device into limited number (four) of known gestures.
Techniques-wise, ~\cite{vanderveenJointSmoothingProof} theoretically observed that stationarity of multipath parameters improves multipath estimation accuracy. \cite{pandharkar2011estimating} removed the effect of time offset between laser source and camera receiver on the ToF measurements by considering differential ToF between different rays. ~WiCapture\xspace builds on these techniques and compensates not only for time offset but also frequency offset between the source and the receiver
|
\section{Introduction}
The synergy between competing nonlinearities
in the Schr\"odinger equation can give rise to very interesting dynamics
\cite{PhysRevLett.102.203903,setzpfandt2009competing}, including, for instance, solitons
\cite{PhysRevA.83.053838,Laudyn:15} and phase transitions
\cite{PhysRevLett.116.163902,0295-5075-98-4-44003}.
In this paper, we provide novel insights on the (focusing) cubic- (defocusing) quintic model, which has
been thoroughly studied in the context of nonlinear optics
\cite{mihalache1988exact,pushkarov1996bright,dimitrevski1998analysis},
where it was shown that large power solitons have neat similarities with regular liquids,
thereby motivating the term ``liquid light'' \cite{michinel2002liquid}.
The same equation
has been applied
in other frameworks too, see {\it e.g},
\cite{josserand1997coalescence,muryshev2002dynamics,khaykovich2006deviation,carretero2008nonlinear,davydova2003two}.
The cubic-quintic equation is an appropriate model for the propagation of
light in certain optical materials,
see for instance \cite{smektala2000non} and references in \cite{caplan2012existence}.
It has also been used as an approximation to the process of filamentation
\cite{piekara1974analysis,centurion2005dynamics,novoa2010filamentation}.
Recent experimental advances reinforce the significance of new theoretical studies.
Despite damping, (limited) soliton propagation has been observed in carbon disulfide
\cite{falcao2013robust}. Furthermore, the droplet-like behavior of cubic-quintic propagation
has been demonstrated in atomic gases at
low optical powers \cite{wu2013cubic,wu2015solitons},
using quantum coherence and interference as proposed in
\cite{michinel2006turning,alexandrescu2009liquidlike}.
Other setups in which the fifth order nonlinearity can be enhanced through quantum
effects comprise Rydberg atoms \cite{bai2016enhanced} and quantum dots
\cite{peng2014tunneling,tian2015giant}.
Confinement and guiding of light in a (defocusing) cubic- (focusing) quintic
has also been reported \cite{reyna2016guiding}.
In the cubic-quintic model, there is a one-parameter family of
form-preserving
traveling dark wave solutions within a critical bright background,
which was computed in
\cite{paredes2014coherent} following the numerical methods of \cite{chiron2016travelling}.
For small velocities, it consists of vortex-antivortex pairs of charges $\pm 1$ (we will make
a usual abuse of language and refer to ``velocity'' for what in the optical setup corresponds
to the propagation angle with respect to the axis).
For larger subsonic velocities, the solutions are
rarefaction pulses, namely dark blobs without vorticity.
The fainter the pulse is, the faster it moves within the bright background.
This family of solutions is similar to the one existing for third order
defocusing nonlinearity
\cite{jones1982motions,jones1986motions,berloff2004motions,bethuel2009travelling}.
Rarefaction pulses
should not be confused with the unstable quiescent bubbles of
\cite{barashenkov1988soliton,barashenkov1989stability}.
A separate issue is how these dark soliton-like excitations can be generated dynamically.
In the context of Bose-Einstein condensates (BEC), they have been generated by phase imprinting
\cite{proud2016jones}. In the framework of superfluids, it was shown that they can appear
when the fluid flows past an obstacle \cite{josserand1995cavitation}, a process that in optics
can be mimicked by the nonlinear interaction with an incoherently coupled beam
\cite{feijoo2014drag} and in BECs with a laser beam (see \cite{mironov2010structure}
and references therein).
A remarkable result of \cite{paredes2014coherent} is that, for the liquid of light, rarefaction
pulses can
be generated by interference in the collision of two bright solitons of very different sizes and powers.
The analogy with bubbles in fluids motivates the usage of the term cavitation for this kind of
process.
The produced caviton excitation propagates within the large soliton and can exit it
becoming a bright soliton again. This bright-dark-bright conversion is familiar in one dimension,
see {\it e.g.} \cite{kim2000soliton,garralon2013numerical}, but it is a distinctive
feature of the cubic-quintic equation
in two dimensions. This peculiarity facilitates the creation of dark traveling waves in a controlled manner
from initial conditions comprising only bright solitons.
With three initial bright solitons, two separate traveling waves can be created within the same
fluid.
The natural question that we address in the
present paper is how these traveling waves interact with each other.
It would be really interesting to
implement this kind of processes in experimental setups as those described in
\cite{falcao2013robust,wu2013cubic,wu2015solitons}.
For the case of defocusing cubic nonlinearity, the dynamics of the dark excitations
in a nontrivial background was analyzed in
\cite{smirnov2012dynamics,mironov2012propagation,mironov2013scattering}
and their interaction with a single vortex in
\cite{smirnov2015scattering}.
In section II, we fix notation and review some features of the cubic-quintic model.
In section III, we show
that vortex-antivortex pairs can be produced by a soliton-soliton collision.
Sections IV-VI describe the result of our simulations concerning dark wave interactions.
We discuss in turn the collision of two vortex-antivortex pairs, that of a rarefaction pulse
with a vortex-antivortex and that of two rarefaction pulses.
In section VII we outline our conclusions and make some final remarks.
The supplemental material \cite{suppl} contains animations for all of the examples of dynamical evolution
that are presented along the paper and a few extra illustrative cases.
\section{Solitons and traveling waves}
In this section we briefly review well-known results concerning the cubic-quintic
model in order to provide the basic ingredients for the following.
In the paraxial approximation, the canonical
equation
governing the wave amplitude $\psi(x,y,z)$ reads:
\begin{equation}
i\partial_z \psi = - (\partial_x^2 + \partial_y^2)\psi - (|\psi|^2 - |\psi|^4) \psi .
\label{CQeq}
\end{equation}
A refractive index of the form $n=n_0+n_2 I + n_4 I^2$ has been assumed, where $I \sim |\psi|^2$ is the
intensity. It is straightforward to check that the equation in terms
of physical quantities can be rescaled to the dimensionless variables of Eq. (\ref{CQeq})
without loss of generality
as long as $n_2>0$, $n_4<0$.
There are stable solitary waves
of the form $\psi= e^{i\beta z}f(r)$ with $\lim_{r\to \infty} f(r)=0$
which we laxly
call bright solitons, as it is customary in the literature.
The numerical study of \cite{dimitrevski1998analysis,michinel2002liquid}
shows that there are solutions for $0<\beta < \beta_{cr} = \frac{3}{16}$.
The power $P=2\pi \int r f(r)^2 dr$ grows monotonically with $\beta$ in the range
$P_0 < P < \infty$ where $P_0$ is the minimal value that leads to self-trapping.
For small $\beta$, the function $f(r)$ is bell-shaped. Near the
$\beta_{cr}$
eigenvalue cutoff
\cite{prytula2008eigenvalue}, $f(r)$ tends to a flat-top profile.
This means that $f \approx \Psi_{cr} = \frac{\sqrt3}{2}$ for $r<r_{sol}$ and
around the soliton radius $r_{sol}$ there is a quick drop to
$f \approx 0$ for $r>r_{sol}$. This limit is the liquid-like phase, in which the
soliton resembles a fluid with constant density and fixed surface tension
subject to the Young-Laplace equation \cite{novoa2009pressure}.
We will use these bright solitons to define the initial conditions of
simulations in the following sections, by considering:
\begin{eqnarray}
\psi|_{z=0}&=& f_1(|{\bf x}|) + f_2(|{\bf x}-{\bf x_2}|)\exp\left(i\frac{{\bf v_2}\cdot {\bf x}}{2}
+ i \phi_2\right)
+ \nonumber\\
&+& f_3(|{\bf x}-{\bf x_3}|)\exp\left(i\frac{{\bf v_3}\cdot {\bf x}}{2}
+ i \phi_3\right)
\label{init}
\end{eqnarray}
where the ${\bf x_i}$ are the initial positions of the solitons,
${\bf v_i}$ their initial velocities and $\phi_i$ their initial phases.
Boldface symbols are two-dimensional vectors.
The $f_i(.)$ are the soliton profiles, where $f_1$ is a flat-top soliton,
corresponding to the liquid where the dynamics takes place
and $f_2$, $f_3$ are smaller solitons that dynamically generate the dark excitations.
In Fig. \ref{fig1}, we plot the profiles of the particular solitons that will be used in
all the examples below.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig1.pdf}
\end{center}
\caption{(Color online) Radial profiles of the three solitons used to define the initial conditions in
the examples below. Their powers $P=2\pi \int_0^\infty r f(r)^2 dr$ are
$P|_{\beta=0.15}=86.0$, $P|_{\beta=0.1802}=2055.5$, $P|_{\beta=0.1856}=30620$.}
\label{fig1}
\end{figure}
Let us now turn to the dark traveling waves \cite{paredes2014coherent}.
They are form-preserving solutions of Eq. (\ref{CQeq}) moving at constant speed $U$
in, say, the $x$-direction, embedded in an infinite liquid.
Inserting the ansatz $ \psi(x,y,z)=e^{i\,\beta_{cr}\,z}\Psi(\eta,y) $ \cite{jones1982motions}.
where $\eta=x-Uz$, we can write:
\begin{equation}
iU\partial_\eta \Psi= (\partial_\eta^2 + \partial_y^2)\Psi +
\left(|\Psi|^2 - |\Psi|^4-\frac{3}{16}\right)\Psi
\label{EQtravel}
\end{equation}
subject to the boundary condition
$
\lim_{\eta^2 + y^2 \to \infty}\Psi = \Psi_{cr}=\frac{\sqrt3}{2}$.
There is a family of solutions parameterized by $0<U<\frac{\sqrt3}{2}$.
For small $U$ they are vortex-antivortex pairs, with
$|\psi|^2=0$ at the phase singularities. When $U$ grows, the vortex and antivortex
merge into a rarefaction pulse, whose $|\psi|^2$ is nowhere vanishing.
It is important to remark that the transition is completely smooth and, roughly, one can think of the
rarefaction pulse as a bound state of vortex and antivortex. In fact, under nontrivial
dynamical evolution both kinds of eigenstates can transform into each
other \cite{mironov2012propagation,smirnov2015scattering}.
An interesting quantity is the current density which, in the hydrodynamical picture,
represents the flow of the fluid.
\begin{equation}
{\bf j}= \frac{1}{i} (\psi^* {\bf \nabla} \psi - \psi {\bf \nabla} \psi^*)
\label{flow}
\end{equation}
The ${\bf j}$ is essential to understand how the dark excitation modifies the
medium around it and therefore to understand
the interaction between traveling waves. In figure \ref{fig2}, we depict
this quantity for three examples of traveling waves.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig2.pdf}
\end{center}
\caption{(Color online) Three numerically computed traveling waves with $U=0.11$
(vortex-antivortex pair, panel (a)), $U=0.35$ (rarefaction pulse, panel (b)) and
$U=0.71$ (faint rarefaction pulse, panel (c)).
The dark excitations are moving rightwards.
The color scale displays
the density $|\psi|^2$ and the arrows represent the current density ${\bf j}$.}
\label{fig2}
\end{figure}
Momentum and energy are conserved quantities defined by:
\begin{eqnarray}
p &=& \frac{1}{2i}\int \left[(\Psi^* - \Psi_{cr})\partial_x \Psi
-(\Psi - \Psi_{cr})\partial_x \Psi^* \right] dxdy
\nonumber
\\
E&=& \int |{\bf \nabla} \Psi|^2 dxdy +
\frac13
\int |\Psi|^2\left(|\Psi|^2-\Psi_{cr}^2 \right)^2 dxdy
\label{pE2}
\end{eqnarray}
Within the family of solutions, one can check that $U=\partial E/\partial p$
and three virial identities are satisfied \cite{jones1982motions,jones1986motions,paredes2014coherent}.
The analyisis in the coming sections results from the numerical integration of
Eq. (\ref{CQeq}) with initial conditions (\ref{init}). The computations are done
using a standard split-step beam propagation method
\cite{agrawal2007nonlinear}.
The evolution associated to the
non-derivative terms is computed with a fourth order
Runge-Kutta method. The plotted figures are built using grids of 800$\times$600
points. We have checked convergence of the method by comparing results with different
grids in $(x,y)$ and steps in $z$.
\section{Coherent generation of vortex-antivortex pairs}
In \cite{paredes2014coherent}, it was shown that a rarefaction pulse
can appear when two coherent bright solitons meet with appropriate relative
velocity and phase.
Roughly speaking, destructive interference generates a void at the collision point
which can acquire the necessary velocity thanks to the incoming momentum.
Although, definitely, an exact solution of
(\ref{EQtravel}) is not realized in the dynamical process,
the resulting robust dark excitation can indeed be identified with
a traveling wave solution. This fact was checked in \cite{paredes2014coherent} by
comparing the dispersion relations. Even if the size of the medium
(the large soliton) is not infinite, it can support the traveling wave if it
is much larger than the dark structure.
In this section, we show that a similar process can result in the formation of
a vortex-antivortex pair. In fact, the difference with \cite{paredes2014coherent}
is simply that the incomnig soliton has to be larger. What happens is that
during a collision in phase opposition, an elongated dark region is created. It cannot be stable
because there are no rarefaction pulse solutions of similar size. Consequently, it
evolves and decays through a snake instability giving rise to the separate vortex
and antivortex, which move forward together with a given velocity $U$.
Since the resulting configuration is not exactly equal to the stationary solution,
the dark regions can change, reconnect and split again. However
the vortex-antivortex profile becomes apparent after long enough propagation in $z$.
An example is depicted in figure \ref{fig3}.
Obviously, the third soliton of (\ref{init}) is not included in the initial condition.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig3.pdf}
\end{center}
\caption{(Color online) The encounter of two bright solitons giving rise to a traveling vortex-antivortex
pair. Initial conditions have ${\bf x_2}=(-180,0)$, ${\bf v_2}=(0.2,0)$, $\phi_2=5$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller one has $\beta_2=0.1802$.
The color code for $|\psi|^2$ is as in Fig. \ref{fig2} and
the range of the axes is $x \in [-270,270]$, $y \in [-190,190]$.
}
\label{fig3}
\end{figure}
In figure \ref{fig4}, we expose the phase structure
of the wavefunction of the example at a particular propagation distance $z$.
The plots prove that the two dark spots of figure \ref{fig3} correspond indeed to a vortex-antivortex
pair.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig4.pdf}
\end{center}
\caption{(Color online) Phase structure of the wavefunction. The plot corresponds to $z=2000$,
panel (f) of
figure \ref{fig3}. The region of the vortex-antivortex has been enlarged .
Panel (a) depicts $|\psi|^2$ with the same color scale of Fig. \ref{fig2}
and the arrows are a quiver diagram for ${\bf j}$ showing flows similar
to figure \ref{fig2}. Panel (b)
corresponds to the interference pattern
with a plane wave: $|\psi(x,y,z=2000)+7 \exp{(-100 i y)}|^2$. The fork-like
structures prove the existence of a vortex-antivortex pair with charges
$\pm 1$.
}
\label{fig4}
\end{figure}
Concerning the reconversion into a bright soliton \cite{paredes2014coherent},
we notice that it can
take place
when the excitation
reaches the boundary of the liquid of light as a single dark pulse. On the other hand,
when it does so as a vortex-antivortex pair, two waves propagating in opposite
directions along the
edge of the large soliton get excited \cite{suppl}.
It must be emphasized that the generation of vortex and antivortex is
only one of the possible qualitative outcomes that emerge depending on
the relative velocity and phase. As in \cite{paredes2014coherent},
the droplets can simply coalesce into one. The collision can also result
in rarefaction pulses of different energies and speeds.
For low velocities, part of the energy can bounce back evolving into
a smaller bright soliton.
In all cases, surface and bulk sound waves are excited during the process.
If the collision is very violent, the large
soliton can be severely distorted, ceasing to be a liquid-like approximately homogeneous medium.
We close this section by noting that there are vortex solutions of the cubic-quintic
equation (\ref{CQeq}) of the form
$\psi=e^{i\beta z}e^{il\theta}f(r)$ with $\lim_{r\to\infty}f(r)=0$, where
$l$ is the topological charge and $\theta$ is the polar angle.
Their profiles and stability have been studied in
\cite{quiroga1997stable,towers2001stability,berezhiani2001dynamics,
malomed2002stability,michinel2004square}
and their collisional dynamics in \cite{paz2004collisional}.
We remark that the vortices that we are studying in this paper
as solutions of Eq. (\ref{EQtravel}) are different objects:
they live within the vorticity-less liquid of light and they only exist in pairs and
moving with a finite velocity.
\section{Collisions of vortex-antivortex pairs}
Let us start illustrating the interactions
by computing the head-on encounter of two vortex pairs created as
described in section III. A typical example is displayed in Fig. \ref{fig5}.
The result is an exchange in which the vortex of each pair recombines with the
antivortex of the other one (the exchange of a single vortex with a vortex-antivortex
pair was described in \cite{smirnov2015scattering} with cubic nonlinear potential.)
The solitary waves come out perpendicular to
the incoming direction.
This can be understood in terms of the flow lines of Eq. (\ref{flow}), considering that, during the
approach, each pair generates a smooth inhomogeneity in the background in which the other
one propagates \cite{mironov2013scattering}.
For instance, the antivortex on the top right (see the panels (c) and (d) of Fig. \ref{fig5})
feels the flow lines generated by the phase structure of the
vortex on the top left (see Fig. \ref{fig4})
and is pushed upwards. Conversely, the vortex in the
bottom right turns downwards because of the antivortex in the bottom left. Since these bends
tend to associate again vortex and antivortex, the propagation can continue after the exchange.
In Fig. \ref{fig5}, we have considered slightly different phases for the initial solitons in order to
show that a perfect symmetry is not needed for this process.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig5.pdf}
\end{center}
\caption{Head-on encounter of two vortex-antivortex pairs resulting in an exchange mode.
Initial conditions are defined by Eq. \ref{init} with
$-{\bf x_2}={\bf x_3}=(180,0)$, ${\bf v_2}=-{\bf v_3}=(0.2,0)$, $\phi_2=5.3$, $\phi_3=5$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=\beta_3=0.1802$.
Color code and axes are defined as in Fig. \ref{fig3}.
}
\label{fig5}
\end{figure}
Similar exchanges can happen for collisions at angles. Figure \ref{fig6} depicts an example
where the incoming excitations are perpendicular to each other.
In this case, the vortex moving downwards and the antivortex
moving leftwards attract each other and coalesce into a dark blob which can be considered
an excited version of a rarefaction pulse. It comes out heading the top right of the plot and
is finally reconverted into a (highly excited) bright soliton when it reaches the edge of the
medium.
The remaining vortex and antivortex eventually couple to each other and continue to propagate
towards the bottom left. Notice that the velocity of this pair is much lower than that of the
aforementioned rarefaction pulse, as expected from the stationary solutions
characterized in section II.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig6.pdf}
\end{center}
\caption{Encounter at a right angle of two vortex-antivortex pairs resulting in an exchange mode.
Initial conditions are defined by Eq. \ref{init} with
${\bf x_2}=(0,180)$, ${\bf x_3}=(180,0)$, ${\bf v_2}=(0,-0.2)$, ${\bf v_3}=(-0.2,0)$, $\phi_2=\phi_3=5$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=\beta_3=0.1802$.
The color code is defined as in Fig. \ref{fig2}.
The range of the axes is $x,y \in [-270,270]$.
}
\label{fig6}
\end{figure}
The simulation of Fig. \ref{fig6} is also interesting because it shows other generic features
of the dynamics, which can be better appreciated in the animation presented in the
supplemental material \cite{suppl}. In particular, we must emphasize that the evolution of the
dark excitations is not elastic, in the sense that some energy is radiated away in the form
of sound waves. Moreover, faint rarefaction pulses, of small energy regarding Eq. (\ref{pE2}),
can be generated. These radiation processes take place during collisions and also during the
relaxation of the coherently generated dark bubbles towards their stationary
vortex pair form.
We also remark that, for encounters like that of Fig. \ref{fig6}, small changes in the initial
conditions can determine how the dark regions combine and greatly affect the outcoming
pulses. For instance, if we just change ${\bf v_3}$ from (-0.2,0) to (-0.21,0), therefore breaking
the symmetry, between both incoming vortex pairs,
the one moving horizontally arrives first. Instead of performing a exchange with the other vortex,
it merges with the antivortex, creating an elongated void of net vorticity -1. This snake-like structure
starts rotating and eventually decays emitting a rarefaction pulse. We present this evolution
in \cite{suppl}. Thus, the encounter gives rise to a vortex-antivortex pair and a rarefaction pulse, just
as in Fig. \ref{fig6}, but their resulting propagation directions are rather different.
This simulation also shows that, when there is an eventual dark-bright reconversion, the outgoing
dark soliton does not necessarily come out with the same propagation direction as the dark
blob which generates it.
We close this section by considering a head-on encounter in which the vortices of each pair
meet each other (instead of heading an antivortex as in Fig. \ref{fig5}). This can be accomplished
by slightly shifting the $y$-position of the bright solitons defined in the initial conditions.
An example is depicted in Fig. \ref{fig7}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig7.pdf}
\end{center}
\caption{Encounter of two vortex-antivortex pairs, shifted with respect to each
other along the direction transverse to propagation, resulting in pseudo-elastic
scattering.
Initial conditions are defined by Eq. (\ref{init}) with
${\bf x_2}=(-180,-10)$, ${\bf x_3}=(180,10)$, ${\bf v_2}=(0.2,0)$, ${\bf v_3}=(-0.2,0)$, $\phi_2=\phi_3=5$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=\beta_3=0.1802$.
Color code and axes range are defined as in Fig. \ref{fig3}.
}
\label{fig7}
\end{figure}
This evolution can be qualitatively understood noting that
the vortices repel each other and therefore are slowed down while the
antivortices continue advancing. This induces a rotation of the whole dark structure, which eventually
breaks down resulting in two separate pulses which come out at an angle,
different from the incoming one.
This is a kind of pseudo-elastic collision.
Notice, however, that the scattered pulses cannot be neatly considered vortex-antivortex as the
incoming ones. Vortex pairs and rarefaction pulses can be cleanly defined for stationary situations but
in dynamical evolutions like the present one, the separation between both is not obvious
and they can even transform into each other, as noticed in \cite{smirnov2015scattering}
in a different but somewhat related scenario.
\section{Collision of a rarefaction pulse with a vortex-antivortex pair}
We now consider the encounter of a rarefaction pulse with a vortex-antivortex pair.
An illustrative case is sketched in Fig. \ref{fig8}. In the example, the dark regions moving in opposite
directions pass near each other but do not experience a direct contact. They keep their distinct
identities during the whole evolution and therefore this process is very similar to a elastic scattering.
The pulses continue their propagation away from each other and therefore we call this a flyby
mode, following \cite{smirnov2015scattering}.
In the figure, it can be appreciated that the propagation of the rarefaction pulse is rotated by a small
angle when both waves meet (the horizontal dashed line has been included in the plots to guide the
eye). Again, this is due to the flow lines defined in Eq. (\ref{flow}) and
represented in Fig. \ref{fig2}, whose structure explains why the caviton turns upwards.
The vortex-antivortex pair is also affected by the encounter, by since its energy and momentum
(\ref{pE2}) is quite larger that that of the rarefaction pulse, it is much harder to appreciate the
diversion. Notice that this flyby mode is only relevant for a narrow window of
the scattering impact parameter. If the caviton pulse moves far from the dipolar structure, the
phase gradients are tiny and their effect is negligible. On the other hand, if both waves are too near,
the dark regions recombine giving rise to more complicated evolutions, as we show in the next example.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig8.pdf}
\end{center}
\caption{(Color online) Flyby encounter of a rarefaction pulse with a vortex-antivortex pair.
The pulse trajectory is modified because of the flows generated by the vortex-antivortex phase
structure. The horizontal dashed line marks $y=0$, the path that the rarefaction pulse would
follow in the absence of other excitations.
Initial conditions are defined by Eq. (\ref{init}) with
${\bf x_2}=(-180,27)$, ${\bf x_3}=(240,0)$, ${\bf v_2}=(0.2,0)$, ${\bf v_3}=(-0.5,0)$, $\phi_2=4.9$,
$\phi_3=4$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=0.1802$
and $\beta_3=0.15$.
Color code and axes range are defined as in Fig. \ref{fig3}.
}
\label{fig8}
\end{figure}
The initial conditons in Fig. \ref{fig9} resemble those of Fig. \ref{fig8}, but the
initial $y$-displacement of the bright solitons is slightly smaller, yielding
a smaller impact parameter for the collision of the dark waves.
In this case, the dark regions associated to the antivortex and the rarefaction pulse
come into contact and merge, initially giving rise to a large blob of vorticity -1.
Since the vortex-antivortex pair has the larger momentum
and energy, the subsequent evolution can be roughly described as an absorption of the
rarefaction pulse by the pair, which becomes highly excited, but continues its propagation rigthwards.
This structure slowly relaxes towards the stationary vortex-antivortex solution by the emission of
sound waves and faint rarefaction pulses \cite{suppl}.
In \cite{suppl}, we also present a simulation in which the vortex pair and the caviton approach each
other with zero impact parameter. Roughly, the dynamics can be understood in terms of the
previous discussion: when the dark regions touch each other, the rarefaction pulse is swallowed by
the vortex-antivortex which, albeit excited, continues its propagation.
We have checked that this kind of
qualitative
behavior is quite generic, regardless of the incoming angles and velocities.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig9.pdf}
\end{center}
\caption{(Color online)
Inelastic collision of a rarefaction pulse and a vortex-antivortex pair.
The rarefaction pulse touches the antivortex, blends with it and, eventually, also gets
connected to the dark region around the vortex. The resulting structure can
be considered as a highly excited vortex-antivortex pair which continues to propagate within
the liquid of light.
Initial conditions are defined by Eq. (\ref{init}) with
${\bf x_2}=(-180,18)$, ${\bf x_3}=(240,0)$, ${\bf v_2}=(0.2,0)$, ${\bf v_3}=(-0.5,0)$, $\phi_2=4.9$,
$\phi_3=5$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=0.1802$
and $\beta_3=0.15$.
Color code and axes range are defined as in Fig. \ref{fig3}.
}
\label{fig9}
\end{figure}
\section{Collisions of rarefaction pulses}
Let us now discuss the case of two interacting rarefaction pulses.
First of all, we notice the existence of flyby modes, similar to those described in the previous section,
when the impact parameter is not too large but enough to avoid direct contact.
It is also worth commenting on the
dynamics of head-on collisions.
The most common result is illustrated in Fig. \ref{fig10}. When the pulses meet, a larger dark blob is
created with, possibly, a bright spot inside (see panel (d) of Fig. \ref{fig10}.)
Then, two rarefaction pulses appear again and continue their propagation. During the encounter,
part of the energy is radiated away and, therefore, the pulses after the collision are slightly
fainter and faster. Thus, in this respect, the rarefaction pulses behave as dark quasi-solitons.
We remark that this happens for symmetric encounters as the one of the figure or asymmetric
ones with pulses of different energies.
As expected, when the cavitons reach the edge of the large soliton, they can be reconverted in
bright solitons again. In fact, the simulation of Fig. \ref{fig10} can be interpreted as a
bright-dark-bright-dark-bright transformation of the propagating excitation \cite{suppl}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig10.pdf}
\end{center}
\caption{(Color online)
Symmetric head-on encounter of two rarefaction pulses which cross each other, losing a fraction of their
energy in the process.
Initial conditions are defined by Eq. (\ref{init}) with
$-{\bf x_2}={\bf x_3}=(180,0)$, ${\bf v_2}=-{\bf v_3}=(0.5,0)$, $\phi_2=\phi_3=6$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=\beta_3=0.15$.
Color code and axes range are defined as in Fig. \ref{fig3}.
}
\label{fig10}
\end{figure}
Curiously, the picture changes completely if the initial conditions are properly fine tuned.
Figure \ref{fig11} depicts an example in which the rarefaction pulses annihilate each other and
their energy is radiated in the form of a circular sound wave.
Visibly, the behavior of the rarefaction pulses in this case totally differs from that of form
preserving solitons. As a matter of fact, the seemingly antagonistic character of Figs.
\ref{fig10} and \ref{fig11} can be continuously connected, by noticing
that in all head-on encounters the outgoing energy is shared by a bulk wave and two rarefaction pulses.
In Fig. \ref{fig10}, most of the energy goes to the latter whereas in Fig. \ref{fig11} it is mostly acquired by
the former, while other initial conditions lead to intermediate possibilities.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig11.pdf}
\end{center}
\caption{(Color online)
Symmetric head-on encounter of two rarefaction pulses resulting
in a fully inelastic collision. The pulses annihilate each other and yield all their energy to a
circular sound wave.
Initial conditions are defined by Eq. (\ref{init}) with
$-{\bf x_2}={\bf x_3}=(180,0)$, ${\bf v_2}=-{\bf v_3}=(0.5,0)$, $\phi_2=\phi_3=4.8$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=\beta_3=0.15$.
Color code and axes range are defined as in Fig. \ref{fig3}.
}
\label{fig11}
\end{figure}
Finally, we comment on the encounter of rarefaction pulses at angles. Figure \ref{fig12}
illustrates this case by considering a perpendicular concurrence.
As in the previous cases, the dark regions combine producing a dark blob, which is larger than the
incoming ones. However, in this case this blob can survive and, in a loose sense, propagate in the direction
required by momentum conservation.
Thus, the simulation of Fig. \ref{fig12} can be neatly portrayed as the merging of two rarefaction pulses
into a more energetic one.
Similarly to all of the presented examples, part of the energy is radiated away during the process.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig12.pdf}
\end{center}
\caption{(Color online)
Two rarefaction pulses collide perpendicularly and merge.
Initial conditions are defined by Eq. (\ref{init}) with
${\bf x_2}=(0,180),{\bf x_3}=(180,0)$, ${\bf v_2}=(0,-0.5)$, $ {\bf v_3}=(-0.5,0)$, $\phi_2=\phi_3=5$.
The large soliton is the one with $\beta_1=0.1856$ and the smaller ones have $\beta_2=\beta_3=0.15$.
The color code is defined as in Fig. \ref{fig2}.
The range of the axes is $x,y \in [-270,270]$.
}
\label{fig12}
\end{figure}
We close the section by noticing that there is a second typical qualitative behavior, which we show
in the last animation of \cite{suppl}. What happens there is that the dark blob splits giving
rise again to two rarefaction pulses (we emphasize that, even if in \cite{suppl} it may seem that
the dark pulses coming out of the collision
propagate almost in parallel, they are not vortex and antivortex).
Roughly, this last possibility can be thought of as another example of quasi-elastic scattering or
as a bounce of the pulses against each other.
\section{Summary and outlook}
In this work, we have numerically analyzed Eq. (\ref{CQeq}), reporting on a number of novel
qualitative phenomena for the cubic-quintic model in 1+2 dimensions.
The interplay of diffraction with focusing and defocusing nonlinear effects endows the cubic-quintic
nonlinear Schr\"odinger equation with an extremely rich phenomenology. In particular, there are
dark traveling waves and
bright solitons, which for large powers become liquid-like.
Noticeably, the dark and bright stationary and stable solitary waves can transform into each other
during evolution. In particular, a bright soliton can excite a rarefaction pulse when it meets
a bright soliton of larger power \cite{paredes2014coherent}. We have shown that a vortex-antivortex
pair can be generated in a similar way. The process, however, is not as clean as in the
previous case. The incoming soliton has to be larger and gives rise to a more pronounced distortion
of the flat-top soliton. Moreover, the vortex and antivortex are not generated directly, but only
as the end result of a snake instability of an initial dark blob. Thus, the radiation of part of the excess
energy is essential in approaching the stationary vortex-antivortex solution.
When a strong enough rarefaction pulse reaches the border of the liquid of light, it typically generates an outgoing
bright soliton. On the other hand, the vortex-antivortex pair excites a
couple of
surface waves propagating in opposite directions.
The possibility of creating the dark states by interference and nonlinear evolution has allowed us to
propose numerical experiments concerning their scattering with initial conditions which only
include bright solitons, see Eq. (\ref{init}).
We have made a qualitative analysis of the encounters between vortex-antivortex pairs and rarefaction
pulses. In brief, our results can be summarized as follows:
\begin{itemize}
\item If the vortex of a pair meets an antivortex of another pair and viceversa, they tend to get
exchanged resulting in two new pairs with different propagation directions.
\item If the impact parameter of a collision is large enough and the dark regions do not touch each
other, there are elastic flyby modes and the propagation direction of each wave is altered because of the flow
lines associated to the opposite structure.
\item When a vortex or a rarefaction pulse touches a vortex-antivortex pair, an excited dark blob is created.
It propagates for a while and eventually decays approaching the stationary states. The end result is strongly
dependent on initial conditions.
\item Rarefaction pulses which collide head-on typically cross each other, losing some energy by
radiating sound waves. In particular situations, the radiation can take most of the energy. If the
pulses collide at an angle, they can merge into a larger rarefaction pulse or scatter quasielastically.
\end{itemize}
This list does not exhaust the possibilities but it certainly provides a qualitative description for most
of the collisions between dark traveling waves.
It is tempting to interpret the traveling waves as quasiparticles and to try to understand collisions
in terms of their energy-momentum conservation, Eq. (\ref{pE2}).
Implicitly, this has been our point of view when using the words ``elastic'' and
``inelastic''.
Notice that $p$ and $E$ as a whole are conserved in a collision. Nevertheless, if we only take into account
the dark traveling waves, the conservation breaks down, as it obvious from figure \ref{fig10}.
The main reason is that sound waves take a sizable fraction of energy and momentum in many
processes. Moreover, as we have already emphasized, the dark waves typically appear in
excited form and therefore the velocity-momentum and
dispersion relations that can be deduced from the stationary
solutions only apply approximately. Excited dark states have complicated
dynamics and cannot always be easily identified with their stationary counterparts.
Thus, the quasiparticle interpretation is illustrative but it should be clear that
it is just a qualitative
rough description.
Our results open some interesting possibilities. First of all, it would be nice to realize the
described phenomena in optical setups along the lines of
\cite{falcao2013robust,wu2013cubic,wu2015solitons}. It would also be desirable to study similar effects
in two dimensions for the cubic defocusing nonlinearity, since it is relevant for Bose-Einstein experiments
like \cite{proud2016jones}, see also \cite{verma2016snake} and references therein.
Moreover,
it would be worth considering the three dimensional cubic-quintic case,
which supports top-flat stable spatiotemporal solitons
\cite{desyatnikov2000three,jovanoski2001light} and vortices
\cite{desyatnikov2000three,mihalache2002stable}.
Their collsional dynamics has been analyzed in
\cite{hong2008energy,adhikari2016elastic}
but the dynamics of dark traveling waves has not been described yet.
Using the cubic defocusing Schr\"odinger equation,
interesting dynamical
analysis of the interplay of rarefaction pulses, vortex rings and vortex lines in 1+3 dimensions
have been presented in the context of Bose-Einstein condensates
\cite{berloff2002evolution,berloff2004interactions,komineas2005collisions}
and superfluids
\cite{caplan2014scattering}. It would be desirable to make contact with these analyses
in the cubic-quintic case.
Finally, we remark that our setup has partial
similarities with other physical systems as, {\it e.g.}, the scattering by impurities in superfluids as
recently modeled in \cite{pshenichnyuk2016inelastic}. It could be worth exploring analogies between
different frameworks.
\acknowledgments
We thank David N\'ovoa and Jos\'e Ram\'on Salgueiro for
useful comments.
This work is supported by grants FIS2014-58117-P
from Ministerio de Econom\'\i a y Competitividad and grants GPC2015/019
and EM2013/002 from Xunta de Galicia.
The work of D. F. is supported by the FPU Ph.D. program
|
\section{Introduction}
The time evolution of non-equilibrium states in many-particle quantum systems continues to stay as an open issue. As a quantum system, we mean here an interacting and closed quantum system with infinitely many degrees of freedom, in the absence of decoherence or dissipation effects. In the thermodynamic limit, one can argue that parts of the system act as `reservoirs' for other parts, to assure at long times, the relaxation to equilibrium of arbitrary (non-equilibrium) initial states. This property is recognize to be a consequence of \emph{quantum ergodicity}, a concept that was addressed by von Neumann in 1929 \cite{vonneumann}, and is most relevant for the foundations of Statistical Mechanics. In order to better understand the above assumption, one can test non-equilibrium properties in models that are exactly solvable, and eventually find evidence of non-ergodic behavior~\cite{noneq}. Spin chains are considered as useful examples of non trivial quantum interacting systems that may fulfill the above requirements. The relaxation of inhomogeneous initial states has been studied in a number of articles \cite{berimJETP,berimLowT,berimPhysA,berimPRB,karevski,anomalous}, using spin chains with cyclic boundary conditions as prototype systems. In the thermodynamic limit, the relaxation mechanism is driven exclusively by quantum fluctuations and interference effects. A slow relaxation is obtained from contributions of stationary points, where the interference is constructive. Degenerate stationary points present anomalous behavior, reducing even more the damping of the initial excitations as a function of the parameters of the model. But in spite of being very slow, the relaxation process leads those systems to a final homogeneous and stationary state at asymptotically very long times \cite{berimPRB}, a situation that we call \emph{equilibrium state}, in agreement with the ergodic hypothesis. All the systems referred above are amenable of exact analytical solutions, they all have full translational symmetry and are solved with periodic boundary conditions. In the present contribution, we also study the dynamics of a spin chain, but we have broken the translational symmetry, considering a semi-infinite system with an open boundary, where we have attached an impurity. The impurity is characterized by two parameters, its magnetic moment and the coupling with the chain, which may be different from those in the bulk of the system. In the thermodynamic limit (semi-infinite chain), the spectrum has a continuous branch, that we call `the band', and depending on values of the impurity parameters, localized states may be split off from the band. This latter fact is determinant for the phenomena we are dealing with here, that is the preparation of an inhomogeneous (out-of-equilibrium) state that does not relax to equilibrium at asymptotically very long times.
For definiteness, in this work we study the quantum dynamics of a semi-infinite \emph{XY} spin chain with an impurity. This problem can be solved exactly, using the diagonalization method of Ref.~\cite{tsukernik}. In the above approach, the Jordan-Wigner transformation maps the spin Hamiltonian into a free-fermion problem, which can be solved in closed form through a unitary transformation. Due to the absence of translational symmetry, one has to rely in calculations performed in the site representation (real space). Assuming a quite general inhomogeneous initial state for the spin chain, exact results are obtained for the long-time dependence of the magnetization at the different sites of the chain. The stationary phase method has been employed to estimate this long-time behavior. Non-ergodic time evolution is observed, when two localized impurity levels are split off from the band (one above and one below the band edges). For an arbitrary inhomogeneous initial state, the time evolution displays strong oscillations in the site magnetizations. Most of the oscillations are quickly damped by destructive interference, stationary point contributions relax more slowly, as in the examples previously mentioned in Ref.~\cite{berimPhysA,berimPRB,anomalous}, and at asymptotically very long times it only remains an undamped oscillation, with a frequency that is identified as the Rabi frequency of the two localized levels. If the inhomogeneity of the initial state is chosen near the open boundary, the amplitude of the oscillation is paramount at the first sites of the chain. It decreases in magnitude when going to the bulk of the sample, but it is not damped in time.
The phenomenon reported here is similar to the one observed experimentally in $1-D$ Bose gases confined by light traps \cite{cradle}. After the preparation of out-of-equilibrium initial states, non-ergodic behavior is observed, with the absence of damping after thousands of atom collisions. Concerning spin systems, proper experimental setups are nowadays available. Thanks to the progress of experimental techniques, it is possible to create quantum spin chains~\cite{brune}, for both, ferromagnetic and antiferromagnetic couplings ~\cite{gambardella1,hirjibehedin}. Devices which are candidates for building a quantum computer, such as Josephson junction arrays, mimic the physics of quantum spin chains~\cite{makhlin}. Inhomogeneous initial excitations can be produced experimentally in those systems, by locally manipulating gate voltages, tunneling barriers or magnetic fields.
In the next Section of this paper, we discuss the model used in the calculation, along with the method of solution. After this, we schematically describe the time evolution for a particular (but typical) inhomogeneous initial state. Results for the long-time behavior of the magnetization are displayed in the last Section, where we state the final conclusions.
\section{The model}
We consider a semi-infinite quantum spin chain with an impurity at one end of the chain, as depicted in Fig.~\ref{chain}.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{chain}
\caption{(Color online) Semi-infinite quantum spin chain with an impurity `atom' at site \(n=0\). A uniform external magnetic field $H$ is applied along the \(z\)-axis .}
\label{chain}
\end{figure}
The figure is self-explanatory, $\mu$ is the magnetic moment in the bulk of the chain, where spins are coupled with exchange $J$. The corresponding quantities for the impurity are $(\mu',J')$. For convenience, to follow the solution of Ref.~\cite{tsukernik}, we define the parameters:
\begin{equation}
\eta=J'/J\qquad\Delta=-\left(\mu'-\mu\right)H/J~.
\end{equation}
We model our spin chain using a \emph{XY}-Hamiltonian for spin $s=1/2$ with nearest neighbor interactions in the presence of a uniform transverse magnetic field $H$:
\begin{equation}
\begin{aligned}
\mathcal H\!=\!&-J'\!\!\left(S_0^{\mathstrut x} S_1^{\mathstrut x}\!+\!S_0^{\mathstrut y} S_1^{\mathstrut y}\right)\!-\!J\!\!\sum_{n=1}^{N-1}\!\!\left(S_{n\vphantom1}^{\mathstrut x} S_{n+1}^{\mathstrut x}\!+\!S_{n\vphantom1}^{\mathstrut y} S_{n+1}^{\mathstrut y}\right)\\
&-\mu' H S_0^{\mathstrut z}-\mu H\sum_{n=1}^N S_{n\vphantom1}^{\mathstrut z},
\end{aligned}
\label{ham}
\end{equation}
and we consider the case of a semi-infinite spin chain (\(N\to\infty\)) for which the hamiltonian \eqref{ham} can be exactly diagonalized~\cite{tsukernik}. The infinite number of degrees of freedom avoids revivals of the initial quantum state, which may be present in finite size systems~\cite{revival}. For details of the solution, we refer the reader to the paper~\cite{tsukernik}. For completeness, we briefly summarize the procedure. We firstly use the Jordan-Wigner transformation, to express spin-1/2 operators in terms of fermion annihilation and creation operators \(c_n^{\mathstrut}\) and \(c_n^\dag\):
\begin{equation}
S_0^+=c_0^\dag,\qquad S_n^+=c_n^\dag\prod_{l=0}^{n-1}\left(1-2c_{l}^\dag c_{l}\right),\quad n\geqslant1,
\label{jw}
\end{equation}
where $S_n^+$ are the spin ladder operators. The resulting fermion Hamiltonian can be diagonalized by a unitary transformation that preserves the anti-commuting properties:
\begin{equation}
c_n^{\mathstrut}=\sum_{\lambda=1} u_{n\vphantom1}^{(\lambda)} b_\lambda^{\mathstrut},\quad n\geqslant0,
\label{transfunit}
\end{equation}
where the coefficients \(u_{n\vphantom1}^{(\lambda)}\) (wave functions) are derived from the Heisenberg equations of motion for the quasiparticle operators $b_\lambda$, proviso that the Hamiltonian acquires the quasi-particle form:
\begin{equation}
\mathcal H=\sum_{\lambda=1} \epsilon_\lambda~ b_\lambda^\dag\ b_\lambda ,
\end{equation}
with $\lambda$ being the `good quantum number' characterizing the energy spectrum. As in any eigenvalue problem, the equations of motion lead to a set of coupled linear equations in finite difference for the wave functions $u_{n}^{\lambda}$ (which is infinite in the thermodynamics limit). This set of equations is solved by recurrence methods, yielding the eigenvalues ${\epsilon_\lambda}$ as well as the corresponding wave functions $u_{n}^{\lambda}$~\cite{tsukernik}. As a general rule, one gets a continuous spectrum corresponding to the bulk of the chain. Depending on values of the impurity parameters, one can split off two, one, or none localized levels from the band. The different regions of solution in the impurity parameter space are displayed in Fig.~\ref{regs}, with coordinates $\Delta$ and $\eta^2$.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{regs}
\caption{Characteristic regions for the presence of localized levels in the spectrum. See explanations in the main text.}
\label{regs}
\end{figure}
The symbols \emph{(i)} and\emph{(iii)} are used to denote borders, and \emph{(ii)} is the crossing at $\Delta=0,~\eta^2=2$. The four characteristic regions of solutions are labeled by \emph{(iv)}, \emph{(v)}, with $\Delta>0$, \emph{(v)}, with $\Delta<0$, and \emph{(vi)}. There are no bound states in region \emph{(iv)}, while there is one bound level in regions \emph{(v)}; with energy lower than the band, for $\Delta<0$, and higher than the band for $\Delta>0$. In region \emph{(vi)} there are two localized levels, one above and one below the continuous spectrum.
With this final remark, we end the description of the exact solution obtained in Ref.~\cite{tsukernik}. With the above solution in hand, we turn to the calculations of the dynamical properties of the chain, when an inhomogeneous state is prepared as initial state. The detailed calculations for the four regions, the borders and the crossing will be published in a separate contribution~\cite{next}. Here, we will focus in a few relevant examples.
\section{Time evolution}
We assume an initial state for the chain in which each `atom' is in one of the two eigenstates of the \(z\)-component of spin, \(\left|1/2\right\rangle\) (spin up) or \(\left|-1/2\right\rangle\) (spin down). The system initial state then reads
\begin{equation}
\left|\Psi(0)\right\rangle=\left|m_z^{(0)}(0), m_z^{(1)}(0), \dots, m_z^{(N)}(0)\right\rangle,
\label{estin}
\end{equation}
with
\begin{equation}
m_z^{(n)}(0)=\pm1/2,\quad n\geqslant0.
\label{valsestin}
\end{equation}
The above is an eigenstate of $S^z$, the $z$-component of the total spin, but it is not an eigenstate of Hamiltonian (\ref{ham}), so it displays a non trivial dynamics. To describe the latter, we calculate the \(z\)-axis magnetization for each site \(n\) at an arbitrary instant of time \(t\). Denoting this quantity by \(\left\langle S_{n\vphantom1}^{\mathstrut z}\right\rangle_{\!\!t}\), we have
\begin{equation}
\left\langle S_{n\vphantom1}^{\mathstrut z}\right\rangle_{\!\!t}=\left\langle\Psi(t)\,\middle|\,S_{n\vphantom1}^{\mathstrut z}\,\middle|\,\Psi(t)\right\rangle,
\label{defmagn}
\end{equation}
where \(\left|\Psi(t)\right\rangle=e^{-\frac i\hbar\mathcal H t}\left|\Psi(0)\right\rangle\) is the system state at time \(t\). The Schr\"{o}dinger expectation value $\left\langle S_{n\vphantom1}^{\mathstrut z}\right\rangle$ can be written in the Heisenberg picture, with
\begin{equation}
\left\langle S_{n\vphantom1}^{\mathstrut z}\right\rangle_{\!\!t}=\left\langle S_{n\vphantom1}^{\mathstrut z}(t)\right\rangle_{\!\!0},\quad S_{n\vphantom1}^{\mathstrut z}(t)=e^{\frac i\hbar\mathcal H t}\,S_{n\vphantom1}^{\mathstrut z}\,e^{-\frac i\hbar\mathcal H t}~,
\label{defmagnestin}
\end{equation}
where the average in the latter is taken with the initial state.
The Heisenberg operator \(S_{n\vphantom1}^{\mathstrut z}(t)\) can be expressed in terms of the quasiparticle operators \(b_\lambda^{\mathstrut}\) and \(b_\lambda^\dag\), whose trivial time evolution
\begin{equation}
b_\lambda^{\mathstrut}(t)=b_\lambda^{\mathstrut} e^{-\frac i\hbar\epsilon_\lambda^{\mathstrut} t},\quad b_\lambda^\dag(t)=b_\lambda^\dag e^{\frac i\hbar\epsilon_\lambda^{\mathstrut} t},
\label{bsHeins}
\end{equation}
allows us to obtain the explicit time dependence of the magnetization. Subsequently, using the result for arbitrary \(t\), we address the question of obtaining the magnetization for very large values of time, i.e., the long-time tails. In the thermodynamic limit ($N\rightarrow\infty$), this is done using the stationary phase method~\cite{SPM}, to estimate the dominant behavior at asymptotically long times. Once the initial state is given in the form (\ref{estin}), the site magnetization can be written as\footnote{Several equivalent exact expressions for the site-magnetization can be obtained~\cite{next}, and they are more or less convenient, depending on the structure of the initial state. Relation (\ref{3rdresger}) given in the text is the most compact one.}:
\begin{equation}
\left\langle S_{n\vphantom1}^{\mathstrut z}\right\rangle_{\!\!t}=\sum_{m=0}^Nm_z^{(m)}(0)S_{n,m}(t),\quad n\geqslant0~,
\label{3rdresger}
\end{equation}
where all the interference effects are contained in the factor $S_{n,m}(t)$, which is written as
\begin{equation}
S_{n,m}(t)=\left|\sum_{\lambda=1}~u_{n\vphantom1}^{(\lambda)} {u_{m\vphantom1}^{(\lambda)}}^*e^{-\frac i\hbar\epsilon_\lambda^{\mathstrut} t}\right|^2~,
\label{Snm}
\end{equation}
with \{$u_{m}^{\lambda}$\} being the wave functions obtained through the diagonalization process (see Eq. (\ref{transfunit})). In the general case, the sum (\ref{Snm}) includes the continuous as well as the discrete spectra. Exact asymptotic series can be obtained for the long-time behavior of $S_{n,m}(t)$. In general, dominant terms of the series come from contributions of stationary points and they show a very slow relaxation, in the form of a power law $t^{-\alpha}$, with $\alpha\geq1$~\cite{next}. There are remarkable exceptions that we want to discuss here. To better illustrate the above phenomena, we will choose the specific initial state given below:
\begin{equation}
\left|\Psi(0)\right\rangle=\left|\uparrow\right\rangle_{\!0}\left|\downarrow\right\rangle_{\!1}\left|\uparrow\right\rangle_{\!2}\left|\uparrow\right\rangle_{\!3}\dots\left|\uparrow\right\rangle_{\!N}~,
\label{specific}
\end{equation}
where all the spins are pointing up, except the second which is coupled to the impurity and pointing down. For this state, the site-magnetization acquires the very simple form
\begin{equation}
\left\langle S_{n\vphantom1}^{\mathstrut z}\right\rangle_{\!\!t}=\frac12-S_{n,1}(t),\quad n\geqslant0~.
\label{magnestinesp}
\end{equation}
Results for the time evolution of this state are shown in the next Section, where we also give the final comments.
\section{Results and discussion}
We will present results and figures corresponding to time evolution for values of parameters in regions (\emph{iv}), (\emph{v}) and (\emph{vi}) of Fig.~\ref{regs}, when the initial state is prepared in the form given by (\ref{specific}).
\subsection{The band}
For region (\emph{iv}), there are no bound states. The relaxation is driven by states within the continuum, where the interference is in general, destructive. There are two stationary points, which correspond to the edges of the band, which yield a slower evolution in the form of a damped oscillation that relaxes as $ 1/t^3$. This relaxation is due to constructive interference of states in the neighborhood of stationary points. The frequency of the oscillation is the band width, \emph{i.e.} $\omega=2J/\hbar$. In Fig.~\ref{plotreg4n0123}, we display the long-time evolution, for the first sites of the chain ($n=0,1,2,3$).
\begin{figure}[h!]
\includegraphics[width=\linewidth]{plotreg4n0123}
\caption{(Color online) Long-time behavior of the magnetization at sites $n=0,1,2,3$, for parameters in region (\emph{iv}) of Fig.~\ref{regs}. The examples shown correspond to the values \(\Delta=0\), \(\eta^2=1\). The unit of time is given by \(\tau=\hbar/\left|J\right|\).}
\label{plotreg4n0123}
\end{figure}
It is evident from the figure, that at asymptotically long-times, there is only one frequency and all oscillations are damped in time. The magnetization of the extrapolated final state is homogeneous and stationary.
\subsection{One impurity state}\label{one}
In region (\emph{v}), one bound state is split off from the band. There is a symmetry between the two subregions $\Delta>0$ and $\Delta<0$, where the localized state splits from above or below the band, respectively (assuming $J>0$, and if $J$ changes sign, the statement is reversed). For the dynamics, it makes no difference the sign of $J$, and the oscillatory dominant term has two different frequencies superposed. Now the time evolution of the magnetization shows a peculiar property, with oscillations decaying as $1/t^{3/2}$ to a \emph{constant value} which depends on the site. The two frequencies that appear in the long-time behavior are given by the difference between the localized level and the edges of the band:
\begin{equation}
\omega_{\pm}=\frac{\epsilon_i\pm J}{\hbar}~,
\label{frequencies}
\end{equation}
where $\epsilon_i$ is the energy of the bound state. Thus, one gets a `low' and a `high' frequency oscillation in the asymptotic behavior. An example is shown in Fig.~\ref{plotreg5n0123}:
\begin{figure}[h!]
\includegraphics[width=\linewidth]{plotreg5n0123}
\caption{(Color online) Long-time behavior of the magnetization at sites $n=0,1,2,3$, for the initial state \eqref{specific}. Values of parameters are \(\left|\Delta\right|=1/2\), \(\eta^2=2\), which correspond to region (\emph{v}). Relaxation is very slow, and due to the time scale, only the oscillation with the `high frequency' is apparent. The unit of time is given by \(\tau=\hbar/\left|J\right|\).}
\label{plotreg5n0123}
\end{figure}
The magnetization extrapolated at $t\rightarrow \infty$ is stationary, but inhomogeneous, showing a non constant profile near the impurity, at asymptotically very long times.
\subsection{Two impurity states}
Region (\emph{vi}) is the characteristic region of parameter space with the presence of two bound states, one above and one below the continuous band. The stationary phase methods yields the dominant terms of the asymptotic series in the form of oscillations with five different frequencies, two `low', two `high', and the highest which corresponds to the Rabi frequency of the two bound states. The `low' and `high' frequencies are similar to those obtained in subsection \ref{one} for the case of one impurity, except that here we have two localized levels. The Rabi frequency is $\omega_{Rabi}=|\epsilon_1-\epsilon_2|/\hbar$~, where $\epsilon_1$ and $\epsilon_2$ are the energies of the bound states. When $\Delta=0$, the levels split symmetrically from the band edges, and the number of frequencies is reduced to three. The `high' and `low' frequency oscillations are damped in time by a factor proportional to $1/t^{3/2}$, as in the previous case. What is \emph{remarkable} in this case is the fact that the Rabi oscillation is not damped in time, and the system is never stationary, nor homogeneous. Examples are shown in Fig.~\ref{plotreg6n0123}:
\begin{figure}[h!]
\includegraphics[width=\linewidth]{plotreg6n0123}
\caption{(Color online) Long-time behavior of the magnetization at sites $n=0,1,2,3$, for the initial state \eqref{specific}. Values of parameters are \(\Delta=0\), \(\eta^2=3\). Due to the time scale, only the Rabi frequency is visible. Oscillations are undamped. The unit of time is given by \(\tau=\hbar/\left|J\right|\).}
\label{plotreg6n0123}
\end{figure}
\subsection{Final discussions}
In summary, for the same initial state, the non-ergodic behavior appears by changing the impurity parameters in the quantum model. The absence of damping at asymptotically long times, occurs when two localized modes are separated from the continuum, with the impurity parameters in region (\emph{vi}), as shown in Fig.~\ref{regs}. The quantum interference between the two bound states yields magnetization oscillations with the Rabi frequency of the levels. The oscillations settle over very long times with constant amplitude. This effect may have potential applications in quantum computation. In fact, spin chains have been suggested as efficient channels for the short-range transport of information in a quantum computing device~\cite{bose}. Even in finite-size chains, one could find precursory imprints of the undamped oscillation. In addition of the systems mentioned in the Introduction, the effect can also be investigated with polar atoms or molecules trapped in optical lattices, systems that simulate the physics of spin chains. The $XY$ model can be realized by proper combinations of microwave excitations and the many-body interactions induced by the radiation~\cite{simulation}.
\section*{Acknowledgements} \label{Sec-acknowledgements}
The authors are grateful to S\~ao Paulo Research Foundation (\textbf{FAPESP}, Brazil) for financial support through the project No. 2009/53826-8.
\section*{References}
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
With the deluge of data in many modern application areas, there is pressing need for scalable computational algorithms for inference from such data, including uncertainty quantification (UQ). Somewhat surprisingly, even as the volume of data increases, uncertainty often remains sizable. Examples in which this phenomenon occurs include financial fraud detection \citep{ngai2011application}, disease mapping \citep{wakefield2007disease} and online click-through tracking \citep{wang2010click}. Bayesian approaches provide a useful paradigm for quantifying uncertainty in inferences and predictions in these and other settings.
The standard approach to Bayesian posterior computation is Markov chain Monte Carlo (MCMC) and related sampling algorithms. However, conventional MCMC algorithms often scale poorly in problem size and complexity. Due to its sequential nature, the computational cost of MCMC is the product of two factors: the evaluation cost at each sampling iteration and the total number of iterations needed to obtain an acceptably low Monte Carlo (MC) error. While a substantial literature has developed focusing on decreasing computational cost per iteration (\cite{minsker2014robust,maclaurin2014firefly,
srivastava2015wasp,conrad2015accelerating} among others), {very little has been done
to reduce the} {number of iterations needed to produce a desired MC error in posterior summaries}
{as the problem size grows.}
{A major concern in applying MCMC algorithms in big data problems is that the level of autocorrelation in the MCMC path may increase with the size of the data. Markov chains with high autocorrelation
tend to produce a low {\em effective sample size (ESS)} per unit computational time, {which is informally known as the {\em slow mixing} problem}. The ESS is designed to compare the information content in the sampling iterations relative to a gold standard Monte Carlo algorithm that collects independent samples. If the number of effective samples in 1,000 iterations is only 10, then the MCMC algorithm will need to be run 100 times as long as the gold standard algorithm to obtain the same MC error in posterior summaries. Such a scenario is not unusual in big data problems, leading MCMC algorithms to face a {\em double burden}, with the time per iteration increasing and it becoming necessary to collect more and more iterations.}
{This double burden has led many members of the machine learning community to abandon MCMC in favor of more easily scalable alternatives, such as variational approximations. Unfortunately these approaches lack theoretical guarantees and often badly under-estimate posterior uncertainty. Hence, there has been substantial interest in recent years in designing scalable MCMC algorithms. The particular focus of this paper is on a popular and broad class of Data Augmentation (DA)-MCMC algorithms. DA-MCMC algorithms are used routinely in many classes of models, with the algorithms of \cite{albert1993bayesian} for probit models and \cite{polson2013bayesian} for logistic models particularly popular. Our focus is on improving the performance of such algorithms in big data settings in which issues can arise in terms of both the time per iteration and the mixing. The former problem can be addressed using any of a broad variety of existing approaches, and we focus here on the slow
mixing problem.}
{ \cite{johndrow2016inefficiency} discovered that popular DA-MCMC algorithms have small effective sample sizes in large data settings involving imbalanced data. For example, data may be binary with a high proportion of zeros. A key insight is that the reason for this problem is a discrepancy in the rates at which Gibbs step sizes and the width of the high-probability region of the posterior converge to zero as $n$ increases. In particular, the conditional posterior given the augmented data may simply be too concentrated relative to the marginal posterior, with this problem amplified as the data sample size increases. There is a rich literature on methods for accelerating mixing in DA-MCMC algorithms using tricks ranging from reparameterization to parameter-expansion \citep{liu1999parameter,meng1999seeking,papaspiliopoulos2007general}. However, we find that such approaches fail to address the miscalibration problem and have no impact on the worsening mixing rate with increasing data sample size $n$.}
{The focus of this article is on proposing a general new class of algorithms for addressing the fundamental miscalibration problem that leads to worsening mixing of DA-MCMC with $n$. In particular, the key idea underlying our proposed class of {\em calibrated} DA (CDA) algorithms is to introduce auxiliary parameters that change the variance of full conditional distributions for one or more parameters. These auxiliary parameters can adapt with the data sample size $n$ to fundamentally address the key problem causing the worsening mixing with $n$. In general, the invariant measure of CDA-MCMC does not correspond exactly to the true joint posterior distribution of interest. Instead, we can view CDA-MCMC as representing a computationally more efficient perturbation of the original Markov chain. { The perturbation error can be eliminated using Metropolis-Hastings. Compared to other adaptive Metropolis-Hastings algorithms,
which often require carefully chosen multivariate proposals and complicated adaptation with multiple chains \citep{tran2016adaptive}, CDA-MCMC only
requires a simple modification to Gibbs sampling steps. We show the auxiliary parameters can be efficiently adapted for each type of data augmentation, via minimizing
the difference between Fisher information of conditional and
marginal distributions.}}
\section{Calibrated Data Augmentation} \label{sec:cda}
Data augmentation Gibbs samplers alternate between sampling latent data $z$ from their conditional posterior distribution given model parameters $\theta$ and observed data $y$, and sampling parameters $\theta$ given $z$ and $y$; either of these steps can be further broken down into a series of full conditional sampling steps but we focus for simplicity on algorithms of the form:
\be \label{eq:da}
z \mid \theta, y &\sim \pi(z;\theta,y) \\
\theta \mid z,y &\sim f(\theta;z,y),
\ee
where $f$ belongs to a location-scale family, such as the Gaussian. Popular data augmentation algorithms are designed so that both of these sampling steps can be conducted easily and efficiently; e.g., sampling the latent data for each subject independently and then drawing $\theta$ simultaneously (or at least in blocks) from a multivariate Gaussian or other standard distribution. This effectively avoids the need for tuning, which is a major issue for Metropolis-Hastings algorithms, particularly when $\theta$ is high-dimensional.
Data augmentation algorithms are particularly common for generalized linear models (GLMs), with $\bb E(y_i \mid x_i, \theta) = g^{-1}(x_i \theta)$ and a conditionally Gaussian prior distribution chosen for $\theta$. We focus in particular on Poisson log-linear, binomial logistic, and binomial probit as motivating examples.
Consider a Markov kernel $K((\theta,z);\cdot)$ with invariant measure $\Pi$ and update rule of the form \eqref{eq:da}, and a Markov chain $(\theta_t,z_t)$ on a state space $\Theta \times \mc Z$ evolving according to $K$. We will abuse notation in writing $\Pi(d\theta) = \int_{z \in \mc Z} \Pi(d\theta,dz)$. The lag-1 autocorrelation for a function $g : \Theta \to \bb R$ at stationarity can be expressed as the Bayesian fraction of missing information (\cite{papaspiliopoulos2007general}, \cite{rubin2004multiple}, \cite{liu1994fraction})
\be
\gamma_g &= 1- \frac{\bb E[\var(g(\theta) \mid z)]}{\var(g(\theta))}, \label{eq:missinginfo}
\ee
where the integrals in the numerator are with respect to $\Pi(d\theta,dz)$ and in the denominator with respect to $\Pi(d\theta)$. Let
\be
L_2(\Pi) = \left\{ g : \Theta \to \bb R, \int_{\theta \in \Theta} \{g(\theta)\}^2 \Pi(d\theta) < \infty \right\}
\ee
be the set of real-valued, $\Pi$ square-integrable functions. The \emph{maximal autocorrelation}
\be
\gamma = \sup_{g \in L^2(\Pi)} \gamma_g = 1- \inf_{g \in L^2(\Pi)} \frac{\bb E[\var(g(\theta) \mid z)]}{\var(g(\theta))}
\ee
is equal to the geometric convergence rate of the data augmentation Gibbs sampler (\cite{liu1994fraction}). For $g(\theta) = \theta_j$ a coordinate projection, the numerator of the last term of \eqref{eq:missinginfo} is, informally, the average squared step size for the augmentation algorithm at stationarity in direction $j$, while the denominator is the squared width of the bulk of the posterior in direction $j$. Consequently, $\gamma$ will be close to 1 whenever the average step size at stationarity is small relative to the width of the bulk of the posterior.
The purpose of CDA is to introduce additional parameters that allow us to control the step size relative to the posterior width -- roughly speaking, the ratio in \eqref{eq:missinginfo} -- with greater flexibility than reparametrization or parameter expansion. The flexibility gains are achieved by allowing the invariant measure to change as a result of the introduced parameters. The additional parameters, which we denote $(r,b)$, correspond to a collection of reparametrizations, each of which defines a proper (but distinct) likelihood $L_{r,b}(\theta;y)$, and for which there exists a Gibbs update rule of the form \eqref{eq:da}. In general, $b$ will correspond to a location parameter and $r$ a scale parameter that are tuned to increase $\bb E[\var(g(\theta) \mid z)]\{\var(g(\theta))\}^{-1}$, although the exact way in which they enter the likelihood and corresponding Gibbs update depend on the application. The reparametrization also has the property that $L_{1,0}(\theta;y) = L(\theta;y)$, the original likelihood. The resulting Gibbs sampler, which we refer to as CDA Gibbs, has $\theta$-marginal invariant measure $\Pi_{r,b}(\theta;y) \propto L_{r,b}(\theta;y) \Pi^0(\theta)$, where $\Pi^0(\theta)$ is the prior. Ultimately, we are interested in $\Pi_{1,0}(\theta;y)$, so we use CDA Gibbs as an efficient proposal for Metropolis-Hastings. That is, we propose $\theta^*$ from $Q(\theta;\cdot)$ where
\be \label{eq:Q}
Q_{r,b}(\theta;A) = \int_{(\theta^*,z) \in A \times \mc Z} \pi_{r,b}(z;\theta,y) f_{r,b}(\theta^*;z,y) dz d\theta^*
\ee
for $A \subseteq \Theta$, where $\pi_{r,b}$ and $f_{r,b}$ denote the conditional densities of $z$ and $\theta$ in the Gibbs sampler with invariant measure $\Pi_{r,b}$. By tuning working parameters during an adaptation phase to reduce the autocorrelations and increase the Metropolis-Hastings acceptance rate, we can select values of the working parameters that yield a computationally efficient algorithm. Tuning is facilitated by the fact that the M-H acceptance ratios using this proposal kernel have a convenient form, which is a nice feature of using Gibbs to generate M-H proposals.
\begin{remark} \label{rem:accrat}
The CDA M-H acceptance ratio is given by
\be
1 \wedge \frac{L(\theta';y) \Pi^0(\theta') Q_{r,b}(\theta;\theta')}{L(\theta;y) \Pi^0(\theta) Q_{r,b}(\theta';\theta)} = 1 \wedge \frac{L(\theta';y)L_{r,b}(\theta;y)}{L(\theta;y)L_{r,b}(\theta';y)} \label{eq:mh-accrat}
\ee
\end{remark}
A general strategy for tuning is given in Section \ref{sec:tuning}.
We give a basic convergence guarantee that holds for the CDA M-H under weak assumptions on $L_{r,b}$, which is based on \cite[Theorem 3, also pp. 214]{roberts1994simple}. Basically, one needs $\Pi(\cdot) \ll \Pi_{r,b}(\cdot)$ for all $r,b$, where for two probability measures $\mu,\nu$, $\mu(\cdot) \ll \nu(\cdot)$ means $\mu$ is absolutely continuous with respect to $\nu$.
\begin{remark}[Ergodicity] \label{rem:ergodic}
Assume that $\Pi(d\theta)$ and $\Pi_{r,b}(d\theta)$ have densities with
respect to Lebesgue measure on $\bb R^p$, and that \\ $K_{r,b}((\theta,z);(\theta',z'))>0 \,\forall\, ((\theta,z),(\theta',z')) \in (\Theta \times \mc Z) \times (\Theta \times \mc Z)$. Then,
\begin{itemize}
\item For fixed $r,b$, CDA Gibbs is ergodic with invariant measure $\Pi_{r,b}(d\theta,dz)$.
\item A Metropolis-Hastings algorithm with proposal kernel $Q_{r,b}(\theta';\theta)$ as defined in \eqref{eq:Q} with fixed $r,b$ is ergodic with invariant measure $\Pi(d\theta)$.
\end{itemize}
\end{remark}
Proofs are located in the Appendix.
\subsection{Initial Example: Probit with Intercept Only}
We use a simple example to illustrate CDA. Consider an intercept-only probit
\be
y_i \sim \Bern(p_i), \quad p_i = \Phi(\theta) \quad i=1,\ldots,n
\ee
and improper prior $\Pi^0(\theta) \propto 1$. The basic data augmentation algorithm \citep{tanner1987calculation,albert1993bayesian} has the update rule
\be
z_i \mid \theta, y_i &\sim \left\{ \begin{array}{cc} \No_{[0,\infty)}( \theta,1) & \text{ if } y_i = 1 \\ \No_{(-\infty,0]}( \theta,1) & \text{ if } y_i = 0 \end{array} \right. \quad i=1,\ldots,n\\
\theta \mid z, y &\sim \No\left( n^{-1} \sum_i z_i, n^{-1} \right),
\ee
where $\No_{[a,b]}(\mu,\sigma^2)$ is the normal distribution with mean $\mu$ and variance $\sigma^2$ truncated to the interval $[a,b]$. \cite{johndrow2016inefficiency} show that when $\sum_i y_i = 1$, $\var(\theta_t \mid \theta_{t-1})$ is approximately $n^{-1} \log n$, while the width of the high probability region of the posterior is order $(\log n)^{-1}$, leading to slow mixing.
As the conditional variance $\mbox{var}(\theta\mid z,y)$ is independent of $z$, we introduce a scale parameter $r$ in the update for $z$, then adjust the conditional mean by a location parameter $b$. This is equivalent to changing the scale of $z_i \mid\theta,y_i$ from $1$ to $r$ and the mean from $\theta$ to $\theta+b$. These adjustments yield
\be
\mbox{pr}(y_i = 1 | \theta, r, b) =& \int_{0}^{\infty} \frac{1}{\sqrt{2 \pi r} } \exp\left(-\frac{(z_i-\theta-b)^2}{2 r^2} \right) dz_i
\\ = & \Phi\bigg( \frac{\theta+b}{\sqrt{r}}\bigg),
\label{eq:prop-marginal-probit-intercept}
\ee
leading to the modified data augmentation algorithm
\be \label{eq:cda-probit-intercept}
z_i \mid \theta, y_i &\sim \left\{ \begin{array}{cc} \No_{[0,\infty)}( \theta+b,r) & \text{ if } y_i = 1 \\ \No_{(-\infty,0]}( \theta+b,r) & \text{ if } y_i = 0 \end{array} \right. \quad i=1,\ldots,n\\
\theta \mid z,y &\sim \No\left(n^{-1} \sum_i(z_i-b), n^{-1} r \right).
\ee
To achieve step sizes consistent with the width of the high posterior probability region, we need $n^{-1} r \approx (\log n)^{-1}$, so $r \approx n/\log n$. To preserve the original target, we use (\ref{eq:cda-probit-intercept}) to generate an M-H proposal $\theta^*$. By Remark \ref{rem:accrat}, the M-H acceptance probability is given by \eqref{eq:mh-accrat} with $L_{r,b}(\theta;y_i) = \Phi\big( ({\theta+b}){r}^{-1/2}\big) ^{y_i} \Phi\big( - ({\theta+b}){r}^{-1/2}\big)^{(1-y_i)}$ and $L(\theta;y_i) = L_{1,0}(\theta;y_i)$. Setting $r_i=1$ and $b_i=0$ leads to acceptance rate of $1$, which corresponds to the original Gibbs sampler.
To illustrate, we consider $\sum_i y_i =1$ and $n=10^4$. Letting $r = n/\log n$, we then choose the $b_i$'s to increase the acceptance rate in the M-H step. In this simple example, it is easy to compute a ``good'' value of $b_i$, since $b_i = -3.7 (\sqrt r -1)$ results in $\mbox{pr}(y_i = 1) = \Phi(-3.7) = n^{-1}\sum_i y_i \approx 10^{-4}$ in the proposal distribution, centering the proposals near the MLE for $p_i$.
We perform computation for these data with different values of $r$ ranging from $r=1$ to $r=5,000$, with $r=1,000 \approx n/\log n$ corresponding to the theoretically optimal value. Figure~\ref{probit_demo_intercept_proposal} plots autocorrelation functions (ACFs) for these different samplers without M-H adjustment. Autocorrelation is very high even at lag 40 for $r=1$, while increasing $r$ leads to dramatic improvements in mixing. There are no further gains in increasing $r$ from the theoretically optimal value of $r=1,000$ to $r=5,000$. Figure~\ref{probit_demo_intercept_density} shows kernel-smoothed density estimates of the posterior of $\theta$ without M-H adjustment for different values of $r$ and based on long chains to minimize the impact of Monte Carlo error; the posteriors are all centered on the same values but with variance increasing somewhat with $r$. With M-H adjustment such differences are removed; the M-H step has acceptance probability close to one for $r=10$ and $r=100$, about 0.6 for $r=1,000$, and 0.2 for $r=5,000$.
\begin{figure}[H]
{\caption{Autocorrelation functions (ACFs) and kernel-smoothed density estimates for different CDA samplers in intercept-only probit model.}}
\subfigure[ACF for CDA without M-H adjustment.]{\label{probit_demo_intercept_proposal}
\includegraphics[width=0.3\linewidth]{probit_demo_acf_prop.pdf}
\quad
\subfigure[Posterior density estimates without M-H adjustment.]{\label{probit_demo_intercept_density}
\includegraphics[width=0.3\linewidth]{density_probit.pdf}}
\quad
\subfigure[ACF for CDA with M-H adjustment]{\label{probit_demo_intercept_posteriorsample}
\includegraphics[width=0.3\linewidth]{probit_demo_acf.pdf}}
}
\label{probit_demo_intercept}
\end{figure}
\section{Specific Algorithms} \label{sec:algos}
In this section, we describe CDA algorithms for general probit and logistic regression, and describe a general strategy for tuning $r,b$.
\subsection{Probit Regression}
Consider the probit regression:
\be
y_i \sim \Bern(p_i), \quad p_i = \Phi(x_i \theta) \quad i=1,\ldots,n
\ee
with improper prior $\Pi^0(\theta) \propto 1$. The data augmentation sampler \citep{tanner1987calculation, albert1993bayesian} has the update rule
\be
z_i \mid \theta, x_i, y_i &\sim \left\{ \begin{array}{cc} \No_{[0,\infty)}(x_i \theta,1) & \text{ if } y_i = 1 \\ \No_{(-\infty,0]}(x_i \theta,1) & \text{ if } y_i = 0 \end{array} \right. \quad i=1,\ldots,n\\
\theta \mid z, x, y &\sim \No((X'X)^{-1} X'z, (X'X)^{-1}).
\ee
\cite{liu1999parameter} and \cite{meng1999seeking}, among others, previously studied this algorithm and proposed to rescale $\theta$ through parameter expansion. However, this modification does not impact the conditional variance of $\theta$ and thus does not directly increase typical step sizes.
Our approach is fundamentally different, since we directly adjust the conditional variance. Similar to the intercept only model, we modify $\mbox{var} (\theta| z)$ by changing the scale of each $z_i$. Since the conditional variance is now a matrix, for flexible tuning, we let $r$ and $b$ vary over index $i$, yielding update rule
\be \label{eq:cda-probit}
z_i \mid \theta, x_i, y_i &\sim \left\{ \begin{array}{cc} \No_{[0,\infty)}(x_i \theta+b_i,r_i) & \text{ if } y_i = 1 \\ \No_{(-\infty,0]}(x_i \theta+b_i,r_i) & \text{ if } y_i = 0 \end{array} \right. \quad i=1,\ldots,n\\
\theta \mid z, X &\sim \No((X'R^{-1}X)^{-1} X'R^{-1}(z-b), (X'R^{-1}X)^{-1}),
\ee
where $R = \diag(r_1,\ldots,r_n)$, $b = (b_1,\ldots,b_n)'$, under the Bernoulli likelihood:
\be
\mbox{pr}(y_i = 1 | \theta, x_i, r_i, b_i) = & \int_{0}^{\infty}
\frac{1}{\sqrt{2 \pi r_i} } \exp\left(-\frac{(z_i-x_i\theta-b_i)^2}{2 r_i}
\right) dz_i \\
= & \Phi\bigg( \frac{x_i\theta+b_i}{\sqrt{r_i}}\bigg).
\label{eq:prop-marginal-probit}
\ee
For fixed $r = (r_1,\ldots,r_n)$ and $b = (b_1,\ldots,b_n)$, \eqref{eq:prop-marginal-probit} defines a proper Bernoulli likelihood for $y_i$ conditional on parameters, and therefore the transition kernel $K_{r,b}((\theta,z);\cdot)$ defined by the Gibbs update rule in \eqref{eq:cda-probit} would have a unique invariant measure for fixed $r,b$, which we denote $\Pi_{r,b}(\theta,z \mid y)$.
For insight into the relationship between $r$ and step size, consider the $\theta$-marginal autocovariance in a Gibbs sampler evolving according to $K_{r,b}$:
\be
\cov_{r,b}(\theta_t \mid \theta_{t-1},X,z,y) &= (X'R^{-1}X)^{-1} \\ & + (X'R^{-1}X)^{-1} X'R^{-1}\cov(z-b | R) R^{-1}X(X'R^{-1}X)^{-1} \\
&\ge (X'R^{-1}X)^{-1}, \label{eq:varlb-probit}
\ee
In the special case where $r_i = r_0$ for all $i$, we have
\be
\cov_{r,b}(\theta_t \mid \theta_{t-1}, X,z,y) &\ge r_0 (X'X)^{-1},
\ee
so that all of the conditional variances are increased by at least a factor of $r_0$. This holds uniformly over the entire state space, so it follows that
\be
\bb E_{\Pi_{r,b}}[\var(\theta_j \mid z)] \ge r_0 \bb E_{\Pi}[\var(\theta_j \mid z)].
\ee
The key to CDA is to choose $r,b$ to make $\bb E_{\Pi_{r,b}}[\var(\theta_j \mid z)]$ close to $\var_{\Pi_{r,b}}(\theta_j \mid z)$, while additionally maximizing the M-H acceptance probability. We defer the choice for $r,b$ and their effects to the last subsection.
\subsection{Logistic Regression}
Calibration was easy to achieve in the probit examples, because $\mbox{var}( \theta | z,y)$ does not involve the latent variable $z$. In cases in which the latent variable impacts the variance of the conditional posterior distribution of $\theta$, we propose to stochastically increase $\mbox{var}(\theta|z,y)$ by modifying the distribution of $z$. We focus on the logistic regression model with
\be
y_i \sim \Bern(p_i), \quad p_i = \frac{\exp(x_i \theta)}{1+\exp(x_i \theta)} \quad i=1,\ldots,n
\ee
and improper prior $\Pi^0(\theta) \propto 1$. For this model, \cite{polson2013bayesian} proposed Polya-Gamma data augmentation:
\be
z_i &\sim {\PG}(1, |\xtheta|) \quad i=1,\ldots,n,\\
\theta &\sim \No \left( (X' Z X)^{-1} X' (y-0.5) , (X' Z X)^{-1} \right),
\ee
where $Z= \diag(z_1,\ldots,z_n)$. This algorithm relies on expressing the logistic regression likelihood as
$$L( \xtheta; y_i )= \int \exp\{ \xtheta (y_i-1/2)\} \exp\bigg\{ -\frac{z_i (\xtheta)^2}{2}\bigg\} \PG(z_i \mid 1,0) dz_i,$$
where $\mbox{PG}(a_1,a_2)$ denotes the Polya-Gamma distribution with parameters $a_1,a_2$, with $\bb{E}z_i= {a_1}/{(2 a_2)}\tanh({a_2}/{2})$.
We replace $\PG(z_i \mid 1,0)$ with $\PG(z_i \mid r_i,0)$ in the step for updating the latent data. Since $\bb E_z \mbox{var}(\theta|z,y)$ lacks closed-form, we focus on the precision matrix $\bb E_z \big( \mbox{var}(\theta|z,y)\big)^{-1} = X' \bb E Z X$. Smaller $r_i$ can lead to smaller $\bb E z_i$, providing a route to calibration. Applying the bias-adjustment term $b_i$ to the linear predictor $\eta_i = x_i\theta$ leads to
\be
L_{r,b}(\xtheta;y_i) = & \int_{0}^{\infty} \exp\{ (\xtheta+b_i)
(y_i-r_i/2)\} \\ & \quad \exp\bigg\{ -\frac{z_i (\xtheta+b_i)^2}{2}\bigg\} \PG(z_i \mid r_i,0) dz_i \\
= & \frac{\exp \{ (x_i \theta + b_i)y_i \}}{\{1+\exp(\xtheta +b_i)\}^{r_i}},
\label{eq:prop-marginal-logit}
\ee
and the update rule for the CDA Gibbs sampler is then
\be
z_i &\sim {\PG}(r_i, |\xtheta+b_i|) \quad i=1,\ldots,n,\\
\theta^* &\sim \No \left( (X' Z X)^{-1} X' (y -r/2- Zb) , (X' Z X)^{-1} \right),
\ee
where $r = (r_1,\ldots,r_n)$. By \eqref{eq:mh-accrat}, the M-H acceptance probability is
\be
1 \wedge \prod_i \frac{ \{1+\exp(\xtheta)\} \{1+\exp(\xtheta^*+b_i)\}^{r_i} } { \{1+\exp(\xtheta^*)\} \{1+\exp(\xtheta+b_i)\}^{r_i} }.
\ee
\subsection{Choice of Calibration Parameters} \label{sec:tuning}
As illustrated in the previous subsection, efficiency of CDA is dependent on a good choice of the calibration parameters $r=(r_1,\ldots,r_n)$ and $b=(b_1,\ldots,b_n)$. We propose a simple and efficient algorithm for calculating ``good'' values of these parameters relying on Fisher information. {Although our choice of calibration parameters relies on large data sample arguments, we find that this calibration approach also works well in smaller data samples.}
{
Our
goal is to adjust the conditional variance under calibration of $(r,b)$ to approximately match the marginal variance under the exact target distribution.
{
The inverses of the following Fisher information provide useful approximation to the two posterior covariances.
\be \left (\mc I_{y\mid \theta}({\theta}) \right)_{i,j} & = \bb E_{y\mid \theta} \left[ \left( \frac{\partial}{\partial \theta_i} \log L(y;\theta) \right) \left( \frac{\partial}{\partial \theta_j} \log L(y;\theta) \right) \right], \\
\left( \mc I_{y\mid \theta,z}({\theta};r,b) \right)_{i,j}& = \bb E_{y\mid \theta,z} \left[ \left( \frac{\partial}{\partial \theta_i} \log L_{r,b}(y,z;\theta) \right)\left( \frac{\partial}{\partial \theta_j} \log L_{r,b}(y,z;\theta) \right) \right]
\ee
}
for $i=1,\ldots,p$, $j=1,\ldots,p$, with $\bb E_{y\mid \theta}$ taken over the distribution of $y$ under the target marginal $L(y;\theta)$ and $\bb E_{y\mid \theta,z}$ taken over the conditional distribution of $y$ under the augmented $L_{r,b}(y,z;\theta)$ with the calibration of $(r,b)$.
Since $\mc I_{y\mid \theta,z}({\theta};r,b)$ depends on random $z$, we marginalize over
the conditional distribution
of $z$ under $L_{r,b}(\theta; y)$ and obtain $\bb E_{z\mid \theta}\mc I_{y\mid \theta,z}(\theta;r,b)$. Via adjusting $r$, one can then minimize the difference between $\mc I_{y\mid \theta}(\theta)$
and $\bb E_{z\mid \theta}\mc I_{y\mid \theta,z}(\theta;r,b)$.}
{
Often, one can avoid computing the full Fisher information. For each class of models under the same data augmentation, they share the same form of conditional likelihoods for $y\mid \eta(\theta)$ given a mapping $\eta(\theta):\bb
R^p\rightarrow \bb R^d $.
For example, all Bernoulli probit models follow $y_{i}\mid \eta_{i}(\theta) \stackrel{iid}{\sim}\Bern( \Phi(\eta_{i}(\theta)) )$, except with different $\eta(\theta)$. The above full Fisher information can be rewritten as
\be\mc I_{y\mid \theta}({\theta}) & = \dot \eta\mc I_{y\mid \theta}(\eta({\theta})) \dot \eta', \\
\mc I_{y\mid \theta,z}({\theta};r,b) & = \dot \eta\mc I_{y\mid \theta,z}(\eta({\theta)};r,b) \dot \eta',
\ee
where $\dot \eta$ denotes the $p$-by-$d$ gradient matrix consisting of the partial
derivative ${\partial
\eta_{k}(\theta)}/{\partial \theta_j}$ of the $k$th output
$\eta_{k}(\theta)$ with respect to $\theta_j$. It suffices to reduce the difference between $\mc I_{y\mid \theta}(\eta({\theta}))$ and $\mc I_{y\mid \theta,z}(\eta({\theta)};r,b)$ instead of the full Fisher
information. The solution is a function of $\eta$, with form invariant to models under the same conditional likelihood of $y\mid \eta$.
}
{
In all CDA algorithms presented in
this article, $\mc I_{y\mid \theta}(\eta({\theta}))$ and $\mc I_{y\mid \theta,z}(\eta({\theta)};r,b)$ are simple diagonal matrices, and
it can be made exactly $\mc I_{y\mid \theta}({\theta})=\bb E_{z\mid \theta}\mc I_{y\mid \theta,z}({\theta};r,b)$ for given $\theta$ with a closed-form
solution. As there could be more complicated
scenarios, we suggest the following. When the difference cannot
be simply eliminated, one could utilize a metric between two matrices, such
as Rao's distance $\{\text{tr}
[\log(A^{-1/2}BA^{-1/2})^{2}]\}^{1/2}$ with $\text{tr}$ as the trace \citep{atkinson1981rao},
and an optimization algorithm to minimize the difference. When the Fisher information
is intractable to compute, one could instead utilize the observed Fisher information
\citep{efron1978assessing},
$$\bigg(\hat{\mc I}_{y\mid \theta}({\theta}) \bigg)_{i,j} = \bigg( \frac{\partial}{\partial \theta_i} \log L(y;\theta) \bigg) \bigg( \frac{\partial}{\partial \theta_j} \log L(y;\theta) \bigg),$$
$$ \bigg( \hat{\mc I}_{y\mid \theta,z}({\theta};r,b) \bigg)_{i,j} = \bigg( \frac{\partial}{\partial \theta_i} \log L_{r,b}(y,z;\theta) \bigg)\bigg( \frac{\partial}{\partial \theta_j} \log L_{r,b}(y,z;\theta) \bigg),$$
with $y$ the observed data.
}
{As the Fisher information matrices depend on $\theta$, we use an adaptation phase to dynamically update $r_t,b_t$ with posterior sample $\theta_t$. Then
we stop adaptation after $\theta_t$ enters the high posterior
density region. This approach is similar to using the frequentist Fisher information evaluated at the maximum-a-posteriori
(MAP) estimate, except it does not require the extra optimization steps for computing
the MAP. It works well empirically in examples we have considered.}
Specifically, we choose $r_{t+1}$ to minimize the difference between $\mc I_{y\mid \theta_{t}}(\theta_{t})$
and $\bb E_{z\mid \theta_{t}}\mc I_{y\mid \theta_{t},z}(\theta_{t};r_{t+1},b_{t})$, or $\mc I^{-1}_{y\mid \theta_{t}}(\theta_{t})$
and $\bb E_{z\mid \theta_{t}}\mc I^{-1}_{y\mid \theta_{t},z}(\theta_{t};r_{t+1},b_{t})$.
Additionally, we set $b_{t+1}$ to minimize the difference between
$L_{1,0}(\theta_t;y)$ and $L_{r_{t+1},b_{t+1}}(\theta_t;y)$. Thus, we use $r$ to adjust the conditional variance based on $L_{r,b}$ to match the marginal variance based on $L$ and $b$ to make $L_{r,b}$ close to $L_{1,0}$ in the neighborhood of $\theta_t$. Intuitively, this will make the target distribution closer to the invariant measure of calibrated Gibbs, and correspondingly increase the MH acceptance rate. {Some illustrative results about the adaptation
are provided in the appendix.} The proposal kernel we describe above is \emph{adaptive}; that is, we have a collection of proposal kernels $\mc Q = \{Q_{r,b}\}_{(r,b) \in \bb R_+ \times \bb R}$, and we choose a different member of $\mc Q$ at each iteration to create the proposal.
In general, ergodicity of adaptive algorithms requires a diminishing adaptation condition \citep{roberts2007coupling}. {For simplicity, we satisfy this condition by stopping adaptation after a tuning phase.}
For a concrete illustration, we first return to the first example of probit regression. Letting $\eta_i = x_i\theta$, we obtain
\be
\mc I_{y\mid \theta}({\theta}) &= \dot\eta\diag\bigg\{\frac{\phi(\eta_i)^2}{ {\Phi(\eta_i)(1- \Phi(\eta_i))}}\bigg\}\dot\eta',\qquad \bb E_{z\mid \theta}\mc I_{y\mid \theta,z}({\theta};r,b)= \dot\eta R^{-1} \dot\eta',
\ee
where $\phi$ is the standard normal density, with $\dot\eta=X'$. {Having $\mc I_{y\mid \theta}({\theta})= \bb
E_{z\mid \theta}\mc I_{y\mid \theta,z}({\theta};r,b)$ and $L_{r,b}(\eta_i;y_i)= L(\eta_i;y_i)
$ yields
\be
r_i &= \frac{\Phi(\eta_i)(1- \Phi(\eta_i))} {\phi(\eta_i)^2},\\
b_i &= \eta_i (\sqrt{r_i}-1).
\ee
For Bernoulli probit models with other forms of $\eta_i$, the solution for tuning parameters remains the same.}
In simulation, we consider a probit regression with an intercept and two predictors $x_{i,1},x_{i,2}\sim \No(1,1)$, with $\theta=(-5,1,-1)'$, generating $\sum y_i=20$ among $n=10,000$. The \cite{albert1993bayesian} DA algorithm mixes slowly (Figure~\ref{probit_reg_trace} and \ref{probit_reg_acf}). We also show the
results of the parameter expansion algorithm (PX-DA) proposed by \cite{liu1999parameter}. PX-DA only mildly reduces the correlation, as it does not solve the small step size problem. For CDA, we tuned $r$ and $b$ for $100$ steps using the Fisher information, reaching a satisfactory acceptance rate of $0.6$ {and leading to dramatically better mixing}.
\begin{figure}[H]
\subfigure[Traceplot for the original DA, parameter expanded DA and CDA algorithms.]{\label{probit_reg_trace
\includegraphics[width=0.45\linewidth]{probit15_trace_plot.pdf}
\qquad
\subfigure[ACF for original DA, parameter expanded DA and CDA algorithms.]{\label{probit_reg_acf
\includegraphics[width=0.45\linewidth]{probit15_acf.pdf}}
}
{\caption{Panel (a) demonstrates in traceplot and panel (b) in autocorrelation the substantial improvement in CDA by correcting the variance mis-match in probit regression with rare event data, compared with the original \citep{albert1993bayesian} and parameter-expanded methods \citep{liu1999parameter}.}}
\end{figure}
For the second example of logistic regression, taking $\eta_i=x_i\theta$, the Fisher information matrices
are:
\begin{eqnarray}
\mc I_{y\mid \theta}({\theta}) =& \dot\eta \diag\bigg\{\frac{\exp(\eta_i)}{ \{1+\exp(\eta_i)\} ^2}\bigg\} \dot\eta', \nonumber \\
\bb E_{z\mid \theta}\mc I_{y\mid \theta,z}({\theta};r,b)= & \dot\eta \diag\bigg\{ \frac{r_i}{2 |\eta_i+b_i|}\tanh\Big(\frac{|\eta_i+b_i|}{2} \Big)\bigg\} \dot\eta' . \nonumber
\end{eqnarray}
where $\dot\eta=X'$.
Setting $\mc I_{y\mid \theta}({\theta})=\bb E_{z\mid \theta}\mc I_{y\mid \theta,z}({\theta};r,b)$ and $ \{1+\exp(\eta_i)\} = \{1+\exp(\eta_i+b_i)\}^{r_i}$
to locally maximize the M-H acceptance rate yields
\be
r_i & =\frac{\exp(\eta_i)}{ \{1+\exp(\eta_i)\} ^2} {2 |\eta_i+b_i|}/ \tanh\Big(\frac{|\eta_i+b_i|}{2} \Big),
\\ b_i & = \log[ \{1+\exp(\eta_i)\}^{1/r_i} -1] - \eta_i.
\ee
{Again, this tuning solution is invariant to the different forms of $\eta$ under
the Bernoulli logistic model.}
To illustrate, we use a two parameter intercept-slope model with $x_1\sim \No(0,1)$ and $\theta=(-9,1)'$. With $n= 10^5$, we obtain rare outcome data with
$\sum y_{i} = 50 $. Besides the original DA algorithm \citep{polson2013bayesian}, we also consider an M-H sampler using a multivariate normal proposal $\theta^*|\theta \sim \No(\theta^*| \theta, {\mc I}^{-1}(\theta))$ with the inverse Fisher information as the covariance. {Similarly, we test an alternative of using DA to generate new $\theta^*$, and scaling to $\theta^{**}=\theta+\alpha(\theta^{*}-\theta)$, with $\alpha\ge 1$, as an M-H proposal. Both M-H with a normal proposal and with scaled proposal suffer from low acceptance rate, unless $\alpha\approx 1$ in the latter (corresponding to almost no adjustment from DA).} For CDA we tuned $r$ and $b$ for $100$ steps, reaching an acceptance rate of $0.8$. Shown in Figure~\ref{logit_random_mixing}, DA and simple M-H mix slowly, exhibiting strong autocorrelation even at lag $40$, while CDA has dramatically better mixing.
\begin{figure}[H]
\subfigure[Traceplots for DA, CDA and M-H with multivariate normal proposal.]
\includegraphics[width=0.45\linewidth]{logit_demo_trace_plot}
\qquad
\subfigure[ACF for DA, CDA and M-H with multivariate normal proposal.]
\includegraphics[width=0.45\linewidth]{logit_demo_acf}}
}
{\caption{Panel (a) demonstrates in traceplot and panel (b) in autocorrelation the substantial improvement of CDA in logistic regression with rare event data, compared with the original DA \citep{polson2013bayesian} and the M-H algorithm with multivariate normal proposal (MH-MVN).\label{logit_random_mixing}}}
\end{figure}
{
\section{Simulation Study: Scaling to Massive $n$}
}
As motivated above, two factors are necessary for obtaining usable posterior samples within a practical time: a low computing cost in each iteration and a high effective sample size within a small number of iterations. We first demonstrate that calibration can solve the latter issue.
To start, we consider a simple Bernoulli logistic regression with a common intercept:
$$y_i\stackrel{iid}{\sim} \Bern\Big(\frac{\exp(\theta)}{1+\exp(\theta)}\Big), \quad i=1,\ldots,n,$$
with a flat improper prior for $\theta$. As the likelihood is \\ $L(y;\theta )= {\exp(\theta\sum y_i)}{(1+\exp(\theta))^{-n}} $, it enjoys efficient computing per iteration
that only involves $1$ Polya-Gamma latent variable. Alternatively, using calibration parameters $(r,b)$, a proposal can be simulated from
\be
z &\sim \PG\left ( n r,\theta+b \right),\\
\theta^* &\sim \No \left( \frac{ \sum y_i - r n/2 -z b }{z}, \frac{1}{z}\right),
\ee
and M-H acceptance step as described above using \\ $L_{r,b}(\theta; y) = {\exp(\theta+b)^{\sum y_i}}{\{ 1+\exp(\theta+b)\}^{-nr}}$.
To have a proper $L_{r,b}(\theta;y)$, we further require $r \ge (\sum y_i-1)/n + \epsilon$ with $\epsilon$ a small positive constant.
Using Fisher information, the parameters are adapted initially for $200$
steps, via
\be
r & =\frac{\exp(\theta)}{ \{1+\exp(\theta)\} ^2} / \left ( \frac{1}{2 |\theta+b|} \tanh\frac{|\theta+b|}{2} \right) \vee \big ( (\sum y_i-1)/n + \epsilon \big), \\
b & =\log[ \{1+\exp(\theta)\}^{1/r} -1] - \theta.
\ee
{To obtain enormous data sample size rare event data}, we fixed $\sum
y_i=1$ and increase $n$ from $10^1$ to a massive $10^{14}$.
Figure~\ref{massive_n_sims}(a) compares the effective sample size per $1,000$
steps using DA and CDA. Surprisingly, the deterioration of DA shows up as early
as $n=10^2$; its slow-down becomes critical at $n=10^4$ with
effective sample size close to $0$. CDA performs exceptionally well, even
at massive $n=10^{14}$ (we stop at $10^{14}$ as $1/n$ reaches the limit of floating
point accuracy).
In more complicated settings, one issue for data augmentation in general is the large number of latent variables to sample in each iteration. A common strategy is to {avoid sampling latent variables for
every observation by approximating the Markov transition kernel using subsamples}
\citep{quiroz2016exact,johndrow2015approximations}. {Different from other
example algorithms, this approximation changes the invariant measure. Finding
a suitable sub-sample size while bounding approximation
error requires careful treatment, which is beyond the scope of
this article. Instead, our goal is to show sub-sampling alone does not address the burden of low ESS issue; whereas }{one can trivially couple our proposed CDA
strategy with such subsampling to scale DA-MCMC up to enormous data sample sizes. We illustrate such
coupling here.}
We consider the same two-parameter intercept-slope model in logistic regression
as described in the last section, except we now vary data sample size from
$n=10^5$ to $10^8$. We simulate Bernoulli outcome \\$y_i\sim \Bern\big(({1+\exp(-x_i\theta)})^{-1}\big)$ based on $x_1\sim \No(0,1)$ and $\theta=(-\theta_0,1)'$. We vary $\theta_0$ and induce $\sum{y_i}\approx
10$ for each $n$. We utilize the sub-sampled-Polya-Gamma algorithm described by \cite{johndrow2015approximations},
and apply CDA to calibrate the variance discrepancy. Since $y$ is highly imbalanced in
the number of $0$ and $1$s, we apply biased-sampling by including all data
with $y_i=1$, while sub-sampling $1\%$ of data with $y_i=0$. {Existing work on applying biased subsampling in logistic regression mainly aims to obtain point estimates \citep{king2001logistic,wang2017optimal},
in this article we present a simple solution for Bayesian inference.}
{Denoting the set of all data with $y_i=1$ as $V_1$ and a random subset with $y_i=0$ as $V_0$, it is sensible to keep
the likelihood contribution from $y_i=1$ unchanged, while adjusting the part from $y_i=0$ via a power of a ratio of $({n-|V_1|})/{|V_0|}$, leading to an approximate likelihood
$$L(\theta;y) = \prod_{i\in V_1}\frac{\exp(x_i\theta)}{ 1+\exp(x_i\theta)} (\prod_{i\in V_0}\frac{1}{ 1+\exp(x_i\theta)}
)^{\frac{n-|V_1|}{|V_0|}}.$$
The number of latent variables is reduced to $|V_0|+|V_1|$; since
$n$ is still large, the slow mixing would remain and calibration is needed.
} The algorithmic details and the calibrated form are presented in the appendix.
Figure~\ref{massive_n_sims}(b) compares the performance of the two approximating
algorithms, one combining CDA and sub-sampling, and one using sub-sampling alone. Clearly, only accelerating each step via sub-sampling does not solve the inefficiency of very low
effective sample size; while using CDA and sub-sampling together can produce {excellent computational performance}.
\begin{figure}[H]
{\caption{CDA maintains high effective sample size, even when scaling
to massive $n$. Panel(a) shows the performance of DA and CDA when $n$ is scaled up to
$10^{14}$; Panel(b) shows the performance of CDA and DA, coupled with sub-sampling
approximation to reduce the number of sampled latent variables. Only accelerating computing time in each iteration (DA-Subsampling) does not solve the scalability issue. \label{massive_n_sims}}}
\subfigure[Effective sample size (with 95\% pointwise credible interval) per $1,000$ steps with different sample
size $n$ from $10$ to $10^{14}$, using logistic regression model with intercept only.]
\includegraphics[width=0.45\linewidth]{simMassiveN}
\qquad
\subfigure[Effective sample size per (with 95\% pointwise credible interval) $1,000$ steps with different $n$
from $10^5$ to $10^{8}$, in logistic regression with slope and intercept, using sub-sampling.]
\includegraphics[width=0.55\linewidth]{simMassiveNSubsampling}}
}
\end{figure}
\section{Co-Browsing Behavior Application}
We apply CDA to an online browsing activity dataset. The dataset contains a two-way table of visit count by users who browsed one of $96$ client websites of interest, and one of the $n=59,792$ high-traffic sites during the same browsing session. We refer to visiting more than one site during the same session as co-browsing. For each of the client websites, it is of large commercial interest to find out the high-traffic sites with relatively high co-browsing rates, so that ads can be more effectively placed. For the computational advertising company, it is also useful to understand the co-browsing behavior and predict the traffic pattern of users. We consider two models for these data.
\subsection{Hierarchical Binomial Model for Co-Browsing Rates}
We initially focus on one client website and analyze co-browsing rates with the high-traffic sites. With the total visit count $N_i$ available for the $i$th high-traffic site, the count of co-browsing $y_i$ can be considered as the result of a binomial trial. With $y_i$ extremely small relative to $N_i$ (ratio $0.00011 \pm 0.00093$), the maximum likelihood estimate $y_i/N_i$ can have poor performance. For example, when $y_i=0$, estimating the rate as exactly $0$ is not ideal. Therefore, it is useful to consider a hierarchical model to allow borrowing of information across high-traffic sites:
\be
y_i \sim \Binom\left(N_i, \frac{\exp(\theta_i)}{1+\exp(\theta_i)}\right), \quad \theta_i\stackrel{iid}{\sim} \No(\theta_0, \sigma^2_0), \quad i=1\ldots n\\
(\theta_0,\sigma^2_0) \sim \pi(\theta_0,\sigma^2_0)
\ee
{We choose weakly informative priors}. Based on expert opinion in quantitative advertising, we use a prior $\theta_0\sim \No(-12,49)$ and uniform prior on $\sigma^2_0$. Similar to the logistic regression, we calibrate the binomial Polya-Gamma augmentation, leading to the proposal likelihood:
\be
L_{r,b}(\theta_i;y_i,N_i, r_i, b_i) = \frac{\exp(\theta_i+b_i)^y_i}{\{ 1+\exp(\theta_i+b_i)\}^{N_ir_i}}
\ee
Conditioned on the latent Polya-Gamma latent variable $z_i$, each proposal $\theta^*_i$ can be sampled from:
\be
z_i &\sim \PG\left( (N_ir_i),\theta_i+b_i \right),\\
\theta_i^* &\sim \No \left( \frac{ y_i - r_i N_i/2 -z_i b_i + \theta_0/\sigma^2_0}{z_i+ 1/\sigma^2_0}, \frac{1}{z_i+ 1/\sigma^2_0}\right),
\ee
and accepted or rejected using an M-H step. We further require $r_i \ge (y_i-1)/N_i + \epsilon$ to have a proper $L_{r,b}(\theta_i;y_i, N_i)$ with $\epsilon$ a small constant. Similar to logistic regression, the auxiliary parameters are chosen as
\be
r_i & =\frac{\exp(\theta_i)}{ \{1+\exp(\theta_i)\} ^2} / \left ( \frac{1}{2 |\theta_i+b_i|} \tanh\frac{|\theta_i+b_i|}{2} \right) \vee \big ( (y_i-1)/N_i + \epsilon \big),\\
b_i &=\log[ \{1+\exp(\theta_i)\}^{1/r_i} -1] - \theta_i\ee
during adaptation. Since $\theta_i$'s are conditionally independent, the calibrated proposal can be individually accepted with high probability for each $i$. This leads to a high average acceptance of $0.9$, despite the high dimensionality of $59,792$ $\theta_i$'s.
After $\theta_i$'s are updated, other parameters are sampled from
$\theta_0 \sim \No\big( (n/\sigma^2 +1/49)^{-1} (\sum_i \theta_i /\sigma^2 -12/49 ), (n /\sigma^2 +1/49)^{-1} \big),$ and $\sigma^2_0 \sim \IG( n/2-1, \sum_i (\theta_i -\theta_0)^2 /2)$.
Figure~\ref{data_binomial} shows the boxplots of the ACFs for all $\theta_i$'s. We compare the result with the original DA \citep{polson2013bayesian} and Hamiltonian Monte Carlo (HMC) provided by the \texttt{STAN} software \citep{carpenter2016stan}. We run DA for $100,000$ steps, HMC for $2,000$ steps and CDA for $2,000$ steps, so that they have approximately the same effective sample size (calculated with the \texttt{CODA} package in \texttt{R}). All of the parameters mix poorly in DA; HMC and CDA lead to significant improvement with autocorrelation rapidly decaying to close to zero within $5$ lags.
Shown in Table~\ref{tab:binomial}, CDA and HMC have very close estimates in posterior means and $95\%$ credible intervals for the parameters, while DA has poor estimates due to critically slow mixing. The difference between HMC and CDA is that, although HMC is slightly more efficient in effective sample size per iteration ($T_{eff}/T$) for this model, it is much more computationally intensive and generates many fewer iterations than CDA within the same budget of computing time. As the result, CDA has the most efficient computing time per effective sample.
\begin{figure}[H]
{\caption{Boxplots of the ACFs show the mixing of the $59,792$ parameters in the hierarchical binomial model, for the original DA\citep{polson2013bayesian}, CDA and HMC. \label{data_binomial}}}
\subfigure[ACFs of the rate parameters $\theta_i$ using DA.]
\includegraphics[width=0.32\linewidth]{binomial_random_acf_da}
\subfigure[ACFs of the rate parameters $\theta_i$ using CDA.]
\includegraphics[width=0.32\linewidth]{binomial_random_acf_cda}}
\subfigure[ACFs of the rate parameters $\theta_i$ using HMC.]
\includegraphics[width=0.32\linewidth]{binomial_random_acf_hmc}}
}
\end{figure}
\begin{table}[H]
\tiny
\centering
\begin{tabular}{|l |r |r| r| r |}
\hline
& DA & CDA & HMC\\
[0.5ex]
$ \sum \theta_i/n$ & -10.03 (-10.16, -9.87)& -12.05 (-12.09, -12.02) & -12.06 (-12.09, -12.01)\\
$ \sum \theta_i^2/n$ & 102.25 (98.92, 105.23)& 153.04 (152.06, 154.05) & 153.17 (152.02, 154.29)\\
$\theta_0$ & -10.03 (-10.17, -9.87)& -12.05 (-12.09,-12.01) & -12.06 (-12.10, -12.01)\\
$\sigma^2$ & 1.60 (1.36, 1.82)& 7.70 (7.49, 7.88) & 7.71 (7.51, 7.91)\\
$T_{eff} / T$ & 0.0085 (0.0013 0.0188) & 0.5013 (0.1101,1.0084) & 0.8404 (0.5149, 1.2470)\\
Avg Computing Time / $T$ & 1.2 sec & 1.2 sec & 6 sec\\
Avg Computing Time / $T_{eff}$ & 140.4 sec & 0.48 sec & 1.3 sec\\
\hline
\end{tabular}
\caption{Parameter estimates (with 95\% credible intervals) and computing speed (ratios among computing time, effective sample sizes $T_{eff}$ and total iterations $T$) of the DA, CDA and HMC in hierarchical binomial model. CDA provides parameter estimates as accurate as HMC, and is more computationally efficient than HMC.}
\label{tab:binomial}
\end{table}
\subsection{Poisson Log-Normal Model for Web Traffic Prediction}
The co-browsing on one high-traffic site and one client site is commonly related to the click-through of users from the former to the latter. Therefore, the count of co-browsing is a useful indication of the click-through traffic. For any given client website, predicting the high traffic sites that could generate the most traffic is of high commercial interest. Therefore, we consider a Poisson regression model. We choose the co-browsing count of one client website as the outcome $y_i$ and the log count of the other $95$ websites as the predictors $x_{ij}=\log (x^*_{ij}+1)$ for $i=1,\ldots ,59792$ and $j=1,\ldots ,95$. A Gaussian random effect is included to account for over-dispersion relative to the Poisson distribution, leading to a Poisson log-normal regression model:
\be
y_i \sim \Poi \left( \exp (\xbeta + \tau_i )\right), \quad \tau_i\stackrel{iid}{\sim} \No(\tau_0, \nu^2), \quad i=1\ldots n\\
\beta \sim \No(0, I \sigma_\beta^2), \quad \tau_0 \sim \No(0,\sigma_\tau^2) \quad \nu^2\sim \pi(\nu^2).
\ee
We assign a weakly informative prior for $\beta$ and $\tau_0$ with $ \sigma_\beta^2=\sigma_\tau^2=100$. For the over-dispersion parameter $\nu^2$, we assign a non-informative uniform prior.
When $\beta$ and $\tau$ are sampled separately, the random effects $\tau = \{\tau_1,\ldots, \tau_n\}$ can cause slow mixing. Instead, we sample $\beta$ and $\tau$ jointly. Using $\tilde X = [ I_n || X ]$ as a $n \times (n+p)$ juxtaposed matrix, and $\eta_i=\xbeta + \tau_i$ for the linear predictor, the model can be viewed as a linear predictor with $n+p$ coefficients, and $\theta= \{\tau, \beta\}'$ can be sampled jointly in a block. The reason for improved mixing with blocked sampling can be found in \cite{liu1994collapsed}.
We now focus on the mixing behavior of data augmentation. We first review data augmentation for the Poisson log-normal model. \cite{zhou2012lognormal} proposed to treat $\Poi(\eta_i)$ as the limit of the negative binomial $\NB\big(\lambda,{\eta_i}/{(\lambda+\eta_i)}\big)$ with $\lambda\rightarrow \infty$, and used moderate $\lambda=1,000$ for approximation. The limit can be simplified as (omitting constant):
\be
L(\eta_i;y_i)=\frac{ \exp(y_i \eta_i \} }{\exp\{\exp(\eta_i)\}} = \lim_{\lambda\rightarrow\infty}\frac{\exp(y_i \eta_i)}{\{1+ \exp(\eta_i )/\lambda\}^{\lambda }}.
\label{eq:pos_approx}
\ee
With finite $\lambda$ approximation, the posterior can be sampled via Polya-Gamma augmented Gibbs sampling:
\be
z_i \mid \eta_i \sim \PG ( & \lambda, \eta_i -\log \lambda) \quad i=1\ldots n\\
\theta \mid z, y \sim \No \bigg( & \Big(\tilde X' Z \tilde X+
\begin{bmatrix} 1/\nu^2 \cdot I_n & 0\\ 0 & 1/\sigma^2_{\beta} \cdot I_p
\end{bmatrix}\Big)^{-1} \\
& \quad \{ \tilde X' \big ( y - \lambda/2 + Z \log \lambda\big) + \begin{bmatrix} \tau_0/\nu^2 1_n \\ 0_p \end{bmatrix} \} , \\
& \Big(\tilde X' Z \tilde X+ \begin{bmatrix} 1/\nu^2 \cdot I_n & 0\\ 0 & 1/\sigma^2_{\beta} \cdot I_p \end{bmatrix}\Big)^{-1} \bigg),
\ee
where $Z = \diag\{ z_1, \ldots, z_n\}$, $1_n = \{1, \ldots 1\}'$ and $0_p = \{0, \ldots 0\}'$.
However, this approximation-based data augmentation is inherently problematic. For example, setting
$\lambda = 1,000$ leads to large approximation error. As in \eqref{eq:pos_approx}, the approximating denominator has $(1+\exp\left(\eta_i)/\lambda\right)^\lambda= \exp \{ \exp(\eta_i) + \bigO(\exp(2\eta_i)/\lambda) \}$; for moderately large $\eta_i \approx 10$, $\lambda$ needs to be at least $10^9$ to make $\exp(2\eta_i)/\lambda$ close to $0$. This large error cannot be corrected with an additional M-H step, since the acceptance rate would be too low. On the other hand, it is not practical to use a large $\lambda$ in a Gibbs sampler, as it would create extremely large $z_i$ (associated with small conditional covariance for $\theta$), resulting in slow mixing.
We use CDA to solve this dilemma. We first choose a very large $\lambda$ ($10^9$) to control the approximation error, then use a small fractional $r_i$ multiplying to $\lambda$ for calibration. This leads to a proposal likelihood similar to the logistic CDA:
\be
L_{r,b}(\xtheta;y_i)=\frac{\exp(\eta_i -\log \lambda +b_i)^{y_i}}{\{1+ \exp(\eta_i -\log \lambda +b_i)\}^{r_i\lambda }},
\ee
with $r_i \ge (y_i-1)/\lambda + \epsilon$ for proper likelihood, and proposal update rule:
\be
\small
z_i \sim \PG ( & r_i\lambda, \eta_i -\log \lambda + b_i) \quad i=1\ldots n\\
\theta^* \sim \No \bigg( & \Big(\tilde X' Z \tilde X+ \begin{bmatrix}
1/\nu^2 \cdot I_n & 0\\ 0 & 1/\sigma^2_{\beta} \cdot I_p
\end{bmatrix}\Big)^{-1} \\
& \quad \Big\{ \tilde X' \big( y - r\lambda/2 + Z \log (\lambda -b )\big) + \begin{bmatrix} \tau_0/\nu^2 1_n \\ 0_p \end{bmatrix} \Big\} , \\
& \Big(\tilde X' Z \tilde X+ \begin{bmatrix} 1/\nu^2 \cdot I_n & 0\\ 0 & 1/\sigma^2_{\beta} \cdot I_p \end{bmatrix}\Big)^{-1} \bigg)
\ee
Letting $\eta_i^* = \tilde X \theta^*$, the proposal is accepted with probability (based on Poisson density and the approximation $L_{r,b}(\xtheta;y_i)$):
\be
1 \wedge \prod_i \frac{ \exp \{ \exp (\eta_i)\}}{ \exp \{ \exp (\eta_i^*)\}} \frac {{\{1+ \exp(\eta_i^{*} -\log \lambda +b_i)\}^{r_i\lambda }}}{{\{1+ \exp(\eta_i -\log \lambda +b_i)\}^{r_i\lambda }}}.
\ee
During the tuning, we set
\be
r_i & = \tau_i\exp(\eta_i) / \left( \frac{\lambda } {2|\eta_i + b_i - \log\lambda|} \tanh\frac{ |\eta_i + b_i - \log\lambda|}{2} \right ) \vee \big( (y_i-1)/\lambda + \epsilon \big), \\
b_i &=\log[ \exp \{ \exp(\eta_i - \log\lambda -\log r_i) \}-1] -\eta_i + \log\lambda.
\ee
After $\theta$ is updated, the other parameters can be sampled via
\be
\tau_0 &\sim \No\left( (n/ \nu^2 + 1/ \sigma^2_\tau)^{-1} \sum_i
\tau_i/\nu^2 , (n/ \nu^2 + 1/ \sigma^2_\tau)^{-1} \right)\\
\nu^2 &\sim
\IG ( n/2-1, \sum_i (\tau_i-\tau_0)^2 /2)
\ee
We ran the basic DA with $\lambda=1,000$ approximation, CDA with $\lambda=10^9$ and HMC. We ran DA for $200,000$ steps, CDA for $2,000$ steps and HMC for $20,000$ steps so that they have approximately the same effective sample size. For CDA, we used the first $1,000$ steps for adapting $r$ and $b$. Figure~\ref{data_poisson} shows the mixing of DA, CDA and HMC. Even with small $\lambda = 1,000$ in DA, all of the parameters mix poorly; HMC seemed to be affected by the presence of random effects, and most of parameters remain highly correlated within $40$ lags; CDA substantially improves the mixing. Table~\ref{table:Poisson} compares all three algorithms. CDA has the most efficient computing time per effective sample, and is about $30-300$ times more efficient than the other two algorithms.
\begin{figure}[H]
{\caption{CDA significantly improves the mixing of the parameters in the Poisson log-normal. \label{data_poisson}}}
\subfigure[Autocorrelation of the parameters from DA.]
\includegraphics[width=0.3\linewidth]{poisson_acf_da}
\subfigure[Autocorrelation of the parameters from CDA.]
\includegraphics[width=0.3\linewidth]{poisson_acf_cda}}
\subfigure[Autocorrelation of the parameters from HMC.]
\includegraphics[width=0.3\linewidth]{poisson_acf_hmc}}
}
\end{figure}
To evaluate the prediction performance, we use another co-browsing count table for the same high traffic and client sites, collected during a different time period. We use the high traffic co-browsing count $x_{ij}^{\dagger*}$ and their log transform
$x^\dagger_{ij} = \log( x_{ij}^{\dagger*} +1 )$ for the $j = 1,\ldots,95$ clients to predict the count for the client of interest $y_i^\dagger$, over the high traffic site $i=1,\ldots, 59792$. We carry out prediction using $\hat y_i^\dagger= \bb E_{ \beta, \tau \mid y,x} y_{i}^\dagger =\bb E_{ \beta, \tau \mid y,x}\exp( x_{i}^\dagger\beta + \tau_i)$ on the client site. The expectation is taken over posterior sample $\beta, \tau \mid y,x$ with training set $\{y,x\}$ discussed above. Cross-validation root-mean-squared error $\big(\sum_i(\hat y_i^\dagger - y_i^\dagger)^2/n\big)^{1/2}$ between the prediction and actual count $ y_i^\dagger$'s is computed. Shown in Table~\ref{table:Poisson}, slow mixing in DA and HMC cause poor estimation of the parameters and high prediction error, while CDA has significantly lower error.
\begin{table}[H]
\tiny
\centering
\begin{tabular}{|l |r |r| r| r |}
\hline
& DA & CDA & HMC\\
[0.5ex]
$\sum \beta_j / 95$ & 0.072 (0.071, 0.075)& -0.041 (-0.042, -0.038) & -0.010 (-0.042, -0.037) \\
$\sum \beta_j^2 / 95$ & 0.0034 (0.0033, 0.0035)& 0.231 (0.219 0.244) & 0.232 (0.216 0.244) \\
$\sum\tau_i/n$ & -0.405 (-0.642, -0.155)& -1.292 (-2.351, -0.446) & -1.297 (-2.354, -0.451) \\
$\sum\tau_i^2/n$ & 1.126 (0.968, 1.339)& 3.608 (0.696, 7.928) & 3.589 (0.678, 8.011) \\
Prediction RMSE & 33.21 & 8.52 & 13.18\\
$T_{eff} / T$ & 0.0037 (0.0011 0.0096) & 0.3348 (0.0279, 0.699) & 0.0173 (0.0065, 0.0655) \\
Avg Computing Time / $T$ & 1.3 sec & 1.3 sec & 56 sec\\
Avg Computing Time / $T_{eff}$ & 346.4 sec & 11.5 sec & 3240.6 sec\\
\hline
\end{tabular}
\caption{Parameter estimates, prediction error and computing speed of the DA, CDA and HMC in Poisson regression model.}
\label{table:Poisson}
\end{table}
\section{Discussion}
Data augmentation (DA) is a technique routinely used to enable implementation of simple Gibbs samplers, avoiding the need for expensive and complex tuning of Metropolis-Hastings algorithms.
Despite the convenience, DA can slow down mixing when the conditional posterior variance given the augmented data is substantially smaller than the marginal variance. When the data sample size is massive, this problem arises when the rates of convergence of the augmented and marginal posterior differ, leading to
critical mixing problems. There is a very rich literature on strategies for improving mixing rates of Gibbs samplers, with centered or non-centered re-parameterizations \citep{papaspiliopoulos2007general} and parameter-expansion \citep{liu1999parameter} leading to some improvements. However, existing approaches {do not} solve large sample mixing problems in not addressing the fundamental rate
mismatch issue.
To tackle this problem, we propose to calibrate data augmentation and use a parameter to directly adjust the conditional variance (which is associated with step size). CDA adds a little cost due to the likelihood evaluation, which is often negligible as compared to the random number generation. In this article, we demonstrate that calibration is generally applicable when $\theta \mid z$ belongs to the location-scale family. We expect it to be extensible to any conditional distribution with a variance or scale.
As both CDA and HMC involve M-H steps, we draw some further comparison between the two. Both methods rely on finding a good proposal by searching a region far from the current state. One key difference lies in the computing efficiency. Although HMC is more generally applicable beyond data augmentation, it is computationally intensive since Hamiltonian dynamics often requires multiple numeric steps. CDA only requires one step of calibrated Gibbs sampling, which is often much more efficient leveraging on existing data augmentation algorithms.
|
\section{Introduction}
The physics of compact stars have always drawn the interest to the researchers. The theoretical idea of black holes as exotic objects whose gravitational fields are so strong that even light cannot escape have originated as early as in the $18$th century. But round that time it remained only as a mere textbook curiosity. Compact objects as real astrophysical objects where only considered in the early $1930$'s. Chadwick's discovery of a neutron in 1932 sparked renewed interest in the gravitational collapse of compact stars, following which Baade and Zwicky suggested the notion of a neutron star while investigating supernovae observations. Most importantly Einstein's theory of general relativity laid the foundation of our understanding of compact stars. Early research works on the structure of compact objects have been done by Oppenheimer, Volkoff and Tolman in $1939$ \cite{Oppenheimer,Tolman}. They successfully derived the equations of relativistic stellar structures from Einstein's field equations and discovered the existence of a limit to the mass of stable degenerate relativistic stars.\par
In the past few years cooperative efforts of radio and optical astronomers have revealed a great deal of observational data which in turn has provided physicists new theoretical insights into the physics of compact objects. Fascinating new observational data of the stellar objects Her X-1, Cyg X-2, 4U 1820-30, SAX J 1808.4-3658, 4U 1728-34, PSR 0943+10 and RX J185635-3754 have further contributed to this area of research. More sophisticated data have challenged us to investigate more precise and complicated approach to stellar modeling. In comparison to normal stars of comparable mass, compact objects have much smaller radii and hence, much stronger surface gravity. The interior structures of compact objects have extremely high densities and involve phases which are not yet clearly understood. Such high densities of the order of $10^{15}g~cm^{-3}$ or even higher, as suggested by some recent developments, causes the nuclear fluids in the interior of a compact star to be anisotropic as suggested by Ruderman \cite{Ruderman}. For an anisotropic fluid, the radial and tangential pressures are not equal (i.e. $p_r \neq p_t$) and the anisotropy factor $\Delta = p_t - p_r$ increases rapidly with the increase in radial distance but vanishes at the centre of our stellar model. The cause of anisotropy may be due to a variety of reasons like presence of a solid core or type $3A$ fluid or type $P$ superfluid \cite{Kippenhahn} or may result from different kind of phase transition, rotation, magnetic stress, pion condensation etc. \cite{Sokolov,Sawyer}. In an early work Ponce de Le\'{o}n \cite{Ponce} obtained two new exact analytical solutions to Einstein's field equations for static fluid sphere with anisotropic pressures. In another notable work Herrera \& Santos \cite{Herrera} provided an exhaustive review on the subject of anisotropic fluids. B\"{o}hmer and Harko \cite{Bohmer} obtained upper and lower bounds of the physical parameters like mass-radius ratio, anisotropy, red-shift and total energy for anisotropic fluid in the presence of the cosmological constant. In addition to that, we would like to mention the works of Mak and Harko \cite{Mak} and Sharma and Maharaj \cite{Sharma} which suggest that anisotropy is a sufficient condition in the study of dense nuclear matter. By assuming pressure anisotropy Bhar \cite{pb1,pb2,pb3} obtained anisotropic model of compact star in $(3+1)D$ space-time. Bhar {\em et al.} \cite{pb4} studied the behavior of static spherically symmetric relativistic objects with locally anisotropic matter distribution considering the Tolman VII form for the gravitational potential $g_{rr}$ in curvature coordinates together with the linear relation between the energy density and the radial pressure. A new class of interior solutions for anisotropic stars admitting conformal motion in higher dimensional non-commutative space-time by choosing a particular density distribution function of Lorentzian type as provided by Nazari and Mehdipour \cite{meh} was obtained by Bhar et al \cite{pb5}. A large number of works on the modeling of compact object in the platform of KB metric were done by several researchers \cite{varela,kalam,hos,rah}.\par
In this paper we have investigated a new relativistic model of compact star of embedding class one. Here we have used the well known fact that the {\it n} dimensional manifold {\it $V_n$} can be embedded in a pseudo-Euclidean space of {\it m} dimensions where $ m=n(n+1)/2$. The embedding class of {$V_n$} is the minimum number $m-n$ of extra dimensions which needs to be added to manifold {\it $V_n$}. Note that in a metric of embedding class one the metric functions $\lambda$ and $\nu$ are dependent on each other. Due to such exceptional relationship, it is possible to generate either Schwarzschild interior solution \cite{sch1} or Kohlar-Chao solution \cite{koh1} for a neutral perfect fluid distribution. However, a very recently Bhar et al. \cite{Piyali1} and Maurya et al. \cite{Maurya1,Maurya2,Maurya3} have investigated an anisotropic compact star model of embedding class one. Some other notable works in this field have been done extensively by Singh et al.\cite{Singh1,Singh2,Singh3,sing1,sing2,sing3}.\par
The paper has been divided in the following sections; in section $2$ the field equations have been solved using the Karmakar condition.In section $3$ a particular model have been discussed, next in section $4$ we have worked on various physical acceptability conditions has must satisfied by our model. The exterior space-time and boundary conditions have been investigated in section $5$. After that in section $6$ the different properties of the solution have been discussed in details. Next in section $7$ we have analyzed the stability of our stellar model in various conditions. Lastly in section $8$ we have discussed some concluding remarks.
\section{Field Equations and Karmarkar's Condition}
In the canonical coordinate $(x^{\mu})\equiv (t,~r,~\theta,~\phi)$, the interior of a static and spherically symmetry object is described by line element
\begin{equation}\label{line}
ds^{2}=e^{\nu(r)}dt^{2}-e^{\lambda(r)}dr^{2}-r^{2}\left(d\theta^{2}+\sin^{2}\theta d\phi^{2} \right)
\end{equation}
Where $\nu$ and $\lambda$ are functions of the radial coordinate `$r$' only.\par
The Einstein field equations for anisotropic fluid distribution are given as (in the unit $G=c=1$)
\begin{equation}
-8\pi T^\mu_\xi = \mathcal{R}^\mu_\xi-{1\over 2}\mathcal{R}~g^\mu_\xi
\end{equation}
where
\begin{eqnarray}
T^\mu_\xi & = & \rho v^\mu v_\xi + p_r \chi_\xi \chi^\mu + p_t(v^\mu v_\xi -\chi_\xi \chi^\mu-g^\mu_\xi) ~,\label{t}
\end{eqnarray}
the energy-momentum tensor, $\mathcal{R}^\mu_\xi$ is the Ricci tensor, $\mathcal{R}$ is the scalar curvature, $p_r$ and $p_t$ denote radial and transverse pressures respectively, $\rho$ the density distribution , $v^\mu$ the four velocity and $\chi^\mu$ is the unit space-like vector in the radial direction.\par
Now for the line element (\ref{line}) and the matter distribution (\ref{t}) in Einstein Field equations (assuming $G=c=1$) take this form,
\begin{eqnarray}
\frac{1-e^{-\lambda}}{r^{2}}+\frac{e^{-\lambda}\lambda'}{r}&=&8\pi\rho,\label{f1}\\
\frac{e^{-\lambda}-1}{r^{2}}+\frac{e^{-\lambda}\nu'}{r}&=&8\pi p_{r},\label{f2}\\
e^{-\lambda}\left(\frac{\nu''}{2}+\frac{\nu'^{2}}{4}-\frac{\nu'\lambda'}{4}+\frac{\nu'-\lambda'}{2r} \right)&=&8\pi p_t.\label{f3}
\end{eqnarray}
where $`\prime'$ represents differentiation with respect to the radial coordinate $r$. Using the Eqs. (\ref{f2}) and (\ref{f3}) we obtain the anisotropy parameter
\begin{eqnarray}
\Delta&=&p_t-p_r \nonumber\\
&&= \frac{e^{-\lambda}}{8\pi}\left[{\nu'' \over 2}-{\lambda' \nu' \over 4}+{\nu'^2 \over 4}-{\nu'+\lambda' \over 2r}+{e^\lambda-1 \over r^2}\right]\nonumber\\\label{del}
\end{eqnarray}
Now if the space-time (\ref{line}) satisfies the Karmarkar condition \cite{kar}
\begin{equation}
\mathcal{R}_{1414}=\frac{\mathcal{R}_{1212}\mathcal{R}_{3434}+ \mathcal{R}_{1224}\mathcal{R}_{1334}}{\mathcal{R}_{2323}}\label{con}
\end{equation}
with $\mathcal{R}_{2323}\neq 0$ \cite{pandey}, it represents the space time of embedding class $1$.\\
Now the above components of $\mathcal{R}_{hijk}$ for metric (\ref{line}) are:
\begin{eqnarray}
\mathcal{R}_{2323}&=&\,r^2\,\sin^{2}\theta~\left[1-e^{-\lambda}\right] \nonumber\\
\mathcal{R}_{1212}&=&\frac{1}{2}\,\lambda'\,r \nonumber\\
\mathcal{R}_{1334} &=& R_{1224} \sin^2 \theta = 0 \nonumber\\
\mathcal{R}_{1414}&=& -e^{\nu}\left[\frac{1}{2}~\nu''+\frac{1}{4}~{\nu'}^{2}-\frac{1}{4}~\lambda'~\nu'\right] \nonumber\\
\mathcal{R}_{3434}&=& -\frac{r}{2}\sin^{2}\theta~ \nu'e^{\nu-\lambda} \nonumber
\end{eqnarray}
Using the above expression eq.(\ref{con}) gives the following differential equation
\begin{equation}
{2\nu'' \over \nu'}+\nu'={\lambda' e^\lambda \over e^\lambda-1}\label{dif1}
\end{equation}
On integration we get the following relationship between the metric potentials $\nu$ and $\lambda$ as
\begin{equation}
e^{\nu}=\left(A+B\int \sqrt{e^{\lambda}-1}~dr\right)^2\label{nu1}
\end{equation}
where $A$ and $B$ are constants of integration.
By using (\ref{nu1}) we can rewrite (\ref{del}) as
\begin{eqnarray}
\Delta = {\nu' \over 4e^\lambda}\left[{2\over r}-{\lambda' \over e^\lambda-1}\right]~\left[{\nu' e^\nu \over 2rB^2}-1\right] \label{del1}
\end{eqnarray}
We have to solve the Einstein field equations (\ref{f1})-(\ref{f3}) with the help of equation (\ref{nu1}). One can notice that we have four equations with $5$ unknowns namely $\lambda,\nu,\rho,p_r$ and $p_t$.
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{metric}
\end{minipage}
\caption{The metric potentials are plotted against $r$ by taking $a=0.0.08,\,b=-0.00098,\,A=-0.786839,\,B=0.0369469$ for the compact star 4U 1538-52}\label{metric}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{rho}
\end{minipage}
\caption{Matter density is plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{rho1}
\end{figure}
\section{A particular Model}
To generate the model let us assume the expression for the $g_{rr}$ metric
potential as,
\begin{equation}\label{elambda}
e^{\lambda}=1 + \frac{a^2 r^2}{(1 + b r^2)^4}
\end{equation}
Where $a$ and $b$ are constants having the dimensions {\em length$^{-1}$} and {\em length$^{-2}$} respectively. In a very recent work Singh \& Pant \cite{sing1} obtained a model of compact star of embedding class I by choosing the metric potential as $e^{\lambda}=1 + a^2 r^2(1 + b r^2)^n$. They have only analyze the results for positive values of $n$. In our present paper we follow the above {\em ansatz} with $n=-4$.\\
Solving eqns. (\ref{nu1}) and (\ref{elambda}) we obtain the expression for the metric coefficient $e^{\nu}$ as,
\begin{equation}\label{enu}
e^{\nu}=\left[A -\frac{a B}{2 b (1 + b r^2)}\right]^2
\end{equation}
Using eqns.(\ref{elambda}) and (\ref{enu})the expressions for matter density, radial and transverse pressure are obtained as,
\begin{eqnarray}
8\pi \rho&=&\frac{a^2\big[a^2 r^2 +(1 + b r^2)^3 (3-5 b r^2)\big]}{\big[a^2 r^2 + (1 +
b r^2)^4\big]^2} \label{rho}\\
8\pi p_r&=&\frac{a \big[a^2 B - 2 a A b (1 + b r^2) + 4 b B (1 + b r^2)^3\big]}{\big[
2 A b (1 + b r^2)-aB\big]\big[a^2 r^2 + (1 + b r^2)^4\big]} \label{pr}\\
8\pi p_t&=&\frac{a (1 + b r^2)^3}{\big[
2 A b (1 + b r^2)-a B\big]\big[a^2 r^2 + (1 + b r^2)^4\big]^2}\times\nonumber\\
&&\bigg[a^2 B (1 - b r^2) +
4 b B f_1(r)-
2 a A b f_2(r)\bigg]\label{pt}
\end{eqnarray}
where,
\begin{eqnarray*}
f_1(r)&=& (1 - b r^2) (1 + b r^2)^3\\
f_2(r)&=&(1 + b r^2) (1 - 3 b r^2)
\end{eqnarray*}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{pr}
\end{minipage}
\caption{The radial and transverse pressure are plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{pr1}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{delta}
\end{minipage}
\caption{The anisotropic factor is plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{delta}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{dp}
\end{minipage}
\caption{The density and pressure gradients are plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{dp}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{ratio}
\end{minipage}
\caption{$p_r/\rho$ and $p_t/\rho$ are plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{ratio}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{trace}
\end{minipage}
\caption{$(p_r+2p_t)/\rho$ is plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{trace}
\end{figure}
\section{Physical acceptability conditions}\label{Sec4}
For the well behaved nature of the solution, the following conditions should be satisfied \cite{abreu}:
\begin{description}
\item[(i)] The metric potentials should be free from singularities inside the radius of the star moreover the fluid sphere should satisfy $e^{\nu(0)}=$ constant, and $e^{-\lambda(0)}=1$.
\item[(ii)] The density $\rho$ and pressures $p_r,\,p_t$ should be positive inside the fluid configuration and should decreasing outward.
\item[(iii)] The radial pressure $p_r$ must be vanishing but the tangential pressure $p_t$ may not necessarily vanish at the boundary $r=r_\Sigma$. However, the radial pressure is equal to the tangential pressure at the center of the fluid sphere, i.e., pressure anisotropy vanishes at the center, $\Delta(0)=0$ \cite{bl,iva} and $\displaystyle\Delta(r=r_\Sigma)=p_t(r_\Sigma)>0$ \cite{bh}.
\item[(iv)] The radial pressure gradient $dp_r/dr\leq0$~for $0\leq r \leq r_\Sigma$.
\item[(v)] The density gradient $d\rho/dr\leq0$ for $0\leq r\leq r_\Sigma$.
\item[(vi)] A physically acceptable fluid sphere must satisfy the causality conditions, the radial and tangential adiabatic speeds of sound should less than the speed of light. In the unit $c=1$ the causality conditions take the form $0<v_{sr}^{2}=dp_r/d\rho\leq 1$ and $0<v_{st}^{2}=dp_t/d\rho\leq 1$.
\item[(vii)] The interior solution should satisfy either
\begin{itemize}
\item strong energy condition (SEC) $\rho-p_r-2p_t\geq0,\,\rho-p_r\geq0,\,\rho-p_t\geq0$ or
\item dominant energy condition (DEC) $\rho\geq p_r$ and $\rho\geq p_t$.
\end{itemize}
\item[(viii)] The interior solution should continuously match with the exterior Schwarzschild solution.
\end{description}
Conditions (iv) and (v) imply that pressure and density should be maximum at the center and monotonically decreasing towards the surface.\par
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{ec}
\end{minipage}
\caption{The energy conditions are potted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{ec}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{mass}
\end{minipage}
\caption{The mass function is plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{mass1}
\end{figure}
\section{Exterior space-time and boundary condition}
To fix the values of the constants $a,\,b,\,A$ and $B$ we match our interior spacetime to the exterior Schwarzschild line element given by,
\begin{eqnarray}
ds^{2}&=&\left(1-\frac{2m}{r}\right)dt^{2}-\left(1-\frac{2m}{r}\right)^{-1}dr^{2}\nonumber\\
&&-r^{2}(d\theta^{2}+\sin^{2}\theta d\phi^{2})
\end{eqnarray}
outside the event horizon $r>2m$, $m$ being the mass of the black hole.\par
Using the continuity of the metric coefficients $e^{\nu},e^{\lambda}$ across the boundary we get the following three equations
\begin{eqnarray}
1 + \frac{a^2 r_{\Sigma}^2}{(1 + b r_{\Sigma}^2)^4}&=&\left(1-\frac{2m}{r_{\Sigma}}\right)^{-1} \label{b1}\\
1-\frac{2m}{r_{\Sigma}}&=&\left[A -\frac{a B}{2 b (1 + b r_{\Sigma}^2)}\right]^2 \label{b2}
\end{eqnarray}
and $p_r(r=r_{\Sigma})=0$ gives,
\begin{equation}\label{b3}
a^2 B - 2 a A b (1 + b r_{\Sigma}^2) + 4 b B (1 + b r_{\Sigma}^2)^3=0
\end{equation}
Solving eqns. (\ref{b1})-(\ref{b3})we get,
\begin{eqnarray}
a&=& \frac{(1+br_{\Sigma}^2)^2}{r_{\Sigma}}\sqrt{\frac{2m/r_{\Sigma}}{1-2m/r_{\Sigma}}} \\
B^2&=& \frac{m}{2r_{\Sigma}^3} \\
A&=& \frac{\big[a^2+4b(1+br_{\Sigma}^2)^3\big] B}{2ab(1+br_{\Sigma}^2)}
\end{eqnarray}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{zs}
\end{minipage}
\caption{The compactness factor and surface redshift are plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{zs}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{z}
\end{minipage}
\caption{The gravitational redshift factor is plotted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{z}
\end{figure}
\section{Properties of the solution}
\subsection{Singularity free nature of the model parameter}
At the center of the star the expressions for metric potentials are obtained as,
\[e^{\lambda}|_{r=0}=1,\, e^{\nu}|_{r=0}=\left(A-\frac{aB}{2b}\right)^2\]
which are constants and
\[(e^{\lambda})'|_{r=0}=0,\, (e^{\nu})'|_{r=0}=0.\]
The central density and central pressure are obtained as,
\begin{eqnarray}
8\pi \rho_c&=&3a^2\\
\label{pc}8\pi p_r(r=0)=8\pi p_t(r=0)&=&\frac{a \big[a^2 B - 2 a A b + 4 b B \big]}{\big[
2 A b -aB\big]}\nonumber\\
\end{eqnarray}
To satisfy Zeldovich's condition at the interior, $p_r/\rho$ at
the center must be $\leq 1$. Therefore
\begin{equation}\label{zel}
\frac{a^2 B - 2 a A b + 4 b B}{3a\big[
2 A b -aB\big]}\leq1
\end{equation}
From eqns. (\ref{pc}) and (\ref{zel}) we get,
\begin{eqnarray}
\frac{a^2+b}{2ab} \leq \frac{A}{B} < \frac{a^2+4b}{2ab}
\end{eqnarray}
Differentiating eqns. (\ref{rho})-(\ref{pt}) we get the density and pressure gradient as,
\begin{eqnarray}
8\pi\frac{d\rho}{dr}&=&-\frac{2 a^2 r f_3(r)}{(a^2 r^2 + (1 + b r^2)^4)^3}\label{drho}\\
8\pi\frac{dp_r}{dr}&=&\frac{2ar}{[a^2 r^2+(1 + b r^2)^4]^2 \{a B -
2 A b (1 + b r^2)\}^2}\times \nonumber\\
&&\bigg[a^5 B^2 - 4 a^4 A b B (1 + b r^2) - 16 A b^3 B (1 + b r^2)^7\nonumber\\
&&- 8 a^2 A b^2B f_4(r)
+4 a^3 b (1 + b r^2)^2 f_5(r) +\nonumber\\
&&4 a b^2 (1 + b r^2)^5 f_6(r) \bigg]\label{dpr}\\
8\pi\frac{dp_t}{dr}&=&\frac{4 a r (1 +
b r^2)^2}{\big[a B -
2 A b (1 + b r^2)\big]^2 \big[a^2 r^2 + (1 + b r^2)^4\big]^3}\times\nonumber\\
&&\bigg[-8 A b^3 B (2-b r^2) (1 + b r^2)^8 -
2 a^2 A b^2 B f_7(r) \nonumber \\
&& + a^5 B^2 \big[1 - b r^2 (1-b r^2)\big] -
2 a b^2 (1 + b r^2)^6 f_8(r) \nonumber \\
&&-2 a^4 A b B (1 + b r^2)\big[2-b r^2(3-5 b r^2)\big] \nonumber\\
&&+a^3 b (1 + b r^2)^2 f_9(r)\bigg]\label{dpt}
\end{eqnarray}
where,
\begin{eqnarray*}
f_3(r)&=&a^4 r^2+
20 b (1-b r^2) (1 + b r^2)^6 +\\
&&(a + a b r^2)^2[5-
b r^2 (2-17 b r^2)]\\
f_4(r)&=&(1 + b r^2)^3 (3 + b r^2)\\
f_5(r)&=& A^2 b + B^2 (2 - b r^2)\\
f_6(r)&=&4 A^2 b + B^2 (1 + b r^2)\\
f_7(r)&=&(1 + b r^2)^4 \big[11 - b r^2(8-b r^2)\big]\\
f_8(r)&=&B^2 (-3 + b r^2) (1 + b r^2) +
4 A^2 b (-2 + 3 b r^2)\\
f_9(r)&=&4 A^2 b (1- b r^2(2-3 b r^2))\\
&&+B^2(1 + b r^2)\big[7-b r^2(9-8 b r^2)\big]
\end{eqnarray*}
The profiles of the metric coefficients, matter density, radial and transverse pressure are plotted in Figs. \ref{metric}, \ref{rho1} and fig. \ref{pr1} respectively which shows that metric coefficients, $\rho$, $p_r$ and $p_t$ all are positive and free from central singularity. The radial pressure $p_r$ vanishes at the boundary of the star whereas the transverse pressure and matter density are still positive inside the stellar interior as well as at the boundary. $\rho,\,p_r$ and $p_t$ all are monotonic decreasing function of $r$. The monotonic decreasing condition is verified by Fig.~\ref{dp}. The anisotropic factor is plotted against $r$ in Fig.~\ref{delta}. The profile shows that $\Delta>0$ inside the stellar configuration and therefore the anisotropic force is repulsive in nature and according to Gokhroo and Mehra \cite{gm} it is necessary to construct the compact object. Both $p_r/\rho$ and $p_t/\rho$ are monotonic decreasing function of r and lies in the range $0<p_r/\rho,\,p_t/\rho<1$ (Fig.\ref{ratio}) which verifies that the underlying fluid distribution is non-exotic in nature. $(p_r+2p_t)/\rho$ is plotted against $r$ in fig. \ref{trace}, which is monotonic decreasing function of $r$ and $0<(p_r+2p_t)/\rho<1.$ All the energy conditions are satisfied by our present model which is shown in fig.~\ref{ec}.
\subsection{Mass-radius relation}
Using the relationship $e^{-\lambda}=1-\frac{2m}{r}$ the mass function is obtained as
\begin{equation}\label{mass}
m(r)=\frac{a^2r^3}{2\big[(1+br^2)^4+a^2r^2\big]},
\end{equation}
the compactness factor and surface red-shift are obtained as,
\begin{eqnarray}
u(r) &=& \frac{a^2r^2}{2\big[(1+br^2)^4+a^2r^2\big]} \\
z_s &=& \Big[\frac{(1+br_\Sigma^2)^4}{(1+br_\Sigma^2)^4+a^2r_\Sigma^2}\Big]^{-\frac{1}{2}}-1
\end{eqnarray}
The gravitational red-shift of the stellar configuration is given by
\begin{eqnarray}
z &=& e^{-\nu/2}-1 = \left[A - \frac{a B}{2 b (1 + b r^2)}\right]^{-1}-1
\end{eqnarray}
The profile of mass function is plotted in fig.~\ref{mass1}, which is monotonic increasing in nature and regular at the center of the star as well as positive inside the stellar interior. Following the concept of Buchdahl \cite{buch} for a compact star the ratio of mass to the radius can not be arbitrarily large. The maximally allowable mass-to-radius ratio $(M/r_{\Sigma})$ for an isotropic fluid sphere should lie in the range $ 2M/r_{\Sigma} <8/9$ (in the units $c=G=1$). The compactification factor and surface red-shift are plotted in figs. \ref{zs}. Both the profiles are monotonic increasing in nature. On the other hand the profile of gravitational red-shift is plotted against $r$ in fig.~\ref{z}, which is monotonically decreasing in nature.
\section{Stability}
\subsection{Stability under three different forces}
For checking the static equilibrium of our present model under three forces $viz$ gravitational force, hydrostatics force and anisotropic force we consider the following equation
\begin{equation}\label{tov1}
-\frac{M_G(r)(\rho+p_r)}{r}e^{\frac{\nu-\lambda}{2}}-\frac{dp_r}{dr}+\frac{2}{r}(p_t-p_r)=0,
\end{equation}
proposed by Tolman-Oppenheimer-Volkoff and called the TOV equation and $M_G(r)$ can derived from the Tolman-Whittaker
formula and the Einstein's field equations and is defined by
\begin{equation}
M_G(r)=\frac{1}{2}re^{\frac{\lambda-\nu}{2}}~\nu'
\end{equation}
represents the gravitational mass within the radius $r$.
Plugging the expression of $M_G(r)$ in equation $(\ref{tov1})$, we obtain
\begin{equation}\label{tov2}
-\frac{\nu'}{2}(\rho+p_r)-\frac{dp_r}{dr}+\frac{2}{r}(p_t-p_r)=0.
\end{equation}
The above equation can be rewritten as,
\begin{equation}
F_g+F_h+F_a=0,
\end{equation}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{tov}
\end{minipage}
\caption{The variation of three different forces acting on the system are potted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{tov}
\end{figure}
where $F_g, F_h$ and $F_a$ represents the gravitational, hydrostatics and anisotropic forces respectively.
Using the Eqs. (\ref{rho})-(\ref{pt}), the expression for $F_g,F_h$ and $F_a$ can be written as,
\begin{eqnarray}
F_g &=&-\frac{\nu'}{2}(\rho+p_r)\nonumber\\
&=& -4 b B r (a + a b r^2)^2 \frac{f_{10}(r)}{f_{11}(r)}\\
F_h &=&-\frac{dp_r}{dr}\\
F_a &=&\frac{2\Delta}{r}
\end{eqnarray}
where,
\begin{eqnarray*}
f_{10}(r)&=&2 b B (1 + b r^2)^4 - 2 a A b (1 + b r^2) (-1 + 3 b r^2) \\
&&+a^2 B (-1 + 5 b r^2)\\
f_{11}(r)&=&[a B - 2 A b (1 + b r^2)]^2 [a^2 r^2 + (1 + b r^2)^4]^2
\end{eqnarray*}
The profile of three different forces are plotted in fig. \ref{tov}. The figure shows that hydrostatics and anisotropic force are positive and is dominated by the gravitational force which is negative to keep the system in static equilibrium.
\subsection{Relativistic adiabatic index}
For a relativistic anisotropic sphere the stability is related to the adiabatic index $\Gamma$, the ratio of two specific heats, defined by Chan et al.\cite{chan1},
\begin{equation}
\Gamma_r=\frac{\rho+p_r}{p_r}\frac{dp_r}{d\rho} ~~;~~ \Gamma_t=\frac{\rho+p_t}{p_t}\frac{dp_t}{d\rho}
\end{equation}
Now $\Gamma>4/3$ gives the condition for the stability of a Newtonian sphere and $\Gamma =4/3$ being the condition for a neutral equilibrium proposed by Bondi \cite{bondi64}. This condition changes for a relativistic isotropic sphere due to the regenerative effect of pressure, which renders the sphere more unstable. For an anisotropic general relativistic sphere the situation becomes more complicated, because the stability will depend on the type of anisotropy.
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{gamma}
\end{minipage}
\caption{The relativistic adiabatic indices are potted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{f9}
\end{figure}
\subsection{Causality condition}
The radial and tangential speeds of sound for our model of compact star are obtained as,
\begin{eqnarray}
v_r^2 &=& \frac{dp_r}{d\rho}=\Big(\frac{dp_r}{dr}\Big)/\Big(\frac{d\rho}{dr}\Big) \\
v_t^2 &=& \frac{dp_t}{d\rho}=\Big(\frac{dp_t}{dr}\Big)/\Big(\frac{d\rho}{dr}\Big)
\end{eqnarray}
\begin{figure}[!htb]\centering
\begin {minipage}{0.5\textwidth}
\includegraphics[width=\linewidth]{sv}
\end{minipage}
\caption{The stability factor are potted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{sv}
\end{figure}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{sv1}
\end{minipage}
\caption{The stability factor are potted against $r$ for the values mentioned in Fig. \ref{metric}.}\label{sv1}
\end{figure}
Where the expression for $\frac{d\rho}{dr},\,\frac{dp_r}{dr}$ and $\frac{dp_t}{dr}$ are shown in the expressions (\ref{drho})-(\ref{dpt}). For a physically acceptable model of relativistic anisotropic star the radial and transverse velocity of sound should lie in the range $0<v_r^2,v_t^2<1$ which is known as causality condition. With the help of graphical representation we have shown that our model satisfies the causality condition (please refer fig. \ref{sv})
To check the stability of anisotropic stars under the radial perturbations Herrera \cite{her} introduced the concept of
``cracking" and using this concept of cracking it was proved by Abreu et al. \cite{abreu} that the region of an anisotropic fluid sphere where $-1 \leq v_t^2-v_r^2\leq0$ is potentially stable but the region where $0 < v_t^2-v_r^2 \leq 1$ is potentially unstable. From Fig.\ref{sv} it is clear that our model satisfies this condition. So we conclude that our model is potentially stable. Moreover $0 <v_r^2 < 1$ and $0 <v_t^2 < 1$ therefore according to Andr\'{e}asson\cite{andre} $|v_r^2 - v_t^2|<1$ which is also clear from Fig.\ref{sv1}.
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{m-rh-c}
\end{minipage}
\caption{The variation mass with central density (0$-6.74\times 10^{17}~g/cm^3$) for the values mentioned in Fig. \ref{metric} and by assuming $R=7.8~km$. The mass becomes saturated to 3.9 $M_\odot$ when the $\rho_c$ reaches about $4.04\times 10^{17}~g/cm^3$ and onward.}\label{mrho}
\end{figure}
\subsection{Harrison-Zeldovich-Novikov static stability criterion}
Chandrasekhar \cite{chan64}, Harrison \cite{har65} etc. determined the eigen-frequencies of all the fundamental modes to analyzed the stability of stars. However, Harrison \cite{har65} and Zeldovich \& Novikov \cite{zel} simplify into much simpler calculations and reduced it to much more simpler formalism. They have assumed that the adiabatic index of a pulsating star is same as in a slowly deformed matter. This leads to a stable configuration only if the mass of the star is increasing with central density i.e. $dM/d\rho_c > 0$ and unstable if $dM/d\rho_c < 0$ or $dM/d\rho_c = 0$.
In our solution, the mass as a function of central density can be written as
\begin{equation}
M = {8\pi \rho_c R^3/3 \over 2\big(8\pi \rho_c R^2/3+\{1+bR^2\}^4 \big)}\label{mrho1}
\end{equation}
which gives us (for a given radius)
\begin{eqnarray}
{d M \over d \rho_c} &=& 12 \pi R^3 (b R^2+1)^4 \Big[3 b^4 R^8+12 b^3 R^6+18 b^2 R^4 \nonumber \\
&& +12 b R^2+8 \pi R^2 \rho_c+3\Big]^{-2}> 0.
\end{eqnarray}
This condition can be further confirmed by Fig. \ref{mrho}. However, when the mass becomes saturated with increase in central density i.e. $dM/d\rho_c=0$ or decrease with increase in $\rho_c$ i.e. $dM/d\rho_c<0$, the unstable configuration is triggered.
\begin{table}[h]
\tbl{The values of the constants calculated from our model for few well-know compact star candidates.}
{\begin{tabular}{@{}cccccc@{}} \toprule
Compact Star & $a$ &$b$&$A$&$B$ \\
\colrule
4U~1538-52&0.08&-0.00098&-0.786839&0.0369469 \\
PSR~J1614-2230 &0.10&-0.00116&-1.75292&0.0539164 \\
Vela~X-1 &0.09&-0.00125&-1.13613&0.0489534 \\
Cen-X3 &0.09&-0.00088&-1.60872&0.0444594\\
\botrule
\end{tabular} \label{ta1}}
\end{table}
\begin{table}[h]
\tbl{Optimization of masses and radii of few well-know compact star candidates.}
{\begin{tabular}{@{}cccccc@{}} \toprule
Compact Star &$R~$(km) & $M/M_{\odot}$ &R& $M/M_{\odot}$&References \\
&observed&observed&calculated&calculated\\
\colrule
4U~1538-52&$7.866\pm0.21$&$0.87\pm0.07$&7.8&0.88&Gangopadhyay et al.\cite{gan} \\
PSR~J1614-2230 &$9.69\pm0.2$&$1.97\pm0.04$&9.7& 1.97&Gangopadhyay et al.\cite{gan} \\
Vela~X-1 &$9.56\pm0.08$&$1.77\pm0.08$&9.56 & 1.77&Gangopadhyay et al.\cite{gan}\\
Cen-X3 &$9.178\pm0.13$&$1.49\pm0.08$&9.18&1.5&Gangopadhyay et al.\cite{gan}\\
\botrule
\end{tabular} \label{ta1}}
\end{table}
\begin{table}[h]
\tbl{central density, surface density, central pressure, compactness and surface redshift of few well-know compact star candidates calculated from our model.}
{\begin{tabular}{@{}cccccc@{}} \toprule
Compact Star & central density &Surface density& central pressure & compactness & surface redshift\\
& $gm/cm^3$ &$gm/cm^3$& $dyne/cm^{2}$ & & \\
\colrule
4U 1538-52& $1.03\times10^{15}$ &$7.84\times10^{14}$&$8.192\times10^{35}$&0.166 &0.224\\
PSR~ J1614-2230& $1.611\times10^{15}$&$7.502\times10^{14}$&$4.42\times10^{35}$& 0.299&0.579\\
Vela~X-12 & $1.30\times10^{15}$&$7.62\times10^{14}$&$2.88\times10^{35}$& 0.273&0.484\\
Cen X - 3&$1.30\times10^{15}$&$7.26\times10^{14}$&$1.90\times10^{35}$&0.241&0.389\\
\botrule
\end{tabular} \label{ta1}}
\end{table}
\begin{figure}[!htb]\centering
\begin {minipage}{0.6\textwidth}
\includegraphics[width=\linewidth]{eos}
\end{minipage}
\caption{The variation of radial and transverse pressure with respect to density are plotted against $r$. }\label{eos1}
\end{figure}
\section{Discussion}
A perfect fluid distribution satisfying Karmarkar condition is either given by the Schwarzschild interior solution \cite{sch1} or Kohlar-Chao solution \cite{koh1}. To avoid such a perfect fluid distribution we have taken a different form of metric potential $e^\lambda$ along with some anisotropy in pressure and then proceeded with the Karmarkar condition. Some of the key features of the anisotropic stellar model are as follows:
\begin{description}
\item[(i)] The metric potentials are regular and monotone increasing away from the center of the star (Fig. \ref{metric}). They are free from singularities inside the radius of the star and moreover the fluid sphere satisfy $e^{\nu}(r=0)=$ constant, and $e^{-\lambda}(r=0)=1$ at the stellar center.
\item[(ii)] From the profiles of density $\rho$ and pressures $p_r,~p_t$ in Figs. \ref{rho1} and fig. \ref{pr1} respectively, we have found that they are positive inside the fluid configuration and free from any central singularity.
\item[(iii)] The radial pressure $p_r$ vanishes at the boundary of the star but the tangential pressure $P_t$ remain positive at the stellar center and at the boundary $r=r_\Sigma$. However, the radial pressure is equal to the tangential pressure at the center of the fluid sphere, i.e., pressure anisotropy factor vanishes at the center, $\Delta(r=0)=0$ as is evident from the profile of anisotropy $\Delta$ in Fig.~\ref{delta}. Also one can note from it that $\Delta>0$ inside the stellar configuration causing the anisotropic force to be repulsive in nature.
\item[(iv)] The radial pressure gradient $dp_r/dr\leq0$~for $0\leq r \leq r_\Sigma$. The density gradient $d\rho/dr\leq0$ for $0\leq r\leq r_\Sigma$. Both $p_r/\rho$ and $p_t/\rho$ are monotonic decreasing function of r and lies in the range
$0<p_r/\rho,\,p_t/\rho<1$ (fig.~\ref{ratio}) which verifies that the underlying fluid distribution is non-exotic in nature.
\item[(v)] All the energy conditions are satisfied by our present model which is shown in fig.~\ref{ec}. The ratio of $(p_r + 2p_t)/\rho$ plotted in Fig.~\ref{trace} is monotonic decreasing function of $r$ and less than $1$.
\item[(vi)] The profiles of mass function suggest that it is positive, regular and monotone increasing in nature.The compactification factor and surface red-shift are monotone increasing while the profile of gravitational red-shift suggest that it is monotone decreasing.
\item[(vii)] The mass of the stellar object increases with the increase in central density (for a given radius) or equivalently $dM/d\rho_c<0$ and thus the static stability criterion hold good till $\rho_c=4.04\times 10^{17}~g/cm^3$. This imply that our solution is not only well-behaved but also represents static and stable. The mass becomes saturated to 3.9 $M_\odot$ when the $\rho_c$ reaches about $4.04\times 10^{17}~g/cm^3$ and onward. Hence the region where $\rho_c>4.04\times 10^{17}~g/cm^3$ satisfy $dM/d\rho=0$ and thus belong to unstable region.
\end{description}
One can note that to solve the Einstein's field equations we have not chosen a particular equation of state. Instead we have assumed $g_{rr}$ and used Karmakar condition to find the another metric coefficient. Then by plugging the expressions of the metric coefficients into the Einstein's field equations we are able to find the expressions for $\rho,\,p_r$ and $p_t$ using an algebraic calculation without solving any differential equations. In the earlier works where the researchers used the KB metric to model a compact stars also found the the expressions for $\rho,\,p_r$ and $p_t$ using an algebraic calculations only \cite{pb1,pb2,varela,kalam,hos,rah}. From the expressions of $\rho,\,p_r$ and $p_t$ we can't obtain a particular equation of state , i.e., a relationship between the radial pressure $p_r$ and the matter density $\rho$ or between the transverse pressure $p_t$ and the matter density $\rho$. To see the trend of the EoS, we plot a graph for $p_r$ and $p_t$ vs $\rho$ in Fig.~\ref{eos1}.\par
Throughout the paper all the plots are drawn for the compact star 4U 1538-52. We have calculated the values of the constants for some well-known compact stars which are presented in table~1. Recently an improved
method was used by Rawls et al.\cite{raw} for determining the mass of neutron stars such as
Vela X-1, SMC X-1, Cen X-3 etc. in eclipsing X-ray pulsar
binaries. To analyze the published
data for these systems they used a numerical code based on Roche geometry
with various optimizers which they supplemented with
new spectroscopic and photometric data for $4U 1538-52$. This allows them to calculate an improved value
for the neutron star masses. Their derived values are $(0.87\pm0.07 M_{\odot})$ for 4U 1538-52
$(1.77\pm.08M_{\odot})$ for Vela X-1 and $(1.49\pm0.08M_{\odot})$ for Cen X-3. Freire et al.\cite{freire} obtained the mass of the compact star PSRJ 1614-2230 as $(1.97\pm0.04M_{\odot})$. Gangopadhyay et al.\cite{gan} estimated the radius of the compact star $7.866\pm0.21$ for 4U~1538-52, $9.69\pm0.2$ for
PSR~J1614-2230
$9.56\pm0.08$ for Vela~X-1 and
$9.178\pm0.13$ for Cen-X3. By using suitable values of the constants we have calculated the mass and radius of above mentioned compact stars which is presented in table $2$ and it is clear that calculated values of mass and radius of these stars is well fitted with the observational data. Moreover we have obtained the central density, surface density central pressure, compactness and surface redshift in table~3. The central density is in the order of $10^{15}g cm^{-3}$, surface density is in the order of $10^{14}~g cm^{-3}$, compactness factor is $<4/9$ and surface redshift $z_s\leq1.$\par
Hence we conclude that our proposed model may be used to
describe the interior of a superdense star corresponding to the exterior Schwarzschild line element.
|
\section{introduction}
Protoplanetary discs - the birthplaces of planets - are found around young stars which are themselves formed in clusters. The discs are thus externally irradiated by other cluster members, in particular by the most massive stars. Strong irradiation of discs close to O stars is well established, for example from observations of proplyds in Orion \citep{1996AJ....111.1977M, 1998ApJ...499..758J, 1998AJ....115..263O, 2000AJ....119.2919B, 2001AJ....122.2662O, 2002ApJ...566..315H}. For some time there has also been the theoretical expectation that protoplanetary discs might be significantly affected by more canonical radiation field strengths found in a cluster environment \citep[e.g.][]{2001MNRAS.325..449S, 2004ApJ...611..360A, 2011PASP..123...14H, 2016MNRAS.457.3593F, 2016MNRAS.tmp.1384H}. This is now being directly supported by recent observations, such as those by \cite{2016ApJ...826L..15K} who identify proplyds irradiatied by a 3000\,G$_0$\footnote{G$_0$ - the Habing unit - is a measure of the UV field local to our solar system and has the value $1.6\times10^{-3}$\,erg\,cm$^{-2}$\,s$^{-1}$.} radiation field -- approximately 100 times weaker than the field strengths irradiating classical proplyds \citep{1999ApJ...515..669S}. To date, however, there is no direct observational inference of externally driven mass loss from discs at lower, but more typical, radiation field strengths in the range $1<$\,G$_0$\,$<1000$.
IM Lupi is a roughly Solar mass \citep{2009A&A...501..269P} young \citep[$\sim0.5-1$\,Myr;][]{2012A&A...544A.131M} M0 star situated at a distance of $\sim161$\,pc \citep{2016arXiv160904172G} in the vicinity of the Lupus 2 cloud. Although CO emission likely associated with the disc is detected out to $\sim1000$\,AU, it is only detected in millimetre continuum out to about 313\,AU so the gas disc is more extended than the dust \citep{2007A&A...462..211L, 2008A&A...489..633P, 2009A&A...501..269P, 2016ApJ...832..110C}. The disc is also very massive, with estimates of 0.1 and 0.17\,M$_\odot$ from \cite{2008A&A...489..633P} and \cite{2016ApJ...832..110C} respectively. The mass accretion rate is currently about $10^{-8}$\,M$_\odot$\,yr$^{-1}$ \citep{2016arXiv161207054A}. The most recent analysis of this system by \cite{2016ApJ...832..110C} combined new $^{12}$CO, $^{13}$CO and C$^{18}$O ALMA observations with a broad array of modelling resources to provide a very comprehensive chemical and radiative transfer model of IM Lup, which could describe many features of the disc very successfully. They also included the effect of external irradiation on the composition and thermal structure of the disc. Based on their modelling efforts and from geometrical arguments based on \textit{HIPPARCOS} data they estimate a low UV field incident upon the disc of only about 4\,G$_0$. There was, however, a diffuse halo of low velocity CO emission about the disc that their model failed to explain. They suggested that this halo might be a remnant structure rather than being material driven out of the disc by photoevaporation. The photoevaporation interpretation was disfavoured based on the inferred low UV field and outer disk temperatures, which were well below those which had been previously considered by external photoevaporation models \citep{2004ApJ...611..360A, 2016MNRAS.457.3593F}. However, since this regime is previously unexplored it is difficult to conclude this with any certainty.
In this letter we use photochemical-dynamical models to investigate the external irradiation of IM Lup by the weak UV radiation field expected. We aim to determine the expected mass loss rate and flow properties and to determine whether the CO halo could be explained by such a flow. Ultimately we aim to determine whether low radiation field strengths can drive efficient mass loss and whether IM Lup offers an opportunity to observationally probe externally driven mass loss in the modest radiation regime.
\vspace{-10pt}
\section{Modelling the external photoevaporation of IM Lup}
\subsection{Numerical method and disc construction}
We directly model the photoevaporative outflow, driven by external irradiation, using a radiation hydrodynamics and photodissociation region chemistry code \textsc{torus-3dpdr}, for which key relevant papers are \cite{2012MNRAS.420..562H, 2015MNRAS.448.3156H, 2015MNRAS.453.2277H, 2015MNRAS.454.2828B}. This code was used to run models of externally irradiated discs in benchmark scenarios where there are semi--analytic solutions in \cite{2016MNRAS.tmp.1384H} - validating the approach. The details of the method are also discussed in the latter paper.
In summary we perform calculations of the photodissociation region (PDR) chemistry in sequence with hydrodynamics using operator splitting. The PDR chemistry network is a reduced version of the \textsc{umist} network \citep{2013A&A...550A..36M} including 33 species and 330 reactions, and was derived such that it gives temperatures that do not differ appreciably ($\sim$10\,per cent) from the much more substantial (and computationally expensive) full network. {We do not include polycyclic aromatic hydrocarbons (PAHs), since although they are a key heating mechanism in PDRs, they are observed to be depleted towards discs} \citep{2006A&A...459..545G, 2010ApJ...714..778O}. Our models will therefore yield mass loss rates lower than models that would include PAHs. Because we compute steady state flow profiles we are permitted to perform the PDR calculations relatively infrequently, as the same steady state profile will always eventually result.
Following \cite{2004ApJ...611..360A, 2016MNRAS.457.3593F}, our models are 1D spherical {(see Figure 1 of the latter paper)}. This is believed to be justified because the mass loss is expected to be dominated from the disc outer edge since: i) the material there is least gravitationally bound and ii) the density falls off vertically in a disc more rapidly than radially. This method also assumes that the incident (exciting) UV field approaches inwards radially and cooling line photons escape outwards radially - so every other direction is infinitely optically thick.
We employ a fixed structure for the disc itself, which acts as an inner boundary condition to the radiatively driven flow. Interior to some outer disc radius $R_{\textrm{d}}$ we do not allow the conditions to evolve over time. For these fixed disc conditions we use the parameters derived by \cite{2016ApJ...832..110C}. The disc's gas surface density profile follows that of \cite{1974MNRAS.168..603L}, i.e.
\begin{equation}
\Sigma_{\textrm{g}}(R) = \Sigma_{c}\left(\frac{R}{R_c}\right)^{-\gamma}\exp\left[-\left(\frac{R}{R_c}\right)^{2-\gamma}\right],
\label{Lynden-BellPringle}
\end{equation}
where \cite{2016ApJ...832..110C} find $\Sigma_c = 25$\,g\,cm$^{-2}$, $R_c =100$\,AU, and $\gamma=1$.
The scale height is set by
\begin{equation}
H(R) = H_{100}\left(\frac{R}{100\,\textrm{AU}}\right)^{\psi},
\end{equation}
where \cite{2016ApJ...832..110C} find $\psi=1.15$ and $H_{100}=12$\,AU.
For the dust we assume a cross section of $\sigma_{\textrm{FUV}} = 5.04\times10^{-23}$\,cm$^{-2}$, dust to gas ratio of $d/g = 10^{-4}$ and the maximum grain size $s_{\textrm{max}}=1\,\mu$m, which are all representative of the kind of dust parameters in the flow found by \cite{2016MNRAS.457.3593F}. We assume that the disc outer edge is sufficiently far from the parent star that the temperature there is only 10\,K. The outer dynamical boundary condition in our models is free-outflow, no-inflow and the inner condition set by the disc properties at $R_{\textrm{d}}$ as described above. The mid-plane number density in the discs of these 1D spherical models is
\begin{equation}
n(R) = \frac{1}{\mu m_{\textrm{H}}} \frac{\Sigma_{\textrm{g}}(R)}{\sqrt{2\pi}H(R)}.
\label{midN}
\end{equation}
The radial extent of our simulation grid -- 10$^{17}$\,cm -- was chosen such that the critical radius in the flow \citep{2016MNRAS.457.3593F} is captured, which we check using the approach detailed in section 5.3.2 of \cite{2016MNRAS.tmp.1384H}. We use an adaptive grid with a maximum number of cells of 2048 and therefore a maximum resolution of 3.25\,AU. We run each model for 1\,Myr, though steady state flows are established long before this.
\vspace{-15pt}
\section{Results and discussion}
\subsection{Disc photoevaporation and evolution}
We ran a grid of photoevaporation models for different disc outer radii and incident radiation field strengths. We chose disc radii in 50\,AU intervals from 350 to 800\,AU and radiation field strengths of 0.5, 1, 2, 4, 8 and 16\,G$_0$. We compute the mass loss rate from our models following \cite{2004ApJ...611..360A}:
\begin{equation}
\dot{M} = 4\pi R^2 \rho \dot{R} \mathcal{F},
\end{equation}
where $\mathcal{F}$ is the fraction of solid angle subtended by the disc outer edge
\begin{equation}
\mathcal{F} = \frac{H_d}{\sqrt{H_d^2+R_{d}^2}},
\end{equation}
and $H_d$ is the scale height at the disc outer edge $R_d$.
We compute the average of this quantity over the entire flow (note that $\mathcal{F}$ is constant for a given disc outer radius). A summary of the mass loss rates from our grid of models is shown in Figure \ref{GridMLR}. For large discs (like IM Lup) where material at the outer edge is not so gravitationally bound, substantial mass loss rates ($\sim10^{-8}$\,M$_\odot$\,yr$^{-1}$) can be driven even when the incident radiation field strength is very modest. Note that the 4\,G$_0$ field expected to be irradiating IM Lup and driving this mass loss is $\sim10^3$ times weaker than that irradiating the proplyds observed by \cite{2016ApJ...826L..15K} and $\sim10^4$ times weaker than the proplyds in the ONC \citep[e.g.][]{2000AJ....119.2919B, 2001AJ....122.2662O, 2002ApJ...566..315H}
The current mass accretion rate in this system was recently computed using new X--shooter data to be $10^{-8}$\,M$_{\odot}$\,yr$^{-1}$ with an uncertainty of 0.35 dex by \cite{2016arXiv161207054A}. The external photoevaporative mass loss rate for UV fields $\geq4$\,G$_0$ is hence expected to be of order or greater than the mass accretion rate.
\begin{figure}
\includegraphics[width=7.7cm]{./MLRGRID.pdf}
\vspace{-10pt}
\caption{Log mass loss rate as a function of disc outer radius for different incident UV fields. }
\vspace{-10pt}
\label{GridMLR}
\end{figure}
We fit the mass loss rate as a function of radius, which we feed into the \cite{2007MNRAS.376.1350C} secular evolutionary code to examine the disc evolution. The viscous timescale of this disc is of order 10\,Myr, so the main evolutionary behaviour is truncation of the disc by external photoevaporation. Figure \ref{Router} shows the evolution of the disc outer edge as a function of time for different incident UV field strengths {(note that for models that drop below an outer radius of 350\,AU we compute additional photoevaporation models to estimate the mass loss rate at these smaller radii)}. In all cases the disc outer edge {rapidly retreats to some stagnation radius in less than 1\,Myr, after which it varies in size only slowly}. The {mean radius over 10\,Myr} as a function of incident UV field is given in Figure \ref{Rstag}, showing strong variation for fields $<8$\,G$_0$. A key point is that because the observed CO emission is currently extended out to beyond 1000\,AU, even an extremely weak UV radiation field would be expected to truncate this very rapidly. The observed CO emission therefore either has to be part of a photoevaporative flow, or part of some \textit{much} denser envelope that is resilient against the effects of the incident radiation field. Because IM Lup is very young ($\leq 1$\,Myr), its outer edge may still be in the process of retreating towards the stagnation radius. Another interesting point is that due to the disc's long viscous timescale IM Lup is likely to remain unusually large at the stagnation radius (perhaps > 300\,AU) for many Myr, unless some other mechanism further truncates the disc.
\begin{figure}
\includegraphics[width=7.7cm]{./RDEVO_new2.pdf}
\vspace{-10pt}
\caption{The outer disc radius of IM Lup as a function of time according to our evolutionary models that include external photoevaporation.}
\vspace{-10pt}
\label{Router}
\end{figure}
\begin{figure}
\includegraphics[width=7.7cm]{./RSTAG_new2.pdf}
\vspace{-10pt}
\caption{{The mean radius of IM Lup over 10\,Myr as a function of the incident UV field strength. }}
\vspace{-10pt}
\label{Rstag}
\end{figure}
\vspace{-10pt}
\subsection{Could external photoevaporation explain the CO halo?}
Our models imply that even in the presence of a weak UV field substantial mass loss is being induced from IM Lup by external photoevaporation; however, being 1D, they are difficult to directly compare with the real observed data of \cite{2016ApJ...832..110C}. Generating synthetic observations from 1D models has the limitation that some assumption about the vertical density, temperature and compositional structure is required. Nevertheless we make an optimistic attempt at comparison. We assume that the disc (the boundary condition of the dynamical models) is hydrostatic. In the flow region we use our simulation results and assume that at a given spherical radius there is a constant density, isothermal, isochemical flow, with scale height set by assuming that $H/R$ beyond the disc outer edge is constant. We produce synthetic data cubes using the comoving frame molecular line radiative transfer components of \textsc{torus}, detailed in \cite{2010MNRAS.407..986R}. These cubes are then azimuthally averaged in the same manner used to produce the results in Figure 12 of \cite{2016ApJ...832..110C}.
Because our synthetic observations are based on 1D models, and there is a large array of possible parameters, we do not aim to fit the CO observations. Furthermore, given that we are comparing with $^{12}$CO our synthetic observations will be particularly limited in components of the flow that are optically thick (which vary for each model but we generally find are interior to about 800--900\,AU). Rather then, we aim to demonstrate that even weak external photoevaporation is capable of producing substantial emission at large radii, such as that observed in the CO halo of IM Lup.
Figure \ref{synthObservations} shows a collection of approximate synthetic emission profiles from our photoevaporation models, as well as emission profiles from a selection of the models from \cite{2016ApJ...832..110C}. The latter models modify the incident UV field but do not permit radial dynamical evolution and thus impose the surface density profile given by equation \ref{Lynden-BellPringle}. As a result the extent of the CO emission is significantly less than that observed. Conversely our external photoevaporation models do show emission comparable in extent and magnitude to the observations.
In Figure \ref{extent} we plot the radial extent of the CO emission in our models as a function of the disc outer radius, with different lines representing different incident UV fluxes. The fitted gas extent from \citet{2016ApJ...832..110C} is 1200\,AU which we take to be ``the extent'' of IM Lup for our comparison here, though in practice the detection is marginal beyond 1000\,AU. From our models the extent is the point at which the flux drops below $2$\,mJy\,beam$^{-1}$\,km\,s$^{-1}$, which is the background as calculated using the average of the first and last velocity channels in the synthetic data cube. Most of our models have an extent $1000-1300$\,AU. Generally the models that have extent consistent with the observations have disc outer radii which imply that the disc outer edge is still retreating.
If the observed extent were known with higher certainty, we could use it in conjunction with Figures \ref{Router} and \ref{extent} to constrain the minimum disc outer edge and hence maximum age. For example if we knew that the observed extent was 1200\,AU then linearly interpolating Figure \ref{extent} would yield minimum disc outer radii of {430, 450 and 530\,AU} for incident UV fields of 4, 2 and 1\,G$_0$ respectively. Using our evolutionary models from Figure \ref{Router}, these minimum disc outer radii would correspond to approximate maximum IM Lup ages of {0.3, 0.8 and 0.8} \,Myr respectively -- so all would be conceivable given the uncertain 0.5--1\,Myr estimate for the age of IM Lup. Future higher sensitivity observations might offer such a constraint.
\begin{figure}
\includegraphics[width=7.8cm]{./plotMULTI_varR_2_moreturb.pdf}
\includegraphics[width=7.8cm]{./plotMULTI_varG_2_moreturb.pdf}
\includegraphics[width=7.8cm]{./plotMULTI_cleeves_1to3.pdf}
\vspace{-10pt}
\caption{Azimuthally averaged emission profiles from our approximate synthetic observations, compared to the observed data points (with 1\,$\sigma$ error bars) from Cleeves et al. (2016). The upper panel varies the disc outer radius for a radiation field of 4\,G$_0$. The middle panel varies the incident radiation field strength upon a {450\,AU disc}. The bottom panel shows a collection of models from Cleeves et al. (2016). }
\label{synthObservations}
\end{figure}
\begin{figure}
\hspace{-10pt}
\vspace{-15pt}
\includegraphics[width=8cm]{./extentMod_new.pdf}
\vspace{-10pt}
\caption{The approximate radial extent of CO emission in our models as a function of disc outer radius. The horizontal line represents the radial extent from the best fit model of Cleeves et al. (2016). }
\vspace{-10pt}
\label{extent}
\end{figure}
\vspace{-15pt}
\section{Summary and conclusions}
We model the external photoevaporation of the large protoplanetary disc IM Lup. This disc has a large CO ``halo'' that was identified in recent ALMA observations by \cite{2016ApJ...832..110C} which could not be explained by hydrostatic chemical and radiative transfer models which assumed that the surface density at large radius was an extrapolation of the form given in equation \ref{Lynden-BellPringle}. We find that although the radiation field irradiating IM Lup is very weak ($<10^4$ times the UV field irradiating the proplyds near O stars in Orion), the disc is sufficiently large that the weakly gravitationally bound material at the disc outer edge can be efficiently photoevaporated. Specifically a 4\,G$_0$ radiation field induces mass loss of $\sim10^{-8}$\,M$_{\odot}$\,yr$^{-1}$ which is comparable to the current accretion rate onto the star.
Having a $\sim$10\,Myr viscous timescale, the effect of this mass loss is to rapidly (<1\,Myr) truncate the disc outer edge down to some stagnation radius. The stagnation radius ranges from about 600\,AU for an irradiating UV field of 0.5\,G$_0$ down to about {300-350\,AU} for fields $8-16$\,G$_0$. In the absence of other external influences the disc {only evolves slowly away from the stagnation radius over 10\,Myr}. {Once gas from the disc cannot be delivered to the outer edge at a rate sufficient to supply the photoevaporative wind the disc is expected to shrink rapidly}.
We also generated approximate synthetic observations from our models, which are able to explain the radial extent of CO emission about IM Lup. Our scenarios that are consistent with the observed extent of CO emission of IM Lup generally imply that its disc outer radius is still in the process of being truncated. More generally we demonstrate that even weak external fields can lead to significant extended emission from large discs, which hydrostatic models are unable to achieve.
\vspace{-15pt}
\section*{Acknowledgements}
TJH is funded by an Imperial College London Junior Research Fellowship. This work has also been supported by the DISCSIM project, grant agreement 341137 funded by the European Research Council under ERC-2013-ADG. LIC acknowledges the support of NASA through Hubble Fellowship grant HST-HF2-51356.001-A awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555.
This work was undertaken on the COSMOS Shared Memory system at DAMTP, University of Cambridge operated on behalf of the STFC DiRAC HPC Facility. This equipment is funded by BIS National E-infrastructure capital grant ST/J005673/1 and STFC grants ST/H008586/1, ST/K00333X/1.
This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2013.00694. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
\vspace{-18pt}
\bibliographystyle{mn2e}
|
\section{Introduction}
One important problem for many high-stakes sequential decision making under uncertainty domains, including robotics, health care, education, and dialogue systems, is estimating the performance of a new policy without requiring it to be deployed. To address this, off-policy policy evaluation (OPE) algorithms use historical data collected from executing one policy (called the behavior policy), to predict the performance of a new policy (called the evaluation policy). Importance sampling (IS) is one powerful approach that can be used to evaluate the potential performance of a new policy \citep{precup2000eligibility}. In contrast to model based approaches to OPE \cite{hallak2015off}, importance sampling provides an unbiased estimate of the performance of the evaluation policy. In particular, importance sampling is robust to partial observability, which is often prevalent in real-world domains.
Unfortunately, importance sampling estimates of the performance of the evaluation policy can be inaccurate when the horizon of the problem is long: the variance of IS estimators can grow exponentially with the number of sequential decisions made in an episode.
This is a serious limitation for applications that involve decisions made over tens or hundreds of steps, like dialogue systems where a conversation might require dozens of responses, or intelligent tutoring systems that make dozens of decisions about how to sequence the content shown to a student.
Due to the importance of OPE, there have been many recent efforts to improve the accuracy of importance sampling.
For example, \citet{Dudik2011} and \citet{jiang2016doubly} proposed doubly robust importance sampling estimators that can greatly reduce the variance of predictions when an approximate model of the environment is available.
\citet{Thomas2016} proposed an estimator that further integrates importance sampling and model-based approaches, and which can greatly reduce mean squared error.
These approaches trade-off between the bias and variance of model-based and importance sampling approaches, and result in strongly consistent estimators. Unfortunately, in long horizon settings, these approaches will either create estimates that suffer from high variance or exclusively rely on the provided approximate model, which can have high bias.
Other recent efforts that estimate a \textit{value function} using off-policy data rather than just the performance of a policy \cite{NIPS2015_5807,munos2016safe, hallak2015generalized} also suffer from bias if the input state description is not Markovian (such as if the domain description induces partial observability).
To provide better off policy estimates in long horizon domains, we propose leveraging temporal abstraction. In particular, we analyze using options-based policies (policies with temporally extended actions) \cite{sutton1999between} instead of policies over primitive actions.
We prove that the we can obtain an exponential reduction in the variance of the resulting estimates, and in some cases, cause the variance to be independent of the horizon. We also demonstrate this benefit with simple simulations. Crucially, our results can be equivalently viewed as showing that using options can drastically reduce the amount of historical data required to obtain an accurate estimate of a new evaluation policy's performance.
We also show that using options-based policies can result in special cases which can lead to significant reduction in estimation error through dropping importance sampling weights. Furthermore, we generalize the idea of dropping weights and derive a covariance test that can be used to automatically determine which weights to drop. We demonstrate the potential of this approach by constructing a new importance sampling algorithm called Incremental Importance Sampling~(INCRIS) and show empirically that it can significantly reduce estimation error.
\section{Background}
We consider an agent interacting with a Markov decision process (MDP) for a finite sequence of time steps. At each time step the agent executes an action, after which the MDP transitions to a new state and returns a real valued reward. Let $s \in S$ be a discrete state, $a \in A$ be a discrete action, and $r$ be the reward bounded in $[0, R_{\max}]$).
The transition and reward dynamics are unknown and are denoted by the transition probability $T(s' | s, a)$ and reward density $R(r | s, a)$.
A primitive policy maps histories to action probabilities, i.e., $\pi(a_t | s_1,a_1,r_1,\ldots,s_t)$ is the probability of executing action $a_t$ at time step $t$ after encountering history $s_1,a_1,r_1,\ldots,s_t$. The return of a trajectory $\tau$ of $H$ steps is simply the sum of the rewards $G(\tau) = \sum_{t=1}^H r_t$. Note we consider the undiscounted setting where $\gamma=1$. The value of policy $\pi$ is the
expected return when running that policy: $V_\pi = \mathbb{E}_{\pi}(G(\tau))$.
Temporal abstraction can reduce the computational complexity of planning and online learning \cite{sutton1999between,mann2013advantage,mankowitz2014time,brunskill2014pac}. One popular form of temporal abstraction is to use sub-policies, in particular options \citep{sutton1999between}. Let $\Omega$ be the space of trajectories. $o$, an option, consists of $\pi$, a primitive policy (a policy over primitive actions), $\beta: \Omega \rightarrow [0,1]$, a termination condition where $\beta(\tau)$ is the probability of stopping the option given the current partial trajectory $\tau \in \Omega$ from when this option began, and $I \subset S$, an input set where $s \in I$ denotes the states where $o$ is allowed to start. Primitive actions can be considered as a special case of options, where the options always terminate after a single step. $\mu(o_t | s_1,a_1,\ldots,s_t)$ denotes the probability of picking option $o_t$ given history $(s_1,a_1,\ldots,s_t)$ when the previous option has terminated, according to options-based policy $\mu$. A high-level trajectory of length $k$ is denoted by $T = (s_1, o_1, v_1, s_2, o_2, v_2, \dots, s_k, o_k, v_k)$ where $v_t$ is the sum of the rewards accumulated when executing option $o_t$.
In this paper we will consider batch, offline, off-policy evaluation of policies for sequential decision making domains using both primitive action policies and options-based policies. We will now introduce the general OPE problem using primitive policies: in a later section we will combine this with options-based policies.
In OPE we assume access to historical data, $D$, generated by an MDP, and a behavior policy ,$\pi_b$. $D$ consists of $n$ trajectories, $\{\tau^{(i)}\}_{i=1}^n$. A trajectory has length $H$, and is denoted by $\tau^{(i)} = (s^{(i)}_1, a^{(i)}_1, r^{(i)}_1, s^{(i)}_2, a^{(i)}_2, r^{(i)}_2, \dots, s^{(i)}_H, a^{(i)}_H, r^{(i)}_H)$. In off-policy evaluation, the goal is to use the data $D$ to estimate the value of an evaluation policy $\pi_e$: $V_{\pi_e}$. As $D$ was generated from running the behavior policy $\pi_b$, we cannot simply use the Monte Carlo estimate. An alternative is to use importance sampling to reweight the data in $D$ to give greater weight to samples that are likely under $\pi_e$ and lesser weight to unlikely ones. We consider per-decision importance sampling (PDIS) \cite{precup2000eligibility}, which gives the following estimate of the value of $\pi_e$:
\begin{align}
\operatorname{PDIS}(D) &= \frac{1}{n} \sum_{i = 1}^n \left( \sum_{t=1}^H \rho^{(i)}_t r^{(i)}_t \right) ,&
\rho^{(i)}_t &= \prod_{u=1}^t \frac{\pi_e(a^{(i)}_u | s^{(i)}_u)}{\pi_b(a^{(i)}_u | s^{(i)}_u)},
\end{align}
where $\rho^{(i)}_t$ is the weight given to the rewards to correct due to the difference in distribution. This estimator is an unbiased estimator of the value of $\pi_e$:
\begin{align}
\mathbb{E}_{\pi_e}(G(\tau)) &= \mathbb{E}_{\pi_b}(PDIS(\tau)),
\end{align}
where $\mathbb{E}_{\pi}(\dots)$ is the expected value given that the trajectories $\tau$ are generated by $\pi$.
For simplicity, hereafter we assume that primitive and options-based policies are a function
only of the current state, but our results apply also when the they are a function of the history. Note that importance sampling does not assume that the states in the trajectory are Markovian, and is thus robust to error in the state representation, and in general, robust to partial observability as well.
\section{Importance Sampling and Long Horizons}
We now show how the amount of data required for importance sampling to obtain a good off-policy estimate can scale exponentially with the problem horizon. Notice that in the standard importance sampling estimator, the weight is the product of the ratio of action probabilities. We now prove that this can cause the variance of the policy estimate to be exponential in $H$.\footnote{These theorems can be seen as special case instantiations of Theorem 6 in \citep{pmlr-v38-li15b} with simpler, direct proofs.}
\begin{theorem}
The mean squared error of the PDIS estimator can be $\Omega(2^H)$. {\bf Proof.} See appendix.
\label{thm:mse_h}
\end{theorem}
Equivalently, this means that achieving a desired mean squared error of $\epsilon$ can require a number of trajectories that scales exponentially with the horizon. A natural question is whether this issue also arises in a weighted importance sampling \citep{precup2001off}, a popular (biased) approach to OPE that has lower variance. We show below that the long horizon problem still persists.
\begin{theorem}
It can take $\Omega(2^H)$ trajectories to shrink the MSE of weighted importance sampling (WIS) by a constant. {\bf Proof.} See appendix.
\end{theorem}
\section{Combining Options and Importance Sampling}
We will show that one can leverage the advantage of options to mitigate the long horizon problem. If the behavior and evaluation policies are both options-based policies, then the PDIS estimator can be exponentially more data efficient compared to using primitive behavior and evaluation policies.
Due to the structure in options-based policies, we can decompose the difference between the behavior policy and the evaluation policy in a natural way. Let $\mu_b$ be the options-based behavior policy and $\mu_e$ be the options-based evaluation policy. First, we examine the probabilities over the options. The probabilities $\mu_b(o_t | s_t)$ and $\mu_e(o_t | s_t)$ can differ and contribute a ratio of probabilities as an importance sampling weight. Second, the underlying policy, $\pi$, for an option, $o_t$, present in both $\mu_b$ and $\mu_e$ may differ, and this also contributes to the importance sampling weights. Finally, additional or missing options can be expressed by setting the probabilities over missing options to be zero for either $\mu_b$ or $\mu_e$. Using this decomposition, we can easily apply PDIS to options-based policies.
\begin{theorem}
\label{thm:main}
Let $\mathcal{O}$ be the set of options that have the same underlying policies between $\mu_b$ and $\mu_e$. Let $\overline{\mathcal{O}}$ be the set of options that have changed underlying policies. Let $k^{(i)}$ be the length of the $i$-th high level trajectory from data set $D$. Let $j^{(i)}_t$ be the length of the sub-trajectory produced by option $o^{(i)}_t$. The PDIS estimator applied to $D$ is
\begin{align}
PDIS(D) &= \frac{1}{n} \sum_{i = 1}^n \left( \sum_{t=1}^{k^{(i)}} w^{(i)}_t y^{(i)}_t \right) & w^{(i)}_t &= \prod_{u=1}^t \frac{\mu_e(o^{(i)}_u | s^{(i)}_u)}{\mu_b(o^{(i)}_u | s^{(i)}_u)},
\\
y^{(i)}_t &= \begin{cases}
v^{(i)}_t \text{ if } o^{(i)}_t \in \mathcal{O} \\
\sum_{b=1}^{j^{(i)}_t} \rho^{(i)}_{t,b} r^{(i)}_{t,b} \text{ if } o^{(i)}_t \in \overline{\mathcal{O}}
\end{cases} & \rho^{(i)}_{t,b} &= \prod_{c=1}^{j^{(i)}_t} \frac{\pi_e(a^{(i)}_{t,c} | s^{(i)}_{t,c}, o^{(i)}_t)}{\pi_b(a^{(i)}_{t,c} | s^{(i)}_{t,c}, o^{(i)}_t)},
\end{align}
where $r^{(i)}_{t,b}$ is the $b$-th reward in the sub-trajectory of option $o^{(i)}_t$ and similarly for $s$ and $a$.
\end{theorem}
\begin{proof}
This is a straightforward application of PDIS to the options-based policies using the decomposition mentioned.
\end{proof}
Theorem \ref{thm:main} expresses the weights in two parts: one part comes from the probabilities over options which is expressed as $w^{(i)}_t$, and another part comes from the underlying primitive policies of options that have changed with $\rho^{(i)}_{t,b}$. We can immediately make some interesting observations below.
\begin{corollary}
\label{cor:case1}
If no underlying policies for options are changed between $\mu_b$ and $\mu_e$, and all options have length at least $J$ steps, then the worst case variance of PDIS is exponentially reduced from $\Omega(2^H)$ to $\Omega(2^{(H/J)})$
\end{corollary}
Corollary \ref{cor:case1} follows from Theorem \ref{thm:main}. Since no underlying policies are changed, then the only importance sampling weights left are $w^{(i)}_t$. Thus we can focus our attention only on the high-level trajectory which has length at most $H/J$. Effectively, the horizon has shrunk from $H$ to $H/J$, which results in an exponential reduction of the worst case variance of PDIS.
\begin{corollary}
\label{cor:case2}
If the probabilities over options are the same between $\mu_b$ and $\mu_e$, and a subset of options $\overline{\mathcal{O}}$ have changed their underlying policies, then the worst case variance of PDIS is reduced from $\Omega(2^H)$ to $\Omega(2^K)$ where $K$ is an upper bound on the sum of the lengths of the options.
\end{corollary}
Corollary \ref{cor:case2} follows from Theorem \ref{thm:main}. The options whose underlying policies are the same between behavior and evaluation can effectively be ignored, and cut out of the trajectories in the data. This leaves only options whose underlying policies have changed, shrinking down the horizon from $H$ to the length of the leftover options. For example, if only a single option of length 3 is changed, and the option appears once in a trajectory, then the horizon can be effectively reduced to just 3. This result can be very powerful, as the reduced variance becomes independent of the horizon $H$.
\section{Experiment 1: Options-based Policies}
This experiment illustrates how using options-based policies can significantly improve the accuracy of importance-sampling-based estimators for long horizon domains. Since importance sampling is particularly useful when a good model of the domain is unknown and/or the domain involves partial observability, we introduce a partially observable variant of the popular Taxi domain \cite{dietterich2000hierarchical} called NoisyTaxi for our simulations (see figure \ref{fig:taxi}).
\subsection{Partially Observable Taxi}
\begin{SCfigure}[4][h]
\label{fig:taxi}
\centering
\includegraphics[scale=0.4]{taxi}
\caption{Taxi Domain \citep{dietterich2000hierarchical}. It is a $5 \times 5$ gridworld (Figure \ref{fig:taxi}). There are 4 special locations: R,G,B,Y. A passenger starts randomly at one of the 4 locations, and its destination is randomly chosen from one of the 4 locations. The taxi starts randomly on any square. The taxi can move one step in any of the 4 cardinal directions N,S,E,W, as well as attempt to pickup or drop off the passenger.
Each step has a reward of $-1$. An invalid pickup or dropoff has a $-10$ reward and a successful dropoff has a reward of 20.}
\end{SCfigure}
In NoisyTaxi, the location of the taxi and the location of the passenger is partially observable. If the row location of the taxi is $c$, the agent
observes $c$ with probability 0.85, $c+1$ with
probability 0.075 and $c-1$ with probability
0.075 (if adding or subtracting 1 would cause
the location to be outside the grid, the resulting
location is constrained to still lie in the grid).
The column location of the taxis is observed
with the same noisy distribution.
Before the taxi successfully picks up the passenger, the observation of the location of the passenger has a probability of 0.15 of switching randomly to one of the four designated locations. After the passenger is picked up, the passenger is observed to be in the taxi with 100\% probability (e.g. no noise while in the taxi).
\subsection{Experimental Results}
We consider $\epsilon$-greedy option policies, where with
probability $1-\epsilon$ the policy samples the optimal option,
and probability $\epsilon$ the policy samples a random option.
Options in this case are $n$-step policies, where ``optimal''
options involve taking $n$-steps of the optimal (primitive action)
policy, and ``random'' options involve taking $n$ random primitive
actions.\footnote{We have also tried using more standard options that navigate to a specific destination, and the experiment results closely mirror those shown here.} Our behavior policies $\pi_b$ will use $\epsilon=0.3$ and our
evaluation policies $\pi_e$ use $\epsilon=0.05$. We investigate how
the accuracy of estimating $\pi_e$ varies as a function both
of the number of trajectories and the length of the options $n=1,2,3$.
Note $n=1$ corresponds to having a primitive action policy.
Empirically, all behavior policies have essentially the same performance. Similarly all evaluation policies have essentially the same performance. We first collect data using the behavior policies, and then use PDIS to evaluate their respective evaluation policies.
Figure \ref{fig:graph1} compares the MSE (log scale) of the PDIS estimators for the evaluation policies.
\begin{SCfigure}[2][h]
\label{fig:graph1}
\centering
\includegraphics[scale=0.3]{graph1}
\caption{Comparing the MSE of PDIS between primitive and options-based behavior and evaluation policy pairs. Note the y-axis is a log scale. Our results show that PDIS for the options-based evaluation policies are an order of magnitude better than PDIS for the primitive evaluation policy. Indeed, Corollary \ref{cor:case1}
shows that the $n$-step options policies are effectively reducing
the horizon by a factor of $n$ over the primitive policy. As
expected, the options-based policies that use 3-step options have the lowest MSE.}
\end{SCfigure}
\section{Going Further with Options}
Often options are used to achieve a specific sub-task in a domain. For example in a robot navigation task, there may be an option to navigate to a special fixed location. However one may realize that there is a faster way to navigate to that location, so one may change that option and try to evaluate the new policy to see whether it is actually better. In this case the old and new option are both always able to reach the special location; the only difference is that the new option could get there faster. In such a case we can further reduce the variance of PDIS. We now formally define this property.
\begin{definition}
Given behavior policy $\mu_b$ and evaluation policy $\mu_e$, an option $o$ is called \textbf{stationary}, if the distribution of the states on which $o$ terminates is always the same for $\mu_b$ and $\mu_e$. The underlying policy for option $o$ can differ for $\mu_b$ and $\mu_e$; only the termination state distribution is important.
\end{definition}
A stationary option may not always arise due to solving a sub-task. It can also be the case that a stationary option is used as a way to perform a soft reset. For example, a robotic manipulation task may want to reset arm and hand joints to a default configuration in order to minimize sensor/motor error, before trying to grasp a new object.
Stationary options allows us to point to a step in a trajectory where we know the state distribution is fixed. Because the state distribution is fixed, we can partition the trajectory into two parts. The beginning of the second partition would then have state distribution that is independent of the actions chosen in the first partition. We can then independently apply PDIS to each partition, and sum up the estimates. This is powerful because it can halve the effective horizon of the problem.
\begin{theorem}
\label{thm:special}
Let $\mu_b$ be an options-based behavior policy. Let $\mu_e$ be an options-based evaluation policy. Let $\mathcal{O}$ be the set of options that $\mu_b$ and $\mu_e$ use. The underlying policies of the options in $\mu_e$ may be arbitrarily different from $\mu_b$.
Let $o_1$ be a stationary option. We can decompose the expected value as follows. Let $\tau_1$ be the first part of a trajectory up until and including the first occurrence of $o_1$. Let $\tau_2$ be the part of the trajectory after the first occurrence of $o_1$ up to and including the first occurrence of $o_2$. Then
\begin{align}
\mathbb{E}_{\mu_e}(G(\tau)) &= \mathbb{E}_{\mu_b}(PDIS(\tau)) = \mathbb{E}_{\mu_b}(PDIS(\tau_1)) + \mathbb{E}_{\mu_b}(PDIS(\tau_2))
\end{align}
{\bf Proof.} See appendix.
\end{theorem}
Note that there are no conditions on how the probabilities over options may differ, nor on how the underlying policies of the non-stationary options may differ. This means that, regardless of these differences, the trajectories can be partitioned and PDIS can be independently applied. Furthermore, Theorem \ref{thm:main} can still be applied to each of the independent applications of PDIS. Combining Theorem \ref{thm:special} and Theorem \ref{thm:main} can lead to more ways of designing a desired evaluation policy that will result in a low variance PDIS estimate.
\section{Experiment 2: Stationary Options}
We now demonstrate Theorem \ref{thm:special} empirically on NoisyTaxi. In NoisyTaxi, we know that a primitive $\epsilon$-greedy policy will eventually
pick up the passenger (though it may take a very long time depending on $\epsilon$). Since the starting location of the passenger is uniformly random, the location of the taxi immediately after picking up the passenger is also uniformly random, but over the four pickup locations. This implies that, regardless of the $\epsilon$ value
in an $\epsilon$-greedy policy, we can view executing that $\epsilon$-greedy
policy until the passenger is picked up as a new "PickUp-$\epsilon$" option that
always terminates in the same state distribution.
Given this argument, we can use Theorem \ref{thm:special} to decompose any NoisyTaxi trajectory into the part before the passenger is picked up, and the part after the
passenger is picked up, estimate the expected reward for each, and
then sum. As picking up the passenger is often the halfway point in a trajectory (depending on the locations of the passenger and the destination),
we can perform importance sampling over two, approximately half
length, trajectories.
More concretely, we consider two $n=1$ options (e.g. primitive action)
$\epsilon$-greedy policies. Like in the prior subsection, the behavior
policy has $\epsilon=0.3$ and the evaluation policy has $\epsilon=0.05$.
We compare performing normal PDIS to estimate the value of the evaluation
policy to estimating it using partitioned-PDIS using Theorem \ref{thm:special}. See Figure \ref{fig:graph2} for results.
\begin{SCfigure}[2][h]
\label{fig:graph2}
\centering
\includegraphics[scale=0.3]{graph2}
\caption{Comparing MSE of Normal PDIS and PDIS that uses Theorem \ref{thm:special}. We gain an order of an order of magnitude reduction in MSE (labeled Partitioned-PDIS). Note this did not require that the primitive policy used options: we merely used the fact that if there are subgoals in the domain where the agent is likely to go through with a fixed state distribution, we can leverage that to decompose the value of a long horizon into the sum over multiple shorter ones. Options is one common way this will occur, but as we see in this example, this can also occur in other ways.
}
\end{SCfigure}
\section{Covariance Testing}
The special case of stationary options can be viewed as a form of dropping certain importance sampling weights from the importance sampling estimator. With stationary options, the weights before the stationary options are dropped when estimating the rewards thereafter. By considering the bias incurred when dropping weights, we derive a general rule involving covariances as follows. Let $W_1 W_2 r$ be the ordinary importance sampling estimator for reward $r$ where the product of the importance sampling weights are partitioned into two products $W_1$ and $W_2$ using some general partitioning scheme such that $\mathbb{E}(W_1) = 1$. Note that this condition is satisfied when $W_1,W_2$ are chosen according to commonly used schemes such as fixed timesteps (not necessarily consecutive) or fixed states, but can be satisfied by more general schemes as well. Then we can consider dropping the product of weights $W_1$ and simply output the estimate $W_2 r$:
\begin{align}
\mathbb{E}(W_1 W_2 r) &= \mathbb{E}(W_1)\mathbb{E}(W_2 r) + \operatorname{Cov}(W_1, W_2 r) \\
&= \mathbb{E}(W_2 r) + \operatorname{Cov}(W_1, W_2 r) \label{eq:covtest}
\end{align}
This means that if $\operatorname{Cov}(W_1, W_2 r) = 0$, then we can drop the weights $W_1$ with no bias. Otherwise, the bias incurred is $\operatorname{Cov}(W_1, W_2 r)$. Then we are free to choose $W_1,W_2$ to balance the reduction in variance and the increase in bias.
\subsection{Incremental Importance Sampling~(INCRIS)}
Using the Covariance Test (eqn \ref{eq:covtest}) idea, we propose a new importance sampling algorithm called Incremental Importance Sampling~(INCRIS). This is a variant of PDIS where for a reward $r_t$, we try to drop all but the most recent $k$ importance sampling weights, using the covariance test to optimize $k$ in order to lower MSE.
Let $\pi_b$ and $\pi_e$ be the behavior and evaluation policies respectively (they may or may not be options-based policies). Let $D = \{ \tau^{(1)}, \tau^{(2)}, \dots, \tau^{(n)} \}$ be our historical data set generated from $\pi_b$ with $n$ trajectories of length $H$. Let $\rho_t = \frac{\pi_e(a_t | s_t)}{\pi_b(a_t | s_t)}$. Let $\rho^{(i)}_t$ be the same but computed from the $i$-th trajectory. Suppose we are given estimators for covariance and variance. See algorithm \ref{alg:covtest} for details.
\begin{algorithm}[h]
\begin{algorithmic}[1]
\STATE \textbf{Input:} $D$
\FOR{$t=1$ to $H$}
\FOR{$k=0$ to $t$}
\STATE $A_k = \prod_{j=1}^{t-k} \rho_j$
\STATE $B_k = \prod_{j=t-k+1}^{t} \rho_j$
\STATE Estimate $\operatorname{Cov}(A_k, B_k r_t)$ and denote $\widehat{C}_k$
\STATE Estimate $\operatorname{Var}(B_k r_t)$ and denote $\widehat{V}_k$
\STATE Estimate MSE with $\widehat{MSE}_k = \widehat{C}_k^2 + \widehat{V}_k$
\ENDFOR
\STATE $k' = \operatorname{argmin}_{k} \widehat{MSE}_k$
\STATE Let $\widehat{r_t} = \frac{1}{n} \sum_{i=1}^n B_k^{(i)} r_t$
\ENDFOR
\STATE \textbf{Return} $\sum_{t=1}^H \widehat{r_t}$
\end{algorithmic}
\caption{INCRIS}
\label{alg:covtest}
\end{algorithm}
\subsection{Strong Consistency}
In the appendix, we provide a proof that INCRIS~is strongly consistent. We now give a brief intuition for the proof. As $n$ goes to infinity, the estimates for the MSE get better and better and converge to the bias. We know that if we do not drop any weights, we get an unbiased estimate and thus the smallest MSE estimate will go to zero. Thus we get more and more likely to pick $k$ that correspond to unbiased estimates.
\section{Experiment 3: Incremental Importance Sampling}
To evaluate INCRIS, we constructed a simple MDP that exemplifies to properties of domains for which we expect INCRIS\ to be useful.
Specifically, we were motivated by the applications of reinforcement learning methods to type 1 diabetes treatments \citep{Bastani2014,Thomas2017} and digital marketing applications \citep{Theocharous2015}.
In these applications there is a natural place where one might divide data into episodes: for type 1 diabetes treatment, one might treat each day as an independent episode, and for digital marketing, one might treat each user interaction as an independent episode.
However, each day is not actually independent in diabetes treatment---a person's blood sugar in the morning depends on their blood sugar at the end of the previous day. Similarly, in digital marketing applications, whether or not a person clicks on an ad might depend on which ads they were shown previously (e.g., someone might be less likely to click an ad that they were shown before and did not click on then).
So, although this division into episodes is reasonable, it does not result in episodes that are completely independent, and so importance sampling will not produce consistent estimates (or estimates that can be trusted for high-confidence off-policy policy evaluation \citep{Thomas2015}).
To remedy this, we might treat all of the data from a single individual (many days, and many page visits) as a single episode, which contains nearly-independent subsequences of decisions.
To model this property, we constructed an MDP with three states, $s_1,s_2,$ and $s_3$ and two actions, $a_1$ and $a_2$. The agent always begins in $s_1$, where taking action $a_1$ causes a transition to $s_2$ with a reward of $+1$ and taking action $a_2$ causes a transition to $s_3$ with a reward of $-1$. In $s_2$, both actions lead to a terminal absorbing state with reward $-2+\epsilon$, and in $s_3$ both actions lead to a terminal absorbing state with reward $+2$. For now, let $\epsilon = 0$.
This simple MDP has a horizon of $2$ time steps---after two actions the agent is always in a terminal absorbing state.
To model the aforementioned examples, we modified this simple MDP so that whenever the agent would transition to the terminal absorbing state, it instead transitions back to $s_1$.
After visiting $s_1$ fifty times, the agent finally transitions to a terminal absorbing state.
Furthermore, to model the property that the fifty sub-episodes within the larger episode are not completely independent, we set $\epsilon=0$ initially, and $\epsilon = \epsilon + 0.01$ whenever the agent enters $s_2$.
This creates a slight dependence across the sub-episodes.
For this illustrative domain, we would like an importance sampling estimator that assumes that sub-episodes are independent when there is little data in order to reduce variance.
However, once there is enough data for the variances of estimates to be sufficiently small relative to the bias introduced by assuming that sub-episodes are independent, the importance sampling estimator should automatically begin considering longer sequences of actions, as INCRIS\ does.
We compared INCRIS\ to ordinary importance sampling (IS), per-decision importance sampling (PDIS), weighted importance sampling (WIS), and consistent weighted per-decision importance sampling (CWPDIS). The behavior policy selects actions randomly, while the evaluation policy selects action $a_1$ with a higher probability than $a_2$.
In Figure \ref{fig:toyResults} we report the mean squared errors of the different estimators using different amounts of data.
\begin{SCfigure}[2][h]
\label{fig:toyResults}
\centering
\includegraphics[scale=0.3]{LHDomainResults.pdf}
\caption{Performance of different estimators on the simple MDP that models properties of the diabetes treatment and digital marketing applications. The reported mean squared errors are the sample mean squared errors from $128$ trials. Notice that INCRIS\ achieves an order of magnitude lower mean squared error than all of the other estimators, and for some $n$ it achieves two orders of magnitude improvement over the unweighted importance sampling estimators.
}
\end{SCfigure}
\section{Conclusion}
We have shown that using options-based behavior and evaluation policies allow for lower mean squared error when using importance sampling due to their structure. Furthermore, special cases may naturally arise when using options, such as when options terminate in a fixed state distribution, and lead to greater reduction of the mean squared error.
We examined options as a first step, but in the future these results may be extended to full hierarchical policies (like the MAX-Q framework). We also generalized naturally occurring special cases with covariance testing that leads to dropping out weights in order to improve importance sampling predictions. We showed an instance of covariance testing in the algorithm INCRIS, which can greatly improve estimation accuracy for a general class of domains, and hope to derive more powerful estimators based on covariance testing that can apply to even more domains in the future.
\newpage
\section*{Acknowledgements}
The research reported here was supported in part by an ONR Young Investigator award, an NSF CAREER award, and by the Institute of Education Sciences, U.S. Department of Education. The opinions expressed are those of the authors and do not represent views of NSF, IES or the U.S. Dept. of Education.
\section{Proof of Theorem 1}
Because PDIS is an unbiased estimator of an evaluation policy's performance, its MSE is equal to its variance. To prove the theorem statement, we provide an existence proof by constructing a sample MDP where, given a particular behavior policy, there is an evaluation policy whose estimate under PDIS will have a variance that scales exponentially with the horizon $H$.
Consider a discrete state and action Markov decision process. The horizon is $H$ and the MDP has $2H + 1$ states and 2 actions. The states form two chains: the top chain has length $H+1$ and the bottom chain has length $H$. Label the states of the top chain as $x_1, \dots, x_{H+1}$, and the states of the bottom chain be $y_1, \dots, y_H$. The start state is $x_1$. An episode halts after $H$ steps.
The two actions are $a_1,a_2$. Taking action $a_1$ in the top chain deterministically transitions to the next state in the top chain i.e. from $x_i$ to $x_{i+1}$. Taking action $a_2$ in the top chain deterministically transitions to the corresponding state in the bottom chain i.e. from $x_i$ to $y_i$. The reward is zero everywhere except a reward of 1 is received for executing action $a_1$ at state $x_H$. The optimal policy is to always take action $a_1$.
Let the behavior policy $\pi_b$ be uniformly random i.e. there is always a probability of $1/2$ of picking either action. The evaluation policy is the optimal policy, $\pi_e(s)=a_1$ for all states.
Since the only nonzero reward is the single reward of $1$ at $x_H$ for action $a_1$, and it is only possible to reach that state by taking action $a_1$ for $H$ steps, PDIS reduces to a sum only over trajectories consisting solely of $H$ steps of action $a_1$, whose weights are
$\rho = \prod_{u=1}^H \frac{\pi_e(a_1 | s^{(i)}_u)}{\pi_b(a_1 | s^{(i)}_u)} = 2^H$.
The PDIS estimate of the evaluation policy is a scaled Binomial distribution where with probability $p=\frac{1}{2^H}$ a trajectory's weighted return is $2^H$ and zero otherwise.
Thus the variance of the PDIS estimate of $\pi_e$ is $\frac{1}{n} (2^H - 1)$ for $n$ trajectories, which is $\Omega(2^H)$ .
\section{Proof of Theorem 2}
We prove the above statement by constructing a MDP and selecting a behavior and target policy which will result in the stated MSE dependence on the horizon. We consider the same MDP, $\pi_b$, and $\pi_e$ as used in the proof above. For this particular MDP, the only weight that matters is the weight associated with the single final reward of the correct trajectory, so per-decision importance sampling and ordinary importance sampling are equivalent.
If the optimal trajectory does not appear in the historical data, then WIS is undefined. This is because the weight of any nonoptimal trajectory is 0, so dividing by the sum of the weights is undefined. However if we change $\pi_e$ from deterministically picking $a_1$ to picking $a_1$ with arbitrarily high probability, then the weights of nonoptimal trajectories will be arbitrarily close to zero, resulting in a WIS estimate of 0. Thus we define WIS to estimate a value of 0 when the optimal trajectory does not appear in the data. Any optimal trajectory that appears in your data will have a weight of $2^H$. Then because the weights of nonoptimal trajectories are 0, the WIS estimate will be exactly 1. Thus as soon as WIS sees at least one correct trajectory it will have the perfect estimate, otherwise the estimate will be 0. The WIS estimate is a Bernoulli distribution where the probability of 1 is the probability of at least one optimal trajectory appearing in the data.
Since the WIS estimate is Bernoulli, its variance is bounded by a constant. Furthermore the variance is small. Thus we take a closer look at the bias, since MSE is the sum of the variance and bias squared. First we compute the probability the WIS returns 1. This is the probability of at least one optimal trajectory appearing, which is equivalent to one minus the probability of no optimal trajectory appearing: $1 - \left(1- \frac{1}{2^H} \right)^n$. Thus the expected value of the WIS estimate is $1 - \left(1- \frac{1}{2^H} \right)^n$. Then the bias is $\left(1- \frac{1}{2^H} \right)^n$. Let the bias be $B$. We will compute how much data is needed to compensate for the increase in the bias when $H$ increases. Rearranging and solving for $n$ (using a taylor approximation) we get $n = \frac{\log B}{\log \left(1- \frac{1}{2^H} \right)} \approx \frac{\log B}{- \frac{1}{2^H}} \approx \Omega(2^H)$. Thus we need an exponential number of trajectories to compensate for the increase in bias when the horizon $H$ is increased. Since MSE consists partly of biased squared, we would need even more data to compensate for a squared increase in bias, but for simplicity we still use an exponential bound.
\section{Proof of Theorem 4}
Let $t^*$ be the timestep when $o_1$ terminates. Then
\begin{align}
& \mathbb{E}_{\mu_e}(J(\tau)) \\
&= \mathbb{E}_{\mu_e}\left( J(\tau_1) + J(\tau_2) \right) \\
&= \mathbb{E}_{\mu_e} (J(\tau_1)) + \mathbb{E}_{\mu_e} (J(\tau_2)) \\
&= \mathbb{E}_{\mu_e} (J(\tau_1)) + \mathbb{E}_{\mu_e} \left( \mathbb{E}_{\mu_e} (J(\tau_2) | s_{t^*} = s) \right) \label{eqn:totalexp}\\
&= \mathbb{E}_{\mu_e} (J(\tau_1)) + \sum_{s \in \mathcal{S}} \Pr(s_{t^*} = s | \mu_e) \left( \mathbb{E}_{\mu_e} (J(\tau_2) | s_{t^*} = s) \right) \\
&= \mathbb{E}_{\mu_b} (PDIS(\tau_1)) + \sum_{s \in \mathcal{S}} \Pr(s_{t^*} = s | \mu_e) \left( \mathbb{E}_{\mu_b} (PDIS(\tau_2) | s_{t^*} = s) \right) \label{eqn:PDIS} \\
&= \mathbb{E}_{\mu_b} (PDIS(\tau_1)) + \sum_{s \in \mathcal{S}} \Pr(s_{t^*} = s | \mu_b) \left( \mathbb{E}_{\mu_b} (PDIS(\tau_2) | s_{t^*} = s) \right) \label{eqn:sameterm} \\
&= \mathbb{E}_{\mu_b} (PDIS(\tau_1)) + \mathbb{E}_{\mu_b} \left( \mathbb{E}_{\mu_b} (PDIS(\tau_2) | s_{t^*} = s) \right)\\
&= \mathbb{E}_{\mu_b} (PDIS(\tau_1)) + \mathbb{E}_{\mu_b} (PDIS(\tau_2)) \label{eqn:totalexp2}
\end{align}
where eqn \ref{eqn:totalexp} follows from the law of total expectation, eqn \ref{eqn:PDIS} follows from using PDIS with a fixed initial state distribution $s_{t^*}=s$, eqn \ref{eqn:sameterm} follows because $s_{t^*}$ is the terminating state for option $o_{1}$ whose terminating state distribution stayed the same between $\mu_b$ and $\mu_e$, and eqn \ref{eqn:totalexp2} follows from the law of total expectation.
\section{Strong Consistency of INCRIS}
Given strongly consistent estimators for covariance and variance (e.g. sample covariance and sample variance), we show that INCRIS~is consistent by showing that $\sum_{t=1}^H \widehat{r_t} \overset{\text{a.s.}}{\longrightarrow} \mathbb{E}_{\pi_e}\left( \sum_{t=1}^H r_t \right)$, i.e. the total expected value under the evaluation policy. Since we have a finite sum, it is sufficient to show that for all $t$, $\widehat{r_t} \overset{\text{a.s.}}{\longrightarrow} \mathbb{E}_{\pi_e}(r_t)$.
For any $k$, because we have strongly consistent covariance and variance estimators, as $n \rightarrow \infty$ we have that $\widehat{V}_k \overset{\text{a.s.}}{\longrightarrow} \operatorname{Var}(B_k r_t)$ and since $\operatorname{Var}(B_k r_t)$ converges to zero, we also have that $\widehat{V}_k \overset{\text{a.s.}}{\longrightarrow} 0$. Similarly, $\widehat{C}_k \overset{\text{a.s.}}{\longrightarrow} \operatorname{Cov}(A_k, B_k r_t)$. Therefore $\widehat{MSE}_k \overset{\text{a.s.}}{\longrightarrow} \left( \operatorname{Cov}(A_k, B_k r_t) \right)^2$.
By eqn. \ref{eq:covtest} we have that $\operatorname{Cov}(A_k, B_k r_t) = \mathbb{E}(A_k B_k r_t) - \mathbb{E}(B_k r_t)$, which is the bias of $\mathbb{E}(A_k r_t)$.
Let $K^* = \{ k | \operatorname{Cov}(A_k, B_k r_t) = 0 \}$. Notice that $t \in K^*$ since $\widehat{r_t} = \frac{1}{n} \sum_{i=1}^n \prod_{j=1}^{t} \rho^{(i)}_j r_t$ is the ordinary importance sampling estimator, which is unbiased.
We want to show that as $n \rightarrow \infty$, the algorithm eventually picks $k' \in K^*$ and so $\widehat{r}_t$ is an unbiased estimate. To do so, let $(\Omega, \Sigma, p)$ be the probability space. Since for all $k$, $\widehat{MSE}_k \overset{\text{a.s.}}{\longrightarrow} \left( \operatorname{Cov}(A_k, B_k r_t) \right)^2$, then $p(G) = 1$ where $G = \{ \omega \in \Omega | \forall k \lim_{n \rightarrow \infty} \widehat{MSE}_k = \left( \operatorname{Cov}(A_k, B_k r_t) \right)^2 \}$.
Now we can restrict our focus to only events $\omega \in G$. First, let
\[\epsilon_{gap} = \min_{k | \operatorname{Cov}(A_k, B_k r_t) > 0} \left( \operatorname{Cov}(A_k, B_k r_t) \right)^2\]
which is the smallest nonzero MSE. Since $\lim_{n \rightarrow \infty} \widehat{MSE}_k = \left( \operatorname{Cov}(A_k, B_k r_t) \right)^2$, by definition of limit, we have that there exists $n_0$ such that for all $n \geq n_0$, $|\widehat{MSE}_k - \left( \operatorname{Cov}(A_k, B_k r_t) \right)^2| < \frac{\epsilon_{gap}}{3}$. By the definition of $\epsilon_{gap}$, for $n \geq n_0$, $k' \in K^*$.
We have shown that the algorithm eventually picks $k' \in K^*$. Next we will show that this implies $\widehat{r_t} \overset{\text{a.s.}}{\longrightarrow} \mathbb{E}_{\pi_e}(r_t)$.
By the strong law of large numbers, for any $k$, $\left( \frac{1}{n} \sum_{i=1}^n B_k^{(i)} r_t \right) \overset{\text{a.s.}}{\longrightarrow} \mathbb{E}(B_{k} r_t)$. Then we know that $p(G') = 1$ where $G' = \{ \omega \in \Omega | \lim_{n \rightarrow \infty} \left( \frac{1}{n} \sum_{i=1}^n B_{k}^{(i)} r_t \right) = \mathbb{E}(B_{k} r_t) \}$. Then $p(G \cap G') = 1$, and we can restrict our focus to $\omega \in (G \cap G')$. We have already shown that when $\omega \in (G \cap G')$, there exists $n_0$ such that for all $n \geq n_0$, $k' \in K^*$. But we also know that when $\omega \in (G \cap G')$, $\lim_{n \rightarrow \infty} \left( \frac{1}{n} \sum_{i=1}^n B_{k'}^{(i)} r_t \right) = \mathbb{E}(B_{k'} r_t)$ for any $k'$, so for $k' \in K^*$, $\lim_{n \rightarrow \infty} \left( \frac{1}{n} \sum_{i=1}^n B_{k'}^{(i)} r_t \right) = \mathbb{E}(B_{k'} r_t) = \mathbb{E}(B_{t} r_t) = \mathbb{E}_{\pi_e}(r_t)$. Therefore, for all $\epsilon > 0$, we will always be able to find some $n_0$ large enough such that for all $n \geq n_0$, when $\omega \in (G \cap G')$, we have $k' \in K^*$ and $\widehat{r_t} = \left( \frac{1}{n} \sum_{i=1}^n B_{k'}^{(i)} r_t \right)$ and therefore $|\widehat{r_t} - \mathbb{E}(B_{t} r_t)| < \epsilon$.
Thus we have shown $\widehat{r_t} \overset{\text{a.s.}}{\longrightarrow} \mathbb{E}_{\pi_e}(r_t)$, and so INCRIS~is strongly consistent.
|
\subsection*{#1}}{\qed\medskip}
\theoremstyle{definition}
\newtheorem{discussion}{Discussion}
\icmltitlerunning{Density Level Set Estimation on Manifolds with DBSCAN}
\begin{document}
\twocolumn[
\icmltitle{Density Level Set Estimation on Manifolds with DBSCAN}
\begin{icmlauthorlist}
\icmlauthor{Heinrich Jiang}{a}
\end{icmlauthorlist}
\icmlaffiliation{a}{Google}
\icmlcorrespondingauthor{Heinrich Jiang}{<EMAIL>}
\icmlkeywords{}
\vskip 0.3in
]
\printAffiliationsAndNotice{}
\begin{abstract}
We show that DBSCAN can estimate the connected
components of the $\lambda$-density level set $\{ x : f(x) \ge \lambda\}$ given $n$ i.i.d. samples from an unknown density $f$.
We characterize the regularity of the level set boundaries using parameter $\beta > 0$ and analyze the estimation error under the Hausdorff metric.
When the data lies in $\mathbb{R}^D$ we obtain a rate of $\widetilde{O}(n^{-1/(2\beta + D)})$, which matches known lower bounds
up to logarithmic factors. When the data lies on an embedded unknown $d$-dimensional manifold in $\mathbb{R}^D$, then we obtain a rate of
$\widetilde{O}(n^{-1/(2\beta + d\cdot \max\{1, \beta \})})$.
Finally, we provide adaptive parameter tuning in order to attain these rates with no a priori knowledge
of the intrinsic dimension, density, or $\beta$.
\end{abstract}
\section{Introduction}
\input{Introduction.tex}
\section{Overview}
\input{Overview.tex}
\section{The connection to neighborhood graphs} \label{dbscan-knn}
\input{Theory.tex}
\section{Manifold Setting} \label{manifoldsetting-section}
\input{ManifoldTheory.tex}
\section{Consistency and Rates} \label{consistency-section}
\input{Consistency.tex}
\section{Removal of False Clusters} \label{pruning-section}
\input{Pruning.tex}
\section{Adaptive Parameter Tuning} \label{adaptive-section}
\input{Parameters.tex}
\section{Full Dimensional Setting} \label{fulldimensionalsetting-section}
\input{fulldimension}
\section{Conclusion}
\input{Conclusion}
\section*{Acknowledgements}
The author is grateful to Samory Kpotufe for insightful discussions and to the anonymous reviewers for their useful feedback.
\section{Proof of Ball Volume Result}
\begin{proof}[Proof of Lemma~\ref{ballvolume}]
For the lower bound, we have by Lemma 5.3 of \cite{lowerBoundBall} that
\begin{align*}
\text{vol}_d(B(x, r) \cap M) \ge v_d\cdot r^d\cdot \cos\left(\arcsin\left(\frac{r}{2\tau}\right)\right) \ge v_d\cdot r^d\cdot (1 - r^2\tau^2),
\end{align*}
where the last inequality holds for $r < 1/\tau$. For the upper bound, we have by \cite{upperBoundBall} that
\begin{align*}
\text{vol}_d(B(x, r) \cap M) \le v_d\cdot r^d\cdot \left( \frac{1}{2\sqrt{1 - 2r/\tau} - 1}\right)^d \le v_d\cdot r^d\cdot (1 + 4dr/\tau),
\end{align*}
where the last inequality holds when $r < \tau/4d$.
\end{proof}
\section{Proof of Density Estimation Results}
\begin{proof}[Proof of Lemma~\ref{fk_upper_bound}]
Choose $r$ such that
\begin{align*}
(1 + 4dr/\tau) \cdot v_d \cdot r^d \cdot (f(x) + \epsilon)
= \frac{k}{n} - C_{\delta, n} \frac{\sqrt{k}}{n}
\le v_d \cdot \hat{r}(\epsilon, x)^d \cdot (f(x) + \epsilon).
\end{align*}
Thus, $r \le \hat{r}(\epsilon, x)$ and hence
\begin{align*}
\mathcal{F}(B(x, r)) \le \text{vol}_d(B(x, r) \cap M) \cdot (f(x) + \epsilon)
\le (1 + 4dr/\tau) \cdot v_d \cdot r^d \cdot (f(x) + \epsilon)
= \frac{k}{n} - C_{\delta, n} \frac{\sqrt{k}}{n},
\end{align*}
where the second inequality follows by Lemma~\ref{ballvolume}.
Thus, by Lemma~\ref{ballbounds}, we have $\mathcal{F}_n(B(x, r)) < k/n$ and hence $r_k(x) > r$. Therefore,
\begin{align*}
f_k(x) < \frac{k}{nv_dr^d} = \frac{1 + 4dr/\tau}{1 - C_{\delta, n}/\sqrt{k}} (f(x) + \epsilon)
\le \left( 1 + 2\frac{C_{\delta, n}}{\sqrt{k}} + 4dr/\tau \right) (f(x) + \epsilon)
\le \left( 1 + 3\frac{C_{\delta, n}}{\sqrt{k}} \right) (f(x) + \epsilon).
\end{align*}
The last inequality holds because $r \le (k/(n\cdot v_d \cdot (f(x) + \epsilon)))^{1/d}$ and choosing $C_1$ appropriately we have $4rd/\tau \le
4d (k/(n\cdot v_d \cdot \lambda_0))^{1/d} /\tau < C_{\delta, n}/\sqrt{k}$.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{fk_lower_bound}]
Let $x \in X$ such that $f(x) \ge \lambda_0$. Choose $r$ such that
\begin{align*}
(1 - \tau^2 r^2) \cdot v_d \cdot r^d \cdot (f(x) - \epsilon) = \frac{k}{n} + C_{\delta, n} \frac{\sqrt{k}}{n}.
\end{align*}
We have $r \le c (2k/(n\cdot v_d \cdot (f(x) - \epsilon)))^{1/d}$ for some constant $c$ depending on $d$ and $\tau$.
Thus, choosing $C_2$ approrpiately, we have $\tau^2r^2 < \frac{1}{4}$. Thus,
\begin{align*}
v_d \cdot r^d \cdot (f(x) - \epsilon) \le \frac{4}{3} \left( \frac{k}{n} + C_{\delta, n} \frac{\sqrt{k}}{n} \right) \le v_d \cdot \check{r}(\epsilon, x)^d \cdot (f(x) - \epsilon).
\end{align*}
Thus $r < \check{r}(\epsilon, x)$ and we obtain:
\begin{align*}
\mathcal{F}(B(x, r)) \ge \text{vol}_d(B(x, r) \cap M) \cdot (f(x) - \epsilon)
\le (1 - \tau^2 r^2) \cdot v_d \cdot r^d \cdot (f(x) - \epsilon)
= \frac{k}{n} + C_{\delta, n} \frac{\sqrt{k}}{n}.
\end{align*}
Thus, by Lemma~\ref{ballbounds}, we have $\mathcal{F}_n(B(x, r)) \ge k/n$ and hence $r_k(x) \le r$.
Therefore,
\begin{align*}
f_k(x) &\ge \frac{k}{n\cdot v_d\cdot r^d} = \frac{1 - \tau^2 r^2}{ 1 + C_{\delta, n}/\sqrt{k}} (f(x) - \epsilon) \ge (1 - \tau^2 r^2 - C_{\delta,n}/\sqrt{k}) \cdot (f(x) - \epsilon) \ge \left(1 - 3 \cdot \frac{C_{\delta, n}}{\sqrt{k}} \right)\cdot (f(x) - \epsilon),
\end{align*}
where the last inequality follows $\tau^2r^2 \le \tau^2 c^2 (2k/(n\cdot v_d \cdot \lambda_0))^{2/d} \le \frac{C_{\delta, n}}{\sqrt{k}}$ where
the latter holds for $C_2$ chosen appropriately. The result follows.
\end{proof}
\section{Proof of Separation Result}
\begin{proof} [Proof of Lemma~\ref{separation}]
Define $\bar{r} := \min \{r_c, r_s, \frac{1}{2} \tau\}$.
We show that $(M \backslash \mathcal{X}_C) \cap X$ and $C^{\oplus \bar{r}} \cap X$ are disconnected in $G(k, \varepsilon)$.
To do this, it suffices to show that:
\begin{itemize}
\item (1) $G(k, \varepsilon)$ has no point in $\mathcal{X}_C \backslash C^{\oplus \bar{r}}$.
\item (2) $G(k, \varepsilon)$ has no point in $S_C^{\oplus r_s/2}$.
\item (3) $G(k, \varepsilon)$ has no edge connecting a point in $C^{\oplus \bar{r}}$ to a point in $\mathcal{X}_C \backslash C^{\oplus \bar{r}}$.
\end{itemize}
We begin by showing (1). Define $\lambda' := \lambda - \check{C}_\beta(\bar{r}/2)^\beta$. Thus, for $x \in C^{\oplus \bar{r}} \cap X$, we have $\hat{r}(\lambda' - f(x), x) \ge \bar{r}/2$. Thus the conditions for Lemma~\ref{fk_upper_bound} are satisfied as long as $K_l$ and $K_u$ are appropriately large and small, respectively. Hence,
\begin{align*}
\sup_{x \in X \cap (\mathcal{X}_C \backslash C^{\oplus \bar{r}})} f_k(x)
\le \left(1 + 3\cdot \frac{C_{\delta, n}}{\sqrt{k}}\right) (\lambda - \check{C}_\beta (\bar{r}/2)^\beta)
< \lambda - \frac{C_{\delta, n}^2}{\sqrt{k}} \cdot \lambda,
\end{align*}
where the last inequality holds for $K_l$ sufficiently large. Thus, it is now clear that $r_k(x) > \varepsilon$ for $x \in X \cap (\mathcal{X}_C \backslash C^{\oplus \bar{r}})$, showing (1). Next, if $x \in S_C^{\oplus r_s/2}$, then $\hat{r}(\lambda' - f(x), x) \ge \bar{r}/2$ and the same holds for sample points in $S_C^{\oplus r_s/2}$ implying (2).
To show (3), it suffices to show that any such edge will have length less than $r_s$ since $S_C^{\oplus r_s/2}$ separates $C^{\oplus \bar{r}}$ and $\mathcal{X}_C \backslash C^{\oplus \bar{r}}$ by length at least $r_s$. Indeed, for $x \in X \cap C^{\oplus \bar{r}}$,
\begin{align*}
\mathcal{F}(B(x, \bar{r})) \ge \text{vol}_d(M \cap B(x, \bar{r})) \inf_{x' \in M \cap B(x, 2\bar{r})} f(x')
\ge (1 - \tau^2\bar{r}^2) \cdot v_d\cdot \bar{r}^d (f(x) - \hat{C}_\beta\bar{r}^\beta)
\ge \frac{k}{n} + C_{\delta, n}\frac{\sqrt{k}}{n},
\end{align*}
where the second inequality follows from Lemma~\ref{ballvolume}, and the last inequality holds when $K_u$ is sufficiently small.
By Lemma~\ref{ballbounds}, we have $r_k(x) \le \bar{r} < r_s$, establishing (3). Thus, $(M \backslash \mathcal{X}_C) \cap X$ and $C^{\oplus \bar{r}} \cap X$ are disconnected in $G(k, \varepsilon)$.
It is easy to see that $C^{\oplus \bar{r}}$ contains a sample point and thus there exists a connected component, $\overline{C}$
of $G(k, \varepsilon)$ such that $\overline{C} \subseteq \mathcal{X}_C$.
Now, by Lemma~\ref{dbscan}, we have that there exists $\widehat{C} \in \widehat{\mathcal{C}}_\lambda$ such that
$\widehat{C} = \{ x \in X : d(x, C) \le \varepsilon \}$. Since $\overline{C}$ has no intersection with $S_C^{\oplus r_s/2}$, it follows that as long as $\varepsilon < r_s/2$, then $\widehat{C} \subseteq \mathcal{X}_C$. For $K_u$ chosen sufficiently small, we have $\varepsilon < r_s/2$, as desired.
\end{proof}
\section{Proof of Connectedness}
\begin{proof} [Proof of Lemma~\ref{connectedness}]
Define $A := C^{\oplus r_n(C)}$.
In light of Lemma~\ref{dbscan}, it suffices to show that $A \cap X$ is connected in $G(k, \varepsilon)$.
Define $r_o := (k/(2nv_d ||f||_\infty))^{1/d}$. We next show that for each $x \in A^{\oplus \varepsilon}$,
we have a sample point in $B(x, r_o)$. Indeed, for any $z \in B(x, r_0/3) \cap \mathcal{N}$ (while taking $K_l$ sufficiently large so that $r_0/3 > 1/n$):
\begin{align*}
\mathcal{F}(B(z, r_o/2)) &\ge \text{vol}_d(B(z, r_o/2) \cap M) \inf_{x' \in B(z, r_o/2 + \varepsilon )} f(x') \\
&\ge (1 - \tau^2 (r_o/2)^2) \cdot v_d \cdot (r_o/2)^d \cdot (\lambda - \hat{C}_\beta (r_o /2+ \varepsilon + r_n(C))^\beta) \\
&\ge (1 - \tau^2 (r_o/2)^2) \cdot v_d \cdot (r_o/2)^d \cdot (\lambda - \hat{C}_\beta (3\cdot \varepsilon)^\beta) \ge C_{\delta, n} \frac{\sqrt{d \log n}}{n},
\end{align*}
with the last inequality holding when $K_u$ is chosen sufficiently small so that the $\tau^2 (r_o/2)^2$ and $\hat{C}_\beta (3\cdot \varepsilon)^\beta$ terms
become small enough, and $K_l$ is chosen sufficiently large so that the $(r_o/2)^d$ factor will be large enough.
Thus by Lemma~\ref{ballbounds} we have that with probability at least $1 - \delta$, $B(z, r_o/2) \subseteq B(x, r_o)$ contains a sample.
Now, let $x$ and $x'$ be two points in $A\cap X$. We show there exists a path $x = x_0,x_1,...,x_p = x'$ such that $||x_i - x_{i+1}|| < r_o$ and $x_i \in B(A, r_o)$. For arbitrary $\gamma \in (0, 1)$, we can choose $x = z_0, z_1,...,z_p = x'$ where $||z_{i+1} - z_i|| \le \gamma r_o$. Next, for $K_l$ sufficiently large, there exists $\gamma$ sufficiently small such that
\begin{align*}
\left(1 - \tau^2 \frac{(1-\gamma)^2r_o^2}{4}\right) \cdot v_d\left( \frac{(1-\gamma)r_o}{2}\right)^d \inf_{z \in B(A, r_0)} f(z) \ge \frac{C_{\delta, n}\sqrt{d\log n}}{n}.
\end{align*}
Therefore by Lemma~\ref{ballbounds}, there exists a sample point $x_i$ in $B(z_i, (1-\gamma)r_o/2)$ and
\begin{align*}
||x_{i+1} - x_i|| \le ||x_{i+1} - z_{i+1}|| + ||z_{i+1} - z_i|| + ||z_i -x_i|| \le r_o.
\end{align*}
\noindent All that remains is to show $(x_i, x_{i+1}) \in G(k, \varepsilon)$.
We see that $x_i \in B(A, r_o)$ for each $i$ and for any $x \in B(A, r_o)$, we have
\begin{align*}
\mathcal{F}(B(x, \varepsilon)) &\ge (1 - \tau^2 \varepsilon^2) \cdot v_d\cdot \varepsilon^d \inf_{x' \in B(x, r_o + \varepsilon)} f(x')
\ge (1 - \tau^2 \varepsilon^2) \cdot v_d \cdot \varepsilon^d \cdot (\lambda - \hat{C}_\beta (2r_o + \varepsilon)^\beta)
\ge \frac{k}{n} + \frac{C_{\delta, n} \sqrt{k}}{n},
\end{align*}
where the last inequality holding when $K_u$ is chosen sufficiently small so that the $\tau^2 \varepsilon^2$ and $\hat{C}_\beta (2r_o +\varepsilon)^\beta$
become small enough, and $K_l$ is chosen sufficiently large so that the $\varepsilon^d$ factor will be large enough.
Therefore, $r_k(x_i) \le \varepsilon$ and so $x_i \in G(k, \varepsilon)$ for all $x_i$. Finally, $||x_{i+1} - x_i|| \le r_o \le \varepsilon$. Hence, $(x_i, x_{i+1}) \in G(k, \varepsilon)$. The result immediately follows.
\end{proof}
\section{Proof of Theorem~\ref{pruning}}
\begin{proof} [Proof Sketch]
Define $x \in \widehat{C}$ as a core-point if $r_k(x) \le \varepsilon$. For any core point $x \in \widehat{C}$, we have
$f_k(x) \ge \lambda - \frac{C_{\delta, n}^2}{\sqrt{k}}\lambda$. With $K_l, K_u$ chosen appropriately, we can apply Lemma~\ref{fk_upper_bound} with $\epsilon = \frac{C_{\delta, n}}{\sqrt{k}} \lambda$ to obtain
\begin{align*}
f(x) &\ge \frac{f_k(x)}{1 + 3 \cdot C_{\delta, n}/\sqrt{k}} - \epsilon
\ge \left(1 - 5\cdot \frac{C_{\delta, n}}{\sqrt{k}}\right) \cdot \lambda.
\end{align*}
Defining $r := (5\lambda \cdot C_{\delta, n} / (\check{C}_\beta \sqrt{k}))^{1/\beta}$, we have that $x \in C^{\oplus r}$ for some $C \in \mathcal{C}_\lambda$ when $r < \min \{r_c, (\gamma/\hat{C}_\beta)^{1/\beta} \}$, which is achieved as long as $K_l$ is chosen sufficiently large.
Since any core point of $\widehat{C}$ is in $C^{\oplus r}$, then in
light of Theorem~\ref{hausdorfferror} and Lemma~\ref{dbscan}, it suffices to show that
separate CCs of $\mathcal{C}_\lambda$ do not get merged in $G(k, \tilde\varepsilon)$ (separation) and that
$C^{\oplus r} \cap X$ appear in the same connected component of $G(k, \tilde\varepsilon)$ (connectedness).
These follow from Lemma~\ref{separation} and~\ref{connectedness}, but with the minor modification that $\varepsilon$ is replaced by $\tilde{\varepsilon}$, which requires $K_l$ and $K_u$ to be adjusted accordingly to hold. Otherwise the arguments are the same.
\end{proof}
\section{Full Dimensional Setting}
The analysis for this situation is largely the same as under the manifold assumption. We will only highlight the main difference, which is
in the density estimation bounds.
We can utilize such bounds from \cite{optimalknn}, which are repeated here.
\begin{lemma} [Lemma 3 of \cite{optimalknn}]
Suppose that $k \ge 4 C_{\delta, n}^2$. Then with probability at least $1-\delta$, the following holds for all $x\in \mathbb{R}^d$ and $\epsilon > 0$.
\begin{align*}
f_k(x) < \left(1 + 2 \frac{C_{\delta, n}}{\sqrt{k}} \right) (f(x) + \epsilon),
\end{align*}
provided $k$ satisfies $v_d \cdot \hat{r}(x, \epsilon) \cdot (f(x) + \epsilon) \ge \frac{k}{n} + C_{\delta, n}\frac{\sqrt{k}}{n}$.
\end{lemma}
\begin{lemma} [Lemma 4 of \cite{optimalknn}]
Suppose that $k \ge 4 C_{\delta, n}^2$. Then with probability at least $1-\delta$, the following holds for all $x\in \mathbb{R}^d$ and $\epsilon > 0$.
\begin{align*}
f_k(x) \ge \left(1 -\frac{C_{\delta, n}}{\sqrt{k}} \right) (f(x) - \epsilon),
\end{align*}
provided $k$ satisfies $v_d \cdot \check{r}(x, \epsilon) \cdot (f(x) - \epsilon) \ge \frac{k}{n} - C_{\delta, n}\frac{\sqrt{k}}{n}$.
\end{lemma}
Unlike Lemma~\ref{fk_upper_bound} and~\ref{fk_lower_bound}, these results don't require $k \lesssim n^{2/(2+d)}$ and $k \lesssim n^{4/(4+d)}$, respectively.
This allows us to take $k \lesssim n^{2\beta/(2\beta + d)}$ rather than $k \lesssim \min \{n^{2/(2 + d)}, n^{2\beta/(2\beta + d)} \} = n^{2\beta'/(2\beta' + d)}$
in the analysis. Otherwise, the analysis is nearly identical up to constant factors.
\subsection{Level-Set Conditions}
Much of the results will depend on the behavior of level set boundaries. Thus, we require sufficient drop-off at the boundaries, as well as separation between the CCs at a particular level set. We give the following notion of separation.
\begin{definition}
$A, A'$ are $r$-separated in $M$ if there exists a set $S$ such that every path from $A$ to $A'$ intersects $S$ and
$\sup_{x \in M \cap (S + B(0, r))} f(x) < \inf_{x \in A \cup A'} f(x)$.
\end{definition}
Define the following shorthands for distance from a point to a set, the intersection of $M$ with a neighborhood around a set under the Euclidean distance,
and the largest Euclidean distance from a point in a set to its closest sample point.
\begin{definition}
$d(x, A) := \inf_{x' \in A} |x - x'|$,
$C^{\oplus r} := \{x \in M : d(x, C) \le r \}$,
$r_n(C) := \sup_{c \in C} d(c, X)$.
\end{definition}
We have the following mild assumptions which ensures that the CCs can be separated from the rest of the density
by sufficiently wide valleys and there is sufficient decay around the level set boundaries.
\begin{assumption} [Separation Conditions] \label{clusterProperties}
Let $\lambda > 0$ and $\mathcal{C}_\lambda$ be a CCs of $\{ x \in M : f(x) \ge \lambda \}$.
There exists $\check{C}_\beta, \hat{C}_\beta, \beta, r_s, r_c > 0$ and $0 < \lambda_0 < \lambda$ such that the following holds:
For each $C \in \mathcal{C}_\lambda$, there exists $A_C$, a connected component of $M^{\lambda_0} := \{ x \in M : f(x) \ge \lambda_0 \}$
such that:
\vspace{-0.2cm}
\begin{itemize}
\setlength\itemsep{0.0em}
\item $A_C$ separates $C$ by a valley: $A_C$ does not intersect with any other CC in $\mathcal{C}_\lambda$;
$A_C$ and $M^{\lambda_0} \backslash A_C$ are $r_s$-separated by some $S_C$.
\item $C^{\oplus r_c} \subseteq A_C$.
\item $\beta$-regularity: For $x \in C^{\oplus r_c} \backslash C$, we have
\begin{align*}
\check{C}_{\beta} \cdot d(x, C)^{\beta} \le \lambda - f(x) \le \hat{C}_{\beta} \cdot d(x, C)^{\beta}.
\end{align*}
\end{itemize}
\end{assumption}
\begin{remark}
We can choose any $0 < \beta < \infty$. The $\beta$-regularity assumption appears in e.g. \cite{adaptive}. This is very general and
also allows us to make a separate global
smoothness assumption
\end{remark}
\begin{remark}
We currently characterize the smoothness w.r.t. the Euclidean distance. One could alternatively use the geodesic distance
on $M$, $d_M(p, q)$. It follows from Proposition 6.3 of \citet{lowerBoundBall} that when $|p - q| < \tau/4$, we have $|p - q| \le d_M(p ,q) \le 2 |p - q|$. Since the distances we deal in our analysis with are of such small order,
these distances can thus essentially be treated as equivalent. We use the Euclidean distance throughout the paper for simplicity.
\end{remark}
\begin{remark}
For the rest of this paper, it will be understood that Assumptions~\ref{manifold},~\ref{density}, and~\ref{clusterProperties} hold.
\end{remark}
We can define a region which isolates $C$ away from other clusters of $\{ x \in M : f(x) \ge \lambda \}$.
\begin{definition}
$\mathcal{X}_{C} := \{ x : \exists \text{ a path }\mathcal{P} \text{ from } x \text{ to } x'\in C \text{ such that } \mathcal{P} \cap S_C = \emptyset\}$.
\end{definition}
\subsection{Parameter Settings}\label{parameters}
Fix $\lambda > 0$ and $\delta > 0$. Let $k$ satisfy the following
\begin{align*}
K_l \cdot (\log n)^2 \le k \le K_u \cdot (\log n)^{2d/(2+d)} \cdot n^{2\beta' / (2\beta' + d)},
\end{align*}
where $\beta' := \min\{1, \beta \}$, and $K_l$ and $K_u$ are positive constants depending on $\delta, \check{C}_\beta, \hat{C}_\beta, \beta, \tau, d, ||f||_\infty, \lambda_0, r_s, r_c$
which are implicit in the proofs later in this section.
The remainder of this section will be to show that DBSCAN(minPts, $\varepsilon$) with
\begin{align*}
&\text{minPts} = k, \text{ } \varepsilon = \left(\frac{k}{n\cdot v_d \cdot (\lambda - \lambda \cdot C_{\delta, n}^2/\sqrt{k})} \right)^{1/d}
\end{align*}
will consistently estimate each CC of
$\{ x \in M : f(x) \ge \lambda \}$.
Throughout the text, we denote $\widehat{\mathcal{C}_\lambda}$ as the clusters returned by DBSCAN under this setting.
\subsection{Separation and Connectedness}
Take $C \in \mathcal{C}_\lambda$. We show that DBSCAN will return an estimated CC
$\widehat{C}$, such that $\widehat{C}$ does not contain any points outside of $\mathcal{X}_C$.
Then, we show that $\widehat{C}$ contains all the sample points in $C$.
The proof ideas used are similar to that of standard results in cluster trees estimation; they can be found in the appendix.
\begin{lemma} [Separation] \label{separation}
There exists $K_l$ sufficiently large and $K_u$ sufficiently small such that the following holds with probability at least $1 - \delta$. Let $C \in \mathcal{C}_\lambda$. There exists $\widehat{C} \in \widehat{\mathcal{C}_\lambda}$ such that
$\widehat{C} \subseteq \mathcal{X}_C$.
\end{lemma}
\begin{lemma} [Connectedness]\label{connectedness}
There exists $K_l$ sufficiently large and $K_u$ sufficiently small such that the following holds with probability at least $1 - \delta$. Let $C \in \mathcal{C}_\lambda$. If there exists $\widehat{C} \in \widehat{\mathcal{C}_\lambda}$ such that
$\widehat{C} \subseteq \mathcal{X}_C$, then $C^{\oplus r_n(C)} \cap X \subseteq \widehat{C}$.
\end{lemma}
\begin{remark}
These results allow $C$ to have any dimension between $0$ to $d$ since we reason with $C^{\oplus r_n(C)}$, which contains samples, instead of simply $C$.
\end{remark}
\subsection{Hausdorff Error}
We give the estimation rate under the Hausdorff metric.
\begin{definition} [Hausdorff Distance]
\begin{align*}
d_{\text{Haus}}(A, A') = \max \{ \sup_{x \in A} d(x, A'), \sup_{x' \in A'} d(x', A) \}.
\end{align*}
\end{definition}
\begin{theorem} \label{hausdorfferror}
There exists $K_l$ sufficiently large and $K_u$ sufficiently small such that the following holds with probability at least $1 - \delta$.
For each $C \in \mathcal{C}_\lambda$, there exists $\widehat{C} \in \widehat{\mathcal{C}_\lambda}$ such that
\begin{align*}
d_{\text{Haus}}(C, \widehat{C}) \le 2\cdot (4\lambda / \check{C}_\beta )^{1/\beta} \cdot C_{\delta, n}^{2/\beta} \cdot k^{-1/2\beta}.
\end{align*}
\end{theorem}
\begin{proof}
For $K_l$ and $K_u$ appropriately chosen, we have Lemma~\ref{separation} and Lemma~\ref{connectedness} hold. Thus we have for $C \in \mathcal{C}_\lambda$, there exists $\widehat{C} \in \widehat{\mathcal{C}_\lambda}$ such that
\begin{align*}
C^{\oplus r_n(C)} \cap X \subseteq
\widehat{C} \subseteq \bigcup_{\substack{ x\in \mathcal{X}_C \cap X\\ f_k(x) \ge \lambda - \frac{C_{\delta, n}^2}{\sqrt{k}} \lambda}} B(x, \varepsilon) \cap M.
\end{align*}
Define $\bar{r} := \left(\frac{4\lambda \cdot C_{\delta, n}^2}{\check{C}_\beta\cdot \sqrt{k}}\right)^{1/\beta}$. We show that $d_{\text{Haus}}(C, \widehat{C}) \le \bar{r}$, which involves two directions to show from the Hausdroff metric: that $\max_{x \in \widehat{C}} d(x, C) \le \bar{r}$ and $\sup_{x \in C} d(x, \widehat{C}) \le \bar{r}$.
We start by proving $\max_{x \in \widehat{C}} d (x, C) \le \bar{r}$.
Define $r_0 = \bar{r}/2$. We have
\begin{align*}
r_0 = \frac{1}{2} \left(\frac{4\cdot C_{\delta, n}^2}{\check{C}_\beta \cdot \sqrt{k}} \right)^{1/\beta} \ge \left(\frac{k}{v_d n \lambda_0}\right)^{1/d} \ge \varepsilon,
\end{align*}
where the first inequality holds when $K_u$ is chosen sufficiently small, and
the last inequality holds because $\lambda_0 < \lambda - \frac{C_{\delta, n}^2}{\sqrt{k}} \lambda$.
Hence $r_0 + \varepsilon \le \bar{r}$. Therefore, it suffices to show
\begin{align*}
\sup_{ x \in (\mathcal{X}_C \backslash C^{\oplus r_0}) \cap X} f_k(x) < \lambda - \frac{C_{\delta, n}^2}{\sqrt{k}} \lambda.
\end{align*}
We have that for $x \in (\mathcal{X}_C \backslash C^{\oplus r_0/2}) \cap X$, $f(x) \le \lambda - \check{C}_\beta(r_0/2)^\beta := \lambda'$.
Thus, for any $x \in (\mathcal{X}_C \backslash C^{\oplus r_0}) \cap X$ and letting $\epsilon = \lambda' - f(x)$, we have
\begin{align*}
\hat{r}(\epsilon, x) &\ge r_0 / 2
\ge (4\lambda_0 C_{\delta, n}/(\sqrt{k} \cdot \check{C}_{\beta}))^{1/\beta} / 2.
\end{align*}
For $K_u$ chosen sufficiently small, the last
equation will be large enough (i.e. of order $(k/v_d n \lambda)^{1/d}$) so that the conditions
of Lemma~\ref{fk_upper_bound} hold. Thus, applying this for each $x \in (\mathcal{X}_C \backslash C^{\oplus r_0}) \cap X$, we obtain
\begin{align*}
\sup_{x \in (\mathcal{X}_C \backslash C^{\oplus r_0}) \cap X} f_k(x) < \left(1 + 3 \frac{C_{\delta, n}}{\sqrt{k}} \right) (\lambda - \check{C}_\beta (r_0/2)^\beta).
\end{align*}
We have the r.h.s. is at most $\lambda - \frac{C_{\delta, n}^2}{\sqrt{k}} \lambda$ for $K_u$ chosen appropriately
and the first direction follows.
We now turn to the other direction, that $\sup_{x \in C} d (x, \widehat{C}) \le \bar{r}$. Let $x \in C$. Then there exists sample point $x' \in B(x, r_n(C))$ by definition of $r_n$ and we have that $x' \in \widehat{C}$. Finally, $r_n(C) \le \bar{r}$ for $K_l$ sufficiently large, and thus $|x' - x| \le \bar{r}$. The result follows.
\end{proof}
\begin{remark}
When taking $k \approx n^{2\beta'/(2\beta' + d)}$, we obtain the error rate of $d_{\text{Haus}}(C, \widehat{C}) \approx n^{-1/(2\beta + d\cdot \max\{1, \beta\})}$, ignoring logarithmic factors.
When $0 < \beta \le 1$, this matches the known lower bound established in Theorem 4 of \citet{tysbakovMinimax}.
However, we do not obtain this rate when $\beta > 1$. In this case, the density estimation error will be of order at least $n^{-1/(2+d)}$ due in part to
the error from resolving the geodesic balls with Euclidean balls. This does not arise in the full dimensional setting, which will be described later.
\end{remark}
\subsection{Setup}
We make the following regularity assumptions which are standard among works
on manifold learning e.g. \citep{manifold07, manifold12, balakrishnan2013cluster}.
\begin{assumption}\label{manifold}
$\mathcal{F}$ is supported on $M$ where:
\vspace{-0.3cm}
\begin{itemize}
\setlength\itemsep{0.0em}
\item $M$ is a $d$-dimensional smooth compact Riemannian manifold without boundary embedded in compact subset $\mathcal{X} \subseteq \mathbb{R}^D$.
\item The volume of $M$ is bounded above by a constant.
\item $M$ has condition number $1/\tau$, which controls the curvature and prevents self-intersection.
\end{itemize}
\vspace{-0.3cm}
Let $f$ be the density of $\mathcal{F}$ with respect to the uniform measure on $M$.
\end{assumption}
\begin{assumption}\label{density}
$f$ is continuous and bounded.
\end{assumption}
\subsection{Basic Supporting Bounds}
The following result bounds the empirical mass of Euclidean balls to the true mass under $f$.
It is a direct consequence of Lemma 6
of \citet{balakrishnan2013cluster}.
\begin{lemma} [Uniform convergence of empirical Euclidean balls (Lemma 6 of \citet{balakrishnan2013cluster})] \label{ballbounds}
Let $\mathcal{N}$ be a minimal fixed set such that each point in $M$ is at most distance $1/n$ from some point in $\mathcal{N}$.
There exists a universal constant $C_0$ such that the following holds with probability at
least $1 - \delta$.
For all $x \in X \cup \mathcal{N}$,
\begin{align*}
\mathcal{F}(B) \ge C_{\delta, n} \frac{\sqrt{d \log n}}{n} &\Rightarrow \mathcal{F}_n(B) > 0\\
\mathcal{F}(B) \ge \frac{k}{n} + C_{\delta, n} \frac{\sqrt{k}}{n} &\Rightarrow \mathcal{F}_n(B) \ge \frac{k}{n} \\
\mathcal{F}(B) \le \frac{k}{n} - C_{\delta, n}\frac{\sqrt{k}}{n} &\Rightarrow \mathcal{F}_n(B) < \frac{k}{n}.
\end{align*}
where $C_{\delta, n} = C_0 \log(2/\delta) \sqrt{d \log n}$, $\mathcal{F}_n$ is the empirical distribution, and $k \ge C_{\delta, n}$.
\end{lemma}
\begin{remark}
For the rest of the paper, many results are qualified to hold with probability at least $1 - \delta$. This is
precisely the event in which Lemma~\ref{ballbounds} holds.
\end{remark}
\begin{remark}
If $\delta = 1/n$, then $C_{\delta, n} = O((\log n)^{3/2})$.
\end{remark}
Next, we need the following bound on the volume of the intersection Euclidean ball and $M$; this is required to get a handle on the true mass of the ball under
$\mathcal{F}$ in later arguments. The upper and lower bounds follow from \citet{upperBoundBall} and Lemma 5.3 of \citet{lowerBoundBall}. The proof is given in the appendix.
\begin{lemma} [Ball Volume] \label{ballvolume}
If $0 < r < \min\{\tau/4d, 1/\tau\}$, and $x \in M$ then
\begin{align*}
v_d r^d (1 - \tau^2 r^2) \le \text{vol}_d(B(x, r) \cap M) \le v_d r^d (1 + 4dr/\tau).
\end{align*}
where $v_d$ is the volume of a unit ball in $\mathbb{R}^d$ and $\text{vol}_d$ is the volume w.r.t. the uniform measure on $M$.
\end{lemma}
\subsection{$k$-NN Density Estimation}
Here, we establish density estimation rates for the $k$-NN density estimator
in the manifold setting. This builds on work in density estimation on manifolds e.g. \citep{hendriks90,pelletier05,ozakin09,kim13,berry17}; thus,
it may be of independent interest.
The estimator is defined as follows
\begin{definition} [k-NN Density Estimator] \label{kNNdensity}
\begin{align*}
f_k(x) := \frac{k}{n\cdot v_d\cdot r_k(x)^d}.
\end{align*}
\end{definition}
The following extends previous work of \citet{optimalknn} to the manifold case.
The proofs can be found in the appendix.
\begin{lemma}[$f_k$ upper bound]\label{fk_upper_bound} Suppose that Assumptions~\ref{manifold} and~\ref{density} hold.
Define the following which charaterizes how much the density increases locally in $M$:
\begin{align*}
\hat{r}(\epsilon, x) &:=\sup\left\{r : \sup_{x' \in B(x, r) \cap M} f(x') - f(x) \le \epsilon \right\}.
\end{align*}
Fix $\lambda_0 > 0$ and $\delta > 0$ and suppose that $k \ge C_{\delta, n}^2$.
Then there exists constant $C_1 \equiv C_1 (\lambda_0, d, \tau)$ such that if
\begin{align*}
k \le C_1 \cdot C_{\delta, n}^{2d/(2+d)} \cdot n^{2/(2+d)},
\end{align*}
then the following holds with probability at least $1 - \delta$
uniformly in $\epsilon > 0$ and $x \in X$ with $f(x) + \epsilon \ge \lambda_0$:
\begin{align*}
f_k(x) < \left(1 + 3 \cdot \frac{C_{\delta, n}}{\sqrt{k}} \right)\cdot (f(x) + \epsilon),
\end{align*}
provided $k$ satisfies $v_d\cdot \hat{r}(\epsilon, x)^d \cdot (f(x) + \epsilon) \ge \frac{k}{n} - C_{\delta, n}\frac{\sqrt{k}}{n}$.
\end{lemma}
\begin{lemma}[$f_k$ lower bound] \label{fk_lower_bound} Suppose that Assumptions~\ref{manifold} and~\ref{density} hold.
Define the following which charaterizes how much the density decreases locally in $M$:
\begin{align*}
\check{r}(\epsilon, x) &:=\sup\left\{r : \sup_{x' \in B(x, r) \cap M} f(x) - f(x') \le \epsilon \right\}.
\end{align*}
Fix $\lambda_0 > 0$ and $0 < \delta < 1$ and suppose $k \ge C_{\delta, n}$.
Then there exists constant $C_2 \equiv C_2(\lambda_0, d, \tau)$ such that if
\begin{align*}
k \le C_2 \cdot C_{\delta, n}^{2d/(4 + d)}\cdot n^{4 / (4 + d)},
\end{align*}
then with probability at least $1-\delta$, the following holds uniformly
for all $\epsilon > 0$ and $x \in X$ with $f(x) - \epsilon \ge \lambda_0$:
\begin{align*}
f_k(x) \ge \left(1 - 3 \cdot \frac{C_{\delta, n}}{\sqrt{k}} \right)\cdot (f(x) - \epsilon),
\end{align*}
provided $k$ satisfies $v_d\cdot \check{r}(\epsilon, x)^d \cdot (f(x) - \epsilon) \ge \frac{4}{3}\left(\frac{k}{n} + C_{\delta, n}\frac{\sqrt{k}}{n} \right)$.
\end{lemma}
\begin{remark}
We will often bound the density of points with low density. In low-density regions,
there is less data and thus we require more points to get a tight bound. However, in many cases
a tight bound is not necessary; thus the purposes of $\epsilon$ is to allow some slack.
The higher the $\epsilon$, the easier it is for the lemma conditions to be satisified.
In particular, if $f$ is $\alpha$-H\"older continuous (i.e. $|f(x) - f(x')| \le C_\alpha |x - x'|^\alpha$), we have $\hat{r}(\epsilon, x), \check{r}(\epsilon, x) \ge (\epsilon / C_{\alpha})^{1/\alpha}$.
\end{remark}
\subsection{Determining $d$}
Knowing the manifold dimension $d$ is necessary to tune the parameters
as described in Section~\ref{parameters}. There has been much work done on estimating the
intrinsic dimension as many learning procedures (including this one) require $d$ as an input.
Such work in intrinsic dimension estimation include
\citep{intrinsicKegl,intrinsicBickel,intrinsicHein}.
\citet{intrinsicKnnOld} and more recently \citet{intrinsicKnnNew} take a $k$-nearest neighbor approach.
We work with the estimate of a dimension at a point proposed in the latter work:
\begin{align*}
\hat{d}(x) = \frac{\log 2}{\log (r_{2k}(x)/ r_k (x))}.
\end{align*}
The main result of \citet{intrinsicKnnNew} gives a high-probability bound for a single sample $X_1 \in X$.
Here we give a high-probability bound under
more mild smoothness assumptions which hold uniformly for all samples above some density-level given our new knowledge of $k$-NN density estimation rates. This may
be of independent interest.
\begin{theorem} \label{estimating_d}
Suppose that $f$ is $\alpha$-H\"older continuous for some $0 < \alpha \le 1$.
Choose $\bar{\lambda}_0 > 0$ and $\delta > 0$. Then there exists constants $C_1, C_2$ depending on $\delta, C_\alpha, \alpha, \tau, d, \bar{\lambda}_0$ such that if $k$ satisfies
\begin{align*}
C_1 \cdot (\log n)^2 \le k \le C_2 \cdot n^{2\alpha / (2\alpha + d)},
\end{align*}
then with probability at least $1 - \delta$,
\begin{align*}
|\hat{d}(x) - d| \le 20d\cdot||f||_\infty\cdot \frac{C_{\delta, n}}{\sqrt{k}},
\end{align*}
uniformly for all $x \in X$ with $f_k(x) \ge \bar{\lambda}_0$.
\end{theorem}
\vspace{-0.3cm}
\begin{proof}
We have for $x \in X$ such that if $f_k(x) \ge \bar{\lambda}_0$, then $f(x) \ge \lambda_0 := \bar{\lambda}_0 / 2$ by Lemma~\ref{fk_upper_bound} for $C_1$ chosen appropriately large and $C_2$ chosen appropriately small.
\begin{align*}
\hat{d}(x) = \frac{\log 2}{\log (r_{2k}(x)/ r_k (x))} = \frac{d\log 2}{\log 2 + \log (f_{k}(x)/ f_{2k} (x))}.
\end{align*}
We now try to get a handle on $f_k(x) / f_{2k}(x)$ and show it is sufficiently close to $1$.
Applying Lemma~\ref{fk_upper_bound} and~\ref{fk_lower_bound} with $\epsilon = \frac{C_{\delta, n}}{\sqrt{k}} f(x)$ and $C_1$, $C_2$ appropriately chosen so that the conditions for the two Lemmas hold (remember that here we have $\hat{r}(\epsilon, x), \check{r}(\epsilon, x) \ge (\epsilon / C_{\alpha})^{1/\alpha}$), we obtain
\begin{align*}
\frac{f_k(x)}{f_{2k}(x)} &\ge \frac{(1 - 3C_{\delta, n}/\sqrt{k}) (1 - C_{\delta, n}/\sqrt{k}) \cdot f(x)}{(1 + 3C_{\delta, n}/\sqrt{k}) (1 + C_{\delta, n}/\sqrt{k}) \cdot f(x)}\\
&\ge 1 - 9 \cdot \frac{C_{\delta, n}}{\sqrt{k}},
\end{align*}
where the last inequality holds when $C_1$ is chosen sufficiently large so that $C_{\delta, n}/\sqrt{k}$ is sufficiently small. On the other hand, we similarly obtain (for $C_1$ and $C_2$ appropriately chosen):
\begin{align*}
\frac{f_k(x)}{f_{2k}(x)} &\le \frac{(1 + 3C_{\delta, n}/\sqrt{k}) (1 + C_{\delta, n}/\sqrt{k}) \cdot f(x)}{(1 - 3C_{\delta, n}/\sqrt{k})(1 - C_{\delta, n}/\sqrt{k})\cdot f(x)}\\
&\le 1 + 9 \cdot \frac{C_{\delta, n}}{\sqrt{k}}.
\end{align*}
It is now clear that by the expansion $\log(1 - r) = - r - r^2/2 - r^3 /3 - \cdots$, and for $K_l$ chosen sufficently large so that $C_{\delta, n}/\sqrt{k}$ is sufficiently small, we have
\begin{align*}
\left| \log \left(\frac{f_{k}(x)}{ f_{2k} (x)} \right) \right| \le 10 \cdot \frac{C_{\delta, n}}{\sqrt{k}}.
\end{align*}
The result now follows by combining this with the earlier established expression for $\hat{d}(x)$, as desired.
\end{proof}
\begin{remark}
In \citet{intrinsicKnnNew}, it is the case that $\alpha = 1$; under this setting, we match their bound with an error rate of $n^{1/(2+d)}$ with $k \approx n^{2/(2+d)}$ being the
optimal choice for $k$ (ignoring log factors).
\end{remark}
\subsection{Determining $k$}
After determining $d$, the next parameter we look at is $k$. In particular, to obtain the optimal rate, we
must choose $k\approx n^{2\beta'/(2\beta' + d)}$ without knowledge of $\beta$. We present a consistent
estimator for $\beta$.
We need the following definition. The first characterizes how much $f$ varies in balls of a certain radius along
the boundaries of the $\lambda$-level set (where $\partial \mathcal{C}_\lambda$ denotes the boundary of $\mathcal{C}_\lambda$). The second is meant to be an estimate of the first, which can be computed from the data alone.
The final is our estimate of $\beta$.
\begin{align*}
D_{r} &= \inf_{x_0 \in \partial \mathcal{C}_\lambda} \sup_{x \in B(x_0, r)} |\lambda - f(x)|\\
\hat{D}_{r, k} &= \min_{\substack{x_0 \in X \\ B(x_0, r) \cap X \neq \emptyset}} \max_{x \in B(x_0, r)\cap X} |\lambda - f_k(x)| \\
\hat{\beta} &= \log_r ( \hat{D}_{r, k})
\end{align*}
The next is a result of how $\hat{D}_{r, k}$ estimates $D_r$.
\begin{lemma} \label{alpha_bound} Suppose that $f$ is $\alpha$-H\"older continuous for some $0 < \alpha \le 1$.
Let $k = \lfloor (\log n)^5 \rfloor$ and $r = 1 / \sqrt{\log n}$. Then there exists positive constants $\tilde{C}$ and $N$ depending on $d, \tau, \alpha, C_\alpha, \lambda_0, ||f||_\infty, r_c$
such that when $n \ge N$, then the following holds with probability at least $1 - 1/n$.
\vspace{-0.1cm}
\begin{align*}
|D_r - \hat{D}_{r, k}| \le \tilde{C} / (\log n)^2.
\end{align*}
\end{lemma}
\vspace{-0.5cm}
\begin{proof}[Proof sketch]
Suppose that the value of $D_r$ is attained at $x_0 = p$ and the value of $\hat{D}_{r, k}$ is attained at $x_0 = q$.
Let $y, z$ be the points that maximize $|\lambda - f(x)|$ on $B(p, r)$ and $B(q, r)$, respectively. Let $\hat{y}, \hat{z}$ be the sample points that maximize $|\lambda - f_k(x)|$ on
$B(p, r)$ and $B(q, r)$, respectively. Now, we have
\vspace{-0.2cm}
\begin{align*}
&D_r - \hat{D}_{r, k} = |\lambda - f(y)| - |\lambda - f_k(\hat{z})|\\
&\le |\lambda - f(z)| - |\lambda - f_k(\hat{z})| \le |f(z) - f_k(\hat{z})| \\
&\le \max \{ f(z) - f_k(z), f_k(\hat{z}) - f(\hat{z}) \}.
\end{align*}
Now let $z'$ be the closest sample point to $z$ in $B(q, r)$. Then,
\begin{align*}
&\le \max \{ f(z') - f_k(z'), f_k(\hat{z}) - f(\hat{z}) \} + |f(z) - f(z')| \\
&+ |f_k(z) - f_k(z')| \le \max_{x \in X, f(x) \ge \lambda_0} |f(x) - f_k(x)| \\
&+ C_\alpha|z - z'|^\alpha+ |f_k(z) - f_k(z')|.
\end{align*}
\vspace{-0.2cm}
On the other hand, we have
\begin{align*}
&\hat{D}_{r, k} - D_r = |\lambda - f_k(\hat{z})| - |\lambda - f(y)| \\
&\le |\lambda - f_k(\hat{y})| - |\lambda - f(y)| \le |f(y) - f_k(\hat{y})| \\
&\le \max \{ f(y) - f_k(y), f_k(\hat{y}) - f(\hat{y}) \}.
\end{align*}
\vspace{-0.2cm}
Let $y'$ be the closest sample point to $y$ in $B(p, r)$. Then,
\begin{align*}
&\le \max \{ f(y') - f_k(y'), f_k(\hat{y}) - f(\hat{y}) \} + |f(y) - f(y')|\\
& + |f_k(y) - f_k(y')| \le \max_{x \in X, f(x) \ge \lambda_0} |f(x) - f_k(x)| \\
&+ C_\alpha |y-y'|^\alpha + |f_k(y) - f_k(y')|.
\end{align*}
Thus it suffices to bound $\max_{x \in X, f(x) \ge \lambda_0} |f(x) - f_k(x)|, |y-y'|, |z-z'|,
|f_k(y) - f_k(y')|, |f_k(z) - f_k(z')|$. First take $\delta = 1/n$ and use
Lemma~\ref{fk_upper_bound} and~\ref{fk_lower_bound} for $\max_{x \in X, f(x) \ge \lambda_0} |f(x) - f_k(x)|$.
Using Lemma~\ref{ballbounds}, we can show that $r_n := |y - y'| \lesssim (\log n/n)^{1/d}$.
Next we bound $|f_k(y) - f_k(y')|$. $y' \in X$ so we have guarantees on its $f_k$ value.
Note that $r_k(y') - r_n \le r_k(y) \le r_k(y')+ r_n$. Let $r_k = r_k(y')$. This implies that
$ f_k(y') (r_k / (r_k + r_n))^d \le f_k(y) \le f_k(y') (r_k / (r_k - r_n))^d$.
Now since $r_k \approx (k / n)^{1/d}$, we have $|f_k(y) - f_k(y')| \lesssim \log n / k$.
The same holds for the bounds related to $z, z'$.
\end{proof}
\begin{theorem} [$\hat{\beta} \rightarrow \beta$ in probability] \label{estimating_alpha} Suppose
$f$ is $\alpha$-H\"older continuous for some $\alpha$ with $0 < \alpha \le \beta'$.
Let $k = \lfloor (\log n)^5 \rfloor$ and $r = 1 / \sqrt{\log n}$. Then for all $\epsilon > 0$,
\vspace{-0.3cm}
\begin{align*}
\lim_{n \rightarrow \infty} \mathbb{P}\left(|\hat{\beta} - \beta| \ge \epsilon\right) = 0.
\end{align*}
\end{theorem}
\vspace{-0.5cm}
\begin{proof}
Based on the $\beta$-regularity assumption, we have for $r < r_c$:
\vspace{-0.5cm}
\begin{align*}
\check{C}_\beta r^\beta \le D_r \le \hat{C}_\beta r^\beta.
\end{align*}
Combining this with Lemma~\ref{alpha_bound}, we have with probability at least $1 - 1/\sqrt{n}$ that
\begin{align*}
\check{C}_\beta r^\beta - \tilde{C} / (\log n)^2 \le \hat{D}_{r, k} \le \hat{C}_\beta r^\beta + \tilde{C} / (\log n)^2.
\end{align*}
Thus with probability at least $1 - 1/n$,
\begin{align*}
\beta - \hat{\beta} &\ge \frac{\log (1 - \tilde{C}/(\hat{D}_{r, k} \cdot (\log n^2))) }{\log r} - \frac{\log \hat{C}_\beta}{\log r} \\
\beta - \hat{\beta} &\le \frac{\log (1 + \tilde{C}/(\hat{D}_{r, k} \cdot (\log n^2))) }{\log r} + \frac{\log \check{C}_\beta}{\log r}.
\end{align*}
It is clear that these expressions go to $0$ as $n\rightarrow \infty$ and the result follows.
\end{proof}
\begin{remark}\label{adaptive_parameter}
We can then take $k = n^{\hat{\beta'}/(2\hat{\beta'} + d)}$ with $\hat{\beta'} = \min\{1, \hat{\beta} - \epsilon_0\}$ for some $\epsilon_0 > 0$ so that $\hat{\beta'} < \beta'$ for $n$ sufficiently large and thus $k$ lies in the allowed ranges described in Section~\ref{parameters} asymptotically.
The settings of $\varepsilon$ and $\text{MinPts}$ are implied by this choice of $k$ and our estimate of $d$.
\end{remark}
\subsection{Rates with Data-driven Tuning}
Putting this all together, along with Theorems~\ref{hausdorfferror} and~\ref{pruning}, gives us the following consequence about level set recovery with adaptive tuning. It shows that we can obtain rates arbitrarily close to those obtained as if the smoothness parameter $\beta$ and intrinsic dimension were known.
\begin{corollary} \label{nearoptimal}
Suppose that $0 < \delta < 1$ and $f$ is $\alpha$-H\"older continuous for some $0 < \alpha \le 1$ and suppose
the data-driven choices of parameters described in Remark~\ref{adaptive_parameter} are used for DBSCAN.
For any $\epsilon > 0$, there exists $N_{\epsilon, \delta, f} \equiv N(\epsilon, \delta, f)$ and $C_\delta \equiv C_\delta(\delta, f)$ such that the following holds. If $n \ge N_{\epsilon, \delta, f}$, then with probability at least $1 - \delta$ simulatenously for each $C \in \mathcal{C}_\lambda$, there exists $\widehat{C} \in \widehat{\mathcal{C}_\lambda}$ such that
\begin{align*}
d_{\text{Haus}}(C, \widehat{C}) \le C_{\delta} \cdot n^{-\frac{1}{2\beta + d\max \{1, \beta\}}+ \epsilon}.
\end{align*}
Moreover, using Algorithm~\ref{pruning}, there is a one-to-one correspondence between $\mathcal{C}_\lambda$ and $\widehat{\mathcal{C}_\lambda}$.
\end{corollary}
|
\section{\uppercase{Approach}}
\label{sec:approach}
An overview of the visual-interactive system is shown in Figure \ref{fig:titleImage}.
Figure \ref{fig:workflow} illustrates the interplay of the technical components assembled to a workflow.
In Sections \ref{approach:feedback}, \ref{approach:model}, and \ref{approach:result}, we describe the three core components in detail, after we discuss data characteristics and abstractions in Section \ref{sec:approach_data}.
\input{secondaryDataTable.tex}
\subsection{Data Characterization}
\label{sec:approach_data}
\subsubsection{Data Source}
Various web references provide data about soccer players with information differing in its scope and depth.
For example some websites offer information about market price values or sports betting statistics, while other sources provide statistics about pass accuracy in every detail.
Our prior requirement to the data is its public availability to guarantee the reproducibility of our experiments.
In addition, the information about players should be comprehensible for broad audiences and demonstrate the applicability.
Finally, the attributes should be of mixed types (numerical, categorical, boolean).
This is why ~Wikipedia\footnote{Wikipedia, https://en.wikipedia.org/wiki/Main\_Page, last accessed on September 22th, 2016} the free encyclopedia serves as our primary data source.
The structured information about players presented at Wikipedia is retrieved from ~DBpedia\footnote{DBpedia, http://wiki.dbpedia.org/about, last accessed on September 22th, 2016}.
We access ~DBpedia with ~SPARQL \cite{sparql2008} the query language for RDF\footnote{RDF, https://www.w3.org/RDF, last accessed on September 22th, 2016} recommended by W3C\footnote{RDF, https://www.w3.org, last accessed on September 22th, 2016}.
We focused on the Europe's five top leagues (Premier League in England, Seria A in Italy, Ligue 1 in France, Bundesliga in Germany, and LaLiga in Spain).
Overall, we gathered 2,613 players engaged by the teams of respective leagues.
\subsubsection{Data Abstraction}
Table \ref{tab:primary} provides an overview of the available information about soccer players.
Important attributes for the player (re-)identification are the unique name in combination with the nationality and the current team.
Moreover, a various numerical and categorical information is provided for similarity modeling.
Table \ref{tab:secondary} depicts the secondary data (i.e., attributes deduced from primary data).
Our strategy for the extraction of additional information is to obtain as much meaningful attributes as possible.
One benefit of our approach will be a weighing of all involved attributes, making the selection of relevant features for downstream analyses an easy task.
This strategy is inspired by user-centered design approaches in different application domains where we asked domain experts about the importance of attributes (features) for the similarity definition process \cite{BernardWKMSK13,bernardIJoDL2015}.
One of the common responses was \emph{``everything can be important!''}
In the usage scenarios, we demonstrate how the similarity model will weight the importance of primary and secondary attributes with respect to the learned pairs of labeled players.
\begin{figure*}[t]
\vspace{-2mm}
\centering
\includegraphics[width=1.0\textwidth]{workflow.pdf}
\caption{Workflow of the approach. Users assign similarity scores for pairs of players in the feedback interface. In the backend, the feedback is interpreted (blue) and delegated to the similarity model. Active learning support suggests players to improve the model (orange). A kNN-search supports the use case of the workflow shown in the result visualization (purple).}
\label{fig:workflow}
\end{figure*}
\subsubsection{Preprocessing}
One of the data-centered challenges was the sparsity of some data attributes.
This phenomenon can often be observed when querying less popular instances of concepts from DBpedia.
To tackle this challenge we removed attributes and instances from the data set containing only little information.
Remaining missing values were marked with missing value indicators, with respect to the type of attribute.
For illustration purposes, we also removed players without an image in Wikipedia.
The final data set consists of 1,172 players.
An important step in the preprocessing pipeline is normalization.
By default, we rescaled every numerical attribute into the relative value domain to foster metrical comparability.
\begin{figure*}[t]
\vspace{-2mm}
\centering
\includegraphics[width=1.0\linewidth]{useCaseInternational3.png}
\caption{Similarity model learned with stars in the European soccer scene. The history provides an overview of ten labeled pairs of players. The ranking of weighted attributes assigns high correlations to the vertical position, player size, and national games. Karim Benzema served as the query player: all retrieved players share quite similar attributes with Benzema.}
\label{fig:useCaseInternational}
\end{figure*}
\subsection{Visual-Interactive Learning Interface}
\label{approach:feedback}
One of the primary views of the approach enables users to give feedback about individual instances.
Examples of the feedback interface can be seen in the Figures \ref{fig:titleImage} and \ref{fig:useCaseInternational}.
The interface for the definition of similarity between soccer players shows two players in combination with a slider control in between.
The slider allows the communication of similarity scores between the two players.
We decided for a quasi-continuous slider, in accordance to the continuous numerical function to be learned.
However, one of possible design alternative would propose a feedback control with discrete levels of similarity.
Every player is represented with an image (when available and permitted), a flag icon showing the player's nationality, as well as textual labels for the player name and the current team.
These four attributes are also used for compact representations of players in other views of the tool.
The visual metaphor of a soccer field represents the players' main positions.
In addition, a list-based view provides the details about the players' attribute values.
The feedback interface combines three additional functionalities most relevant for the visual-interactive learning approach.\\
First, users need to be able to define and select players of interest.
This supports the idea to grasp detailed feedback about instances matching the users' expert knowledge \cite{SG10,HNHWH12}.
For this purpose a textual query interface is provided in combination with a combobox showing players matching a user-defined query.
In this way, we combine query-by-sketch and the query-by-example paradigm for the straightforward lookup of known players.\\
The second ingredient for an effective active learning approach is the propagation of instances to the user reducing the remaining model uncertainty.
One crucial design decision determined that \emph{users should always be able to label players they actually know}.
Thus, we created a solution for the candidate selection combining automated suggestions by the model with the preference of users.
The feedback interface provides two sets of candidate players, one set is located at the left and the other one right of the interface.
Replacing the left feedback instance with one of the suggested players at the left will reduce the remaining entropy regarding the current instance at the right, and vice-versa.
However, we are aware that other strategies for proposing unlabeled instances exist.
Two of the obvious alternative strategies for labeling players would be a) providing a global pool of unlabeled players (e.g., in combination with drag-and-drop) or b) offering pairs of instances with low confidence.
While these two strategies may be implemented in alternative designs, we recall the design decision that users need to know the instances to be labeled.
In this way, we combine a classical active learning paradigm with the user-defined selection of players matching their expert knowledge.\\
Finally, the interface provides a history functionality for labeled pairs of players at the bottom of the feedback view (cf. Figure \ref{fig:titleImage}).
For every pair of players images are shown and the assigned similarity score is depicted in the center.
\subsection{Similarity Modeling}
\label{approach:model}
\subsubsection{Similarity Learning}
The visual-interactive learning interface provides feedback about the similarity of pairs of instances.
Thus, the feedback propagated to the system is according to the learning function $f(i_1, i_2) = y$ whereas $y$ is a numerical value between 0 (unsimilar) and 1 (very similar).
Similarity learning is designed as a two-step approach. First, every attribute (feature) of the data set is correlated with the user feedback. Second, pairwise distances are calculated for any given instance of the data set.
The correlation of attributes is estimated with Pearson's correlation coefficient.
Pairwise distances between categorical attributes are transformed into the numerical space with the Kronecker delta function.
The correlation for a given attribute is then estimated between the labeled pairs of instances provided by the user and the distance in the value domain obtained by that attribute.
In the current state of the approach every attribute is correlated independently to reduce computation time and to maximize interpretability of the resulting weights.
The result of this first step of the learning model is a weighting of the attributes that is proportional to the correlation.
In a second step, the learning model calculates distances between any pair of instances.
As the underlying data may consist of mixed attribute types, different distance measures are used for different types of attributes.
For numerical data we employ the (weighted) Euclidean distance.
For categorical attributes we choose the Goodall distance \cite{boriah2008similarity} since it is sensitive to the probability of attribute values and less frequent observations are assigned higher scores.
The weighted Jaccard distance \cite{cha2005enhancing} is used for binary attributes.
The Jaccard distance neglects negative matches (both inputs false), which might be advantageous for many similarity concepts, i.e. the absence of an attribute in two items does not add to their similarity \cite{sneath1973numerical}.
After all distance measures have been computed in separate distance matrices all matrices are condensed into a single distance matrix by a weighted sum, where the weights represent the fraction of the sum of weights for each attribute type.
\subsubsection{Active Learning Strategy}
We follow an interactive learning strategy that allows for keeping the user in the loop.
To support the iterative nature, we designed an active learning strategy that fosters user input for instances for which no or little information is available yet.
As a starting point for active learning the user selects a known instance from the database.
Note that this is important as the user needs a certain amount of knowledge about the instance to make similarity assessments in the following (see Section \ref{approach:feedback}).
After an instance has been selected, we identify the attribute with the highest weight.
Next, we estimate the \emph{farthest} neighbors to the selected item under the given attribute for which no similarity assessments exist so far.
A set of respective candidates is then presented to the user.
This strategy is useful as it identifies pairs of items for which the system cannot make assumptions so far.
The user can now select one or more proposed items and add similarity assessments.
By adding assessments the coverage of the attribute space is successfully improved especially in sparse areas where little information was available so far.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\columnwidth]{result5.png}
\caption{Nearest neighbor search for Lionel Messi. Even if superstars are difficult to replace, the set of provided nearest neighbors is quite reasonable. Keisuke Honda may be surprising, nevertheless Honda has similar performance values in the national team of Japan.}
\label{fig:result}
\end{figure}
\subsubsection{Model Visualization}
Visualizing the \emph{output} of algorithmic models is crucial, e.g., to execute downstream analysis tasks (see Section \ref{approach:result}).
In addition, we visualize the \emph{current state of the model} itself.
In this way, designers and experienced users can keep track of the model improvement, its quality improvement, and its determinism.
The core black-box information of this two-step learning approach is the set of attribute weights representing the correlation between attributes and labeled pairs of instances.
Halfway right in the tool, we make the attribute weights explicit (between the feedback interface and the model result visualization), as it can be seen in the title figure.
An enlarged version of the model visualization is shown at the left of Figure \ref{fig:result} where the model is used to execute a kNN search for Lionel Messi.
From top to bottom the list of attributes is ranked in the order of their weights.
It is a reasonable point of discussion whether the attribute weights should be visualized to the final group of users of such a system.
A positive argument (especially in this scientific context) is the transparency of the system which raises trust and allows the visual validation.
However, a counter argument is biasing users with information about the attribute/feature space.
Recalling that especially in complex feature spaces users do not necessarily know any attribute in detail, it may be a valid design decision to exclude the model visualization from the visual-interactive system.
\subsection{Result Visualization -- Visual-Interactive NN Search}
\label{approach:result}
We provide a visual-interactive interface for the visualization of the model output (see Figure \ref{fig:result}).
A popular use case regarding soccer players is the identification of similar players for a reference player, e.g., when a player is replaced in a team due to an upcoming transfer event.
Thus, the interface of the result visualization will provide a means to query for similar soccer players.
We combine a query interface (query-by-sketch, query-by-example) with a list-based visualization of retrieved players.
The retrieval algorithm is based on a standard k-NN search (k nearest neighbors) using the model output.
For every list element of the result set a reduced visual representation of a soccer player is depicted, including the player's image, name, nationality, position, and team.
Moreover, we show information about three attributes contributing to the current similarity model significantly.
Finally, we depict rank information as well as the distance to the query for ever element.
The result visualization rounds up the functionality.
Users can train individual similarity models of soccer player similarity and subsequently perform retrieval tasks.
From a more technical perspective, the result visualization closes the feedback loop of the visual-interactive learning approach.
In this connection, users can analyze retrieved sets of players and give additional feedback for weakly learned instances.
An example can be seen in Figure \ref{fig:result} showing a retrieved result set for Lionel Messi used as an example query.
\section{\uppercase{Conclusion}}
\label{sec:conclusion}
We presented a tool for the visual-interactive similarity search for complex data objects by example of soccer players.
The approach combines principles from active learning, information visualization, visual analytics, and interactive information retrieval.
An algorithmic workflow accepts labels for instances and creates a model reflecting the similarity expressed by the user.
Complex objects including numerical, categorical, and boolean attribute types can be included in the algorithmic workflow.
Visual-interactive interfaces ease the labeling process for users, depict the model state, and represent output of the similarity model.
The latter is implemented by means of an interactive information retrieval technique.
While the strategy to combine active learning with visual-interactive interfaces enabling users to label instances of interest is special, the application by example of soccer players is, to the best of our knowledge, unique.
Domain experts are enabled to express expert knowledge about similar players, and utilize learned models to retrieve similar soccer players.
We demonstrated that only very few labels are needed to train meaningful and robust similarity models, even if the data set was unlabeled at start.
Future work will include additional attributes about soccer players, e.g., market values or variables assessing the individual player performance.
In addition, it would be interesting to widen the scope and the strategy to other domains, e.g., in design study approaches.
Finally, the performance of individual parts of the algorithmic workflow may be tested against design alternatives in future experiments.
\section{\uppercase{Discussion}}
\label{sec:discussion}
In the evaluation section, we demonstrated the applicability of the approach from different perspectives.
However, we want to shed light on aspects that allow alternative design decisions, or may be beneficial subjects to further investigation.
\paragraph{Similarity vs. Distance}
Distance measures are usually applied to approximate similarity relationships.
This is also the case in our work.
We are, however, aware that metric distances can in general not be mapped directly to similarities, especially when the dimension of the data becomes high and the points in the feature space move far away from each other.
Finding suitable mappings between distance and similarity is a challenging topic that we will focus on in future research.
\paragraph{Active Learning Strategy}
The active learning support of this approach builds on the importance (weights) of attributes to suggest new learning instances to be queried.
Thus, we focus on a scalable solution that takes the current state of the model into account and binds suggestions to previously-labeled instances.
Alternative strategies may involve other intrinsic aspects of the data (attributes or instances) or the model itself.
For example statistical data analysis, distributions of value domains, or correlation tests could be considered.
Other active learning strategies may be inspired by classification approaches, i.e., models learning categorical label information.
Concrete classes of strategies involve uncertainty sampling or query by committee \cite{settles2009}.
\paragraph{Numerical vs. Categorical}
This research effort explicitly addressed a complex data object with mixed data, i.e., objects characterized by numerical, categorical, and boolean attributes.
This class of objects is widespread in the real-world, and we argue that it is worth to address this additional analytical complexity.
However, coping with mixed data can benefit from a more in-depth investigation at different steps of the algorithmic pipeline.
\paragraph{Usability}
We presented a technique that actually works but has not been throughoutly evaluated with users.
Will users be able to interact with the system?
We did cognitive walkthroughs and created the designs in a highly interactive manner.
Still, the question arises whether domain experts will appreciate the tool and be able to work with it in an intuitve way.
\section{\uppercase{Evaluation}}
\label{sec:usageScenario}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\columnwidth]{useCaseArtificialWeight3.png}
\caption{Experiment with a mental model based on teams and player age. It can be seen that for ten labeled pairs of players the system is able to grasp this mental model.}
\label{fig:useCaseArtificialWeight}
\end{figure}
Providing scientific proof for this research effort is non-trivial since the number of comparable approaches is scarce.
Moreover, we address the challenge of dealing with data sets which are completely unlabeled at start, making classical quantitative evaluations with ground truth test data impossible.\\
In the following, we demonstrate and validate the applicability of the approach with different strategies.
In a first proof of concept scenario a similarity model is trained for an explicitly known mental model, answering the question whether the similarity model will be able to capture a human's notion of similarity.
Second, we assess the effectiveness of the approach in two usage scenarios. We demonstrate how the tool can be used to learn different similarity models, e.g., to replace a player in the team by a set of relevant candidates.
Finally, we report on experiments for the quantification of model efficiency.
\subsection{Proof of Concept - Fixed Mental Model}
\label{eval:1}
The first experiment assesses whether the similarity model of the system is able to grasp the mental similarity model of a user.
As an additional constraint, we limit the number of labeled pairs to ten, representing the requirement of very fast model learning.
As a proof of concept, we predefine a mental model and express it with ten labels.
In particular, we simulate a fictitious user who is only interested in the age of players, as well as their current team.
In other words, a numerical and a categorical attribute defines the mental similarity model of the experiment.
If two players are likely identical with respect to these two attributes (age +-1) the user assigns the similarity score $1.0$.
If only one of the two attributes match, the user feedback is $0.5$ and if both attributes disagree a pair of players is labeled with the similarity score $0.0$.
The ten pairs of players used for the experiment are shown in Figure \ref{fig:useCaseArtificialWeight}.
In addition to the labeled pairs, the final attribute weights calculated by the system are depicted.
Three insights can be identified.
First, it becomes apparent that the two attributes with the highest weights exactly match the pre-defined mental model.
Second, the number of national games, the number of national games per year, and the number of league games also received weights.
Third, the set of remaining attributes received zero weights.
While the first insight validates the experiment, the second insight sheds light on attributes correlated with the mental model.
As an example, we hypothesize that the age of players is correlated with the number of games.
This is a beneficial starting point for downstream feature selection tasks, e.g., when the model is to be implemented as a static similarity function.
Finally, the absence of weights for most other attributes demonstrates that only few labels are needed to obtain a precise focus on relevant attributes.
\subsection{Usage Scenario 1 - Top Leagues}
\label{eval:2}
The following usage scenario demonstrates the effectiveness of the approach.
A user with much experience in Europe's top leagues (Premier League, Seria A, Ligue 1, Bundesliga, LaLiga) rates ten pairs of prominent players with similarity values from very high to very low.
The state of the system after ten labeling iterations can be seen in Figure \ref{fig:useCaseInternational}.
The history view shows that high similarity scores are assigned to the pairs:
Mats Hummels vs. David Luiz,
Luca Toni vs. Claudio Pizarro,
Dani Alves vs. Juan Bernat,
Toni Kroos vs. Xabi Alonso,
Eden Hazard vs. David Silva,
and Jamie Vardi vs. Marco Reus.
Comparatively low similarity values are assigned to the pairs
Philipp Lahm vs. Ron-Robert Zieler,
Sandro Wagner cs. Bastian Oczipka,
Manuel Neuer vs. Marcel Heller, and
Robert Huth vs. Robert Lewandowski.
The set of player instances resembles a vast spectrum of nationalities, ages, positions, as well as numbers of games and goals.
The resulting leaning model depicts high weights to the vertical position on the field, the player size, the national games per year.
Further, being a national player and the goals for the respective national teams contribute to the global notion of player similarity.
In the result visualization, the user chose Karim Benzema for the nearest neighbor search.
The result set (Edison Cavani, Robert Lewandowski, Salomon Kalou, Claudio Pizarro, Klass-Jan Huntelaar, Pierre-Emerick Aubameyang) represents the learned model quite well.
All these players are strikers, in a similar age, and very successful in their national teams.
In contrast, there is not a single player listed in the result who does not stick to the described notion of similarity.
In summary, this usage scenario demonstrates that the tool was able to reflect the notion of similarity of the user with a very low number of training instances.
\subsection{Usage Scenario 2 - National Trainer}
\label{eval:2b}
In this usage scenario, we envision to be a national trainer.
Our goal is to engage a similarity model which especially resembles the quality of soccer players, but additionally takes the nationality of players into account.
As a result, similar players coming from the same country are classified similar.
The reason is simple; players from foreign countries cannot be positioned in a national team.
Figure \ref{fig:titleImage} shows the history of the ten labeled pairs of players.
We assigned similar scores to two midfielders from Belgium, two strikers from France, two defenders from Germany, two goalkeepers from Germany, and two strikers from Belgium.
In addition, four pairs of players with different nationalities were assigned with considerably lower similarity scores, even if the players play at very similar positions.
The result visualization shows the search result for Dimitri Payet who was used as a query player.
In this usage scenario, the result can be used to investigate alternative players for the French national team.
With only ten labels, the algorithm retrieves exclusively players from France, all having expertise in the national team, and all sharing Dimitri Payets position (offensive midfielder).
\subsection{Quantification of Efficiency}
\label{eval:3}
In the final evaluation strategy, we conduct an experiment to yield quantitative results for the efficiency.
We assess the `speed' of the convergence of the attribute weighting for a given mental model, i.e. how many learning iterations the model needs to achieve stable attribute weights.
The independent variable of the experiment is the number of learned iterations, i.e., the number of instances already learned by the similarity model.
The dependent variable is the change of the attribute weights of the similarity model between two learning iterations, assessed by the quantitative variable $\Delta w$.
To avoid other degrees of freedom, we fix the mental model used in the experiment.
For this purpose, a small group of colleagues all having an interest in soccer defined labels of similarity for 50 pairs of players.\\
To guarantee robustness and generalizability, we run the experiment 100 times.
Inspired by cross-validation, the set of training instances is permuted in every run.
The result is depicted in Figure \ref{fig:weight_deltas_50}.
Obviously, the most substantial difference of attribute weights is in the beginning of the learning process between the 1st and the 2nd learning iteration ($\Delta w = 0.36$).
In the following, the differences significantly decrease before reaching a saturation point approximately after the 5th iteration.
For the 6th and later learning iterations $\Delta w$ is already below $0.1$ and $0.03$ after the 30th iteration.
To summarize, the approach only requires very few labeled instances to produce a robust learning model.
This is particularly beneficial when users have very limited time, e.g., important experts in the respective application field.
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\textwidth]{weight_deltas_50.png}
\caption{Quantification of efficiency. The experiment shows differences in the attribute weighting between consecutive learning iterations. A saturation point can be identified, approximately after the 5th labeled pair of instances.}
\label{fig:weight_deltas_50}
\end{figure*}
\section{\uppercase{Introduction}}
\label{sec:introduction}
The way how similarity of data objects is defined and represented in an analytical system has a decisive influence on the results of the algorithmic workflow for downstream data analysis.
From an algorithmic perspective the notion of object similarity is often implemented with distance measures resembling an inverse relation to similarity.
Many data mining approaches necessarily require the definition of distance measures, e.g., for conducting clustering or dimension reduction.
In the same way, most information retrieval algorithms carry out indexing and retrieval tasks based on distance measures.
Finally, the performance of many supervised and unsupervised machine learning methods depends on meaningful definitions of object similarity.
The classical approach for the definition of object similarity includes the identification, extraction, and selection of relevant attributes (features), as well as the definition of a distance measure and optionally a mapping from distance to similarity.
Furthermore, many real-world examples require additional steps in the algorithmic pipeline such as data cleansing or normalization.
In practice, quality measures such as precision and recall are used to assess the quality of the similarity models and the classifiers built upon them.
In this work, we strengthen the connection between the \emph{notion of similarity of individual users} and its adoption to the algorithmic definition of object similarity.
Taking the example of soccer players from European soccer leagues, a manager may want to identify previously unknown soccer players matching a reference player, e.g., to occupy an important position in the team lineup.
This is contrasted by a national coach who is also interested in selecting a good team.
However, the national coach is independent from transfer fees and salaries while his choice is limited to players of the respective nationality.
The example sheds light on various remaining problems that many classical approaches are confronted with.
First, in many approaches designers do not know beforehand which definition of object similarity is most meaningful.
Second, many real-world approaches require multiple definitions of similarity for being usable for different users or user groups.
Moreover, it is not even determined that the notion of similarity of single users remains constant.
Third, the example of soccer players implicitly indicates that definition of similarity becomes considerably more difficult for high-dimensional data.
Finally, many real-word objects consist of mixed data, i.e. attributes of numerical, categorical, and binary type. However, most current approaches for similarity measurement are limited to numerical attributes.
We hypothesize that it is desirable to shift the definition of object similarity from an offline preprocessing routine to an integral part of future analysis systems.
In this way the individual notions of similarity of different users will be reflected more comprehensively.
The precondition for the effectiveness of such an approach is a means that enables users to communicate their notion of similarity to the system.
Logically, such a system requires the functionality to grasp and adopt the notion of similarity communicated by the user.
Provided that users are able to conduct various data analysis tasks relying on object similarity in a more dynamic and individual manner.
This requirement shifts the definition of similarity towards active learning approaches.
Active learning is a research field in the area of semi-supervised learning where machine learning models are trained with as few user feedback as possible, learning models that are as generalizable as possible.
Beyond classical active learning, the research direction of this approach is towards visual-interactive learning allowing users to give feedback for those objects they have precise knowledge about.
We present a visual-interactive learning system that learns the similarity of complex data objects on the basis of user feedback.
The use case of soccer players will serve as a relevant and intuitive example.
Overall this paper makes three primary contributions.
\emph{First}, we present a visual-interactive interface that enables users to select two soccer players and to submit feedback regarding their subjective similarity.
The set of labeled pairs of players is depicted in a history visualization for lookup and reuse.
\emph{Second}, a machine learning model accepts the pairwise notions of similarity and learns a similarity model for the entire data set.
An active learning model identifies player objects where user feedback would be most beneficial for the generalization of the learned model, and propagates them to a visual-interactive interface.
\emph{Third}, we present a visual-interactive retrieval interface enabling users to directly submit example soccer players to query for nearest neighbors.
The interface combines both validation support as well as a downstream application of model results.
The results of different types of evaluation techniques particularly assess the efficiency of the approach.
In many cases it takes only five labeled pairs of players to learn a robust and meaningful model.\\
The remaining paper is organized as follows.
Section \ref{sec:rw} shows related work.
We present our approach in Section \ref{sec:approach}.
The evaluation results are described in Section \ref{sec:usageScenario}, followed by a discussion in Section \ref{sec:discussion} and the conclusion in Section \ref{sec:conclusion}.
\section{\uppercase{Related Work}}
\label{sec:rw}
The contributions of this work are based on two core building blocks, i.e., visual-interactive interfaces (information visualization, visual analytics) and algorithmic similarity modeling (metric learning).
We provide a subsection of related work for both fields.
\subsection{Visual-Interactive Instance Labeling}
\label{subsec:instanceLabeling}
We focus on visual-interactive interfaces allowing users to submit feedback about the underlying data collection.
In the terminology of the related work, a data element is often referred to as an \emph{instance}, the feedback for an instance is called a \emph{label}.
Different types of labels can be gathered to create some sort of learning model.
Before we survey existing approaches dealing with similarity in detail, we outline inspiring techniques supporting other types of labels.
Some techniques for learning similarity metrics are based on \emph{rules}.
The approach of \cite{FTKW08} allows users to create rules for ranking images based on their visual characteristics.
The rules are then used to improve a distance metric for image retrieval and categorization.
Another class of interfaces facilitates techniques related to \emph{interestingness} or relevance feedback strategies, e.g., to improve retrieval performance \cite{salton1997improving}.
One popular application field is evaluation, e.g,. to ask users which of a set of image candidates is best, with respect to a pre-defined quality criterion \cite{NW_DPID2016}.
In the visual analytics domain, relevance feedback and interestingness-based labeling has been applied to learn users' notions of interest, e.g., to improve the data analysis process.
Behrisch et al. \cite{BKSS14} present a technique for specifying features of interest in multiple views of multidimensional data.
With the user distinguishing relevant from irrelevant views, the system deduces the preferred views for further exploration.
Seebacher et al. \cite{SSJK16} apply a relevance feedback technique in the domain of patent retrieval, supporting user-based labeling of relevance scores.
Similar to our approach, the authors visualize the weight of different modalities (attributes/features).
The weights are subject to change with respect to the iterative nature of the learning approach.
In the visual-interactive image retrieval domain the Pixolution Web interface\footnote{Pixolution, http://demo.pixolution.de, last accessed on September 22th, 2016} combines tag-based and example-based queries to adopt users' notions of interestingness.
Recently the notion of interestingness was adopted to prostate cancer research.
A visual-interactive user interface enables physicians to give feedback about the well-being status of patients \cite{vahc2015}.
The underlying active-learning approach calculates the numerical learning function by means of a regression tree.
\emph{Classification} tasks require categorical labels for the available instances.
Ware et al. \cite{ware2001interactive} present a visual interface enabling users to build classifiers in a visual-interactive way. The approach works well for few and well-known attributes, but requires labeled data sets for learning classifiers.
Seifert and Granitzer's \cite{SG10} approach outlines user-based selection and labeling of instances as meaningful extension of classical active learning strategies \cite{settles2009}.
The authors point towards the potential of combining active learning strategies with information visualization which we adopt for both the representation of instances and learned model results.
H\"{o}ferlin et al. \cite{HNHWH12} define \emph{interactive} learning as an extension, which includes the direct manipulation of the classifier and the selection of instances. The application focus is on building ad-hoc classifiers for visual video analytics.
Heimerl at al. \cite{HKBE12} propose an active learning approach for learning classifiers for text documents. Their approach includes three user interfaces: basic active learning, visualization of instances along the classifier boundary, and interactive instance selection.
Similar to our approach the classification-based visual analytics tool by Janetzko et al. \cite{Janetzko2014} also applies to the soccer application domain.
In contrast to our application goal, the approach supports building classifiers for interesting events in soccer games by taking user-defined training data into consideration.
User-defined labels for relevance feedback, interestingness, or class assignment share the idea to bind a single label to an instance, reflecting the classical machine learning approach ($f(i) = y$).
However, functions for learning the concept of similarity require a label representing the relation of pairs or groups of instances, e.g., in our case, $f(i_1, i_2) = y$, where y represents a similarity score in this case.
Visual-interactive user interfaces supporting such learning functions have to deal with this additional complexity.
A workaround strategy often applied for the validation of information retrieval results shows multiple candidates and asks the user for the most similar instances with respect to a given query.
We neglect this approach since our users do not necessarily have the knowledge to give feedback for any query instance suggested by the system.
Rather, we follow a user-centered strategy where users themselves have an influence on the selection of pairs of instances.
Another way to avoid complex learning functions is allowing users to explicitly assign weights to the attributes of the data set \cite{ware2001interactive,jeong2009ipca}.
The drawback of this strategy is the necessity of users knowing the attribute space in its entirety.
Especially when sophisticated descriptors are applied for the extraction of features (e.g., Fourier coefficients) or deep learned features, explicit weighting of individual features is inconceivable.
Rather, our approach applies an implicit attribute learning strategy.
While the similarity model indeed uses weighted attributes for calculating distances between instances (see Section \ref{rw:simiModeling}), an algorithmic model derives attribute weights based on the user feedback at object-level.
We conclude with a visual-interactive feedback interface where users are enabled to align small sets of instances on a two-dimensional arrangement \cite{bernardWSCG2014}.
The relative pairwise distances between the instances are then used by the similarity model.
We neglect strategies for arranging small sets of more than two instances in 2D since we explicitly want to include categorical and boolean attributes.
It has been shown that the interpretation of relative distances for categorical data is non-trivial \cite{vmv2014_ds}.
\subsection{Similarity Modeling}
\label{rw:simiModeling}
Aside from methods that employ visual interactive interfaces for learning the similarity between objects from user input as presented in the previous section, methods for the autonomous learning of similarity relations have been introduced~\cite{kulis2012metric,bellet2013survey}.
Human similarity perception is a psychologically complex process which is difficult to formalize and model mathematically. It has been shown previously that the human perception of similarity does not follow the rules of mathematical metrics, such as identity, symmetry and transitivity~\cite{tversky1977features}. Nevertheless, today most approaches employ distance metrics to approximate similarity estimates between two items (e.g., objects, images, etc.).
Common distance metrics are Euclidean distance (L2 distance) and Manhattan distance (L1 distance) \cite{yu2008distance}, as well as warping or edit distance metrics.
The edit distance was, e.g., applied to the soccer domain in a search system where users can sketch trajectories of player movement \cite{vda2016}.
To better take human perception into account and to better adapt the distance metric to the underlying data and task an increasingly popular approach is to \emph{learn} similarity or distance measures from data (metric learning). For this purpose different strategies have been developed.
In \emph{linear} metric learning the general idea is to adapt a given distance function (e.g., a Mahalanobis-like distance function) to the given task by estimating its parameters from underlying training data
The learning is usually performed in a supervised or weakly-supervised fashion by providing ground truth in the form of (i) examples of similar and dissimilar items (positive and negative examples), (ii) continuous similarity assessments for pairs of items (e.g., provided by a human) and (iii) triplets of items with relative constraints, such as A is more similar to B and C \cite{bellet2013survey,xing2003distance}. During training the goal is to find parameters of the selected metric that maximizes the agreement between the distance estimates and the ground truth, i.e., by minimizing a loss function that measures the differences to the ground truth. The learned metric can then be used to better cluster the data or to improve the classification performance in supervised learning.
Similar to these approaches, we also apply a linear model.
Instead of learning the distance metric directly, we estimate the Pearson correlation between the attributes and the provided similarity assessments.
In this way, the approach is applicable even to small sets of labeled pairs of instances.
The weights explicitly model the importance of each attribute and, as a by-product, enable the selection of the most important features for downstream approaches.
To facilitate the full potential, we apply weighted distance measures for internal similarity calculations, including measures for categorical \cite{boriah2008similarity} and boolean \cite{cha2005enhancing} attributes.
\input{primaryDataTable.tex}
In \emph{non-linear} metric learning, one approach is to learn similarity (kernels) directly without explicitly selecting a distance metric. The advantage of kernel-based approaches is that non-linear distance relationships can be modeled more easily.
For this purpose the data is first transformed by a non-linear kernel function.
Subsequently, non-linear distance estimates can be realized by applying linear distance measurements in the transformed non-linear space \cite{abbasnejad2012survey,torresani2006large}.
Other authors propose \emph{multiple kernel learning}, which is a parametric approach that tries to learn a combination of predefined base kernels for a given task \cite{gonen2011multiple}.
Another group of non-linear approaches employs neural networks to learn a similarity function \cite{norouzi2012hamming,salakhutdinov2007learning}.
This approach has gained increasing importance due to the recent success of deep learning architectures \cite{chopra2005learning,Zagoruyko_2015_CVPR,bell2015learning}. The major drawback of these methods is that they require huge amounts of labeled instances for training which is not available in our case.
The above methods have in common that the learned metric is applied globally to all instances in the dataset. An alternative approach is \emph{local metric learning} that learns specific similarity measures for subsets of the data or even separate measures for each data item \cite{frome2007learning,weinberger2009distance,noh2010generative}. Such approaches have advantages especially when the underlying data has heterogeneous characteristics. A related approach are per-exemplar classifiers which even allow to select different features (attributes) and distance measures for each item.
Per-exemplar classification has been applied successfully for different tasks in computer vision \cite{malisiewicz2011ensemble}. While our proposed approach to similarity modeling operates in a global manner, our active learning approach exploits local characteristics of the feature space by analyzing the density of labeled instances in different regions for making suggestions to the user.
The approaches above mostly require large amounts of data as well as ground truth in terms of pairs or triplets of labeled instances.
Furthermore, they rely on numerical data (or at least non-mixed data) as input. We propose an approach for metric learning for unlabeled data (without any ground truth) with mixed data types (categorial, binary, and numerical), which is also applicable to small datasets and data sets with initially no labeled instances.
For this purpose, we combine metric learning with active learning \cite{yang2012bayesian} and embed it in an interactive visualization system for immediate feedback.
Our approach allows the generation of useful distance metrics from a small number of user inputs.
|
\section{Introduction}
\label{sec.int}
Complex organic molecules, defined in astrochemistry as molecules that
consist of six or more atoms of which at least one is a carbon atom,
are widely found in star-forming regions \citep{herbstdishoeck2009}. A special category of complex molecules is that of the prebiotics, molecules that can be linked via their chemical structures or reactivity to life bearing molecules, such as amino-acids and sugars. Methyl isocyanate, CH$_{3}$NCO, also
known as isocyanomethane, is a molecule that falls in this category,
because of its structural similarity with a peptide bond (Figure
\ref{fig.structure}). This type of bond connects amino-acids to form
proteins and as such is interesting because it connects to chemistry relevant to the formation of the building blocks of life.
The majority of identified complex molecules has mainly been detected toward
high-mass hot cores, such as Orion KL and Sgr B2
\citep[e.g.,][]{blake1987,nummelin2000,belloche2013,belloche2014,tercero2013,neill2014,crockett2014},
but over the past decades detections toward low mass, sun-like,
protostars such as IRAS 16293-2422 (hereafter IRAS16293) have been
regularly reported as well. IRAS16293 (d = 120 pc) is considered as a protostellar template
for low-mass sources and is particularly rich in organic
molecules \citep{vandishoeck1995,cazaux2003,bottinelli2004,kuan2004,bisschop2008,jaber2014}. \citet{jorgensen2012} demonstrated the capabilities of the Atacama Large Millimeter/submillimeter Array (ALMA) with the detection of the prebiotic molecule glycolaldehyde \citep[see also][for a history of chemical studies of this source]{jorgensen2016}. More recently, other complex molecules (acetone, propanal, ethylene oxide) were identified in the framework of the PILS survey \citep{lykke2017}. Even the deuterated isotopologues of several complex molecules were detected toward this source \citep{parise2003,coutens2016,jorgensen2016}.
\begin{figure}
\begin{center}
\includegraphics[width=\hsize]{CH3NCO_peptide_bond.png}
\caption{Structure of methyl isocyanate (left), isocyanic acid (middle) and the generalized structure of a peptide bond (right). In each structure the components of the peptide bond are highlighted in red. R$_{1}$ and R$_{2}$ are different molecular functional groups, which can, among other possibilities, be a methyl group as for methyl isocyanate.}
\label{fig.structure}
\end{center}
\end{figure}
Unlike other molecules such as isocyanic acid (HNCO) and formamide
(NH$_{2}$CHO)\citep{bisschop2007,lopez-sepulcre2015,coutens2016},
which have a similar peptide-like structure as methyl isocyanate,
CH$_{3}$NCO only recently started to attract attention in the
astrochemistry community. This activity was inspired by a report from
\citet{goesmann2015} that CH$_3$NCO may be abundantly present on the
surface of comet 67P/Churyumov-Gerasimenko, as measured with the {\it
Cometary Sampler and Composition} (COSAC) instrument of Rosetta's
Philae lander. Its measured high abundance of 1.3\% with respect to
H$_2$O and CH$_3$NCO/HNCO=4.3 was based on the assumption that the mass peak recorded at
$m/z=57$ is dominated by CH$_3$NCO, as COSAC's low mass resolution does
not allow discrimination of different species with nearly identical mass. Recent
measurements with the {\it Rosetta Orbiter Spectrometer for Ion and
Neutral Analysis} (ROSINA) at much higher mass resolution showed
that the CH$_3$NCO coma abundance is significantly lower (K.\ Altwegg,
private communication). Nevertheless, the possibility of linking
complex molecules in star- and planet-forming regions with those
present in comets triggered the search for methyl isocyanate in the
interstellar medium.
Interstellar CH$_3$NCO was first detected toward Sgr B2(N) by
\citet{halfen2015} at low rotational temperatures of $\sim$25 K with a
column density ratio of $N$(HNCO)/$N$(CH$_{3}$NCO) = 35--53 depending
on the specific velocity component. \citet{cernicharo2016} detected
methyl isocyanate toward Orion KL at $N$(HNCO)/$N$(CH$_{3}$NCO)
$\simeq$ 15 and $T_{\rm ex}$ = 150 K. Their observations toward the
cold prestellar core B1-b did not yield a detection of the molecule
down to a upper column density limit of $< 2 \times$ 10$^{11}$
cm$^{-2}$ or $N$(HNCO)/$N$(CH$_{3}$NCO) > 42 \citep[based on a HNCO detection toward the same source by][]{lopez-sepulcre2015}. In the same
paper, publically accessible Sgr B2 observations from \citet{belloche2013} were
re-analysed with an updated spectral line list and, interestingly,
yielded a detection of warm methyl isocyanate at $T_{\rm ex} \simeq$
200 K and $N$(HNCO)/$N$(CH$_{3}$NCO) $\simeq$ 40. Attempts to also
identify the methyl isocyanate isomer CH$_{3}$CNO were unsuccessful
down to $N$(CH$_{3}$NCO)/$N$(CH$_{3}$CNO) $> 100$.
The astrochemical origin of methyl isocyanate is not yet understood
and this is partly due to the limited number of laboratory studies
that have been performed. \citet{hendersongudipati2015} tentatively
assigned a mass fragmentation peak to CH$_{3}$NCO after VUV
irradiating solid-state mixtures of NH$_{3}$:CH$_{3}$OH. In other
experiments by \citet{ruzi2012} UV irradiation of frozen
$n$-methylformamide (CH$_{3}$NHCHO) also yielded methyl isocyanate,
although it was concluded to represent a minor product channel.
A number of formation routes have been hypothesized by
astrochemists. \citet{halfen2015} postulated gas-phase formation by
HNCO or HOCN methylation:
\begin{equation}
\label{eq.halfen1}
\rm
HNCO/HOCN(g) + CH_{3}(g) \rightarrow CH_{3}NCO(g) + H(g)
\end{equation}
\noindent
or reactions of HNCO or HOCN with protonated methane, followed by electron recombination:
\begin{equation}
\label{eq.halfen2}
\rm
HNCO/HOCN(g) + CH_{5}^{+}(g) \rightarrow CH_{3}NCOH^{+}(g) + H_{2}(g)
\end{equation}
\begin{equation}
\label{eq.halfen3}
\rm
CH_{3}NCOH^{+}(g) + e^{-} \rightarrow CH_{3}NCO(g) + H(g)
\end{equation}
\citet{cernicharo2016} favoured solid-state formation mechanisms based
on the detection of CH$_{3}$NCO toward hot cores and its non-detection
in the cold dark cloud B1-b. Particularly, the methylation of HNCO has been
mentioned as a possible route to form methyl isocyanate in the solid-state, i.e., on the surface of an icy dust grain:
\begin{equation}
\label{eq.cernicharo1}
\rm
HNCO(s) + CH_{3}(s) \rightarrow CH_{3}NCO(s) + H(s)
\end{equation}
\citet{belloche2017} used the grain-surface radical-addition reaction
CH$_3$ + NCO $\rightarrow$ CH$_{3}$NCO in their models, with most of
the NCO formed via H-abstraction of HNCO:
\begin{equation}
\label{eq.belloche1}
\rm
HNCO(s) + H(s) \rightarrow NCO(s) + H_{2}(g)
\end{equation}
\begin{equation}
\label{eq.belloche2}
\rm
NCO(s) + CH_{3}(g,s) \rightarrow CH_{3}NCO(s)
\end{equation}
These postulated routes require the reactants to be present in
sufficient amounts. Gaseous HNCO is detected in high abundances in
protostellar environments and has been imaged in IRAS16293, showing it
to be prominent in both source A and B
\citep{bisschop2008,coutens2016}. It likely results from sublimation
of OCN$^-$, known to be a major ice component in low-mass protostellar
envelopes \citep{vanbroekhuizen2005}. A significant abundance of
CH$_3$ gas is a more speculative assumption since the molecule can only be observed
by infrared spectroscopy and has so far only been seen in diffuse gas
toward the Galactic Center \citep{feuchtgruber2000}. Alternatively,
CH$_3$ radicals can be produced {\it in situ} in ices by
photodissociation of known abundant ice components like CH$_4$ or
CH$_3$OH and then react with HNCO or OCN$^-$. This is the solid-state
route that is investigated here.
In this work we present the first detection of methyl isocyanate
toward both components of the low mass protobinary IRAS16293 on scales
of $< 100$ AU. An independent detection toward source B is also reported by \citet{martin-domenech2017}. A set of systematic laboratory experiments is presented in order to validate the solid-state formation routes of
CH$_{3}$NCO. The observational work is presented in Section
\ref{sec.obs} and the laboratory work in Section
\ref{sec.lab}. In Section \ref{sec.dis} the results of the
observations and laboratory experiments are compared and discussed in
the context of recent astrochemical models. The conclusions of this
paper are given in Section \ref{sec.con}.
\section{Observations}
\label{sec.obs}
\subsection{The ALMA PILS survey}
We searched for methyl isocyanate in the Protostellar Interferometric
Line Survey (PILS) data, an unbiased spectral survey of the low-mass
protostellar binary IRAS16293 with ALMA. A full description and data
reduction of the survey is presented in
\citet{jorgensen2016}. Briefly, this survey covers a spectral range
from 329.147 to 362.896 GHz and was obtained with both the 12m array
and the Atacama Compact Array (ACA). The beam size ranges between
$\sim$0.4$\arcsec$ and 0.7$\arcsec$ depending on the configuration at
the time of the observations. The rms of the combined data sets is
about 7--10 mJy beam$^{-1}$ channel$^{-1}$, i.e., approximately 4--5 mJy
beam$^{-1}$ km s$^{-1}$. To facilitate the analysis, the combined data set used
in this paper was produced with a circular restoring beam of
0.5$\arcsec$ at a spectral resolution of 0.2 km s$^{-1}$.
Two positions are analysed in this study. The first position is offset
by one beam diameter ($\sim$0.5$\arcsec$) from the continuum peak of
source B in the south west direction ($\alpha_{\rm J2000}$=16$^{\rm h}$32$^{\rm m}$22$\fs$58, $\delta_{\rm J2000}$=-24$\degr$28$\arcmin$32.8$\arcsec$) \citep[see high resolution images in][]{baryshev2015}. Source B presents narrow
lines (FWHM $\sim$ 1 km\,s$^{-1}$). This position is found to be
optimal for line identifications, as the lines are particularly bright,
do not have strong absorption features toward the bright continuum of
source B, and do not suffer from high continuum optical depth
\citep{coutens2016,lykke2017,jorgensen2016}. In this paper, we also
analyse source A, which exhibits broader lines than source B making the
line identification quite challenging \citep{pineda2012}. The
linewidth varies, however, depending on the spatial separation from
this source. With an average FWHM of $\sim$ 2.5 km\,s$^{-1}$, the
position $\alpha_{\rm J2000}$=16$^{\rm h}$32$^{\rm m}$22$\fs$90,
$\delta_{\rm J2000}$=$-24 \degr 28 \arcmin 36.2 \arcsec$ appears to be
one of the best positions to search for new species toward source A
(0.3$''$ offset). At this position, the emission is centered at $v_{\rm
LSR}$ of $\sim$0.8 km\,s$^{-1}$, blueshifted from the source A
velocity of $v_{\rm LSR}=3.2$ km s$^{-1}$
\citep{jorgensen2011}.
\citet{bisschop2008} found HNCO and other
nitrogen containing species such as CH$_3$CN to be more prominent toward source A
than source B. Consequently, it is also interesting to check whether there exists a small
scale chemical differentiation among the N-bearing species for the two
sources.
\subsection{Results}
Methyl isocyanate is detected toward both components A and B of
IRAS16293. The identification is based on spectroscopic data from the
Cologne Database for Molecular Spectroscopy (CDMS,
\citealt{muller2001,muller2005}), taken from \citet{cernicharo2016} as
well as from \citet{koput1986}. CH$_3$NCO is an asymmetric rotor with
the NCO group lying at an angle of 140$^o$ and a low-lying ($\sim182$
cm$^{-1}$) C-N-C bending mode $\nu_b$. Moreover, the barrier to
internal rotation of the CH$_3$ group is low, only 21 cm$^{-1}$. The
pure rotational spectrum therefore has $A$ and $E$ torsional states
and vibrationally excited transitions can become detectable at temperatures
of a few hundred K. The labelling of the states used here refers to the quantum
numbers $J$ (rotational angular momentum), $K_a$ and $K_c$ (projection of angular momentum on the respective inertial axes) with internal rotation interactions
indicated by the quantum number $m$, with $m$=0 and $\pm$3 for the $A$
states and $m$=1, -2 and 4 for the $E$ states
\citep{halfen2015,cernicharo2016}.
Using the CASSIS software\footnote{\url{http://cassis.irap.omp.eu/}},
we have been able to identify 43 unblended lines of CH$_3$NCO in the bending ground
state ($\nu_b$ = 0) with upper energy levels $E_{\rm up}$ ranging from
320 to 670 K toward source B (see Table \ref{table_obs}). Figure
\ref{fig_CH3NCO_vb0_B} shows the unblended lines detected toward this
component as well as the LTE modeling for two different excitation
temperatures at 100 and 300 K. Both excitation temperatures allow to reproduce the observations; only predicted transitions are observed. Our data are not sensitive to any
cold CH$_3$NCO component since lines with low $E_{\rm up}$ values are
missing in the spectral range of the PILS survey. Among the complex molecules that were detected and analysed towards source B, some (acetaldehyde, ethylene oxide) exhibit a relatively low excitation temperature of $\sim$125 K \citep{lykke2017}, while others (formamide, isocyanic acid, methanol, methyl formate, glycolaldehyde, ethylene glycol) show a higher excitation temperature of $\sim$ 300 K \citep{jorgensen2012,coutens2016,jorgensen2016}. Their spatial distribution is however rather similar and it is not possible to determine to which category CH3NCO belongs. The methyl isocyanate column density
is not very sensitive to the exact value of the excitation
temperature: assuming the same source size of 0.5$\arcsec$ as used in
previous PILS studies \citep{coutens2016,jorgensen2016,lykke2017}, the CH$_{3}$NCO
column density is found to be about 3\,$\times$\,10$^{15}$ cm$^{-2}$
and 4\,$\times$\,10$^{15}$ cm$^{-2}$ for $T_{\rm ex}$ = 300 and 100 K, respectively. All lines are optically thin. It is expected that for the same assumptions all column densities toward source B are accurate to better than 30\%.
Toward source A, most of the lines are blended due to the larger
linewidths ($\sim$ 2.5 km\,s$^{-1}$). We can, however, identify 11
unblended lines of CH$_3$NCO (see Figure \ref{fig_CH3NCO_vb0_A}). A
column density of $\sim$ 6\,$\times$\,10$^{15}$ cm$^{-2}$ and $\sim$ 9\,$\times$\,10$^{15}$ cm$^{-2}$ (assuming a
source size of 0.5$\arcsec$) is in good agreement with
observations for excitation temperatures of 300 and 100 K, respectively, again with an
uncertainty of about 30\%.
At high excitation temperatures, rotational levels in the first excited
bending state ($\nu_b$=1) may be populated as well and predictions for possible transitions are shown
in Figure \ref{fig.ch3nco_vb1}. For $T_{\rm ex}$ = 300 K, some faint
lines can indeed be tentatively attributed to CH$_{3}$NCO $\nu_b$=1 transitions
toward source B. An integrated intensity map of one of the
brightest CH$_{3}$NCO lines, the 39$_{0,39,0}$--38$_{0,38,0}$ transition at 336339.9 MHz, is presented in Figure
\ref{fig_map_MIC}. Similarly to other complex molecules, the emission
is quite compact with a size of $\sim$60 AU radius and centred near the two sources, with little difference between them
\citep{coutens2016,jorgensen2016,lykke2017}. For source B, the emission
is somewhat offset due to the continuum becoming optically thick on
source.
\begin{figure}
\begin{center}
\includegraphics[width=\hsize]{Map_CH3NCO_paper.pdf}
\caption{Integrated intensity map of the CH$_{3}$NCO 39$_{0,39,0}$--38$_{0,38,0}$ transition at 336339.9 MHz and $E_{\rm up}$ = 323.7 K between 1.7 and 3.7 km\,s$^{-1}$. The positions of the continuum of source A (South East source) and source B (North West source) are indicated with red crosses, while the positions studied in this paper are indicated with red circles.
The contour levels start at 5$\sigma$ with additional steps of 5$\sigma$. The circular restoring beam of 0.5$\arcsec$ size is indicated in grey in the right hand lower corner.}
\label{fig_map_MIC}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=17.8cm]{Model_CH3NCO_B_source_compar_new_Eup_1.pdf}
\includegraphics[width=17.8cm]{Model_CH3NCO_B_source_compar_new_Eup_2.pdf}
\caption{Black: detected unblended lines of CH$_3$NCO toward source B. Red solid: best-fit model for $T_{\rm ex}$ = 300 K. Green dashed: best-fit model for $T_{\rm ex}$ = 100 K. The $E_{\rm up}$ values of the lines are indicated in blue in the right upper part of each panel.}
\label{fig_CH3NCO_vb0_B}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=18cm]{Model_CH3NCO_A_source_compar_Eup.pdf}
\caption{Black: detected unblended lines of CH$_3$NCO toward source A. Red solid: best-fit model for $T_{\rm ex}$ = 300 K. Green dashed: best-fit model for $T_{\rm ex}$ = 100 K. The $E_{\rm up}$ values of the lines are indicated in blue in the right upper part of each panel.}
\label{fig_CH3NCO_vb0_A}
\end{center}
\end{figure*}
\begin{table*}
\centering
\caption{Comparison of molecular abundance ratios toward different sources}
\label{tab.comp_obs}
\begin{tabular}{lccccc}
\hline \hline
Source & HNCO/CH$_3$NCO & CH$_3$CN/CH$_3$NCO & CH$_3$OH/CH$_3$NCO & Reference \\
\hline
IRAS16293 B & 12 & 8 & 3333 & \multirow{2}{*}{This work}\\
IRAS16293 A & 4 & 9 & 3200 & \\
\hline
Orion KL A & 15 & 10 & 400 & \multirow{3}{*}{\citet{cernicharo2016}}\\
Orion KL B & 15 & 25 & 225 & \\
Sgr B2(N1) & 40 & 50 & 40 & \\
\hline
Sgr B2(N2) & 9 & 10 & 182 & \citet{belloche2016,belloche2017} \\
\hline
\multirow{2}{*}{Comet 67P/C-G} & \multirow{2}{*}{>0.2} & \multirow{2}{*}{>0.2} & \multirow{2}{*}{--} & \citet{goesmann2015} \\
& & & & Altwegg et al. in prep. \\
\hline
\end{tabular}
\\
\end{table*}
We also searched for spectral evidence of two isomers of methyl isocyanate - acetonitrile oxide (CH$_3$CNO, \citealt{winnewisser1982}) and methyl cyanate (CH$_3$OCN, \citealt{sakaizumi1990,kolesnikova2016}) - but the corresponding transitions were not found in the PILS data set. From this non-detection
$3\sigma$ upper limits of 5\,$\times$\,10$^{13}$
cm$^{-2}$ and 5\,$\times$\,10$^{14}$ cm$^{-2}$ are derived, respectively (assuming an excitation temperature of 100 K). These isomers are consequently less abundant than methyl
isocyanate by at least two and one orders of magnitude, in agreement with recent
findings toward Orion KL \citep{cernicharo2016,kolesnikova2016}.
In view of the important role that HNCO and HOCN may play in the
formation of CH$_{3}$NCO, we also searched for spectral signatures of these precursor species. The
analysis of the PILS data for HNCO toward source B is presented in
\citet{coutens2016}. The HNCO lines are optically thick, so lines of
the isotopologue HN$^{13}$CO were analysed and the HNCO column density
was derived assuming a $^{12}$C/$^{13}$C ratio of 68
\citep{milam2005}.
To get precise abundance ratios, we re-analysed the HN$^{13}$CO data using the same dataset (with the circular restoring beam of 0.5$\arcsec$) and obtained a HNCO column density of 3.7\,$\times$\,10$^{16}$ cm$^{-2}$, which is within the 30\% uncertainty range. The resulting HNCO/CH$_3$NCO abundance ratio is about 12
toward source B with an uncertainty of less than a factor 2.
Within the error margins, this is similar to the value derived in Orion KL
\citep{cernicharo2016} or Sgr B2(N2) \citep{belloche2016,belloche2017}.
For source A, the column density of HNCO is estimated to be
about 3.4\,$\times$\,10$^{16}$ cm$^{-2}$ ($T_{\rm ex}$ = 100 K).
The corresponding HNCO/CH$_3$NCO ratio is consequently about 4 toward this
component, with a somewhat larger uncertainty of a factor of 3 due to the difficulty of deriving a precise column density for HNCO because of line blending. Overall, we can
conclude that the two HNCO/CH$_3$NCO ratios are comparable toward the two components of the binary.
Whereas HNCO is readily identified, this is not the case for HOCN. No
HOCN lines were detected and a $3\sigma$ upper limit of
3\,$\times$\,10$^{13}$ cm$^{-2}$ is derived toward source B. HOCN is
consequently less abundant than HNCO and CH$_{3}$NCO by at least 3 and 2 orders
of magnitude, respectively. Therefore, it is highly unlikely that the
gas-phase formation pathway involving HOCN, as proposed by
\citet{halfen2015} in Eq.\ \ref{eq.halfen1} and \ref{eq.halfen2}, contributes significantly to the
production of methyl isocyanate in this source.
We have also determined the abundance ratios of CH$_{3}$NCO with
respect to CH$_3$OH and CH$_3$CN for comparison with other sources
in which methyl isocyanate has been detected (see Table
\ref{tab.comp_obs}). The column density of CH$_3$OH was estimated based on CH$_3^{18}$OH by J\o rgensen et al. (in prep.) for the one beam offset position toward source B ($\sim$ 1\,$\times$\,10$^{19}$ cm$^{-2}$) using a $^{16}$O/$^{18}$O ratio of 560 \citep{wilson1994}. With the same assumptions, we estimate a column density of CH$_3$OH of $\sim$ 2\,$\times$10$^{19}$ cm$^{-2}$ in source A. CH$_3$CN was analysed by Calcutt et al. (in prep.) toward both source A ($\sim$
8\,$\times$\,10$^{16}$ cm$^{-2}$) and source B ($\sim$
3\,$\times$\,10$^{16}$ cm$^{-2}$). Similarly to the HNCO/CH$_3$NCO ratio, the
abundance ratio of CH$_3$CN/CH$_3$NCO is comparable to the values
found in Orion KL \citep{cernicharo2016} and lower than
toward Sgr B2(N1), but again comparable to Sgr B2(N2) \citep{belloche2016,belloche2017}. Methanol is, however, clearly more abundant than methyl isocyanate toward IRAS16293 than toward Orion KL and Sgr B2.
The HNCO/CH$_3$NCO and CH$_3$CN/CH$_3$NCO abundance ratios derived in
IRAS16293 are much larger than the lower limits found in comet 67P
($>$0.2). A proper comparison awaits the new results derived with the
ROSINA instrument (Altwegg et al. in prep.).
\section{Laboratory experiments}
\label{sec.lab}
Grain surface formation routes of complex molecules have been investigated
experimentally for many years using cryogenic set-ups to mimic specific chemical processes under fully controlled laboratory conditions \citep[see][for
reviews]{herbstdishoeck2009,linnartz2015}. In the present work the
formation of methyl isocyanate is investigated by energetically processing solid-state
CH$_{4}$:HNCO mixtures with vacuum-UV radiation. VUV irradiated methane ice is known to efficiently yield methyl
radicals \citep{bossa2015}, and these radicals are expected to further react through surface diffusion with HNCO
to form CH$_{3}$NCO, as proposed by \citet{goesmann2015}
and \citet{cernicharo2016}, reaction \ref{eq.cernicharo1}. The Cryogenic Photoproduct Analysis Device
2 (CryoPAD2) in the Sackler Laboratory for Astrophysics is used to perform the measurements to investigate the role of methylation of HNCO in methyl isocyanate formation. A short description of this set-up, experimental procedure and
reference data is given below.
Note that in these experiments CH$_4$ is used as the parent of CH$_3$
but in interstellar space methyl radicals active in the ice may also originate from CH$_{3}$OH dissociation \citep{oberg2009a} or from direct CH$_{3}$ accretion from the gas-phase. The main aim of this section is to
investigate whether or not the proposed solid-state reaction
as shown in reaction \ref{eq.cernicharo1} can proceed at temperatures typical for dark cloud conditions in the ISM, i.e., independent of the exact precursor species.
\subsection{Set-up and method}
CryoPAD2 consists of a central chamber operating under ultra-high
vacuum conditions ($P \simeq$ 10$^{-10}$ mbar), to which various
instruments are attached. On the top a cryostat is mounted which cools a
gold-coated reflective surface down to 12 K. The temperature of this
surface can be controlled through resistive heating and varied
between 12 and 300 K, with an absolute temperature accuracy better than 1 K. Pure and mixed gases are prepared separately in a
gas-mixing system which is attached to a high-precision leak valve to
the main chamber and which doses the gases onto the cooled surface. The
deposited ice samples are irradiated with VUV radiation from a Microwave Discharge
Hydrogen-flow Lamp (MDHL), which emits radiation peaking at 121 nm and
a region between 140 to 160 nm, corresponding to photon energies in
the range of 7.5 to 10.2 eV \citep{chen2014,ligterink2015}. The total lamp flux
is (1.1$\pm$0.1)\,$\times$\,10$^{14}$ photons s$^{-1}$ that is determined using a calibrated NIST diode. CryoPAD2 possesses two diagnostic tools to monitor the VUV induced processes in the ice. The beam of a
Fourier-Transform IR Spectrometer (FTIRS) is directed under grazing
incidence onto the gold-coated surface, in order to perform Reflection
Absorption IR Spectroscopy (RAIRS). The sample preparation and chemical
changes under influence of VUV radiation are monitored by recording vibrational fingerprint spectra of molecules on the surface. To decrease the pertubing role of atmospheric CO$_{2}$ and H$_{2}$O, the system is purged with filtered dry air. The second instrument is
a highly sensitive Quadrupole Mass Spectrometer (QMS), with an ionization element at 70~eV, which is able to trace gas-phase
molecules in the chamber that are released from the ice surface upon linear heating during a Temperature Programmed Desorption (TPD) experiment. TPD is a very useful method that allows to identify desorbing species through their specific desorption temperature and mass fragmentation pattern. Unambiguous identifications become possible through the use of isotopologues and searching for the corresponding mass shifts of specific fragments. Obviously, TPD comes with the destruction of the ice.
In the experiments CH$_{4}$ (Linde Gas, 99.995\% purity),
$^{13}$CH$_{4}$ (Sigma-Aldrich, 99\% purity) and HNCO are used. HNCO
is produced from cyuranic acid (Sigma-Aldrich, 98\% purity), the solid
trimer of HNCO, following the protocol described in
\citet{vanbroekhuizen2004}. Impurities of the HNCO production process are
removed by freeze-thaw cycles to obtain a HNCO purity of >99\%.
For the experiments samples of pure HNCO and methane, and mixtures of
$^{13}$CH$_{4}$/CH$_{4}$:HNCO at 5:1 ratio are prepared. This ratio
is within a factor of 2 of that observed for interstellar ices
\citep{oberg2011} but is particularly chosen to create a large abundance of CH$_3$
radicals to test whether reaction \ref{eq.cernicharo1} proceeds or not. Homogeneously mixed ices are grown on the surface at 20 K and irradiated with a total fluence
of $\sim$3.3$\times$10$^{17}$ photons. During irradiation of the
sample, IR spectra are continuously recorded at 1 cm$^{-1}$
resolution. After the irradiation TPD is started, while still
recording IR spectra.
The strongest vibrational features of solid methyl isocyanate are
found between 2320 and 2280 cm$^{-1}$ ($\sim$4.34 $\mu$m) for the
-N=C=O antisymmetric stretching vibration and overtone 2$\nu_{7}$ CH$_{3}$
rocking mode. \citet{sullivan1994} lists these at 2320, 2280, 2270, 2240 and 2228 cm$^{-1}$, with 2280 cm$^{-1}$ being the strongest band. \citet{zhou2009} positioned all bands around 2300 cm$^{-1}$ and \citet{reva2010} puts the band for methyl isocyanate in a nitrogen matrix at 2334.7, 2307.9, 2288.9, 2265.2 and 2259.7 cm$^{-1}$, finding the strongest transition at 2288.9 cm$^{-1}$. The region around these bands is used to monitor CH$_{3}$NCO formation in the ice. We focus on the region between 2400 -- 2100 cm$^{-1}$. From previous experiments it is known that CO$_{2}$ (2340 cm$^{-1}$),
OCN$^{-}$ (2165 cm$^{-1}$) and CO (2135 cm$^{-1}$) are produced from
HNCO (2265 cm$^{-1}$) upon irradiation and these photo-products also have spectral features in the region characterizing methyl isocyanate
\citep{raunier2004,vanbroekhuizen2004}. Other known products of HNCO
irradiation are formamide, urea and formaldehyde, which do not have any interfering IR
features in the region of interest. Energetic processing of methane
does not yield products that show up in the region of
interest. Products that are formed from methane are mainly ethane and
to a lesser extent ethene and ethyn
\citep{bennett2006,bossa2015}. These species are seen in our
spectra at their appropriate frequencies. The abundant formation of ethane demonstrates that CH$_3$ is
indeed produced in the experiments, since ethane is the direct product of methyl-radical recombination.
In order to identify methyl isocyanate in the gas-phase using TPD, the mass
fragmentation pattern available from the NIST database\footnote{NIST
Mass Spec Data Center, S.E. Stein, director, "Mass Spectra" in NIST
Chemistry WebBook, NIST Standard Reference Database Number 69,
Eds. P.J. Lindstrom and W.G. Mallard, National Institute of
Standards and Technology, Gaithersburg MD, 20899,
http://webbook.nist.gov.} is used. The fragmentation pattern at 70~eV comprises unique peaks at $m/z$ = 57 and 56 (hereafter also called the
primary and secondary mass peak), which have a $m/z$ 57:56 ratio of
5:2 and these will be used as main TPD mass tracers. Known products of pure HNCO and methane irradiation does not have a
mass fragmentation peak at $m/z$ = 57 (see also Appendix \ref{sec.mass_CH4_HNCO}).
It is important to mention that methyl isocyanate is severely toxic and specialised laboratories and equipment are needed to work with this substance. This complicates the extensive use of this species and for this reason additional experiments, starting from the pure precursor, have not been performed.
\subsection{Results -- IR Spectra}
Figure \ref{fig.FTIR_plot} presents the IR spectra recorded during the first 1$\times$10$^{17}$ photons
irradiating the $^{12/13}$CH$_{4}$:HNCO samples. All
spectra are normalized to the HNCO peak. Three known spectroscopic
features of CO$_{2}$, OCN$^{-}$ and CO (blue) show up during
irradiation. Also, two new features become visible around 2300 cm$^{-1}$
(red), which do not show up while processing samples of pure HNCO or
CH$_{4}$. Also a clear red shift of about 10 cm$^{-1}$ of the two features is seen between the
sample of $^{12}$CH$_{4}$ and $^{13}$CH$_{4}$, moving transitions at 2322 and
2303 cm$^{-1}$ to 2313 and 2294 cm$^{-1}$. These spectroscopic features
are therefore the result of a product formed in the reaction between methane and
isocyanic acid, and, since they are found close to known
CH$_{3}$NCO features (given by \citet{sullivan1994} and \citet{reva2010}), are plausibly identified with methyl isocyanate. Another feature is seen in the wing of the
HNCO peak around 2235 cm$^{-1}$, which does not clearly shift with
methane isotopologues. The origin of this band is unclear.
Bandstrength values for methyl isocyanate are not available from the literature, however, a rough indication of the amount of formed methyl isocyanate versus lost HNCO can be given by making the assumption that the bandstrength of the NCO antisymmetric stretch vibration of methyl isocyanate equals that of the corresponding vibration of HNCO. To obtain the ratio, the integrated area of the 2303 cm$^{-1}$ feature is divided by the integrated loss area of the 2265 cm$^{-1}$ HNCO band for a number of spectra. A ratio of $N$(HNCO)/$N$(CH$_{3}$NCO) = 100 -- 200 is found, which is about an order of magnitude higher than the ratio observed toward IRAS16293. It should be noted, that it is not $a priori$ clear whether solid state laboratory and gas phase astronomical abundances can be directly compared (see e.g. Chuang et al. 2017). It is likely that ongoing gas phase reactions, for one or the other species, change the solid state to gas phase ratios. It is also possible that non-linear RAIRS effects can offset the column density \citep{teolis2007} or that the CH$_{3}$NCO bandstrength is significantly different. In a same manner CH$_{3}$NCO photodestruction may affect the overall abundances.
\begin{figure}
\begin{center}
\includegraphics[width=\hsize]{FTIR_plot.pdf}
\caption{IR spectra taken over time for the $^{12}$CH$_{4}$:HNCO (A) and $^{13}$CH$_{4}$:HNCO (B) mixture. HNCO and the products CO$_{2}$, OCN$^{-}$ and CO are listed (blue). Spectroscopic features that coincide with CH$_{3}$NCO are found at the red lines and show a clear shift with the $^{13}$CH$_{4}$ isotopologue. One unidentified peak is found in the right wing of the HNCO peak, indicated by a question mark. }
\label{fig.FTIR_plot}
\end{center}
\end{figure}
\subsection{Results -- temperature programmed desorption}
The desorption temperature of methyl isocyanate has not been reported in the
literature, but the TPD traces of our experiments on UV processed CH$_{4}$:HNCO ices do show the combined
release of the primary and secondary masses of methyl isocyanate
$m$=57 and 56 at 145 K (Figure \ref{fig.mass_trace}). The verification experiment of pure HNCO did not show these masses and only $m/z$ = 56 (and no 57) was seen after the irradiation of pure methane ice, releasing at 105 K (Figure \ref{fig.mass_CH4_HNCO} in the Appendix). At the 145 K desorption peak the primary over secondary
mass ratio is around 1-1.5, lower than the value of 2.5 suggested by
NIST upon 70 eV electron impact ionization. The NIST calibration values are a good indicator of the values to be expected, but are to some extent set-up specific. In such cases, the expected reactant can be deposited directly and the fragmentation pattern can be studied adapted to the set-up in use. This is unfortunately not possible here due to the aforementioned toxicity of methyl isocyanate. Moreover, contributions from other reaction products cannot be fully excluded. In
Figure \ref{fig.mass_trace}B a second desorption peak is found around
205 K, which is seen only as a shoulder in panel A. It is unlikely that this peak is associated with methyl isocyanate, instead it shows that the chemical network involves the formation of other species as well. In fact, since the additional product(s) have a mass fragmentation
pattern which contributes to the secondary mass of methyl isocyanate,
this could explain why the primary/secondary mass ratio does not
exactly match with that given by NIST. Additionally, the unidentified product of pure methane irradiation could be trapped in the ice and contribute to other desorption peaks. Therefore the 145 K desorption peak is still consistent with methyl isocyanate. There is also no
other candidate molecule with a primary mass of 57 in the NIST
database that could plausibly explain the TPD spectra.
\begin{figure}
\begin{center}
\includegraphics[width=\hsize]{mass_trace.pdf}
\caption{TPD trace of the primary (A) and secondary (B) masses of $^{12/13}$CH$_{3}$NCO. Both the primary and secondary mass are seen being released around 145 K (blue line). The secondary mass shows another release peak around 205 K (green), which is suspected to also contribute to first release peak, thus altering respectively the $m/z$ 57/56 and 58/57 ratio.}
\label{fig.mass_trace}
\end{center}
\end{figure}
CryoPAD2 offers the unique feature to combine IR and TPD data, i.e., at the moment a specific ice feature starts thermally desorbing, the IR signal starts decreasing and simultaneously the mass signal is increasing. In the measurements presented here, this effect is not as clearly visible as in previous studies (e.g. \citet{oberg2009a}), but this may be partly due to the low final S/N of the feature, making it difficult to link RAIRS and TPD signals. The present data are as accurate as possible, but obviously their interpretation would benefit from focused experiments determining the thermal desorption peak of pure CH$_{3}$NCO and the IR bandstrengths.
\section{Linking observations and laboratory data}
\label{sec.dis}
Methyl isocyanate is detected in a significant abundance toward both components of
IRAS16293 with an excitation temperature of at least 100 K. The two
isomers, CH$_{3}$CNO and CH$_{3}$OCN, are not detected. Abundance
ratios with respect to HNCO and CH$_{3}$CN are found to be similar to
those found for Orion KL and comparable within a factor of a few to
Sgr B2, making a common formation pathway in these three sources
likely. Moreover, no chemical differentiation between source A and B
is found. Gas-phase production of CH$_{3}$NCO via HOCN (reaction
\ref{eq.halfen1}) can be seen as insignificant due to the low upper
limit on this molecule, but a gas-phase route involving HNCO cannot be excluded.
The laboratory experiments on the formation of CH$_{3}$NCO
strongly suggest that a solid-state formation scheme is possible, starting from two astronomically relevant precursor species, CH$_{4}$ and HNCO. The proposed reaction
CH$_{3}$ + HNCO $\rightarrow$ CH$_{3}$NCO + H indeed seems to proceed, confirming a solid state reaction pathway, but other routes such as hydrogen stripping CH$_{3}$ + HNCO
$\rightarrow$ CH$_{4}$ + NCO may take place in parallel. Reactions
involving the OCN$^{-}$ anion, which is abundantly formed in
these experiments, provide alternative routes. Irradiation of a sample of
OCN$^{-}$:CH$_{4}$, where OCN$^{-}$ is formed via the acid-base
reaction of HNCO with NH$_{3}$, will be interesting to study in future
work since this anion is also a well known component of interstellar ice \citep{boogert2015}.
\citet{belloche2017} have incorporated reactions \ref{eq.belloche1} and \ref{eq.belloche2} (see Section \ref{sec.int}) into the large
gas-grain model MAGICKAL \citep{garrod2013} and simulated the cold
collapse + warm-up phase of a hot core region. Although tailored to
the high-mass source SgrB2(N) many of the model features are also
expected to be valid for the case of IRAS16293. In this model,
HNCO is formed by reaction of NH + CO, a route that has been
demonstrated experimentally by \citet{fedoseev2015,fedoseev2016}. Depending on assumptions about the barriers of the grain surface reactions involved in the formation of
HNCO and CH$_3$NCO, abundance ratios HNCO/CH$_3$NCO of 3-50 following
ice sublimation are readily found in the models, consistent with the
IRAS16293 observations.
\section{Conclusions}
\label{sec.con}
The main observational and experimental conclusions are listed below:
\begin{enumerate}
\item Methyl isocyanate is detected for the first time toward a low mass protostar, IRAS16293, on solar system scales (emission radius of 60 AU). Column densities of $\sim$ (3--4)\,$\times$\,10$^{15}$ and $\sim$ (6--9)\,$\times$\,10$^{15}$ cm$^{-2}$ are obtained toward source B and A, respectively, yielding $N$(HNCO)/$N$(CH$_{3}$NCO) = 12 and 4, with no significant variation between the two sources.
\item The abundance ratios of CH$_3$NCO relative to the N-bearing species HNCO and CH$_3$CN are similar to those found toward Orion KL and deviate by at most an order of magnitude from Sgr B2.
\item The isomers of methyl isocyanate, CH$_{3}$CNO and CH$_{3}$OCN, are not detected. These species are less abundant than CH$_3$NCO by at least a factor 100 and 10, respectively.
\item HOCN is not found down to <3 $\times$ 10$^{13}$ cm$^{-2}$, giving $N$(HOCN)/$N$(HNCO) > 1000, which makes this an insignificant gas-phase precursor to methyl isocyanate in IRAS16293.
\item Laboratory experiments strongly suggest that it is possible to form CH$_{3}$NCO on an icy surface, irradiating CH$_{4}$ and HNCO as astronomically relevant precursor species with VUV light, generating methyl radicals as reactive intermediates to form methyl isocyanate.
\end{enumerate}
The detection of CH$_{3}$NCO adds to the growing list
of complex molecules known to be present around solar mass protostars,
showing that the ingredients for prebiotic molecules are
abundant. Future deeper searches for even more complex molecules
relevant for the origin of life are warranted.
\section*{Acknowledgements}
This paper makes use of the following ALMA data:
ADS/JAO.ALMA\#2013.1.00278.S. ALMA is a partnership of ESO
(representing its member states), NSF (USA) and NINS (Japan), together
with NRC (Canada) and NSC and ASIAA (Taiwan), in cooperation with the
Republic of Chile. The Joint ALMA Observatory is operated by ESO,
AUI/NRAO and NAOJ. Astrochemistry in Leiden is supported by the European Union A-ERC
grant 291141 CHEMPLAN, by the Netherlands Research School for
Astronomy (NOVA) and by a Royal Netherlands Academy of Arts and Sciences
(KNAW) professor prize. CryoPAD2 was realized with NOVA and NWO (Netherlands Organisation for Scientific Research) grants. The work of A.C. was funded by the STFC grant ST/M001334/1. The group of J.K.J. acknowledges support from a Lundbeck Foundation Group Leader Fellowship, as well as the ERC under the European Union's Horizon 2020 research and innovation programme through ERC Consolidator Grant S4F
(grant agreement No 646908). Research at the Centre for Star and Planet Formation is funded by the Danish National Research Foundation.
The authors of this paper thank A. Belloche (Max-Planck-Institut f\"ur Radioastronomie) for useful input to the paper. We would like to thank the referee for the constructive comments.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{sec:introduction}
Highly dynamic networks are made of dynamic (often mobile) entities such as vehicles, drones, or robots.
It is generally assumed, in these networks, that the set of entities (nodes) is constant, while the set of communication
links varies over time. Many classical assumptions do not hold in these networks. For example, the topology may be
disconnected at any instant. It may also happen that an edge present at some time never appears again in the future.
In fact, of all the edges that appear at least once, one can distinguish between two essential sets: the set of {\em
recurrent} edges, which always reappear in the future (or remain present), and the set of {\em non recurrent} edges
which eventually disappear in the future. The static graph containing the union of both edge sets is called the {\em
footprint} of the network~\cite{CFQS12}, and its restriction to the recurrent edges is the {\em eventual footprint}
of the network~\cite{DLLP16}.
It is not clear, at first, what assumptions seem reasonable in a highly dynamic network. Special cases have been considered recently, such as
always-connected dynamic
networks~\cite{OW05}, $T$-interval connected networks~\cite{KLO10}, or networks the edges of which correspond to pairwise interactions obeying a uniform random scheduler (see e.g.~\cite{AADFP06,MCS11}).
Arguably, one of the weakest possible assumption is that any pair of nodes be able to communicate infinitely often through {\em temporal paths} (or journeys). Interestingly enough, this property was identified more than three decades ago by Awerbuch and Even~\cite{AE84} and remained essentially ignored afterwards. The corresponding class of dynamic networks (Class~5 in~\cite{CFQS12}---here referred to as {\ensuremath{{\cal TC^{\cal R}}}}\xspace for consistency with various notations~\cite{DKP15,GCLL15,AGMS15}) is however one of the most general and it actually includes the three aforementioned cases.
Dubois {\it et al.}~\cite{DKP15} observe that class ${\ensuremath{{\cal TC^{\cal R}}}}\xspace$ is actually the set of dynamic networks whose {\em eventual footprint} is connected. In other words, it is more than reasonable to assume that some of the edges are recurrent and their union does form a {\em connected} spanning subgraph.
Solving classical problems such as symmetry-breaking tasks relative to this particular set thus makes sense, as the nodes can rely forever on the corresponding solution, even though intermittently~\cite{CF13b,DKP15}.
Unfortunately, it is impossible for a node to distinguish between the set of recurrent edges and the set of non recurrent edges. So, the best the nodes can do is to compute a solution relative to the footprint, hoping that this solution still makes sense in the eventual footprint, whatever it is. (Whether, and how the nodes can learn the footprint itself is discussed later on.)
This context suggests a particular form of heredity which we call {\em robustness}. In classical terms, robustness can be formulated as the fact that a given property must be inherited by all the connected spanning subgraphs of the original graph. Significantly, this concept admits several possible interpretations, including the dynamic interpretation developed here. A more conventional, almost direct interpretation is that some edges in a classical (static) network are subject to permanent failure at some point, and the network is to be operated so long as it remains connected. While this interpretation is more intuitive and familiar, we insist on the fact that the dynamic interpretation of robustness is what makes its study compelling, for this notion arises naturally in class {\ensuremath{{\cal TC^{\cal R}}}}\xspace, which is one of the most general class of dynamic networks imaginable. The reader may adopt either interpretation while going through the paper, keeping in mind that our results apply to both contexts and are therefore quite general.
\noindent\textbf{Contributions.}
We investigate the concept of {\em robustness} of a property, with a focus on the {\em maximal independent set} (MIS) problem, which consists of selecting a subset of nodes none of which are neighbors (independence) and such that no further node can be added to it (maximality).
As it turns out, a robust MIS may or may not exist, and if it exists, it may or may not be computable locally depending on the considered graph ({\it resp.} footprint).
\begin{figure}[h]
\centering
\subfloat[]{
\label{fig:mis-a}
\begin{tikzpicture}
\tikzstyle{every node}=[circle, inner sep=1.2pt, fill=darkgray]
\path (0,0) node (a) {};
\path (a)+(0:1) node (b) {};
\path (a)+(-60:1) node (c) {};
\draw (a)--(b)--(c)--(a);
\tikzstyle{every node}=[circle, draw, red, inner sep=2pt]
\path (a) node {};
\end{tikzpicture}
}
~
\subfloat[]{
\label{fig:mis-b}
\begin{tikzpicture}
\tikzstyle{every node}=[circle, inner sep=1.2pt, fill=darkgray]
\path (0,0) node (a) {};
\path (a)+(-30:1) node (b) {};
\path (b)+(0:1) node (c) {};
\path (b)+(-60:1) node (d) {};
\path (c)+(30:1) node (e) {};
\draw (a)--(b)--(c)--(e);
\draw (b)--(d)--(c);
\tikzstyle{every node}=[circle, draw, red, inner sep=2pt]
\path (a) node {};
\path (c) node {};
\end{tikzpicture}
}
~
\subfloat[]{
\label{fig:mis-c}
\begin{tikzpicture}
\tikzstyle{every node}=[circle, inner sep=1.2pt, fill=darkgray]
\path (0,0) node (a) {};
\path (a)+(-30:1) node (b) {};
\path (b)+(0:1) node (c) {};
\path (b)+(-60:1) node (d) {};
\path (c)+(30:1) node (e) {};
\draw (a)--(b)--(c)--(e);
\draw (b)--(d)--(c);
\tikzstyle{every node}=[circle, draw, red, inner sep=2pt]
\path (a) node {};
\path (d) node {};
\path (e) node {};
\end{tikzpicture}
}
~
\subfloat[]{
\label{fig:mis-d}
\begin{tikzpicture}
\tikzstyle{every node}=[circle, inner sep=1.2pt, fill=darkgray]
\path (0,0) node (a) {};
\path (a)+(0:1) node (b) {};
\path (b)+(-90:1) node (c) {};
\path (c)+(-180:1) node (d) {};
\draw (a)--(b)--(c)--(d)--(a);
\tikzstyle{every node}=[circle, draw, red, inner sep=2pt]
\path (a) node {};
\path (c) node {};
\end{tikzpicture}
}
\vspace{-0.25cm}
\caption{\label{fig:examples} Four examples of MISs in various graphs (resp. footprints).}
\end{figure}
For example, if the graph is a triangle (see Figure~\ref{fig:mis-a}), then only one MIS exists up to isomorphism, consisting of a single node. However, this set is no longer maximal in one of the possible connected spanning subgraphs ({e.g.}\xspace after removing an adjacent edge to the selected node). Therefore, the triangle graph
admits no robust MIS. Some graphs do admit a robust MIS, but not all of the MISs are robust.
Figures~\ref{fig:mis-b} and~\ref{fig:mis-c} show two MISs
in the bull graph, only one of which is robust. Finally,
some graphs like the square graph (Figure~\ref{fig:mis-d}) are such that {\em all} MISs are robust. Although the last two examples seem to suggest that robust MISs are related to {\em maximum} MISs, being maximum is actually neither a necessary nor a sufficient condition.
In this paper, we characterize the class of graphs such that all MISs are robust, denoted \ensuremath{{\cal RMIS^\forall}}\xspace.
We prove that \ensuremath{{\cal RMIS^\forall}}\xspace consists {\em exactly} of the union of complete bipartite graphs and a new class of graphs called {\em sputniks}, which contains among others all the trees (for which any property is trivially robust). While the sufficient side is easy to establish, proving that these graphs are the only ones is more difficult. Interestingly,
while the best known algorithms for deterministic distributed MIS in general graphs are superlogarithmic in the number of nodes $n$, namely they take $2^{\mathcal{O}(\sqrt{\log n})}$ rounds~\cite{PS96} (better randomized algorithms are known~\cite{LW11}), graphs in \ensuremath{{\cal RMIS^\forall}}\xspace turn out to be specific enough to find an MIS (robust by definition) by using only information available within a sublogarithmic distance. We present an algorithm that first settles specific subsets of the networks using information available within {\em constant} distance, the residual instance being a disjoint union of trees. The residual instance can then be given to state-of-the-art algorithms like Barenboim and Elkin's for graphs of bounded arboricity~\cite{BE10}, which is known to use only information within distance $\mathcal{O}(\log n/\log \log n)$. An added benefit of this reduction is that any further progress on the MIS problem on trees will automatically transpose to robust MISs in \ensuremath{{\cal RMIS^\forall}}\xspace. (Note that we deliberately do not use the terms ``rounds'' or ``time'', due to the non equivalence of locality and time in the context of a footprint.)
Next, we turn our attention to general graphs and ask whether a robust MIS can be found (if one exists) using only local information. We answer negatively, proving an $\Omega(n)$ lower bound on the locality of the problem. This result implies a separation between the MIS problem and the robust MIS problem in general graphs, since the former is feasible within $2^{\mathcal{O}(\sqrt{\log n})}$ hops~\cite{PS96}. It also implies that no strategy is essentially better than collecting the network at a single node and subsequently solving the problem in an offline manner. Motivated by this observation, we consider the offline problem of finding a robust MIS in a given graph if one exists (and rejecting otherwise).
The trivial strategy amounts to enumerating all MISs until a robust one is found, however there may be exponentially many MISs in general graphs (Moon and Moser~\cite{MM65}, see also~\cite{F87,GGG88} for an extension to the case of connected graphs). We present a polynomial time algorithm for computing a robust MIS in any given graph (if one exists). Our algorithm relies on a particular decomposition of the graph into a tree of biconnected components (${\cal ABC}$-tree), along which constraints are propagated about the MIS status of special nodes in between the components. The inner constraints of non trivial components are solved by reduction to the 2-SAT problem (which {\em is} tractable). As a by-product, the set of instances for which a robust MIS is found characterizes the existential analogue of \ensuremath{{\cal RMIS^\forall}}\xspace, that is the class \ensuremath{{\cal RMIS^\exists}}\xspace of all graphs that {\em admit} a robust MIS. (Whether a closer characterization exists is left as an open question.)
\noindent\textbf{Further discussion on the dynamic interpretation.}
As pointed out, in a dynamic network there is no way to distinguish between recurrent and non recurrent edges,
therefore the nodes cannot learn the {\em eventual} footprint \cite{DLLP16} (this observation is the very basis of the notion of
robustness). Now, what about the union of both types of edges, that is, the footprint itself? Clearly, the footprint can never be {\em decided} in a definitive sense by the nodes, since some edges may appear arbitrary late for the first time. However, it is also clear that every edge of the footprint {\em will} eventually appear; thus, over time the nodes can learn the footprint in a stabilized way, by updating their representation as new edges are detected. It is therefore possible to update some structure or property that eventually relates to the correct footprint. (Alternatively, one may assume simply that prior information about the footprint is given to the nodes, or that an oracle informs the nodes once every edge of the footprint has appeared.) Again, the reader is free to ignore the dynamic interpretation if the static one makes for a sufficient motivation.
\noindent\textbf{Outline.} Section~\ref{sec:definitions} presents the main definitions and concepts. Then, we characterize in Section~\ref{sec:forallMIS} the class \ensuremath{{\cal RMIS^\forall}}\xspace and present a dedicated MIS algorithm that requires only information up to a sublogarithmic number of hops. Section~\ref{sec:existsMIS} establishes the non-locality of the problem in general and describes a tractable algorithm that computes a robust MIS in a given graph if one exists. Section~\ref{sec:conclusion} concludes with some remarks.
\section{Main concepts and definitions}
\label{sec:definitions}
Many of the concepts presented in the introduction, including that of temporal paths, footprint, or classes of dynamic networks are not defined here. The authors believe that the informal descriptions given in introduction are sufficient to understand the dynamic interpretation of the results. (If that is not the case, the reader is referred to~\cite{CFQS12} for thorough definitions using the time-varying graph formalism.) Our results themselves are formulated using standard concepts of graph theory, making them independent from both interpretations.
\subsection{Basic definitions}
Let $G=(V,E)$ be an undirected graph, with $V$ the set of nodes (vertices) and $E$ the set of bidirectional communication links (edges). We denote by $n=|V|$ the number of nodes in the graph, and by $D$ the {\em diameter} of the graph, that is, the length of the longest shortest path in $G$ over all possible pairs of nodes. We denote by $N(v)$ the neighborhood of a vertex $v$, which is the set of vertices $\{w : \{v,w\}\in E\}$. The degree of a vertex $v$ is $|N(v)|$. A vertex is {\em pendant} if it has degree~$1$. A {\em cut vertex} (or {\em articulation point}) is a vertex whose removal disconnects the graph. A {\em cut edge} (or {\em bridge}) is an edge whose removal disconnects the graph. We say that an edge is {\em removable} if it is not a cut edge.
A {\em spanning connected subgraph} of a graph $G=(V_G,E_G)$ is a graph $H=(V_H, E_H)$ such that $V_H = V_G$, $E_H \subseteq E_G$, and $H$ is connected.
In the most general variant, we define {\em robustness} as follows.
\begin{definition}[Robustness]
\label{def:robustness}
A property $P$ is said to be {\em robust} in $G$ if and only if it is satisfied in every connected spanning subgraph of $G$ (including $G$ itself).
\end{definition}
In other words, a robust property holds even after an arbitrary number of edges are removed without disconnecting the graph. Robustness is a special case of hereditary property, and more precisely a special case of decreasing monotone property (see for instance~\cite{K88}).
In this paper,
we focus on the {\em maximal independent set} (MIS) problem. An MIS is a set of nodes such that no two nodes in the set are neighbors and the set is maximal for the inclusion relation.
Following Definition~\ref{def:robustness}, a robust MIS in a graph $G$ (RMIS, for short) is an MIS that remains {\em maximal} and {\em independent} in every connected spanning subgraph of $G$. Observe that independence is stable under the removal of edges; therefore, it is sufficient that the MIS be maximal in all these subgraphs in order to be an RMIS. We define two classes of graphs related to the robustness of MISs.
\begin{definition}[\ensuremath{{\cal RMIS^\forall}}\xspace]
This class is the set of all graphs in which all MISs are robust.
\end{definition}
\vspace{-0.4cm}
\begin{definition}[\ensuremath{{\cal RMIS^\exists}}\xspace]
This class is the set of all graphs that admit at least one robust MIS.
\end{definition}
We define the distributed problem of computing an RMIS in a given graph as follows.
\begin{definition}[{\sc RobustMIS}\xspace problem]
\label{def:rmis-problem}
Given a graph $G$ and an algorithm ${\cal A}$ executed at every node of $G$, ${\cal A}$ solves {\sc RobustMIS}\xspace on $G$ iff every node eventually terminates by outputting IN or OUT, and the set of nodes outputting IN forms an RMIS on $G$. Algorithm ${\cal A}$ solves {\sc RobustMIS}\xspace in a class of graphs ${\cal C}$ iff for all $G \in {\cal C}$, ${\cal A}$ solves {\sc RobustMIS}\xspace on $G$.
\end{definition}
Finally, let us define two classes of graphs that turn out to be closely related to RMISs, namely {\em complete bipartite graphs} and {\em sputnik graphs}. The latter is introduced here for the first time.
\begin{definition}[Complete bipartite graph]
A complete bipartite graph is a graph $G=(V_1 \cup V_2,E)$ such that $V_1 \cap V_2 = \emptyset$ and $E = V_1 \times V_2$. In words, the vertices can be partitioned into two sets $V_1$ and $V_2$ such that every vertex in $V_1$ shares an edge with every vertex in $V_2$ (completeness), and these are the only edges (bipartiteness).
\end{definition}
\vspace{-0.4cm}
\begin{definition}[Sputnik]
A graph is a {\em sputnik} iff every vertex belonging to a cycle also has a pendant neighbor.
\LONG{
(An example of sputnik is shown in Figure~\ref{fig:LNF}.)
}
\end{definition}
\subsection{Computational model}
Based on the chosen interpretation of our results, the base graph in the above definitions refers
either to the footprint of a dynamic network, or to the network itself.
In the dynamic case, the actual timing of the edges is arbitrary, so the classical equivalence between time and locality in synchronous network does not hold. Nonetheless, we rely on the ${\cal LOCAL}$ model~\cite{L92,NS95} to describe the algorithms.
To avoid confusion between locality and time in the dynamic case, we always state the complexities in terms of locality, saying that an algorithm (or problem) is $\mathcal{O}(f(G))$-local if it can be solved in $\mathcal{O}(f(G))$ rounds in the ${\cal LOCAL}$ model. (Other terminologies include saying that such problems are in $LD(f(G))$\,\cite{FKP11}.)
For completeness, let us recall the main features of the $\mathcal{LOCAL}$ model.
In this model, the nodes operate in synchronous discrete rounds and they wake up simultaneously. In each round, a node can exchange messages of arbitrary
size with its neighbors and perform some local (typically unrestricted) computation.
The complexity of an algorithm over a class of graphs is the maximum number of rounds, taken over all graphs of this class, performed until all nodes have terminated. In the dynamic interpretation of our results, the algorithms are seen as being {\em restarted} every time the local knowledge of the footprint changes.
\section{Characterization of \ensuremath{{\cal RMIS^\forall}}\xspace and locality of {\sc RobustMIS}\xspace}
\label{sec:forallMIS}
In this section, we show that \ensuremath{{\cal RMIS^\forall}}\xspace, the class of graphs in which all MISs are robust, corresponds exactly to the union of complete bipartite graphs and sputnik graphs. Then we present an algorithm that solves {\sc RobustMIS}\xspace in \ensuremath{{\cal RMIS^\forall}}\xspace using information available only within a sublogarithmic number of hops in $n$.
\subsection{Characterization of \ensuremath{{\cal RMIS^\forall}}\xspace}
We first show that all MISs are robust in complete bipartite graphs and in sputnik graphs. Due to space limitation, the proofs of the two following lemmas are postponed to Appendix \ref{sec:missing}. They follow easily from the very definition of RMIS and of these classes of graphs.
\begin{lemma}
\label{lem:BK}
All MISs are robust in complete bipartite graphs.
\end{lemma}
\vspace{-0.4cm}
\begin{lemma}
\label{lem:sputniks}
All MISs are robust in sputnik graphs.
\end{lemma}
We now prove the stronger result that if a graph is such that all possible MISs are robust, then it {\em must} be either a bipartite complete graph or a sputnik.
\begin{lemma}
\label{lem:necessary}
If $G$ is not a sputnik, and yet every MIS in $G$ is robust, then $G$ is bipartite complete.
\end{lemma}
\begin{proof}
If $G$ is not a sputnik, then some node $u$ in a cycle $C \subseteq G$ has no pendant neighbor. In general, $u$ may be an articulation point, and so the graph $G \setminus \{u\}$ may result in several components. Let $X_1, X_2, \dots$ be the resulting components with vertex $u$ back in each of them. In particular, let $X_1$ be the one that contains $C$ and observe that $X_1$ contains at least $3$ vertices (cycle). The other components, if they exist, all contain at least two vertices other than $u$ (otherwise $u$ would have a pendant neighbor).
\noindent\textbf{Claim 1:} If all MISs in $G$ are robust, then all neighbors of $u$ in $X_1$ have the same neighborhood.
We prove this claim by contradiction.
Let two neighbors $v_1, v_2$ of $u$ be such that $N(v_1) \ne N(v_2)$. We will show that at least one MIS is not robust. Without loss of generality, assume that some vertex $x$ belongs to $N(v_1) \setminus N(v_2)$. Then we can build an MIS that contains both $v_2$ and $x$ (as a special case, $x$ may be the same vertex as $v_2$, but this is not a problem). For each of the components $X_{i\ge 2}$, choose an edge $\{u,w_i\} \in X_i$ and add another neighbor of $w_i$ to the MIS (such a neighbor exists, as we have already seen). One can see that $u$, $v_1$ and all $w_i$ can no longer enter the MIS because they all have neighbors in it. Now, choose the remaining elements of the MIS arbitrarily. We will show that the resulting MIS is not robust, by consider the removal of edges as follows. In all components $X_{i\ge 2}$, remove {\em all} edges incident to $u$ except $\{u,w_i\}$; and in $X_1$, remove all edges incident to $u$ except $\{u,v_1\}$. The resulting graph remains connected, by definition, since each of the $X_i \setminus \{u\}$ is connected. And yet, $u$ no longer has a neighbor in the MIS, which contradicts robustness.\hfill$\blacksquare$
Now, Claim 1 implies that none of $u$'s neighbors in $X_1$ has a pendant neighbor (since their neighborhoods are the same). As a result, the arguments that applied to $u$ because of its absence of pendant neighbors, apply in turn to $u$'s neighbors in $X_1$. In particular, it means that if some node $v$ is neighbor to $u$ in $X_1$, then all neighbors of $v$ (including $u$) must have the same neighborhood. Therefore, $u$ cannot be an articulation point and we are left with the single component $X_1$, in which all neighbors of $u$ have the same neighbors and these neighbors in turn have the same neighbors, which implies that the graph is complete bipartite.
\end{proof}
Based on Lemmas~\ref{lem:BK}, \ref{lem:sputniks}, and \ref{lem:necessary}, we conclude with the following theorem.
\begin{theorem}
\label{th:forallMIS}
All MISs are robust in a graph $G$ if and only if $G$ is complete bipartite or sputnik.
\end{theorem}
\subsection{{\sc RobustMIS}\xspace is locally solvable in \ensuremath{{\cal RMIS^\forall}}\xspace}
\label{sec:forallMIS-local}
\begin{figure}
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\tikzstyle{every node}=[defnode]
\path (1.5,10.3) node [] (v0) {};
\path (3.08,10.48) node [] (v3) {};
\path (1.9,8.36) node [] (v5) {};
\path (3.78,7.8) node [] (v9) {};
\path (6.48,7.38) node [] (v16) {};
\path (8.02,7.36) node [] (v17) {};
\path (7.32,10.8) node [] (v18) {};
\path (5.4,7.52) node [] (v19) {};
\path (8.88,7.4) node [] (v20) {};
\path (9.64,7.42) node [] (v24) {};
\path (10.38,7.44) node [] (v25) {};
\path (2.18,9.66) node [] (v1) {};
\path (3.26,9.68) node [] (v2) {};
\path (2.88,8.76) node [] (v4) {};
\path (6.56,8.86) node [] (v11) {};
\path (7.32,9.64) node [] (v12) {};
\path (8.02,8.86) node [] (v13) {};
\path (6.9,8.12) node [] (v14) {};
\path (7.66,8.1) node [] (v15) {};
\path (4.5,8.32) node [] (v8) {};
\path (9.98,8.24) node [] (v23) {};
\path (4.44,9.66) node [] (v6) {};
\path (9.02,9.82) node [] (v21) {};
\path (5.06,9.02) node [] (v7) {};
\path (5.52,9.92) node [] (v10) {};
\path (9.92,9.08) node [] (v22) {};
\tikzstyle{every path}=[];
\draw [darkgray] (v0)--(v1);
\draw [darkgray] (v1)--(v2);
\draw [darkgray] (v2)--(v4);
\draw [darkgray] (v4)--(v1);
\draw [darkgray] (v5)--(v4);
\draw [darkgray] (v2)--(v3);
\draw [darkgray] (v2)--(v6);
\draw [darkgray] (v7)--(v6);
\draw [darkgray] (v9)--(v8);
\draw [darkgray] (v8)--(v7);
\draw [darkgray] (v6)--(v10);
\draw [darkgray] (v10)--(v11);
\draw [darkgray] (v11)--(v12);
\draw [darkgray] (v12)--(v13);
\draw [darkgray] (v13)--(v11);
\draw [darkgray] (v11)--(v14);
\draw [darkgray] (v14)--(v15);
\draw [darkgray] (v15)--(v13);
\draw [darkgray] (v16)--(v14);
\draw [darkgray] (v17)--(v15);
\draw [darkgray] (v18)--(v12);
\draw [darkgray] (v19)--(v11);
\draw [darkgray] (v20)--(v13);
\draw [darkgray] (v21)--(v13);
\draw [darkgray] (v22)--(v21);
\draw [darkgray] (v23)--(v22);
\draw [darkgray] (v23)--(v24);
\draw [darkgray] (v23)--(v25);
\tikzstyle{every node}=[mis]
\path (v18) node [] {};
\path (v3) node [] {};
\path (v0) node [] {};
\path (v5) node [] {};
\path (v9) node [] {};
\path (v16) node [] {};
\path (v17) node [] {};
\path (v19) node [] {};
\path (v20) node [] {};
\path (v24) node [] {};
\path (v25) node [] {};
\path (v6) node [] {};
\path (v21) node [] {};
\path (v18) coordinate[xshift=5pt,yshift=5pt] (v18ne);
\path (v3) coordinate[yshift=5pt] (v3n);
\path (v0) coordinate[xshift=-7pt,yshift=6pt] (v0nw);
\path (v5) coordinate[xshift=-5pt,yshift=-4pt] (v5sw);
\path (v9) coordinate[yshift=-5pt] (v9s);
\path (v19) coordinate[yshift=-5pt] (v19s);
\path (v16) coordinate[yshift=-5pt] (v16s);
\path (v17) coordinate[yshift=-5pt] (v17s);
\path (v20) coordinate[yshift=-5pt] (v20s);
\path (v24) coordinate[yshift=-5pt] (v24s);
\path (v25) coordinate[xshift=5pt,yshift=-5pt] (v25se);
\path (v25) coordinate[xshift=5pt,yshift=5pt] (v25ne);
\path (v24) coordinate[yshift=5pt] (v24n);
\path (v20) coordinate[yshift=5pt] (v20n);
\path (v17) coordinate[yshift=5pt] (v17n);
\path (v16) coordinate[yshift=5pt] (v16n);
\path (v19) coordinate[yshift=5pt] (v19n);
\path (v9) coordinate[yshift=5pt] (v9n);
\path (v5) coordinate[xshift=4pt,yshift=4pt] (v5ne);
\path (v0) coordinate[xshift=4pt,yshift=-3pt] (v0se);
\path (v3) coordinate[yshift=-5pt] (v3s);
\path (v18) coordinate[xshift=5pt,yshift=-5pt] (v18se);
\draw[thick,blue,dotted,rounded corners] (v18ne)--(v3n)--(v0nw)--(v5sw)--(v9s)--(v19s)--(v16s)--(v17s)--(v20s)--(v24s)--(v25se)--(v25ne)--(v24n)--(v20n)--(v17n)--(v16n)--(v19n)--(v9n)--(v5ne)--(v0se)--(v3s)--(v18se)--cycle;
\path (v1) coordinate[xshift=-5pt,yshift=-3pt] (v1sw);
\path (v4) coordinate[xshift=-3pt,yshift=-5pt] (v4sw);
\path (v8) coordinate[yshift=-4pt] (v8s);
\path (v14) coordinate[yshift=-5pt] (v14s);
\path (v15) coordinate[yshift=-5pt] (v15s);
\path (v23) coordinate[xshift=5pt,yshift=-5pt] (v23se);
\path (v23) coordinate[xshift=5pt,yshift=4pt] (v23ne);
\path (v13) coordinate[xshift=4pt,yshift=4pt] (v13ne);
\path (v12) coordinate[yshift=7pt] (v12n);
\path (v11) coordinate[xshift=-4pt,yshift=4pt] (v11nw);
\path (v8) coordinate[yshift=4pt] (v8n);
\path (v2) coordinate[xshift=2pt,yshift=5pt] (v2n);
\path (v1) coordinate[xshift=-5pt,yshift=5pt] (v1nw);
\draw[thick,green,dashed,rounded corners] (v1sw)--(v4sw)--(v8s)--(v14s)--(v15s)--(v23se)--(v23ne)--(v13ne)--(v12n)--(v11nw)--(v8n)--(v2n)--(v1nw)--cycle;
\draw[thick,red!60] (v6)+(.57,-.05) circle (.7);
\draw[thick,red!60] (v21)+(.48,-.3) circle (.7);
\tikzstyle{every node}=[]
\path (v12n) node[right=5pt, green] {N};
\path (v5) node[below=10pt, blue] {P};
\path (v10) node[right=5pt, red] {F};
\end{tikzpicture}
\vspace{-0.5cm}
\end{center}
\caption{A sputnik and its sets $P$ (dotted set), $N$ (dashed set), and $F$ (plain set).}\label{fig:LNF}
\end{figure}
We now prove that computing deterministically an RMIS in class \ensuremath{{\cal RMIS^\forall}}\xspace can be done locally, by presenting a distributed algorithm that computes a (regular) MIS using only information available within $o(\log n)$-hops a sublogarithmic number of hops in $n$. By definition of the class, this MIS is robust.
Informally, the algorithm proceeds as follows (due to space limitations, the pseudo-code and the formal proof of the algorithm are moved to Appendix \ref{sec:missing}). Class \ensuremath{{\cal RMIS^\forall}}\xspace consists of exactly the union of bipartite complete graphs and sputniks (Theorem~\ref{th:forallMIS}). First, the nodes decide if the graph is complete bipartite by looking within a constant number of hops (three). If so, membership to the MIS is decided according to some convention ({e.g.}\xspace all nodes in the same part as the smallest identifier are in the MIS). Otherwise, the graph {\em must} be a sputnik and every node decides (without more information) which of the following three cases it falls into: 1) it is a pendant node (set $P$ in Figure~\ref{fig:LNF}), 2) it is not a pendant node but has at least one pendant neighbor (set $N$), or 3) none of the two cases apply (set $F$). In the first case, it enters the MIS, while in the second it decides not to. We prove that the set of nodes falling into the third case does form a disjoint union of trees, each of which can consequently be solved by state-of-the-art algorithms. In particular, Barenboim and Elkin~\cite{BE10} present a $\mathcal{O}(\log n/\log \log n)$-local algorithm that solves MIS in graphs of bounded arboricity (and a fortiori trees).
On the negative side, we show (using standard arguments) that Linial's $\Omega(\log^* n)$ lower bound for $3$-coloring~\cite{L92} in cycles extends to {\sc RobustMIS}\xspace in class \ensuremath{{\cal RMIS^\forall}}\xspace, leading to the following theorem.
\begin{theorem}\label{th:localityforall}
{\sc RobustMIS}\xspace is $\Omega(\log^* n)\cap\mathcal{O}(\log n / \log \log n)$-local in class \ensuremath{{\cal RMIS^\forall}}\xspace.
\end{theorem}
\section{Nonlocality of {\sc RobustMIS}\xspace in general graphs and global resolution}
\label{sec:existsMIS}
In this section, we prove that the problem of computing deterministically an RMIS in general graphs, if one exists, is {\em not} local. Precisely, we first observe that {\em deciding} whether an RMIS exists is not a local problem; then, we prove a $\Omega(n)$ lower bound on the distance at which it might be necessary to look to solve the problem if an RMIS exists, where $n$ is the diameter of the network. Motivated by this result, we present an offline algorithm that compute an RMIS, in polynomial time, if one exists. It can be used in a strategy where all the information about the network is collected at one node (or several, the algorithm being deterministic).
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=.7]
\tikzstyle{every node}=[defnode]
\path (-1,1) node [] (ma) {};
\path (0,0) node [] (ba) {};
\path (0,2) node [] (ha) {};
\path (1,1) node [] (mb) {};
\path (3,1) node [] (m0) {};
\path (4,0) node [] (b0) {};
\path (4,2) node [] (h0) {};
\path (5,1) node [] (m1) {};
\path (6,0) node [] (b3) {};
\path (6,2) node [] (h3) {};
\path (7,0) node [] (b3b) {};
\path (7,2) node [] (h3b) {};
\path (8,1) node [] (m4) {};
\path (9,0) node [] (b4) {};
\path (9,2) node [] (h4) {};
\path (10,1) node [] (m5) {};
\path (12,1) node [] (m7) {};
\path (13,0) node [] (b7) {};
\path (13,2) node [] (h7) {};
\path (14,1) node [] (m8) {};
\tikzstyle{every node}=[]
\path (ma) node[left] {$\beta_k$};
\path (ba) node[below] {$\gamma_k$};
\path (ha) node[above] {$\alpha_k$};
\path (mb) node[] {};
\path (m0) node[above,xshift=-5pt] {$\beta_1$};
\path (b0) node[below] {$\gamma_1$};
\path (h0) node[above,xshift=-2pt] {$\alpha_1$};
\path (m1) node[right,xshift=2pt] {$\beta_0$};
\path (b3) node[below] {$\gamma_0$};
\path (h3) node[above,xshift=2pt] {$\alpha_0$};
\path (b3b) node[below] {$c_0$};
\path (h3b) node[above] {$a_0$};
\path (m4) node[left] {$b_0$};
\path (b4) node[below] {$c_1$};
\path (h4) node[above] {$a_1$};
\path (m5) node[above,xshift=5pt] {$b_1$};
\path (m7) node[] {};
\path (b7) node[below] {$c_k$};
\path (h7) node[above] {$a_k$};
\path (m8) node[right] {$b_k$};
\path (mb) node[right=9pt] {$\dots$};
\path (m5) node[right=9pt] {$\dots$};
\draw (ma)--(ba)--(mb)--(ha)--(ma);
\draw (m0)--(b0)--(m1)--(h0)--(m0);
\draw (m1)--(b3)--(b3b)--(m4)--(h3b)--(h3)--(m1);
\draw (m4)--(b4)--(m5)--(h4)--(m4);
\draw (m7)--(b7)--(m8)--(h7)--(m7);
\tikzstyle{every node}=[mis]
\path (ha) node {};
\path (ba) node {};
\path (h0) node {};
\path (b0) node {};
\path (h3) node {};
\path (b3) node {};
\path (m4) node {};
\path (m5) node {};
\path (m7) node {};
\path (m8) node {};
\end{tikzpicture}
\vspace{-0.65cm}
\end{center}
\caption{The graph $G_k$ ($k\in\mathbb{N}$) and one of its two possible robust MISs.}\label{fig:gk}
\end{figure}
\subsection{{\sc RobustMIS}\xspace is non local in general graphs}\label{sub:rmisexists}
Let us first observe that the problem of deciding whether an RMIS exists is not local. Consider two graphs $G_1$ and $G_2$ which respectively consist of a $\mathcal{O}(n)$-long path and to a lollipop graph ({i.e.}\xspace a graph joining a $\mathcal{O}(n)$-long path to a clique of size $\mathcal{O}(n)$). Then, clearly, a node at one extremity of $G_1$ and the (unique) pendant node of $G_2$ cannot distinguish their $o(n)$ neighborhood (even with identifiers, which could be exactly the same in this neighborhood) whereas $G_1$ admits an RMIS and $G_2$ does not. We go further and prove that, even if some RMISs do exist, then finding one is non local.
To prove this result, we exhibit an infinite family of graphs $(G_k)_{k\in\mathbb{N}}$, each of which has diameter linear in $k$ (and $n$).
We first show through Lemmas~\ref{lem:gk-robust} and~\ref{lem:gk-only} that every $G_k$ admits only {\em two} RMISs $M_1$ and $M_2$ which are {\em complements} of each other; that is $M_2 = \overline{M_1} = V \setminus M_1$.
Intuitively, these MISs are such that two nodes at distance $\mathcal{O}(n)$ must take opposite decisions, although they have the same view up to distance $\mathcal{O}(n)$. (The real proof is more complex and involves showing that identifiers do not help either.) As a result, the nodes may have to collect information up to distance $\Omega(n)$ in order to decide consistently.
Let $(G_k)_{k\in\mathbb{N}}$ be an infinite famility of graphs defined as follows. Graph
$G_0=(V_0,E_0)$ is such that $V_0=\{a_0,b_0,c_0,\alpha_0,\beta_0,\gamma_0\}$
and $E_0$ induces a cycle $a_0$-$b_0$-$c_0$-$\gamma_0$-$\beta_0$-$\alpha_0$-$a_0$ as shown in Figure~\ref{fig:gk}.
Then $G_k=(V_k,E_k)$ is obtained
from $G_{k-1}=(V_{k-1},E_{k-1})$ as follows: $V_k=V_{k-1}\cup\{a_k,b_k,c_k,\alpha_k,\beta_k,\gamma_k\}$
and $E_k=E_{k-1}\cup\{\{\beta_{k-1},\alpha_{k}\},\{\beta_{k-1},\gamma_{k}\},$
$\{\alpha_{k},\beta_{k}\},\{\gamma_{k},\beta_{k}\},\{b_{k-1},a_{k}\},
\{b_{k-1},c_{k}\},\{a_{k},b_{k}\},\{c_{k},b_{k}\}\}$.
For any $k$, define $M_1$ as the set of nodes $\{\alpha_i,\gamma_i,b_i|i\le k\}$ and $M_2=\{a_i,c_i,\beta_i|i \le
k\}$. Observe that $M_2 = V_k\setminus M_1$ (written $\overline{M_1}$). Set $M_1$ is illustrated in Figure~\ref{fig:gk}.
\begin{lemma}\label{lem:gk-robust}
For any $k\geq 0$,
$M_1$ and
$M_2$
are RMISs in $G_k$.
\end{lemma}
\begin{proof}
We prove this for $M_1$. The same holds symmetrically for $M_2$.
First, observe that $M_1$ is a valid MIS: no two of its nodes are neighbors by construction (independence) and all nodes in $V_k\setminus M_1$ have neighbors in $M_1$ (maximality).
Now, to obtain a connected spanning subgraph of $G_k$, one
can remove from $E_k$ at most one edge from each simple cycle of $G_k$. Since any node of
$V_k\setminus M_1$ has a number of neighbors in $M_1$ strictly greater than the number of simple cycles it belongs to, $M_1$ is robust.
\end{proof}
\begin{lemma}\label{lem:gk-only}
For any $k\geq 0$, $M_1$ and $M_2$ are the only two RMISs in $G_k$, implying that nodes $b_k$ and $\beta_k$ must take opposite decisions in all RMISs.
\end{lemma}
\begin{proof}
We say that an edge $e \in E$ is \emph{critical} with respect to some MIS $M$ in $G_k$
if $e$ is removable ({i.e.}\xspace not a cut edge) and $M$ is no longer maximal in $(V_k,E_k\setminus\{e\})$. The existence of a critical edge implies that the considered MIS is not robust.
Let us now consider an RMIS $M$ in $G_k$. We prove several claims
on $M$.
\noindent\textbf{Claim 1:} If $\alpha_0\in M$, then $\{\gamma_0,b_0\} \subseteq M$ and $\{\beta_0,c_0,a_0\} \subseteq \overline{M}$.
If $\alpha_0\in M$, then $\{\beta_0,a_0\} \subseteq \overline{M}$ (independence). It also holds that $b_0\in M$, otherwise the edge $\{\alpha_0,a_0\}$ is critical (robustness). It follows that $c_0\notin M$ (independence) and $\gamma_0\in M$ (maximality).
\noindent\textbf{Claim 2:} If $\alpha_0\notin M$, then $\{\gamma_0,b_0\} \subseteq \overline{M}$ and $\{\beta_0,c_0,a_0\} \subseteq M$. \hfill {\it (symmetric to Claim~1)}
\noindent\textbf{Claim 3:} If $\alpha_0\in M$, then $\{\alpha_i,\gamma_i\}\subseteq M$ and $\beta_i\notin M$ for all $i\le k$.
By contradiction, if $\beta_i\in M$ for some $i$, then $\{\alpha_i,\gamma_i\} \subseteq \overline{M}$ (independence). Let $i$ be smallest possible, then edges $\{\alpha_i,\beta_i\}$ and
$\{\beta_i,\gamma_i\}$ are critical {\em w.r.t.}\xspace $M$ (recall that, if
$i>0$, $\beta_{i-1}\notin M$ by construction and $\beta_0\notin M$
by Claim 1), which contradicts robustness. Therefore, $\beta_i\notin M$.
The maximality of $M$ allows us to conclude.
\noindent\textbf{Claim 4:} If $\alpha_0\notin M$, then $\{\alpha_i,\gamma_i\}\subseteq \overline{M}$ and $\beta_i\in M$ for all $i\le k$. \hfill {\it (symmetric to Claim 3)}
\noindent\textbf{Claim 5:} If $\alpha_0\in M$, then $\{a_i,c_i\}\subseteq \overline{M}$ and $b_i\in M$ for all $i\le k$.
By contradiction, if $b_i\notin M$ for some $i$, then $\{a_i,c_i\}\subseteq \overline{M}$ (independence). Let $i$ be smallest possible (recall that, if
$i>0$, $b_{i-1}\in M$ by construction and $b_0\in M$
by Claim 1). Let $w$ be $b_0$ if $i=0$ and be $b_{i-1}$ otherwise.
The edges $\{w,a_i\}$ and $\{w,\gamma_i\}$
are then critical, which contradicts robustness. Therefore, $b_i\in M$.
The independence of $M$ allows us to conclude.
\noindent\textbf{Claim 6:} If $\alpha_0\notin M$, then $\{a_i,c_i\}\subseteq M$ and $b_i\notin M$ for all $i\le k$. \hfill {\it (symmetric to Claim 5)}
Claims 1 to 6 imply that $M=M_1$ if $\alpha_0\in M$ and $M=M_2$ otherwise.
\end{proof}
Finally, we relate these results to the locality of the {\sc RobustMIS}\xspace problem.
\begin{theorem}\label{th:localityexists}
{\sc RobustMIS}\xspace requires the nodes to use information up to distance $\Omega(n)$ in $G_k$.
\end{theorem}
\begin{proof}
The proof would be straightforward in an anonymous network, due to the fact that $b_k$ and $\beta_k$ have indistinguishable structural neighborhoods ({\em a.k.a.}\xspace views~\cite{YK96}) up to distance $\mathcal{O}(n)$, and yet, they must take different decisions (Lemma~\ref{lem:gk-only}). Unique identifiers make the argument more complicated, since $b_k$ and $\beta_k$ do have unique {\em labeled} views ({i.e.}\xspace views taking into account identifiers) even at distance $0$.
Let us call $b_k$ and $\beta_k$ the {\em extremities} of the network. Observe that the distance between both extremities is larger than $4k$.
Let ${\cal L}_1$, ${\cal L}_2$, and ${\cal L}_3$ be three possible labeling functions that assign unique identifiers to the neighborhood of an extremity up to distance $k$ (say) and such that the three labelings have no identifier in common. Let $G_k^1$ be the labeled graph whose structure is isomorphic to $G_k$, in which the neighborhood of $b_k$ is labeled according to ${\cal L}_1$ and the neighborhood of $\beta_k$ is labeled according to ${\cal L}_2$; the rest of the nodes are labeled arbitrarily. Let $G_k^2$ be defined similarly, but using ${\cal L}_3$ instead of ${\cal L}_2$ in the neighborhood of $\beta_k$. Finally, let $G_k^3$ be defined similarly, but using ${\cal L}_2$ in the neighborhood of $b_k$ and ${\cal L}_3$ in the neighborhood of $\beta_k$. Now, if $b_k$ and $\beta_k$ use only information up to distance $k$, then they must take identical decisions in at least one of the three labeled graphs, contradicting Lemma~\ref{lem:gk-only}.
\end{proof}
\subsection{A global algorithm to compute a robust MIS (if one exists)}
We now describe an algorithm that tests constructively whether an RMIS exists in a graph $G$.
Our algorithm relies on the construction of an auxiliary tree called {\em \ensuremath{\mathcal{ABC}}-tree\xspace}, which represents a particular decomposition of
the graph based on biconnected components (it is neither a block-cut tree, nor a bridge tree, but a mix of these two types of decomposition). Roughly speaking, our algorithm works by propagating constraints about the MIS along
the \ensuremath{\mathcal{ABC}}-tree\xspace .
Each non-trivial component is solved on the way up by means of a reduction of its constraints to 2-SAT (which {\em is} polynomial-time solvable).
In the following, we describe how the \ensuremath{\mathcal{ABC}}-tree\xspace is built over $G$. It is followed by
an informal presentation of the algorithm---due to the lack of space, the formal algorithm has been moved to
Appendix~\ref{sec:pseudo} and its proof is presented in Appendix~\ref{sec:bigproof}.
\noindent\textbf{Decomposition of $G$.}
In the context of this section, we call {\em biconnected component} (or simply {\em component}) in $G$ a maximal
subgraph $H\subseteq G$ such that the removal of any node in $H$ does not disconnect $H$ ({i.e.}\xspace $H$ is $2$-{\it vertex}-connected).
By abuse of notation, we write $u \in H$ if $u$ is a vertex of the subgraph $H$.
We consider here a mix of the so-called block-cut tree and bridge tree and refer to it as the {\it \ensuremath{\mathcal{ABC}}-tree\xspace}.
Let \ensuremath{\mathcal{S}}\xspace be the set of biconnected components of $G$ (see Figure~\ref{fig:ex_compo} for an illustration).
\begin{figure}
\begin{center}
\includegraphics[width=.55\textwidth] {example_composant.pdf}
\hspace{-35pt}
\includegraphics[width=.47\textwidth]{example_abctree.pdf}
\end{center}
\vspace{-0.65cm}
\caption{\label{fig:ex_compo}
Decomposition of a graph into biconnected components (left) and the corresponding \ensuremath{\mathcal{ABC}}-tree\xspace (right).
Nodes with shapes~$\circ$ and~{\scriptsize $\Diamond$} correspond respectively to pendant vertices ($\in \ensuremath{\mathcal{P}}\xspace$) and articulation points ($\in \ensuremath{\mathcal{A}}\xspace$).
Nodes with shape~{\tiny $\Box$} correspond to bridge edges ($\in \ensuremath{\mathcal{B}}\xspace$), and nodes with shape~{\tiny $\bigcirc$} show components that contain at least $3$ vertices ($\in \ensuremath{\mathcal{C}}\xspace$).
}
\end{figure}
Two adjacent components either share a common articulation point or they are linked by a bridge.
For instance, node $10$ in Figure~\ref{fig:ex_compo} is common to components $J$ and $M$, and edges $\{6,7\}$ and $\{8,21\}$ are bridges between $D$ and $J$, and $J$ and $K$ respectively.
Let \ensuremath{\mathcal{A}}\xspace be the set of all articulation points (whether or not they are shared) and \ensuremath{\mathcal{B}}\xspace the set of bridges. Let \ensuremath{\mathcal{P}}\xspace be the set of all pendant vertices, which form singleton components---in
Figure~\ref{fig:ex_compo}, $\ensuremath{\mathcal{P}}\xspace = \{4,5,7,12,20,24\}$. Finally, let \ensuremath{\mathcal{C}}\xspace be the set of components that contain three or more vertices---$\ensuremath{\mathcal{C}}\xspace = \{H,J,K,M,N\}$.
Remark that if \ensuremath{\mathcal{C}}\xspace is empty, then $G$ is acyclic.
The ${\cal ABC}$ decomposition (or \ensuremath{\mathcal{ABC}}-tree\xspace) of $G$, is the graph ${\ensuremath{T_\mathcal{C}}}\xspace=(V', E')$ such that $V' = \ensuremath{\mathcal{A}}\xspace \cup \ensuremath{\mathcal{B}}\xspace \cup \ensuremath{\mathcal{C}}\xspace \cup \ensuremath{\mathcal{P}}\xspace$ and $E'$ is
defined by the two following rules:
$(i)$ $\forall a \in \ensuremath{\mathcal{A}}\xspace$, $\forall c \in \ensuremath{\mathcal{C}}\xspace$, $\{a,c\} \in E'$ if and only if $a \in c$; and
$(ii)$ $\forall b=\{u,v\} \in \ensuremath{\mathcal{B}}\xspace$, $\{b,u\} \in E'$ and $\{b,v\} \in E'$.
Figure~\ref{fig:ex_compo} (right side) shows the \ensuremath{\mathcal{ABC}}-tree\xspace corresponding to the graph of the left side.
\noindent\textbf{Algorithm.}
The algorithm works on {\ensuremath{T_\mathcal{C}}}\xspace, the \ensuremath{\mathcal{ABC}}-tree\xspace made over the four sets \ensuremath{\mathcal{A}}\xspace, \ensuremath{\mathcal{B}}\xspace, \ensuremath{\mathcal{C}}\xspace, and \ensuremath{\mathcal{P}}\xspace with respect to $G$.
If the set \ensuremath{\mathcal{C}}\xspace is empty, it means that $G$ is acyclic. In that case, $G$ trivially admits an RMIS,
which is returned by the algorithm.
Otherwise ($\ensuremath{\mathcal{C}}\xspace \neq \emptyset$), a component vertex $C \in \ensuremath{\mathcal{C}}\xspace$ is arbitrarily selected to be the root of {\ensuremath{T_\mathcal{C}}}\xspace,
denoted by \ensuremath{\mathcal{R_{\CT}}}\xspace.
Then, the classical concepts in oriented trees, such as {\em children}, {\em parent},
{\em descendant}, {\em subtree}, or {\em leaf} apply to the vertices of {\ensuremath{T_\mathcal{C}}}\xspace.
For ease of reading, we abuse the term ``{\em admit}'' by saying that a subtree $\hat{T}$ of {\ensuremath{T_\mathcal{C}}}\xspace ``{\em admits an
RMIS}'' if the subgraph of $G$ corresponding to $\hat{T}$ admits an RMIS.
At the higher level, the algorithm proceeds within two phases. Based on {\ensuremath{T_\mathcal{C}}}\xspace, the first one is called the {\em
labeling} phase.
Initiated from the leaves of the tree, it evaluates whether the subgraph of $G$ corresponding to the current
subtree admits a robust MIS or not and labels each subtree according to that.
If it does admit an RMIS, it may impose some constraints about the membership of
the higher nodes. For instance, a robust MIS of the subtree may exist only if the articulation point leading up to the parent belongs to it.
Then, the goal of the labeling phase consists of propagating (and memorizing within labels) such constraints up through subtle interactions among the
various types of vertices (namely, pendant nodes, articulation point, bridge, or component) leading up to the root.
Besides, the inner topological configuration of a single component may also impose non-trivial
constraints for the existence of a robust MIS. Intuitively, it must have properties that relate to (but are slightly more complex than)
bipartiteness.
The second (short) phase of the algorithm is called the {\em deciding} phase.
It simply consists of deciding whether the graph admits a robust MIS or not considering
the label of the root of the \ensuremath{\mathcal{ABC}}-tree\xspace.
\noindent\textbf{Labeling Phase.}
As already mentioned, a given subtree $\hat{T}$ of {\ensuremath{T_\mathcal{C}}}\xspace may or may not admit an RMIS.
Intuitively, the global decision depends on topological contraints, established over $\hat{T}$. Obviously, those constraints
influence the possible topological organization of a global RMIS toward the parent of $\hat{T}$. So, this mecanism
involves in a crucial way the unique $x \in \hat{T}$ such that $x \in \ensuremath{\mathcal{A}}\xspace \cup \ensuremath{\mathcal{P}}\xspace$ through which $\hat{T}$
is connected to the remainder of {\ensuremath{T_\mathcal{C}}}\xspace. In the following $x$ is called
the {\em attachment point} of $\hat{T}$. In other words and more conveniently, the attachment point of the root $r$ of $\hat{T}$ is
either $r$ itself if $r \in \ensuremath{\mathcal{A}}\xspace \cup \ensuremath{\mathcal{P}}\xspace$ or the parent of $r$ if $r \in \ensuremath{\mathcal{B}}\xspace \cup \ensuremath{\mathcal{C}}\xspace$.
For instance, in
Figure~\ref{fig:ex_compo}, assuming that \ensuremath{\mathcal{R_{\CT}}}\xspace is the component vertex {\tt M}, then, $4$ is the
attachment point of itself ($4 \in \ensuremath{\mathcal{P}}\xspace$), $8$ is the one of both $\{8,14\}$ and $\{8,21\}$, while $10$ is the
attachment point of $\{10,28\}$ ($\in \ensuremath{\mathcal{B}}\xspace$), $M$ ($\in \ensuremath{\mathcal{C}}\xspace$), and itself ($\in \ensuremath{\mathcal{A}}\xspace)$.
Constraint transmission takes place from the leaves to the root by tagging
each vertices with the following labels: \fbox{\footnotesize{\textsl{PI}}}\xspace for Possibly In\xspace (meaning that $\hat{T}$ admits an RMIS that includes $x$, the attachment point of $\hat{T}$); \fbox{\footnotesize{\textsl{PO}}}\xspace for Possibly Out\xspace (meaning that $\hat{T}$ admits an RMIS that does not include $x$); \fbox{\footnotesize{\textsl{PE}}}\xspace for Possibly External\xspace (meaning that $\hat{T}$ is not tagged \fbox{\footnotesize{\textsl{PO}}}\xspace, and it admits an RMIS that does not include $x$ assuming that another node $x'$, external to $\hat{T}$
belongs to the RMIS); and \fbox{\footnotesize{\textsl{N}}}\xspace for Negative\xspace (meaning that none of the three other tags is applicable to $\hat{T}$). An extra label, \fbox{\footnotesize{\textsl{E}}}\xspace is used at the root (see below). Note that the algorithm associates to each label a set of vertices that is used to store a robust MIS of $\hat{T}$ satisfying the constraint of the label. Also, remark that \fbox{\footnotesize{\textsl{PO}}}\xspace and \fbox{\footnotesize{\textsl{PE}}}\xspace are mutually exclusive.
Furthermore, a vertex can be tagged with more than a single tag, namely either \fbox{\footnotesize{\textsl{PI}}}\xspace and \fbox{\footnotesize{\textsl{PO}}}\xspace
(together), or \fbox{\footnotesize{\textsl{PI}}}\xspace and \fbox{\footnotesize{\textsl{PE}}}\xspace (together).
The analysis consists of recursively tagging each vertex $v \in {\ensuremath{T_\mathcal{C}}}\xspace$ from the leaves to the root. Let us first
consider $v$ as a leaf. There are two cases: either $v \in \ensuremath{\mathcal{P}}\xspace$ or $v \in \ensuremath{\mathcal{C}}\xspace$. In the former case,
$v$ is tagged with both \fbox{\footnotesize{\textsl{PI}}}\xspace and \fbox{\footnotesize{\textsl{PE}}}\xspace. (Indeed, $v$ being a pendant node, $v$ can or cannot be in the RMIS
depending or its unique neighbor.) For instance, in
Figure~\ref{fig:ex_compo}, assuming that $\ensuremath{\mathcal{R_{\CT}}}\xspace = M$, the vertices
$4$, $5$, and $20$ are tagged \fbox{\footnotesize{\textsl{PI}}}\xspace and \fbox{\footnotesize{\textsl{PE}}}\xspace.
If $v \in \ensuremath{\mathcal{C}}\xspace$, then the algorithm checks whether $v$ must be tagged \fbox{\footnotesize{\textsl{N}}}\xspace, \fbox{\footnotesize{\textsl{PI}}}\xspace, \fbox{\footnotesize{\textsl{PO}}}\xspace, or \fbox{\footnotesize{\textsl{PE}}}\xspace. For instance, in
the same example, $N$ is tagged \fbox{\footnotesize{\textsl{PI}}}\xspace and \fbox{\footnotesize{\textsl{PO}}}\xspace. Indeed, $N$ is a square (sub)graph (refer to Figure~\ref{fig:ex_compo}) and
its attachment point is $28$. Clearly, it admits two possible RMISs: either $\{31,29\}$ or $\{28,30\}$. In former
case, $28$ does not belong to the RMIS (implying the label of type \fbox{\footnotesize{\textsl{PO}}}\xspace); in the latter, $28$ belongs to the RMIS (implying the label of type \fbox{\footnotesize{\textsl{PI}}}\xspace).
The actual finding is solved through reduction to the 2-SAT problem described below.
From now on, consider that $v$ is an internal vertex ({i.e.}\xspace $v \notin \ensuremath{\mathcal{P}}\xspace$).
Provided that none of its descendants is tagged \fbox{\footnotesize{\textsl{N}}}\xspace, an internal vertex $v \in {\ensuremath{T_\mathcal{C}}}\xspace$ is analyzed according of its
type (whenever a vertex has two tags, each corresponding rule is applied) as follows.
Consider first the case where $v \in \ensuremath{\mathcal{B}}\xspace$. If its (unique) descendant $u$ is tagged \fbox{\footnotesize{\textsl{PI}}}\xspace,
then $v$ is tagged
\fbox{\footnotesize{\textsl{PO}}}\xspace; if $u$ is tagged \fbox{\footnotesize{\textsl{PE}}}\xspace, then $v$ is tagged \fbox{\footnotesize{\textsl{PI}}}\xspace; if $u$ is tagged \fbox{\footnotesize{\textsl{PO}}}\xspace, then
$v$ is tagged \fbox{\footnotesize{\textsl{PI}}}\xspace, and if possible ($v$ is not already tagged \fbox{\footnotesize{\textsl{PO}}}\xspace), also \fbox{\footnotesize{\textsl{PE}}}\xspace.
In the example (Figure~\ref{fig:ex_compo}), the vertices $\{3,4\}$, $\{3,5\}$, and $\{18,20\}$ are
all tagged \fbox{\footnotesize{\textsl{PI}}}\xspace and \fbox{\footnotesize{\textsl{PO}}}\xspace.
When $v \in \ensuremath{\mathcal{A}}\xspace$, let $D_v$ be the set of descendant vertices of $v$. If every vertex $u \in D_v$ is
tagged \fbox{\footnotesize{\textsl{PI}}}\xspace, then $v$ is also tagged \fbox{\footnotesize{\textsl{PI}}}\xspace; if every vertex $u \in D_v$ is
tagged \fbox{\footnotesize{\textsl{PE}}}\xspace, then $v$ is also tagged \fbox{\footnotesize{\textsl{PE}}}\xspace; if every vertex $u \in D_v$ is
tagged either \fbox{\footnotesize{\textsl{PO}}}\xspace or \fbox{\footnotesize{\textsl{PE}}}\xspace and there exists
$u' \in D_v$ tagged \fbox{\footnotesize{\textsl{PO}}}\xspace, then $v$ is tagged \fbox{\footnotesize{\textsl{PO}}}\xspace.
For instance, vertices $3$, $11$, and $18$ are all tagged \fbox{\footnotesize{\textsl{PI}}}\xspace and \fbox{\footnotesize{\textsl{PO}}}\xspace.
If $v \in \ensuremath{\mathcal{C}}\xspace\setminus\{\ensuremath{\mathcal{R_{\CT}}}\xspace\}$, then, as for the leaves, $v$ is analysed using the 2-SAT reduction described below.
However, by contrast with a leaf vertex, the analysis introduces extra
clauses to the 2-SAT expression, according to the labels of its descendants.
If $v = \ensuremath{\mathcal{R_{\CT}}}\xspace$ ($\in \ensuremath{\mathcal{C}}\xspace$), then the algorithm operates the last 2-SAT reduction that checks
the existence of an RMIS, again accoring to the labels of its descendants. \ensuremath{\mathcal{R_{\CT}}}\xspace is then tagged
either with \fbox{\footnotesize{\textsl{N}}}\xspace ($G$ admits no RMIS) or with \fbox{\footnotesize{\textsl{E}}}\xspace ($G$ admits at least one RMIS).
\noindent\textbf{Testing a component vertex.}
The finding process is mainly based on the resolution of a 2-SAT expression.
Let us first consider an internal $C$ component---leaf and root components are special cases that are addressed later.
The procedure first consider $C'$ that is equal to $C$ to which edges linking vertices both tagged \fbox{\footnotesize{\textsl{PO}}}\xspace are removed. Doing so,
$C'$ may be split into several components. If $C'$ is not bipartite, then the component is tagged with \fbox{\footnotesize{\textsl{N}}}\xspace. Otherwise, for each maximal connected component $C_i$, one part of the bipartition is arbitrarily
chosen in which each vertex $v$ ($\in C'$) is labelled with a label $\ell(v)$ equal to $x_i$.
The vertices of the other part are labelled $\neg x_i$.
All those labels form a 2-SAT expression to which the tags coming up from the subtrees are included.
For instance, a node $v \in \ensuremath{\mathcal{A}}\xspace$ that is tagged
\fbox{\footnotesize{\textsl{PI}}}\xspace forces the label $\ell(v)$ of the corresponding vertex $v \in C_i$ to {\tt true}. Also, the edge $\{u,v\}$ that was
removed from $C$ to $C'$ also forces the labels corresponding to $u$ and $v$ to be mutually exclusive
($\neg\ell(u)\vee\neg\ell(v)$), meaning that at most one of the two can be included into the RMIS, but not both.
Since a vertex of {\ensuremath{T_\mathcal{C}}}\xspace can be tagged with one or two tags, the satisfiability of the 2-SAT expression is evaluated
assuming first that the attachment point $x$ of $C$ belongs to the RMIS (Tag~\fbox{\footnotesize{\textsl{PI}}}\xspace, $\ell(x) = \mathtt{true}$).
Next, it is evaluated assuming that the attachment point $x$ of $C$ does not belong to the RMIS
(Tag~\fbox{\footnotesize{\textsl{PO}}}\xspace, $\ell(x) = \mathtt{false}$).
If $C$ could not be tagged \fbox{\footnotesize{\textsl{PO}}}\xspace ({i.e.}\xspace the expression could not be satisfied assuming $\ell(x) = \mathtt{false}$),
then it can still be tagged \fbox{\footnotesize{\textsl{PE}}}\xspace. This is done
by temporarily adding an aerial $\mathsf{g}$ at the attachment point $x$ of $C$ ({i.e.}\xspace a virtual extra vertex $y$ with
the corresponding edge $\{x,y\}$) and repeating the whole above process with $C \cup \mathsf{g}$.
Note this process is also performed at the leaves and at the root ($C = \ensuremath{\mathcal{R_{\CT}}}\xspace$). However, in both cases, the process is
simpler. Indeed, since the leaves have no descendant, it is sufficient to check whether $C$ is bipartite or not. For
the root (that has no attachment point), it is sufficient to check whether the 2-SAT expression is satisfied or not.
\noindent\textbf{Deciding phase.}
This phase simply consists in testing the label of \ensuremath{\mathcal{R_{\CT}}}\xspace (attributed
in the labeling phase). If \ensuremath{\mathcal{R_{\CT}}}\xspace is labeled with \fbox{\footnotesize{\textsl{N}}}\xspace, the algorithm rejects. Otherwise (\ensuremath{\mathcal{R_{\CT}}}\xspace
is labeled with \fbox{\footnotesize{\textsl{E}}}\xspace), the algorithm returns the set associated to the label of \ensuremath{\mathcal{R_{\CT}}}\xspace that
is a robust MIS of $G$ thanks to the work done in the labeling phase.
\section{Conclusion}
\label{sec:conclusion}
This paper is dedicated to showing the actual impact of robustness in highly dynamic distributed systems.
A property is robust if and only if it is satisfied in every connected spanning subgraphs of a given graph.
Focusing on the minimal independent set (MIS) problem, we proved the existence of a significant complexity gap between graphs where \emph{all} MIS are robust (building a robust MIS is then a \emph{local} problem) and graphs where \emph{some} MIS are robust (building a robust MIS is then a \emph{global} problem).
We are convinced that robustness is a key property of highly dynamic systems
to achieving stable structures in such unstable environments.
The complete characterization of the class \ensuremath{{\cal RMIS^\exists}}\xspace is left open, as well as the study of similar symmetry breaking tasks.
\newpage
\pagenumbering{roman}
\setcounter{page}{1}
\input{arxiv.bbl}
\newpage
|
\section{Introduction}
In the early 1970s Joan Birman and Hugh Hilden wrote a series of now-classic papers on the interplay between mapping class groups and covering spaces. The initial goal was to determine a presentation for the mapping class group of $S_2$, the closed surface of genus two (it was not until the late 1970s that Hatcher and Thurston \cite{HT} developed an approach for finding explicit presentations for mapping class groups).
The key innovation by Birman and Hilden is to relate the mapping class group $\Mod(S_2)$ to the mapping class group of $S_{0,6}$, a sphere with six marked points. Presentations for $\Mod(S_{0,6})$ were already known since that group is closely related to a braid group.
The two surfaces $S_2$ and $S_{0,6}$ are related by a two-fold branched covering map $S_2 \to S_{0,6}$:
\begin{center}
\includegraphics[scale=.65]{hi}
\end{center}
The six marked points in the base are branch points. The deck transformation is called the \emph{hyperelliptic involution} of $S_2$, and we denote it by $\iota$. Every element of $\Mod(S_2)$ has a representative that commutes with $\iota$, and so it follows that there is a map
\[
\Theta : \Mod(S_2) \to \Mod(S_{0,6}).
\]
The kernel of $\Theta$ is the cyclic group of order two generated by (the homotopy class of) the involution $\iota$. One can check that each generator for $\Mod(S_{0,6})$ lifts to $\Mod(S_2)$ and so $\Theta$ is surjective. From this we have a short exact sequence
\[ 1 \to \langle \iota \rangle \to \Mod(S_2) \stackrel{\Theta}{\to} \Mod(S_{0,6}) \to 1,\]
and hence a presentation for $\Mod(S_{0,6})$ can be lifted to a presentation for $\Mod(S_2)$.
But wait---the map $\Theta$ is not a priori well defined! The problem is that elements of $\Mod(S_2)$ are only defined up to isotopy, and these isotopies are not required to respect the hyperelliptic involution. The first paper by Birman and Hilden proves that in fact all isotopies can be chosen to respect the involution. Birman and Hilden quickly realized that the theory initiated in that first paper can be generalized in various ways, and they wrote a series of paper on the subject, culminating in the paper \emph{On Isotopies of Homeomorphisms of Riemann Surfaces} \cite{BH73}, published in \emph{Annals of Mathematics} in 1973.
In the remainder of this article, we will discuss the history of the Birman--Hilden theory, including generalizations by MacLachlan--Harvey and the second author of this article, we will give several applications, explain three proofs, and discuss various open questions and new directions in the theory. As we will see, the Birman--Hilden theory has had influence on many areas of mathematics, from low-dimensional topology, to geometric group theory, to representation theory, to algebraic geometry and more, and it continues to produce interesting open problems and research directions.
\bigskip
\noindent\emph{The other article by Birman and Hilden.} Before getting on with our main business, we would be remiss not to mention the other paper by Birman and Hilden \cite{BH75}, the 1975 paper \emph{Heegaard splittings of branched coverings of $S^3$}, published in \emph{Transactions of the American Mathematical Society} (there is also the corresponding research announcement, \emph{The homeomorphism problem for $S^3$}, published two years earlier \cite{BH3mfd73}). In this paper, Birman and Hilden discuss the relationship between branched covers and Heegaard splittings of 3-manifolds. Their results cover a lot of territory. For instance:
\begin{itemize}
\item they prove that every closed, orientable 3-manifold of Heegaard genus 2 is a two-fold branched covering space of $S^3$ branched over a 3-bridge knot or link;
\item they give an algorithm for determining if a Heegaard splitting of genus two represents $S^3$;
\item they prove that any simply connected two-fold cover of $S^3$ branched over the closure of a braid on three strands is itself $S^3$; and
\item they disprove a conjecture of Haken that among all simply connected 3-manifolds, and among all group presentations for their fundamental groups arising from their Heegaard splittings, the presentations for $\pi_1(S^3)$ have a certain nice property.
\end{itemize}
While this paper has also been influential and well-cited, and in fact relies on their work on surfaces, we will restrict our focus in this article to the work of Birman and Hilden on mapping class groups.
\section{Statements of the main theorem}\label{state}
Let $p: S \to X$ be a covering map of surfaces, possibly branched, possibly with boundary. We say that $f : S \to S$ is \emph{fiber preserving} if for each $x \in X$ there is a $y \in X$ so that $f(p^{-1}(x)) = p^{-1}(y)$; in other words, as the terminology suggests, $f$ takes fibers to fibers.
Given two homotopic fiber-preserving homeomorphisms of $S$, we can ask if they are homotopic through fiber-preserving homeomorphisms. If the answer is yes for all such pairs of homeomorphisms, we say that the covering map $p$ has the \emph{Birman--Hilden property}. An equivalent formulation of the Birman--Hilden property is: whenever a fiber-preserving homeomorphism is homotopic to the identity, it is homotopic to the identity through fiber-preserving homeomorphisms.
We are now ready to state the main theorems of the Birman--Hilden theory. There are several versions, proved over the years by various authors, each generalizing the previous. The first version is the one that appears in the aforementioned 1973 \emph{Annals of Mathematics} paper by Birman and Hilden and also in the accompanying research announcement \emph{Isotopies of Homeomorphisms of Riemann surfaces} \cite{BHannouncement}. Throughout, we will say that a surface is \emph{hyperbolic} if its Euler characteristic is negative.
\begin{theorem}[Birman--Hilden]\label{bh}
Let $p : S \to X$ be a finite-sheeted regular branched covering map where $S$ is a hyperbolic surface. Assume that $p$ is either unbranched or is solvable. Then $p$ has the Birman--Hilden property.
\end{theorem}
If we apply Theorem~\ref{bh} to the branched covering map $S_2 \to S_{0,6}$ described earlier, then it exactly says that the map $\Theta : \Mod(S_2) \to \Mod(S_{0,6})$ is well defined.
It is worthwhile to compare our Theorem~\ref{bh} to what is actually stated by Birman and Hilden. In their paper, they state two theorems, each of which is a special case of Theorem~\ref{bh}. Their Theorem~1 treats the case of regular covers where each deck transformation fixes each preimage of each branch point in $X$. This clearly takes care of the case of unbranched covers, and also the case of certain solvable branched covers (on one hand a finite group of homeomorphisms of a surface that fixes a point must be a subgroup of a dihedral group, and on the other hand there are solvable---even cyclic---branched covers that do not satisfy the condition of Theorem~1). Birman and Hilden's Theorem~2 deals with the general case of solvable covers, which includes some unbranched covers.
In early 1973 MacLachlan and Harvey \cite{MH} published a paper called \emph{On Mapping Class Groups and Covering Spaces}, in which they give the following generalization of Theorem~\ref{bh}.
\begin{theorem}[MacLachlan--Harvey]\label{mh}
Let $p : S \to X$ be a finite-sheeted regular branched covering map where $S$ is a hyperbolic surface. Then $p$ has the Birman--Hilden property.
\end{theorem}
MacLachlan and Harvey's work was contemporaneous with the work of Birman and Hilden cited in Theorem~\ref{bh}, and was subsequent to the original paper by Birman and Hilden on the hyperelliptic case. Their approach is completely different, and is framed in terms of Teichm\"uller theory.
The 2014 Ph.D. thesis of the second author of this article is a further generalization \cite{winarski}. For the statement, a preimage of a branch point is \emph{unramified} if some small neighborhood is mapped injectively under the covering map, and a cover is \emph{fully ramified} if no branch point has an unramified preimage.
\begin{theorem}[Winarski]\label{win}
Let $p: S \to X$ be a finite-sheeted branched covering map where $S$ is a hyperbolic surface, and suppose that $p$ is fully ramified. Then $p$ has the Birman--Hilden property.
\end{theorem}
Note that all regular covers are fully ramified and also that all unbranched covers are fully ramified. Thus Theorem~\ref{win} indeed implies Theorems~\ref{bh} and~\ref{mh}. In Section 2.3 of her paper, Winarski gives a general construction of irregular branched covers that are fully ramified. Thus there are many examples of covering spaces that satisfy the hypotheses of Theorem~\ref{win} but not those of Theorem~\ref{mh}.
We will briefly remark on the assumption that $S$ is hyperbolic. It is not hard to construct counterexamples in the other cases. For instance suppose $S$ is the torus $T^2$ and $p : S \to X$ is the branched cover corresponding to the hyperelliptic involution of $T^2$. In this case $X$ is the sphere with four marked points. Rotation of $T^2$ by $\pi$ in one factor is a fiber-preserving homeomorphism homotopic to the identity, but the induced homeomorphism of $X$ acts nontrivially on the marked points and hence is not homotopic to the identity. Thus this cover fails the Birman--Hilden property. One can construct a similar example when $S$ is the sphere $S^2$ and $p : S^2 \to X$ is the branched cover induced by a finite order rotation.
\section{Restatement of the main theorem}
We will now give an interpretation of the Birman--Hilden property---hence all three theorems above---in terms of mapping class groups. Here, the \emph{mapping class group} of a surface is the group of homotopy classes of orientation-preserving homeomorphisms that fix the boundary pointwise and preserve the set of marked points (homotopies must also fix the boundary and preserve the set of marked points).
Let $p: S \to X$ be a covering map of surfaces, possibly branched. We treat each branch point in $X$ as a marked point, and so homeomorphisms of $X$ are assumed to preserve the set of branch points. Let $\LMod(X)$ denote the subgroup of the mapping class group $\Mod(X)$ consisting of all elements that have representatives that lift to homeomorphisms of $X$. This group is called the \emph{liftable mapping class group} of $X$.
Let $\SMod(S)$ denote the subgroup of $\Mod(S)$ consisting of the homotopy classes of all fiber-preserving homeomorphisms. Here we emphasize that two homeomorphisms of $S$ are identified in $\SMod(S)$ if they differ by an isotopy that is not necessarily fiber preserving (so that we have a subgroup of $\Mod(S)$). We also emphasize that preimages of branch points are not marked. Fiber-preserving homeomorphisms are also called \emph{symmetric homeomorphisms}; these are exactly the lifts of liftable homeomorphisms of $X$. The group $\SMod(S)$ is called the \emph{symmetric mapping class group} of $S$.
Let $D$ denote the subgroup of $\SMod(S)$ consisting of the homotopy classes of the deck transformations (it is a fact that nontrivial deck transformations represent nontrivial mapping classes).
\begin{proposition}\label{mcg}
Let $p : S \to X$ be a finite-sheeted branched covering map where $S$ is a hyperbolic surface without boundary. Then the following are equivalent:
\begin{itemize}
\item $p$ has the Birman--Hilden property,
\item the natural map $\LMod(X) \to \SMod(S)/D$ is injective,
\item the natural map $\SMod(S) \to \LMod(X)$ is well defined, and
\item $\SMod(S)/D \cong \LMod(X)$.
\end{itemize}
\end{proposition}
The proposition is straightforward to prove. The main content is the equivalence of the first two statements. The other statements, while useful in practice, are equivalent by rudimentary abstract algebra. Using the proposition, one obtains several restatements of Theorems~\ref{bh}, \ref{mh}, and~\ref{win} in terms of mapping class groups.
Birman and Hilden also proved that for a regular cover $\SMod(S)$ is the normalizer in $\Mod(S)$ of the deck group $D$ (regarded as a subgroup of $\Mod(S)$), and so we can also write the last statement in Proposition~\ref{mcg} as
\[
N_{\Mod(S)}(D) / D \cong \LMod(X).
\]
Birman and Hilden only stated the result about normalizers in the case where the deck group is cyclic. However, by combining their argument with Kerckhoff's resolution of the Nielsen realization problem \cite{kerckhoff} one obtains the more general version.
There is also a version of Proposition~\ref{mcg} for surfaces with boundary. Since the mapping class group of a surface with boundary is torsion free, the deck transformations do not represent elements of $\Mod(S_g)$. And so in this case we can simply replace $D$ with the trivial group. For example, in the presence of boundary the Birman--Hilden property is equivalent to the statement that $\SMod(S) \cong \LMod(X)$. This will become especially important in the discussion of braid groups below.
\section{Application to presentations of mapping class groups}
The original work on the Birman--Hilden theory concerns the case of the hyperelliptic involution and is reported in the 1971 paper \emph{On the mapping class groups of closed surfaces as covering spaces} \cite{BH71}. We will explain how Theorem~\ref{bh} specializes in this case and helps to give presentations for the associated symmetric mapping class group and the full mapping class group in genus two.
Consider the covering space $S_g \to S_{0,2g+2}$ induced by a hyperelliptic involution of $S_g$. In general a \emph{hyperelliptic involution} of $S_g$ is a homeomorphism of order two that acts by $-I$ on $H_1(S_g;\Z)$; we remark that the hyperelliptic involution is unique for $S_1$ and $S_2$ but there are infinitely many hyperelliptic involutions of $S_g$ when $g \geq 3$, all conjugate in the group of homeomorphisms.
Theorem~\ref{bh} and Proposition~\ref{mcg} give an isomorphism
\[
\SMod(S_g) / \langle \iota \rangle \cong \LMod(S_{0,2g+2}).
\]
In the special case of the hyperelliptic involution we have $\LMod(S_{0,2g+2}) = \Mod(S_{0,2g+2})$. Indeed, we can check directly that each half-twist generator for $\Mod(S_{0,2g+2})$ lifts to a Dehn twist in $S_g$.
In the case $g=2$ we further have
\[
\SMod(S_2) = \Mod(S_2).
\]
In other words, every mapping class of $S_2$ is symmetric with respect to the hyperelliptic involution. The easiest way to see this is to note that each of the Humphries generators for $\Mod(S_2)$ is a Dehn twist about a curve that is preserved by the hyperelliptic involution. We thus have the following isomorphism:
\[ \Mod(S_2) / \langle \iota \rangle \cong \Mod(S_{0,6}). \]
Simple presentations for $\Mod(S_{0,n})$ were found by Magnus, and so from his presentation for $\Mod(S_{0,6})$ Birman and Hilden use the above isomorphism to derive the following presentation for $\Mod(S_2)$. The generators are the Humphries generators for $\Mod(S_2)$, and we denote them by $T_1,\dots, T_5$. The relations are: \begin{align*}
[T_i,T_j]&=1 \hspace*{.25in}\text{ for } |i-j|>2 \\
T_iT_{i+1}T_i=T_{i+1}T_i&T_{i+1} \hspace*{.25in}\text{ for } 1 \leq i \leq 4 \\
(T_1T_2T_3T_4T_5)^6&=1 \\
(T_1T_2T_3T_4T_5T_5T_4T_3T_2T_1)^2&=1 \\
\left[T_1T_2T_3T_4T_5T_5T_4T_3T_2T_1,T_1\right]&=1
\end{align*}
The first two relations are the standard braid relations from $B_6$, the next relation describes the kernel of the map $B_6 \to \Mod(S_{0,6})$, and the last two relations come from the two-fold cover: the mapping class
\[
T_1T_2T_3T_4T_5T_5T_4T_3T_2T_1
\]
is the hyperelliptic involution. This presentation is the culmination of a program begun by Bergau and Mennicke \cite{BG}, who approached the problem by studying the surjective homomorphism $B_6 \to \Mod(S_2)$ that factors through the map $\Mod(S_{0,6}) \to \Mod(S_2)$ used here.
Birman used the above presentation to give a normal form for elements of $\Mod(S_2)$ and hence a method for enumerating 3-manifolds of Heegaard genus two \cite{birman}.
As explained by Birman and Hilden, the given presentation for $\Mod(S_2)$ generalizes to a presentation for $\SMod(S_g)$. The latter presentation has many applications to the study of $\SMod(S_g)$. It was used by Meyer \cite{meyer72} to show that if a surface bundle over a surface has monodromy in $\SMod(S_g)$ then the signature of the resulting 4-manifold is zero; see also the related work of Endo \cite{endo}. Endo and Kotschick used the Birman--Hilden presentation to show that the second bounded cohomology of $\SMod(S_g)$ is nontrivial \cite{EK01}. Also, Kawazumi \cite{kawazumi} used it to understand the low-dimensional cohomology of $\SMod(S_g)$.
In 1972 Birman and Chillingworth published the paper \emph{On the homeotopy group of a non-orientable surface} \cite{BC}. There, they determine a generating set for the mapping class group (= homeotopy group) of an arbitrary closed non-orientable surface using similar ideas, namely, they exploit the associated orientation double cover and pass information through the Birman--Hilden theorem from the orientable case. They also find an explicit finite presentation for the mapping class group of a closed non-orientable surface of genus three, which admits a degree two cover by $S_2$.
One other observation from the 1971 paper is that $\Mod(S_2)$ is both a quotient of and a subgroup of $\Mod(S_{2,6})$. To realize $\Mod(S_2)$ as a quotient, we consider the map $\Mod(S_{2,6}) \to \Mod(S_2)$ obtained by forgetting the marked points/punctures; this is a special case of the Birman exact sequence studied by Birman in her thesis \cite{birmanthesis}. And to realize $\Mod(S_2)$ as a subgroup, we use the Birman--Hilden theorem: since every element of $\Mod(S_2)$ has a symmetric representative that preserves the set of preimages of the branch points in $S_{0,6}$ and since isotopies between symmetric homeomorphisms can also be chosen to preserve this set of six points, we obtain the desired inclusion. Birman and Hilden state that ``the former property is easily understood but the latter much more subtle.'' As mentioned by Mess \cite{mess}, the inclusion $\Mod(S_2) \to \Mod(S_{2,6})$ can be rephrased as describing a multi-section of the universal bundle over moduli space in genus two.
\section{More applications to the genus two mapping class group}
In the previous section we saw how the Birman--Hilden theory allows us to transport knowledge about the mapping class group of a punctured sphere to the mapping class group of a surface of genus two. As the former are closely related to braid groups, we can often push results about braid groups to the mapping class group. Almost every result about mapping class groups that is special to genus two is proved in this way.
A prime example of this is the result of Bigelow--Budney \cite{BB} and Korkmaz \cite{KL} which states that $\Mod(S_2)$ is linear, that is, $\Mod(S_2)$ admits a faithful representation into $\GL_N(\C)$ for some $N$. Bigelow and Krammer independently proved that braid groups were linear, and so the main work is to derive from this the linearity of $\Mod(S_{0,n})$. They then use the isomorphism $\Mod(S_2) / \langle \iota \rangle \cong \Mod(S_{0,6})$ to push the linearity up to $\Mod(S_2)$.
A second example is from the thesis of Whittlesey, published in 2000. She showed that $\Mod(S_2)$ contains a normal subgroup where every nontrivial element is pseudo-Anosov \cite{whittlesey}. The starting point is to consider the Brunnian subgroup of $\Mod(S_{0,6})$. This is the intersection of the kernels of the six forgetful maps $\Mod(S_{0,6}) \to \Mod(S_{0,5})$, so it is obviously normal in $\Mod(S_{0,6})$. She shows that all nontrivial elements of this group are pseudo-Anosov and proves that the preimage in $\Mod(S_2)$ has a finite-index subgroup with the desired properties.
We give one more example. In the 1980s, before the work of Bigelow and Krammer, Vaughan Jones discovered a representation of the braid group defined in terms of Hecke algebras \cite{jones}. As in the work of Bigelow--Budney and Korkmaz, one can then derive a representation of $\Mod(S_{0,2g+2})$ and then---using the Birman--Hilden theory---of $\SMod(S_g)$. When $g=2$ we thus obtain a representation of $\Mod(S_2)$ to $\GL_5(\Z[t,t^{-1}])$. This representation was used by Humphries \cite{humphries} to show that the normal closure in $\Mod(S_2)$ of the $k$th power of a Dehn twist about a nonseparating curve has finite index if and only if $|k| \leq 3$.
There are many other examples, such as the computation of the asymptotic dimension of $\Mod(S_2)$ by Bell and Fujiwara \cite{BF} and the determination of the minimal dilatation in $\Mod(S_2)$ by Cho and Ham \cite{CH}; the list goes on, but so must we.
\section{Application to representations of the braid group}\label{rep}
The Birman--Hilden theorem gives an important relationship between braid groups and mapping class groups. This is probably the most oft-used application of their results.
Let $S_g^1$ the orientable surface of genus $g$ with one boundary component and let $D_{2g+1}$ denote the closed disk with $2g+1$ marked points in the interior. Consider the covering space $S_g^1 \to D_{2g+1}$ induced by a hyperelliptic involution of $S_g^1$. It is well known that $\Mod(D_{2g+1})$ is isomorphic to the braid group $B_{2g+1}$. As in the closed case, it is not hard to see that $\LMod(D_{2g+1}) = \Mod(D_{2g+1})$ (again, each of the standard generators for $B_{2g+1}$ lifts to a Dehn twist).
One is thus tempted to conclude that $\SMod(S_g^1) / \langle \iota \rangle \cong B_{2g+1}$. But this is not the right statement, since $\iota$ does not represent an element of $\Mod(S_g^1)$. Indeed, for surfaces with boundary we insist that homeomorphisms and homotopies fix the boundary pointwise (otherwise we would not have the isomorphism $\Mod(D_{2g+1}) \cong B_{2g+1}$). Therefore, the correct isomorphism is:
\[ \SMod(S_g^1) \cong B_{2g+1} \]
The most salient aspect of this isomorphism is that there is an injective homomorphism
\[ B_{2g+1} \to \Mod(S_g^1). \]
The injectivity here is sometimes attributed to Perron--Vannier \cite{PV}. It is possible that they were the first to observe this consequence of the Birman--Hilden theorem but the only nontrivial step is the Birman--Hilden theorem.
In the case of $g=1$ the representation of $B_3$ is onto $\Mod(S_1^1)$, and so
\[ \Mod(S_1^1) \cong B_3. \]
Similarly we have
\[ \Mod(S_1^2) \cong B_4 \times \Z. \]
The point here is that $B_4$ surjects onto $\SMod(S_1^2)$ and the latter is almost isomorphic to $\Mod(S_1^2)$; the extra $\Z$ comes from the Dehn twist about a single boundary component.
One reason that the embedding of $B_{2g+1}$ in $\Mod(S_g^1)$ is so important is that if we compose with the standard symplectic representation
\[ \Mod(S_g^1) \to \Sp_{2g}(\Z) \]
then we obtain a representation of the braid group
\[ B_{2g+1} \to \Sp_{2g}(\Z) .\]
This representation is called the standard symplectic representation of the braid group. It is also called the \emph{integral Burau representation} because it is the only integral specialization of the Burau representation besides the permutation representation. The symplectic representation is obtained by specializing the Burau representation at $t=-1$, while the permutation representation is obtained by taking $t=1$.
The image of the integral Burau representation has finite index in the symplectic group: it is an extension of the level two symplectic group by the symmetric group on $2g+1$ letters. The projection onto the symmetric group factor is the standard symmetric group representation of the braid group. See A'Campo's paper \cite{acampo} for details.
The kernel of the integral Burau representation is known as the hyperelliptic Torelli group. This group is well studied, as it describes the fundamental group of the branch locus of the period mapping from Teichm\"uller space to the Siegel upper half-space; see, for instance, the paper by Brendle, Putman, and the first author of this article \cite{BMP} and the references therein.
There are plenty of variations on the given representation. Most important is that if we take a surface with two boundary components $S_g^2$ and choose a hyperelliptic involution, that is, an order two homeomorphism that acts by $-I$ on the first homology of the surface, then the quotient is $D_{2g+2}$ and so we obtain an isomorphism:
\[ \SMod(S_g^2) \cong B_{2g+2}.\]
Also, since the inclusions $S_g^1 \to S_{g+1}$ and $S_g^2 \to S_{g+1}$ induce injections $\SMod(S_g^1) \to \Mod(S_{g+1})$ and $\SMod(S_g^2) \to \Mod(S_{g+1})$ we obtain embeddings of braid groups into mapping class groups of closed surfaces.
In the 1971 paper Birman and Hilden discuss the connection with representations of the braid group. They point out the related fact that $B_{2g+2}$ surjects onto $\SMod(S_g)$ (this follows immediately from their presentation for the latter). In the special case $g=1$ this becomes the classical fact that $B_4$ surjects onto $\SMod(S_1) = \Mod(S_1) \cong \SL_2(\Z)$. We can also derive this fact from our isomorphism $\Mod(S_1^1) \cong B_3$, the famous surjection $B_4 \to B_3$, and the surjection $\Mod(S_1^1) \to \Mod(S_1)$ obtained by capping the boundary.
One useful application of the embeddings of braid groups in mapping class groups is that we can often transport relations from the former to the latter. In fact, almost all of the widely-used relations in the mapping class group have interpretations in terms of braids. This is especially true in the theory of Lefschetz fibrations; see for instance the work of Korkmaz \cite{korkmaz01} and Hamada \cite{hamada} and of Baykur and Van Horn-Morris \cite{BV}.
\section{Application to a question of Magnus}
The last application we will explain is beautiful and unexpected. It is the resolution of a seemingly unrelated question of Magnus about braid groups.
As mentioned in the previous section, the braid group $B_n$ is isomorphic to the mapping class group of a disk $D_n$ with $n$ marked points. Let us write $D_n^\circ$ for the surface obtained by removing from $D_n$ the marked points. There is then a natural action of $B_n$ on $\pi_1(D_n^\circ)$ (with base point on the boundary). The latter is isomorphic to the free group $F_n$ on $n$ letters. Basic algebraic topology tells us that this action is faithful. In other words, we have an injective homomorphism:
\[
B_n \to \Aut(F_n).
\]
This is a fruitful way to view the braid group; for instance, since the word problem in $\Aut(F_n)$ is easily solvable, this gives a solution to the word problem for $B_n$.
Let $F_{n,k}$ denote the normal closure in $F_n$ of the elements $x_1^k,\dots,x_n^k$. The quotient $F_n/F_{n,k}$ is isomorphic to the $n$-fold free product $\Z/k\Z \ast \cdots \ast \Z/k\Z$. Since the elements of $B_n$ preserve the set of conjugacy classes $\{[x_1],\dots,[x_n]\}$, there is an induced homomorphism
\[
B_n \to \Aut(F_n/F_{n,k}).
\]
Let $B_{n,k}$ denote the image of $B_n$ under this map. Magnus asked:
\begin{quote}
\emph{Is $B_n$ isomorphic to $B_{n,k}$?}
\end{quote}
In other words, is the map $B_n \to \Aut(F_n/F_{n,k})$ injective?
In their \emph{Annals paper}, Birman and Hilden answer Magnus' question in the affirmative. Here is the idea. Let $H_{n,k}$ denote the kernel of the map
\[ F_n \to \Z/k\Z \]
where each generator of $F_n$ maps to 1. The covering space of $D_n^\circ$ corresponding to $H_{n,k}$ is a $k$-fold cyclic cover $S^\circ$. If we consider a small neighborhood of one of the punctures in $D_n^\circ$, the induced covering map is equivalent to the connected $k$-fold covering space of $\C \setminus \{0\}$ over itself (i.e. the one induced by $z \mapsto z^k$). As such, we can ``plug in'' to $S^\circ$ a total of $n$ points in order to obtain a surface $S$ and a cyclic branched cover $S \to D_n$. The fundamental group of $S^\circ$ is $H_{n,k}$ by definition. It follows from Van Kampen's theorem that $\pi_1(S) \cong H_{n,k}/F_{n,k}$. Indeed, a simple loop around a puncture in $S^\circ$ projects to a loop in $D_n^\circ$ that circles the corresponding puncture $k$ times.
As in the case of the hyperelliptic involution, we can check directly that each element of $B_n$ lifts to a fiber-preserving homeomorphism of $S$. Therefore, to answer Magnus' question in the affirmative it is enough to check that the map $B_n \to \Aut \pi_1(S)$ is injective. Suppose $b \in B_n$ lies in the kernel. Then the corresponding fiber-preserving homeomorphism of $S$ is homotopic, hence isotopic, to the identity. By the Birman--Hilden theorem (the version for surfaces with boundary), $b$ is trivial, and we are done.
Bacardit and Dicks \cite{BD} give a purely algebraic treatment of Magnus' question; they credit the argument to Crisp and Paris \cite{CP}. Another algebraic argument for the case of even $k$ was given by D.L. Johnson \cite{johnson}. Yet another combinatorial proof was given in 1992 by Kr\"uger \cite{kruger}.
\section{A famous (but false) proof of the Birman--Hilden theorem}
When confronted with the Birman--Hilden theorem, one might be tempted to quickly offer the following easy proof: given the branched cover $p : S \to X$, the fiber-preserving homeomorphism $f : S \to S$, the corresponding homeomorphism $\bar f : X \to X$, and an isotopy $H : S \times I \to S$ from $f$ to the identity map, we can consider the composition $p \circ H$, which gives a homotopy from $\bar f$ to the identity. Then, since homotopic homeomorphisms of a surface are isotopic, there is an isotopy from $\bar f$ to the identity, and this isotopy lifts to a fiber-preserving isotopy from $f$ to the identity. Quod erat demonstrandum.
This probably sounds convincing, but there are two problems. First of all, the composition $p \circ H$ is really a homotopy between $p \circ f$ and $p$ which are maps from $S$ to $X$; since $H$ is not fiber preserving, there is no way to convert this to a well-defined homotopy between maps $X \to X$. The second problem is that $H$ might send points that are not preimages of branch points to preimages of branch points; so even if we could project the isotopy, we would not obtain a homotopy of $X$ that respects the marked points.
In the next two sections will will outline proofs of the Birman--Hilden theorem in various cases. The reader should keep in mind the subtleties uncovered by this false proof.
\section{The unbranched (= easy) case}
Before getting to the proof of the Birman--Hilden theorem, we will warm up with the case of unbranched covers. This case is much simpler, as all of the subtlety of the Birman--Hilden theorem lies in the branch points. Still the proof is nontrivial, and later we will prove the more general case by reducing to the unbranched case.
In 1972 Birman and Hilden published the paper \emph{Lifting and projecting homeomorphisms} \cite{BH72}, which gives a quick proof of Theorem~\ref{bh} in the case of regular unbranched covers. Following along the same lines, Aramayona, Leininger, and Souto generalized their proof to the case of arbitrary (possibly irregular) unbranched covers \cite{ALS}. We will now explain their proof.
Let $p : S \to X$ be an unbranched covering space of surfaces, and let $f : S \to S$ be a fiber-preserving homeomorphism that is isotopic to the identity. Without loss of generality, we may assume that $f$ has a fixed point. Indeed, if $f$ does not fix some point $x$, then we can push $p(x)$ in $X$ by an ambient isotopy, and lift this isotopy to $S$ until $x$ is fixed. As a consequence, $f$ induces a well-defined action $f_\star$ on $\pi_1(S)$. Since $f$ is isotopic to the identity, $f_\star$ is the identity. If $\bar f$ is the corresponding homeomorphism of $X$, then it follows that $\bar f_\star$ is the identity on the finite-index subgroup $p_\star(\pi_1(S))$ of $\pi_1(X)$. From this, plus the fact that roots are unique in $\pi_1(X)$, we conclude that $\bar f_\star$ is the identity. By basic algebraic topology, $\bar f$ is homotopic to the identity, and hence it is isotopic to the identity, which implies that $f$ is isotopic to the identity through fiber-preserving homeomorphisms, as desired.
\section{Three (correct) proofs of the Birman--Hilden theorem}
In this section we present sketches of the proofs of all three versions of the Birman--Hilden theorem given in Section~\ref{state}. We begin with the original proof by Birman and Hilden, which is a direct attack using algebraic and geometric topology. Then we explain the proof of MacLachlan and Harvey's Teichm\"uller theoretic approach, and finally the combinatorial topology approach of the second author, which gives a further generalization.
\p{The Birman--Hilden proof: Algebraic and geometric topology} As in the statement of Theorem~\ref{bh}, let $p : S \to X$ be a regular branched covering space where $S$ is a hyperbolic surface. As in Theorem~1 of the \emph{Annals} paper by Birman and Hilden, we make the additional assumption here that each deck transformation for this cover fixes each preimage of each branch point in $X$. Theorem~\ref{bh} will follow easily from this special case. Let $f$ be a fiber-preserving homeomorphism of $S$ and assume that $f$ is isotopic to the identity.
Let $x$ be the preimage in $S$ of some branch point in $X$. The first key claim is that $f(x)=x$ \cite[Lemma 1.3]{BH73}. Thus if we take the isotopy $H$ from $f$ to the identity and restrict it to $x$, we obtain an element $\alpha$ of $\pi_1(S,x)$. Birman and Hilden argue that $\alpha$ must be the trivial element. The idea is to argue that $\alpha$ is fixed by each deck transformation (this makes sense since the deck transformations fix $x$), and then to argue that the only element of $\pi_1(S)$ fixed by a nontrivial deck transformation is the trivial one (to see this, regard $\alpha$ as an isometry of the universal cover $\mathbb{H}^2$ and regard a deck transformation as a rotation of $\mathbb{H}^2$).
Since $\alpha$ is trivial, we can deform it to the trivial loop, and by extension we can deform the isotopy $H$ to another isotopy that fixes $x$ throughout. Proceeding inductively, Birman and Hilden argue that $H$ can be deformed so that it fixes all preimages of branch points throughout the isotopy. At this point, by deleting branch points in $X$ and their preimages in $S$, we reduce to the unbranched case.
Finally, to prove their Theorem~2, which treats the case of solvable covers, Birman and Hilden reduce it to Theorem~1 by factoring any solvable cover into a sequence of cyclic covers of prime order. Such a cover must satisfy the hypotheses of Theorem~1.
It would be interesting to use the Birman--Hilden approach to prove the more general theorem of Winarski. There is a paper by Zieschang from 1973 that uses similar reasoning to Birman and Hilden and recovers the result of MacLachlan and Harvey \cite{zieschang73}.
\p{MacLachlan and Harvey's proof: Teichm\"uller theory} We now explain the approach of MacLachlan and Harvey. Let $p : S \to X$ be a regular branched covering space where $S$ is a hyperbolic surface. We will give MacLachlan and Harvey's argument for Theorem~\ref{mh} and at the same time explain why the argument gives the more general result of Theorem~\ref{win}.
The mapping class group $\Mod(S)$ acts on the Teichm\"uller space $\Teich(S)$, the space of isotopy classes of complex structures on $S$ (or conformal structures on $S$, or hyperbolic structures on $S$, or algebraic structures on $S$). Let $X^\circ$ denote the complement in $X$ of the set of branch points. There is a map $\Xi : \Teich(X^\circ) \to \Teich(S)$ defined by lifting complex structures through the covering map $p$ (one must apply the removable singularity theorem to extend over the preimages of the branch points).
The key point in the proof is that $\Xi$ is injective. One way to see this is to observe that Teichm\"uller geodesics in $\Teich(X^\circ)$ map to Teichm\"uller geodesics in $\Teich(S)$ of the same length. Indeed, the only way this could fail would be if we had a Teichm\"uller geodesic in $\Teich(X^\circ)$ where the corresponding quadratic differential had a simple pole (= 1-pronged singularity) at a branch point and some pre-image of that branch point was unramified (1-pronged singularities are only allowed at marked points, and preimages of branch points are not marked). This is why the most natural setting for this argument is that of Theorem~\ref{win}, namely, where $p$ is fully ramified.
Let $Y$ denote the image of $\Xi$. The symmetric mapping class group $\SMod(S)$ acts on $Y$ and the kernel of this action is nothing other than $D$. The liftable mapping class group $\LMod(X)$ acts faithfully on $\Teich(X^\circ)$ and hence---as $\Xi$ is injective---it also acts faithfully on $Y$. It follows immediately from the definitions that the images of $\SMod(S)$ and $\LMod(X)$ in the group of automorphisms of $Y$ are equal. It follows that $\SMod(S)/D$ is isomorphic to $\LMod(X)$, as desired.
\p{Winarski's proof: Combinatorial topology} Let $p : S \to X$ be a fully ramified branched covering space where $S$ is a hyperbolic surface. To prove Theorem~\ref{win} we will show that $\Phi : \LMod(X) \to \SMod(S)/D$ is injective.
Suppose $f \in \LMod(X)$ lies in the kernel of $\Phi$. Let $\varphi$ be a representative of $f$. Since $\Phi(f)$ is trivial we can choose a lift $\tilde \varphi : S \to S$ that is isotopic to the identity; thus $\tilde \varphi$ fixes the isotopy class of every simple closed curve in $S$. The main claim is that $\varphi$ fixes the isotopy class of every simple closed curve in $X$. From this, it follows that $f$ has finite order in $\LMod(X)$. Since $\ker(\Phi)$ is torsion free \cite[Prop 4.2]{winarski}, the theorem will follow.
So let us set about the claim. Let $c$ be a simple closed curve in $X$, and let $\tilde c$ be its preimage in $S$. By assumption $\tilde \varphi(\tilde c)$ is isotopic to $\tilde c$ and we would like to leverage this to show $\varphi(c)$ is isotopic to $c$. There are two stages to the argument: first dealing with the case where $\varphi(c)$ and $c$ are disjoint, and then in the case where they are not disjoint we reduce to the disjoint case.
If $\varphi(c)$ and $c$ are disjoint, then $\tilde \varphi( \tilde c)$ and $\tilde c$ are disjoint. Since the latter are isotopic, they co-bound a collection of annuli $A_1,\dots,A_n$. Then, since orbifold Euler characteristic is multiplicative under covers, we can conclude that $p(\cup A_i)$ is an annulus with no branch points (branch points decrease the orbifold Euler characteristic), and so $c$ and $\varphi(c)$ are isotopic.
We now deal with the second stage, where $\varphi(c)$ and $c$ are not disjoint. In this case $\tilde \varphi(\tilde c)$ and $\tilde c$ are not disjoint either, but by our assumptions they are isotopic in $S$. Therefore, $\tilde \varphi(\tilde c)$ and $\tilde c$ bound at least one bigon.
Consider an innermost such bigon $B$. Since $B$ is innermost, $p(B)$ is an innermost bigon bounded by $\varphi(c)$ and $c$ in $X$ (the fact that $B$ is innermost implies that $p|B$ is injective). If there were a branch point in $p(B)$ then since $p$ is fully ramified, this would imply that $B$ was a $2k$-gon with $k > 1$, a contradiction. Thus, we can apply an isotopy to remove the bigon $p(B)$ and by induction we reduce to the case where $\varphi(c)$ and $c$ are disjoint.
For an exposition of Winarski's proof in the case of the hyperelliptic involution, see the book by Farb and the first author of this article \cite{primer}.
\begin{center}
\begin{figure}
\includegraphics{fuller}
\caption{The simple threefold cover of $S_g$ over the sphere. The gray curves divide $S_g$ into three regions, each serving as a ``fundamental domain'' for the cover. The preimages in $S_g$ of the branch points in $S^2$ are shown as dots but are treated as unmarked points in $S_g$.}
\label{fuller}
\end{figure}
\end{center}
\section{Open questions and new directions}
One of the most striking aspects of the Birman--Hilden story is the breadth of open problems related to the theory and the constant discovery of related directions. We have already mentioned a number of questions arising from the theory.
Perhaps the most obvious open problem is the following.
\begin{question}
Which branched covers of surfaces have the Birman--Hilden property?
\end{question}
Based on the discussion in Section~\ref{state} above, one might hope that all branched coverings---at least where the cover is a hyperbolic surface---have the Birman--Hilden property. However, this is not true. Consider, for instance, the simple threefold cover $p : S_g \to X$, where $X$ is the sphere with $2g+4$ branch points (this cover is unique). As shown in Figure~\ref{fuller} we can find an essential curve $a$ in $X$ whose preimage in $S_g$ is a union of three homotopically trivial simple closed curves. It follows that the Dehn twist $T_a$ lies in the kernel of the map $\LMod(X) \to \SMod(S_g)$ and so $p$ does not have the Birman--Hilden property. See Fuller's paper for further discussion of this example and the relationship to Lefschetz fibrations \cite{fuller}.
Berstein and Edmonds generalized this example by showing that no simple cover of degree at least three over the sphere has the Birman--Hilden property \cite{BE}, and Winarski further generalized this by proving that no simple cover of degree at least three over any surface has the Birman--Hilden property \cite{winarski}.
Having accepted the fact that not all covers have the Birman--Hilden property, one's second hope might be that a cover has the Birman--Hilden property if and only if it is fully ramified. However, this is also false. Chris Leininger \cite{leininger} has explained to us how to construct a counterexample using the following steps. First, let $S$ be a surface and let $z$ be a marked point in $S$. Let $p : \tilde S \to S$ be a characteristic cover of $S$ and let $\tilde z$ be one point of the full preimage $p^{-1}(z)$. Ivanov and McCarthy \cite{IM} observed that there is an injective homomorphism $\Mod(S,z) \to \Mod(\tilde S,p^{-1}(z))$ where for each element of $\Mod(S,z)$, we choose the lift to $\tilde S$ that fixes $\tilde z$. Aramayona--Leininger--Souto \cite{ALS} proved that the composition of the Ivanov--McCarthy homomorphism with the forgetful map $\Mod(\tilde S,p^{-1}(z)) \to \Mod(\tilde S, \tilde z)$ is injective. If we then take a regular branched cover $S' \to \tilde S$ with branch locus $\tilde z$, the resulting cover $S' \to S$ is not fully ramified but it has the Birman--Hilden property.
Here is another basic question.
\begin{question}\label{smodq}
For which cyclic branched covers of $S_g$ over the sphere is $\SMod(S_g)$ equal to a proper subgroup of $\Mod(S_g)$? When is it finite index?
\end{question}
Theorem 5 in the \emph{Annals} paper by Birman and Hilden states for a cyclic branched cover $S \to X$ over the sphere we have $\LMod(X) =\Mod(X)$. Counterexamples to this theorem were recently discovered by Ghaswala and the second author (see the erratum \cite{erratum}), who wrote a paper \cite{GW} classifying exactly which branched covers over the sphere have $\LMod(X)=\Mod(X)$. Theorem 6 in the paper by Birman and Hilden states that for a cyclic branched cover of $S_g$ over the sphere with $g \geq 3$ the group $\SMod(S_g)$ is a proper subgroup of $\Mod(S_g)$. The proof uses their Theorem 5, so Question~\ref{smodq} should be considered an open question. Of course this question can be generalized to other base surfaces besides the sphere and other types of covers. For simple branched covers over the sphere (which, as above, do not have the Birman--Hilden property) Berstein and Edmonds \cite{BE} proved that $\SMod(S_g)$ is equal to $\Mod(S_g)$.
\medskip
We can also ask about the Birman--Hilden theory for orbifolds and 3-manifolds.
\begin{question}
Which covering spaces of two-dimensional orbifolds have the Birman--Hilden property?
\end{question}
Earle proved some Birman--Hilden-type results for orbifolds in his recent paper \cite{earle}, which he describes as a sequel to his 1971 paper \emph{On the moduli of closed Riemann surfaces with symmetries} \cite{earle71}.
\begin{question}
Which covering spaces of 3-manifolds enjoy the Birman--Hilden property?
\end{question}
Vogt proved that certain regular unbranched covers of certain Seifert-fibered 3-manifolds have the Birman--Hilden property \cite{vogt}. He also explains the connection to understanding foliations in codimension two, specifically for foliations of closed 5-manifolds by Seifert 3-manifolds.
A specific 3-manifold worth investigating is the connect sum of $n$ copies of $S^2 \times S^1$; call it $M_n$. The outer automorphism group of the free group $F_n$ is a finite quotient of the mapping class group of $M_n$. Therefore, one might obtain a version of the Birman--Hilden theory for the outer automorphism group of $F_n$ by developing a Birman--Hilden theory for $M_n$.
\begin{question}
Does $M_n$ enjoy the Birman--Hilden property? If so, does this give a Birman--Hilden theory for free groups?
\end{question}
For example, consider the hyperelliptic involution $\sigma$ of $F_n$, the outer automorphism that (has a representative that) inverts each generator of $F_n$. This automorphism is realized by the homeomorphism of $M_n$ that reverses each $S^1$-factor. The resulting quotient of $M_n$ is the 3-sphere with branch locus the $(n+1)$-component unlink. This is in consonance with the fact that the centralizer of $\sigma$ in the outer automorphism group of $F_n$ is the palindromic subgroup and that the latter is closely related to the configuration space of unlinks in $S^3$; see the paper by Collins \cite{collins}.
\medskip
Next, there are many questions about the hyperelliptic Torelli group and its generalizations. As discussed in Section~\ref{rep}, the hyperelliptic Torelli group is the kernel of the integral Burau representation of the braid group. With Brendle and Putman, the first author of this article proved \cite{BMP} that this group is generated by the squares of Dehn twists about curves that surround an odd number of marked points in the disk $D_n$.
\begin{question}
Is the hyperelliptic Torelli group finitely generated? Is it finitely presented? Does it have finitely generated abelianization?
\end{question}
There are many variants of this question. By changing the branched cover $S \to D_n$, we obtain many other representations of (the liftable subgroups of) the braid group. Each representation gives rise to its own Torelli group. Except for the hyperelliptic involution case, very little is known. One set of covers to consider are the superelliptic covers studied by Ghaswala and the second author of this article \cite{GW17}.
Another aspect of this question is to determine the images of the braid groups in $\Sp_{2N}(\Z)$ under the various representations of (finite index subgroups of) the braid group arising from various covers $S \to D_n$. By work of McMullen \cite{mcmullen} and Venkatarmana \cite{venkataramana}, it is known that when the degree of the cover is at least three and $n$ is more than twice the degree, the image has finite index in the centralizer of the image of the deck group.
\begin{question}
For which covers $S \to D_n$ does the associated representation of the braid group have finite index in the centralizer of the image of the deck group?
\end{question}
There are still many aspects to the Birman--Hilden theory that we have not touched upon. Ellenberg and McReynolds \cite{EM} used the theory to prove that every algebraic curve over $ \mathbb{\bar Q}$ is birationally equivalent over $\C$ to a Teichm\"uller curve. Nikolaev \cite{nikolaev} uses the embedding of the braid group into the mapping class group to give cluster algebraic representations of braid groups. Kordek applies the aforementioned result of Ghaswala and the second author of this article to deduce information about the Picard groups of various moduli spaces of Riemann surfaces \cite{kordek}. A Google search for ``Birman--Hilden'' yields a seemingly endless supply of applications and connections (the \emph{Annals} paper has 139 citations on Google Scholar at the time of this writing). We hope that the reader is inspired to learn more about these connections and pursue their own developments of the theory.
\p{Acknowledgments.} The authors would like to thank John Etnyre, Tyrone Ghaswala, Allen Hatcher, and Chris Leininger for helpful conversations.
\bibliographystyle{plain}
|
\section{Acknowledgement}
I am very grateful to Vít Jelínek for many consultations, valuable advice and proofreading of this work.
\input{bibliography}
\end{document}
\section{Introduction}
Permutations of numbers or other finite sets are a~very deeply and frequently studied
combinatorial and algebraic object. There are two main structures on permutations investigated
in modern mathematics: groups, closed under the composition operator, and hereditary pattern-avoiding classes,
closed under the relation of containment. This paper is one of several texts
exploring the relation between the two notions by applying the composition operator to
permutation classes. That is, given two classes $\mathcal{A}$ and $\mathcal{B}$, we denote by $\mathcal{A} \circ \mathcal{B}$ the
class of all permutations which can be written as a~composition of a~permutation from $\mathcal{A}$
and a~permutation from $\mathcal{B}$.
The oldest results combining permutation classes and groups that we know of are due
to Atkinson and Beals \cite{AtkinsonBeals01}, who consider the permutation classes
whose permutations of length $n$ form a~subgroup of $S_n$ for every $n$ and completely
characterise the types of groups which may occur this way. These results were recently
refined and extended by Lehtonen and Pöschel \cite{Lehtonen16, LehtonenPoschel16}.
In an earlier version of their paper, Atkinson and Beals \cite{AtkinsonBeals99} also deal with composing permutation classes,
showing that compositions of many pairs of finitely based classes are again finitely based.
Some permutation classes characterise permutations which can be sorted
by some sorting machine such as a~stack. In this view, a~composition of two permutation
classes can characterise permutations sortable by two corresponding sorting machines connected
serially. For example, Atkinson and Stitt \cite[Section 6.4]{AtkinsonStitt02} introduce
the pop-stack, a~sorting machine which sorts precisely the layered permutations (see Section \ref{ch4} for a~definition),
and consider the class of permutations which can be sorted by two pop-stacks in series, i.e.
which can be written as a~composition of two layered permutations.
Using their more general results they calculate its generating function and enumerate its basis.
Albert et al. \cite{AlbertEtAl07} give more enumerative results on compositions of classes
in terms of sorting machines.
In the present paper, we study a~different question connected to compositions of classes;
namely whether a~permutation of a~given class $\mathcal{C}$ can always be written as a~composition
of two or more permutations from its subclasses, i.e. whether $\mathcal{C} \subseteq \mathcal{C}_1 \circ \mathcal{C}_2 \circ \cdots \circ \mathcal{C}_k$
for some $\mathcal{C}_1, \ldots, \mathcal{C}_k \subsetneq \mathcal{C}$. If this is true, we say that the class $\mathcal{C}$ is \emph{composable}
and we refer to this property of $\mathcal{C}$ as \emph{composability}.
The paper is organised as follows. In Section \ref{ch1} we supply all the necessary
definitions and facts about permutation classes. In Section \ref{ch2}
we introduce composability and give some basic results. In Section \ref{ch3}
we explore composability of the class $\textnormal{Av}(k\cdots21)$. In Section \ref{ch4} we explore
composability of various classes of layered patterns. Finally in Section \ref{ch5}
we give several additional miscellaneous results.
\section{Preliminaries} \label{ch1}
For a~positive integer $n$ we let $[n]$ denote the set $\{1, 2, \ldots, n\}$. A~\emph{permutation of order}
$n$ is a~bijective function $\pi \colon [n] \longrightarrow [n]$. We
denote the order of a~permutation $\pi$ by $|\pi|$.
We may also interpret a~permutation $\pi$ as a~sequence $\pi(1), \pi(2), \ldots, \pi(n)$
of distinct elements of $[n]$, or as a~diagram in an $n \times n$ square in the plane,
consisting of points $\{(i,\pi(i)); 1 \leq i \leq n\}$. For $n \geq 0$
let $\mathcal{S}_n$ denote the set of all permutations of order $n$.
If $\pi$ and $\sigma$ are two permutations of order $n$ we define their \emph{composition} $\pi \circ \sigma$
as $(\pi \circ \sigma)(i) = \pi(\sigma(i))$ for every $i \in [n]$.
We define two more permutation operators. The \emph{sum} $\pi \oplus \sigma$
of permutations $\pi \in \mathcal{S}_k$ and $\sigma \in \mathcal{S}_l$ is the permutation
$\pi(1), \pi(2), \ldots, \pi(k), \sigma(1) + k, \sigma(2) + k, \ldots, \sigma(l) + k.$
The \emph{skew sum} $\pi \ominus \sigma$ is the permutation
$\pi(1) + l, \pi(2) + l, \ldots, \pi(k) + l, \sigma(1), \sigma(2), \ldots, \sigma(l).$
For example, $3127645 = 312 \oplus 4312$ and $6547123 = 3214 \ominus 123$ (see Figure \ref{fig_sum}).
\begin{figure}[h]
\centering
\subfloat[][$3127645 = 312 \oplus 4312$] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.4cm,y=0.4cm]
\clip(0.42,0.52) rectangle (15.46,15.45);
\draw (1,1)-- (15,1);
\draw (1,1)-- (1,15);
\draw (1,15)-- (15,15);
\draw (15,15)-- (15,1);
\draw (19,1)-- (33,1);
\draw (33,1)-- (33,15);
\draw (33,15)-- (19,15);
\draw (19,15)-- (19,1);
\draw (1,7)-- (15,7);
\draw (19,7)-- (33,7);
\draw (7,1)-- (7,15);
\draw (27,15)-- (27,1);
\begin{scriptsize}
\fill [color=black] (2,6) circle (2pt);
\fill [color=black] (4,2) circle (2pt);
\fill [color=black] (6,4) circle (2pt);
\fill [color=black] (8,14) circle (2pt);
\fill [color=black] (10,12) circle (2pt);
\fill [color=black] (12,8) circle (2pt);
\fill [color=black] (14,10) circle (2pt);
\fill [color=black] (20,12) circle (2pt);
\fill [color=black] (22,10) circle (2pt);
\fill [color=black] (24,8) circle (2pt);
\fill [color=black] (26,14) circle (2pt);
\fill [color=black] (28,2) circle (2pt);
\fill [color=black] (30,4) circle (2pt);
\fill [color=black] (32,6) circle (2pt);
\end{scriptsize}
\end{tikzpicture}
}
\subfloat[][$6547123 = 3214 \ominus 123$] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.4cm,y=0.4cm]
\clip(18.45,0.52) rectangle (33.56,15.45);
\draw (1,1)-- (15,1);
\draw (1,1)-- (1,15);
\draw (1,15)-- (15,15);
\draw (15,15)-- (15,1);
\draw (19,1)-- (33,1);
\draw (33,1)-- (33,15);
\draw (33,15)-- (19,15);
\draw (19,15)-- (19,1);
\draw (1,7)-- (15,7);
\draw (19,7)-- (33,7);
\draw (7,1)-- (7,15);
\draw (27,15)-- (27,1);
\begin{scriptsize}
\fill [color=black] (2,6) circle (2pt);
\fill [color=black] (4,2) circle (2pt);
\fill [color=black] (6,4) circle (2pt);
\fill [color=black] (8,14) circle (2pt);
\fill [color=black] (10,12) circle (2pt);
\fill [color=black] (12,8) circle (2pt);
\fill [color=black] (14,10) circle (2pt);
\fill [color=black] (20,12) circle (2pt);
\fill [color=black] (22,10) circle (2pt);
\fill [color=black] (24,8) circle (2pt);
\fill [color=black] (26,14) circle (2pt);
\fill [color=black] (28,2) circle (2pt);
\fill [color=black] (30,4) circle (2pt);
\fill [color=black] (32,6) circle (2pt);
\end{scriptsize}
\end{tikzpicture}
}
\caption{An example of sums and skew sums}
\label{fig_sum}
\end{figure}
In addition, we will sometimes write $\pi_1 \oplus \pi_2 \oplus \cdots \oplus \pi_k$ as $\bigoplus_{i=1}^k\pi_i$.
\subsection{Permutation classes}
Two sequences of numbers $s_1, s_2, \ldots, s_n$ and $r_1, r_2, \ldots, r_n$ are
\emph{order-isomorphic} if for any two indices $i,j \in [n]$ it holds that $s_i < s_j$
if and only if $r_i < r_j$.
We define the following partial ordering on the set of all permutations.
We say that $\pi$ is \emph{contained in} $\sigma$ and write $\pi \leq \sigma$ if
$\sigma$ has a~subsequence of length $|\pi|$ order-isomorphic to $\pi$.
See the example of containment in Figure \ref{fig_contain}. On the other hand,
if $\pi \nleq \sigma$, we say that $\sigma$ \emph{avoids} $\pi$.
\begin{figure}[ht]
\centering
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.5cm,y=0.5cm]
\clip(0.69,0.64) rectangle (13.4,13.35);
\draw (1,1)-- (1,13);
\draw (1,1)-- (13,1);
\draw (13,1)-- (13,13);
\draw (13,13)-- (1,13);
\draw(4,8) circle (0.24cm);
\draw(6,6) circle (0.24cm);
\draw(12,10) circle (0.24cm);
\begin{scriptsize}
\fill [color=black] (2,2) circle (3pt);
\fill [color=black] (4,8) circle (3pt);
\fill [color=black] (6,6) circle (3pt);
\fill [color=black] (8,12) circle (3pt);
\fill [color=black] (10,4) circle (3pt);
\fill [color=black] (12,10) circle (3pt);
\end{scriptsize}
\end{tikzpicture}
\caption{The permutation 213 is contained in 143625.}
\label{fig_contain}
\end{figure}
A set $\mathcal{C}$ of permutations is called a~\emph{permutation class} if for every $\pi \in \mathcal{C}$ and every $\sigma \leq \pi$
we have $\sigma \in \mathcal{C}$. We say that $\mathcal{C}$ \emph{avoids} a~permutation $\sigma$ if $\sigma \notin \mathcal{C}$, i.e.
every $\pi \in \mathcal{C}$ avoids $\sigma$.
Permutation classes are often described by the patterns they avoid. If $B$ is any set of permutations,
we denote by $\textnormal{Av}(B)$ the set of all permutations avoiding every element of $B$.
Observe that $\mathcal{C}$ is a~permutation class if and only if $\mathcal{C} = \textnormal{Av}(B)$ for some set $B$.
Indeed, if $\mathcal{C}$ is a~permutation class, then $\mathcal{C} = \textnormal{Av}(\mathcal{S} \setminus \mathcal{C})$,
and if $\sigma \leq \pi \in \mathcal{C}$, then $\pi$ avoids all permutations of $B$ and clearly $\sigma$ avoids them too.
If $\mathcal{C} = \textnormal{Av}(B)$ and $B$ is an anti-chain with respect to containment, we call $B$ the \emph{basis} of $\mathcal{C}$.
Also if $B = \{\pi_1, \pi_2, \ldots, \pi_k\}$ is finite, we write just $\textnormal{Av}(\pi_1, \ldots, \pi_k)$
instead of $\textnormal{Av}(\{\pi_1, \ldots, \pi_k\})$. Finally, if $\mathcal{C} = \textnormal{Av}(\pi)$ for a~single
permutation $\pi$, we say that $\mathcal{C}$ is a~\emph{principal class}.
Let $s_1, s_2, \ldots, s_k$ be $k$ finite sequences of numbers. We denote their concatenation
by $s_1s_2\cdots s_k$.
If a~sequence $s$ can be constructed by interleaving $s_1, s_2, \ldots, s_k$ in some (not necessarily unique) way,
we say that $s$ is a~\emph{merge of} or it is \emph{merged from} $s_1, s_2, \ldots, s_k$.
We define $\mathcal{I}_k$ resp. $\mathcal{D}_k$ to be the class of all permutations merged from at most $k$ increasing resp.
decreasing subsequences. Also let $\mathcal{I} = \mathcal{I}_1$ and $\mathcal{D}=\mathcal{D}_1$, i.e. $\mathcal{I} = \textnormal{Av}(21)$ is the set of all increasing
permutations and $\mathcal{D} = \textnormal{Av}(12)$ is the set of all decreasing permutations, and for convenience let
$\mathcal{I}_0 = \mathcal{D}_0 = \mathcal{S}_0$.
The classes $\mathcal{I}_k$ and $\mathcal{D}_k$ are well-known examples of principal classes.
\begin{fact}[Vatter \cite{Vatter15}] \label{basicfact}
$\mathcal{I}_{k-1} = \textnormal{Av}(k\cdots21)$ and $\mathcal{D}_{k-1} = \textnormal{Av}(12\cdots k)$ for any positive integer $k$.
\qed
\end{fact}
Next we recall a~known and important property of infinite permutation classes which will become
useful in the upcoming sections.
\begin{fact}[Atkinson, Beals \cite{AtkinsonBeals01}] \label{fact_infinite}
Let $\mathcal{C}$ be an infinite permutation class. Then either $\mathcal{I} \subseteq \mathcal{C}$ or $\mathcal{D} \subseteq \mathcal{C}$.
\qed
\end{fact}
\subsection{Splittability}
In this section we shortly introduce another concept which has been recently used
to derive enumerative results on permutation classes and which we will also utilize in our work.
A permutation $\pi$ is \emph{merged from permutations} $\alpha$ and $\beta$ if we can
color the elements of $\pi$ with red and blue such that the red subsequence is order-isomorphic
to $\alpha$ and the blue sequence is order-isomorphic to $\beta$. Given two
permutation classes $\mathcal{A}$ and $\mathcal{B}$ we define their \emph{merge} denoted by $\mathcal{A} \odot \mathcal{B}$
as the class of all permutations which can be merged from a~(possibly empty) permutation from $\mathcal{A}$
and a~(possibly empty) permutation from $\mathcal{B}$. For example, it is easy to see that
\begin{equation*} \mathcal{I}_k = \underbrace{\mathcal{I} \odot \mathcal{I} \odot \cdots \odot \mathcal{I}}_{k\times}. \end{equation*}
We say that a~class $\mathcal{C}$ is \emph{splittable} if it has two proper subclasses $\mathcal{A}$ and $\mathcal{B}$ such that $\mathcal{C} \subseteq \mathcal{A} \odot \mathcal{B}$.
We refer the reader to the work of Jelínek and Valtr \cite{JelinekValtr13} for an exhaustive study of splittability.
\section{The notion of composability} \label{ch2}
In the following sections we provide definitions of the key notions of this work
as well as basic facts and observations.
\subsection{Composing permutation classes}
We define the \emph{composition} of two permutation classes $\mathcal{A}$ and $\mathcal{B}$
as the set $\mathcal{A} \circ \mathcal{B} = \{ \pi \circ \varphi; \pi \in \mathcal{A}, \varphi \in \mathcal{B}, |\pi| = |\varphi|\}$.
\begin{lemma} Let $\mathcal{A}$ and $\mathcal{B}$ be arbitrary permutation classes.
\begin{enumerate}[(a)]
\item $\mathcal{A} \circ \mathcal{B}$ is again a~permutation class.
\item Composing permutation classes is associative, i.e. $(\mathcal{A} \circ \mathcal{B}) \circ \mathcal{C} = \mathcal{A} \circ (\mathcal{B} \circ \mathcal{C})$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\alpha \circ \beta = \pi \in \mathcal{A} \circ \mathcal{B}$, so that $\alpha \in \mathcal{A}$ and $\beta \in \mathcal{B}$.
Then a~permutation contained in $\pi$ at indices $i_1 < \cdots < i_r$ is composed
of $\alpha' \leq \alpha$ and $\beta' \leq \beta$ such that $\beta'$ is contained at indices
$i_1, \ldots, i_r$ in $\beta$ and $\alpha'$ is contained at indices $\beta(i_1), \ldots, \beta(i_2)$ in $\alpha$.
Associativity follows from associativity of permutation composition.
\end{proof}
Having verified associativity of the composition operator we can now define the composition of more than two classes in a~natural
inductive way:
\begin{equation*} \mathcal{C}_1 \circ \mathcal{C}_2 \circ \cdots \circ \mathcal{C}_k = (\mathcal{C}_1 \circ \mathcal{C}_2 \circ \cdots \circ \mathcal{C}_{k-1}) \circ \mathcal{C}_k. \end{equation*}
We will also sometimes use the power notation $\underbrace{\mathcal{C} \circ \mathcal{C} \circ \cdots \circ \mathcal{C}}_{k \times} = (\mathcal{C})^k$.
We continue by proving several simple lemmas about composing permutations merged from few increasing sequences.
\begin{lemma} \label{lemma_kl} $\mathcal{I}_k \circ \mathcal{I}_l \subseteq \mathcal{I}_{kl}$ for any integers $k,l \geq 0$.
\end{lemma}
\begin{proof}
Choose $\pi \in \mathcal{I}_k$ and $\varphi \in \mathcal{I}_l$, partition $\varphi$ into $l$ increasing sequences and choose
one of them at indices $i_1 < \cdots < i_r$. Then $\varphi(i_1) < \cdots < \varphi(i_r)$ and so
$\pi(\varphi(i_1)), \ldots, \pi(\varphi(i_r))$ is a~subsequence of $\pi$ and therefore it can be partitioned
into at most $k$ increasing sequences since that is the property of $\pi$. This is true for the image of
each of the $l$ increasing subsequences in $\varphi$ and therefore $\pi \circ \varphi$ can
be partitioned into at most $k\cdot l$ increasing subsequences.
\end{proof}
Since $\mathcal{D} \circ \mathcal{D} = \mathcal{I}$, the argument of the previous proof can be repeated to show that $\mathcal{D}_k \circ \mathcal{D}_l \subseteq \mathcal{I}_{kl}$.
We can generalise this even more.
\begin{lemma} \label{lemma_extrakl} Let $k,l,m,n$ be any non-negative integers. Then
\begin{equation*} (\mathcal{I}_k \odot \mathcal{D}_m) \circ (\mathcal{I}_l \odot \mathcal{D}_n) \subseteq \mathcal{I}_{kl + mn} \odot \mathcal{D}_{kn + ml}. \end{equation*}
\end{lemma}
\begin{proof}
Use the approach identical to that of Lemma \ref{lemma_kl}.
\end{proof}
\subsection{Composability}
The main problem we are addressing in this work is whether permutations in a~given permutation class
can be constructed by composing permutations from two or more smaller classes. We formalise this as follows.
A permutation class $\mathcal{C}$ is said to be \emph{composable from classes} $\mathcal{C}_1, \ldots, \mathcal{C}_k$ if
$\mathcal{C} \subseteq \mathcal{C}_1 \circ \cdots \circ \mathcal{C}_k$. A~class $\mathcal{C}$ is \emph{$k$-composable}, if it is composable from its $k$
proper subclasses $\mathcal{C}_1, \ldots, \mathcal{C}_k$. A~class $\mathcal{C}$ is \emph{composable}, if it is $k$-composable for some $k \geq 2$.
Using this terminology, our goal is thus answering the question whether a~given permutation class is composable.
Clearly, for every class $\mathcal{C}$ we have $\mathcal{C} \subseteq \mathcal{C} \circ \mathcal{I}$. For an infinite class we have either $\mathcal{I} \subseteq \mathcal{C}$,
which implies $\mathcal{C} \subseteq \mathcal{C} \circ \mathcal{C}$, or $\mathcal{D} \subseteq \mathcal{C}$, which implies $\mathcal{I} \subseteq \mathcal{C} \circ \mathcal{C}$ and $\mathcal{C} \subseteq \mathcal{C} \circ \mathcal{C} \circ \mathcal{C}$.
Restricting ourselves to proper subclasses in the definition of a~composable class is motivated
by these trivial inclusions.
We begin the exploration of composability by proving the following result which implies that unlike splittability,
$k$-composability for $k > 2$ does not imply $2$-composability.
\begin{thm} \label{evencomposable}
Let $\mathcal{C}$ be an infinite permutation class such that $\mathcal{I} \nsubseteq \mathcal{C}$. Then $\mathcal{C}$
is not $2k$-composable for any positive integer $k$.
\end{thm}
\begin{proof}
Since $\mathcal{C}$ does not contain $\mathcal{I}$, there is an integer $n$ such that $\mathcal{C}$ avoids $12\cdots n(n+1)$
and therefore $\mathcal{C} \subseteq \mathcal{D}_n$ by Fact \ref{basicfact}.
Now let $\mathcal{A}_1, \mathcal{B}_1, \mathcal{A}_2, \mathcal{B}_2, \ldots, \mathcal{A}_k, \mathcal{B}_k$ be proper subclasses of $\mathcal{C}$
and suppose that $\mathcal{C} \subseteq \mathcal{A}_1 \circ \mathcal{B}_1 \circ \mathcal{A}_2 \circ \mathcal{B}_2 \circ \cdots \circ \mathcal{A}_k \circ \mathcal{B}_k$.
Since all these classes are subsets of $\mathcal{D}_n$, Lemma \ref{lemma_extrakl}
implies $\mathcal{A}_i \circ \mathcal{B}_i \subseteq \mathcal{I}_{n^2}$ for every $i \in [k]$.
Using Lemma \ref{lemma_extrakl} again we get that
\begin{equation*}\mathcal{A}_1 \circ \mathcal{B}_1 \circ \mathcal{A}_2 \circ \mathcal{B}_2 \circ \cdots \circ \mathcal{A}_k \circ \mathcal{B}_k
\subseteq \mathcal{I}_{n^2} \circ \cdots \circ \mathcal{I}_{n^2} \subseteq \mathcal{I}_{n^{2k}},\end{equation*}
therefore, according to our assumption, $\mathcal{C} \subseteq \mathcal{I}_{n^{2k}}$, which means that $\mathcal{C}$
does not contain a~decreasing permutation of length $n^{2k}+1$ by Fact \ref{basicfact}.
But since $\mathcal{C}$ is infinite and does not contain $\mathcal{I}$, it has to contain $\mathcal{D}$
according to Fact \ref{fact_infinite}, which is a~contradiction.
\end{proof}
\subsection{Properties of symmetries}
In this section we explore how composability is preserved under
some of the usual symmetrical maps.
For a~permutation $\pi$ of length $n$ we define $\pi^r$ to be the \emph{reverse} of $\pi$, i.e. $\pi^r(k) = \pi(n-k+1)$,
and $\pi^c$ to be the \emph{complement} of $\pi$, i.e. $\pi^c(k) = n-\pi(k)+1$.
For a~permutation class $\mathcal{A}$ we define
the \emph{inverse class}
$\mathcal{A}^{-1} = \{\pi^{-1}; \pi \in \mathcal{A}\}$,
the \emph{reverse class} $\mathcal{A}^r = \{\pi^r; \pi \in \mathcal{A}\}$, and the \emph{complementary class} $\mathcal{A}^c = \{\pi^c; \pi \in \mathcal{A}\}$.
\begin{figure}[h]
\centering
\subfloat[][$14352$] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.6cm,y=0.6cm]
\clip(-0.18,-0.72) rectangle (5.14,4.7);
\draw (0,4.5)-- (0,-0.5);
\draw (0,-0.5)-- (5,-0.5);
\draw (5,-0.5)-- (5,4.5);
\draw (5,4.5)-- (0,4.5);
\draw (5.5,4.5)-- (5.5,-0.5);
\draw (5.5,-0.5)-- (10.5,-0.5);
\draw (10.5,-0.5)-- (10.5,4.5);
\draw (10.5,4.5)-- (5.5,4.5);
\draw (11,4.5)-- (11,-0.5);
\draw (11,-0.5)-- (16,-0.5);
\draw (16,-0.5)-- (16,4.5);
\draw (16,4.5)-- (11,4.5);
\draw (16.5,4.5)-- (16.5,-0.5);
\draw (16.5,-0.5)-- (21.5,-0.5);
\draw (21.5,-0.5)-- (21.5,4.5);
\draw (21.5,4.5)-- (16.5,4.5);
\begin{scriptsize}
\fill [color=black] (0.5,0) circle (2.0pt);
\fill [color=black] (1.5,3) circle (2.0pt);
\fill [color=black] (2.5,2) circle (2.0pt);
\fill [color=black] (3.5,4) circle (2.0pt);
\fill [color=black] (4.5,1) circle (2.0pt);
\fill [color=black] (6,0) circle (2.0pt);
\fill [color=black] (8,2) circle (2.0pt);
\fill [color=black] (9,1) circle (2.0pt);
\fill [color=black] (7,4) circle (2.0pt);
\fill [color=black] (10,3) circle (2.0pt);
\fill [color=black] (15.5,0) circle (2.0pt);
\fill [color=black] (14.5,3) circle (2.0pt);
\fill [color=black] (13.5,2) circle (2.0pt);
\fill [color=black] (12.5,4) circle (2.0pt);
\fill [color=black] (11.5,1) circle (2.0pt);
\fill [color=black] (17,4) circle (2.0pt);
\fill [color=black] (17.5,1) circle (2.0pt);
\fill [color=black] (19,2.08) circle (2.0pt);
\fill [color=black] (20,0) circle (2.0pt);
\fill [color=black] (21,3) circle (2.0pt);
\end{scriptsize}
\end{tikzpicture}
}
\subfloat[][$(14352)^{-1}$] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.6cm,y=0.6cm]
\clip(5.36,-0.72) rectangle (10.7,4.7);
\draw (0,4.5)-- (0,-0.5);
\draw (0,-0.5)-- (5,-0.5);
\draw (5,-0.5)-- (5,4.5);
\draw (5,4.5)-- (0,4.5);
\draw (5.5,4.5)-- (5.5,-0.5);
\draw (5.5,-0.5)-- (10.5,-0.5);
\draw (10.5,-0.5)-- (10.5,4.5);
\draw (10.5,4.5)-- (5.5,4.5);
\draw (11,4.5)-- (11,-0.5);
\draw (11,-0.5)-- (16,-0.5);
\draw (16,-0.5)-- (16,4.5);
\draw (16,4.5)-- (11,4.5);
\draw (16.5,4.5)-- (16.5,-0.5);
\draw (16.5,-0.5)-- (21.5,-0.5);
\draw (21.5,-0.5)-- (21.5,4.5);
\draw (21.5,4.5)-- (16.5,4.5);
\begin{scriptsize}
\fill [color=black] (0.5,0) circle (2.0pt);
\fill [color=black] (1.5,3) circle (2.0pt);
\fill [color=black] (2.5,2) circle (2.0pt);
\fill [color=black] (3.5,4) circle (2.0pt);
\fill [color=black] (4.5,1) circle (2.0pt);
\fill [color=black] (6,0) circle (2.0pt);
\fill [color=black] (8,2) circle (2.0pt);
\fill [color=black] (9,1) circle (2.0pt);
\fill [color=black] (7,4) circle (2.0pt);
\fill [color=black] (10,3) circle (2.0pt);
\fill [color=black] (15.5,0) circle (2.0pt);
\fill [color=black] (14.5,3) circle (2.0pt);
\fill [color=black] (13.5,2) circle (2.0pt);
\fill [color=black] (12.5,4) circle (2.0pt);
\fill [color=black] (11.5,1) circle (2.0pt);
\fill [color=black] (17,4) circle (2.0pt);
\fill [color=black] (17.5,1) circle (2.0pt);
\fill [color=black] (19,2.08) circle (2.0pt);
\fill [color=black] (20,0) circle (2.0pt);
\fill [color=black] (21,3) circle (2.0pt);
\end{scriptsize}
\end{tikzpicture}
}
\subfloat[][$(14352)^{r}$] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.6cm,y=0.6cm]
\clip(10.92,-0.72) rectangle (16.16,4.7);
\draw (0,4.5)-- (0,-0.5);
\draw (0,-0.5)-- (5,-0.5);
\draw (5,-0.5)-- (5,4.5);
\draw (5,4.5)-- (0,4.5);
\draw (5.5,4.5)-- (5.5,-0.5);
\draw (5.5,-0.5)-- (10.5,-0.5);
\draw (10.5,-0.5)-- (10.5,4.5);
\draw (10.5,4.5)-- (5.5,4.5);
\draw (11,4.5)-- (11,-0.5);
\draw (11,-0.5)-- (16,-0.5);
\draw (16,-0.5)-- (16,4.5);
\draw (16,4.5)-- (11,4.5);
\draw (16.5,4.5)-- (16.5,-0.5);
\draw (16.5,-0.5)-- (21.5,-0.5);
\draw (21.5,-0.5)-- (21.5,4.5);
\draw (21.5,4.5)-- (16.5,4.5);
\begin{scriptsize}
\fill [color=black] (0.5,0) circle (2.0pt);
\fill [color=black] (1.5,3) circle (2.0pt);
\fill [color=black] (2.5,2) circle (2.0pt);
\fill [color=black] (3.5,4) circle (2.0pt);
\fill [color=black] (4.5,1) circle (2.0pt);
\fill [color=black] (6,0) circle (2.0pt);
\fill [color=black] (8,2) circle (2.0pt);
\fill [color=black] (9,1) circle (2.0pt);
\fill [color=black] (7,4) circle (2.0pt);
\fill [color=black] (10,3) circle (2.0pt);
\fill [color=black] (15.5,0) circle (2.0pt);
\fill [color=black] (14.5,3) circle (2.0pt);
\fill [color=black] (13.5,2) circle (2.0pt);
\fill [color=black] (12.5,4) circle (2.0pt);
\fill [color=black] (11.5,1) circle (2.0pt);
\fill [color=black] (17,4) circle (2.0pt);
\fill [color=black] (17.5,1) circle (2.0pt);
\fill [color=black] (19,2.08) circle (2.0pt);
\fill [color=black] (20,0) circle (2.0pt);
\fill [color=black] (21,3) circle (2.0pt);
\end{scriptsize}
\end{tikzpicture}
}
\subfloat[][$(14352)^{c}$] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.6cm,y=0.6cm]
\clip(16.38,-0.72) rectangle (21.82,4.7);
\draw (0,4.5)-- (0,-0.5);
\draw (0,-0.5)-- (5,-0.5);
\draw (5,-0.5)-- (5,4.5);
\draw (5,4.5)-- (0,4.5);
\draw (5.5,4.5)-- (5.5,-0.5);
\draw (5.5,-0.5)-- (10.5,-0.5);
\draw (10.5,-0.5)-- (10.5,4.5);
\draw (10.5,4.5)-- (5.5,4.5);
\draw (11,4.5)-- (11,-0.5);
\draw (11,-0.5)-- (16,-0.5);
\draw (16,-0.5)-- (16,4.5);
\draw (16,4.5)-- (11,4.5);
\draw (16.5,4.5)-- (16.5,-0.5);
\draw (16.5,-0.5)-- (21.5,-0.5);
\draw (21.5,-0.5)-- (21.5,4.5);
\draw (21.5,4.5)-- (16.5,4.5);
\begin{scriptsize}
\fill [color=black] (0.5,0) circle (2.0pt);
\fill [color=black] (1.5,3) circle (2.0pt);
\fill [color=black] (2.5,2) circle (2.0pt);
\fill [color=black] (3.5,4) circle (2.0pt);
\fill [color=black] (4.5,1) circle (2.0pt);
\fill [color=black] (6,0) circle (2.0pt);
\fill [color=black] (8,2) circle (2.0pt);
\fill [color=black] (9,1) circle (2.0pt);
\fill [color=black] (7,4) circle (2.0pt);
\fill [color=black] (10,3) circle (2.0pt);
\fill [color=black] (15.5,0) circle (2.0pt);
\fill [color=black] (14.5,3) circle (2.0pt);
\fill [color=black] (13.5,2) circle (2.0pt);
\fill [color=black] (12.5,4) circle (2.0pt);
\fill [color=black] (11.5,1) circle (2.0pt);
\fill [color=black] (17,4) circle (2.0pt);
\fill [color=black] (18,1) circle (2.0pt);
\fill [color=black] (19,2) circle (2.0pt);
\fill [color=black] (20,0) circle (2.0pt);
\fill [color=black] (21,3) circle (2.0pt);
\end{scriptsize}
\end{tikzpicture}
}
\caption{Symmetries of the permutation 14352}
\end{figure}
It is clear that all these class operators are involutory, i.e. $(\mathcal{A}^{-1})^{-1} = \mathcal{A}$, $(\mathcal{A}^r)^r = \mathcal{A}$
and $(\mathcal{A}^c)^c = \mathcal{A}$. The following simple lemma describes how these operators relate to composition.
\begin{lemma} \label{lemma_basicsym}
Let $\mathcal{A}, \mathcal{A}_1, \mathcal{A}_2, \ldots, \mathcal{A}_k$ be permutation classes. Then
\begin{enumerate}[(a)]
\item $(\mathcal{A}_1 \circ \mathcal{A}_2 \circ \cdots \circ \mathcal{A}_k)^{-1} = \mathcal{A}_k^{-1} \circ \cdots \circ \mathcal{A}_2^{-1} \circ \mathcal{A}_1^{-1}$,
\item $\mathcal{A}^r = \mathcal{A} \circ \mathcal{D}$ and $\mathcal{A}^c = \mathcal{D} \circ \mathcal{A}$,
\item $(\mathcal{A}^r)^c = (\mathcal{A}^c)^r = \mathcal{D} \circ \mathcal{A} \circ \mathcal{D}$.
\end{enumerate}
\end{lemma}
\begin{proof}
(a): If $\pi_i \in \mathcal{A}_i$ for every $i \in [k]$, then by the property of inverse elements in a~group we have
$(\pi_1 \circ \pi_2 \circ \cdots \circ \pi_k)^{-1} = \pi_k^{-1} \circ \cdots \circ \pi_2^{-1} \circ \pi_1^{-1}$.
(b): Let $\alpha \in \mathcal{A}$ and $\delta \in \mathcal{D}$ be permutations of order $n$. By definition $\delta(k) = n - k + 1$
for every $k \in [n]$. Therefore $\alpha(\delta(k)) = \alpha(n-k+1) = \alpha^r(k)$ and
$\delta(\alpha(k)) = n - \alpha(k) + 1 = \alpha^c(k)$ for every $k \in [n]$.
(c): Apparent from (b).
\end{proof}
Using this lemma we derive several composability criteria for symmetries of a~given class,
the first of which requires no further proof as it is an immediate consequence of Lemma \ref{lemma_basicsym}.
\begin{cor} \label{cor_inversion}
Let $\mathcal{A}$ be a~permutation class. Then the following statements are equivalent:
\begin{enumerate}[(a)]
\item $\mathcal{A}$ is composable,
\item $\mathcal{A}^{-1}$ is composable,
\item $(\mathcal{A}^r)^c$ is composable.
\end{enumerate}
\end{cor}
The case of the reverse and complementary operators is more complicated and requires additional assumptions.
\begin{lemma}\label{lemma_symcomp}
If $\mathcal{A}$ is a~$k$-composable class and $\mathcal{I} \subsetneq \mathcal{A}$, then both $\mathcal{A}^r$ and $\mathcal{A}^c$ are $(2k-1)$-composable.
\end{lemma}
\begin{proof}
Let $\mathcal{A}$ be composable from its proper subclasses $\mathcal{A}_1, \mathcal{A}_2, \ldots, \mathcal{A}_k$.
Then
\begin{equation*} \mathcal{A}^r = \mathcal{A} \circ \mathcal{D} \subseteq \mathcal{A}_1 \circ \mathcal{A}_2 \circ \cdots \circ \mathcal{A}_k \circ \mathcal{D} =
(\mathcal{A}_1^r \circ \mathcal{D}) \circ (\mathcal{A}_2^r \circ \mathcal{D}) \circ \cdots \circ (\mathcal{A}_k^r \circ \mathcal{D}) \circ \mathcal{D}.\end{equation*}
It holds that $\mathcal{D} \circ \mathcal{D} = \mathcal{I}$, so we have
\begin{equation*} \mathcal{A}^r \subseteq \mathcal{A}_1^r \circ \mathcal{D} \circ \mathcal{A}_2^r \circ \mathcal{D} \circ \cdots \circ \mathcal{A}_k^r.\end{equation*}
Clearly $\mathcal{A}_i^r \subsetneq \mathcal{A}^r$ and since $\mathcal{I} \subsetneq \mathcal{A}$, we have $\mathcal{D} \subsetneq \mathcal{A}^r$,
so the proper subclass criterion is met and $\mathcal{A}^r$ is therefore $(2k-1)$-composable.
Analogously we show that
\begin{equation*} \mathcal{A}^c \subseteq \mathcal{A}_1^c \circ \mathcal{D} \circ \mathcal{A}_2^c \circ \mathcal{D} \circ \cdots \circ \mathcal{A}_k^c. \qedhere \end{equation*}
\end{proof}
\section{On permutations avoiding a decreasing sequence} \label{ch3}
Recall that $\mathcal{I}_k = \textnormal{Av}((k+1)\cdots21)$ is the class of permutations merged from $k$ increasing sequences,
or equivalently those avoiding a~decreasing sequence of length $k+1$.
In this section, we prove that $\mathcal{I}_k$ is 2-composable and show several
examples of how $\mathcal{I}_k$ can be composed from two proper subclasses.
\subsection{Vertical and horizontal merge}
Let $\mathcal{C}_1, \ldots, \mathcal{C}_k$ be any permutation classes. We define the \emph{vertical merge} of these classes as the class
of permutations that can be written as a~concatenation $s_1s_2\cdots s_k$ of $k$ (possibly empty) sequences
such that $s_i$ is order-isomorphic
to a~permutation of $\mathcal{C}_i$. We write this class as $\mathcal{V}(\mathcal{C}_1, \ldots, \mathcal{C}_k)$. In addition,
if $\mathcal{C}_1 = \mathcal{C}_2 = \cdots = \mathcal{C}_k = \mathcal{I}$, we let $\mathcal{V}_k$ denote the class $\mathcal{V}(\mathcal{C}_1, \ldots, \mathcal{C}_k)$. Similarly we define
the $\emph{horizontal merge}$ of these classes as the class of permutations that
can be written as a~merge of $k$ (possibly empty) sequences $s_1, s_2, \ldots, s_k$
such that each $s_i$ is order-isomorphic to $\pi_i \in \mathcal{C}_i$ and every element of $s_i$
is smaller than every element of $s_{i+1}$ for $1 \leq i \leq k-1$. Note that this implies
that each $s_i$ uses a set of consecutive integers.
We let $\mathcal{H}(\mathcal{C}_1, \ldots, \mathcal{C}_k)$ denote the horizontal merge of classes $\mathcal{C}_1, \ldots, \mathcal{C}_k$ and if
$\mathcal{C}_1 = \mathcal{C}_2 = \cdots = \mathcal{C}_k = \mathcal{I}$ we write $\mathcal{H}_k = \mathcal{H}(\mathcal{C}_1, \ldots, \mathcal{C}_k)$.
Alternatively, we can observe that $\pi \in \mathcal{V}(\mathcal{C}_1, \ldots, \mathcal{C}_k)$ resp. $\pi \in \mathcal{H}(\mathcal{C}_1, \ldots,\mathcal{C}_k)$
if and only if its plot in $\mathbb{R}^2$ can be separated by vertical resp.
horizontal lines into at most $k$ parts , $i$-th of them containing a~sequence order-isomorphic
to a~permutation in $\mathcal{C}_i$ (see Figure \ref{figure_VH}), hence the names of the classes.
\begin{figure}[h]
\centering
\subfloat[][An element of the vertical merge $\mathcal{V}_k$]{
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm]
\clip(1.66,0.44) rectangle (8.4,5.5);
\draw (3,1)-- (7,1);
\draw (7,1)-- (7,5);
\draw (7,5)-- (3,5);
\draw (3,5)-- (3,1);
\draw (3.84,5)-- (3.84,1);
\draw (4.86,1)-- (4.86,5);
\draw (6,5)-- (6,1);
\begin{scriptsize}
\fill [color=black] (3.24,1.44) circle (1.5pt);
\fill [color=black] (3.48,2.56) circle (1.5pt);
\fill [color=black] (3.66,4.26) circle (1.5pt);
\fill [color=black] (4.06,1.14) circle (1.5pt);
\fill [color=black] (4.28,1.94) circle (1.5pt);
\fill [color=black] (4.5,3.18) circle (1.5pt);
\fill [color=black] (4.74,4.78) circle (1.5pt);
\fill [color=black] (5.04,1.36) circle (1.5pt);
\fill [color=black] (5.22,2.18) circle (1.5pt);
\fill [color=black] (5.38,2.82) circle (1.5pt);
\fill [color=black] (5.62,3.94) circle (1.5pt);
\fill [color=black] (5.82,4.54) circle (1.5pt);
\fill [color=black] (6.18,1.64) circle (1.5pt);
\fill [color=black] (6.3,2.4) circle (1.5pt);
\fill [color=black] (6.58,3.44) circle (1.5pt);
\fill [color=black] (6.84,4.1) circle (1.5pt);
\end{scriptsize}
\end{tikzpicture}
}
\subfloat[][An element of the horizontal merge $\mathcal{H}_k$]{
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm]
\clip(6.8,0.44) rectangle (13.22,5.24);
\draw (6,5)-- (6,1);
\draw (8,1)-- (12,1);
\draw (12,1)-- (12,5);
\draw (12,5)-- (8,5);
\draw (8,5)-- (8,1);
\draw (8,4.12)-- (12,4.12);
\draw (8,2.96)-- (12,2.96);
\draw (8,2.14)-- (12,2.14);
\begin{scriptsize}
\fill [color=black] (8.72,1.12) circle (1.5pt);
\fill [color=black] (9.26,1.38) circle (1.5pt);
\fill [color=black] (10.22,1.68) circle (1.5pt);
\fill [color=black] (11.34,1.96) circle (1.5pt);
\fill [color=black] (8.34,2.24) circle (1.5pt);
\fill [color=black] (9.86,2.56) circle (1.5pt);
\fill [color=black] (10.94,2.86) circle (1.5pt);
\fill [color=black] (8.1,3.06) circle (1.5pt);
\fill [color=black] (8.96,3.42) circle (1.5pt);
\fill [color=black] (10.42,3.68) circle (1.5pt);
\fill [color=black] (11.88,3.98) circle (1.5pt);
\fill [color=black] (8.62,4.2) circle (1.5pt);
\fill [color=black] (9.66,4.48) circle (1.5pt);
\fill [color=black] (11.24,4.78) circle (1.5pt);
\end{scriptsize}
\end{tikzpicture}
}
\caption{Examples of vertical and horizontal merges}
\label{figure_VH}
\end{figure}
In addition we define $\mathcal{H} = \mathcal{H}_2$ and $\mathcal{V} = \mathcal{V}_2$ for future convenience.
We continue by observing an important connection between the horizontal and vertical merge.
\begin{lemma}\label{lemma_VHinvert}
Let $\mathcal{C}_1, \ldots, \mathcal{C}_k$ be any permutation classes. Then
\begin{equation*} \mathcal{H}(\mathcal{C}_1, \ldots, \mathcal{C}_k) = \left(\mathcal{V}(\mathcal{C}_1^{-1}, \ldots, \mathcal{C}_k^{-1})\right)^{-1}. \end{equation*}
\end{lemma}
\begin{proof}
If $\pi \in \mathcal{V}(\mathcal{C}^{-1}_1, \ldots, \mathcal{C}^{-1}_k)$, we have that $\pi = s_1s_2\cdots s_k$ such that $s_i$ is order-isomorphic
to $\pi_i \in \mathcal{C}_i^{-1}$. For every $i \in [k]$, $\pi^{-1}$ contains a~set of consecutive integers
on indices $(s_i)_1, (s_i)_2, \ldots, (s_i)_{|s_i|}$ and the sequence at these indices is order-isomorphic
to $\pi_i^{-1} \in \mathcal{C}_i$.
The opposite inclusion is equally straightforward.
\end{proof}
When composed with any other class $\mathcal{A}$, the classes $\mathcal{H}_k$, $\mathcal{V}_k$ and $\mathcal{I}_k$ can be viewed
as a~unary operator transforming $\mathcal{A}$ in a~specific way.
We formalise this approach in the following lemma.
\begin{lemma} \label{lemma_behaviour}
Let $\mathcal{A}$ be an arbitrary permutation class. Then
\begin{enumerate}[(a)]
\item $\mathcal{A} \circ \mathcal{H}_k$ is precisely the class of permutations which can be obtained from a~permutation of $\mathcal{A}$
by dividing it into at most $k$ contiguous subsequences and interleaving them in any way,
\item $\mathcal{A} \circ \mathcal{V}_k$ is precisely the class of permutations which can be obtained from a~permutation of $\mathcal{A}$
by dividing it into at most $k$ subsequences and concatenating them,
\item $\mathcal{A} \circ \mathcal{I}_k$ is precisely the class of permutations which can be obtained from a~permutation of $\mathcal{A}$
by dividing it into at most $k$ subsequences and interleaving them in any way.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\alpha \in \mathcal{A}$, $\eta \in \mathcal{H}_k$, $\nu \in \mathcal{V}_k$ and $\iota \in \mathcal{I}_k$.
(a):
Consider the permutation $\alpha \circ \eta \in \mathcal{A} \circ \mathcal{H}_k$. Then $\eta$ is merged from $k$ (possibly empty) sequences
of consecutive integers $s_1, \ldots, s_k$. Now we define $k$ sequences $r_1, \ldots, r_k$
such that $|s_i| = |r_i|$ and $(r_i)_j = \alpha((s_i)_j)$ for every $i \in [k]$ and $j \in [|s_i|]$
. Every $r_i$ is a~contiguous subsequence of $\alpha$ and at the same time $\alpha \circ \eta$
is merged from $r_1, \ldots, r_k$.
On the other hand, if a~permutation $\pi$ is obtained from $\alpha \in \mathcal{A}$ by dividing it into $k$~contiguous subsequences
$r_1, \ldots, r_k$ and merging them in some way, we define $k$~sequences $s_1, \ldots, s_k$
such that $s_i$ is the sequence of indices of the elements of $r_i$ in $\alpha$. Then by definition $\alpha((s_i)_j) = (r_i)_j$
for any suitable $i$~and $j$, and since we divided $\alpha$
into contiguous subsequences, each $s_i$ is a~sequence of consecutive integers.
Now consider the permutation $\eta$ created by replacing the subsequence $r_i$ by the sequence $s_i$ in $\pi$ for every $i$.
Then $\eta$ is merged from $s_1, \ldots, s_k$, which are sequences of consecutive integers, therefore $\eta \in \mathcal{H}$.
At the same time, for any $m \in \{1, 2, \ldots, |\pi|\}$ there are indices $i$~and $j$~such that $\pi(m) = (r_i)_j =
\alpha((s_i)_j) = \alpha(\eta(m))$, where the last equality holds because we replaced $(r_i)_j$ by $(s_i)_j$ when
constructing $\eta$ from $\pi$. Therefore $\pi = \alpha \circ \eta$.
(b):
Consider the permutation $\alpha \circ \nu \in \mathcal{A} \circ \mathcal{V}_k$. The permutation $\nu$ is formed by concatenating $k$ increasing sequences
$s_1, \ldots, s_k$. Define $k$ sequences $r_1, \ldots, r_k$ such that $|r_i| = |s_i|$ and $(r_i)_j= \alpha((s_i)_j)$.
Each $r_i$ is a~subsequence of $\alpha$ and at the same time $\alpha \circ \nu = r_1r_2\cdots r_k$.
On the other hand, if a~permutation $\pi$ is obtained from $\alpha \in \mathcal{A}$ by dividing it into $k$ subsequences
$r_1, \ldots, r_k$ and then concatenating them, we define $k$ sequences $s_1, \ldots, s_k$ such that
$s_i$ is the sequence of indices of elements of $r_i$ in $\alpha$. Thus every $s_i$ is an increasing sequence
and $\alpha((s_i)_j) = (r_i)_j$.
Consider a permutation $\nu$ created by replacing the subsequence $r_i$ by the sequence $s_i$ in $\pi$ for every $i$.
Since $\pi$ is a concatenation of $r_1, \ldots, r_k$, we get that $\nu$ is a concatenation of $s_1, \ldots, s_k$
and thus $\nu \in \mathcal{V}_k$.
Also, for any $m \in \{1, 2, \ldots, |\pi|\}$ there are indices $i$~and $j$~such that $\pi(m) = (r_i)_j =
\alpha((s_i)_j) = \alpha(\nu(m))$, where the last equality holds because we replaced $(r_i)_j$ by $(s_i)_j$ when
constructing $\nu$ from $\pi$. Therefore $\pi = \alpha \circ \nu$.
(c): The proof is similar to the proofs of (a) and (b).
\end{proof}
\subsection{Composability results}
Using the machinery introduced in the previous section
we now prove a~key lemma which we will use to show several composability results.
\begin{lemma}\label{important_lemma}
Let $\mathcal{C}_1, \mathcal{C}_2, \ldots, \mathcal{C}_k$ be arbitrary permutation classes.
Then \begin{equation*}\mathcal{C}_1 \odot \mathcal{C}_2 \odot \cdots \odot \mathcal{C}_k \subseteq \mathcal{V}(\mathcal{C}_1, \ldots, \mathcal{C}_k) \circ \mathcal{H}_k.\end{equation*}
\end{lemma}
\begin{proof}
Consider a~permutation $\pi \in \mathcal{C}_1 \odot \cdots \odot \mathcal{C}_k$ and divide
it into $k$ sequences $s_1, \ldots, s_k$ such that $s_i$ is isomorphic to a~permutation from $\mathcal{C}_i$.
The permutation $\nu = s_1s_2\cdots s_k$ then lies in $\mathcal{V}(\mathcal{C}_1, \ldots, \mathcal{C}_k)$,
which together with Lemma \ref{lemma_behaviour}(a) implies $\pi \in \mathcal{V}(\mathcal{C}_1, \ldots, \mathcal{C}_k) \circ \mathcal{H}_k$.
\end{proof}
By reformulating the previous statement we immediately get the following.
\begin{cor}\label{cor_splitcompose}
Let $\mathcal{A}$, $\mathcal{B}$ and $\mathcal{C}$ be permutation classes such that $\mathcal{C} \subseteq \mathcal{A} \odot \mathcal{B}$. Then $\mathcal{C} \subseteq \mathcal{V}(\mathcal{A},\mathcal{B}) \circ \mathcal{H}$.
\end{cor}
Using what has already been shown in this section it is now elementary to show that $\mathcal{I}_k$ is 2-composable.
\begin{thm} \label{thm_Ik}
The class $\mathcal{I}_k$ is 2-composable for every $k \geq 2$. In particular, $\mathcal{I}_k \subseteq \mathcal{V}_k \circ \mathcal{H}_k$.
\end{thm}
\begin{proof}
Trivially $\mathcal{V}_k \subsetneq \mathcal{I}_k$ and $\mathcal{H}_k \subsetneq \mathcal{I}_k$.
Next we recall that
\begin{equation*}\mathcal{I}_k = \underbrace{\mathcal{I} \odot \cdots \odot \mathcal{I}}_{k\times}\end{equation*}
and use Lemma \ref{important_lemma} for $\mathcal{C}_1 = \mathcal{C}_2 = \cdots = \mathcal{C}_k = \mathcal{I}$.
\end{proof}
We proceed by proving a~result in some sense opposite to that of Lemma \ref{lemma_kl}, namely
we show that $\mathcal{I}_k$ may be constructed from smaller $\mathcal{I}_a, \mathcal{I}_b$ using composition.
\begin{thm}\label{thm_k+l-1}
$\mathcal{I}_{k+l-1} \subseteq \mathcal{I}_k \circ \mathcal{I}_l$ for all integers $k, l \geq 2$.
\end{thm}
\begin{proof}
Consider a~permutation $\pi \in \mathcal{I}_{k+l-1}$, merged from two sequences $a$ and $b$
such that $a$ is merged from $k$ increasing sequences $s_1, \ldots, s_k$ and $b$ is merged from $l-1$ increasing sequences
$s_{k+1}, \ldots, s_{k+l-1}$. Let $c$ be the increasing sequence created by sorting the elements of $b$.
Consider a~permutation $\sigma$ created by merging the sequences $a$ and $c$ so that $c$ and $s_k$ form a~single
increasing sequence. Clearly $\sigma \in \mathcal{I}_k$ and sequences $s_{k+1}, \ldots, s_{k+l-1}$ are subsequences
of $\sigma$, since they are increasing and therefore were not affected by sorting $b$.
According to Lemma \ref{lemma_behaviour}(c) the class $\mathcal{I}_k \circ \mathcal{I}_l$ contains all permutations we can create from $\sigma$ by dividing
it into $l$ subsequences and merging them in any way. It is therefore enough to find a~way to divide $\sigma$
into $l$ subsequences which can be merged into $\pi$. A~simple choice of $l$ such subsequences is
$a, s_{k+1}, \ldots, s_{k+l-1}$.
\end{proof}
This theorem raises the question whether we could construct a~bigger class from given $\mathcal{I}_k$ and $\mathcal{I}_l$.
\begin{question}
Given positive integers $k$ and $l$, what is the largest integer $m = m(k,l)$ such that $\mathcal{I}_m \subseteq \mathcal{I}_k \circ \mathcal{I}_l$?
\end{question}
So far we have shown that $m(k,l) \leq kl$ (Lemma \ref{lemma_kl}) and that \\
$m(k,l) \geq k+l-1$ (Theorem \ref{thm_k+l-1}). It is also not difficult to show the sharp
inequality $m(k,l) < kl$ by constructing a permutation $\pi \in \mathcal{I}_{kl} \setminus (\mathcal{I}_k \circ \mathcal{I}_l)$.
\section{Classes of layered patterns}
\label{ch4}
In this section we cover classes of permutations which can be written as a~sum or as a~skew sum
of increasing or decreasing permutations. Among these classes we provide
infinitely many examples of composable classes as well as several examples of classes which are
not composable.
Let $\iota_k$ denote the increasing permutation of order $k$ and
$\delta_k$ denote the decreasing permutation of order $k$.
A permutation is \emph{layered} if it is a~sum of decreasing permutations
which are then called \emph{layers}. We let $\mathcal{L}$ denote the class of all layered permutations. We let $\mathcal{L}_k$ denote
the class of permutations which are sums of at most $k$ layers.
The complement of a~layered permutation is clearly a~skew sum of increasing permutations and we
call such a~permutation \emph{co-layered}. The class $\mathcal{L}^c$ consists of precisely the co-layered permutations.
\begin{figure}[h]
\centering
\subfloat[][Layered permutation] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.5cm,y=0.5cm]
\clip(-4.12,-6.72) rectangle (8.12,5.62);
\draw [line width=1.2pt] (-4,5.5)-- (-4,-6.5);
\draw [line width=1.2pt] (-4,-6.5)-- (8,-6.5);
\draw [line width=1.2pt] (8,-6.5)-- (8,5.5);
\draw [line width=1.2pt] (8,5.5)-- (-4,5.5);
\draw [line width=1.2pt] (8.5,5.5)-- (8.5,-6.5);
\draw [line width=1.2pt] (8.5,-6.5)-- (20.5,-6.5);
\draw [line width=1.2pt] (20.5,-6.5)-- (20.5,5.5);
\draw [line width=1.2pt] (20.5,5.5)-- (8.5,5.5);
\begin{scriptsize}
\fill [color=black] (-3.5,-4) circle (2.0pt);
\fill [color=black] (-2.5,-5) circle (2.0pt);
\fill [color=black] (-1.5,-6) circle (2.0pt);
\fill [color=black] (-0.5,-2) circle (2.0pt);
\fill [color=black] (0.5,-3) circle (2.0pt);
\fill [color=black] (1.5,4) circle (2.0pt);
\fill [color=black] (2.5,3) circle (2.0pt);
\fill [color=black] (3.5,2) circle (2.0pt);
\fill [color=black] (4.5,1) circle (2.0pt);
\fill [color=black] (5.5,0) circle (2.0pt);
\fill [color=black] (6.5,-1) circle (2.0pt);
\fill [color=black] (7.5,5) circle (2.0pt);
\fill [color=black] (9,3) circle (2.0pt);
\fill [color=black] (10,4) circle (2.0pt);
\fill [color=black] (11,5) circle (2.0pt);
\fill [color=black] (12,-1) circle (2.0pt);
\fill [color=black] (13,0) circle (2.0pt);
\fill [color=black] (14,1) circle (2.0pt);
\fill [color=black] (15,2) circle (2.0pt);
\fill [color=black] (16,-3) circle (2.0pt);
\fill [color=black] (17,-2) circle (2.0pt);
\fill [color=black] (18,-6) circle (2.0pt);
\fill [color=black] (19,-5) circle (2.0pt);
\fill [color=black] (20,-4) circle (2.0pt);
\end{scriptsize}
\end{tikzpicture}
}
\subfloat[][Co-layered permutation] {
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.5cm,y=0.5cm]
\clip(8.42,-6.72) rectangle (20.66,5.62);
\draw [line width=1.2pt] (-4,5.5)-- (-4,-6.5);
\draw [line width=1.2pt] (-4,-6.5)-- (8,-6.5);
\draw [line width=1.2pt] (8,-6.5)-- (8,5.5);
\draw [line width=1.2pt] (8,5.5)-- (-4,5.5);
\draw [line width=1.2pt] (8.5,5.5)-- (8.5,-6.5);
\draw [line width=1.2pt] (8.5,-6.5)-- (20.5,-6.5);
\draw [line width=1.2pt] (20.5,-6.5)-- (20.5,5.5);
\draw [line width=1.2pt] (20.5,5.5)-- (8.5,5.5);
\begin{scriptsize}
\fill [color=black] (-3.5,-4) circle (2.0pt);
\fill [color=black] (-2.5,-5) circle (2.0pt);
\fill [color=black] (-1.5,-6) circle (2.0pt);
\fill [color=black] (-0.5,-2) circle (2.0pt);
\fill [color=black] (0.5,-3) circle (2.0pt);
\fill [color=black] (1.5,4) circle (2.0pt);
\fill [color=black] (2.5,3) circle (2.0pt);
\fill [color=black] (3.5,2) circle (2.0pt);
\fill [color=black] (4.5,1) circle (2.0pt);
\fill [color=black] (5.5,0) circle (2.0pt);
\fill [color=black] (6.5,-1) circle (2.0pt);
\fill [color=black] (7.5,5) circle (2.0pt);
\fill [color=black] (9,3) circle (2.0pt);
\fill [color=black] (10,4) circle (2.0pt);
\fill [color=black] (11,5) circle (2.0pt);
\fill [color=black] (12,-1) circle (2.0pt);
\fill [color=black] (13,0) circle (2.0pt);
\fill [color=black] (14,1) circle (2.0pt);
\fill [color=black] (15,2) circle (2.0pt);
\fill [color=black] (16,-3) circle (2.0pt);
\fill [color=black] (17,-2) circle (2.0pt);
\fill [color=black] (18,-6) circle (2.0pt);
\fill [color=black] (19,-5) circle (2.0pt);
\fill [color=black] (20,-4) circle (2.0pt);
\end{scriptsize}
\end{tikzpicture}
}
\caption{Examples of layered and co-layered patterns}
\end{figure}
We start by proving that $\mathcal{L}_2$ is not composable using a~counting argument. As it turns out,
proper subclasses of $\mathcal{L}_2$ are asymptotically too small to build the entire $\mathcal{L}_2$ class
using composition.
\begin{thm} \label{l2}
The class $\mathcal{L}_2$ is not composable.
\end{thm}
\begin{proof}
Suppose that $\mathcal{L}_2 \subseteq \mathcal{C}_1 \circ \mathcal{C}_2 \circ \cdots \circ \mathcal{C}_k$ such that $\mathcal{C}_i \subsetneq \mathcal{L}_2$ for every $i \in [k]$.
Each of these subclasses avoids at least one permutation of $\mathcal{L}_2$. In other words
for every $\mathcal{C}_i$ there is a~$\pi_i \in \mathcal{L}_2$ such that $\mathcal{C}_i \subseteq \mathcal{L}_2 \cap \textnormal{Av}(\pi_i)$.
Considering a~sufficiently large $n$ so that
$\pi_i \leq \delta_n \oplus \delta_n$ for every $i \in [k]$
we get that $\mathcal{C}_i \subseteq \mathcal{L}_2 \cap \textnormal{Av}(\delta_n \oplus \delta_n)$ for every $i$, in other words
every permutation in these subclasses has one of its two layers shorter than $n$.
It follows that for a~fixed integer $N$ there are at most $2(n-1)$ permutations of order $N$
in any $\mathcal{C}_i$, therefore there are at most $(2n-2)^k$ permutations in $\mathcal{C}_1 \circ \cdots \circ \mathcal{C}_k$.
But $\mathcal{L}_2$ contains $N$ permutations of order $N$ for any $N$, therefore we obtain
a contradiction by choosing $N > (2n-2)^k$.
\end{proof}
The number of permutations of order $n$ in $\mathcal{L}_2$ is linear in $n$ while any proper subclass contains
only constantly many permutations of fixed order. We can use the same approach using the asymptotic
jump from polynomial to exponential functions to show that a~different class of permutations cannot be composable.
Namely, let $\mathcal{F}_2$ be the class of layered permutations with layers of size 1~or 2.
\begin{thm}\label{f2}
The class $\mathcal{F}_2$ is not composable.
\end{thm}
\begin{proof}
Suppose $\mathcal{F}_2$ is composable from $k$ of its proper subclasses $\mathcal{C}_1, \mathcal{C}_2, \ldots, \mathcal{C}_k$.
We choose a~permutation from $\mathcal{F}_2 \setminus C_i$ for every $i$ and we select $n$ large enough so that
every chosen permutation is contained in $\pi = \sum_{i=1}^n\delta_2$. Then if $\mathcal{C} = \mathcal{F}_2 \cap \textnormal{Av}(\pi)$,
we get that $\mathcal{F}_2 \subseteq (\mathcal{C})^k$. Every permutation in $\mathcal{C}$ contains fewer than $n$ layers of size 2, otherwise
it would contain $\pi$. Clearly there are at most $N^a$ permutations of $\mathcal{F}_2$ that have order $N$
and exactly $a$ layers of size 2. Therefore $\mathcal{C}$ contains at most $N^1 + N^2 + \cdots + N^{n-1} \leq nN^n$
permutations of order $N$ and the composition $(\mathcal{C})^k$ then contains at most $n^kN^{nk}$ permutations of
order $N$, which is a~number polynomial in $N$. As mentioned in \cite[Chapter 4]{Vatter15},
the number of permutations of order $N$ of $\mathcal{F}_2$
is counted by the Fibonacci numbers which grow exponentially, therefore there is $N$ large enough so that
$\mathcal{F}_2$ has more permutations of order $N$ than $(\mathcal{C})^k$.
Note that this result also follows immediately from the theorem of Kaiser and Klazar (\cite[3.4]{KaiserKlazar03}),
which states that if the number of permutations of order $n$ in a~permutation class is less than the $n$-th
Fibonacci number for at least one value of $n$, then it is eventually polynomial in $n$. This implies
that every class counted by the Fibonacci numbers is uncomposable.
\end{proof}
The argument used in the proofs above cannot be used for $\mathcal{L}_3$,
so we need a~different approach to show that this class too is not composable. We will make use of the following
property of $\mathcal{L}_2 \cup \mathcal{L}_2^r$.
\begin{lemma} \label{l2group}
$(\mathcal{L}_2 \cup \mathcal{L}_2^r) \cap \mathcal{S}_n$ is a~subgroup of $\mathcal{S}_n$ for every $n$, i.e. it is closed under composition.
\end{lemma}
\begin{proof}
In this proof, we consider an additive group structure on the set $[n]$
with the neutral element $n$ and an operator $+_n$ defined as
\begin{equation*} a +_n b = 1 + (a + b - 1)\mod n. \end{equation*}
First we prove that $\mathcal{L}_2^r \cap \mathcal{S}_n$ by itself is a~subgroup of $\mathcal{S}_n$.
Observe that $\mathcal{L}_2^r \cap \mathcal{S}_n$ contains exactly permutations $\pi$ such that there
is a~shifting number $k$ with $\pi(i) = i+_nk$ for every $i \in [n]$.
Indeed, if $\pi = \iota_a \ominus \iota_b$ then for any $i \in [n]$ we have
$\pi(i) = i+_nb$ and conversely if $\pi(i) = i+_nk$
for every $i\in [n]$ then $\pi = \iota_{n-k} \ominus \iota_k$. Now for two permutations $\pi, \sigma \in \mathcal{L}^r_2$
with shifting numbers $k,l$ respectively we have $\pi(\sigma(i)) =
i +_n l +_n k$ for any $i \in [n]$, therefore $\pi \circ \sigma \in \mathcal{L}_2^r$ since it has
a shifting number $k+_nl$.
It trivially holds that $\mathcal{L}_2 \circ \mathcal{D} = \mathcal{L}_2^r = \mathcal{L}_2^c = \mathcal{D} \circ \mathcal{L}_2$. Considering
$\pi, \sigma \in (\mathcal{L}_2 \cup \mathcal{L}_2^r) \cap \mathcal{S}_n$ it remains to distinguish the following four cases:
\begin{enumerate}[(i)]
\item $\pi \in \mathcal{L}_2^r$ and $\sigma \in \mathcal{L}_2^r$, then $\pi \circ \sigma \in \mathcal{L}_2^r$ by the discussion above,
\item $\pi \in \mathcal{L}_2^r$ and $\sigma \in \mathcal{L}_2$, then $\pi \circ \sigma = (\pi \circ \sigma^r) \circ \delta_n \in \mathcal{L}_2$,
\item $\pi \in \mathcal{L}_2$ and $\sigma \in \mathcal{L}_2^r$, then $\pi \circ \sigma = (\delta_n \circ \pi^c) \circ \sigma =
\delta_n \circ (\pi^r \circ \sigma) \in \mathcal{L}_2$,
\item $\pi \in \mathcal{L}_2$ and $\sigma \in \mathcal{L}_2$, then $\pi \circ \sigma = \pi^r \circ (\delta_n \circ \delta_n) \circ \sigma^c =
\pi^r \circ \sigma^r \in \mathcal{L}^r_2$.
\qedhere
\end{enumerate}
\end{proof}
\begin{thm}\label{l3}
The class $\mathcal{L}_3$ is not composable.
\end{thm}
\begin{proof}
Suppose that $\mathcal{L}_3 \subseteq \mathcal{C}_1 \circ \mathcal{C}_2 \circ \cdots \circ \mathcal{C}_k$ such that $\mathcal{C}_i \subsetneq \mathcal{L}_3$ for any $i$.
Using the same initial argumentation as in the proof of Theorem \ref{l2} we get
that there is an $n$ such that $\mathcal{L}_3 \subseteq (\mathcal{L}_3 \cap \textnormal{Av}(\delta_n \oplus \delta_n \oplus \delta_n))^k$,
meaning that every permutation of $\mathcal{L}_3$ can be composed from $k$ permutations having at least one
of the three layers shorter than $n$.
Let $\pi_i \in \mathcal{C}_i$ for $1 \leq i \leq k$ and $\pi = \pi_1 \circ \pi_2 \circ \cdots \circ \pi_k$.
We now claim that it is possible to remove at most $(n-1)k$ elements from $\pi$
to obtain a~two-layered or a~two-co-layered permutation. We will prove this by induction on $k$.
The case $k=1$ is easy since $\pi = \pi_1$ avoids $\delta_n \oplus \delta_n \oplus \delta_n$,
so it has a~layer of length shorter than $n$ whose removal creates a~two-layered pattern.
For $k > 1$ let $\sigma = \pi_1 \circ \cdots \circ \pi_{k-1}$ and $\pi = \sigma \circ \pi_k$.
Let all these permutations have the order $N$.
By the induction hypothesis, there are $a$ indices $i_1, \ldots i_a$ such
that $a \geq N - (n-1)(k-1)$ and $\sigma$ restricted to these indices has the two-layer
or the two-co-layer pattern. Also there are $b$ indices $j_{1}, \ldots, j_{b}$
such that $b \geq N - (n-1)$ and $\pi_k$ restricted to these indices forms the two-layer
or the two-co-layer pattern.
Let us now restrict the function $\sigma \circ \pi_k$ to the set
$S = \{\pi_k^{-1}(i_1), \ldots, \pi_k^{-1}(i_a)\} \cap \{j_{1}, \ldots, j_{b}\}$ whose size is at least $N - (n-1)k$.
Then both $\pi_k(S)$ and $\sigma(\pi_k(S))$ are still two-layer or two-co-layer patterns,
which implies the same for their composition according to Lemma \ref{l2group}.
Therefore $\pi$ restricted to $S$ forms a~two-layer or two-co-layer pattern
and $N - |S| \leq (n-1)k$ which completes the induction step.
Consequently, any permutation of order $N$ in $\mathcal{C}_1 \circ \mathcal{C}_2 \circ \cdots \circ \mathcal{C}_k$
contains a~two-layered or a~two-co-layered pattern of size at least $N - k(n-1)$.
But choosing $N = 3(k(n-1)+1)$ and considering the permutation $\bigoplus_{i=1}^{3}\delta_{k(n-1)+1} \in \mathcal{L}_3$
we obtain a~contradiction.
\end{proof}
If we allow more than three but still constantly many layers, we always get a~composable class.
\begin{thm} \label{l4}
The class $\mathcal{L}_k$ is $3$-composable for every $k \geq 4$.
\end{thm}
\begin{proof}
We will show that $\mathcal{L}_k \subseteq \mathcal{L}_{k-1} \circ \mathcal{L}_{k-2} \circ \mathcal{L}_{k-1}$.
If $\pi \in \mathcal{L}_k$ of order $n$ has fewer than $k$ layers, then $\pi = \pi \circ \delta_n \circ \delta_n \in \mathcal{L}_{k-1} \circ \mathcal{L}_{k-2} \circ \mathcal{L}_{k-1}$.
Otherwise $\pi$ has at least 4~layers and has the form $\pi = \delta_a \oplus \delta_b \oplus \delta_c \oplus \delta_d \oplus \pi'$
for some positive $a$, $b$, $c$, $d$. Since for every layered $\sigma$ we have $\sigma \circ \sigma \circ \sigma = \sigma$
it is not hard to check that
\begin{equation*}\pi = (\delta_{a+b} \oplus \delta_c \oplus \delta_d \oplus \pi') \circ
(\delta_{a+b} \oplus \delta_{c+d} \oplus \pi') \circ
(\delta_a \oplus \delta_b \oplus \delta_{c+d} \oplus \pi') \in \mathcal{L}_{k-1} \circ \mathcal{L}_{k-2} \circ \mathcal{L}_{k-1}.
\end{equation*}
The situation is represented in Figure \ref{figure_l4}.
\end{proof}
\begin{figure}[h]
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.74cm,y=0.74cm]
\clip(2,1.8) rectangle (20.6,6.18);
\draw (2,2)-- (6,2);
\draw (6,2)-- (6,6);
\draw (6,6)-- (2,6);
\draw (2,6)-- (2,2);
\draw [dash pattern=on 1pt off 1pt] (2,2.61)-- (6,2.61);
\draw [dash pattern=on 1pt off 1pt] (2.61,2)-- (2.61,6);
\draw [dash pattern=on 1pt off 1pt] (3.62,6)-- (3.62,2);
\draw [dash pattern=on 1pt off 1pt] (4.86,2)-- (4.86,6);
\draw [dash pattern=on 1pt off 1pt] (2,4.86)-- (6,4.86);
\draw [dash pattern=on 1pt off 1pt] (6,3.62)-- (2,3.62);
\draw (6.86,2.01)-- (10.86,2.01);
\draw (10.86,2.01)-- (10.86,6.01);
\draw (10.86,6.01)-- (6.86,6.01);
\draw (6.86,6.01)-- (6.86,2.01);
\draw [dash pattern=on 1pt off 1pt] (6.86,2.63)-- (10.86,2.63);
\draw [dash pattern=on 1pt off 1pt] (7.47,2.01)-- (7.47,6.01);
\draw [dash pattern=on 1pt off 1pt] (8.47,6.01)-- (8.47,2.01);
\draw [dash pattern=on 1pt off 1pt] (9.71,2.01)-- (9.71,6.01);
\draw [dash pattern=on 1pt off 1pt] (6.86,4.87)-- (10.86,4.87);
\draw [dash pattern=on 1pt off 1pt] (10.86,3.63)-- (6.86,3.63);
\draw (16.43,2.01)-- (20.43,2.01);
\draw (20.43,2.01)-- (20.43,6.01);
\draw (20.43,6.01)-- (16.43,6.01);
\draw (16.43,6.01)-- (16.43,2.01);
\draw [dash pattern=on 1pt off 1pt] (16.43,2.63)-- (20.43,2.63);
\draw [dash pattern=on 1pt off 1pt] (17.05,2.01)-- (17.05,6.01);
\draw [dash pattern=on 1pt off 1pt] (18.05,6.01)-- (18.05,2.01);
\draw [dash pattern=on 1pt off 1pt] (19.29,2.01)-- (19.29,6.01);
\draw [dash pattern=on 1pt off 1pt] (16.43,4.87)-- (20.43,4.87);
\draw [dash pattern=on 1pt off 1pt] (20.43,3.63)-- (16.43,3.63);
\draw (11.63,1.99)-- (15.63,1.99);
\draw (15.63,1.99)-- (15.63,5.99);
\draw (15.63,5.99)-- (11.63,5.99);
\draw (11.63,5.99)-- (11.63,1.99);
\draw [dash pattern=on 1pt off 1pt] (11.63,2.61)-- (15.63,2.61);
\draw [dash pattern=on 1pt off 1pt] (12.24,1.99)-- (12.24,5.99);
\draw [dash pattern=on 1pt off 1pt] (13.24,5.99)-- (13.24,1.99);
\draw [dash pattern=on 1pt off 1pt] (14.48,1.99)-- (14.48,5.99);
\draw [dash pattern=on 1pt off 1pt] (11.63,4.85)-- (15.63,4.85);
\draw [dash pattern=on 1pt off 1pt] (15.63,3.61)-- (11.63,3.61);
\draw (6,4.2) node[anchor=north west] {$ = $};
\draw (10.9,4.2) node[anchor=north west] {$\circ$};
\draw (15.7,4.2) node[anchor=north west] {$\circ$};
\draw (2,2.61)-- (2.61,2);
\draw (2.61,3.62)-- (3.62,2.61);
\draw (3.62,4.86)-- (4.86,3.62);
\draw (4.86,6)-- (6,4.86);
\draw (6.86,3.63)-- (8.47,2.01);
\draw (8.47,4.87)-- (9.71,3.63);
\draw (9.71,6.01)-- (10.86,4.87);
\draw (11.63,3.61)-- (13.24,1.99);
\draw (13.24,5.99)-- (15.63,3.61);
\draw (16.43,2.63)-- (17.05,2.01);
\draw (17.05,3.63)-- (18.05,2.63);
\draw (18.05,6.01)-- (20.43,3.63);
\end{tikzpicture}
\caption{$\delta_a \oplus \delta_b \oplus \delta_c \oplus \delta_d=(\delta_{a+b} \oplus \delta_c \oplus \delta_d) \circ
(\delta_{a+b} \oplus \delta_{c+d}) \circ
(\delta_a \oplus \delta_b \oplus \delta_{c+d})
$}
\label{figure_l4}
\end{figure}
This theorem raises the question whether $\mathcal{L}_k$ could be 2-composable for $k \geq 4$. Our work
from Section \ref{ch2} quickly determines that this is not the case.
\begin{prop}
$\mathcal{L}_k$ for $k\geq 4$ is not 2-composable. In particular, it is not $n$-composable for
any even number $n$.
\end{prop}
\begin{proof}
Since $\mathcal{L}_k$ is an infinite class which does not contain $\mathcal{I}$ the statement directly follows from Theorem \ref{evencomposable}.
\end{proof}
We have now covered the classes $\mathcal{L}_k$ for all $k \geq 2$.
It remains to consider the class~$\mathcal{L}$, which we show to be uncomposable. Before we proceed with the
proof, we introduce an additional useful concept.
We call a~subsequence $s$ of a~permutation $\pi$ a~\emph{block} if $s$ is either an increasing
or a~decreasing contiguous subsequence of consecutive integers.
We then call $\pi$ a~$k$\emph{-block} if it is a~concatenation of at most $k$ blocks (see Figure \ref{figure_blocks}).
\begin{figure}[ht]
\centering
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.6cm,y=0.6cm]
\clip(0.25,-6.39) rectangle (10.17,3.7);
\draw (0.5,3.5)-- (0.5,-6);
\draw (0.5,-6)-- (10,-6);
\draw (10,-6)-- (10,3.5);
\draw (10,3.5)-- (0.5,3.5);
\draw (0.5,1)-- (10,1);
\draw (0.5,-1)-- (10,-1);
\draw (0.5,-4.5)-- (10,-4.5);
\draw (2.5,3.5)-- (2.5,-6);
\draw (6,-6)-- (6,3.5);
\draw (8.5,3.5)-- (8.5,-6);
\begin{scriptsize}
\fill [color=black] (1,0.5) circle (1.5pt);
\fill [color=black] (1.5,0) circle (1.5pt);
\fill [color=black] (2,-0.5) circle (1.5pt);
\fill [color=black] (3,-4) circle (1.5pt);
\fill [color=black] (3.5,-3.5) circle (1.5pt);
\fill [color=black] (4,-3) circle (1.5pt);
\fill [color=black] (4.5,-2.5) circle (1.5pt);
\fill [color=black] (4.96,-2.1) circle (1.5pt);
\fill [color=black] (5.46,-1.58) circle (1.5pt);
\fill [color=black] (6.5,3) circle (1.5pt);
\fill [color=black] (6.92,2.5) circle (1.5pt);
\fill [color=black] (7.5,2) circle (1.5pt);
\fill [color=black] (8,1.5) circle (1.5pt);
\fill [color=black] (9.02,-5.64) circle (1.5pt);
\fill [color=black] (9.54,-5.1) circle (1.5pt);
\end{scriptsize}
\end{tikzpicture}
\caption{An example of a 4-block}
\label{figure_blocks}
\end{figure}
\begin{lemma} \label{lemma_blocks}
Let $\pi \in S_n$ be a~$k$-block and let $\sigma \in S_n$ be an $l$-block. Then $\pi \circ \sigma$ is a~$(k\cdot l)$-block.
\end{lemma}
\begin{proof}
Choose a~block of $\sigma$ at indices $a, a+1, \ldots, a+b$. Then the sequence
$$\pi(\sigma(a)), \pi(\sigma(a+1)), \ldots, \pi(\sigma(a+b))$$
is a~contiguous subsequence of either
$\pi(1), \pi(2), \ldots, \pi(n)$ or $\pi(n), \ldots, \pi(2), \pi(1)$
and therefore is a~concatenation of at most $k$ blocks since $\pi$ itself is a~$k$-block.
This is true for each of the $l$ blocks of $\sigma$,
therefore $\pi \circ \sigma$ is a~$(k\cdot l)$-block.
\end{proof}
Now we can prove our main result.
\begin{thm}\label{l}
The class $\mathcal{L}$ is not composable.
\end{thm}
\begin{proof}
Every subclass of $\mathcal{L}$ is determined by one or more forbidden layered permutations.
If $\mathcal{L}$ is composable from $k$ subclasses, we may choose one forbidden layered permutation from each of them
and then choose $n$ large enough so that $\pi = \bigoplus_{i=1}^{n+1}\delta_{n+1}$ contains all of the chosen patterns.
That way, $\mathcal{L} \subseteq \mathcal{C}^k$ where $\mathcal{C} = \textnormal{Av}(\pi) \cap \mathcal{L}$.
Clearly every permutation in $\mathcal{C}$ has at most $n$ layers longer than $n$, otherwise it would contain $\pi$.
Our goal is to show that permutations in $\mathcal{C}^k$ are somehow very close to patterns composed from permutations
that have a~constant number of
non-trivial layers and all other layers are just of size 1. Given a layered permutation we call
a layer of length at most $n$ a \emph{short} layer and a layer of length more than $n$ a \emph{long} layer.
We say that two permutations $\alpha$ and $\beta$ are $(c,l)$\emph{-close}, if $|\alpha(i) - \beta(i)| \leq c$
for every index $i$ with at most $l$ exceptions.
For $\sigma \in \mathcal{C}$ we denote by $N(\sigma)$ the permutation created from $\sigma$ by replacing every short layer
by the corresponding number of layers of size 1, i.e. flipping the short layers into increasing blocks.
We can now formally state our goal: we shall prove that for any $\sigma_1, \sigma_2, \ldots, \sigma_k \in \mathcal{C}$
the permutations $\sigma_k \circ \cdots \circ \sigma_2 \circ \sigma_1$ and $N(\sigma_k) \circ \cdots \circ N(\sigma_2) \circ N(\sigma_1)$
are $(2nk, 8n^2k^2)$-close. We will prove this by induction on $k$.
If $k=1$, we have to show that $\sigma_1$ and $N(\sigma_1)$ are $(2n,8n^2)$-close. Since $N(\sigma)$ is created
by manipulating layers of $\sigma$ of length at most $n$ in place, every element of $\sigma$ is shifted by at most $n$, so they
are even $(n,0)$-close, thus the first step of induction is done.
If $k \geq 2$, suppose that $\sigma = \sigma_k \circ \cdots \circ \sigma_2 \circ \sigma_1$ and
$\nu = N(\sigma_k) \circ \cdots \circ N(\sigma_2) \circ N(\sigma_1)$
are $(2nk, 8n^2k)$-close and we shall prove the statement for $k+1$.
Given a~layered permutation and one of its layers of size $l+1$ at indices $i, i+1, \ldots, i+l$,
we say that a~number $u$ is \emph{in the area of influence} of this layer if $i \leq u \leq i+l$.
Given a long layer of $\sigma_{k+1}$ or $N(\sigma_{k+1})$ (their long layers are the same),
there are at most $4nk$ indices $u$ such that $|\sigma(u) - \nu(u)| \leq kn$ and
$\nu(u)$ is in the area of influence of this layer and $\sigma(u)$ is not: at most $2nk$ to the left and to the right of the layer.
Similarly there are at most $4nk$ indices $u$ such that $|\sigma(u) - \nu(u)| \leq kn$ and $\sigma(u)$ is in the area
of influence of the considered layer and $\nu(u)$ is not. Since there are at most $n$ long layers, we get that
in total there are at most $8n^2k$ indices $u$ such that $|\sigma(u) - \nu(u)| \leq kn$ and one of $\{\sigma(u), \nu(u)\}$
is in the area of influence of a long layer while the other is not in that area.
By the induction hypothesis, there are at most $8n^2k^2$ indices $u$ such that $|\sigma(u) - \nu(u)| > nk$.
Together with the at most $8n^2k$ indices from the previous paragraph we get $8n^2k^2 + 8n^2k = 8n^2k(k+1) \leq 8n^2(k+1)^2$
indices at which we will allow $\sigma_{k+1} \circ \sigma$ and $N(\sigma_{k+1}) \circ \nu$ to differ arbitrarily in our
proof that these two permutations are $(2n(k+1), 8n^2(k+1)^2)$-close. It remains to show
$|\sigma_{k+1}(\sigma(u)) - N(\sigma_{k+1}(\nu(u))| \leq n(k+1)$ for all the remaining indices $u$ to complete the induction step.
For other indices $u \in \{1, 2, \ldots, |\sigma|\}$ not considered so far it holds that $|\nu(u) - \sigma(u)| \leq 2nk$
and that either $\nu(u)$ and $\sigma(u)$ are both in the area of influence of the same long layer of $\sigma_{k+1}$
or they are in areas of influence of short or trivial layers. In the latter case the value of $\sigma(u)$ changes by
at most $n$ after applying $\sigma_{k+1}$ to it and similarly the value of $\nu(u)$ changes by at most $n$ after applying
$N(\sigma_{k+1})$ to it, thus
$|N(\sigma_{k+1})(\nu(u)) - \sigma_{k+1}(\sigma(u))| \leq 2nk + 2n \leq n(k+1)$. In the former case
it is enough to realise that for a given decreasing permutation $\delta_a$ it holds that
$\delta_a(x\pm y) = \delta_a(x) \mp y$, thus if $\sigma(u)$ differs by
$y$ from $\nu(u)$ and they are in the area of influence of the same long layer, after applying $\sigma_{k+1}$ (or $N(\sigma_{k+1})$, which
is the same for the big layers) the values still differ by $y \leq 2nk \leq 2n(k+1)$, which finishes
the induction step.
Notice that for $\sigma \in \mathcal{C}$ the permutation $N(\sigma)$ is a~$(2n)$-block
according to the definition above.
Thus by Lemma \ref{lemma_blocks}
we get that by composing $k$ such permutations we get a~permutation which is a~$(2n)^k$-block.
As a~result we get that each permutation from $(\mathcal{C})^k$ is $(c,l)$-close
to a~$C$-block for suitable fixed constants $c,l,C$. Notice
now that every $C$-block avoids the $(C+1)$-block $\gamma = 214365\cdots(2C+2)(2C+1)$,
so every permutation from $(\mathcal{C})^k$ is $(c,l)$-close to a~permutation avoiding $\gamma$. We
can construct a~layered permutation which is not $(c,l)$-close to any $\gamma$-avoider as follows.
Choose a~layered permutation with $C+1$ layers of length $l+2c+1$ and consider a~permutation
$(c,l)$-close to it. Then in every layer there are at least $2c+1$ elements whose value
changed by at most $c$; therefore there exist at least two elements which remained in decreasing order.
Choosing these two elements from every layer forms an occurrence of $\gamma$.
Since $\mathcal{L}$ contains a~permutation which is not $(c,l)$-close to $\gamma$ and $(\mathcal{C})^k$
does not contain such permutations, we get that $\mathcal{L} \nsubseteq (\mathcal{C})^k$, achieving contradiction.
\end{proof}
Preceding results and Lemma \ref{lemma_symcomp} imply the following corollary.
\begin{cor}
The classes of co-layered permutations $\mathcal{L}_2^c$, $\mathcal{L}_3^c$ and $\mathcal{L}^c$ are not composable.
\end{cor}
\section{Other results}
\label{ch5}
In the final section of this work we collect several miscellaneous results concerning composability.
First we provide more examples of composable classes, and then we fininish by presenting
several additional examples of uncomposable classes.
\subsection{Composable principal classes}
In this section, we use results of Section \ref{ch3} and of \cite{JelinekValtr13}
to prove that many classes avoiding a~single decomposable pattern (a permutation which can
be written as a~non-trivial sum of smaller permutations) are composable.
We will base our proof on the following splittability result of Jelínek and Valtr \cite{JelinekValtr13}.
\begin{lemma}[Jelínek, Valtr \cite{JelinekValtr13}]\label{splitlemma}
Let $\alpha, \beta, \gamma$ be three nonempty permutations and let $\pi \in \textnormal{Av}(\alpha \oplus \beta \oplus \gamma)$.
Then $\pi$ can be merged from two sequences $(a)_{i=1}^n$ and $(c)_{i=1}^m$ such that $a$ avoids $\alpha \oplus \beta$,
$c$ avoids $\beta \oplus \gamma$ and for any $i \in [n]$ and $j \in [m]$ either $\pi^{-1}(a_i) < \pi^{-1}(c_j)$ or $a_i < c_j$.
\end{lemma}
\begin{thm}\label{thm_52}
If $\alpha$ and $\gamma$ are any non-empty permutations and $\beta = \delta_n$ for a~positive integer $n$, then
$$\textnormal{Av}(\alpha \oplus \beta \oplus \gamma) \subseteq (\mathcal{V}(\textnormal{Av}(\alpha \oplus \beta),\textnormal{Av}(\beta\oplus \gamma))
\cap \textnormal{Av}(\alpha \oplus \beta \oplus \gamma)) \circ \mathcal{H}.$$
In particular, $\textnormal{Av}(\alpha \oplus \delta_n \oplus \gamma)$ is 2-composable whenever $\alpha \oplus \delta_n \oplus \gamma
\notin \mathcal{H}$.
\end{thm}
\begin{proof}
Let $\mathcal{C} = \textnormal{Av}(\alpha \oplus \beta \oplus \gamma)$, $\mathcal{A} = \textnormal{Av}(\alpha \oplus \beta)$ and $\mathcal{B} = \textnormal{Av}(\beta \oplus \gamma)$.
Lemma \ref{splitlemma} and Corollary \ref{cor_splitcompose} immediately imply that
$\mathcal{C} \subseteq \mathcal{V}(\mathcal{A},\mathcal{B}) \circ \mathcal{H}.$
Let $\pi \in \mathcal{C}$ be merged from sequences $a$ and $c$ as in Lemma \ref{splitlemma}
and let $\sigma = ac$. We have to show that $\sigma \in \mathcal{C}$. Suppose for a~contradiction that $\sigma$
contains a~copy of $\alpha \oplus \beta \oplus \gamma$. Let $b$ be the decreasing subsequence of $\sigma$
representing the occurrence of $\beta$. Then $b$ cannot be
contained entirely in $a$ or in $c$ since that would create a~copy of $\alpha \oplus \beta$ in $a$
or of $\beta \oplus \gamma$ in $c$. Thus if $\beta = 1$ the contradiction is reached immediately.
If $|\beta| > 1$, we would like to show that $b$ is also a~subsequence of $\pi$. Assume it is not,
therefore there are elements $b_i$ and $b_j$ with $i < j$ such that they appear in reverse
order in $\pi$. That can only be achieved if $b_i$ is in $a$ and $b_j$ is in $c$,
which together with $b_i > b_j$ contradicts the properties of $a$ and $c$ from Lemma \ref{splitlemma}.
It follows that the entire occurrence of $\alpha \oplus \beta \oplus \gamma$ is also contained in $\pi$,
which is a~contradiction, thus $\mathcal{C} \subseteq (\mathcal{V}(\mathcal{A}, \mathcal{B}) \cap \mathcal{C}) \circ \mathcal{H}$.
To prove that $\mathcal{C}$ is really 2-composable for $\alpha \oplus \beta \oplus \gamma \notin \mathcal{H}$ it remains
to verify that $\mathcal{V}(\mathcal{A},\mathcal{B}) \cap \mathcal{C}$ and $\mathcal{H}$ are proper subclasses of $\mathcal{C}$. Clearly
$\mathcal{V}(\mathcal{A},\mathcal{B}) \cap \mathcal{C} \subseteq \mathcal{C}$ and the condition $\alpha \oplus \beta \oplus \gamma \notin \mathcal{H}$ implies
$\mathcal{H} \subseteq \mathcal{C}$, so it remains to show that the inclusions are proper.
Consider the permutation $(\alpha \oplus \beta) \ominus (\beta \oplus \gamma)$ which is clearly in $\mathcal{C}$
and not in $\mathcal{V}(\mathcal{A},\mathcal{B})$. For the class $\mathcal{H}$ we use the results of Atkinson, who showed in \cite[Proposition 3.4]{Atkinson99}
that the class $\mathcal{H}$ has
a basis of size 3~and therefore it cannot be equal to a~principal class.
\end{proof}
Note that for the case $\beta = 1$ we get
\begin{equation*} \mathcal{V}(\textnormal{Av}(\alpha \oplus 1), \textnormal{Av}(1 \oplus \gamma)) \subsetneq \textnormal{Av}(\alpha \oplus 1 \oplus \gamma), \end{equation*}
and thus we may omit the intersection with $\textnormal{Av}(\alpha \oplus 1 \oplus \gamma)$ in the formula of Theorem \ref{thm_52}.
Indeed, if a~permutation is concatenated of two parts, first avoiding $\alpha \oplus 1$ and the second avoiding $1 \oplus \gamma$,
such a~permutation cannot contain an occurrence of $\alpha \oplus 1 \oplus \gamma$ since one of the two parts would contain
the middle 1~and thus the forbidden pattern.
\subsection{More uncomposable classes}
So far we have used classes such as $\mathcal{V}$ or $\mathcal{H}$ to prove that other classes are composable.
In this section, we will show that these classes, and classes similar to them, are themselves uncomposable.
We call a~permutation $\eta \in \mathcal{H}$ \emph{alternating} if $\eta(2i-1) < \eta(2i) > \eta(2i+1)$ for all
possible values of $i$. We will use the following simple observation about alternating permutations in $\mathcal{H}$.
\begin{figure}[ht]
\centering
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.5cm,y=0.5cm]
\clip(-2.5,-3.09) rectangle (5.52,4.99);
\draw (-2,4.5)-- (-2,-2.5);
\draw (-2,-2.5)-- (5,-2.5);
\draw (5,-2.5)-- (5,4.5);
\draw (5,4.5)-- (-2,4.5);
\draw (-2,1.5)-- (5,1.5);
\begin{scriptsize}
\fill [color=black] (-1.5,-2) circle (1.5pt);
\fill [color=black] (-0.5,2) circle (1.5pt);
\fill [color=black] (0.5,-1) circle (1.5pt);
\fill [color=black] (1.5,3) circle (1.5pt);
\fill [color=black] (2.5,0) circle (1.5pt);
\fill [color=black] (3.5,4) circle (1.5pt);
\fill [color=black] (4.5,1) circle (1.5pt);
\end{scriptsize}
\end{tikzpicture}
\caption{The alternating permutation of length 7}
\end{figure}
\begin{obs}\label{obs_alt}
Every permutation from $\mathcal{H}$ is contained in an alternating permutation from $\mathcal{H}$.
\end{obs}
\begin{prop} \label{VHstuff}
The classes $\mathcal{V}$, $\mathcal{V}^c$, $\mathcal{V}(\mathcal{D}, \mathcal{I})$, $\mathcal{V}(\mathcal{I}, \mathcal{D})$, $\mathcal{H}$, $\mathcal{H}^c$, $\mathcal{H}(\mathcal{I}, \mathcal{D})$ \\
and $\mathcal{H}(\mathcal{D}, \mathcal{I})$ are not composable.
\end{prop}
\begin{proof}
We will show the proof for the class $\mathcal{H}$, the same approach can be applied to every mentioned
horizontal merge and the result is transferred by inversion to the vertical merges by to Corollary \ref{cor_inversion}.
Suppose that $\mathcal{H}$ is composable from its proper subclasses $\mathcal{C}_1, \ldots, \mathcal{C}_k$. Each of $\mathcal{C}_1, \ldots, \mathcal{C}_k$
avoids a permutation of $\mathcal{H}$, thus according to Observation \ref{obs_alt} there is an alternating
permutation $\eta \in \mathcal{H}$ such that $\mathcal{C}_i \subseteq \textnormal{Av}(\eta)$ and therefore if $\mathcal{C} = \textnormal{Av}(\eta) \cap \mathcal{H} \subsetneq \mathcal{H}$
we have $\mathcal{H} \subseteq (\mathcal{C})^k$.
Any permutation $\pi \in \mathcal{C}$ is merged from two sequences $a$ and $b$ of consecutive integers.
We label elements of $\pi$ by $a$ or $b$ depending on which sequence they belong to.
A sequence of elements with alternating labels forms a~copy of an alternating permutation in $\pi$.
The length of the longest sequence of alternating labels in $\pi$ is thus limited by a~constant $N$ determined
by the order of $\eta$, thus $\pi$ can be broken into at most $N$ contiguous parts each having one label.
Since elements labeled with a~single label form a~sequence of consecutive integers, this
implies that $\pi$ is in fact an $N$-block. Since the choice of $\pi$ was arbitrary,
every permutation of $\mathcal{C}$ is an $N$-block and by Lemma \ref{lemma_blocks} every permutation of $(\mathcal{C})^k$
is an $(N^k)$-block. But a~long enough alternating permutation from $\mathcal{H}$ is not an $(N^k)$-block,
therefore $\mathcal{H} \nsubseteq (\mathcal{C})^k$ and the proof is finished.
\end{proof}
\section{Conclusion}
This paper studies the previously unexplored concept of composability of permutation classes.
Given a~permutation class, our main goal is to show, how it can be constructed using smaller permutation classes
and the composition operator, or to prove that this cannot be done. Throughout the paper,
we present both types of results.
On the positive side, Theorems \ref{thm_Ik} and \ref{thm_k+l-1}
show two distinct ways of constructing the class $\textnormal{Av}(k\cdots 21)$,
Theorem \ref{l4} provides infinitely many examples of
classes of layered patterns which can be constructed from simpler subclasses
and Theorem \ref{thm_52} shows that many principal classes
avoiding a~decomposable pattern are composable.
On the negative side, in Theorems \ref{l2}, \ref{f2}, \ref{l3} and \ref{l}
we present four different classes of layered patterns which cannot be constructed from any number
of proper subclasses using composition, and Proposition \ref{VHstuff} provides us with
8 more examples of uncomposable classes.
Composability is similar to splittability in that both these properties describe
how a~bigger class is built from smaller ones. We do not know whether these
two properties are somehow connected; however, our research suggests that this may be the case,
since every composable class we have found so far is also splittable. We have
found examples of splittable yet uncomposable classes, namely the classes $\mathcal{L}_2$ and $\mathcal{L}_3$
introduced in Section \ref{ch4}. The class of all layered permutations is an example of a~both uncomposable
and unsplittable class. The last case remains open and we pose it as a~question for future work.
\begin{question}
Is there a~permutation class which is composable and unsplittable?
\end{question}
In Section 3 we showed that if a class is composable and avoids an increasing pattern, then its reverse
and complement are composable. It remains open whether the avoidance condition is necessary.
\begin{question}
Is there a composable class $\mathcal{A}$ such that $\mathcal{A}^r$ or $\mathcal{A}^c$ is not composable?
\end{question}
Splittability has the property that if a class is splittable, it can be split into two parts. This is not the case
for composability as we showed in
Section 4 where we proved that the class $\mathcal{L}_k$ for $k \geq 4$ is $3$-composable but not 2-composable.
\begin{question}
Is there a $4$-composable class which is not $3$-composable? More generally,
is there a universal constant $K$ such that every composable class is $K$-composable?
\end{question}
Our work may find applications in enumerating permutation classes. Denote by $\textnormal{gr}(\mathcal{A})$ the growth
rate of the class $\mathcal{A}$ as defined e.g. in \cite{Vatter15}.
It is not difficult to see that if $\mathcal{C} \subseteq \mathcal{A} \circ \mathcal{B}$, then $\textnormal{gr}(\mathcal{C}) \leq \textnormal{gr}(\mathcal{A}) \cdot \textnormal{gr}(\mathcal{B})$.
Using this observation one could try to find upper bounds for growth rates of composable permutation classes.
|
\section{Introduction}
Machine learning systems have been successful in many domains, from computer vision \citep{img12} to
speech recognition \citep{speech-nn} and machine translation \citep{sutskever14,bahdanau2014neural,cho2014learning}.
Neural machine translation (NMT) is so successful that for some language pairs it approaches,
\emph{on average}, the quality of human translators \citep{gnmt}.
The words \emph{on average} are crucial though. When a sentence resembles
one from the abundant training data, the translation will be accurate.
However, when encountering a rare word such as \emph{Dostoevsky} (in German,
\emph{Dostojewski}), many models will fail.
The correct German translation of \emph{Dostoevsky} does not appear enough times
in the training data for the model to sufficiently learn its translation.
While more example sentences concerning the famous Russian author might eventually be added to the training data, there
are many other rare words or rare events of other kinds.
This illustrates a general problem with current deep learning models: it is necessary
to extend the training data and re-train them to handle such rare or new events.
Humans, on the other hand, learn in a life-long fashion, often from single examples.
We present a life-long memory module that enables one-shot learning
in a variety of neural networks. Our memory module consists of key-value pairs.
Keys are activations of a chosen layer of a neural network, and values are
the ground-truth targets for the given example.
This way, as the network is trained, its memory increases and becomes more useful.
Eventually it can give predictions that leverage on knowledge from past data with
similar activations. Given a new example, the network writes it to memory
and is able to use it afterwards, even if the example was presented just once.
There are many advantages of having a long-term memory. One-shot learning
is a desirable property in its own right, and some tasks, as we will show
below, are simply not solvable without it. Even real-world tasks where we
have large training sets, such as translation, can benefit from long-term
memory. Finally, since the memory can be traced back to training examples,
it might help explain the decisions that the model
is making and thus improve understandability of the model.
It is not immediately clear how to measure the performance of
a life-long one-shot learning model, since most deep learning
evaluations focus on the average performance and do not have
a one-shot component. We therefore evaluate in a few ways,
to show that our memory module indeed works:
\begin{enumerate}[(1)]
\item We evaluate on the well-known one-shot learning task Omniglot,
which is the only dataset with explicit one-shot learning evaluation.
This dataset is small and does not benefit from life-long learning
capability of our module, but we still exceed the best previous results and set new state-of-the-art.
\item We devise a synthetic task that requires life-long one-shot learning.
On this task, standard models fare poorly while our model can solve
it well, demonstrating its strengths.
\item Finally, we train an English-German translation model that has our
life-long one-shot learning module. It retains very good performance
on average and is also capable of one-shot learning. On the qualitative side,
we find that it can translate rarely-occurring words like Dostoevsky.
On the quantitative side, we see that the BLEU score for the generated translations
can be significantly increased by showing it related translations before evaluating.
\end{enumerate}
\section{Memory Module}
Our memory consists of a matrix $K$ of memory keys, a vector $V$ of memory values,
and an additional vector $A$ that tracks the age of items stored in memory.
Keys can be arbitrary vectors of size \texttt{key-size}, and
we assume that the memory values are single integers representing
a class or token ID.
We define a memory of size \texttt{memory-size} as a triple:
\[ \mathcal{M} = (K_{\text{\texttt{memory-size}} \times \text{\texttt{key-size}}},\
V_{\text{\texttt{memory-size}}},\ A_{\text{\texttt{memory-size}}}). \]
A memory query is a vector of size \texttt{key-size} which we assume
to be normalized, i.e., $\norm{q} = 1$. Given a query $q$, we define the nearest
neighbor of $q$ in $\mathcal{M}$ as any of the keys that maximize the dot product with $q$:
\[ \text{NN}(q, \mathcal{M}) = \argmax_{i}\ q \cdot K[i]. \]
Since the keys are normalized, the above notion
corresponds to the nearest neighbor with respect to cosine similarity.
We will also use the natural extension of it to $k$ nearest
neighbors, which we denote $\text{NN}_k(q, \mathcal{M})$. In our
experiments we always used the set of $k=256$ nearest neighbors.
When given a query $q$, the memory $\mathcal{M} = (K, V, A)$ will compute
$k$ nearest neighbors (sorted by decreasing cosine similarity):
\[ (n_1, \dots, n_k) = \text{NN}_k(q, \mathcal{M}) \]
and return, as the main result, the value $V[n_1]$.
Additionally, we will compute the cosine similarities $d_i = q \cdot K[n_i]$
and return $\text{softmax}(d_1 \cdot t, \dots, d_k \cdot t)$.
The parameter $t$ denotes the inverse of softmax temperature and we set it to $t=40$ in our experiments. In models where the memory output
is again embedded into a dense vector, we multiply the embedded output
by the corresponding softmax component so as to provide a signal
about confidence of the memory.
The forward computation of the memory module is thus very simple,
the only interesting part being how to compute nearest neighbors efficiently,
which we discuss below. But we must also answer the question how the memory is trained.
\paragraph{Memory Loss.}
Assume now that in addition to a query $q$ we are also given the correct desired (supervised) value $v$.
In the case of classification, this $v$ would be the class label.
In a sequence-to-sequence task, $v$ would be the desired output token of the
current time step.
After computing the $k$ nearest neighbors $(n_1, \dots, n_k)$ as above,
let $p$ be the smallest index such that $V[n_p] = v$ and $b$ the smallest
index such that $V[n_b] \neq v$. We call $n_p$ the \emph{positive neighbor} and $n_b$ the \emph{negative neighbor}.
When no positive neighbor is among the top-$k$, we pick any vector from memory with value $v$ instead of $K[n_p]$. We define the memory loss as:
\[ \text{loss}(q,v,\mathcal{M}) = \left[ q \cdot K[n_b] - q \cdot K[n_p] + \alpha\right]_+. \]
Recall that both $q$ and the keys in memory are normalized,
so the products in the above loss term correspond to cosine similarities
between $q$, the positive key, and the negative key. Since cosine similarity
is maximal for equal terms, we want to maximize the similarity to the positive key
and minimize the similarity to the negative one. But once they are far enough apart
(by the margin $\alpha$, $0.1$ in all our experiments), we
do not propagate any loss. This definition and reasoning behind it are almost
identical to the one in \cite{facenet} and similar to many other distance
metric learning works \citep{weinberger, wsabie}.
\paragraph{Memory Update.}
In addition to computing the loss, we will also update the memory $\mathcal{M}$
to account for the fact that the newly presented query $q$ corresponds to $v$.
The update is done in a different way depending on whether the main value
returned by the memory module already is the correct value $v$ or not.
As before, let $n_1 = \text{NN}(q, \mathcal{M})$ be the nearest neighbor to $q$.
If the memory already returns the correct value, i.e., if $V[n_1] = v$,
then we only update the key for $n_1$ by taking the average of
the current key and $q$ and normalizing it:
\[ K[n_1] \leftarrow \frac{q + K[n_1]}{\norm{q + K[n_1]}}. \]
When doing this, we also re-set the age: $A[n_1] \leftarrow 0$.
Otherwise, when $V[n_1] \neq v$, we find a new place in the memory
and write the pair $(q, v)$ there. Which place should we choose? We find
memory items with maximum age, and write to one of those (randomly chosen).
More formally, we pick $n' = \argmax_{i} A[i]+r_i$ where $|r_i|\ll|\mathcal{M}|$ is
a random number that introduces some randomness in the choice so as to
avoid race conditions in asynchronous multi-replica training. We then set:
\[ K[n'] \leftarrow q, \quad V[n'] \leftarrow v, \quad A[n'] \leftarrow 0. \]
With every memory update we also increment the age of all non-updated indices by $1$.
The full operation of the memory module is depicted in Figure~\ref{fig:mem}.
\begin{figure}
\begin{center}
\begin{tikzpicture}[yscale=1.0,xscale=0.68] \small
\node (K) at (-0.5, 0.5) {$K$};
\draw (0, 0) rectangle (8.0, 1.0);
\node (q) at (4.0, -1.0) {$q$};
\draw[dotted] (4.0, -0.8) -- (1.5, 0.0);
\draw[dotted] (4.0, -0.8) -- (3.5, 0.0);
\draw[dotted] (4.0, -0.8) -- (5.5, 0.0);
\draw[dotted] (4.0, -0.8) -- (7.5, 0.0);
\draw[fill=gray!50] (1.0, 0.0) rectangle (2.0, 1.0);
\node (k1) at (1.5, 0.5) {$k_1$};
\draw (1.5, -0.1) -- (1.5, 0.1);
\draw (1.5, 0.9) -- (1.5, 1.1);
\node (n1) at (1.5, 1.5) {$n_1$};
\draw (1.5, 1.9) -- (1.5, 2.1);
\draw[fill=gray!50] (5.0, 0.0) rectangle (6.0, 1.0);
\node (kb) at (5.5, 0.5) {$k_b$};
\draw (5.5, -0.1) -- (5.5, 0.1);
\draw (5.5, 0.9) -- (5.5, 1.1);
\node (nn) at (5.5, 1.5) {$n_b$};
\draw (5.5, 1.9) -- (5.5, 2.1);
\node (V) at (-0.5, 2.0) {$V$};
\draw[thick] (0, 2.0) -- (8.0, 2.0);
\node (v1) at (1.5, 2.5) {$V[n_1]$};
\node (vb) at (5.7, 2.5) {$V[n_b] \neq v$};
\node[anchor=west] (case) at (-1.0, 4.0) {Case 1: $V[n_1] = v$;
$\quad$ Loss = $\left[q \cdot k_b - q \cdot k_1 + \alpha\right]_+$};
\node[anchor=west] (upd) at (-0.5, 3.5) {Update:
$K[n_1] \leftarrow \frac{q+k_1}{\norm{q+k_1}}$
$\quad A[n_1] \leftarrow 0$};
\begin{scope}[xshift=11cm]
\node (K) at (-0.5, 0.5) {$K$};
\draw (0, 0) rectangle (8.0, 1.0);
\node (q) at (4.0, -1.0) {$q$};
\draw[dotted] (4.0, -0.8) -- (1.5, 0.0);
\draw[dotted] (4.0, -0.8) -- (3.5, 0.0);
\draw[dotted] (4.0, -0.8) -- (5.5, 0.0);
\draw[dotted] (4.0, -0.8) -- (7.5, 0.0);
\draw[fill=gray!50] (1.0, 0.0) rectangle (2.0, 1.0);
\node (k1) at (1.5, 0.5) {$k_1$};
\draw (1.5, -0.1) -- (1.5, 0.1);
\draw (1.5, 0.9) -- (1.5, 1.1);
\node (n1) at (1.5, 1.5) {$n_1$};
\draw (1.5, 1.9) -- (1.5, 2.1);
\draw[fill=gray!50] (5.0, 0.0) rectangle (6.0, 1.0);
\node (kb) at (5.5, 0.5) {$k_p$};
\draw (5.5, -0.1) -- (5.5, 0.1);
\draw (5.5, 0.9) -- (5.5, 1.1);
\node (nn) at (5.5, 1.5) {$n_p$};
\draw (5.5, 1.9) -- (5.5, 2.1);
\node (V) at (-0.5, 2.0) {$V$};
\draw[thick] (0, 2.0) -- (8.0, 2.0);
\node (v1) at (1.5, 2.5) {$V[n_1]$};
\node (vb) at (5.7, 2.5) {$V[n_p] = v$};
\node[anchor=west] (case) at (-1.0, 4.0) {Case 2: $V[n_1] \neq v$;
$\quad$ Loss = $\left[q \cdot k_1 - q \cdot k_p + \alpha\right]_+$};
\node[anchor=west] (upd) at (-0.5, 3.5) {Update:
$K[n'] \leftarrow q$ $\quad V[n'] \leftarrow v$
$\quad A[n'] \leftarrow 0$};
\end{scope}
\end{tikzpicture}
\end{center}
\caption{The operation of the memory module on a query $q$ with correct value $v$;
see text for details.}
\label{fig:mem}
\end{figure}
\paragraph{Efficient nearest neighbor computation.}
The most expensive operation in our memory module is
the computation of $k$ nearest neighbors.
This can be done exactly or in an approximate way.
In the exact mode, to calculate the nearest neighbors in $K$ to a mini-batch
of queries $Q = (q_1, \dots, q_b)$, we perform a single matrix multiplication:
$Q \times K^T$. This multiplies the \texttt{batch-size} $\times$ \texttt{key-size}
matrix $Q$ by the \texttt{key-size} $\times$ \texttt{memory-size} matrix $K^T$,
and the result is the \texttt{batch-size} $\times$ \texttt{memory-size} matrix
of all distances, from which we can choose the top-$k$. This procedure is
linear in \texttt{memory-size}, so it can be expensive for very large memory
sizes. But matrix multiplication is very heavily optimized, so in our experiments
on GPUs we find that this operation is not a bottleneck for memory sizes up to
half a million.
If the exact mode is too slow, the $k$ nearest neighbors can be computed
approximately using locality sensitive hashing (LSH). LSH is a hashing
scheme so that near neighbors get similar hashes \citep{indyk1998approximate, lsh}.
For cosine similarity, the computation of
an LSH is very simple. We pick a number of random normalized hash vectors
$h_1, \dots, h_l$. The hash of a query $q$ is a sequence of $l$ bits,
$b_1, \dots, b_l$, such that $b_i = 1$ if, and only if, $q \cdot h_i > 0$.
It turns out that near neighbors will, with high probability, have
a large number of identical bits in their hash. To compute the nearest
neighbors it is therefore sufficient to only look into parts of the memory
with similar hashes. This makes the nearest neighbor computation work in
approximately constant time -- we only need to multiply the query by the hash
vectors, and then only use the nearest buckets.
\subsection{Using the Memory Module}
The memory module presented above can be added to any classification network.
There are two main choices: which layer to use to generate queries,
and how to use the output of the module.
In the simplest case, we use the final layer of a network as query and
the output of the module is directly used for classification. This simplest
case is similar to matching networks \citep{matching_nets}
and our memory module yields good results already
in this setting (see below).
Instead of using the output of the module directly,
it is possible to embed it again into a dense representation and mix it with
other predictions made by the network. To study this setting,
we add the memory module to sequence-to-sequence recurrent neural networks.
As described in detail below, a query to memory is made in every step
of the decoder network. Memory output is embedded again into a dense
representation and combined with inputs from other layers of the network.
\paragraph{Convolutional Network with Memory.}
To test our memory module in a simple setting, we first add it to
a basic convolutional network network for image classification.
Our network consists of two convolutional layers with ReLU non-linearity,
followed by a max-pooling layer, another two convolutional-ReLU layers,
another max-pooling, and two fully connected layers. All convolutions use $3\times 3$
filters with $64$ channels in the first pair, and $128$ in the second.
The fully connected layers have dimension $256$ and dropout applied between them.
The output of the final layer is used as query to our memory module
and the nearest neighbor returned by the memory is used as the final network prediction.
Even this basic architecture yields good results in one-shot learning,
as discussed below.
\begin{figure}
\begin{center}
\includegraphics[width=0.75\textwidth,height=5cm]{bnmt.pdf}
\end{center}
\caption{The GNMT model with added memory module.
On each decoding step $t$, the result of the attention $a_t$ is used to
query the memory. The resulting value is combined with the output of the
final LSTM layer to produce the predicted logits $\hat{y}_t$. See text for further details.}
\label{fig:gnmt}
\end{figure}
\paragraph{Sequence-to-sequence with Memory.}
For large-scale experiments, we add the memory module into a large
sequence-to-sequence model. Such sequence-to-sequence recurrent neural
networks (RNNs) with long short-term memory (LSTM) cells \citep{hochreiter1997}
have proven especially successful at natural language processing (NLP) tasks,
including machine translation \citep{sutskever14,bahdanau2014neural,cho2014learning}.
We add the memory module to the Google Neural Machine Translation (GNMT) model
\citep{gnmt}. This model consists of an encoder RNN, which creates a representation
of the source language sentence, and a decoder RNN that outputs the target language
sentence. We left the encoder RNN unmodified. In the decoder RNN, we use the vector
retrieved by the attention mechanism as query to the memory module.
In the GNMT model, the attention vector is used in all LSTM layers beyond the second one,
so the computation of the other layers and the memory can happen in parallel.
Before the final softmax layer, we combine the embedded memory output with
the output of the final LSTM layer using an additional linear layer,
as depicted in Figure~\ref{fig:gnmt}.
\paragraph{Extended Neural GPU with Memory.}
To test versatility of our memory module, we also add it to the Extended Neural GPU,
a convolutional-recurrent model introduced by \cite{engpu}. The Extended Neural GPU
is a sequence-to-sequence model too, but its decoder is convolutional and the size
of its state changes depending on the size of the input. Again, we leave the encoder
part of the model intact, and extend the decoder part by a memory query.
This time, we use the position one step ahead to query memory, and we put
the embedded result to the output tape, as shown in Figure~\ref{fig:engpu}.
Note that in this model the result of the memory will be processed by two
recurrent-convolutional cells before the corresponding output is produced.
The fact that this model still does one-shot learning confirms that the output
of our memory module can be used deep inside a network, not just near the output layer.
\begin{figure
\begin{center}
\begin{tikzpicture}[yscale=1.0,xscale=0.9] \small
\draw[fill=gray] (0, 0) rectangle (0.2, 3.0);
\node(i1) at (-0.6, 2.9) {$i_1$};
\node (idots) at (-0.6, 1.6) {$\vdots$};
\foreach \y in {0.1, 2.9} {
\draw (-0.35, \y) -- (0.1, \y);
}
\node(in) at (-0.6, 0.1) {$i_n$};
\node (s0) at (0.3, -0.5) {$s_0$};
\cgrupic{CGRU}
\begin{scope}[xshift=1.79cm]
\cgrupic{CGRU}
\end{scope}
\node (s1) at (2.1, -0.5) {$s_1$};
\begin{scope}[xshift=3.59cm]
\cgruhands{CGRU}
\end{scope}
\node (dots) at (3.9, 1.5) {$\dots$};
\begin{scope}[xshift=5.39cm]
\cgrupic{\dcgru}
\end{scope}
\node (d0) at (5.7, -0.5) {$s_n = d_0$};
\draw[fill=gray] (7.2, 3.0) rectangle (7.4, 2.8);
\draw[fill=gray!60] (7.2, 2.8) rectangle (7.4, 2.6);
\begin{scope}[xshift=7.19cm]
\cgrupic{\dcgru}
\end{scope}
\node (d1) at (7.5, -0.5) {$d_1$};
\draw[thick,gray] (6.6, 3.1) -- (6.6, 1.8);
\draw[fill=gray] (9.0, 2.8) rectangle (9.2, 2.6);
\draw[fill=gray!60] (9.0, 2.6) rectangle (9.2, 2.4);
\begin{scope}[xshift=8.99cm]
\cgrupic{\dcgru}
\end{scope}
\node (sp) at (9.3, -0.5) {$d_2$};
\begin{scope}[xshift=10.79cm]
\cgruhands{\dcgru}
\end{scope}
\node (dots) at (11.1, 1.5) {$\dots$};
\draw[fill=gray] (12.6, 0.2) rectangle (12.8, 0.0);
\draw[step=0.2] (12.59, 0) grid (13.2, 3.0);
\node (dn) at (12.9, -0.5) {$d_n$};
\draw (7.5, 3.6) rectangle (8.2, 4.0);
\node (m) at (7.85, 3.8) {$\mathcal{M}$};
\draw[->] (7.3, 2.7) -- (7.7, 3.6);
\draw[->] (7.9, 3.6) -- (8.25, 3.45);
\draw (9.3, 3.6) rectangle (10.0, 4.0);
\node (m) at (9.65, 3.8) {$\mathcal{M}$};
\draw[->] (9.1, 2.5) -- (9.5, 3.6);
\draw[->] (9.7, 3.6) -- (10.05, 3.45);
\draw[->] (7.3, 2.9) -- (7.3, 3.25);
\node(o1) at (7.3, 3.4) {$o_1$};
\draw[->] (9.1, 2.7) -- (9.1, 3.25);
\node(o2) at (9.1, 3.4) {$o_2$};
\node (dots) at (11.1, 3.4) {$\ldots$};
\draw[->] (12.7, 0.1) -- (12.7, 3.25);
\node(o2) at (12.7, 3.4) {$o_n$};
\node(p0) at (6.6, 3.4) {$p_0$};
\draw[thick,gray,->] (6.6, 3.2) -- (6.6, 1.8);
\draw[thick,gray] (6.8, 3.4) -- (7.1, 3.4);
\draw[thick,gray] (7.5, 3.4) -- (8.2, 3.4);
\node(p1) at (8.4, 3.4) {$p_1$};
\draw[thick,gray,->] (8.4, 3.2) -- (8.4, 1.8);
\draw[thick,gray] (8.6, 3.4) -- (8.9, 3.4);
\draw[thick,gray] (9.3, 3.4) -- (10.0, 3.4);
\node(p2) at (10.2, 3.4) {$p_2$};
\draw[thick,gray,->] (10.2, 3.2) -- (10.2, 1.8);
\draw[thick,gray] (10.4, 3.4) -- (10.7, 3.4);
\draw[thick,gray] (11.4, 3.4) -- (11.7, 3.4);
\node(pn) at (12.1, 3.4) {$p_{n-1}$};
\draw[thick,gray,->] (12.0, 3.2) -- (12.0, 1.8);
\end{tikzpicture}
\end{center}
\caption{Extended Neural GPU with memory module.
Memory query is read from the position one below the current
output logit, and the embedded memory value is put at the same
position of the output tape $p$. The network learns to use these
values to produce the output in the next step.}
\label{fig:engpu}
\end{figure}
\section{Related Work}
\paragraph{Memory in Neural Networks.}
Augmenting neural networks with memory has been heavily studied recently.
Many of these approaches design a memory component that is intended as a generalization
of the memory in standard recurrent neural networks. In recurrent networks,
the state passed from one time step to the next can be interpreted as the
network's memory representation of the current example. Moving away from this
fixed-length vector representation of memory to a larger and more versatile form
is at the core of these methods.
Augmenting recurrent neural networks with attention~\citep{bahdanau2014neural} can
be interpreted as creating a large memory component that allows content-based
addressing. More generally, \citet{ntm14} augmented a
recurrent neural network with
a computing-inspired memory component that can be addressed via both
content- and address-based queries. \citet{endend_mem_net} present
a similar augmentation and show the importance of allowing
multiple reads and writes to memory between inputs. These approaches
excel at tasks where it is necessary to store large parts of a
sequential input in a representation that can later be precisely queried.
Such tasks include algorithmic sequence manipulation tasks, natural language
modelling, and question-answering tasks.
The success of these approaches hinges on making the memory component fully
differentiable and backpropagating signal through every access of memory.
In this setting, computational requirements necessitate that the memory be small.
Some attempts have been made at making hard access queries to memory~\citep{reinforceNTM15, xuetal2015}, but it was usually
challenging to match the soft version. Recently, more successful
training for hard queries was reported \citep{hardntm} that makes
use of a curriculum strategy that mixes soft and hard queries at
training time.
Our approach applies hard access as well,
but we encourage the model to make good queries via a special memory loss.
Modifications to allow for large-scale memory in neural networks have been
proposed. The original implementation of memory networks~\citep{mem_nets}
and later work on scaling it~\citep{large_mem_nets, hier_mem_nets}
used memory with size in the millions. The cost of doing so is that the
memory must be fixed prior to training. Moreover, since during the beginning
of training the model is unlikely to query the memory correctly, strong supervision
is used to encourage the model to query memory locations that are useful.
These hints are either given as additional supervising information by the
task or determined heuristically as in \citet{goldilocks}.
All the work discussed so far has either used a memory that is fixed before training
or used a memory that is not persistent between different examples.
For one-shot and lifelong learning, a memory must necessarily be both volatile
during training and persistent between examples. To bridge this gap, \citet{santoro16} propose to partition
training into distinct episodes consisting of a sequence of labelled
examples $\{(x_i, y_i)\}_{i=1}^n$. A network augmented
with a fully-differentiable memory is trained to predict $y_i$ given
the previous sequence $(x_1, y_1, \dots, x_{i-1})$. This way,
the model learns to store important examples with their
corresponding labels in memory and later re-use this information to correctly classify new examples.
This model successfully exhibits one-shot learning on Omniglot.
However, this approach again requires fully-differentiable memory access and thus limits
the size of the memory as well as the length of an episode.
This restriction has recently been alleviated by \citet{jack_rae}.
Their model can utilize large memories, but unlike our work
does not have an explicit cost to guide the formation of memory keys.
For classification tasks
like Omniglot, it is easy to construct short episodes so that they include
a few examples from each of several classes. However, this becomes harder as
the output becomes richer. For example, in the difficult
sequence-to-sequence tasks which we consider,
it is hard to determine which examples would be helpful for correctly
predicting others \emph{a priori}, and so constructing short episodes each
containing examples that are similar and act as hints to each other is intractable.
\paragraph{One-shot Learning.}
While the recent work of \citet{santoro16} succeeded in bridging the gap between
memory-based models and one-shot learning, the field of one-shot learning
has seen a variety of different approaches over time.
Early work utilized Bayesian methods to model data generatively~\citep{feifei, oneshot}.
The paper that introduced the Omniglot dataset~\citep{oneshot} approached
the task with a generative model for strokes. This way, given a
single character image, the probability of a different image
being of the same character may be approximated via standard techniques.
One early neural network approach to one-shot learning was given by Siamese networks~\citep{siamese}.
When our approach is applied to the Omniglot image classification dataset,
the resulting training algorithm is actually similar to that of Siamese networks.
The only difference is in the loss function: Siamese networks utilize a
cross-entropy loss whereas our method uses a margin triplet loss.
A more sophisticated neural network approach is given by~\citet{matching}.
The strengths of this approach are (1) the model architecture utilizes
recent advances in attention-augmented neural networks for set-to-set
learning~\citep{set2set}, and (2) the training algorithm is designed
to exactly match the testing phase (given $k$ distinct images and an
additional image, the model must predict which of the $k$ images
is of the same class as the additional image).
This approach may also be considered as a generalization of previous work
on metric learning.
\section{Experiments}
We perform experiments using all three architectures described above.
We experiment both on real-world data and on synthetic tasks that give
us some insight into the performance and limitations of the memory module.
In all our experiments we use the Adam optimizer \citep{adam} and the
parameters for the memory module remain unchanged ($k=256, \alpha=0.1$).
Good performance with a single set of parameters shows the versatility
of our memory module. The source code for the memory module, together
with our settings for Omniglot, is available on github\footnote{
\url{https://github.com/tensorflow/models/tree/master/learning_to_remember_rare_events}}.
\paragraph{Omniglot.}
The Omniglot dataset~\citep{oneshot} consists of 1623 characters from
50 different alphabets, each hand-drawn by 20 different people.
The large number of classes (characters) with relatively few data per class
(20), makes this an ideal data set for testing one-shot classification.
In the $N$-way Omniglot task setup we pick $N$ unseen character classes,
independent of alphabet. We provide the model with one drawing of each character
and measure its accuracy the $K$-th time it sees the character class.
Our setup is identical to \cite{matching_nets}, so we also augmented
the data set with random rotations by multiples of $90$ degrees and use $1200$
characters for training, and the remaining character classes for evaluation.
We present the results from \cite{matching_nets} and ours in Table~\ref{tab:omniglot}.
Even with a simpler network without batch normalization, we get similar results.
\begin{table
\caption{Results on the Omniglot dataset. Although our model uses only a simple
convolutional neural network, the addition of our memory module allows it to approach
much more complex models on 1-shot and multi-shot learning tasks.}
\begin{center}
\begin{tabular}{|l||c|c|c|c|}
\hline
{\bf Model} & 5-way 1-shot & 5-way 5-shot & 20-way 1-shot & 20-way 5-shot \\ \hline
Pixels Nearest Neighbor & 41.7\% & 63.2\% & 26.7\% & 42.6\% \\
MANN (no convolutions) & 82.8\% & 94.9\% & -- & -- \\
Convolutional Siamese Net & 96.7\% & 98.4\% & 88.0\% & 96.5\% \\
Matching Network & 98.1\% & 98.9\% & 93.8\% & 98.5\% \\
\hline
ConvNet with Memory Module & 98.4\% & 99.6\% & 95.0\% & 98.6\% \\
\hline
\end{tabular}
\end{center}
\label{tab:omniglot}
\end{table}
\paragraph{Synthetic task.}
To better understand the memory module operation and to test what
it can remember, we devise a synthetic task
and train the Extended Neural GPU with and without memory
(we use a small Extended Neural GPU with $32$ channels and memory of size half a million).
To create training and test data for our synthetic task,
we use symbols from the set $S = \{2, \dots, 16000\}$
and first fix a random function $f : S \to S$. The function $f$ is
chosen at random, but fixed and the same for all training and testing examples (we used $40$K training examples).
In our synthetic task, the input is a sequence consisting of
As and Bs with one continuous substring of $7$ digits from
the set $0,1,2,3$. The substring is interpreted as a number
written in base-$4$, e.g., $1982 = 132332_4$, so the string
$132332$ would be interpreted as $1982$. The corresponding
output is created by copying all As and Bs, but mapping
the number through the random function $f$. For instance,
assuming $f(1982) = 3726$, the output corresponding to
$132332$ would be $322032$ as $3726 = 322032_4$.
Here is an example of an input-output pair:
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
{\bf Input} & A & 0 & 1 & 3 & 2 & 3 & 3 & 2 & B & A & B & A & B \\ \hline
{\bf Output} & A & 0 & 3 & 2 & 2 & 0 & 3 & 2 & B & A & B & A & B \\ \hline
\end{tabular}
\end{center}
This task clearly requires memory to store the fixed random function.
Since there are 16K elements to learn, it is hard to memorize,
and each single instance occurs quite rarely.
The raw Extended Neural GPU (or any other sequence-to-sequence model)
are limited by their size. With long training, the small model can
memorize some of the sequences, but it is only a small fraction.
Additionally, there is no direct indication in the data what part
of the input should trigger the production of each output symbol.
For example, to produce the first $3$ output in the above example,
the memory key needs to encode all base-$4$ symbols from the input.
Not just one or two aligned symbols, but a number of them.
Moreover, it should not encode more symbols or it will not generalize
to the test set. Similarly, a basic nearest neighbor classifier
fails on this task. We use sequences of length up to $40$ during
training, but there are only $7$ relevant symbols. The simple
nearest neighbor by Hamming distance will most probably select
some sequence with similar prefix or suffix of As and Bs,
and not the one with the corresponding base-$4$ part.
We also trained a large sequence-to-sequence model with attention
on this task (a 2-layer LSTM model with 256 units in each layer).
This model can memorize the whole training set, but it suffers from
a similar problem as the Hamming nearest neighbor -- it almost doesn't
generalize, its accuracy on the test set is only about $1\%$.
The same model with a memory module generalizes much better,
reaching over $30\%$ accuracy. The Extended Neural GPU with our memory module yields even better results, see Table~\ref{tab:synthetic}.
\begin{table
\caption{Results on the synthetic task.
We report the percentage of fully correct sequences
from the test set, which contains 10000 random examples.
See text for details.}
\begin{center}
\begin{tabular}{|l||c|c|}
\hline
{\bf Model} & Accuracy \\ \hline
Hamming Nearest Neighbor & 0.1\% \\
Baseline Sequence-to-Sequence with Attention & 0.9\% \\
Baseline Extended Neural GPU & 12.2\% \\
\hline
Sequence-to-Sequence with Attention and Memory & 35.2\% \\
Extended Neural GPU with Memory Module & 71.3\% \\
\hline
\end{tabular}
\end{center}
\label{tab:synthetic}
\end{table}
\paragraph{Translation.}
To evaluate the memory module in a large-scale setting we use the GNMT model~\citep{gnmt}
extended with our memory module on the WMT14 English-to-German translation task.
We evaluate the model both qualitatively and quantitatively.
On the qualitative side, we note that our memory-augmented
model can successfully translate rare words like Dostoevsky, unlike the baseline model
which predicts an identity-mapped \emph{Dostoevsky} for the
German translation of Dostoevsky.
On the quantitative side, we use the WMT test set.
We find that in terms of BLEU score, an aggregate measure,
the memory-augmented GNMT is on par
with the baseline GNMT, see Table~\ref{tab:translate}.
To evaluate our memory-augmented model for
one-shot capabilities we split the test set in two.
We take the even
lines of the test set (index starting at 0) as a context set and the
odd lines of the test set as the one-shot evaluation set.
While showing the context set to the model, no
additional training occurs, only memory updates are allowed.
So the weights of the model do not change, but the memory does.
Since the sentences in the test set are highly-correlated to each other
(they come from paragraphs with preserved order),
we expect that if we allow a one-shot capable model to use the context set to update its
memory and then evaluate it on the other half of the test set,
its accuracy will increase.
For our GNMT with memory model, we passed the context set through
the memory update operations 3 times.
As seen in Table~\ref{tab:translate}, the context set indeed
helps when evaluating on the odd lines, increasing
the BLEU score by almost $0.5$.
As further indication that our memory module works properly,
we also evaluate the model after showing the whole test set
as a context set. Note that this is essentially an oracle:
the memory module gets to see all the correct answers,
we do this only to test and debug.
As expected, this increases BLEU score dramatically, by over $8$ points.
\begin{table
\caption{Results on the WMT En-De task.
As described in the text, we split the test set in two (odd lines and even lines) to evaluate the
model on one-shot learning. Given the even test set, the model
can perform better on the odd test set. We also see a dramatic improvement when
the model is provided with the whole test set, validating that the memory module
is working as intended.}
\begin{center}
\begin{tabular}{|l||c|c|}
\hline
{\bf Model} & Full Test & Odd Test \\ \hline
GNMT & 23.25 & 23.17 \\ \hline
GNMT with Memory Module & 23.29 & 23.16 \\ \hline
GNMT with Memory Module and Even Test context & -- & 23.60 \\ \hline
GNMT with Memory Module and Whole Test context & 31.11$^*$ & -- \\ \hline
\end{tabular}
\end{center}
\label{tab:translate}
\end{table}
\section{Discussion}
We presented a long-term memory module that can be used for life-long learning.
It is versatile, so it can be added to different deep learning models and
at different layers to give the networks one-shot learning capability.
Several parts of the presented memory module could be tuned and studied
in more detail. The update rule that averages the query with the correct
key could be parametrized. Instead of returning only the single nearest neighbor
we could also return a number of them to be processed by other layers of
the network. We leave these questions for future research.
The main issue we encountered, though, is that evaluating one-shot learning
is difficult, as standard metrics do not focus on this scenario. In this work,
we adapted the standard metrics to investigate our approach. For example,
in the translation task we used half of the test set as context for the other half,
and we still report the standard BLEU score. This allows us to show that our
module works, but it is only a temporary solution.
Better metrics are needed to accelerate progress of one-shot and life-long learning.
Thus, we consider the present work as just a first step on the way to making
deep models learn to remember rare events through their lifetime.
|
\section{Introduction}
\qquad
For $0 \leq \alpha < n$ and $m > 1$, ($m \in \mathbb{N}$), let $T_{\alpha, m}$ be the integral operator defined by
\begin{equation}
T_{\alpha, m}f(x) = \int_{\mathbb{R}^{n}} \left| x-A_1 y\right|^{-\alpha_1} \cdot \cdot \cdot \left| x-A_m y\right|^{-\alpha_m} f(y) dy, \label{Talfa}
\end{equation}
where $\alpha_1, ..., \alpha_m$ are positive constants such that $\alpha_1 + ... + \alpha_m =n- \alpha$, and $A_1, ..., A_m$ are $n \times n$ invertible matrices such that $A_i \neq A_j$ if $i \neq j$. We observe that for the case $\alpha > 0$, $m =1$, and $A_1 = I$, $T_{\alpha, 1}$ is the Riesz potential $I_{\alpha}$. Thus for $0 < \alpha < n$ the operator $T_{\alpha, m}$ is a kind of generalization of the Riesz potential. The case $\alpha = 0$ and $m > 1$ was studied under the additional assumption that $A_i - A_j$ are invertible if $i \neq j$. The behavior of this class of operators and generalizations of them on the spaces of functions $L^{p}(\mathbb{R}^{n})$, $L^{p}(w)$, $H^{p}(\mathbb{R}^{n})$ and $H^{p}(w)$ was studied in \cite{G-U1}, \cite{G-U2}, \cite{Riveros Urciuolo}, \cite{Riv U}, \cite{R-U} and \cite{R-U2}.
If $0 < \alpha < n$ and $m > 1$, then the operator $T_{\alpha, m}$ has the same behavior that the Riesz potential on $L^{p}(\mathbb{R}^{n})$. Indeed
\[
|T_{\alpha, m}f(x)| \leq C \sum_{j=1}^{m} \int_{\mathbb{R}^{n}} | A_{j}^{-1}x - y |^{\alpha - n} |f(y)| dy = C \sum_{j=1}^{m} I_{\alpha}(|f|)( A_{j}^{-1}x),
\]
for all $x \in \mathbb{R}^{n}$, this pointwise inequality implies that $T_{\alpha, m}$ is a bounded operator from $L^{p}(\mathbb{R}^{n})$ into $L^{q}(\mathbb{R}^{n})$ for $1 < p < \frac{n}{\alpha}$ and $\frac{1}{q} = \frac{1}{p} - \frac{\alpha}{n}$, and it is of type weak $(1, n/n-\alpha)$.
It is well known that the Riesz potential $I_{\alpha}$ is bounded from $H^{p}(\mathbb{R}^{n})$ into $H^{q}(\mathbb{R}^{n})$ for $0 < p \leq 1$ and $\frac{1}{q} = \frac{1}{p} - \frac{\alpha}{n}$ (see \cite{weiss} and \cite{krantz}). In \cite{R-U2}, the author jointly with M. Urciuolo proved the $H^{p}(\mathbb{R}^{n})-L^{q}(\mathbb{R}
^{n})$ boundedness of the operator $T_{\alpha, m}$ and we also showed that the $H^{p}(\mathbb{R})-H^{q}(\mathbb{R})$ boundedness does not hold for $0 < p \leq \frac{1}{1+ \alpha}$, $\frac{1}{q}=\frac{1}{p} - \alpha$ and $T_{\alpha, m}$ with $0 \leq \alpha < 1$, $m=2$, $A_1 = 1$, and $A_2=-1$. This is a significant difference with respect to the case $0 < \alpha < 1$, $n=m=1$, and $A_1 = 1$.
In this note we will prove that if we consider certain singular matrices in (\ref{Talfa}), then such an operator is still bounded from $H^{p}$ into $L^{q}$. More precisely, for $m, n \in \mathbb{N}$, $1< m \leq n$, we write $n = n_1+ ...+ n_m$ where $\{ n_1, ..., n_m \} \subset \mathbb{N}$, we also consider $n \times n$ singular real matrices $A_1, ..., A_m$ such that $$\bigoplus_{i=1}^{m} \bigcap_{1\leq j \neq i \leq m} \mathcal{N}_j = \mathbb{R}^{n},$$
where $\mathcal{N}_j = \{ x : A_j x = 0 \}$, $dim(\mathcal{N}_j)=n-n_j$, $A_1+ ...+ A_m$ is invertible. Given $0 < r <1$ and $n_1, ..., n_m$ such that $n_1+ ... + n_m =n$, let $\alpha_1, ..., \alpha_m$ be positive constants such that $\frac{\alpha_1}{n_1} = ... = \frac{\alpha_m}{n_m}=r$, for such parameters we define the integral operator $T_r$ by
\begin{equation}
T_{r}f(x)= \int_{\mathbb{R}^{n}} \, |x-A_1 y|^{-n_1 + \alpha_1} \cdot \cdot \cdot |x-A_m y|^{-n_m + \alpha_m} f(y) \, dy, \label{T}
\end{equation}
where the matrices $A_{i}'$s are as above.
We observe that the operator defined in (\ref{T}) can be written as in (\ref{Talfa}) taking there the matrices $A_{i}$'s singular. In fact, $T_r = T_{\beta, m}$ with $\beta_{i} = n_i - \alpha_i$ for each $i = 1, 2, ..., m$ and $\beta = nr$.
\qquad
Our main result is the following
\qquad
\textbf{Theorem 1.} \textit{Let $T_r$ be the integral operator defined in $(\ref{T})$. If $0 < r < 1$, $0 < p <\frac{1}{r}$ and $\frac{1}{q}=\frac{1}{p} - r$,
then $T_r$ can be extended to an $H^{p}(\mathbb{R}^{n}) - L^{q}(\mathbb{R}^{n})$ bounded operator.}
\qquad
In Section 2 we state two auxiliary lemmas to get the main result in Section 3. We conclude this note with an example in Section 4.
\qquad
Throughout this paper, $c$ will denote a positive constant, not necessarily the same at each occurrence. The symbol $A \lesssim B$ stands for the inequality $A \leq c B$ for some constant $c$.
\section{Preliminary results}
Let $K$ be a kernel in $\mathbb{R}^{n} \times \mathbb{R}^{n}$, we formally define the integral operator $T_{K}$ by $T_{K}f(x) = \int_{\mathbb{R}^{n}} K(x,y) f(y) dy$.
\qquad
We start with the following
\qquad
\textbf{Lemma 1.} \textit{Let $m, n \in \mathbb{N}$, with $1< m \leq n$, and let $n_1, ..., n_m$ be natural numbers such that $n_1 + ... + n_m = n$. For each $i=1, ..., m$ let $K_i$ be kernels in $\mathbb{R}^{n_i} \times \mathbb{R}^{n_i}$ such that the operator $T_{K_i}$ is bounded from $L^{p}(\mathbb{R}^{n_i})$ into $L^{q}(\mathbb{R}^{n_i})$ with $1 < p,q < \infty$, then the operator $T_{K_1 \otimes \cdot \cdot \cdot \otimes K_m}$ is bounded from $L^{p}(\mathbb{R}^{n})$ into
$L^{q}(\mathbb{R}^{n})$.}
\qquad
\textit{Proof.} Since $\mathbb{R}^{n} = \mathbb{R}^{n_1}\times \cdot \cdot \cdot \times \mathbb{R}^{n_m}$ let $x=(x^{1}, ..., x^{m}) \in \mathbb{R}^{n_1}\times \cdot \cdot \cdot \times \mathbb{R}^{n_m}$, now the operator $T_{K_1 \otimes \cdot \cdot \cdot \otimes K_m}$ is given by
$$T_{K_1 \otimes \cdot \cdot \cdot \otimes K_m}f(x)= \int_{\mathbb{R}^{n_1}\times \cdot \cdot \cdot \times \mathbb{R}^{n_m}} \, K_1(x^{1},y^{1}) \cdot \cdot \cdot K_m(x^{m}, y^{m}) f(y^{1}, ..., y^{m}) \, dy^{1} ... dy^{m}.$$
Using that the kernels $K_i$ define bounded operators for $1 \leq i \leq m$, the lemma follows from an iterative argument and the Minkowski's inequality for integrals.$\blacksquare$
\qquad
\textbf{Lemma 2.} \textit{Let $m, n \in \mathbb{N}$, with $1< m \leq n$, and let $n_1, ..., n_m$ be natural numbers such that $n_1 + ... + n_m = n$. If $A_1, ..., A_m$ are $n \times n$ singular real matrices such that $$\bigoplus_{i=1}^{m} \bigcap_{1\leq j \neq i \leq m} \mathcal{N}_j = \mathbb{R}^{n},$$ where
$\mathcal{N}_j = \{ x : A_j x = 0 \}$, $dim(\mathcal{N}_j)=n-n_j$, and $A_1+ ...+ A_m$ is invertible, then there exist two $n \times n$ invertible matrices $B$ and $C$ such that $B^{-1}A_j C$ is the canonical projection from $\mathbb{R}^{n}$ on $\{ 0 \} \times \cdot \cdot \cdot \times \mathbb{R}^{n_j} \times \cdot \cdot \cdot \times \{ 0 \}$ for each $j=1,...,m$.}
\qquad
\textit{Proof.} It is easy to check that $$\bigoplus_{i=1}^{m} \bigcap_{1\leq j \neq i \leq m} \mathcal{N}_j = \mathbb{R}^{n} \Rightarrow
\bigoplus_{1\leq i \neq k \leq m} \bigcap_{1\leq j \neq i \leq m} \mathcal{N}_j = \mathcal{N}_k$$ so
\begin{equation}
A_{k} \left( \bigcap_{1\leq j \neq k \leq m} \mathcal{N}_j \right) = \mathcal{R}(A_k), \,\,\,\,\,\,\,\, k=1, ..., m, \label{proy}
\end{equation}
since $dim(\mathcal{N}_k) = n - n_k$ then $dim\left( \bigcap_{1\leq j \neq k \leq m} \mathcal{N}_j \right) = dim(\mathcal{R}(A_k)) = n_k$. Let $\{ \gamma_{1}^{k}, ..., \gamma_{n_{k}}^{k} \}$ be a basis of
$\bigcap_{1\leq j \neq k \leq m} \mathcal{N}_j$ thus $\{ \gamma_{1}^{1}, ..., \gamma_{n_{1}}^{1}, ..., \gamma_{1}^{m}, ..., \gamma_{n_{m}}^{m} \}$ is a basis for
$\mathbb{R}^{n}$. Let $C$ be the $n \times n$ matrix which columns are the elements of the above basis. Since $A_1+ ...+ A_m$ is invertible we have that
$B=(A_1+ ...+ A_m)C$ is invertible, so (\ref{proy}) gives that $B^{-1}A_j C$ is the canonical projection from $\mathbb{R}^{n}$ on $\{ 0 \} \times \cdot \cdot \cdot \times \mathbb{R}^{n_j} \times \cdot \cdot \cdot \times \{ 0 \}$ for each $j=1,...,m$. $\blacksquare$
\qquad
\section{The main result}
\qquad
\textit{Proof of the Theorem 1.} We begin by obtaining the $L^{p}-L^{q}$ boundedness of the operator $T_r$ for $1 < p < \frac{1}{r}$ and
$\frac{1}{q}=\frac{1}{p}-r$, and then with this result we will prove the $H^{p}-L^{q}$ boundedness of $T_r$ for $0 < p \leq 1$ and $\frac{1}{q}=\frac{1}{p}-r$.
\qquad
\textit{$L^{p}-L^{q}$ boundedness.} If $A$ is a $n \times n$ invertible matrix we put $f_{A}(x) = f(A^{-1}x)$. Let $B$ and $C$ be the matrices given by Lemma 2, then
$$\left[T_r \left( f_{C}\right) \right]_{B^{-1}} (x) = $$
$$\int_{\mathbb{R}^{n}} \, \left|( Bx - A_1 y ) \right|^{-n_1 + \alpha_1} \cdot \cdot \cdot \left|( Bx - A_m y ) \right|^{-n_m + \alpha_m} f(C^{-1}y) \, dy =$$
$$\left|det (C) \right| \int_{\mathbb{R}^{n}} \, \left|B (x - B^{-1}A_1 C y ) \right|^{-n_1 + \alpha_1} \cdot \cdot \cdot \left|B (x - B^{-1}A_m C y ) \right|^{-n_m + \alpha_m} f(y) \, dy.$$
Since $B$ is invertible, then there exists a positive constant $c$ such that $c|x| \leq |Bx|$ for all $x \in \mathbb{R}^{n}$. Thus
$$\left| \left[T_r \left( f_{C}\right) \right]_{B^{-1}} (x) \right| \leq c
\int_{\mathbb{R}^{n}} \, \left|x - B^{-1}A_1 C y \right|^{-n_1 + \alpha_1} \cdot \cdot \cdot \left|x - B^{-1}A_m C y \right|^{-n_m + \alpha_m} |f(y)| \, dy$$
$$\leq c \int_{\mathbb{R}^{n_1} \times ... \times \mathbb{R}^{n_m}} \, \left|x^{1} - y^{1} \right|^{-n_1 + \alpha_1} \cdot \cdot \cdot \left|x^{m} - y^{m} \right|^{-n_m + \alpha_m} |f(y^{1},...,y^{m})| \, dy^{1}...dy^{m},$$
the second inequality it follows from Lemma 2 and from that $|x^{j}-y^{j}| \leq |x - P_j y|$, where $P_j=B^{-1}A_j C$ is the canonical projection from $\mathbb{R}^{n}$ on $\{0\} \times \cdot \cdot \cdot \times \mathbb{R}^{n_j} \times \cdot \cdot \cdot \times \{0\}$. Since $\gamma(\alpha_j)^{-1}|x^{j}-y^{j}|^{-n_j+\alpha_j}$, for an appropriate constant $\gamma(\alpha_j)$ (see \cite{stein2}, p. 117), is the kernel of the Riesz potential on $\mathbb{R}^{n_j}$, then Theorem 1 (in \cite{stein2}, p. 119) and Lemma 1 give the $L^{p}-L^{q}$ boundedness of the operator $T_r$ for $1 < p < \frac{1}{r}$ and $\frac{1}{q}=\frac{1}{p}-r$.
\qquad
\textit{$H^{p}-L^{q}$ boundedness.} Let $0<p\leq 1.$ We recall that a $p$-atom is a measurable function $a$ supported on a ball $B$ of $\mathbb{R}^{n}$ satisfying
$a)$ $\left\Vert a\right\Vert _{\infty }\leq \left\vert B\right\vert^{-\frac{1}{p}}$
\qquad
$b)$ $\int y^{\beta }a(y)dy=0$ for every multiindex $\beta $ with $\left\vert \beta \right\vert \leq \lfloor n(p^{-1}-1) \rfloor$, ($\lfloor s \rfloor$ denotes the integer part of $s$).
\qquad
Let $0<p\leq 1 < p_0 < \frac{1}{r}$, $0 < r < 1$, and $\frac{1}{q} = \frac{1}{p} - r$. Given $f\in H^{p}(\mathbb{R}^{n}) \cap L^{p_0}(\mathbb{R}^{n})$, from Theorem 2, p. 107, in \cite{stein} we have that there exist a sequence of nonnegative numbers $\{ \lambda_j \}_{j=1}^{\infty}$, a sequence of balls $B_{j}=B(z_j, \delta_j)$ centered at $z_j$ with radius $\delta_j$ and $p$ - atoms $a_j$ supported on $B_j$, satisfying
\begin{equation}
\sum\limits_{j\in \mathbb{N}}\left\vert \lambda _{j}\right\vert ^{p} \leq c\left\Vert f\right\Vert _{H^{p}}^{p}, \label{esti}
\end{equation}
such that $f$ can be decomposed as $f=\sum\limits_{j\in \mathbb{N}}\lambda _{j} a_{j},$ where the convergence is in $H^{p}$ and in $L^{p_0}$ (for the converge in $L^{p_0}$ see Theorem 5 in \cite{Rocha}). So the $H^{p}-L^{q}$ boundedness of $T_r$ will be proved if we show that there exists $c>0$ such that
\begin{equation}
\left\Vert T_r a_j \right\Vert _{L^{q}}\leq c \label{uniform estimate}
\end{equation}
with $c$ independent of the $p$-atom $a_j$. Indeed, since $f = \sum_{j=1}^{\infty} \lambda_j a_j$ in $L^{p_0}$ and $T_{r}$ is a $L^{p_0}-L^{\frac{p_0}{1- r p_0}}$ bounded operator, we have that $|T_{r} f (x)| \leq \sum_{j=1}^{\infty} \lambda_j |T_{r} a_j(x)|$ a.e.$x$,
this pointwise estimate, the inequality in (\ref{uniform estimate}), joint to the inequality
\[
\left( \sum_{j \in \mathbb{N}} | \lambda_j |^{\min\{1, q\}} \right)^{\frac{1}{\min\{1, q\}}} \leq \left( \sum_{j \in \mathbb{N}} | \lambda_j |^{p} \right)^{\frac{1}{p}}
\]
and (\ref{esti}) we obtain $\left\Vert T_r f\right\Vert_q \leq c \left\Vert f\right\Vert _{H^{p}}$, for all $f\in H^{p}(\mathbb{R}^{n}) \cap L^{p_0}(\mathbb{R}^{n})$, so the theorem follows from the density of $ H^{p}(\mathbb{R}^{n}) \cap L^{p_0}(\mathbb{R}^{n})$ in $H^{p}(\mathbb{R}^{n})$.
We will prove the estimate in (\ref{uniform estimate}). We define $D = \max_{1\leq i \leq m} \max_{|y|=1} |A_i (y)|$. Let $a_j$ be an $p$ - atom supported on a ball $B_j = B(z_j, \delta_j)$, for each $1 \leq i \leq m$ let $B_{ji}^{\ast} = B(A_i z_j, 4 D \delta_j)$. Since $T_r$ is bounded from $L^{p_{0}}\left( \mathbb{R}^{n}\right) $ into $L^{q_{0}}\left( \mathbb{R}^{n}\right) $ for $\frac{1}{q_{0}}=\frac{1}{p_{0}}-r,$ $1<p_{0}<\frac{1}{r},$ the H\"{o}lder inequality gives
\begin{equation*}
\int\limits_{\bigcup\limits_{1\leq i\leq m}B_{ji}^{\ast }}\left\vert
T_r a_j(x)\right\vert ^{q}dx\leq \sum\limits_{1\leq i\leq
m}\int\limits_{B_{ji}^{\ast }}\left\vert T_r a_j(x)\right\vert ^{q}dx
\end{equation*}
\begin{equation*}
\leq c\sum\limits_{1\leq i\leq m}\left\vert B_{ji}^{\ast }\right\vert ^{1- \frac{q}{q_{0}}}\left\Vert T_r a_j \right\Vert _{q_{0}}^{q}\leq c \delta_{j}^{n-\frac{
nq}{q_{0}}}\left\Vert a_j \right\Vert _{p_{0}}^{q}
\end{equation*}
\begin{equation*}
\leq c \delta_{j}^{n-\frac{nq}{q_{0}}}\left( \int\limits_{B_j}\left\vert
a_j \right\vert ^{p_{0}}\right) ^{\frac{q}{p_{0}}}\leq c \delta_{j}^{n-\frac{nq}{
q_{0}}} \delta_{j}^{-\frac{nq}{p}} \delta_{j}^{\frac{nq}{p_{0}}}=c.
\end{equation*
We denote $k(x,y)=\left\vert x-A_{1}y\right\vert ^{- n_1 + \alpha
_{1}}...\left\vert x-A_{m}y\right\vert ^{- n_m + \alpha _{m}}$ and we put $N-1 = \lfloor n(p^{-1}-1) \rfloor$. In view of the moment condition of $a_j$ we have
\begin{equation}
T_{r} a_j(x)=\int\limits_{B_j}k(x,y)a_j(y)dy=\int\limits_{B_j}\left( k(x,y)-q_{N, j}\left(
x,y\right) \right) a_j(y)dy,
\end{equation
\newline
where $q_{N, j}$ is the degree $N-1$ Taylor polynomial of the function $
y\rightarrow k(x,y)$ expanded around $z_j$. By the standard estimate of the
remainder term of the taylor expansion, there exists $\xi $ between $y$ and
$z_j$ such that
\[
\left\vert k(x,y)-q_{N, j}\left( x,y\right) \right\vert \lesssim \left\vert
y-z_j \right\vert ^{N}\sum\limits_{k_{1}+...+k_{n}=N}\left\vert \frac
\partial ^{N}}{\partial y_{1}^{k_{1}}...\partial y_{n}^{k_{n}}}k(x,\xi
)\right\vert
\]
\[
\lesssim \left\vert y-z_j \right\vert ^{N}\left(
\prod\limits_{i=1}^{m}\left\vert x-A_{i}\xi \right\vert ^{-n_i + \alpha
_{i}}\right) \left( \sum\limits_{l=1}^{m}\left\vert x-A_{l}\xi \right\vert
^{-1}\right) ^{N}.
\]
Now we decompose $\mathbb{R}^{n} = \bigcup_{i=1}^{m} B_{ji}^{\ast} \cup R_j$, where $R_j = \left( \bigcup_{i=1}^{m} B_{ji}^{\ast} \right)^{c}$, at the same time we decompose $R_j = \bigcup_{k=1}^{m} R_{jk}$ with
$$R_{jk} = \{ x \in R_{j} : |x - A_k z_j| \leq |x - A_i z_j| \,\, for \,\, all \,\, i \neq k \}.$$
If $x \in R_{j}$ then $|x - A_i z_j| \geq 4D \delta_j$, for all $i=1, ..., m$, since $\xi \in B_j$ it follows that $|A_i z_j - A_i \xi | \leq D \delta_j \leq \frac{1}{4} |x - A_i z_j|$ so
$$|x - A_i \xi| = |x - A_i z_j + A_i z_j - A_i \xi| \geq |x - A_i z_j| - |A_i z_j - A_i \xi| \geq \frac{3}{4} |x - A_i z_j|.$$
If $x \in R_j$, then $x \in R_{jk}$ for some $k$, and since $\sum_{i=1}^{m}(-n_i + \alpha_i) = - n(1-r)$ we obtain
$$
\left\vert k(x,y)-q_{N, j}\left( x,y\right) \right\vert \lesssim
\left\vert y-z_j \right\vert ^{N}\left( \prod\limits_{i=1}^{m}\left\vert
x-A_{i}z_j \right\vert ^{- n_i + \alpha _{i}}\right) \left(
\sum\limits_{l=1}^{m}\left\vert x-A_{l}z_j \right\vert ^{-1}\right) ^{N},
$$
$$
\lesssim \left\vert y-z_j \right\vert ^{N} \left\vert x-A_{k}z_j \right\vert ^{-n(1-r) -N}, \,\,\,\,\, if \,\, x \in R_{jk} \,\,\, and \,\,\, y \in B_j.
$$
This inequality allow us to conclude that
\[
\int\limits_{R_j}\left\vert \int\limits_{B_j}K(x,y)a_j (y)dy\right\vert^{q}dx = \int\limits_{R_j}\left\vert \int\limits_{B_j} \left[ K(x,y) - q_{N,j}(x,y) \right] a_j (y)dy\right\vert^{q}dx
\]
\[
\lesssim \sum_{k=1}^{m} \int\limits_{R_{jk}}\left(
\int\limits_{B_j}\left\vert y-z_j \right\vert ^{N} \left\vert x-A_{k}z_j \right\vert ^{-n (1-r) - N}\left\vert
a_j (y)\right\vert dy\right) ^{q}dx
\]
\[
\lesssim \left( \int\limits_{B_j}\left\vert y- z_j \right\vert ^{N}\left\vert
a_j (y)\right\vert dy\right) ^{q} \sum_{k=1}^{m} \int\limits_{\left(B_{jk}^{\ast} \right)^{c}}\left\vert x-A_{k}z_j \right\vert ^{-n (1-r)q - Nq} dx
\]
\[
\lesssim \delta_{j}^{qN-n \frac{q}{p}+nq} \int\limits_{4 D \delta_j}^{\infty }t^{-q\left( n(1-r)+N\right) +n-1}dt \leq c
\]
with $c$ independent of the $p-$atom $a_j$ since $-q\left( n(1-r)+N\right)
+n<0.\blacksquare $
\section{An example}
For $n=m=3$, $n_1 = n_2 = n_3 = 1$ we consider the following $3 \times 3$ singular matrices
\[
A_1= \left(
\begin{array}{ccc}
\,\,\, 4 & \,\,\, 4 & -1 \\
\,\,\, 0 & \,\,\,0 & \,\,\, 0 \\
-4 & -4 & \,\,\, 1 \\
\end{array}
\right), \,\,\,\,\,\, A_2=\left(
\begin{array}{ccc}
\,\,\, 1 & -1 & 0 \\
-2 & \,\,\, 2 & 0 \\
\,\,\, 0 & \,\,\, 0 & 0 \\
\end{array}
\right), \,\,\,\,\,\, A_3 = \left(
\begin{array}{ccc}
\,\,\, 1 & 0 & -1 \\
-3 & 0 & \,\,\, 3 \\
-1 & 0 & \,\,\, 1 \\
\end{array}
\right).
\]
It is clear that
\[
A_1 + A_2 +A_3 = \left(
\begin{array}{ccc}
\,\,\, 6 & \,\,\, 3 & -2 \\
-5 & \,\,\, 2 & \,\,\, 3 \\
-5 & -4 & \,\,\, 2 \\
\end{array}
\right)
\]
is invertible. For each $j=1, 2, 3,$ let $\mathcal{N}_j = \{ x \in \mathbb{R}^{3} : A_j x = 0\}$. A computation gives
$$
\mathcal{N}_1 = \langle (1, 0, 4) , (0, 1, 4) \rangle, \,\,\, \mathcal{N}_2 = \langle (1, 1, 0), (0, 0, 1) \rangle, \,\,\, \mathcal{N}_3 = \langle (1, 0, 1), (0, 1, 0) \rangle,
$$
one can check that
$$
\mathcal{N}_1 \cap \mathcal{N}_2 = \langle (1, 1, 8) \rangle, \,\,\, \mathcal{N}_1 \cap \mathcal{N}_3 = \langle (4, -3, 4) \rangle,
\,\,\, and \,\,\, \mathcal{N}_2 \cap \mathcal{N}_3 = \langle (1, 1, 1) \rangle.
$$
\qquad
We observe that $\mathcal{N}_1 \cap \mathcal{N}_2 \oplus \mathcal{N}_1 \cap \mathcal{N}_3 \oplus \mathcal{N}_2 \cap \mathcal{N}_3 = \mathbb{R}^{3}$. As in the proof of Lemma 2, we define the matrices $C$ and $B$ by
\[
C = \left(
\begin{array}{ccc}
1 & \,\,\, 4 & 1 \\
1 & -3 & 1 \\
1 & \,\,\, 4 & 8\\
\end{array}
\right),
\,\,\, B= ( A_1 + A_2 + A_3 ) \, C = \left(
\begin{array}{ccc}
\,\,\, 7 & \,\,\, 7 & -7 \\
\,\,\, 0 & -14 & \,\,\, 21 \\
-7 & \,\,\, 0 & \,\,\, 7 \\
\end{array}
\right),
\]
both matrices are invertibles with
\[
B^{-1} = \left(
\begin{array}{ccc}
\frac{2}{21} & \frac{1}{21} & -\frac{1}{21} \\\\
\frac{1}{7} & 0 & \,\,\, \frac{1}{7} \\\\
\frac{2}{21} & \frac{1}{21} & \,\,\, \frac{2}{21} \\
\end{array}
\right).
\]
Now it is easy to check that
\[
B^{-1 } A_1 C = \left(
\begin{array}{ccc}
1 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & 0 \\
\end{array}
\right), \,\,\,
B^{-1 } A_2 C = \left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 0 \\
\end{array}
\right), \,\,\,
B^{-1 } A_3 C = \left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & 1 \\
\end{array}
\right).
\]
So, from Theorem 1, it follows that the operator $T_r$ defined by
\[
T_rf(x) = \int_{\mathbb{R}^{3}} |x-A_1 y |^{-1+r} |x -A_2 y |^{-1+r} |x - A_3 y |^{-1+r} f(y) dy,
\]
with $0 < r < 1$, is a bounded operator from $H^{p}(\mathbb{R}^{3})$ into $L^{q}(\mathbb{R}^{3})$ for $0 < p < 1/r$ and $\frac{1}{q} = \frac{1}{p} - r$.
\qquad
|
\section{Introduction}
\label{s.intro}
\subsection{Motivation and goal}
\label{ss.motivation}
Two models play a central role in the world of stochastic multi-type population dynamics:
\begin{itemize}
\item[(1)]
The Moran model and its limit for large populations, the Fleming-Viot measure-valued
\emph{diffusion}.
\item[(2)]
The Cannings model and its limit for large populations, the Cannings measure-valued
\emph{jump process} (also called the generalized Fleming-Viot process).
\end{itemize}
The Cannings model accounts for situations in which \emph{resampling} is such that a
random positive fraction of the population in the next generation inherits the type of a
random single individual in the current generation, even in the infinite population limit
(see Cannings~\cite{C74}, \cite{C75}). In order to describe a setting where this effect
has a geographical structure, i.e., where \emph{migration} of individuals is allowed as
well, different models have been proposed in Limic and Sturm~\cite{LS06}, Blath,
Etheridge and Meredith~\cite{BEM07}, Barton, Etheridge and V\'eber~\cite{BEV10},
Berestycki, Etheridge and V\'eber~\cite{BEV13}, and Greven, den Hollander, Kliem and
Klimovsky~\cite{GHKK14}. The behaviour of these models has been studied in detail
and its dependence on the geographic space is fairly well understood.
The type space is typically chosen to be a compact Polish space $E$. In~\cite{GHKK14}, we focused
on the case where the geographic space is the hierarchical group $\Omega_N$ of order $N$, since
this allowed us to carry out a full \emph{renormalization analysis}. In the \emph{hierarchical mean-field
limit} $N \to \infty$, the migration can be chosen in such a way that it approximates migration on the
geographic space $\mathbb{Z}^2$, a possibility that was exploited by Sawyer and Felsenstein~\cite{SF83}
(see also Dawson, Gorostiza and Wakolbinger~\cite{DGW04}).
We analyze the model introduced in \cite{GHKK14}, but add the effect that the Cannings resampling
mechanism is controlled by \emph{catastrophic events} on a small time scale, for which it is appropriate
to assume that the rate of occurrence has a spatially inhomogeneous structure. This leads us to
consider spatial Cannings models with block resampling in \emph{random environment}, i.e., both
the form and the overall rate of the block resampling mechanism depend on the geographic location.
\begin{remark}\label{r.325}
{\rm In a catastrophic event, a part of the population is killed in a large spatial area and is subsequently
replenished via a rapid recolonization, resulting in a bottleneck effect consisting of compression and
subsequent expansion of the descendants of a single ancestor. The mechanisms behind such events
are functions of the background environment, which is inhomogeneous in space but constant in time.
It would be interesting to derive our continuum model (defined in Section~\ref{ss.random}) from an
individual-based model with \textit{two time scales}: the catastrophic events happen on a fast time
scale, while the migration and resampling happen on a slow time scale. Moreover, in our individual-based
model we do \emph{reshuffling} before resampling, which must be motivated likewise. Carrying out
the details of such a derivation would merit a paper in its own right.} \hfill $\square$
\end{remark}
The {\em goal} of the present paper is three-fold:
\begin{itemize}
\item[(1)]
\emph{Construction} of the hierarchical Cannings process in random environment via a well-posed
martingale problem and derivation of a \emph{duality relation} with a hierarchical spatial coalescent
in random environment.
\item[(2)]
Analysis of the longtime behaviour, in particular, the \emph{dichotomy} between a multi-type equilibrium
and a mono-type equilibrium.
\item[(3)]
Scaling analysis of a collection of r\emph{enormalized processes} obtained by looking at the evolution
of blocks averages on successive space-time scales in the hierarchical mean-field limit and the
consequences for universality classes of the mono-type cluster formation.
\end{itemize}
We are particularly interested in \emph{new effects caused by the random environment}.
The mathematical tools we will exploit are the duality of the hierarchical Canning process in random
environment with a hierarchical spatial coalescent in random environment, and the scaling of the block
averages towards a mean-field process in random environment called the McKean-Vlasov process.
This in turn will lead us to study two independent hierarchical random walks in the same random
environment, and to analyze the orbit of iterations of non-linear transformations arising from random
M\"obius transformations that link the behaviour on successive hierarchical scales.
\subsection{Summary of the main results}
\label{ss.summary}
In an earlier paper, we introduced and studied a system of hierarchically interacting measure-valued
random processes that arises as the continuum limit of a large population of individuals subject to
migration, reshuffling and resampling \cite{GHKK14}. More precisely, individuals live in colonies labelled
by $\Omega_N$, the hierarchical group of order $N$, and are subject to \emph{migration} based on a
sequence of migration coefficients $\underline{c}=(c_k)_{k\in\mathbb{N}_0}$ and to \emph{resampling} based on a
sequence of resampling measures $\underline{\Lambda}= (\Lambda_k)_{k\in\mathbb{N}_0}$, both acting on blocks of colonies
(= macro-colonies) on all hierarchical scales $k\in\mathbb{N}_0$ simultaneously. The resampling mechanism
is such that a random positive fraction of the population in a block of colonies inherits the type of a
random single individual in that block, even in the infinite population limit, which is why we refer to
our system as the hierarchical Cannings process. Before resampling in a block takes place, all individuals
in that block are relocated uniformly. This relocation is called \emph{reshuffling} and means that
resampling is done in a locally ``panmictic'' manner.
In the present paper, we study a version of the hierarchical Cannings process in \emph{random
environment}, namely, the resampling measures in different blocks are chosen randomly with
mean $\underline{\Lambda}$ and are kept fixed in time, i.e., we consider the \emph{quenched} version of the
system. We construct the hierarchical Cannings process in random environment via a well-posed
martingale problem, and establish duality with a system of coalescing hierarchical random walks
with block coalescence in random environment. We study the long-time behaviour of the process,
in particular, we give a necessary and sufficient condition on $\underline{c}$ and $\underline{\Lambda}$ under which
almost sure convergence to a \emph{multi-type equilibrium} occurs (= coexistence), as opposed
to a \emph{mono-type equilibrium} (= clustering). The equilibrium depends on the environment,
but the condition on $\underline{c}$ and $\underline{\Lambda}$ for its occurrence does not.
To obtain more detailed information on the evolution of the system, we consider the \emph{hierarchical
mean-field limit} $N \to \infty$. In this limit, with the help of a \emph{renormalization analysis}, we
obtain a full picture of the space-time scaling behaviour on all hierarchical scales simultaneously.
Our main result is that, on each hierarchical scale $k \in \mathbb{N}_0$, the $k$-block averages on time
scale $N^k$ converge to a random process that is a superposition of a Cannings process with a
resampling measure equal to the associated $k$-block resampling measure (which depends on
the environment) and an additional Fleming-Viot process with volatility $d_k$, reflecting the
macroscopic impact of the lower-order resampling and of the drift of strength $c_k$ towards
the limiting $(k+1)$-block average (which is constant on the limiting time scale). It turns out that
$d_k$ is a function of $c_l$ and $\Lambda_l$ for all $0 \leq l < k$, and of the law of the random
environment. Thus, \emph{it is through the volatility that the renormalization manifests itself}.
We show that \emph{the random environment makes the system less volatile}, i.e., $d_k$ is strictly
smaller than its corresponding value for the homogenous system where the resampling measures
are replaced by their mean. We investigate how $d_k$ scales as $k \to \infty$, which leads to
\emph{various different cases} depending on the choice of $\underline{c}$ and $\underline{\Lambda}$. We find that if migration
and resampling occur with comparable rates on all hierarchical scales, then the lower volatility
persists in the limit as $k\to\infty$. The renormalization transformation connecting $d_{k+1}$ to
$d_k$ turns out to be a non-linear transformation arising from a \emph{random M\"obius transformation}.
The scaling behaviour of the iterates of these transformations is studied in detail. We find that if
clustering occurs, then the random environment slows down the growth of the clusters, i.e.,
enhances the diversity of types. We find five \emph{universality classes of cluster formation} in
the regime of clustering. These are linked to the different cases of scaling behaviour of $d_k$.
We find that if the growth of the clusters is \emph{rapid}, then the rate of growth depends on
the realisation of the environment, while if the growth is \emph{slow}, then the effect of the
environment averages out. The latter happens e.g.\ in the \emph{critical regime} where the
system is barely clustering.
\subsection{Outline}
\label{ss.outline}
Sections~\ref{s.model}--\ref{s.randomwalk} deal with the model for finite $N$, while
Sections~\ref{s.mkvrand}--\ref{s.clustering} deal with the hierarchical mean-field limit
$N\to\infty$. In Section~\ref{s.model} we define the hierarchical Cannings process and
its dual. In Section~\ref{s.results} we state our main theorems and summarize the effects
of the random environment. Section~\ref{s.dual} contains the proof of existence and
uniqueness of the hierarchical Cannings process and its dual, and establishes convergence
to an equilibrium. Section~\ref{s.randomwalk} proves the dichotomy between coexistence
(multi-type equilibrium) versus clustering (mono-type equilibrium), and provides the parameter
range for both. Section~\ref{s.mkvrand} contains the multi-scale analysis for the evolution
of block averages on successive space-time scales in the hierarchical mean-field limit, proves
the dichotomy in that limit, and identifies the renormalization transformations connecting the
successive scales. Section~\ref{s.completeproof} analyzes the orbit of the iterations of
these transformations and identifies various different cases for the scaling of the volatility of
the block averages. Section~\ref{s.clustering} links these cases to the universality classes
of cluster formation.
\section{The model}
\label{s.model}
In this section, we define the hierarchical Cannings process in random environment and construct
its dual: a spatial coalescent in random environment. We begin in Section~\ref{ss.reviewC} by
recalling the process without random environment introduced in~\cite{GHKK14}. In
Section~\ref{ss.random} we explain how the random environment is added.
\subsection{The hierarchical Cannings process}
\label{ss.reviewC}
In Sections~\ref{sss.hg}--\ref{sss.resh}, we recall the definition of the hierarchical Canning
process given in ~\cite{GHKK14}. In Section~\ref{sss.hierarCan} we add the random environment
and indicate how the definition needs to be modified.
\subsubsection{The hierarchical group of order $N$}
\label{sss.hg}
The {\em hierarchical group $\Omega_N$} of order $N$ is the set
\be{ag30_a}
\Omega_N = \Big\{\eta=(\eta^l)_{l\in\mathbb{N}_0} \in\{0,1,\ldots, N-1\}^{\mathbb{N}_0}
\colon\, \sum_{l\in\mathbb{N}_0} \eta^l < \infty\Big\}, \qquad N\in\mathbb{N}\backslash\{1\},
\end{equation}
endowed with the addition operation $+$ defined by $(\eta+\zeta)^l=\eta^l+\zeta^l \textrm{ (mod $N$)}$,
$l\in\mathbb{N}_0$. In other words, $\Omega_N$ is the direct sum of the cyclical group of order $N$ (a fact
that is important for the application of Fourier analysis). The group $\Omega_N$ is equipped with
the ultrametric distance $d_{\Omega_N}(\cdot,\cdot)$ defined by
\be{ag31}
d_{\Omega_N}(\eta,\zeta)=d_{\Omega_N}(0,\eta-\zeta)
= \min\{k\in\mathbb{N}_0 \colon\, \eta^l=\zeta^l \,\, \forall\, l \geq k\},
\qquad \eta,\zeta\in\Omega_N.
\end{equation}
Let
\be{block-definition}
B_k(\eta) = \{\zeta\in\Omega_N\colon\, d_{\Omega_N}(\eta,\zeta) \leq k\},
\qquad \eta\in\Omega_N,\,k\in\mathbb{N}_0
\end{equation}
denote the $k$-block around $\eta$ (i.e., the ball of hierarchical radius $k$ around $\eta$), which
we think of as a {\em macro-colony}. The geometry of $\Omega_N$ is explained in Fig.~\ref{fig-hierargr}.
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{hierarchical-group.pdf}
\caption{\small Close-ups of a 1-block, a 2-block and a 3-block in the hierarchical group of order $N=3$.
The elements of the group are the leaves of the tree (indicated by $\Box$'s). The hierarchical distance
between two elements is the graph distance to the most recent common ancestor: $d_{\Omega_3}
(\eta,\zeta) = 2$ for $\eta$ and $\zeta$ in the picture.}
\label{fig-hierargr}
\end{figure}
In what follows, we consider a system of individuals organized in colonies labelled by $\Omega_N$.
Initially each colony has $M$ individuals, each carrying a type drawn from a Polish type space $E$
that is compact. Subsequently, individuals are subject to block migration (Section~\ref{sss.mg})
and block reshuffling-resampling (Section~\ref{sss.resh}). In the continuum-mass limit $M\to\infty$,
the evolution converges to the hierarchical Cannings process (Section~\ref{sss.hierarCan}).
\subsubsection{Block migration}
\label{sss.mg}
We introduce migration on $\Omega_N$ through a random walk kernel.
For that purpose, we introduce a sequence of migration rates
\be{ckdef}
\underline{c} = (c_k)_{k\in\mathbb{N}_0} \in (0,\infty)^{\mathbb{N}_0},
\end{equation}
and we let the individuals \emph{migrate} as follows:
\begin{itemize}
\item
Each individual, for every $k\in\mathbb{N}$, chooses at rate $c_{k-1}/N^{k-1}$ the block of radius $k$ around
its present location and jumps to a location chosen uniformly at random in that block.
\end{itemize}
The transition kernel of the random walk thus performed by the individuals is
\be{32b}
a^{(N)}(\eta,\zeta) = \sum_{k \geq d_{\Omega_N}(\eta,\zeta)} \frac{c_{k-1}}{N^{2k-1}},
\qquad \eta,\zeta\in\Omega_N,\,\eta\neq\zeta, \qquad a^{(N)}(\eta,\eta)=0.
\end{equation}
\begin{remark}
\label{degreeretr}
{\rm The behaviour of the random walk in \eqref{32b} is known in great detail. Dawson, Gorostiza
and Wakolbinger~\cite{DGW05} showed that it is recurrent if and only if $\sum_{k\in\mathbb{N}_0} (1/c_k)
= \infty$. They introduced the concept of \emph{degree of recurrence/transience} $\gamma_N$
\cite[Definition 2.1.1]{DGW05}, which in the special case where $c_k=c^k$ equals $\gamma(N)
= \log c/\log(N/c)$. Note that
\begin{equation}
\gamma(N) \left\{\begin{array}{ll}
<0, &\quad c<1 \text{ (strongly recurrent)},\\
=0, &\quad c=1 \text{ (critically recurrent)},\\
>0, &\quad c>1 \text{ (transient)}.
\end{array}
\right.
\end{equation}
This is the same as for simple random walk on $\mathbb{Z}^d$ with Hausdorff dimension $d=d(N)=(2\log N)/
\log(N/c)$ (when we allow for a continuum of dimensions). In particular, $d=d(N)=2$ for $c=1$.}
\end{remark}
Throughout the paper, we assume that
\be{ak:recurrence-cond}
\limsup_{k\to\infty} \tfrac{1}{k} \log c_k < \log N.
\end{equation}
This guarantees that the total migration rate per individual is finite.
\subsubsection{Block reshuffling-resampling}
\label{sss.resh}
The idea of the Cannings resampling mechanism is to allow reproduction with an offspring that
is of a size comparable to the whole population. Since we have introduced a spatial structure,
we now allow, on all hierarchical levels $k$ simultaneously, a reproduction event where each
individual treats the $k$-block around its present location as a \emph{macro-colony} and uses
it for its resampling. More precisely, we choose a sequence of resampling measures
\be{ag41}
\underline{\Lambda} = \big(\Lambda_k)_{k\in\mathbb{N}_0} \in \mathcal{M}_f([0,1])^{\mathbb{N}_0},
\end{equation}
where $\mathcal{M}_f([0,1])$ denotes the set of finite non-negative measures on $[0,1]$, satisfying
\be{ag}
\Lambda_0(\{0\})=0, \qquad \int_{(0,1]} \frac{\Lambda_0(\mathrm{d} r)}{r} = \infty,
\end{equation}
and
\begin{equation}
\label{ak1000}
\Lambda_k(\{0\})=0, \qquad \int_{(0,1]} \frac{\Lambda_k(\mathrm{d} r)}{r^2} < \infty. \qquad k \in \mathbb{N},
\end{equation}
Let $\Lambda^*_k(\mathrm{d} r) = \Lambda_k(\mathrm{d} r)/r^2$, $r \in (0,1]$. Set
\be{lambda-total-masses}
\lambda_k = \Lambda_k((0,1]), \qquad \lambda^*_k = \Lambda^*_k((0,1]), \qquad k\in\mathbb{N}_0,
\end{equation}
and assume that
\be{lambdakdef}
\underline{\lambda} = (\lambda_k)_{k\in\mathbb{N}_0} \in (0,\infty)^{\mathbb{N}_0}.
\end{equation}
We let individuals \emph{reshuffle-resample} by carrying out the following two steps at once:
\begin{itemize}
\item
For every $\eta\in\Omega_N$ and $k\in\mathbb{N}_0$, choose the block $B_k(\eta)$ at rate $1/N^{2k}$.
\item
First, each individual in $B_k(\eta)$ independently is moved to a uniformly random location in
$B_k(\eta)$, i.e., a reshuffling takes place (see Fig.~\ref{fig-reshuffle}). After that, $r$ is drawn
according to the intensity measure $\Lambda^*_k$ and $a$ is drawn according to the current
type distribution in $B_k(\eta)$, and each of the individuals in $B_k(\eta)$ independently is
replaced by an individual of type $a$ with probability $r$.
\end{itemize}
Note that the reshuffling-resampling affects all the individuals in a macro-colony simultaneously
and in the same manner. The reshuffling-resampling occurs at all levels $k\in\mathbb{N}_0$, at a rate that
is fastest in single colonies and gets slower as the level $k$ of the macro-colony increases.
\footnote{Because the reshuffling is done first, the resampling always acts on a uniformly
distributed state (``panmictic resampling''). Reshuffling is a parallel update affecting all individuals
in a macro-colony simultaneously. Therefore it cannot be seen as a migration of individuals equipped
with independent clocks.}
\begin{figure}[htbp]
\label{fig:reshuffling}
\centering
\vspace{0.2cm}
\includegraphics[width=\textwidth]{block-reshuffling.pdf}
\caption{\small
Random reshuffling in a 1-block on the hierarchical lattice of order $N=3$, with $M=3$ individuals
of two types (full circles and fuzzy circles) per colony. \emph{Note}: Typically a random reshuffling
does not preserve the number of individuals per colony, but in the example drawn here it does.}
\label{fig-reshuffle}
\end{figure}
The first conditions in (\ref{ag}) and (\ref{ak1000}) make the resampling a \emph{jump process}.
Later we will add in diffusion by hand. The second condition in (\ref{ag}) guarantees that the population
has a well-defined genealogy and that after a positive finite time most of the population at a site
descends from a finite number of ancestors (see Pitman~\cite{P99}). The second condition in
(\ref{ak1000}) is needed to guarantee that in finite time a macro-colony is affected by finitely many
reshuffling-resampling events, otherwise the resampling cannot be properly defined.
Throughout the paper, we assume that
\be{ak:lambda-growth-condition}
\limsup_{k\to\infty} \tfrac{1}{k}\,\log \lambda^*_k < \log N.
\end{equation}
Note that each of the $N^k$ colonies in a $k$-block can trigger reshuffling-resampling in that block,
and for each colony the block is chosen at rate $N^{-2k}$. Therefore, \eqref{ak:lambda-growth-condition}
guarantees that the total resampling rate per individual is bounded.
\subsubsection{The generator and the martingale problem}
\label{sss.hierarCan}
We are now ready to formally define the hierarchical Cannings process in terms of a martingale
problem. The process arises as the continuum-mass limit of the individual-based model described
in Sections~\ref{sss.hg}--\ref{sss.resh}. Namely, in each colony of size $M$, instead of recording
the \emph{numbers} of individuals of a given type we record the \emph{empirical distribution} of the
types and pass to the limit $M\to\infty$.
Let $\mathcal{P}(E)$ denote the set of probability measure on $E$ equipped with the topology of weak
convergence. We equip the set $\mathcal{P}(E)^{\Omega_N}$ with the product topology to get a state
space that is Polish. Let $\mathcal{F} \subset C_\mathrm{b} \big(\mathcal{P}(E)^{\Omega_N},\mathbb{R}\big)$ be
the algebra of functions of the form
\be{ak:multi-level-test-functions}
\begin{aligned}
&F(x) = \int_{E^n} \left(\bigotimes_{m=1}^n x_{\eta_m}\big(\mathrm{d} u^m\big)\right)
f\big(u^1,\ldots,u^n\big),
\quad x = (x_{\eta})_{\eta\in\Omega_N}\in\mathcal{P}(E)^{\Omega_N},\\
&n \in \mathbb{N}, \quad f \in C_{\mathrm{b}}(E^n,\mathbb{R}),
\quad \eta_1,\ldots,\eta_n \in \Omega_N.
\end{aligned}
\end{equation}
The linear operator for the martingale problem
\be{ak:multi-level-generator}
L^{(\Omega_N)}\colon\,\mathcal{F} \to C_\mathrm{b}\big(\mathcal{P}(E)^{\Omega_N},\mathbb{R}\big)
\end{equation}
has two parts,
\be{ak:multi-level-generator-decomposition}
L^{(\Omega_N)}=L^{(\Omega_N)}_{\mathrm{mig}}+L^{(\Omega_N)}_{\mathrm{res}}.
\end{equation}
The \emph{migration operator} is given by
\be{ak:multi-level-migration}
(L^{(\Omega_N)}_{\mathrm{mig}} F)(x)
= \sum_{\eta,\zeta \in \Omega_N} a^{(N)}(\eta,\zeta)
\int_E (x_{\zeta} - x_{\eta})(\mathrm{d} a)\,\frac{\partial F(x)}{\partial x_\eta}[\delta_a]
\end{equation}
and the \emph{reshuffling-resampling operator} by
\be{ak:multi-level-resampling-global}
\begin{aligned}
(L^{(\Omega_N)}_{\mathrm{res}} F)(x)
&= \sum_{\eta \in \Omega_N} \sum_{k\in\mathbb{N}_0} N^{-2k} \int_{(0,1]} \Lambda^*_k (\mathrm{d} r)
\int_E y_{\eta,k}(\mathrm{d} a) \left[F\left(\Phi_{r,a,B_k(\eta)}(x)\right)-F(x)\right]\\
&\qquad + \sum_{\eta \in \Omega_N} (L^{d_0}_{\eta} F)(x),
\end{aligned}
\end{equation}
where
\be{ak:blockaverage}
y_{\eta,k} = N^{-k} \sum_{\zeta \in B_k(\eta)} x_\zeta
\end{equation}
is the $k$-block average of the components of $x$ in $B_k(\eta)$, $\Phi_{r,a,B_k(\eta)} \colon\,
\mathcal{P}(E)^{\Omega_N}\to\mathcal{P}(E)^{\Omega_N}$ is the \emph{reshuffling-resampling map} acting as
\be{ak:resampling-mapping-global}
\Big[\big(\Phi_{r,a,B_k(\eta)}\big)(x)\Big]_\zeta =
\begin{cases}
(1-r) y_{\eta,k} + r \delta_a, &\zeta \in B_k(\eta),\\
x_{\zeta}, &\zeta \in \Omega_N \backslash B_k(\eta),
\end{cases}
\quad r \in [0,1],\, a \in E,\, k\in\mathbb{N}_0,\,\eta\in\Omega_N,
\end{equation}
and $L^{d_0}_{\eta}$ is the \emph{Fleming-Viot diffusion operator} with volatility $d_0 \geq 0$,
acting on the colony $x_\eta$, given by
\be{ak:FlemingViot}
(L^{d_0}_{\eta}F)(x) = d_0 \int_E \int_E Q_{x_\eta}(\mathrm{d} u,\mathrm{d} v)\,
\frac{\partial^2 F(x)}{\partial x_\eta^2}[\delta_u,\delta_v]
\end{equation}
with
\be{ak:flemming-viot-kernel}
Q_y(\mathrm{d} u,\mathrm{d} v) = y(\mathrm{d} u)\,\delta_u(\mathrm{d} v) - y(\mathrm{d} u)\,y(\mathrm{d} v), \qquad y \in \mathcal{P}(E),
\end{equation}
the Fleming-Viot diffusion coefficient, and
\be{ak-second-variation}
\frac{\partial^2 F(x)}{\partial x_\eta^2 }[\delta_u,\delta_v]
= \frac{\partial}{\partial x_\eta} \left( \frac{\partial F(x)}{\partial x_\eta} [\delta_u] \right)[\delta_v],
\quad u,v \in E.
\end{equation}
\begin{remark}\label{r.711}
{\rm Note that the right-hand side of \eqref{ak:multi-level-resampling-global} is well-defined
because of assumption \eqref{ak1000}. Indeed, by Taylor-expanding the inner integral in
\eqref{ak:multi-level-resampling-global} in powers of $r$, we get
\be{ak:9999}
\int_E y_{\eta,k}(\mathrm{d} a)
\left[F\left(\Phi_{r,a,B_k(\eta)}(x)\right)-F(x)\right]
= F(y_{\eta,k})-F(x)+O(r^2),
\qquad r \downarrow 0.
\end{equation}
To have a well-defined resampling operator \eqref{ak:multi-level-resampling-global}, the expression
in \eqref{ak:9999} must be integrable with respect to $\Lambda^*_k (\mathrm{d} r)$, which is equivalent to
assumption~\eqref{ak1000}.} \hfill $\square$
\end{remark}
\medskip
The following proposition was proved in \cite{GHKK14}.
\begin{proposition}[{\bf Hierarchical martingale problem}]
\label{P.vecLambda}
\mbox{}\\
For every $x \in \mathcal{P}(E)^{\Omega_N}$, the martingale problem for $(L^{(\Omega_N)}, \mathcal{F},\delta_x)$
is well-posed. \footnote{As part of the definition of the martingale problem, we always require that
the solution has c\`adl\`ag paths and is adapted to the natural filtration.} The unique solution is a
strong Markov process with the Feller property. \hfill $\square$
\end{proposition}
\noindent
The Markov process arising as the solution of this martingale problem is denoted by
\be{XOmegaNdef}
X^{(\Omega_N)}=(X^{(\Omega_N)}(t))_{t\geq 0},
\end{equation}
and is referred to as the $C_N^{\underline{c},\underline{\Lambda}}$-process on $\Omega_N$. Proposition~\ref{P.vecLambda}
does not actually need the second condition in \eqref{ag}.
This condition will be needed only later.
\subsection{The hierarchical Cannings process in random environment}
\label{ss.random}
Our task in this section is to modify the first term in the right-hand side of
\eqref{ak:multi-level-resampling-global} so as to include the effect of a random environment on
the Cannings resampling mechanism. Section~\ref{sss.RE} defines the random environment,
Section~\ref{sss.mp} the modified generator.
\subsubsection{The random environment on the full tree}
\label{sss.RE}
Recall that $\Omega_N$ is the set of \textit{leaves} of the tree in Fig.~\ref{fig-hierargr}.
To introduce the random environment, we need to consider the \emph{full tree}, i.e.,
\be{fulltreedef}
\Omega^{\mathbb{T}}_N = \bigcup_{k\in\mathbb{N}_0} \Omega^{(k)}_N
\quad \text{ with } \quad \Omega^{(k)}_N = \Omega_N / B_k(0),
\end{equation}
where $\Omega_N/B_k(0)$ denotes the quotient group of $\Omega_N$ modulo $B_k(0)$, which
can be identified with the layer of the tree situated at height $k$ above the leaves. Indeed, because
$d_{\Omega_N}$ is an ultrametric distance (recall \eqref{ag31}), for each $k\in\mathbb{N}_0$ the set
$\Omega_N$ decomposes into disjoint balls of radius $k$, which can be labelled by the set
$\Omega^{(k)}_N$. For $\xi \in \Omega^{\mathbb{T}}_N$, we write
\be{heightdef}
|\xi| = \text{ the height of $\xi$ (counting from the leaves)},
\end{equation}
i.e., $|\xi| = k$ when $\xi \in \Omega^{(k)}_N$ for $k\in\mathbb{N}_0$, and we define
\be{Bkequivdef}
B_{|\xi|}(\xi)
\end{equation}
to be the set of sites in $\Omega_N$ that lie below $\xi$ (see Fig.~\ref{fig-hierartree}). We can
define the distance on the layer $\Omega^{(k)}_N$ as the graph distance to the most recent common
ancestor, and the distance on the full tree $\Omega^{\mathbb{T}}_N$ as the \emph{largest} of the
two graph distances to the most recent common ancestor (recall Fig.~\ref{fig-hierargr}).
The latter will be denoted by $d_{\Omega^{\mathbb{T}}_N}$. We write
\begin{equation}\label{e786}
\mathrm{MC}_k(\eta)
\end{equation}
to denote the {\em vertex in $\Omega^{\mathbb{T}}_N$ at height $k \in \mathbb{N}_0$ above $\eta \in \Omega_N$}.
This site carries the rate for the random walk on $\Omega_N$ to become uniformly distributed on the
$k-$ball around $\eta$. Since $\Omega^{\mathbb{T}}_N$ is isomorphic to $\Omega_N \times \mathbb{N}_0$,
we sometimes write $\mathrm{MC}_k(\eta) = \xi = (\eta,k)$.
\begin{figure}[htbp]
\vspace{0.3cm}
\centering \includegraphics[width=\textwidth]{full-tree.pdf}
\caption{\small $\Omega^{\mathbb{T}}_N $ with $N=3$, $\xi \in \Omega^{\mathbb{T}}_N$
with $|\xi| = k = 2$, and $\eta, \zeta \in B_{|\xi|}(\xi)$. The elements of $\Omega^{\mathbb{T}}_N$ are
the vertices of the tree (indicated by $\bullet$'s and $\Box$'s). The elements of $\Omega_N$ are the
leaves of the tree (indicated by $\Box$'s).}
\label{fig-hierartree}
\end{figure}
\medskip
We want to make the reshuffling-resampling \emph{spatially random}. To that end, we let
\be{ap1}
\underline{\Lambda}(\omega) = \big\{\Lambda^\xi(\omega) \colon\,
\xi \in \Omega^{\mathbb{T}}_N\big\}
\end{equation}
be a random field of $\mathcal{M}_f([0,1])$-valued resampling measures indexed by the tree.
\begin{itemize}
\item
Throughout the paper, we \emph{use the symbol $\omega$ to denote the random environment and
the symbol $\mathbb{P}$ to denote the law of $\omega$}.
\end{itemize}
In what follows, we assume that $\Lambda^\xi(\omega)$ is of the form
\be{ap2}
\Lambda^\xi(\omega) = \lambda_{|\xi|} \chi^\xi(\omega),
\end{equation}
where $\underline{\lambda} = (\lambda_k)_{k\in\mathbb{N}_0}$ is a deterministic sequence in $(0,\infty)$ (playing the role of
modulation coefficients) and
\be{chiprop}
\{\chi^\xi(\omega)\colon\,\xi\in\Omega_N^{\mathbb{T}}\}
\end{equation}
is a random field of $\mathcal{M}_f([0,1])$-valued resampling measures that is \emph{stationary under
translations} in $\Omega^{\mathbb{T}}_N$ (i.e., translations sideways and up in Fig.~\ref{fig-hierartree}),
and satisfies the conditions in \eqref{ag} when $\xi=0$ and the conditions in \eqref{ak1000} when
$\xi\neq 0$.
Abbreviate
\be{fdh:masschi}
\rho^\xi(\omega) = \chi^\xi(\omega)((0,1]),
\end{equation}
which is the total mass of $\chi^\xi(\omega)$. Clearly,
\be{rhoseq}
\{\rho^\xi(\omega)\colon\,\xi\in\Omega_N^{\mathbb{T}}\}
\end{equation}
is a random field of $(0,\infty)$-valued total masses that is also stationary under translations in
$\Omega^{\mathbb{T}}_N$. Throughout the paper, we assume that
\be{Aprop}
\mathbb{E}[\rho^\xi(\omega)]=1, \qquad \mathbb{E}[(\rho^\xi(\omega))^2] = C \in (0,\infty),
\end{equation}
and that the sigma-algebra at infinity associated with \eqref{fdh:masschi}, defined by
\be{fdh:tail}
\mathcal{T} = \bigcap_{L \in \mathbb{N}_0} \mathcal{F}_L, \qquad
\mathcal{F}_L = \sigma\Big(\rho^\xi(\cdot) ~\Big|~
\xi \in \Omega^{\mathbb{T}}_N\colon\,d_{\Omega^{\mathbb{T}}_N}(0,\xi) \geq L\Big),
\end{equation}
is trivial, i.e., all its events have probability 0 or 1 under the law $\mathbb{P}$. For one of the theorems
below we need to strengthen \eqref{Aprop} to
\be{Apropalt}
\mathbb{E}[\rho^\xi(\omega)]=1, \qquad
\exists\,\delta>0\colon\,\,\delta \leq \rho^\xi(\omega) \leq \delta^{-1}
\,\,\forall\,\xi\in\Omega_N \text{ for } \mathbb{P}\text{-a.e. } \omega.
\end{equation}
\subsubsection{The generator in random environment}
\label{sss.mp}
Throughout the sequel, we use the symbols $\eta,\zeta$ to denote elements of $\Omega_N$ and
the symbol $\xi$ to denote elements of $\Omega^{\mathbb{T}}_N$.
In random environment, we keep the definitions in Section~\ref{sss.hierarCan}, but we replace
the \emph{reshuffling-resampling operator} $L^{(\Omega_N)}_{\mathrm{res}}$ in
\eqref{ak:multi-level-resampling-global} by
\be{ak:multi-level-resampling-global-RE}
\begin{aligned}
(L^{(\Omega_N)}_{\mathrm{res}}(\omega) F)(x)
&= \sum_{\xi \in \Omega^\mathbb{T}_N} N^{-2|\xi|}
\int_{(0,1]} (\Lambda^\xi(\omega))^*(\mathrm{d} r) \int_E y_{\xi}(\mathrm{d} a)
\left[F\left(\Phi_{r,a,B_{|\xi|}(\xi)}(x)\right)-F(x)\right]\\
&\qquad + \sum_{\eta \in \Omega_N} (L_\eta^{d_0}F)(x)
\end{aligned}
\end{equation}
with $(\Lambda^\xi(\omega))^*(\mathrm{d} r) = \Lambda^\xi(\omega)(\mathrm{d} r)/r^2$, $r \in (0,1]$, where
$y_{\xi} \in \mathcal{P}(E)$ is given by
\be{def-blocks-RE}
y_{\xi}=N^{-|\xi|} \sum_{\zeta\in B_{|\xi|}(\xi)} x_\zeta, \quad \xi \in \Omega^\mathbb{T}_N,
\end{equation}
and $\Phi_{r,a,B_{|\xi|}(\xi)}\colon\,\mathcal{P}(E)^{\Omega_N}\to\mathcal{P}(E)^{\Omega_N}$ is the
\emph{reshuffling-resampling map} acting as
\be{ak:resampling-mapping-global-RE}
\Big[\big(\Phi_{r,a,B_{|\xi|}(\xi)}\big)(x)\Big]_\zeta =
\begin{cases}
(1-r) y_{\xi} + r \delta_a, &\zeta \in B_{|\xi|}(\xi),\\
x_{\zeta}, &\zeta \in \Omega_N \backslash B_{|\xi|}(\xi),
\end{cases}
\qquad r \in [0,1],\, a \in E,\,\xi\in\Omega^\mathbb{T}_N.
\end{equation}
The difference between \eqref{ak:multi-level-resampling-global} and
\eqref{ak:multi-level-resampling-global-RE} is that the resampling in blocks occurs according to
the resampling measure associated with the center of the block, labelled by $\Omega^\mathbb{T}_N$.
The full generator is
\be{ak:resmig}
L^{(\Omega_N)}(\omega) = L^{(\Omega_N)}_{\mathrm{mig}}
+ L^{(\Omega_N)}_{\mathrm{res}}(\omega)
\end{equation}
with $L^{(\Omega_N)}_{\mathrm{mig}}$ the migration operator in \eqref{ak:multi-level-migration}.
\section{Main theorems}
\label{s.results}
In Section~\ref{ss.resN} we present results for fixed $N$, in Section~\ref{ss.hiermfl} for
$N \to \infty$, the hierarchical mean-field limit. In Section~\ref{ss.effectre} we summarize the
effects of the random environment. Throughout the paper, the environment $\omega$ is fixed
and we use the symbol $\mathcal{L}[W]$ to denote the law of a random variable $W$.
\subsection{Results for fixed $N$}
\label{ss.resN}
Section~\ref{sss.wposmart} establishes the well-posedness of the martingale problem,
Section~\ref{sss.coexist} the convergence to an equilibrium that depends on $\omega$.
\subsubsection{Well-posedness of the martingale problem}
\label{sss.wposmart}
We begin by establishing that the martingale problem characterizes the process uniquely and
specifies a strong Markov process.
\begin{theorem}[{\bf Well-posedness of the martingale problem}]
\label{T.wpbasic}
Fix $N\in\mathbb{N} \backslash \{1\}$. For $\mathbb{P}$-a.s.\ $\omega$ and every $x \in \mathcal{P}(E)^{\Omega_N}$, the
$(L^{(\Omega_N)}(\omega),\mathcal{F},\delta_x)$-martingale problem is well-posed.
The unique solution is a strong Markov process with the Feller property. \hfill $\square$
\end{theorem}
The Markov process arising as the solution of the martingale problem is denoted by
\be{XOmegaNomegadef}
X^{(\Omega_N)}(\omega)=(X^{(\Omega_N)}(\omega;t))_{t\geq 0}
= ((X_\eta(\omega,t))_{\eta \in \Omega_N})_{t \geq 0},
\end{equation}
and is referred to as the hierarchical Cannings process on $\Omega_N$ in the environment
$\omega$. Theorem~\ref{T.wpbasic} does not actually need the second
condition in \eqref{ag}. This condition will be needed only later.
\subsubsection{Dichotomy: coexistence versus clustering}
\label{sss.coexist}
We next show that the law of our process converges to a limit law that depends on $\omega$.
\begin{theorem}[{\bf Equilibrium}]
\label{T.ltbeq}
Fix $N \in \mathbb{N}\backslash\{1\}$. Suppose that, under the law $\mathbb{P}$, the law of the initial state
$X^{(\Omega_N)}(\omega;0)$ is stationary and ergodic under translations in $\Omega_N$,
with mean single-coordinate measure
\be{ext}
\theta = \mathbb{E}\big[X_0^{(\Omega_N)} (\omega;0)\big] \in \mathcal{P}(E).
\end{equation}
Then, for $\mathbb{P}$-a.e.\ $\omega$, there exists an equilibrium measure $\nu^N_\theta(\omega) \in
\mathcal{P}(\mathcal{P}(E)^{\Omega_N})$, arising as
\be{ag4.9}
\lim_{t\to\infty} \mathcal{L}\big[X^{(\Omega_N)}(\omega;t)\big] = \nu^N_\theta(\omega),
\end{equation}
satisfying
\be{ag4.2}
\int_{\mathcal{P}(E)^{\Omega_N}} x_0\,\nu^N_\theta(\omega)(\mathrm{d} x) = \theta.
\end{equation}
Moreover, under the law $\mathbb{P}$, $\nu^N_\theta(\omega)$ is stationary and ergodic under translations
in $\Omega_N$. \hfill $\square$
\end{theorem}
\noindent
Note that $\nu^N_\theta(\omega)$ depends on $\omega$ even though its mean single-coordinate
measure $\theta$ (which is determined by the initial state) does not. The proof of Theorem~\ref{T.ltbeq}
is based on a computation with the dual hierarchical Cannings process, which allows us to control
second moments. As we will see in Section~\ref{s.randomwalk}, in random environment this computation
is delicate because it involves two random walks in the same environment, and the \emph{difference}
of these two random walks is \emph{not} a random walk itself, like in the average environment.
Using the stationarity and ergodicity of $\nu^N_\theta$, we next identify the parameter regime for
which $\nu^N_\theta(\omega)$ is a \emph{multi-type equilibrium} (= coexistence given $\omega$),
i.e.,
\be{ag4.8}
\sup_{f \in C_{\mathrm{b}}(E)} \int_{\mathcal{P}(E)^{\Omega_N}}
\nu^N_\theta(\omega)(\mathrm{d} x) \int_E [f(u)-\theta]^2\,x_0(\mathrm{d} u) > 0,
\end{equation}
respectively, a \emph{mono-type equilibrium} (= clustering given $\omega$), i.e.,
\be{ag4.6}
\nu^N_\theta(\omega) = \int_E \delta_{(\delta_u)^{\Omega_N}} \theta(\mathrm{d} u).
\end{equation}
The two regimes are complementary. In the latter regime the system grows mono-type clusters that
eventually cover all finite subsets of $\Omega_N$ (where types may or may not change infinitely
often).
\begin{theorem}[{\bf Dichotomy for finite $N$}]
\label{T.coexcritNfin}
Fix $N \in \mathbb{N} \backslash \{1\}$ and assume \eqref{Apropalt}.
\begin{itemize}
\item[\textup{(a)}]
Let $\mathcal{C}_N=\{\omega\colon\,\text{in $\omega$ coexistence occurs}\}$. Then
$\mathbb{P}(\mathcal{C}_N) \in \{0,1\}$.
\item[\textup{(b)}]
$\mathbb{P}(\mathcal{C}_N)=1$ if and only if
\be{Ndich}
\sum_{k\in\mathbb{N}_0} \frac{1}{c_k+N^{-1}\lambda_{k+1}} \sum_{l=0}^k \lambda_l<\infty.
\end{equation}
\end{itemize}
\hfill $\square$
\end{theorem}
\noindent
Cox and Klenke~\cite{CK00} give a criterion in the clustering regime for when the type at a given
site changes infinitely often. For interacting Fleming-Viot processes they show that this happens
as soon as $\theta$ is not a $\delta$-measure. Because of the reasoning in Section~\ref{s.randomwalk},
we therefore get the following.
\begin{corollary}[{\bf Change of types}]
In the clustering regime, if $\theta \neq \delta_u$ for some $u \in E$, then at every site the type
changes infinitely often. \hfill $\square$
\end{corollary}
\subsection{Results for $N\to\infty$}
\label{ss.hiermfl}
Our remaining theorems capture the space-time scaling behaviour of our process in the
hierarchical mean-field limit $N \to \infty$. In this limit, the degree of recurrence/transience
$\gamma(N)$ tends to $0$ while the Hausdorff dimension $d(N)$ in the metric $\eta\mapsto
e^{|\eta|}$ tends to $2$ (recall Remark~\ref{degreeretr}), so that our process becomes
\emph{near-critical}.
In Section~\ref{sss.mfl} we introduce a key process, called the \emph{McKean-Vlasov process},
which naturally arises in this limit. In Section~\ref{sss.REinfty} we define the random environment
for $N=\infty$. In Section~\ref{sss.blav} we look at the block averages on successive space-time
scales and show that as $N\to\infty$ these converge to a sequence of McKean-Vlasov processes
with \emph{renormalized volatilities}. In Section~\ref{sss.vola} we identify the scaling behaviour
of the volatility on hierarchical scale $k$ in the limit as $k\to\infty$, which leads to various different
cases as a function of $\underline{c}$ and $\underline{\Lambda}$. In Section~\ref{sss.coex} we identify the parameter regimes
that correspond to coexistence, respectively, clustering. In Section~\ref{sss.clust} we link the different
cases of scaling to five universality classes of cluster formation.
\subsubsection{McKean-Vlasov process}
\label{sss.mfl}
We need some definitions and basic facts about the McKean-Vlasov process from \cite{GHKK14}.
Let $\mathcal{F} \subseteq C_\mathrm{b}(\mathcal{P}(E),\mathbb{R})$ be the algebra of functions $F$ of the form
\be{ak:multi-level-test-functions*}
F(y) = \int_{E^n} y^{\otimes n}(\mathrm{d} u)\,f(u),
\qquad y \in \mathcal{P}(E),\, n \in \mathbb{N},\,f \in C_{\mathrm{b}}(E^n,\mathbb{R}).
\end{equation}
For $c,d \in [0,\infty)$, $\Lambda\in\mathcal{M}_f([0,1])$ subject to (\ref{ag}) and $\theta\in\mathcal{P}(E)$, let
$L_\theta^{c,d,\Lambda}\colon\,\mathcal{F} \to C_\mathrm{b}(\mathcal{P}(E),\mathbb{R})$ be the linear operator
\be{generic-interaction-operator}
L_\theta^{c,d,\Lambda} = L^c_\theta + L^d + L^\Lambda
\end{equation}
acting on $F\in \mathcal{F}$ as (recall \eqref{ak:flemming-viot-kernel})
\be{Ldefs}
\begin{aligned}
(L^c_\theta F)(y) &= c \int_E \left(\theta-y\right)(\mathrm{d} a)\,
\frac{\partial F(y)}{\partial y}[\delta_a],\\
(L^dF)(y) &= d \int_E \int_E Q_y(\mathrm{d} u,\mathrm{d} v)\,
\frac{\partial^2 F(y)}{\partial y^2}[\delta_u,\delta_v],\\
(L^\Lambda F)(y) &= \int_{(0,1]} \Lambda^*(\mathrm{d} r) \int_E y(\mathrm{d} a)\,
\big[F\big((1-r)y+ r\delta_a\big)-F(y)\big].
\end{aligned}
\end{equation}
The three parts of $L_\theta^{c,d,\Lambda}$ correspond to: (1) a \emph{drift} towards $\theta$ of
strength $c$ (``immigration-emigration''); (2) a \emph{Fleming-Viot diffusion} with \emph{volatility}
$d$ (``Moran resampling''); (3) a \emph{Cannings process} with \emph{resampling measure}
$\Lambda$ (``Cannings resampling''). This model arises as the $M \to\infty$ limit of an individual-based
model with $M$ individuals at a single site, with {\em immigration} at rate $c$ from a constant source
with type distribution $\theta \in \mathcal{P}(E)$, {\em emigration} at rate $c$ to a cemetery state, diffusive
resampling at rate $d$, and $\Lambda$-resampling.
The following proposition was proved in \cite{GHKK14}.
\begin{proposition}[{\bf McKean-Vlasov martingale problem}]
\label{prop:McKean-Vlasov-well-posedness} $\mbox{}$
\begin{itemize}
\item[\textup{(a)}]
For every $y \in \mathcal{P}(E)$, the martingale problem for $(L_\theta^{c,d,\Lambda}, \mathcal{F},\delta_y)$
is well-posed. The unique solution is a strong Markov process with the Feller property.
\item[\textup{(b)}]
For every $c \in (0,\infty)$, the solution from \textup{(a)} is ergodic in time with unique equilibrium
measure $\nu_\theta^{c,d,\Lambda}$. For $c=0$, the solution from \textup{(a)} is not ergodic in
time, and $\nu_\theta^{0,d,\Lambda}$ is the unique equilibrium measure obtained as the $t\to\infty$
limit with initial state $y=\theta$.
\item[\textup{(c)}]
For $c > 0$,
\begin{equation}
\label{eq:equi-ident-c}
\nu_\theta^{c,d,\Lambda} = \nu_\theta^{1,d/c,\Lambda/c}.
\end{equation}
\end{itemize}
\hfill $\square$
\end{proposition}
Denote by
\be{Zdef}
Z_\theta^{c,d,\Lambda} = \big(Z_\theta^{c,d,\Lambda}(t)\big)_{t\geq 0},
\quad Z^{c,d,\Lambda}_\theta (0) = \theta,
\end{equation}
the solution of the martingale problem in Proposition~\ref{prop:McKean-Vlasov-well-posedness}
for the special choice $y=\theta$. This is called the \emph{McKean-Vlasov process} with parameters
$c,d,\Lambda$ and initial state $\theta$.
\subsubsection{Random environment for $N = \infty$}
\label{sss.REinfty}
In order to be able to pass to the limit $N\to\infty$, we need to define a random environment for
$N=\infty$ in which all the random environments for finite $N$ are embedded. To that end, define
$\Omega_\infty = \oplus_\mathbb{N} \mathbb{N}$, and let (recall \eqref{fulltreedef})
\be{e1096}
\Omega^{\mathbb{T}}_\infty = \bigcup_{k\in\mathbb{N}_0} \Omega^{(k)}_\infty
\quad \text{ with } \quad \Omega^{(k)}_\infty = \Omega_\infty / B_k(0).
\end{equation}
Note that for any $N \in \mathbb{N}$ there is a natural embedding of $\Omega^{\mathbb{T}}_N$ into
$\Omega^{\mathbb{T}}_\infty$. Similarly as in Section~\ref{sss.RE}, we let
\be{reNinfinite}
\underline{\Lambda}(\omega) = \big\{\Lambda^\xi(\omega) \colon\,
\xi \in \Omega^{\mathbb{T}}_\infty \big\}
\end{equation}
be a random field of $\mathcal{M}_f([0,1])$-valued resampling measures index by the full tree, where
$\omega$ again denotes the random environment. We retain the symbol $\mathbb{P}$ for the law of of
$\omega$. As in \eqref{ap2}--\eqref{fdh:tail}, we assume that $\Lambda^\xi(\omega) = \lambda_{|\xi|}
\chi^\xi(\omega)$ where, under the law $\mathbb{P}$, $\{\chi^\xi(\omega)\colon\,\xi \in\Omega^{\mathbb{T}}_\infty\}$
is stationary under translations in $\Omega^{\mathbb{T}}_\infty$, and is such that the total masses
$\rho^\xi(\omega) = \chi^\xi(\omega)((0,1])$ have first moment equal to $1$, second moment finite,
and a trivial sigma-algebra at infinity. For any $N \in \mathbb{N}$, the natural restriction of the random field in
\eqref{reNinfinite} equals the random field in \eqref{ap1}.
\subsubsection{Renormalization via block averages}
\label{sss.blav}
For each $k \in \mathbb{N}_0$, we look at the \emph{$k$-block averages} defined by (recall Fig.~\ref{fig-hierargr})
\be{k-block-average}
Y_{\eta,k}^{(\Omega_N)}(\omega;t) = \frac{1}{N^k}
\sum_{\zeta\in B_k(\eta)} X_{\zeta}^{(\Omega_N)}(\omega;t),
\qquad \eta \in \Omega_N,
\end{equation}
which constitute a {\em renormalization of space} where the component $\eta$ is replaced by the average
of the components in $B_k(\eta)$. After a corresponding {\em renormalization of time} where $t$ is
replaced by $tN^k$, i.e., $t$ is the associated macroscopic time variable, we obtain a {\em renormalized}
interacting system
\be{a3}
\left(\left(Y^{(\Omega_N)}_{\eta,k}(\omega;t N^k)
\right)_{\eta \in \Omega_N}\right)_{t \geq 0},
\qquad k \in \mathbb{N}_0,\,\eta \in \Omega_N,
\end{equation}
which is constant in $B_k(\eta)$ and can be viewed as an interacting system indexed by the set
$\Omega^{(k)}_N$ (see Fig.~\ref{fig-hierargr}). This provides us with a \emph{sequence of
renormalized interacting systems}, which for fixed $N$ are {\em not} Markov.
The key ingredient to study the $ N \to \infty $ limit of \eqref{a3} is the following.
Let $\underline{d}=(d_k)_{k\in\mathbb{N}_0}$ be the sequence of \emph{volatility constants} defined recursively as
\be{diffusion-constants}
d_{k+1} = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{c_k(\mu_k\rho +d_k)}{c_k + (\mu_k\rho + d_k)}\right],
\quad k \in \mathbb{N}_0,
\end{equation}
where
\be{mulambdarel}
\mu_k=\tfrac12\lambda_k, \quad k\in\mathbb{N}_0,
\end{equation}
$\rho$ is the $(0,\infty)$-valued random variable whose law $\mathcal{L}_\rho$ is the same as that of
$\rho^0(\omega)$ under $\mathbb{P}$ (recall (\ref{Aprop}--\ref{fdh:tail})), and $\mathbb{E}_{\mathcal{L}_\rho}$ is expectation
w.r.t.\ $\mathcal{L}_\rho$. For fixed $\underline{c}$, $\underline{\Lambda}$ and $d_0$, the recursion in \eqref{diffusion-constants}
determines $\underline{d}$. The right-hand side is the \emph{average of a random M\"obius transformation
that depends on $\rho$}. Recall that $\rho$ has mean 1.
\paragraph{Heuristics behind the recursion formula for the volatilities.}
In order to understand the recursion formula in \eqref{diffusion-constants}, we consider the 1-block
around the origin $0$ on time scale $Nt$ and let $N\to\infty$. Note that, in this limit, the time scales
for the jumps to different levels separate (recall \eqref{32b}), so that we can focus on each of the time
scales separately.
If we randomly draw two lineages from the 1-block and ask whether they have a common ancestor
some time back (so that they are of the same type), then we get exactly the event that generates
the variance of the 1-block average (otherwise the lineages and their types would be independent
and would have an asymptotically vanishing contribution to the variance). The fact that the lineages
behave like a spatial coalescent follows from the duality introduced in Section~\ref{s.model}. The
lineages have to meet in order to have a common ancestor, which takes them a time of order $Nt$.
Note that triples of lineages have a negligible probability to meet at times of order $Nt$ in the limit of
$N \to \infty$.
If the lineages meet, then they may coalesce. This happens at rate
\begin{equation}\label{e1175}
\lambda^{(\eta,0)}(\omega)=\Lambda^{(\eta,0)}((0,1])(\omega)
\end{equation}
when they both sit at $\eta\in\Omega$. However, they may also move before they coalesce, i.e., make
a migration jump away, which happens with probability $2c_0 /(2c_0+\lambda^{(\eta,0)}(\omega))$.
Hence the effective coalescence rate is $\lambda^{(\eta,0)}(\omega)[2c_0/(2c_0+\lambda^{(\eta,0)}
(\omega))]$. Since the vertex where the lineages meet is uniformly distributed over the 1-block, the
average rate is given by
\begin{equation}\label{e1184}
\mathbb{E}\left[\frac{2c_0\lambda_0\rho(\omega)}{2c_0+\lambda_0\rho(\omega)}\right],
\end{equation}
where we use that $\lambda^{(\eta,0)}(\omega)$ has the same distribution as $\lambda_0\rho(\omega)$
(recall \eqref{ap2}--\eqref{fdh:masschi}). If we would have a diffusive part as well, at constant rate
$2d_0 $, then the lineages would coalesce at the same rate but with $\lambda(\omega)$ replaced
by $2d_0+\lambda(\omega)$. Since the volatility turns out to be equal to this rate, we get the recursion
formula
\begin{equation}\label{e1192}
2d_1=\mathbb{E}\left[\frac{2c_0(2d_0+\lambda_0\varrho(\omega))}
{2c_0+(2d_0+\lambda_0\varrho(\omega))}\right].
\end{equation}
By the same reasoning for $k$-blocks on time scale $tN^k$, we get a heuristic explanation for the
recursion formula in \eqref{diffusion-constants}.
\medskip
Our next theorem states that for each $k\in\mathbb{N}_0$ the $k$-block averages in the limit as $N\to\infty$
evolve according to the McKean-Vlasov process defined in Section~\ref{sss.mfl} with certain
$k$-dependent parameters.
\begin{theorem}[{\bf Hierarchical mean-field limit and renormalization}]
\label{mainth}
Suppose that for each $N$ the random field $X^{(\Omega_N)}(\omega;0)$ is the restriction to
$\Omega_N$ of a random field $X(\omega)$ indexed by $\Omega_\infty = \bigoplus_\mathbb{N} \mathbb{N}$
that is i.i.d.\ with single-component mean $\theta\in\mathcal{P}(E)$. Then, for $\mathbb{P}$-a.e.\ $\omega$ and
every $k\in\mathbb{N}$ and $\eta\in\Omega_\infty$,
\be{macroscopic-behaviour}
\lim_{N\to\infty} \mathcal{L}
\left[\left(Y_{\eta,k}^{(\Omega_N)}(\omega;t N^k)\right)_{t\geq 0}\right]
= \mathcal{L}\left[\left(Z_\theta^{c_k,d_k,\Lambda^{\mathrm{MC}_k(\eta)}
(\omega)}(t)\right)_{t\geq 0}\right],
\end{equation}
where
\be{MCdef}
\mathrm{MC}_k(\eta) = \text{ unique site in $\Omega^\mathbb{T}_\infty$ at height $k$ above
$\eta\in\Omega_\infty$},
\end{equation}
i.e., the label of the block (= macro-colony) of radius $k$ in $\Omega_\infty$ around $\eta\in
\Omega_\infty$ (see \textup{Fig.}~{\rm \ref{fig-hierargr}}). The same is true for $k=0$
when the initial condition for the McKean-Vlasov process in the right-hand side of
\eqref{macroscopic-behaviour} is $Z_\theta^{c_0,d_0,\Lambda^{\eta}(\omega)}(0)
=X^{(\Omega_N)}(\omega;0)$ instead of $Z_\theta^{c_0,d_0,\Lambda^{\eta}(\omega)}(0)
=\theta$. \hfill $\square$
\end{theorem}
Note that among the parameters $c_k,d_k,\Lambda^{\mathrm{MC}_k(\eta)}(\omega)$ of the
limiting McKean-Vlasov process, the volatility $d_k$ is the result of a \emph{self-averaging}
with respect to the random environment up to and including level $k$, as exemplified by
\eqref{diffusion-constants}. \emph{It is through this recursion relation that the renormalization
manifests itself.}
Our next theorem looks at successive block averages simultaneously.
\begin{theorem}[{\bf Multi-scale analysis and the interaction chain}]
\label{T.InteractionChain}
Let $(t_N)_{N\in\mathbb{N}}$ be such that
\begin{equation}\label{e1223}
\begin{aligned}
\lim_{N\to\infty} t_N=\infty \text{ and } \lim_{N\to\infty}t_N/N=0.
\end{aligned}
\end{equation}
Then, for $\mathbb{P}$-a.e.\ $\omega$, every $j\in\mathbb{N}$ and every $\eta\in\Omega_\infty$,
\be{ag:interaction-chain}
\lim_{N\to\infty} \mathcal{L}\left[
\left( Y^{(\Omega_N)}_{\eta, k}(\omega; t_N N^k) \right)_{k = -(j+1), -j, \ldots, 0}\right]
= \mathcal{L}\left[\left(M^{(j)}_{\eta, k}(\omega)\right)_{k = -(j+1), -j, \ldots, 0}\right],
\end{equation}
where $M^{(j)}_\eta(\omega)=(M^{(j)}_{\eta, k}(\omega))_{k = -(j+1), -j, \ldots, 0}$ is the
time-inhomogeneous Markov chain with initial state
\be{ag:interaction-chain-initial}
M^{(j)}_{\eta, -(j+1)}(\omega) = \theta,
\end{equation}
and transition kernel from time $-(k+1)$ to $-k$ given by
\be{ag:interaction-chain-transition-kernel}
K_{\eta, k}(\omega; \theta, \cdot)
= \nu^{c_k, d_k, \Lambda^{\mathrm{MC}_k(\eta)}(\omega)}_\theta(\cdot).
\end{equation}
\hfill $\square$
\end{theorem}
\noindent
The right-hand side of \eqref{e1223} describes the large space-time scaling behaviour of our
hierarchical Cannings process.
\begin{definition}[{\bf Interaction chain}]
\label{def:interaction-chain}
$M^{(j)}_{\eta}(\omega)$ is called the interaction chain at level $j$ at location $\eta \in
\Omega_\infty$ given $\omega$. \hfill $\square$
\end{definition}
\begin{remark}
\label{r.1261}
{\rm Theorem \ref{T.InteractionChain} only specifies the limiting distribution of the one-dimensional
spatial marginals, i.e., the single interaction chains. Similarly as in Dawson, Greven and
Vaillancourt~{\rm \cite[Section~0e]{DGV95}}, it is possible to also specify the joint distribution
of the interaction chains, which can be viewed as a \emph{field of Markov chains indexed by}
$\Omega^{\mathbb{T}}_\infty$.} \hfill $\square$
\end{remark}
An important characteristic of $ M^{(j)}_\eta$ is the variance of $ M^{(j)}_{\eta,0}$, calculated
as
\begin{equation}\label{e1268}
\var \langle M_{\eta,0}^{(j)}, f \rangle
= \prod^{j}_{k=0} \frac{2c_k}{2c_k + \lambda_k\rho_k + 2d_k} \var_\theta(f).
\end{equation}
This shows that a key ingredient for $M_\eta^{(j)}$ is the sequence of volatilities $\underline{d}=(d_k)_{k
\in \mathbb{N}_0}$ and the way this sequence grows or decays. How is this affected by the randomness of
the environment?
Our next theorem shows that the volatility $d_k$ in the random environment can be sandwiched
between the volatility $d^0_k$ in the zero environment ($\mathcal{L}_\rho=\delta_0$, i.e., the system
without resampling) and the volatility $d^1_k$ in the average environment ($\mathcal{L}_\rho=\delta_1$,
i.e., the system with average resampling).
\begin{theorem}[{\bf Randomness lowers volatility}]
\label{T.order}
If $d^0_0=d_0 = d^1_0$, then $d^0_k<d_k < d^1_k$ for all $k \in \mathbb{N}$. \hfill $\square$
\end{theorem}
\subsubsection{Dichotomy for the interaction chain}
\label{sss.coex}
How are the qualitative properties of the Cannings process for large $N$ reflected in the interaction
chain? What about the dichotomy clustering versus coexistence? Before answering these questions
we need to first establish the existence of the entrance law of the interaction chain from level $\infty$,
which we will obtain from the level $j$ interaction chain as limit $j\to\infty$. With this object, we can
address the question of coexistence versus clustering.
\begin{proposition}[{\bf Entrance law of interaction chain exists}]\label{p.1267}\mbox{}\\
The limit as $j\to\infty$ of $M^{(j)}_\eta$ exists. \hfill $\square$
\end{proposition}
The object corresponding to the equilibrium of the stochastic system for finite $N$ in the
hierarchical mean-field limit $N \to \infty$ is the {\em field of entrance laws} of the interaction
chain from level $\infty$ (recall Remark~\ref{r.1261}), in particular, its marginal law $\Pi_\eta
\nu_\theta(\omega)$ at level $0$ in $\eta$, which is element of $\mathcal{P}(\mathcal{P}(E)$.
\begin{definition}[{\bf Entrance law of interaction chain}]
\label{p:interaction-chain-equilibrium}
For $\mathbb{P}$-a.e.\ $\omega$ and all $\eta \in \Omega_\infty$,
\be{hierarch-mean-field-equil}
\lim_{j\to\infty} \mathcal{L}\left[M^{(j)}_{\eta, 0}(\omega)\right]
= \Pi_\eta \nu_\theta(\omega),
\end{equation}
where $\nu_\theta (\omega) \in \mathcal{P}(\mathcal{P}(E))^{\Omega_\infty}$ is the entrance law from
level $\infty$ of the (tree-indexed) interaction chain at level $0$, and $\Pi_\eta\nu_\theta(\omega)$
denotes the projection of $\nu_\theta(\omega)$ on $\eta$.
\hfill $\square$
\end{definition}
Our next theorem is indeed the analogue of Theorem~\ref{T.coexcritNfin} for $N \to \infty$. In this
limit, coexistence and clustering in $\omega$ are defined for $(M^{(\infty)}_{\eta,0})_{\eta\in\Omega_N}$
in the same way as in \eqref{ag4.8}--\eqref{ag4.6}.
\begin{theorem}[{\bf Dichotomy for $N=\infty$}]
\label{T.dicho}
$\mbox{}$
\begin{itemize}
\item[\textup{(a)}]
Let $\mathcal{C}=\{\omega\colon\, \text{in $\omega$ coexistence occurs}\}$.
Then $\mathbb{P}(\mathcal{C}) \in \{0,1\}$.
\item[\textup{(b)}]
$\mathbb{P}(\mathcal{C})=1$ if and only if
\be{inftydich}
\sum_{k\in\mathbb{N}_0} \frac{1}{c_k} \sum_{l=0}^k \lambda_l<\infty.
\end{equation}
\end{itemize}
\hfill $\square$
\end{theorem}
\noindent
Note that condition \eqref{inftydich} is the limit of condition \eqref{Ndich} as $N\to\infty$.
In fact, the two conditions are equivalent when the following \emph{weak regularity condition}
holds:
\be{dichreg}
\text{ either } \quad \limsup_{k\to\infty} \frac{\lambda_{k+1}}{c_k} < \infty \quad \text{ or } \quad
\liminf_{k\to\infty} \left(\frac{\lambda_{k+1}}{c_k} \wedge \frac{\lambda_k}{\lambda_{k+1}}\right)>0.
\end{equation}
An important question is whether the equilibrium measure $\nu_\theta (\omega)$ is the limit
as $N\to\infty$ of the equilibrium measure $\nu_\theta^N (\omega)$ (recall \eqref{ag4.9}). The
answer is yes. We only prove the following.
\begin{corollary}[{\bf Hierarchical mean field limit of equilibrium}]
\label{c.1282}
$\mbox{}$\\
For $ \mathbb{P}$-a.s.\ all $\omega$ and all $\eta \in \Omega_\infty$,
\begin{equation}\label{e1309}
\lim_{N\to\infty} \Pi_\eta \nu^N_\theta = \Pi_\eta \nu_\theta.
\end{equation}
\hfill $\square$
\end{corollary}
\subsubsection{Scaling of the volatility}
\label{sss.vola}
We are interested in the behaviour of the variance of the interaction chain $M^{(j)}$ as $j\to\infty$,
since this allows us to identify {\em universality classes} for the scaling behaviour of our stochastic
system. From the variance formula, we see that $(d_k)_{k \in \mathbb{N}_0}$ is the key input, and so we
study this sequence first. Note that the variance formula really depends on the ratios $d_k/c_k,
\mu_k/d_k$, which we encounter below.
Our next two theorems identify the scaling behaviour of $d_k$ as $k\to\infty$ \emph{in the regime
of clustering}. The first theorem considers the case of \emph{polynomial coefficients}, i.e.,
\be{fdh:regcond}
c_k \sim L_c(k)\,k^a, \qquad \mu_k \sim L_\mu(k)\,k^b, \qquad k \to \infty,
\end{equation}
with $a,b \in \mathbb{R}$ and $L_c,L_\mu$ slowly varying at infinity.
In what follows, we assume that
\be{Kdef}
K = \lim_{k\to\infty} \frac{\mu_k}{c_k} \in [0,\infty], \qquad
L = \lim_{k\to\infty} \frac{k^2\mu_k}{c_k} \in [0,\infty],
\end{equation}
exist and write $ K_k $ and $ L_k $ for the respective sequences. There are \emph{five cases} according
to the values of $K$ and $L$. Four of these, labelled (a)--(d), we can analyze in detail. For the remaining
case, labelled (e), see Remark~\eqref{r:notexhaus}. For cases (c)--(d), we need extra regularity conditions
on $L_c,L_\mu$ in \eqref{fdh:regcond}, for which we refer the reader to \cite[Eqs.~(1.79)--(1.81)]{GHKK14}.
\begin{theorem}[{\bf Scaling of the Fleming-Viot volatility: polynomial coefficients}]
\label{T.scaleFVpol}
$\mbox{}$\\
Under the polynomial scaling assumptions \eqref{fdh:regcond}--\eqref{Kdef}, the following cases apply:
\begin{itemize}
\item[\textup{(a)}]
If $K = \infty$, then $\lim_{k\to\infty} d_k/c_k = 1$.
\item[\textup{(b)}]
If $K \in (0,\infty)$, then $\lim_{k\to\infty} d_k/c_k = M$ with $M \in (0,1)$ the unique solution of the
equation
\be{MK}
M = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{(K\rho+M)}{1+(K\rho+M)}\right].
\end{equation}
\item[\textup{(c)}]
If $K = 0$ and $L = \infty$, then $\lim_{k\to\infty} d_k/\sqrt{c_k\mu_k} = 1$.
\item[\textup{(d)}]
If $K=0$, $L \in [0,\infty)$ and $a \in (-\infty,1)$, then $\lim_{k\to\infty} \sigma_kd_k = M$ with
$\sigma_k = \sum_{l=0}^{k-1} (1/c_l)$ and $M \in [1,\infty)$ given by
\be{M*K}
M = \tfrac12\left[1+\sqrt{1+4L/(1-a)^2}\right].
\end{equation}
\end{itemize}
\hfill $\square$
\end{theorem}
\begin{remark}\label{r:notexhaus}
{\rm It is straightforward to check with the help of \eqref{fdh:regcond}--\eqref{Kdef} that all four
cases {\rm (a)-(d)} correspond to choices of $\underline{c}$ and $\underline{\lambda}$ for which clustering holds, i.e., the
sum in \eqref{inftydich} diverges (note that $\lim_{k\to\infty} \sigma_k=\infty$ in case {\rm (d)}).
However, they are not exhaustive: there is a fifth case (e), corresponding to $K=0$, $L \in [0,\infty)$,
$a=1$ and $\lim_{k\to\infty} \sigma_k=\infty$, for which we have no scaling result. This case lies
at the border of the clustering regime. An example is $c_k \sim k (\log k)^\gamma$, $\gamma
\in (-\infty,1]$, and $\mu_k = k^{-2} c_k$, which we were able to handle in the deterministic model
in \cite{GHKK14}, but cannot handle in the random model treated here.} \hfill $\square$
\end{remark}
The second theorem considers the case of \emph{exponential coefficients}, i.e.,
\be{expregvar}
c_k = c^k\bar{c}_k, \qquad \mu_k= \mu^k\bar{\mu}_k
\end{equation}
with $c,\mu \in (0,\infty)$ and $\bar{c}_k,\bar{\mu}_k$ satisfying \eqref{fdh:regcond} with
exponents $a,b$. We further assume that
\be{barK}
\bar{K} = \lim_{k \to \infty} \frac{\bar{\mu}_k}{\bar{c}_k} \in [0, \infty], \mbox{ we write }
\bar K_k = \frac{\bar \mu_k}{\bar c_k}.
\end{equation}
exists.
\begin{theorem}[{\bf Scaling of the Fleming-Viot volatility: exponential coefficients}]
\label{T.scaleFVexp}
$\mbox{}$\\
Under the exponential scaling assumptions in \eqref{expregvar}--\eqref{barK}, the following cases
apply (cf.\ \textup{Theorem~\ref{T.scaleFVpol}}):
\begin{itemize}
\item[\textup{(A)}]
\textup{[Like Case~(a)]}
$c<\mu$, or $c=\mu$ and $\bar{K}=\infty$: $\lim_{k\to\infty} d_k/c_k=1/c$.
\item[\textup{(B)}]
\textup{[Like Case~(b)]}
$c=\mu$ and $\bar{K} \in (0,\infty)$: $\lim_{k\to\infty} d_k/c_k = \bar{M}/c$ with $\bar{M} \in (0,1)$
the unique solution of the equation
\be{Mexp}
\bar{M} = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{(cK\rho+\bar{M})}{c+(cK\rho+\bar{M})}\right].
\end{equation}
\item[\textup{(C)}]
The case $\bar{K}=0$, with $c=\mu$ or $c>\mu$, splits into three sub-cases:
\begin{itemize}
\item[\textup{(C1)}]
\textup{[Like Case~(b)]}
$c=\mu<1$, $\bar{K}=0$: $\lim_{k\to\infty} d_k/c_k=(1-c)/c$.
\item[\textup{(C2)}]
\textup{[Like Case \textup{(c)}]}
$c=\mu>1$, $\bar{K}=0$, $\sum_{k\in\mathbb{N}_0} \bar{K}_k = \infty$:
\footnote{In \cite{GHKK14} the condition $\sum_{k\in\mathbb{N}_0} \bar{K}_k = \infty$ was mistakenly omitted.}
$\lim_{k\to\infty} d_k/\mu_k
= 1/(\mu-1)$.
\item[\textup{(C3)}]
\textup{[Like Case~(d)]}
$1>c>\mu$ and $\bar{K}=0$, or $1=c>\mu$, $\bar{K}=0$ and $a \in (-\infty,1)$:
$\lim_{k\to\infty} \sigma_k d_k = 1$.
\end{itemize}
\end{itemize}
\hfill $\square$
\end{theorem}
\noindent
The same observation as in Remark~\ref{r:notexhaus} applies. Again, the critical case $a=1$ is
missing in (C3).
\subsubsection{Cluster formation}
\label{sss.clust}
Within the clustering regime it is of interest to study the size of the mono-type regions as a function of
time, i.e., how fast the clusters where one type prevails grow.
This question has been addressed for other population models. For the voter model on $\mathbb{Z}^2$, Cox
and Griffith~\cite{CG86} showed that the radii of the clusters with opinion ``all 1'' or ``all 0'' scale as
$t^{\alpha/2}$ with $\alpha \in [0,1)$, i.e., clusters occur on all scales $\alpha \in [0,1)$. For the model
of hierarchically interacting Fleming-Viot diffusions with $c_k \equiv 1$ (= critically recurrent migration),
Fleischmann and Greven~\cite{FG94} showed that, for all $N \in \mathbb{N} \setminus \{1\}$ and all $\eta \in
\Omega_N$,
\be{allim}
\lim_{t\to\infty}
\mathcal{L}\left[\left(Y^{(\Omega_N)}_{\eta,\lfloor(1-\alpha)t\rfloor}(N^t)\right)_{\alpha \in [0,1)}\right]
= \mathcal{L}\left[\left(Y\left(\log\left(\frac{1}{1-\alpha}\right)\right)\right)_{\alpha \in [0,1)}
\right]
\end{equation}
in the sense of finite-dimensional distributions, where $(Y(t))_{t \in [0,\infty)}$ is the standard
Fleming-Viot diffusion on $\mathcal{P}(E)$. A similar behaviour occurs for other models, e.g.\ branching
models as shown in Dawson and Greven~\cite{DG96}.
The advantage of the hierarchical group is that we can analyze the cluster formation as a function
of $N$ and let $N \to \infty$ to approach the critically recurrent case (recall Remark~\ref{degreeretr}).
We can do this by using the {\em interaction chain}. In \cite{GHKK14} we analysed the Cannings
model in the limit as $N \to \infty$, namely, we proved that for some \textit{level scaling function}
$k\colon\,\mathbb{N}_0\to\mathbb{N}_0$, satisfying $0 \leq k(j) \leq j+1$ and $\lim_{j\to\infty} k(j)=\infty$, we obtained
a \textit{non-trivial clustering limiting law} (henceforth we pick $\eta=0$ and drop it from the notation)
\be{ak:interaction-chain-scaling-limit-homogeneous}
\lim_{j \to \infty} \mathcal{L} \left[ M^{(j)}_{-k(j)} \right]
= \mathcal{L} \big[M^\infty\big]
\end{equation}
for some $M^\infty\in\mathcal{P}(E)$ satisfying $E[M^\infty] = E[X_0^{(\Omega_N)}(0)] = \theta
\in \mathcal{P}(E)$ that is not of the form $M^\infty=\delta_{U}$ for some possibly random $U \in E$. We
will do the same in the random environment $\omega$, namely, our aim is to show that for
$\mathbb{P}$-a.e.\ $\omega$
\be{ak:interaction-chain-scaling-limit}
\lim_{j \to \infty} \mathcal{L} \left[ M^{(j)}_{-k(j)}(\omega) \right]
= \mathcal{L} \big[M^\infty(\omega) \big]
\end{equation}
for some $M^\infty(\omega) \in \mathcal{P}(E)$ satisfying $\mathbb{E}[M^\infty(\omega)] = \theta$ that is not of the
form $M^\infty(\omega) = \delta_{U(\omega)}$ for some possibly random $U(\omega) \in E$.
As in Dawson and Greven~\cite{DG93b,DGV95,DG96}, and similarly as in \eqref{allim}, in order to
obtain the \emph{profile of cluster formation} it is necessary to consider a whole \emph{family of scalings}
$k_\alpha\colon\,\mathbb{N}_0 \to \mathbb{N}_0$, $\alpha \in I$, with $I=\mathbb{N}_0$, $I=[0,\infty)$ or $I=[0,1)$, and with
$j \mapsto k_\alpha(j)$ non-decreasing, $0 \leq k_\alpha(j) \leq j+1$ and $\lim_{j\to\infty} k_\alpha(j)
= \infty$, such that
\be{eq:ak:scaling-interaction-chain}
\lim_{j\to\infty} \mathcal{L} \left[M^j_{-k_\alpha(j)}(\omega)\right]
= \mathcal{L} \left[M^*_{\alpha}(\omega)\right]
\quad \text{ for } \mathbb{P} \text{-a.e. } \omega \text{ and all } \alpha \in I,
\end{equation}
for some non-constant Markov process $M^*=(M^*_\alpha(\omega))_{\alpha \in I}$ on $\mathcal{P}(E)$
that preserves the mean $\theta$. The convergence in \eqref{eq:ak:scaling-interaction-chain} is in the weak
topology on the product space of $\mathcal{P}(E)$ and the space of the environment.
There are five {\em universality classes} of clustering behaviour (see \cite{DG96}):
\begin{definition}[{\bf Clustering classes}]
\label{def:clusuniv}
$\mbox{}$
\begin{itemize}
\item[\textup{(I)}]
{\bf Concentrated clustering} ($M^\ast$ is a Markov chain):
\begin{itemize}
\item[\textup{(I1)}]
$k_\alpha(j) = 0 \vee (j+1 - \alpha)$, $\alpha \in \mathbb{N}_0$, $M^*$ is trapped after one step.
\item[\textup{(I2)}]
$k_\alpha(j) = 0 \vee (j+1 - \alpha)$, $\alpha \in \mathbb{N}_0$, $M^*$ is not trapped.
\end{itemize}
\item[\textup{(II)}]
{\bf Diffusive clustering} ($M^\ast$ is a diffusion process):
\begin{itemize}
\item[\textup{(II1)}]
\emph{Fast clustering}:
$k_\alpha(j) = 0 \vee \lfloor j+1 - \alpha h(j) \rfloor $, $\alpha \in [0,\infty)$, where $h\colon\,\mathbb{N}_0
\to [0,\infty)$ is such that $\lim_{j\to\infty} h(j) = \infty$ and $\lim_{j\to\infty} h(j)/j = 0$.
\item[\textup{(II2)}]
\emph{Moderate clustering}:
$k_\alpha(j) = \lfloor (1-\alpha)(j+1)\rfloor$, $\alpha \in [0,1)$.
\item[\textup{(II3)}]
\emph{Slow clustering}:
$\lim_{j\to\infty} k_\alpha(j)/j=0$, $\alpha \in [0,1)$.
\end{itemize}
\end{itemize}
\end{definition}
\noindent
(The terminology is slightly different from \cite{GHKK14}.) The volume of the clusters at time $t$ in
these five universality classes (arranged in decreasing order of magnitude) equals, respectively,
$N^t$, $ZN^t$, $N^{t-o(t)}$, $N^{Zt}$, $N^{o(t)}$, with $Z \in (0,1)$ some random variable. Note
that slow clustering borders with the regime of coexistence (= no clustering).
Recall (a)-(d) in Theorem~\ref{T.scaleFVpol} and (A)-(C) in Theorem~\ref{T.scaleFVexp}. Recall
that, under the law $\mathbb{P}$, the law of the initial state $X^{(\Omega_N)} (\omega;0)$ is stationary
and ergodic under translations in $\Omega_N^{\mathbb{T}}$, with mean single-coordinate measure
\begin{equation}\label{e1548}
\theta = \mathbb{E}[X_0^{(\Omega_N)}(\omega;0)] \in \mathcal{P}(E).
\end{equation}
The interaction chain on level $j$, arising in the scaling limit $N\to\infty$, starts in $\theta$.
Below this is also assumed for the scaling limit $j\to\infty$.
\begin{theorem}[{\bf Cluster formation}]
\label{T.cluform}
Fix $N \in \mathbb{N} \backslash \{1\}$. The five universality classes in the regime of clustering, linked to the different
cases of scaling behaviour of $ (d_k)_{k \in \mathbb{N}}$, are as follows:
\begin{itemize}
\item[$\bullet$]
{\rm \textbf{(a), (A):}}
The scaling in regime \textup{(I1)} yields \eqref{eq:ak:scaling-interaction-chain} with $I=\mathbb{N}_0$.
The scaling limit $M^*$ is the time-homogeneous Markov chain on $\mathcal{P}(E)$ starting in
$\theta$ with transition kernel $K(\theta, \cdot)$ given by
\be{regi1}
K(\theta, \cdot) = \int_E \theta(\mathrm{d} u) \delta_{\delta_u}(\cdot),
\end{equation}
which satisfies $K_\alpha=K$, for all $\alpha \in \mathbb{N}$.
\item[$\bullet$]
{\rm \textbf{(b), (B), (C1), (C3)[first subcase]:}}
The scaling in regime \textup{(I2)} yields \eqref{eq:ak:scaling-interaction-chain} with $I=\mathbb{N}_0$.
The scaling limit $M^*$ is the time-inhomogeneous Markov chain on $\mathcal{P}(E)$ in random
environment $(\chi_\alpha)_{\alpha \in \mathbb{N}_0}$ starting in $\theta$ with transition kernels
$\{K_\alpha(\theta,\cdot)(\omega)\}_{\alpha \in \mathbb{N}_0}$ given by (recall \eqref{eq:equi-ident-c})
\be{regi2}
K_\alpha(\theta, \cdot)(\omega)
= \nu^{1, \widetilde{M}, 2 \widetilde{K} \chi_\alpha(\omega)}_\theta (\cdot)
\end{equation}
with
\begin{equation}
\label{wtMwtK}
(\widetilde{M}, \widetilde{K}) =
\begin{cases}
(M, K), & \textup{(b)},\\
(\bar{M}/c, \bar{K}), & \textup{(B)},\\
((1-c)/c, 0), & \textup{(C1), (C3)[first subcase]}.
\end{cases}
\end{equation}
In the last two cases, the random environment does not affect the scaling limit, and the scaling is
the same as for the homogeneous environment with the same mean. In the first two cases the
measure-valued process $(\chi_\alpha(\omega))_{\alpha\in\mathbb{N}_0}$ in \eqref{regi2} is constructed
by extending the one-sided stationary random environment $(\chi^{(\eta,k)} (\omega))_{\eta\in\Omega_N,
\,k\in\mathbb{N}_0}$ introduced in \eqref{chiprop} to a two-sided stationary random environment $(\chi^{(\eta,k)}
(\omega))_{\eta\in\Omega_N,\,k\in\mathbb{Z}}$, and defining
\be{regi3}
\mathcal{L}[(\chi_\alpha(\omega))_{\alpha \in \mathbb{N}_0}]
= \mathcal{L}[(\chi^{(\eta,-\alpha)}(\omega))_{\alpha \in \mathbb{N}_0}],
\quad \eta \in \Omega_N,
\end{equation}
which by stationarity does not depend on $\eta$. Furthermore, $\chi_\alpha(\omega)$ is an
$\mathcal{M}_f([0,1])$-valued resampling measure with $\mathbb{E}[\chi_\alpha(\omega)] = \bar \chi_\alpha$
satisfying $\bar \chi_\alpha((0,1])=1$.
\item[$\bullet$]
{\rm \textbf{(c), (C2)[subcase $\lim_{k\to\infty} k\bar{K}_k=\infty$]:}}
The scaling in regime \textup{(II1)} yields \eqref{eq:ak:scaling-interaction-chain} with $I=[0,\infty)$.
The scaling limit $M^*$ is the time-changed standard Fleming-Viot process
\be{clusteringC}
M^*_{\alpha} = Z^{0,1,0}_{\theta} \left( \ell(\alpha) \right), \quad \alpha \in [0,\infty)
\end{equation}
with
\begin{itemize}
\item[--]
{\rm \textbf{(c):}}
$\ell(\alpha) = \alpha$, $h(j) = 1/\sqrt{K_j}$.
\item[--]
{\rm \textbf{(C2)[$\lim_{k\to\infty} k\bar{K}_k = \infty$]:}}
$\ell(\alpha) = \frac{\mu}{\mu-1}\alpha$, $h(j) = 1/K_j$.
\end{itemize}
\item[$\bullet$]
{\rm \textbf{(d), (C2)[subcase $\lim_{k\to\infty} k\bar{K}_k=\bar{N}$], (C3)[second subcase]:}}
The scaling in regime \textup{(II2)} yields \eqref{eq:ak:scaling-interaction-chain} with $I = [0,1)$.
The scaling limit $M^*$ is the time-changed standard Fleming-Viot process
\footnote{\cite{GHKK14} contains a typo: there the time scaling $1/(1-\alpha)^R$ was wrongly
written as $1/(1-\alpha^R)$.}
\be{clusteringD}
M^\ast_{\alpha} = Z^{0,1,0}_{\theta} \left(\log\left(\frac{1}{(1-\alpha)^R}\right)\right),
\quad \alpha \in [0,1)
\end{equation}
with
\begin{itemize}
\item[--]
{\rm \textbf{(d):}}
$R=M(1-a)$.
\item[--]
{\rm \textbf{(C2)[subcase $\lim_{k\to\infty} k\bar{K}_k = \bar{N}$]:}}
$R=\bar{N}\frac{\mu}{\mu-1}$.
\item[--]
{\rm \textbf{(C3)[second subcase]:}}
$R=1-a$.
\end{itemize}
\end{itemize}
For reasons explained in Remark~{\rm \ref{rem:homclus}}, in cases {\rm \textbf{(c), (d), (C2),
(C3)[second subcase]}} only convergence in $\mathbb{P}$-probability and not
$\mathbb{P}$-a.s.\ is obtained.
\hfill $\square$
\end{theorem}
\begin{remark}\label{r.1602}
{\rm We expect that also {\em slow} clustering occurs, namely, in the case (e) that was not treated
in Theorems~\ref{T.scaleFVpol} and \ref{T.scaleFVexp} (recall Remark~\ref{r:notexhaus}).}
\hfill $\square$
\end{remark}
\subsection{Summary of the effects of the random environment}
\label{ss.effectre}
\paragraph{1.}
Theorem~\ref{T.wpbasic} says that the hierarchical Cannings process in random environment is
well-defined for $\mathbb{P}$-a.e.\ $\omega$, while Theorem~\ref{T.ltbeq} shows that it converges to
an $\omega$-dependent equilibrium that preserves the single-component mean.
\paragraph{2.}
Theorem~\ref{mainth} (mono-scale) and Theorem~\ref{T.InteractionChain} (multi-scale) identify
the behaviour of the $k$-block averages in the limit as $N\to\infty$ in terms of the McKean-Vlasov
process with parameters that depend on $\omega$ and $k$. The volatility $d_k$ depends on the
parameters $c_l,\lambda_l$, $0 \leq l < k$, and on the law $\mathcal{L}_\rho$ via the recursion relation
in \eqref{diffusion-constants}, which is a \emph{randomized} version of the recursion relation in
\cite{GHKK14}.
\paragraph{3.}
Theorems~\ref{T.coexcritNfin} and \ref{T.dicho} show that the dichotomy ``coexistence versus
clustering'' is \emph{not} affected by the random environment: the \emph{same} conditions apply to
the homogeneous hierarchical Cannings process studied in \cite{GHKK14}. Apparently, for the nature
of the equilibrium only the large-scale properties of the random environment matter. Since the
resampling measures are stationary under translations with total masses whose sigma-algebra at
infinity is trivial, only the average medium behaviour is relevant. The proof of the dichotomy in
Theorem~\ref{T.coexcritNfin} requires assumption \eqref{Apropalt} rather than assumption
\eqref{Aprop}. We believe this strengthening to be redundant, but a proof would require considerable
extra work.
\paragraph{4.}
Theorem~\ref{T.order} shows that \emph{the effect of the random environment is to lower the
volatility parameter $ d_k $ on every hierarchical scale $k$} compared to the average environment. The intuition behind
this is that the random environment causes fluctuations in the resampling, which in turn reduce the
clustering. The sandwich between the volatilities for the zero environment and the average environment
is useful to control the scaling.
\paragraph{5.}
Theorem~\ref{T.scaleFVpol} (polynomial coefficients) and Theorem~\ref{T.scaleFVexp}
(exponential coefficients) show that for Cases (b) and (B), where \emph{migration and resampling
occur at comparable rates}, the phenomenon of lower volatility $ d_k $ in random environment persists in
the limit as $k\to\infty$: even though the scaling of $d_k$ as $k\to\infty$ is the same as for the
average environment, it has a \emph{different prefactor} (e.g.\ $M$ solving \eqref{MK} is strictly
smaller than $M^*$ solving \eqref{MK} with $\mathcal{L}_\rho$ replaced by $\delta_1$, as is easily shown
by applying Jensen's inequality). For all other cases both the scaling and the prefactor are the same
as for the average environment.
\paragraph{6.}
Theorem~\ref{T.cluform} shows that for Cases (b), (B), (C1), (C3)[first subcase] the scaling of the
clusters in the random environment changes compared to that in the average environment: the
random environment is visible even in the scaling limit. \emph{The effect of the random environment
is to slow down the growth of the clusters}, i.e., to enhance the diversity of types. For all other cases
the scaling of the clusters is the same as for the average environment.
\section{Existence, uniqueness, duality and equilibrium}
\label{s.dual}
In this section, we prove Theorems~\ref{T.wpbasic}--\ref{T.ltbeq}. In Section~\ref{ss.spatialrand}
we construct the dual process with the help of a graphical representation based on Poisson random
measures. In Section~\ref{ss.dualities} we exhibit the duality. In Section~\ref{ss.dualproofs} we
establish the existence and uniqueness of the dual process and show the existence of its equilibrium.
In Section~\ref{ss.dualproofsthms} we use these results to prove Theorems~\ref{T.wpbasic}--\ref{T.ltbeq}.
Theorems~\ref{T.existence-coalescent}--\ref{T.duallongrun} below do not need a separate proof: this
is verbatim the same as the proof for the homogeneous environment given in \cite{GHKK14}.
\subsection{The spatial coalescent in random environment}
\label{ss.spatialrand}
In this section we introduce a hierarchical coalescent process in random environment that will serve
as a dual to the hierarchical Cannings process in random environment.
The coalescent is a Markov process taking values in the set of partitions of $\mathbb{N}$ labelled by the
points of a geographical space. We recall the basic objects and notations, and refer to
\cite[Section~2]{GHKK14} for details.
Let $G$ be a discrete geographical space. Our target geographical space is $G = \Omega_N$.
This will be approximated by a sequence of geographical spaces
\be{Gchoices}
G_{N,K}=\{0,\ldots,N-1\}^K, \quad K\in\mathbb{N},
\end{equation}
which are to be thought of as a sequence of blocks filling $\Omega_N$.
We will also need to consider the mean-field geographical space
\be{Gchoices*}
G = \{0,*\},
\end{equation}
where $\{\ast\}$ is a cemetery location.
The state space of the spatial coalescent is the set of \emph{$G$-labelled partitions} defined as
\be{labelled-partitions}
\Pi_{G,n} = \Big\{ \pi_{G} = \{ (\pi_1, g_1), (\pi_2, g_2), \ldots, (\pi_b, g_b) \}
\colon\,\{ \pi_1, \ldots, \pi_b \} \in \Pi_n,\, g_1, \ldots, g_b \in G, b \in [n] \Big\},
\end{equation}
where $n\in\mathbb{N}$ and
\be{ag19}
\Pi_n = \text{ set of partitions } \pi = \{ \pi_i \subset [n] \}_{i=1}^b
\text{ of $[n]$ into disjoint families $\pi_i$, $i\in[b]$.}
\end{equation}
We equip the set $\Pi_{G,n}$ with the {\em discrete topology}. Let $a$ be a random walk transition
kernel on $G$. When $G = \Omega_N$ we use the hierarchical random walk kernel $a = a^{(N)}$
in \eqref{32b}, when $G = G_{N,K}$ we use the same hierarchical random walk kernel but with
$c_k = 0$ for $k > K$, and when $G = \{ 0, \ast \} $ we use the random walk kernel with
$a(0,\ast) = c$, $a(\ast,0) = 0$.
Given the random environment $\omega$ (recall Section~\ref{sss.RE}), the spatial coalescent
in random environment is the Markov process on state space $\Pi_{G,n}$ with the following dynamics:
\begin{itemize}
\item
\textbf{[Migration]} Each partition block performs an independent random walk on $G$ with random
walk kernel $a^*$, where $a^*(g_1,g_2) = a(g_2,g_1)$, $g_1,g_2 \in G$, is the conjugate random
walk kernel.
\item
\textbf{[Local coalescence]} Independently at each location $g \in G$, the $l$-tuples of the partition
elements at $g$ coalesce into a single partition element at $g$ at rate
\be{eq:lambda-coalescence-rates}
\lambda_{b,l}(\omega) = \int_{(0,1]} r^l (1-r)^{b-l}
\frac{\Lambda^{[g]}(\omega)(\mathrm{d} r)}{r^2},
\end{equation}
where $b$ is the current total number of partition elements and $\Lambda^{[g]}(\omega)$ is the
resampling measure at $g$ in environment $\omega$.
\item
\textbf{[Non-local coalescence with reshuffling]} In the case $G = \Omega_N$, independently at each
location $g \in B_{|\xi|}(\xi)$, $\xi \in \Omega^{\mathbb{T}}_N$, the $l$-tuples of the partition elements
in $B_{|\xi|}(\xi)$ coalesce into a single partition element at $g$ at rate
\be{eq:lambda-coalescence-rates-l}
N^{-2k} \lambda^{(\xi)}_{b,l}(\omega),
\qquad
\lambda^{(\xi)}_{b,l}(\omega) = \int_{(0,1]} r^l (1-r)^{b-l}
\frac{\Lambda^\xi(\omega)(\mathrm{d} r)}{r^2}.
\end{equation}
Subsequently, all the partition elements located in $B_{|\xi|}(\xi)$ are uniformly \emph{reshuffled}, i.e.,
all the partition elements in $B_{|\xi|}(\xi)$ get a new location that is drawn uniformly from $B_{|\xi|}(\xi)$.
\end{itemize}
Note that in the case $G=\Omega_N$ the partition elements of the coalescent perform a hierarchical
random walk on $\Omega_N$ in the environment $\omega$ with migration coefficients given by
(recall \eqref{32b})
\be{cNomega}
c_k(\omega)(N,\eta) = c_k + N^{-1}\,\lambda^{\mathrm{MC}_{k+1}(\eta)}(\omega),
\qquad \eta \in \Omega_N, k \in \mathbb{N}_0,
\end{equation}
where $\mathrm{MC}_k(\eta)$ is the unique site at height $k$ above $\eta\in\Omega_N$ and
$\lambda^{\mathrm{MC}_{k+1}(\eta)}(\omega) = \Lambda^{\mathrm{MC}_{k+1}(\eta)} (\omega)$
$((0,1])$ (recall the notation introduced in Section~\ref{ss.random} and see Fig.~\ref{fig-hierartree}).
The extra term in the right-hand side of \eqref{cNomega} comes from the reshuffling that takes
place prior to the resampling.
The {\em coalescence} rate of two partition elements in $B_{|\xi|}(\xi)$ equals (recall \eqref{fdh:masschi})
\be{lambdaxi}
N^{-2|\xi|} \lambda^\xi(\omega), \qquad
\lambda^\xi(\omega)= \lambda_{|\xi|} \rho^\xi(\omega),
\qquad \xi\in\Omega_N^{\mathbb{T}}.
\end{equation}
We specify the spatial coalescent as a Markov process on $\Pi_G = \cup_{n\in\mathbb{N}} \Pi_{G,n}$ by
providing its generator. To that end, we need a space of test functions on $\Pi_G$. Namely, let
$\mathcal{C}_G$ be the algebra of bounded continuous functions $F\colon\,\Pi_G \to \mathbb{R}$ such that for all
$F \in \mathcal{C}_G$ there exists an $n=n(F) \in \mathbb{N}$ and a bounded function
\be{ghkk1}
F_n\colon\, \Pi_{G,n} \to \mathbb{R}
\end{equation}
with the property that $F(\cdot)=F_n(\cdot\vert_n)$.
Consider the linear operator $L^{(G)*}(\omega) \colon \mathcal{C}_G \to \mathcal{C}_G$
defined as
\be{spatial-coalescent-generator}
L^{(G)*}(\omega) = L^{(G)*}_{\mathrm{mig}} + L^{(G)*}_{\mathrm{coal}}(\omega),
\end{equation}
where the operators $L^{(G)*}_{\mathrm{mig}},L^{(G)*}_{\mathrm{coal}}(\omega)\colon\,
\mathcal{C}_G \to \mathcal{C}_G$ are defined for $\pi_G \in \Pi_G$ and $F \in \mathcal{C}_G$ as
\be{no3}
(L^{(G)*}_{\mathrm{mig}} F) (\pi_G)
= \sum_{i=1}^{b(\pi_{G}|_n)} \sum_{g,f \in G} a^*(g,f)
\big[F_n\big(\textrm{mig}_{g \to f,i}(\pi_{G,n})\big)-F(\pi_G)\big]
\end{equation}
and
\be{add2_a}
\begin{aligned}
&(L^{(G)*}_{\mathrm{coal}}(\omega) F) (\pi_G)
= \sum_{k\in\mathbb{N}_0} \, \sum_{\xi \in \Omega^{(k)}_N} \,\,
\sum_{\substack{J\subset\{i \in [n]\colon\,g_i = g\},\\|J| \geq 2}}\\
&\qquad \left(N^{-2k} \lambda^{(\xi)}_{b(\pi_{G,n},g),|J|}(\omega)
\left[F_n\left(\textrm{resh}_{B_{|\xi|}(\xi),U_{B_{|\xi|}(\xi)}}
\circ \textrm{coal}_{J,g}(\pi_{G,n})\right)-F(\pi_G)\right]\right).
\end{aligned}
\end{equation}
Here, the \textit{migration map} $\textrm{mig}_{g \to f,i}(\pi_{G}\vert_n)$ changes the spatial coordinate
of the $i$-th partition block from $g$ to $f$ (if such a partition element exists), the \textit{coalescence
map} $\textrm{coal}_{J,g}(\pi_{G,n})$ coalesces the partition blocks with indices in $J$ and location
$g$ (if any) into one block, while the \textit{reshuffling map} $\textrm{resh}_{B_{|\xi|}(\xi),U_{B_{|\xi|}(\xi)}}$
independently relocates each partition element located in $B_{|\xi|}(\xi)$ to a new location in
$B_{|\xi|}(\xi)$ that is randomly chosen.
\begin{theorem}[{\bf Existence and uniqueness}]
\label{T.existence-coalescent}
For every $\pi \in \Pi_G$ the $(L^{(G)*}(\omega), \mathcal{C}_G, \delta_\pi)$-mar\-tingale problem is
well-posed. \hfill $\square$
\end{theorem}
We denote the solution of the $(L^{(G)*}(\omega), \mathcal{C}^{(G)},\delta_\pi)$-martingale problem
by
\be{dualdef}
\mathfrak{C}^{(G)}(\omega) = \big( \mathfrak{C}^{(G)}(\omega;t) \big)_{t \geq 0},
\qquad \mathfrak{C}^{(G)}(\omega;0) = \delta_\pi.
\end{equation}
For every $n\in\mathbb{N}$, when restricted to $\Pi_{G,n}$, $\mathfrak{C}^{(G)}(\omega)$ becomes a
strong Markov process $\mathfrak{C}_n^{(G)}(\omega)$ with the Feller property.
\subsection{Dualities}
\label{ss.dualities}
Consider the map
\be{qq:duality}
\begin{aligned}
H^{(n)}_\varphi(x , \pi_{G,n})
& = \int_{E^{b}}
\left(\bigotimes_{i=1}^{b} x_{\eta_{\pi^{-1}(i)}}\big(\mathrm{d} u_i\big)\right)
\varphi\big(u_{\pi(1)},u_{\pi(2)},\ldots,u_{\pi(n)}\big),
\end{aligned}
\end{equation}
where $n\in\mathbb{N}$, $\phi \in C_{\rm b}(\mathcal{P}(E)^n)$, $x = (x_{\eta})_{\eta \in G}\in\mathcal{P}(E)^{G}$, $\pi_{G,n}
\in \Pi_{G,n}$, $b = b(\pi_{G,n}) = |\pi_{G,n}|$.
\begin{theorem}[{\bf Duality}]
\label{T.dual}
Fix $N\in\mathbb{N} \backslash \{1\}$. For each of the choices $G$ in \eqref{Gchoices} and \eqref{Gchoices*},
\be{e1855}
E[H^{(n)}_\varphi(X^G(\omega,t), \Pi_{G,n})]= E [H^{(n)}_\varphi(X^G,C^G(\omega,t))]
\end{equation}
for all $n\in\mathbb{N}$ and $\phi \in C_{\rm b}(\mathcal{P}(E)^n)$, where the same $\omega$ is used on both sides. \hfill $\square$
\end{theorem}
\noindent
This theorem is a consequence of the generator relation
\be{generator-duality}
\left(L^{(G)}(\omega) H^{(n)}_\varphi(\cdot,\pi_{G,n}) \right)(x)
= \left( L^{(G)*}(\omega) H^{(n)}_\varphi(x,\cdot) \right)(\pi_G)
\quad \text{for $\mathbb{P}$-a.e.~$\omega$}.
\end{equation}
This relation has been verified for the homogeneous model in \cite{GHKK14}, but here works
the same.
\subsection{Well-posedness of the martingale problems and equilibria}
\label{ss.dualproofs}
Theorem~\ref{T.wpbasic} can be formulated for geographic spaces that are countable Abelian
groups, in particular, the hierarchical group and the Euclidean lattice. For us the following
generalization suffices.
\begin{theorem}[{\bf Well-posedness}]
\label{T.dualwellpos}
For each of the choices $G$ in \eqref{Gchoices} and \eqref{Gchoices*} the following holds:
For $\mathbb{P}$-a.e.\ $\omega$ and every $\pi \in \Pi_G$, the $(L^{(G)}(\omega), \mathcal{C}^{(G)},
\delta_{\pi})$-martingale problem is well-posed. \hfill $\square$
\end{theorem}
\begin{theorem}[{\bf Equilibrium}]
\label{T.duallongrun}
Fix $N\in\mathbb{N} \backslash \{1\}$. Fix $n\in\mathbb{N}$ and start the $\mathfrak{C}^{(\Omega_N)}(\omega)$-process in
a labelled partition $\{(\pi_i,\eta_i)\}_{i=1}^{n}$, where $\{\pi_i\}_{i=1}^n$ form a partition of $\mathbb{N}$
and $\{\eta_i\}_{i=1}^n$ represent the labels. If $x$ is a random state with mean $\theta \in \mathcal{P}(E)$
whose law is invariant and ergodic under translations, then
\be{ag40}
\lim_{t\to\infty} \mathcal{L} \left[H^{(n)}_\varphi\left(x, \mathfrak{C}^{(\Omega_N)}_n(\omega; t)\right)\right]
= \mathcal{L}\left[H^{(n)}_\varphi\left(\underline{\theta},\mathfrak{C}^{(\Omega_N)}_n(\infty)\right)\right]
\quad \text{for $\mathbb{P}$-a.e.~$\omega$}
\end{equation}
for all $\phi \in C_{\rm b}(\mathcal{P}(E)^n)$. \hfill $\square$
\end{theorem}
In order to prove Theorem~\ref{T.duallongrun}, we follow the argument in \cite[Section 3]{DGV95}).
The partition-valued process converges to a limiting partition. If the locations of the partition elements
would follow a homogeneous random walk, then the key to the argument would be the averaging property one can prove via Fouries analysis
\be{ag14}
\lim_{t\to\infty} \sum_{\zeta \in \Omega_N} p_t(\eta,\zeta) f(\zeta)
= \int_{\Omega_N} f(\xi)\nu(\mathrm{d}\xi) \qquad \forall\,\eta \in \Omega_N,\,f \in C_b(\Omega_N),
\end{equation}
where $p_t(\cdot,\cdot)$ is the time-$t$ transition kernel of the random walk on $\Omega_N$ and
$\nu(\cdot)$ is the Haar measure on $\Omega_N$ (see Evans and Fleischmann~\cite{EF96}).
We need to show that the same holds for our random walk in random environment $\omega$, which
goes as follows.
Place a Poisson clock at every $\xi\in\Omega_N^{\mathbb{T}}\backslash\Omega_N$.
Let the clock at $\xi$ ring at rate
\be{PC}
N^{-(|\xi|-1)} \big[c_{|\xi|-1} + N^{-1}\lambda^\xi(\omega)\big].
\end{equation}
At any moment of time let the random walk look at the ancestral line above its current position
(see Fig.~\ref{fig-hierartree}) and redistribute itself uniformly over the block around its current
position whose hierarchical label corresponds to the height of the first clock on that ancestral line
that rings. The resulting random walk is the same as the hierarchical random walk in environment
$\omega$ with migration coefficients given by \eqref{cNomega}.
Next, let $K_t(\eta)$ be the highest hierarchical level at which prior to time $t$ a Poisson clock
that lies on the ancestral line above $\eta$ has rang. Then at time $t$ the random walk starting
from $\eta$ is uniformly distributed on the $K_t(\eta)$-block around $\eta$. Hence we have
\be{PC1}
\sum_{\zeta \in \Omega_N} p_t^\omega(\eta,\zeta) f(\zeta)
= \sum_{k\in\mathbb{N}_0} P(K_t(\eta)=k) \left[ N^{-k} \sum_{\zeta \in B_k(\eta)} f(\zeta)
+ \sum_{\zeta \in \Omega_N \backslash B_k(\eta)} p_t^\omega(\eta,\zeta) f(\zeta)\right],
\end{equation}
where $p_t^\omega(\cdot,\cdot)$ is the time-$t$ transition kernel of the random walk in $\omega$.
Fix $\eta \in \Omega_N$ and $f \in C_b(\Omega_N)$. The first term between the square brackets
in \eqref{PC1} tends to $\int_{\Omega_N} f(\xi)\nu(\mathrm{d}\xi)$ as $k \to \infty$. The second term is
bounded from above by $\|f\|_\infty\,p_t^\omega (\eta,\Omega_N \backslash B_k(\eta))$, which
tends to zero as $k\to\infty$. Finally, since all Poisson clocks ring at a strictly positive rate, we
have
\be{PC2}
P\left(\lim_{t\to\infty} K_t(\eta) = \infty\right) =1.
\end{equation}
It therefore follows that the right-hand side of \eqref{PC1} tends to the right-hand side of
\eqref{ag14} as $t\to\infty$.
\subsection{Consequences for the Cannings process}
\label{ss.dualproofsthms}
The claims in Theorems~\ref{T.wpbasic}--\ref{T.ltbeq} follow from
Theorems~\ref{T.dualwellpos}--\ref{T.duallongrun}. As argued in \cite{GHKK14}, the proof
follows the strategy for the two-type case given in Evans~\cite[Theorem 4.1]{E97}, which says
that for spatial coalescent processes well-posedness and existence carry over from the dual
process to the original process.
We next prove Corollary~\ref{c.1282}.
\begin{proof}
We analyze both $\nu_\theta^N$ and $\nu_\theta$ with the help of duality relations and show
that the dual representation of the former converges to the dual representation of the latter.
\medskip\noindent
{\bf Step 1: $\nu_\theta$.}
We have to construct a dual process for the entrance law of a Markov chain, namely, the interaction
chain running from level $\infty$ down to level $0$. We consider the process that is dual to the
interaction chain at level $j$. This dual process is a discrete-time Markov chain whose transition
kernel we can determine, for fixed $j$ and in the limit as $j\to\infty$, via an explicit construction.
This dual Markov chain is a spatial coalescent on $\{0,1,\ldots,j\}$, or on $\mathbb{N}$ when we consider
all $j$ simultaneously and are interested in its limit state as $ j\to\infty$.
We first focus on the dual transition kernel at one particular level. In the interaction chain this is
defined via the equilibrium of the McKean-Vlasov process. How did this equilibrium arise? We
consider a {\em mean-field} system of size $N^k$ with parameters $c_k$, $d_k$, $\Lambda_k$
and take the mean-field dual started in $n$ individuals at mutual distance $k$. This dual is shown
to {\em converge}, in the limit as $N\to\infty$ and on time scale $t_N N^k$ with $t_N \to \infty$
and $t_N = o(N)$, to a limiting process that is a coalescent on the geographic space $k\cup
\{\bigtriangleup\}$, with $\bigtriangleup$ a cemetery state, such that the process jumps from
$k$ to $\bigtriangleup$ at rate $c_k$ and does Kingman coalescence at rate $d_k$ and
$\Lambda$-coalescence according to $\Lambda_k(\mathrm{MC}_k(0))(\omega)$. This limiting process
is run for infinite time to obtain the dual transition kernel at the $k$-th step. This partition at
$\bigtriangleup$ is used as input for the next step of the dual with label $k+1$. Altogether this
procedure defines the full Markov chain, i.e., the new site and the new partition element. We
denote the path of the dual Markov chain by
\begin{equation}
\label{e1973}
(\Pi^\infty_k)_{k \in \mathbb{N}_0},\, \mbox{with $ \Pi^\infty_\infty $ its limiting state as } k \to \infty.
\end{equation}
(Recall that partitions are ordered and hence the limiting state exists.) The line of argument is the
same for each $k$. A detailed argument can be found in \cite[Corollary 2.12]{GHKK14}.
We need an explicit description as an $\mathbb{N}$-marked partition-valued process, namely, the above
mentioned random walk, moving one step to the right on $\mathbb{N}$, doing Kingman coalescence at
rate $d_k$ and $\Lambda$-coalescence according to $ \Lambda_k(\mathrm{MC}_k(\eta))$
in state $k$, provided the rate-$c_k$ clock does not ring first.
The dual chain after $j$ steps gives the expression $E_\theta[\langle M^{(j)}_0,f\rangle^n]
(\omega)$, which in the limit as $j\to\infty$ equals $\int^1_0 \Pi_\eta \nu_\theta (\omega)(\mathrm{d} x)\,
\langle x_\eta,f \rangle^n$ by the definition of $\nu_\theta(\omega)$. The dual expectation is the
expression $E[\langle \theta,f\rangle^{\mid \Pi_\infty^\infty \mid}](\omega)$. It therefore suffices
to show that the latter is obtained from the dual representation of $\nu^N_\theta(\omega)$ as
$N\to\infty$.
\begin{remark}
\label{r.1992}
{\rm What is the dual counterpart of Theorem~\ref{mainth}? The connection between the renormalized
system and the interaction chain on the level of the dual is as follows. Consider the dual process for
the $j$-level \emph{hierarchical} system for finite $N$, starting with $n$ partition elements at one site and
letting $t\to\infty$ and $N\to\infty$ in the following way. Consider time scales $(t^k_N)_{k\in\mathbb{N}_0}$
with $t^k_N/N^{k+1} \to 0$ and $ t^k_N/N^k\to\infty$ as $N\to\infty$. Then the coalescent reaches
a partition $\Pi^{k+1}_\infty$, with the remaining partition elements in $B_k$ uniformly distributed.
After that move to the next time scale. Finally, first take $N\to\infty$ and then take
\begin{equation}
\label{e1626}
\Pi^\infty_\infty \mbox{ as the limiting partition element for } j \to \infty.
\end{equation}
By our scaling result in \eqref{ag:interaction-chain}, this object gives us the dual process of the
interaction chain at level $j$.} \hfill $\square$
\end{remark}
\medskip\noindent
{\bf Step 2: $\nu^N_\theta$.}
We return to the representation of $\nu^N_\theta$, respectively, its marginal law at level $0$.
The convergence of the dual chain for the Cannings process on $\Omega_N$, and its limit as
$t\to\infty$ followed by $N\to\infty$ to the dual chain of the interaction chain, will follow from
the fact that the partitions become successively finer and hence converge to a limit partition,
and the fact that the time scales for the random walk to reach distance $k$ separate as $N\to\infty$.
Since the monomials are convergence determining, this will yield the claim.
We have to show that the coalescent on $\Omega_N$, starting with $n$ individuals at site
$0$, converges to a limit process as $t\to\infty$, which we can investigate in the limit as
$N\to\infty$. We need to show that this process has the property that, when we consider the
times where the coalescent makes jumps to the next larger block, we get an embedded Markov
chain with index in $\mathbb{N}_0$, describing the successive maximal jump sizes and values in partitions.
The corresponding partition converges to $ \Pi^N_\infty $ as the index $ k $ tends to infinity. The
claim is that, as $N\to\infty$, this Markov chain converges to a birth process in the first component,
which moves one step to the right, with Kingman-coalescence at rate $d_k$ and
$\Lambda$-coalescence according to $\Lambda_k(\mathrm{MC}_k(0))(\omega)$, provided the
rate-$c_k$ clock does not ring first. This gives $ \Pi^\infty_k $. As $ k \to \infty $ we get \eqref{e1626}.
This is done in Dawson and Greven~\cite{DG96} for the Kingman coalescent, but the necessary
modification is straightforward.
\end{proof}
\section{Dichotomy: coexistence versus clustering}
\label{s.randomwalk}
In this section, we prove Theorem~\ref{T.coexcritNfin}. The question is whether $\mathfrak{C}^{(G)}
(\omega)$, the hierarchical coalescent in the environment $\omega$ defined in \eqref{dualdef},
converges to a single labelled partition element as $t\to\infty$ with probability one. To answer this
question, we have to investigate \emph{whether two tagged partition elements coalesce with
probability one or not}. Recall that, by the projective property of the coalescent, we may focus
on the subsystem of just two dual individuals, because this translates into the same dichotomy
for $\mathfrak{C}_n^{(G)}(\omega)$ for any $n\in\mathbb{N}$, and hence for the entrance law starting
from $n$ partition elements. However, there is additional reshuffling at all higher levels, which is
triggered by a corresponding block-coalescence event. Therefore, we need to consider two coalescing
random walks with slightly {\em adapted} migration coefficients, lacking in particular the random
walk property.
Recall the notation introduced in Sections~\ref{sss.hg}--\ref{sss.resh} and \ref{sss.RE}.
Recall that $\mathrm{MC}_k(\eta)$ is the unique site at height $k$ above $\eta\in\Omega_N$
(see Fig.~\ref{fig-hierartree}). Consider two independent copies
\be{RWs}
Y(\omega) = (Y_t(\omega))_{t\geq 0}, \qquad
Y^\prime(\omega) = (Y^\prime_t(\omega))_{t\geq 0},
\end{equation}
of the hierarchical random walk on $\Omega_N$ in the environment $\omega$ with migration
coefficients given by \eqref{cNomega} and coalescence rates given by \eqref{lambdaxi}. Write
$P^\omega, P^{\omega,\prime}$ for the marginal laws of $Y(\omega),Y^\prime(\omega)$ and
$\bar{P}^\omega = P^\omega \times P^{\omega,\prime}$ for the joint law of the pair $\bar{Y}
(\omega)=(Y(\omega),Y^\prime(\omega))$. Consider the time-$t$ accumulated hazard for
coalescence:
\be{ak12}
H_N(\omega;t) = \sum_{k\in\mathbb{N}_0} N^{-k}
\sum_{\substack{\eta, \eta^\prime \in \Omega_N \\ d_{\Omega_N}(\eta,\eta^\prime) \leq k}}
\lambda^{\mathrm{MC}_k(\eta)}(\omega)
\int^t_0 1_{\{Y_s(\omega) = \eta,Y^\prime_s(\omega) = \eta^\prime\}}\, \mathrm{d} s,
\end{equation}
where we use that $\mathrm{MC}_k(\eta)=\mathrm{MC}_k(\eta^\prime)$ when $d_{\Omega_N}
(\eta,\eta^\prime) \leq k$. The rate $N^{-2k}$ to choose a $k$-block for the coalescence is
multiplied by $N^k$ because all partition elements in that block can trigger a coalescence event,
which explains the factor $N^{-k}$ in \eqref{ak12}.
Let $\lim_{t\to\infty} H_N(\omega;t) = H_N(\omega;\infty)$. We have coalescence of the two random
walks (``common ancestor'') with probability 1 when $H_N(\omega;\infty) = \infty$ $\bar{P}^\omega$-a.s.,
but separation of the two random walks (``different ancestors'') with positive probability when
$H_N(\omega;\infty) < \infty$ $\bar{P}^\omega$-a.s. In Section~\ref{ss.meanhazard} we identify
the dichotomy for the mean hazard $\bar{E}^\omega[H_N(\omega;\infty)]$ combining Fouries analysis
with potential theory of reversible Markov chains to handle the fact that our migration is no longer
a random walk. In Section~\ref{ss.zero-one} we use a zero-one law to show that the same dichotomy
holds for the hazard $H_N(\omega;\infty)$.
\subsection{Mean hazard}
\label{ss.meanhazard}
\bl{meanhazard}
For every $N \in \mathbb{N} \backslash \{1\}$ and $\mathbb{P}$-a.e.\ $\omega$,
\be{dich}
\bar{E}^\omega[H_N(\omega;\infty)] = \infty
\quad \Longleftrightarrow \quad
\sum_{k\in\mathbb{N}_0} \frac{1}{c_k + N^{-1}\lambda_{k+1}} \sum_{l=0}^k \lambda_l = \infty.
\end{equation}
\hfill $\square$
\end{lemma}
\begin{proof}\label{p.1958}
Write
\be{trker}
p^\omega_t(\eta,\zeta) = P^\omega\{Y_t(\omega) = \zeta \mid Y_0(\omega) = \eta\},
\quad \eta, \zeta \in \Omega_N,
\end{equation}
to denote the time-$t$ transition kernel. Let
\be{jointGreen}
G^\omega((0,0),(\eta,\eta^\prime)) = \int_0^{\infty} p^\omega_t(0,\eta)
p^\omega_t(0,\eta^\prime)\, \mathrm{d} t,
\qquad (\eta,\eta^\prime) \in \Omega_N \times \Omega_N,
\end{equation}
denote the Green function for $\bar{Y}(\omega)$. Then \eqref{ak12} gives
\be{eq:expected-hazard}
\begin{aligned}
\bar{E}^\omega[H_N(\omega;\infty)]
&= \sum_{k \in \mathbb{N}_0} N^{-k}
\sum_{\substack{\eta, \eta^\prime \in \Omega_N \\ d_{\Omega_N}(\eta,\eta^\prime) \leq k}}
\lambda^{\mathrm{MC}_k(\eta)}(\omega)\, G^\omega((0,0),(\eta,\eta^\prime))\\
&= \sum_{k \in \mathbb{N}_0} N^{-k} \sum_{\xi \in \Omega_N^{(k)}} \lambda^\xi(\omega)
\sum_{\substack{\eta \in \Omega_N \\ \mathrm{MC}_k(\eta)=\xi}}
\sum_{\eta' \in B_k(\eta)} G^\omega((0,0),(\eta,\eta^\prime))\\
&= \sum_{k \in \mathbb{N}_0} N^{-k} \sum_{\xi \in \Omega_N^{(k)}} \lambda^\xi(\omega)
\sum_{\eta,\eta^\prime \in B_{|\xi|}(\xi)} G^\omega((0,0),(\eta,\eta^\prime))
\end{aligned}
\end{equation}
(recall \eqref{Bkequivdef}). The proof comes in two steps. In Step 1, we pretend that the
$\omega$-dependent term in the right-hand side of \eqref{cNomega} is replaced by its
mean, i.e., the two hierarchical random walks are homogenous with migration coefficients
$\bar{c}_k$ given by
\be{barck}
\bar{c}_k(N) = \mathbb{E}[c_k(\omega)(N,\eta)] = c_k + N^{-1}\,\lambda_{k+1},
\end{equation}
and show that the same dichotomy as in \eqref{dich} holds. In Step 2, we explain why this
replacement does not affect the dichotomy. The Green function of the two homogeneous
hierarchical random walks will be denoted by $G((0,0),(\eta,\eta^\prime))$.
\paragraph{Step 1.}
In what follows, we use the explicit form of the transition kernel $p_t(\eta,\zeta)$, $\eta,\zeta
\in \Omega_N$, for the \emph{homogeneous} hierarchical random walk computed in Dawson,
Gorostiza and Wakolbinger~\cite{DGW05} with the help of Fourier analysis. Namely,
\be{ak:dgw-asympt}
p_t(0,\eta) = \sum_{j \geq k} K_{jk}(N)\,\frac{\exp[-h_j(N) t]}{N^j},
\qquad t \geq 0,\, \eta \in \Omega_N\colon\, d_{\Omega_N}(0,\eta)=k \in \mathbb{N}_0,
\end{equation}
where
\be{Kjkdef}
K_{jk}(N) = \left\{\begin{array}{ll}
0, &j=k=0,\\
-1, &j=k>0,\\
N-1, &\mbox{otherwise},
\end{array}
\right.
\qquad j,k \in \mathbb{N}_0,
\end{equation}
and
\be{hjrjrel}
h_j(N) = \frac{N-1}{N}\,r_j(N) + \sum_{i>j} r_i(N), \qquad j \in \mathbb{N},
\end{equation}
with
\be{ak:106}
r_j(N) = \frac{1}{D(N)} \frac{N-1}{N} \sum_{i \geq j} \frac{\bar{c}_{i-1}(N)}{N^{2i-j-1}},
\qquad j\in\mathbb{N},
\end{equation}
where $D(N)$ is the normalizing constant such that $\sum_{j\in\mathbb{N}} r_j(N)=1$. Note that the
expressions in \eqref{hjrjrel}--\eqref{ak:106} simplify considerably in the limit as $N\to\infty$,
namely, the term with $i=j$ dominates and
\be{Ninfsimp}
h_j(N) \sim r_j(N) \sim \frac{\bar{c}_{j-1}(N)}{D(N) N^{j-1}}, \quad j \in \mathbb{N},
\qquad D(N) \sim \bar{c}_0(N).
\end{equation}
Also note that, because of \eqref{ak:recurrence-cond} and \eqref{ak:lambda-growth-condition},
the following holds:
\begin{equation}
\begin{tabular}{ll}
&\text{For $N\in\mathbb{N} \backslash \{1\}$ the quantities $h_j(N),r_j(N),D(N)$ are bounded from}\\
&\text{above and below by positive finite constants times the right-hand side}\\
&\text{of \eqref{Ninfsimp} uniformly in the index $j$.}
\end{tabular} \label{fdh:Ncomp}
\end{equation}
To compute the sum in \eqref{eq:expected-hazard}, we need to distinguish two cases:
(1) $\xi = 0^k \in \Omega_N^{(k)}$, the unique site in $\Omega_N^{\mathbb{T}}$ at height
$k$ above $0 \in \Omega_N$; (2) $\xi \in \Omega_N^{(k)}\backslash\{0^k\}$.
\paragraph{(1)} \underline{$\xi = 0^k$}. Write
\begin{equation}
\label{eq:ak147}
\sum_{\eta,\eta^\prime \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta^\prime))
= \sum_{0 \leq p,q \leq k} N[p]N[q]\, G((0,0),(\eta^{(p)},\eta^{(q)})),
\end{equation}
where $\eta^{(p)}$ is any site in $\Omega_N$ such that $d_{\Omega_N}(0,\eta^{(p)}) = p$, and
\be{Npdef}
N[p] = |B_p(0)\backslash B_{p-1}(0)| = \left\{\begin{array}{ll}
N^p-N^{p-1}, &p>0,\\
1, &p=0.
\end{array}
\right.
\end{equation}
With the help of \eqref{ak:dgw-asympt} we obtain
\be{eq:ak*}
G((0,0),(\eta^{(p)},\eta^{(q)}))
= \sum_{m \geq p} \sum_{n \geq q} K_{mp}(N) K_{nq}(N)\, N^{-m-n} \frac{1}{h_m(N)+h_n(N)}.
\end{equation}
Inserting \eqref{Kjkdef} and \eqref{Ninfsimp}, we get
\begin{equation}
\mbox{r.h.s. } \eqref{eq:ak*} \sim \frac{\bar{c}_0(N)}{(1+1_{\{p=q\}}) \label{eq:ak**}
\bar{c}_{p \wedge q}(N) N^{p \vee q}}, \qquad N\to \infty,
\end{equation}
where the asymptotics comes from the terms with $m=p+1$ and $n=q+1$.
Combining (\ref{eq:ak147}--\ref{eq:ak**}), we obtain
\be{eq:ak145}
\begin{aligned}
&N^{-k} \sum_{\eta,\eta^\prime \in B_k(0^k)} G((0,0),(\eta,\eta^\prime))\\
&\qquad \sim N^{-k} \left(\sum_{0 \leq p \leq k} N^{2p} \frac{\bar{c}_0(N)}{2 \bar{c}_p(N) N^p}
+ 2 \sum_{0 \leq p < q \leq k} N^{p+q} \frac{\bar{c}_0(N)}{\bar{c}_p(N) N^{q}}\right)
\sim \frac{\bar{c}_0(N)}{2 \bar{c}_k(N)},
\end{aligned}
\end{equation}
where the asymptotics comes from the term with $p=k$.
\paragraph{(2)} \underline{$\xi \in \Omega_N^{(k)}\backslash\{0^k\}$}. Now $p_t(0,\eta)$ is the
same for all $\eta \in B_{|\xi|}(\xi)$, and so we have
\begin{equation}
\label{eq:ak101a}
\begin{aligned}
&N^{-k} \sum_{\xi \in \Omega_N^{(k)}\backslash\{0^k\}} \lambda^\xi(\omega)
\sum_{\eta,\eta^\prime \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta^\prime))\\
&\qquad = N^k \sum_{\xi \in \Omega_N^{(k)}\backslash\{0^k\}}
\lambda^\xi(\omega)\, G((0,0),(\eta^{(k)},\eta^{(k)}))\\
&\qquad \sim N^k \sum_{d\in\mathbb{N}}
\sum_{ \substack{\xi \in \Omega_N^{(k)} \\ d_{\Omega^{(k)}_N}(0^k,\xi) = d} }
\lambda^\xi(\omega)\, \frac{\bar{c}_0(N)}{2 N^{k+d} \bar{c}_{k+d}(N)}\\
&\qquad = \frac{1}{2} \sum_{d\in\mathbb{N}} \frac{\bar{c}_0(N)}{\bar{c}_{k+d}(N)}
\Bigg(\frac{1}{N^{d}}
\sum_{ \substack{\xi \in \Omega_N^{(k)} \\ d_{\Omega^{(k)}_N}(0^k,\xi) = d} }
\lambda^\xi(\omega)\Bigg),
\end{aligned}
\end{equation}
where we use (\ref{eq:ak*}--\ref{eq:ak**}) with $p=q=k$, and $d_{\Omega^{(k)}_N}$ denotes
the distance within $\Omega_N^{(k)}$.
\medskip
Combining \eqref{lambdaxi}, \eqref{eq:expected-hazard}, \eqref{eq:ak145}--\eqref{eq:ak101a},
we arrive at
\be{meanhazfinal}
\bar{E}[H_N(\omega;\infty)] \sim \tfrac12\,\bar{c}_0(N)
\sum_{k\in\mathbb{N}_0} \frac{1}{\bar{c}_k(N)} \sum_{l=0}^k
\lambda_l \Big\{\tfrac12\,1_{\{l=k\}}\,\Theta(\omega;0,k) + 1_{\{l<k\}}\,\Theta(\omega;k-l,l)\Big\},
\end{equation}
where we abbreviate
\be{Thetadef}
\Theta(\omega;a,b) = \frac{1}{N[a]}
\sum_{\substack{\xi\in\Omega^{(b)}_N \\ d_{\Omega_N^{(b)}}(0^b,\xi)=a}}
\rho^\xi(\omega), \qquad a,b \in \mathbb{N}_0.
\end{equation}
Now, by \eqref{Aprop} we have, for some $C<\infty$,
\be{Thetaprop}
\mathbb{E}\big[\Theta(\omega;a,b)\big] = 1, \quad \mathbb{E}\big[\Theta(\omega;a,b)\Theta(\omega;a',b')\big]
\leq C \quad \forall\,a,b,a',b'\in\mathbb{N}_0.
\end{equation}
Because $\{\rho^\xi(\omega)\colon\,\xi \in \Omega_N^{\mathbb{T}}\}$ is stationary, ergodic and
tail trivial (recall \eqref{fdh:tail}), it follows from a standard second-moment estimate that the
sum in the right-hand of \eqref{meanhazfinal} is infinite if and only if its expectation w.r.t.\ $\mathbb{P}$
is infinite. Since
\begin{equation}
\mathbb{E}(\mathrm{r.h.s.}\eqref{meanhazfinal}) = \sum_{k\in\mathbb{N}_0} \frac{1}{\bar{c}_k(N)} \sum_{l=0}^k
\lambda_l \big\{\tfrac12\,1_{\{l=k\}} + 1_{\{l<k\}}\big\},\label{expmeanhazfinal}
\end{equation}
we get the claim in \eqref{dich} for the hierarchical random walk with homogeneous migration
coefficients $\bar{c}_k(N)$ defined in \eqref{barck} (the factor $\tfrac12$ is harmless for the
convergence or divergence of the right-hand side of \eqref{expmeanhazfinal}).
\paragraph{Step 2.}
It remains to show that the same dichotomy holds for the coefficients in \eqref{cNomega} rather
than \eqref{barck}. We start with the observation that the hierarchical random walk in random
environment is \emph{symmetric} and therefore is {\em reversible} with respect to the
{\em Haar measure} on $\Omega_N$. We have the representation (see Bovier and den
Hollander~\cite[Chapter 7]{BdH15})
\be{fdh:jointGreenpottheo}
\begin{aligned}
G^\omega((0,0),(\eta,\eta^\prime))
&= \int_0^{\infty} p_t^\omega(0,\eta) p_t^\omega(0,\eta^\prime)\, \mathrm{d} t\\
&= \frac{P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \infty)}
{a^\omega((\eta,\eta^\prime))\,P^\omega_{(\eta,\eta\prime)}(\hat\tau_{(\eta,\eta\prime)} = \infty)},
\quad (\eta,\eta^\prime) \in \Omega_N \times \Omega_N,
\end{aligned}
\end{equation}
where $a^\omega((a,b)) = \sum_{(c,d)} a^\omega((a,b),(c,d))$ is the total rate at which the random
walk jumps out of $(a,b)$, and
\be{tauabdef}
\begin{aligned}
\tau_{(a,b)} &= \inf\big\{t \geq 0\colon\,Y_t(\omega) = (a,b)\big\},\\
\hat\tau_{(a,b)} &= \inf\big\{t \geq 0\colon\,Y_t(\omega) = (a,b),\,
\exists\,0<s<t\colon\,Y_s(\omega) \neq (a,b) \big\},
\end{aligned}
\end{equation}
are the first hitting time, respectively, the first return time of $(a,b)$. The point of
\eqref{fdh:jointGreenpottheo} is that both the numerator and the denominator can be controlled with
the help of the \emph{Dirichlet Principle}, as follows.
Let
\be{Diridef}
\mathcal{E}^\omega(f,f) = \sum_{(a,b),(c,d)} \frac{a^\omega((a,b),(c,d))}{a^\omega((a,b))}
\,[f((a,b))-f((c,d))]^2
\end{equation}
be the \emph{Dirichlet form} associated with the two random walks in random environment.
By classical potential theory, the escape probability in the denominator of \eqref{fdh:jointGreenpottheo}
is given by the capacity of the pair $(\eta,\eta^\prime)$ and $\infty$,
\be{pot2}
P^\omega_{(\eta,\eta\prime)}(\hat\tau_{(\eta,\eta\prime)} = \infty)
= \mathrm{cap}^\omega((\eta,\eta\prime), \infty)
= \inf_{ {f\colon\,\Omega_N \to [0,1]} \atop {f((\eta,\eta^\prime)) = 1, f(\infty) = 0} } \mathcal{E}^\omega(f,f),
\end{equation}
where $f(\infty)=0$ stands for $\lim_{(\eta,\eta^\prime) \to \infty} f((\eta,\eta^\prime)) = 0$ with
$(\eta,\eta^\prime) \to \infty$ short hand for $d_{\Omega_N}(0,\eta)+d_{\Omega_N}(0,\eta^\prime)
\to \infty$ (recall \eqref{ag31}). The hitting probability in the numerator of \eqref{fdh:jointGreenpottheo}
can also be expressed in terms of capacities after we use a renewal argument. Write
\be{pot3}
\begin{aligned}
P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \infty)
&= \frac{P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \hat\tau_{(0,0)})}
{1-P^\omega_{(0,0)}(\hat\tau_{(0,0)} < \tau_{(\eta,\eta^\prime)})}\\
&= \frac{P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \hat\tau_{(0,0)})}
{P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \hat\tau_{(0,0)})
+ P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} = \hat\tau_{(0,0)} = \infty)}.
\end{aligned}
\end{equation}
We have
\be{pot4}
P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \hat\tau_{(0,0)}) = \mathrm{cap}^\omega((0,0),(\eta,\eta^\prime))
= \inf_{ {f\colon\,\Omega_N \to [0,1]} \atop {f((\eta,\eta^\prime)) = 1, f((0,0)) = 0} } \mathcal{E}^\omega(f,f).
\end{equation}
Moreover,
\be{pot5}
P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} = \hat\tau_{(0,0)} = \infty)
= P^\omega_{(0,0)}(\hat\tau_{(0,0)} = \infty)
- P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} <\infty, \hat\tau_{(0,0)} = \infty).
\end{equation}
The first term equals $\mathrm{cap}^\omega((0,0), \infty)$, while the second term is bounded from above by
$P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} <\infty)$, which tends to zero as $(\eta,\eta^\prime) \to
\infty$ when $G^\omega<\infty$, i.e., when the random walk in random environment is transient.
Below we will show that, under assumption~\eqref{Apropalt}, {\em $G^\omega<\infty$ if and only if $G<\infty$}.
We are now ready to explain why the estimates in Step 1 carry over.
The transition rates of the random walk in random environment are given by
\be{pot6}
a^\omega((a,b)(c,d)) = \left\{\begin{array}{ll}
a^{\omega,(N)}(a,c), &b=d,\\
a^{\omega,(N)}(b,d), &a=c,\\
0, &\text{else.}
\end{array}
\right.
\end{equation}
where $a^{\omega,(N)}$ is the transition kernel in \eqref{32b}, but with $c_k$ replaced by
$c_k(\omega)(N,\eta)$ in \eqref{cNomega}:
\be{pot7}
a^{\omega,(N)}(\eta,\zeta) = \sum_{k \geq d_{\Omega_N}(\eta,\zeta)}
\frac{c_{k-1} + N^{-1} \lambda^{\mathrm{MC}_k(\eta)}(\omega)}{N^{2k-1}},
\qquad \eta,\zeta \in \Omega_N,\,\eta\neq\zeta, \qquad a^{\omega,(N)}(\eta,\eta) = 0.
\end{equation}
By \eqref{lambdaxi}, we have $\lambda^{\mathrm{MC}_k(\eta)}(\omega)
= \lambda_k \rho^{\mathrm{MC}_k(\eta)}(\omega)$.
Assumption \eqref{Apropalt} implies $\delta \leq \lambda^{\mathrm{MC}_k(\eta)}(\omega)/\lambda_k$
$\leq\delta^{-1}$ for all $k\in\mathbb{N}_0$, $\eta\in\Omega_N$ and $\mathbb{P}$-a.e.\ $\omega$, which in turn
implies
\be{pot8}
\delta \leq \frac{a^\omega((a,b)(c,d))}{a((a,b)(c,d))} \leq \delta^{-1}
\quad \forall\,a,b,c,d \in \Omega_N \text{ for } \mathbb{P}\text{-a.e. } \omega,
\end{equation}
where $a$ is the transition kernel in \eqref{32b}, but with $c_k$ replaced by $\bar{c}_k(N)$ in
\eqref{barck}. Inserting these bounds into the formulas for the capacities in \eqref{pot2} and
\eqref{pot4}, and recalling \eqref{fdh:jointGreenpottheo}, we see that
\be{pot9}
\exists\,\delta'>0\colon\quad \delta' \leq
\frac{G^\omega((0,0),(\eta,\eta^\prime))}{G((0,0),(\eta,\eta^\prime))} \leq \delta'^{-1}
\quad \forall\,\eta,\eta^\prime \in \Omega_N \text{ for } \mathbb{P}\text{-a.e. } \omega.
\end{equation}
This shows that the Green function for the random walk in random environment is comparable
to the Green function of the homogeneous random walk. Hence the argument in Step 1 carries over.
Note that $P^\omega_{(0,0)}(\hat\tau_{(0,0)} = \infty)$ in \eqref{pot5} is comparable to
$P_{(0,0)}(\hat\tau_{(0,0)} = \infty)$, which is a strictly positive constant when $G<\infty$.
Consequently, by the observation made below \eqref{pot5}, also $P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)}
= \hat\tau_{(0,0)} = \infty)$ in \eqref{pot5} is comparable to $P_{(0,0)}(\tau_{(\eta,\eta^\prime)}
= \hat\tau_{(0,0)} = \infty)$ when $G,G^\omega<\infty$.
It remains to show that, under assumption~\eqref{Apropalt}, $G^\omega<\infty$ if and only if
$G<\infty$. This is easy. Indeed, if $G<\infty$, then $P_{(0,0)}(\hat\tau_{(0,0)} = \infty)>0$,
hence $P^\omega_{(0,0)}(\hat\tau_{(0,0)} = \infty)>0$, and hence $G^\omega<\infty$ by
\eqref{fdh:jointGreenpottheo} because $P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \infty) \leq 1$.
Conversely, if $G=\infty$, then $P_{(0,0)}(\hat\tau_{(0,0)} = \infty)=0$, hence $P^\omega_{(0,0)}
(\hat\tau_{(0,0)} = \infty)=0$, and hence $G^\omega=\infty$ by \eqref{fdh:jointGreenpottheo}
because $P^\omega_{(0,0)}(\tau_{(\eta,\eta^\prime)} < \infty)>0$.
\hfill $\square$
\end{proof}
\subsection{Zero-one law}
\label{ss.zero-one}
To conclude the proof of the dichotomy in Theorem~\ref{T.coexcritNfin}, we use the following
zero-one law.
\begin{lemma}[{\bf Zero-one law}]
\label{zero-one}
For every $N \in \mathbb{N} \backslash \{1\}$ and $\mathbb{P}$-a.e.\ $\omega$, $H_N(\omega;\infty) = \infty$
if and only if $\bar{E}^\omega[H_N(\omega;\infty)] = \infty$. \hfill $\square$
\end{lemma}
\begin{proof}\label{pr.2231}
The proof comes in five steps.
\paragraph{Step 1.}
For $M,N \in \mathbb{N}$, let $H_N^{(M)}(\omega;\infty)$ denote the truncation of $H_N(\omega;\infty)$
obtained by setting $\lambda_k=0$ for $k>M$ (no resampling in blocks of hierarchical size larger
than $M$). The key to the proof is the following second-moment estimate:
\be{HNM2nd}
\quad \exists\,\,C<\infty\colon\qquad \bar{E}^\omega\big[\big(H_N^{(M)}(\omega;\infty)\big)^2\big]
\leq C \big(\bar{E}^\omega\big[H_N^{(M)}(\omega;\infty)\big]\big)^2
\qquad \forall\,M,N\in\mathbb{N}.
\end{equation}
Before proving \eqref{HNM2nd}, we complete the proof of Theorem~\ref{T.coexcritNfin}.
By Cauchy-Schwarz, for any non-negative random variable $V$ we have
\be{VCS}
\bar{P}^\omega(V>0) \geq (\bar{E}^\omega[V])^2/\bar{E}^\omega[V^2].
\end{equation}
Picking $V=H_N^{(M)}(\omega;\infty)/\bar{E}^\omega[H_N^{(M)}(\omega;\infty)]$ in \eqref{VCS}
and using \eqref{HNM2nd}, we obtain
\be{HNM1}
\bar{P}^\omega\Big(H_N^{(M)}(\omega;\infty)/\bar{E}^\omega\big[H_N^{(M)}(\omega;\infty)\big] > 0\Big)
\geq \frac{1}{C} \qquad \forall\,M,N \in \mathbb{N}.
\end{equation}
Since $H_N^{(M)}(\omega;\infty) \leq H_N(\omega;\infty)$ and the lower bound in \eqref{HNM1} is
uniform in $M$ and $N$, it follows that if $\bar{E}^\omega[H_N(\omega;\infty)] = \lim_{M\to\infty}
\bar{E}^\omega[H_N^{(M)}(\omega;\infty)] = \infty$, then $\bar{P}^\omega(H_N(\omega;\infty)=\infty)
\geq 1/C$. By \eqref{ak12}, $\{\omega\colon\,H_N(\omega;\infty)=\infty\}$ is an element of the
sigma-algebra at infinity defined in \eqref{fdh:tail}, which is trivial. The latter event therefore has
probability either 0 or 1, and since it has positive probability we get the claim.
\paragraph{Step 2.}
Write out (recall \eqref{ak12})
\be{2ndmomentwrite1}
\begin{aligned}
\bar{E}^\omega\big[\big(H_N^{(M)}(\omega;\infty)\big)^2\big]
&= \sum_{k,l=0}^M N^{-k-l} \sum_{ {\eta,\eta'\in\Omega_N} \atop {d_{\Omega_N}(\eta,\eta') \leq k} }
\sum_{ {\zeta,\zeta'\in\Omega_N} \atop {d_{\Omega_N}(\zeta,\zeta') \leq l} }
\lambda^{\mathrm{MC}_k(\eta)}(\omega) \lambda^{\mathrm{MC}_l(\zeta)}(\omega)\\
&\qquad \times \bar{E}^\omega\left[\int_0^\infty
\mathrm{d} s\,1_{\{Y_s(\omega)=\eta,Y'_s(\omega)=\eta'\}} \int_0^\infty
\mathrm{d} u\,1_{\{Y_u(\omega)=\zeta,Y'_u(\omega)=\zeta'\}} \right]\\
& = \sum_{k,l=0}^M N^{-k-l} \sum_{ {\eta,\eta'\in\Omega_N} \atop {d_{\Omega_N}(\eta,\eta') \leq k} }
\sum_{ {\zeta,\zeta'\in\Omega_N} \atop {d_{\Omega_N}(\zeta,\zeta') \leq l} }
\lambda^{\mathrm{MC}_k(\eta)}(\omega) \lambda^{\mathrm{MC}_l(\zeta)}(\omega)\\
&\qquad \times 2\, G^\omega((0,0),(\eta,\eta'))\,G^\omega((\eta,\eta'),(\zeta,\zeta'))\\
&= 2 \sum_{k,l=0}^M N^{-k-l} \sum_{\xi \in \Omega^{(k)}_N} \lambda^{\xi}(\omega)
\sum_{\xi' \in \Omega^{(l)}_N} \lambda^{\xi'}(\omega)\\
&\qquad \times \sum_{\eta,\eta' \in B_{|\xi|}(\xi)} G^\omega((0,0),(\eta,\eta'))
\sum_{\zeta,\zeta' \in B_{|\xi'|}(\xi')} G^\omega((\eta,\eta'),(\zeta,\zeta')).
\end{aligned}
\end{equation}
In what follows, we consider the hierarchical random walk with homogeneous migration coefficients
$\bar{c}_k$ defined in \eqref{barck}. In Step 4 we incorporate the $\omega$-dependence.
\medskip\noindent
Use symmetry to replace $\sum_{k,l=0}^M$ by $2 \sum_{k,l=0}^M 1_{\{k<l\}} + \sum_{k,l=0}^M
1_{\{k=l\}}$. Due to the ultrametricity of the hierarchical distance and the isotropy of the hierarchical
random walk, we have $G((\eta,\eta'),(\zeta,\zeta'))=G((0,0),(\zeta,\zeta'))$ for all $\eta,\eta' \in
B_{|\xi|}(\xi)$ in the following three cases (where $\xi<\xi'$ means that $\xi'$ is an ancestor of $\xi$):
\medskip
\begin{tabular}{ll}
&(1) $k<l$, $\xi \nless \xi'$ and $\zeta,\zeta' \in B_{|\xi'|}(\xi')$.\\
&(2) $k<l$ and $\xi < \xi'$ and $\zeta,\zeta' \in B_{|\xi'|}(\xi') \backslash B_{|\xi|}(\xi)$.\\
&(3) $k=l$, $\xi \neq \xi'$ and $\zeta,\zeta' \in B_{|\xi'|}(\xi')$.
\end{tabular}
\medskip\noindent
Therefore we have
\be{2ndmomentwrite2}
\bar{E}\big[\big(H_N^{(M)}(\omega;\infty)\big)^2\big]
= 2\bar{E}\big[H_N^{(M)}(\omega;\infty)\big]^2 + R
\end{equation}
with $R$ a correction term given by
\be{Rdef}
\begin{aligned}
R
&= 4 \sum_{0 \leq k < l \leq M}
N^{-k-l} \sum_{\xi \in \Omega^{(k)}_N} \lambda^{\xi}(\omega)
\sum_{\xi' \in \Omega^{(l)}_N} \lambda^{\xi'}(\omega) \,1_{\{\xi<\xi'\}}
\sum_{\eta,\eta' \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta'))\\
&\qquad \qquad \times
\sum_{ {\zeta,\zeta' \in B_{|\xi'|}(\xi')} \atop {\zeta \in B_{|\xi|}(\xi)
\text{ and/or } \zeta' \in B_{|\xi|}(\xi)} }
\big[G((\eta,\eta'),(\zeta,\zeta'))-G((0,0),(\zeta,\zeta'))\big]\\
&\quad + 2 \sum_{0 \leq k \leq M} N^{-2k}
\sum_{\xi \in \Omega^{(k)}_N} [\lambda^{\xi}(\omega)]^2
\sum_{\eta,\eta' \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta'))\\
&\qquad\qquad\times \sum_{\zeta,\zeta' \in B_{|\xi|}(\xi)}
\big[G((\eta,\eta'),(\zeta,\zeta'))-G((0,0),(\zeta,\zeta'))\big].
\end{aligned}
\end{equation}
If $R$ would be absent from \eqref{2ndmomentwrite2}, then we would have proved \eqref{HNM2nd}
with $C=2$. Thus, it remains to show that $R$ can only raise the constant. We will do this by showing
that $R \leq O(N^{-2})\,\bar{E}[H_N^{(M)} (\omega;\infty)]^2$ as $N\to\infty$, uniformly in $M$, and
by appealing to the observation made in \eqref{fdh:Ncomp}.
\paragraph{Step 3.}
By translation invariance, $G((\eta,\eta'),(\zeta,\zeta')) = G((0,0),(\zeta-\eta,\zeta'-\eta'))$. By isotropy,
$\sum_{\zeta,\zeta' \in B_{|\xi|}(\xi)} G((0,0),(\zeta-\eta,\zeta'-\eta'))= \sum_{\zeta,\zeta' \in B_{|\xi|}(\xi)}
G((0,0),(\zeta,\zeta'))$ for all $\eta,\eta' \in B_{|\xi|}(\xi)$. Hence, in the first sum in \eqref{Rdef} the term
with $\zeta,\zeta' \in B_{|\xi|}(\xi)$ vanishes, while the second sum in \eqref{Rdef} vanishes altogether,
and so $R$ simplifies to
\be{Rsym1}
\begin{aligned}
R &= 8 \sum_{0 \leq k < l \leq M}
N^{-k-l} \sum_{\xi \in \Omega^{(k)}_N} \lambda^{\xi}(\omega)
\sum_{\xi' \in \Omega^{(l)}_N} \lambda^{\xi'}(\omega) \,1_{\{\xi<\xi'\}}
\sum_{\eta,\eta' \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta'))\\
&\qquad \qquad \times
\sum_{ {\zeta \in B_{|\xi|}(\xi)} \atop {\zeta' \in B_{|\xi'|}(\xi') \backslash B_{|\xi|}(\xi)} }
\big[G((0,0),(\zeta-\eta,\zeta'-\eta'))-G((0,0),(\zeta,\zeta'))\big].
\end{aligned}
\end{equation}
By isotropy, $\sum_{\zeta \in B_{|\xi|}(\xi)} G((0,0),(\zeta-\eta,\zeta'-\eta')) = \sum_{\zeta \in B_{|\xi|}(\xi)}
G((0,0),(\zeta,\zeta'-\zeta))$ for all $\eta,\eta' \in B_{|\xi|}(\xi)$ when $\zeta' \in B_{|\xi'|}(\xi') \backslash
B_{|\xi|}(\xi)$, and so $R$ simplifies further to
\be{Rsym2}
\begin{aligned}
R &= 8 \sum_{0 \leq k < l \leq M}
N^{-k-l} \sum_{\xi \in \Omega^{(k)}_N} \lambda^{\xi}(\omega)
\sum_{\xi' \in \Omega^{(l)}_N} \lambda^{\xi'}(\omega) \,1_{\{\xi<\xi'\}}
\sum_{\eta,\eta' \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta'))\\
&\qquad \qquad \times
\sum_{\zeta \in B_{|\xi|}(\xi)}
\left[ \sum_{\zeta'' \in B_l(0) \backslash B_k(0)} G((0,0),(\zeta,\zeta''))
-\sum_{\zeta' \in B_{|\xi'|}(\xi') \backslash B_{|\xi|}(\xi)}
G((0,0),(\zeta,\zeta'))\right].
\end{aligned}
\end{equation}
If $0 \in B_{|\xi'|}(\xi')$, then $B_l(0) = B_{|\xi'|}(\xi')$, in which case the term between brackets equals
\be{bracket}
\sum_{\zeta' \in B_{|\xi|}(\xi)} G((0,0),(\zeta,\zeta'))
- \sum_{\zeta'' \in B_k(0)} G((0,0),(\zeta,\zeta'')).
\end{equation}
If also $0 \in B_{|\xi|}(\xi) \subset B_{|\xi'|}(\xi')$, then also $B_k(0) = B_{|\xi|}(\xi)$, in which case the
latter difference vanishes. Hence we obtain the bound
\be{Rsym3}
\begin{aligned}
R &\leq 8 \sum_{0 \leq k < l \leq M}
N^{-k-l} \sum_{\xi \in \Omega^{(k)}_N} \lambda^{\xi}(\omega)
\sum_{\xi' \in \Omega^{(l)}_N} \lambda^{\xi'}(\omega) \,1_{\{\xi<\xi'\}}
\sum_{\eta,\eta' \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta'))\\
&\times
\left[1_{\{0 \nleq \xi, \,0 \leq \xi'\}}
\sum_{\zeta,\bar\zeta \in B_{|\xi|}(\xi)} G((0,0),(\zeta,\bar\zeta))
+ 1_{\{0 \nleq \xi'\}}
\sum_{ {\zeta \in B_{|\xi|}(\xi)} \atop {\bar\zeta \in B_l(0)\backslash B_k(0)} }
G((0,0),(\zeta,\bar\zeta))
\right].
\end{aligned}
\end{equation}
The sums over $\eta,\eta'$ and $\zeta,\bar\zeta$ can be computed with the help of \eqref{eq:ak*}.
Recalling \eqref{eq:ak**}--\eqref{eq:ak101a}, we obtain
\be{rest2}
\begin{aligned}
&\sum_{\eta,\eta' \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta'))
= \sum_{\zeta,\bar\zeta \in B_{|\xi|}(\xi)} G((0,0),(\eta,\eta'))\\
&\qquad \sim N^{2k}\,\frac{\bar{c}_0(N)}{2N^{k+d(\xi)}\bar{c}_{k+d(\xi)}(N)}
= N^k \frac{\bar{c}_0(N)}{2N^{d(\xi)}\bar{c}_{k+d(\xi)}(N)}
\end{aligned}
\end{equation}
with $d(\xi) = d_{\Omega_N^{(k)}}(0^k,\xi)$ and
\be{rest3}
\begin{aligned}
&\sum_{ {\zeta \in B_{|\xi|}(\xi)} \atop {\bar\zeta \in B_l(0) \backslash B_k(0)} }
G((0,0),(\zeta,\bar\zeta))
\sim \sum_{ {\zeta \in B_{|\xi|}(\xi)} \atop {\bar\zeta \in B_l(0) \backslash B_k(0)} }
\frac{\bar{c}_0(N)}{N^{l+d'(\xi')}\bar{c}_{d''(\bar\zeta)}(N)}\\
&\qquad = N^{-l+k} \sum_{\bar\zeta \in B_l(0) \backslash B_k(0)}
\frac{\bar{c}_0(N)}{N^{d'(\xi')}\bar{c}_{d''(\bar\zeta)}(N)}
\sim N^k \frac{\bar{c}_0(N)}{N^{d'(\xi')}\bar{c}_l(N)}
\end{aligned}
\end{equation}
with $d'(\xi')= d_{\Omega_N^{(l)}}(0^l,\xi')$ and $d''(\bar\zeta)=d_{\Omega_N}(0,\bar\zeta)$.
Here we use that $\xi \neq 0^k$ when $0 \nleq \xi$ and $\xi' \neq 0^l$ when $0 \nleq \xi'$, and
also that $l+d'(\xi')>d''(\bar\zeta)$ for all $\bar\zeta \in B_l(0)$. Inserting \eqref{rest2}--\eqref{rest3}
into \eqref{Rsym3}, we get
\be{rest4}
\begin{aligned}
R &\leq 8\,[1+o(1)] \sum_{0 \leq k < l \leq M} N^{k-l}
\sum_{\xi \in \Omega^{(k)}_N} \lambda^{\xi}(\omega)
\sum_{\xi' \in \Omega^{(l)}_N} \lambda^{\xi'}(\omega) \,1_{\{\xi<\xi',\,0 \nleq \xi'\}}\\
&\times
\left[
1_{\{0 \nleq \xi, \,0 \leq \xi'\}}
\left(\frac{\bar{c}_0(N)}{2N^{d(\xi)}\bar{c}_{k+d(\xi)}(N)} \right)^2
+ 1_{\{0 \nleq \xi'\}}\,
\frac{\bar{c}_0(N)}{2N^{d(\xi)}\bar{c}_{k+d(\xi)}(N)}
\,\frac{\bar{c}_0(N)}{N^{d'(\xi')}\bar{c}_l(N)}
\right].
\end{aligned}
\end{equation}
\paragraph{Step 4.}
If $0 \nleq \xi$, then $d(\xi) \in \mathbb{N}$. Hence the first part of \eqref{rest4} equals $8\,[1+o(1)]$ times
\be{rest5}
\sum_{0 \leq k < l \leq M} N^{k-l} \lambda_k \lambda_l \sum_{d=1}^l
\left(\frac{\bar{c}_0(N)}{2N^d\bar{c}_{k+d}(N)}\right)^2
\sum_{ {\xi \in \Omega^{(k)}_N} \atop {d(\xi)=d} } \rho^{\xi}(\omega)\rho^{\xi^{l-k}}(\omega),
\end{equation}
where we recall \eqref{lambdaxi} and write $\xi^{l-k}$ to denote the ancestor of $\xi$ at height $l$.
Because $\{\rho^\xi(\omega)\colon\,\xi \in \Omega_N^{\mathbb{T}}\}$ is stationary, ergodic and tail
trivial (recall \eqref{fdh:tail}), the last sum scales as $\sim N^d \mathbb{E}[\rho^{0^k}(\omega)\rho^{0^l}(\omega)]$,
where the expectation is finite because of \eqref{Aprop}. Hence \eqref{rest5} is
\be{rest6}
\leq C[1+o(1)]\,\tfrac14\bar{c}_0(N)^2
\sum_{0 \leq k < l \leq M} N^{k-l} \lambda_k \lambda_l \sum_{d=1}^l
\frac{1}{N^d\bar{c}_{k+d}(N)^2}.
\end{equation}
The last sum scales as $\sim 1/N\bar{c}_{k+1}(N)^2$, and so \eqref{rest5} is
\be{rest7}
\leq C[1+o(1)]\,\tfrac14\bar{c}_0(N)
\sum_{k=0}^M \frac{\lambda_k}{\bar{c}_{k+1}(N)^2}
\sum_{l=k+1}^M \lambda_l N^{k-l-1}
= O(N^{-2})\,\bar{E}[H_N^{(M)}(\omega;\infty)]^2,
\end{equation}
where the equality follows from \eqref{meanhazfinal} with $k,l$ truncated at $M$.
\medskip\noindent
If $0 \nleq \xi'$, then $d'(\xi') \in \mathbb{N}$ and $d(\xi)=l-k+d'(\xi')$. Hence the second part of
\eqref{rest4} equals $8\,[1+o(1)]$ times
\be{rest8}
\sum_{0 \leq k < l \leq M} N^{k-l} \lambda_k \lambda_l \sum_{d'\in\mathbb{N}}
\frac{\bar{c}_0(N)}{2N^{l-k+d'}\bar{c}_{l+d'}(N)}\,\frac{\bar{c}_0(N)}{N^{d'}\bar{c}_l(N)}
\sum_{ {\xi \in \Omega^{(k)}_N} \atop {d(\xi)=l-k+d'} } \rho^{\xi}(\omega)\rho^{\xi^{l-k}}(\omega).
\end{equation}
The last sum is $\leq C[1+o(1)] N^{l-k+d'}$. Hence \eqref{rest8} is
\be{rest9}
\leq C[1+o(1)]\,\tfrac12\bar{c}_0(N)^2
\sum_{0 \leq k < l \leq M} N^{k-l} \lambda_k \lambda_l \sum_{d'\in\mathbb{N}}
\frac{1}{N^{d'}\bar{c}_{l+d'}(N)\bar{c}_l(N)}.
\end{equation}
The last sum scales as $\sim 1/N\bar{c}_{l+1}(N)\bar{c}_l(N)$, and so \eqref{rest8} is
\be{rest10}
\leq C[1+o(1)]\,\tfrac12\bar{c}_0(N)^2
\sum_{k=0}^M \lambda_k \sum_{l=k+1}^M
\frac{\lambda_l}{\bar{c}_{l+1}(N)\bar{c}_l(N)}\,N^{k-l-1}
= O(N^{-2})\,\bar{E}[H_N^{(M)}(\omega;\infty)]^2.
\end{equation}
\paragraph{Step 5.}
We can again use \eqref{pot9} to show that the proof carries over to the random walk in random
environment.
\end{proof}
\medskip\noindent
Lemmas~\ref{meanhazard}--\ref{zero-one} combine to yield Theorem~\ref{T.coexcritNfin}
(recall the discussion at the beginning of this section).
\section{Multi-scale analysis}
\label{s.mkvrand}
In this section we prove Theorem~\ref{mainth}. We first consider a {\em mean-field system}, i.e.,
the geographic space is $G=\{1,\ldots, N\}$ with $N \to \infty$. In Section~\ref{ss.mkvlim} we
look at this system on time scale $t$ (on which the single components evolve) and on time
scale $Nt$ (on which the block average evolves). In Section~\ref{ss.hierarmfl} we use the results
to analyze the system on $\Omega_N$ as $N \to\infty$. Our general strategy runs parallel to that
in \cite{GHKK14} for the homogeneous model. We only point out which {\em new issues} arise.
Thus, this section is \emph{not autonomous},
the principal steps of the arguments are given but not all formulas are repeated, and for an
understanding of the fine details the reader must check the relevant passages in \cite{GHKK14}.
\subsection{The mean-field finite-system scheme}
\label{ss.mkvlim}
As geographic space and transition kernel we take
\be{ag45}
\Omega = \{1,\dots,N\}, \qquad a(i,j) = \frac{1}{N}, \quad i,j \in \Omega.
\end{equation}
As migration rate we take $c_0$, and as resampling measures
\be{ag45alt}
\Lambda^i = \lambda_0\chi^i,\quad i\in \Omega,
\end{equation}
with total masses $\rho^i = \chi^i((0,1])$. We assume that $(\chi^i)_{i \in \mathbb{N}}$ is stationary and
ergodic such that $\varrho^i$ has mean $1$. We also allow a component with Fleming-Viot
resampling at rate $ d_0 $. The corresponding stochastic system is denoted by $(Z^{(N)}(t))_{t \geq 0}$
with $Z^{(N)}(t) = (Z_1^{(N)}(t),\ldots,Z^{(N)}_N(t))$.
We consider time scales $t$ and $Nt$ for the components, respectively, the block average:
\be{ag46}
\begin{aligned}
&(Z^{(N)}(t))_{t \geq 0},\\
&(\bar Z^{(N)}(t))_{t \geq 0} \text{ with }
\bar Z^{(N)}(t) = \frac{1}{N} \sum\limits^N_{i=1} Z^{(N)}_i (Nt).
\end{aligned}
\end{equation}
\begin{theorem}[{\bf [Mean-field finite-system scheme}]
\label{T.McKVl}
Suppose that the initial state is i.i.d.\ with mean measure $\theta \in \mathcal{P}(E)$. Then
\be{ag47}
\lim_{N\to\infty} \mathcal{L} \left[(Z^{(N)}(t))_{t \geq 0}\right]
= \bigotimes_{i \in \mathbb{N}} \mathcal{L} \left[(Z^{c_0, d_0, \Lambda^i}_\theta (t))_{t \geq 0}\right]
\end{equation}
and
\be{ag48}
\lim_{N\to\infty} \mathcal{L} \left[(\bar Z^{(N)}(t))_{t \geq 0}\right]
= \mathcal{L} \left[(Z^{0,d_1,0}_\theta (t))_{t \geq 0}\right],
\end{equation}
where $(Z_\theta^{c,d,\Lambda}(t))_{t\geq 0}$ is the McKean-Vlasov process defined in
Section~{\rm \ref{sss.mfl}}.
\hfill $\square$
\end{theorem}
\begin{proof}\label{pr.2519}
We follow \cite[Section~6]{GHKK14}. The proof of \eqref{ag47} carries over in a straightforward
way. In the proof of \eqref{ag48} a new issue arises: the increasing process of the limit process
incorporates an additional averaging over the random environment controlling the resampling for
the single components. This is handled as follows.
Calculate the generator for a polynomial of $\bar Z^{(N)}(t)$, namely, a function $F$ of the form
\be{ag49}
F(z) = \langle f,z^{\otimes n}\rangle, \quad f \in C_b(E^n,\mathbb{R}),
\end{equation}
applied to a $z \in E$ of the form $z = \frac{1}{N} \sum_{i=1}^N z_i$. This expression can be
expanded in terms of sums of products of monomials of single components. The action of the
generator was calculated and analysed in \cite[Section~6]{GHKK14}. We can argue in the
same way with the following changes. In the action of the generator, integrals are taken with
respect to the random sequence of resampling measures $(\Lambda^i)_{i \in \Omega}$ rather
than a fixed resampling measure $\Lambda$. This entails that for the block average we get a
sum of terms where the random sequence $(\rho^i)_{i\in\Omega}$ appears as weights. This in
turn requires us to change the definition of the set of configurations on which the generator
converges in the limit as $ N\to\infty$ (see \cite[Eq.~(6.41)--(6.42)]{GHKK14}) as follows.
Let $\mathbb{B}^\ast$ be the set of $\underline{x} = (x_i)_{i\in\Omega}\in \mathcal{P}(E)^\mathbb{N}$ with
\begin{align}
\label{ag60}
\lim_{N\to\infty} \mathcal{L}\left[ \frac{1}{N} \sum\limits^N_{i=1} \delta_{(\chi^i(\omega),x_i)} \right]
= \Gamma \in \mathcal{P}\big(\mathcal{M}_f([0,1]) \times \mathcal{P}(E)\big),
\end{align}
where
\begin{align}
\label{ag61}
\Gamma(\,\cdot\,,\mathcal{P}(E)) = \mathcal{L}[\chi^1],\quad \Gamma (\mathrm{d} x_1 \mid \chi^1)
= \nu^{c_0,0,\chi^1}(\mathrm{d} x_1), \quad x_1 \in \mathcal{P}(E).
\end{align}
In order to calculate the sum of the resampling operators as in \cite[Eq.~(6.46)]{GHKK14}, we have
to account for the presence of $\chi^i$, $i \in \Omega$, and invoke the law of large numbers for the
expression in the variance formula, namely, $2c_0/(2c_0+\lambda_0\rho^i +2d_0)$, $i\in\Omega$.
We write the latter as
\be{ag61alt}
\frac{c_0}{c_0+\mu_0\rho^i+d_0}, \qquad i \in \Omega.
\end{equation}
The expressions appearing in the generator, which are averages of local functions of the configuration
and their shifts to any of the $N$ locations, result in the same expression as the one we obtain by
using \eqref{ag61alt} averaged over $i\in\Omega$. In the limit as $N\to\infty$ this leads to the recursion
formula in \eqref{diffusion-constants} for $k=0$. With these changes, the argument runs as in the case
of the homogeneous environment.
\end{proof}
\subsection{The hierarchical mean-field limit}
\label{ss.hierarmfl}
In this section we prove the results claimed in Section~\ref{sss.blav}. The strategy of the proof is to
approximate our system with \emph{infinitely} many hierarchies of components and time scales by
systems with \emph{finitely} many hierarchies of components and time scales, uniformly in $N$. The
latter are analyzed by using the multiscale analysis of the mean-field system. In Section~\ref{sss.2l}
we consider 2-level systems with $N^2$ components, in Section~\ref{sss.kl} $k$-level systems with
$N^k$ components, and in Section~\ref{sss.3l} we pass to the limit $k\to\infty$ of infinitely many
hierarchies. Along the way we make frequent reference to Dawson, Greven and Vaillancourt~\cite{DGV95}
and the work on the homogeneous version of the model in \cite{GHKK14}.
\subsubsection{The $2$-level system on 3 time scales}
\label{sss.2l}
The geographic space is $G_{N,2}=\{0,1,\ldots,N-1\}^2 = G_{N,1}^2$. We pick $d_0>0$,
$c_0,c_1,\mu_0,\mu_1>0$ and $c_k,\mu_k=0$ for $k \geq 2$. We choose the random environment
that is obtained by restricting the random environment of Section~\ref{s.model} to the subtree
corresponding to the 2-block around 0. We show that, on time scales $t$ and $Nt$, we obtain the
same limiting objects as described in Section~\ref{ss.mkvlim}, but with additional volatility and
block resampling.
For the 1-block averages we use the notation
\be{add21}
Y^{(N)}_\eta(t) = N^{-1} \sum_{\sigma \in G_{N,1}} X_{(\sigma,\eta)}^{(N)}(t),
\qquad \eta \in G_{N,1},
\end{equation}
and for the 2-block average (= total average)
\be{add20}
Z^{(N)}(t) = N^{-2} \sum_{(\sigma,\eta) \in G_{N,2}} X_{(\sigma,\eta)}^{(N)}(t)
= N^{-1} \sum_{\eta \in G_{N,1}} Y^{(N)}_\eta(t).
\end{equation}
\begin{proposition}[{\bf [Two-level rescaling}]
\label{P.2lresc}
Under the above assumptions,
\be{ak15new}
\begin{aligned}
\lim_{N\to\infty} \mathcal{L}\left[\left(X^{(N)}_{(\sigma,\eta)}(t)\right)_{t \geq 0}\right]
&= \mathcal{L}\left[\left(Z^{c_0, d_0, \Lambda^{\mathrm{MC}_1
(\cdot,\eta)}(\omega)}_\theta (t)\right)_{t \geq 0}\right]
\qquad \forall\, (\sigma,\eta) \in G_{N,2},\\
\lim_{N\to\infty} \mathcal{L}\left[\left(Y^{(N)}_\eta (Nt)\right)_{t \geq 0}\right]
&= \mathcal{L}\left[\left(Z^{c_1, d_1, \Lambda^{\mathrm{MC}_2
(\cdot,\cdot)}(\omega)}_\theta (t)\right)_{t \geq 0}\right]
\qquad \forall\, \eta \in G_{N,1},\\
\lim_{N\to\infty} \mathcal{L}\left[\left(Z^{(N)} (N^2 t)\right)_{t \geq 0}\right]
&=\mathcal{L}\left[\left(Z^{0,d_2,0}_\theta (t)\right)_{t \geq 0}\right],
\end{aligned}
\end{equation}
with
\be{ak16new}
d_1 = \mathbb{E}_{\mathcal{L}_\rho} \left[\frac{c_0 (\mu_0 \rho(\omega) + d_0)}
{c_0 + (\mu_0 \rho(\omega) + d_0)}\right],
\qquad
d_2 = \mathbb{E}_{\mathcal{L}_\rho} \left[\frac{c_1 (\mu_1 \rho(\omega) + d_1)}
{c_1 + (\mu_1 \rho(\omega) + d_1)}\right].
\end{equation}
\hfill $\square$
\end{proposition}
To prove the above results in the homogeneous environment, we used {\em uniform estimates}
for higher-order perturbations of generators. These no longer hold in the random environment,
due to the unboundedness of the random resampling rates $\rho^{(\cdot,\eta)}(\omega)$. (There
is no problem under assumption \eqref{Apropalt}, and the proof carries over from \cite{GHKK14}.)
To handle this problem we first consider the system where the coefficients $ \lambda^{\mathrm{MC}_k
(\cdot,\eta)} (\omega)$, $k=1,2$, are truncated at level $M<\infty$. For this system we show, with
the help of a coupling argument, that on time scale $N^k t$, $k=1,2$, and averaged over the
random environment and the dynamics, the effect of the truncation goes to zero as $M\to\infty$.
The same holds for the limiting objects, so that we get the claim by using the existence of the
expectation in combination with the stationarity of $\omega$.
To get tightness of the approximating sequence of processes, as in \cite[Eq.~(7.52), p.~117]{GHKK14},
we use the fact that the laws conditioned on the environment $ \omega $ of the averages in
\eqref{add21}--\eqref{add20} are tight. To prove the latter, we use the criterion of Joffe and Metivier
in the form as given in Dawson~\cite[p.~55]{D93}, observing that $\chi^{\mathrm{MC}_k(\cdot,\eta)}
(\omega)$, $\eta \in G_{N,1}$, $k=1,2$, are integrable uniformly in $N$. To check the criterion, we
observe that we can code the information on the random environment into the initial condition of
the process. With this observation, the proof works as for the homogeneous environment.
\subsubsection{The $k$-level system on $k+1$ time scales}
\label{sss.kl}
The reasoning addresses the same points raised above and runs otherwise exactly as in
\cite[Section 7.2]{GHKK14}.
\subsubsection{The infinite-level system on infinitely many time scales}
\label{sss.3l}
The problem is again the extension of the uniform perturbation arguments, which have
to be adapted to guarantee that cutting off higher hierarchical levels leads to an approximation
by finite systems, for which we can apply the reasoning in the previous section, on the relevant
time scales. To get the necessary arguments and estimates we refer the reader to the material
in \cite[Sections 8.1--8.2]{GHKK14}.
The argument used for the homogeneous environment to obtain uniforms bounds does not apply
because the perturbation of the migration and the resampling coming from the hierarchical levels
$\geq k+1$ is unbounded. However, the perturbation terms can be stochastically bounded by a
random variable that has a {\em finite expectation} over the random environment. Again, it suffices to
show with the help of a coupling argument that the stochastic dynamics with $k$ hierarchical
levels approximates the infinite stochastic dynamics on time scales $tN^l$ with $0 \leq l \leq k$.
Apart from that the argument is the same.
\subsection{Dichotomy in the hierarchical mean-field limit}
\label{ss.HMFLdicho}
In this section we prove Theorem~\ref{T.dicho}. First, we argue that the entrance law exists, a fact
that was established in Dawson, Greven and Vaillancourt~\cite{DGV95}[Section 6(a), Proposition 6.2]
for the Fleming-Viot model, based on a variance estimate and the convergence of the sum in the
coexistence criterion. The argument from that paper carries over despite the $\omega$-dependence
of the transition kernels of the interaction chain (read this of from \eqref{e2691} and \eqref{e2749alt} below).
Next, we argue that the dichotomy holds. Here, we again follow the strategy for the homogeneous
environment by calculating the variance of $\langle M^{(j)}_{\eta,0},f \rangle$ for every
$\eta\in\Omega_\infty$ and $f \in C_b(E,\mathbb{R})$ and showing that as $j\to\infty$ this variance
converges to zero, respectively, remains positive, depending on whether the sum in \eqref{inftydich}
is infinite or finite.
The variance formula reads
\begin{equation}\label{e2690}
\var_{\nu_\theta^{c,d,\Lambda}} (\langle\,\cdot\,,f \rangle)
= \frac{2c}{2c+\lambda\rho(\omega)+2d} \var_\theta(f).
\end{equation}
Consequently, by iteration,
\begin{equation}\label{e2691}
\var \langle M_{\eta,0}^{(j)}, f\rangle
= \left[\prod^{j}_{k=0} \frac{2c_k}{2c_k + \lambda_k\rho_k(\omega) + 2d_k}\right]
\var_\theta(f),
\end{equation}
where $\underline{d}=(d_k)_{k \in \mathbb{N}_0} $ is determined by the recursion relation in \eqref{diffusion-constants}.
Taking logarithms, we see that the product tends to a positive limit as $j\to\infty$ if and only if
\begin{equation}
\label{e2749}
\sum_{k\in\mathbb{N}_0} \frac{1}{c_k} (\mu_k\rho_k(\omega) + d_k) < \infty.
\end{equation}
By assumptions \eqref{Aprop}--\eqref{fdh:tail}, the sum converges $\omega$-a.s.\ if and only if
\begin{equation}
\label{e2749alt}
\sum_{k\in\mathbb{N}_0} \frac{1}{c_k} (\mu_k + d_k) < \infty.
\end{equation}
Indeed, the variance of the sum in \eqref{e2749} equals the variance of the $\rho$-field times
$\sum_{k\in\mathbb{N}_0} (\frac{\mu_k}{c_k})^2$, and the latter is bounded from above by the square
of the average of the sum. As shown in \cite[Theorem 3.7(c)]{GHKK14}, the criterion in \eqref{e2749alt}
is the same as the criterion in \eqref{inftydich}.
\section{The orbit of the renormalization transformations}
\label{s.completeproof}
In Section~\ref{ss.volcomp} we show the ordering in Theorem~\ref{T.order}.
In Sections~\ref{ss.scalpol}--\ref{ss.scalexp}, we derive the scaling behaviour
in Theorems~\ref{T.scaleFVpol}--\ref{T.scaleFVexp}.
\subsection{Random environment lowers the volatility}
\label{ss.volcomp}
\begin{proof}[Proof of Theorem~\textup{\ref{T.order}}]
Recall the notation introduced in Section~\ref{ss.random}. Fix $\underline{c}$ and $\underline{\lambda}$.
Let $\underline{d}$ be the solution of the recursion relation in \eqref{diffusion-constants}.
Let $\underline{d}^0,\underline{d}^1$ be the solutions when $\mathcal{L}_\rho$ is replaced by $\delta_0,\delta_1$
(recall that $\rho$ has mean 1 under $\mathcal{L}_\rho$). As initial values take $d^0_0 \leq
d_0 \leq d^1_0$. We use induction on $k$ to show that $d^0_k < d_k < d^1_k$ for
all $k\in\mathbb{N}$.
\begin{figure}[htbp]
\vspace{-.5cm}
\begin{center}
\setlength{\unitlength}{0.4cm}
\begin{picture}(14,10)(0,-1)
\put(-1,0){\line(1,0){14}}
\put(0,-1){\line(0,1){8}}
{\thicklines
\qbezier(0,0)(1,3)(12,4.5)
\qbezier(0,2)(1,4)(12,5)
\qbezier(0,3)(1,5)(12,5.5)
}
\qbezier[80](0,6)(5,6)(12,6)
\put(-1.5,6){$c_k$}
\put(13.5,-.3){$x$}
\end{picture}
\caption{\small Qualitative pictures of $x \mapsto f^0_k(x)$ (bottom), $x \mapsto f_k(x)$ (middle)
and $x \mapsto f^1_k(x)$ (top). All three functions are strictly increasing and strictly concave
on $[0,\infty)$, and tend to $c_k$ as $x \to \infty$.}
\label{fig-Mobius}
\end{center}
\end{figure}
Define (see Fig.~\ref{fig-Mobius})
\be{fdh:f_k}
f^0_k(x) = \frac{c_kx}{c_k+x},
\qquad
f_k(x) = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{c_k(\mu_k\rho+x)}{c_k+(\mu_k\rho+x)}\right],
\qquad f^1_k(x) = \frac{c_k(\mu_k+x)}{c_k+(\mu_k+x)}.
\end{equation}
Because $a \mapsto c_k(\mu_ka+x)/[c_k+(\mu_ka+x)]$ is strictly increasing and strictly concave
on $[0,\infty)$ for all $x \in [0,\infty)$, it follows that $f^0_k(x)<f_k(x) < f^1_k(x)$ for all
$x \in [0,\infty)$. Hence, if $d^0_k \leq d_k \leq d^1_k$, then $d^0_{k+1} = f^0_k(d^0_k)< f_k(d^0_k)
\leq f_k(d_k) = d_{k+1}$ and $d_{k+1} = f_k(d_k) < f^1_k(d_k) \leq f^1_k(d^*_k) = d^1_{k+1}$.
\end{proof}
The same argument proves the claim made in Section~\ref{ss.effectre} that $M<M^*$ for the fixed
points of \eqref{MK} (random environment) and its analogue with $\mathcal{L}_\rho$ replaced by
$\delta_1$ (average environment).
\subsection{Scaling of the volatility: polynomial coefficients}
\label{ss.scalpol}
\begin{proof}[Proof of Theorem~\textup{\ref{T.scaleFVpol}}]\label{pr.2742}
We look at each of the four parameter regimes separately.
Recall \eqref{fdh:regcond}--\eqref{Kdef}.
\medskip\noindent
(a) Let $K_k=\mu_k/c_{k-1}$, $R_k=c_k/c_{k-1}$ and $\mho_k=d_k/c_{k-1}$.
Rewrite \eqref{diffusion-constants} as
\be{fdh:recrew}
\mho_{k+1} = g_k(\mho_k) \quad \text{ with } \quad
g_k(x) = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{(K_k\rho+x)}{R_k+(K_k\rho+x)}\right].
\end{equation}
Since $g_k$ is non-decreasing on $[0,\infty)$, we have the sandwich
\be{fdh:sandwich}
g_k(0) \leq \mho_{k+1} \leq g_k(\infty) = 1.
\end{equation}
We are in the regime where $\lim_{k\to\infty} K_k = K = \infty$ and $\lim_{k\to\infty} R_k = R = 1$.
Hence $\lim_{k\to\infty} g_k(0) = 1$, and so \eqref{fdh:sandwich} yields $\lim_{k\to\infty} d_k/c_k
= \lim_{k\to\infty} \mho_k/R_k = 1/R = 1$.
\medskip\noindent
(b) Again use \eqref{fdh:recrew}. We are in the regime where $\lim_{k\to\infty} K_k = K \in (0,\infty)$
and $\lim_{k\to\infty} R_k = R = 1$. Hence, we see that $g_k$ converges point-wise to $g$ given
by
\be{fdh:flim}
g(x) = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{(K\rho+x)}{R+(K\rho+x)}\right].
\end{equation}
Both $g$ and $g_k$ are strictly increasing and strictly concave on $[0,\infty)$, with $g([0,\infty])
\subseteq [0,1]$ and $g_k([0,\infty]) \subseteq [0,1]$, with unique attracting fixed points $M \in (0,1)$
and $M_k \in (0,1)$, and with $M$ the solution of \eqref{MK}. To show that $\lim_{k\to\infty} \mho_k
= M$, we need two facts.
\begin{lemma}
\label{sk}
Let $s_k = \sup_{x \in [0,1]} |g_k(x) - g(x)|$. Then $\lim_{k\to\infty} s_k = 0$.
\hfill $\square$
\end{lemma}
\begin{proof}\label{pr.2774}
Estimate
\be{sk1}
\begin{aligned}
&\frac{(K_k\rho+x)}{R_k+(K_k\rho+x)}-\frac{(K\rho+x)}{R+(K\rho+x)}
= \frac{R}{R+(K\rho+x)} - \frac{R_k}{R_k+(K_k\rho+x)}\\
&= \frac{(RK_k-R_kK)\rho+(R-R_k)x}{[R+(K\rho+x)][R_k+(K_k\rho+x)]}
\leq \frac{(RK_k-R_kK)}{KR_k} + \frac{R-R_k}{RR_k}\,x.
\end{aligned}
\end{equation}
This gives
\be{sk2}
s_k \leq \frac{R}{K}\left|\frac{K_k}{R_k}-\frac{K}{R}\right| + \left|\frac{1}{R_k} - \frac{1}{R}\right|.
\end{equation}
Let $k\to\infty$ to get the claim.
\end{proof}
\begin{lemma}
\label{contr}
Function $g$ is a strict contraction around $M$, i.e., there exists a $\beta \in (0,1)$ such that
$\sup_{x \in [0,\infty)} (g(x)-M)/(x-M) = \beta$. \hfill $\square$
\end{lemma}
\begin{proof}\label{pr.2796}
Consider the linear function $L(x) = g(0) + [1-\frac{g(0)}{M}]x$, $x \in [0,\infty)$, which satisfies
$L(0)=g(0)$ and $L(M) = M = g(M)$ (see Fig.~\ref{fig-comparison}). Note that $g \geq L$ on $[0,M]$
while $g \leq L$ on $[M,\infty)$. Hence, we have
\be{contr1}
0 \leq \frac{g(x)-M}{x-M} \leq \frac{L(x)-M}{x-M} = 1 - \frac{g(0)}{M}.
\end{equation}
Since $g(0)>0$, we get the claim with $\beta=1-\frac{g(0)}{M}$.
\end{proof}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{0.5cm}
\begin{picture}(14,10)(0,-1.5)
\put(-1,0){\line(1,0){14}}
\put(0,-1){\line(0,1){8}}
{\thicklines
\qbezier(0,2)(1,5)(10,5)
\qbezier(0,2)(4.6,4.6)(9.2,7.2)
}
\qbezier[60](0,0)(3,3)(7,7)
\qbezier[40](4.6,0)(4.6,3)(4.6,6)
\put(10.5,5){$g(x)$}
\put(9.5,7){$L(x)$}
\put(13.5,-.3){$x$}
\put(4.2,-1){$M$}
\put(4.6,4.6){\circle*{0.35}}
\end{picture}
\caption{\small Comparison of $g$ and $L$.}
\label{fig-comparison}
\end{center}
\end{figure}
\noindent
We can now complete the proof as follows. Let $\Delta_k = |\mho_k-M|$. Then
\be{compl1}
\begin{aligned}
\Delta_{k+1} &= |\mho_{k+1}-M| \leq |\mho_{k+1}-g(\mho_k)| + |g(\mho_k)-M|\\
&= |g_k(\mho_k)-g(\mho_k)| + |g(\mho_k)-M| \leq s_k + \beta\Delta_k.
\end{aligned}
\end{equation}
Iteration yields
\be{compl2}
\Delta_{k+1} \leq \sum_{l=0}^k \beta^l s_{k-l} + \beta^{k+1}\Delta_0.
\end{equation}
It follows from Lemma~\ref{sk}--\ref{contr} that $\lim_{k\to\infty} \Delta_k = 0$.
Hence $\lim_{k\to\infty} d_k/c_k = \lim_{k\to\infty} \mho_k/R_k = M/R = M$.
\medskip\noindent
(c--d) Like in Case~(a), the scaling turns out to be the \emph{same} as for the average
environment. The proof is based on a \emph{comparison} between the recursions for the
random environment and the average environment (last two items in \eqref{fdh:f_k}). The
key idea is the following lemma, which can be viewed as a \emph{stability property}.
\begin{lemma}
\label{lem:scal1}
Let $d_0 = d^1_0$. Then, the solution of the recursion $d_{k+1}=f_k(d_k)$, $k\in\mathbb{N}_0$, is
the same as the solution of the recursion $d^1_{k+1}=f^1_k(d^1_k)$, $k\in\mathbb{N}_0$, when in
the latter recursion the coefficient $\mu_k$ is replaced by $\mu_kr_k$ with
\begin{equation}
\label{rkdef}
r_k = \frac{N_k}{D_k}, \qquad N_k = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{c_k\rho}{c_k(1+K_k\rho)+d_k}\right],
\qquad D_k = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{c_k}{c_k(1+K_k\rho)+d_k}\right].
\end{equation}
\hfill $\square$
\end{lemma}
\begin{proof}\label{pr.2860}
Check that
\begin{equation}\label{e2894}
\frac{c_k(\mu_kr_k+d_k)}{c_k+(\mu_kr_k+d_k)}
= \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{c_k(\mu_k\rho+d_k)}{c_k + (\mu_k\rho+d_k)}\right]
= d_{k+1},
\end{equation}
and use induction on $k$.
\end{proof}
\noindent
Since $\rho \mapsto c_k/[c_k(1+K_k\rho)+d_k]$ is non-increasing, we have $N_k \leq D_k
\mathbb{E}_{\mathcal{L}_\rho}[\rho] = D_k$, and so $r_k \leq 1$. The following result shows that $r_k$ tends
to 1 as $k \to \infty$ in Cases (c) and (d).
\begin{lemma}
\label{lem:scal2}
If $\lim_{k\to\infty} K_k = K = 0$, then $\lim_{k\to\infty} r_k=1$.
\hfill $\square$
\end{lemma}
\begin{proof}\label{pr.2879}
For any $C \in (0,\infty)$, we may estimate
\begin{equation}\label{e2913}
N_k \geq \frac{c_k}{c_k(1+K_kC)+d_k}\,\mathbb{E}_{\mathcal{L}_\rho}\left[\rho\,1_{\{\rho \leq C\}}\right],
\qquad D_k \leq \frac{c_k}{c_k+d_k}.
\end{equation}
Since $\lim_{k\to\infty} K_k=0$, we have $\lim_{k\to\infty} (c_k+d_k)/[c_k(1+K_kC)+d_k]
=1$, and hence
\begin{equation}\label{e2919}
\liminf_{k\to\infty} \frac{N_k}{D_k} \geq \mathbb{E}_{\mathcal{L}_\rho}\left[\rho\,1_{\{\rho \leq C\}}\right].
\end{equation}
Now let $C\to\infty$ and use that $\lim_{C\to\infty} \mathbb{E}_{\mathcal{L}_\rho}[\rho\,1_{\{\rho \leq C\}}]
= \mathbb{E}_{\mathcal{L}_\rho}[\rho]=1$ by monotone convergence.
\end{proof}
\medskip\noindent
Lemma~\ref{lem:scal1} implies that the scaling of $d_k$ is the same as the scaling of $d^1_k$
after $\mu_k$ is replaced by $\mu_kr_k$. But the latter scaling was derived in \cite{GHKK14},
and a glance at the results for Cases~(c) and (d) obtained there shows that the scaling is
unaffected by the extra factor $r_k$ because of Lemma~\ref{lem:scal2}.
\end{proof}
A technical remark is in order, for which we refer the reader to \cite[Section~11.3]{GHKK14}.
We have assumed that $k \mapsto \mu_k$ is regularly varying at infinity (recall \eqref{fdh:regcond}).
Because $\lim_{k\to\infty} r_k = 1$, also $k \mapsto r_k\mu_k$ is regularly varying at infinity.
Therefore, $(r_k\mu_k)_{k\in\mathbb{N}_0}$ can be approximated from above and from below by sequences
that have the same scaling behaviour but are \emph{smoothly varying}, i.e., for all $n\in\mathbb{N}$ their
$n$-th order discrete differences are regularly varying as well. This approximation is harmless
because the maps $\underline{c} \mapsto \underline{d}$ and $\underline{\mu} \mapsto \underline{d}$
are component-wise non-decreasing (a fact that is immediate from \eqref{diffusion-constants}), and
so the approximating sequences provide a \emph{sandwich} for the scaling. Now, if the tail exponent
of $r_k\mu_k$ is non-integer, i.e., $b \notin \mathbb{N}$ in \eqref{fdh:regcond}, then for all $n\in\mathbb{N}$ the $n$-th
order discrete differences are \emph{asymptotically monotone}. This observation is important because
it implies that certain sequences arising in \cite[Section~11.3]{GHKK14} have \emph{summable variation},
a property that is crucial for the proof of the scaling. If the tail exponent is integer, i.e., $b \in \mathbb{N}$ in
\eqref{fdh:regcond}, then the asymptotic monotonicity still holds for all $n \leq b$, which turns out to
be enough for the argument.
The extra regularity conditions on $L_c,L_\mu$ in \eqref{fdh:regcond}, which are stated in
\cite[Eqs.\ (1.79)--(1.81)]{GHKK14}, need no modification: $(r_k\mu_k)_{k\in\mathbb{N}_0}$ has the same
slowly varying function $L_\mu$ as $(\mu_k)_{k\in\mathbb{N}_0}$.
\subsection{Scaling of the volatility: exponential coefficients}
\label{ss.scalexp}
\begin{proof}[Proof of Theorem~\textup{\ref{T.scaleFVexp}}]\label{pr.2914}
We look at each of the five parameter regimes (= universality classes) separately.
Recall (\ref{expregvar}--\ref{barK}).
\paragraph{(A)}
Use \eqref{fdh:recrew}. We are in the regime where $\lim_{k\to\infty} K_k=K=\infty$ and
$\lim_{k\to\infty} R_k = c$. The same argument as in the proof of Case~(a) yields
$\lim_{k\to\infty} d_k/c_k = \lim_{k\to\infty} \mho_k/R_k = 1/c$.
\paragraph{(B)}
Let $\bar{K}_k = \bar{\mu}_k/\bar{c}_{k-1}$ and $\bar{R}_k = \bar{c}_k/\bar{c}_{k-1}$. Then
$K_k = c \bar{K}_k$ and $R_k = c \bar{R}_k$ by \eqref{expregvar}, and so \eqref{fdh:recrew}
becomes
\be{fdh:recrewbar}
\mho_{k+1} = \bar{g}_k(\mho_k) \quad \text{ with } \quad
\bar{g}_k(x) = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{(c \bar{K}_k\rho+x)}{c \bar{R}_k+(c \bar{K}_k\rho+x)}\right].
\end{equation}
We are in the regime where $\lim_{k\to\infty} \bar{K}_k = \bar{K} \in (0,\infty)$ and $\lim_{k\to\infty}
\bar{R}_k = \bar{R} = 1$. The same argument as in Case~(b) therefore yields $\lim_{k\to\infty}
d_k/c_k = \lim_{k\to\infty} \mho_k/R_k = \bar{M}/c\bar{R} = \bar{M}/c$ with $\bar{M}$ the unique
attracting fixed point of
\be{fdh:barflim}
\bar{g}(x) = \mathbb{E}_{\mathcal{L}_\rho}\left[\frac{(c\bar{K}\rho+x)}{c+(c\bar{K}\rho+x)}\right],
\end{equation}
which is the analogue of \eqref{fdh:flim}.
\paragraph{(C1)}
This case is the same as Case~(B), but with $\bar{K}=0$. The analogue of \eqref{fdh:barflim} reads
$\bar{g}(x) = x/(c+x)$. Since $\bar{g}$ has $\bar{M}=1-c \in (0,1)$ as unique attracting fixed point,
we can copy the proof of Case~(b) to get $\lim_{k\to\infty} d_k/c_k = \lim_{k\to\infty} \mho_k/R_k
= (1-c)/c\bar{R} = (1-c)/c$. \emph{Note}: In the proof of Case~(b) we used that $g(0)>0$, which fails
here. However, even when $d_0=0$, the iterates $d_k$, $k\in\mathbb{N}$, are bounded away from $0$
because the attracting fixed points of $f_k$, $k\in\mathbb{N}$, are bounded away from $0$. Hence we may
restrict the entire argument to $[\epsilon,1]$ for some $\epsilon>0$ instead of $[0,1]$, and use
that $g(\epsilon)>0$ (recall Fig.~\ref{fig-comparison}).
\paragraph{(C2)}
This case is like Case~(c). Since $\bar{K}=0$, we can copy the proof of Case~(c) and show that
the same scaling holds as in the average environment.
\paragraph{(C3)}
This case is like Case~(d). Since $\bar{K}=0$, we can copy the proof of Case~(d) and show that
the same scaling holds as in the average environment.
\end{proof}
\section{Identification of the universality classes of cluster formation}
\label{s.clustering}
In this section we prove Theorem~\ref{T.cluform}. In Section~\ref{s.size} we deal with cases
(a), (A) and (b), (B), (C1), in Section~\ref{s.order} with cases (c), (C2) and (d), (C3).
The strategy of proof is the same as for the homogeneous environment, except at a few points
where the random environment comes into play seriously. We focus on the necessary
modifications. Like Section~\ref{s.mkvrand}, this section is \emph{not completely autonomous}, and for an
understanding of the fine details the reader must check the relevant passages in \cite{GHKK14}.
Before we begin we recall why we may choose the starting configuration to be identically equal
to $\theta$, the mean of the starting configuration. The initial state and the environment of our
Cannings process are such (recall Theorem~\ref{mainth}) that the scaling limit in
\eqref{macroscopic-behaviour} yields on average $\theta$ on level $j+1$.
\subsection{Random cluster size}
\label{s.size}
\begin{proof}[\bf Proof of cases \textup{(b), (B), (C1), (C3)[first subcase]}]
In Step 1 we give the proof for an i.i.d.\ random environment. In Step 2 we extend the proof to
a stationary and ergodic random environment.
\paragraph{Step 1.}
We consider the set $\mathcal{M}_f([0,1]) \times \mathcal{P}(E)$, describing the environment and the state of a block.
If the random environment is i.i.d., then the sequence
\be{ag50}
\Big(\chi^{(\eta,j+1-\alpha)}(\omega), M^{(j)}_{-(j+1-\alpha)}\Big)_{\alpha \in \mathbb{N}_0}
\end{equation}
is a time-inhomogeneous Markov chain. Let $(K^{\ast,(j)}_\alpha)_{\alpha\in\mathbb{N}_0}$ be its sequence of
transition kernels. (We suppress the index $\eta$ from $M^{(j)}_{\eta,-(j+1-\alpha)}$ because its law is
the same for all $\eta\in\Omega_N$.) It suffices to prove three properties:
\begin{itemize}
\item[(1)]
The sequence of transition kernels $(K^{\ast,(j)}_\alpha)_{\alpha\in\mathbb{N}_0}$ converges as $j \to \infty$
to the sequence $(K^{\ast,\infty}_\alpha)_{\alpha\in\mathbb{N}_0}$ of transition kernels given by
\be{ag51}
K^{\ast,\infty}_\alpha((\chi, \theta), \cdot)
= \mathcal{L} \left[\chi^\alpha \otimes \nu_\theta^{1,\widetilde M, 2 \widetilde K \chi^{\alpha}(\omega)}\right] (\cdot),
\end{equation}
where $\widetilde M,\widetilde K$ are defined in \eqref{wtMwtK} and $(\chi_\alpha(\omega))_{\alpha\in\mathbb{N}_0}$ in
\eqref{regi3}.
\item[(2)] The map
\be{ag52}
((0,\infty) \times (0,\infty) \times \mathcal{P}([0,1])) \ni
(c,d,\Lambda) \mapsto \nu^{c,d, \Lambda}_\theta \in \mathcal{P}(\mathcal{P}(E))
\end{equation}
is continuous.
\item[(3)] The map
\be{ag53}
\mathcal{P}(E) \ni \theta \mapsto \nu^{c, d, \Lambda}_\theta \in \mathcal{P}(\mathcal{P}(E))
\end{equation}
is continuous.
\end{itemize}
Items (1) and (3) imply the convergence of the process in \eqref{ag50}, while item (2) is needed in
the proof item (1).
\begin{proof}[Proof of \textup{(1)--(3)}]\label{pr.1-3}
Here a key is the {\em duality relation} for the McKean-Vlasov limit process. This duality arises as
a special case of our duality relation by choosing a suitable geographic space. This coalescent is
obtained by taking as space $\{0,\ast\}$, where the rates for all transitions in $ \ast $ are zero (cemetery)
state and jumps occur from $0$ to $\ast$ at rate $c$. Kingman coalescence occurs at rate $d$ and
the $\Lambda$-coalescence is given via $\Lambda$ (all as long as we are in $0$). For a detailed discussion,
see \cite[Section 4]{GHKK14}.
With the help of duality we can identify the equilibrium measure $\nu_\theta^{c,d,K\chi}$ by using a
measure-determining sequence of test functions. The parameters $c,d,\chi $ enter via the rate of
jump to the cemetery state (parameter $c$), the rate of pairwise coalescence (parameter $d$),
and the rate of coalescence (parameter $\chi$). In the latter, the ratio $\chi/\chi((0,1])$ determines
the probability for partition elements to coalesce in groups ($\Lambda$-coalescence). In this
equilibrium representation, the coalescent has run for {\em infinite} time.
\medskip\noindent
(1) With $\mathcal{L}$ acting on $ \chi_{j+1-\alpha}(\omega)$, we have
\begin{equation}\label{e3052}
K^{\ast,(j)}_\alpha((\chi,\theta),\cdot) = \mathcal{L} \left[ (\chi^{j+1-\alpha}) \otimes \nu_\theta^{1,d_{j+1-\alpha}/c_{j+1-\alpha}, 2K_{j+1-\alpha} \chi^{j+1-\alpha}(\omega)}\right] (\cdot).
\end{equation}
From Theorems~\eqref{T.scaleFVpol} and \eqref{T.scaleFVexp}, we know that $d_{j+1-\alpha}
/c_{j+1-\alpha}$ and $K_{j+1-\alpha}$ converge to $\widetilde{M}$ and $\widetilde{K}$ as
$j\to\infty$. The point is to show for every $\alpha\in\mathbb{N}_0$ the equilibrium measure in the
right-hand side converges as $j\to\infty$. By the stationarity of the random environment, the
law of $\chi_{j+1-\alpha}(\omega)$ is independent of $j$. Hence (2) and (3) yield the claim.
\medskip\noindent
(2) The continuity in \eqref{ag52} can be deduced from the dual representation in the McKean-Vlasov
limit dynamic, in particular, from the fact that the coalescent has run for infinite time, and depends
continuously on the migration rate $c$ and the Kingman coalescence rate $d$, respectively, the
rates for the $ \Lambda$-coalescence. The coalescent has a monotone
decreasing number of partition elements off the cemetery where all rates are zero and reaches
the cemetery state after a finite time.
This means we have a Markov chain hitting a trap in finite time and therefore depends continuously on the finitely many involved jump-rates.
\medskip\noindent
(3) The continuity in \eqref{ag53} is deduced from the dual representation. We have to show that
the dual expectation depends continuously on $\theta$, which goes as follows. First note that the
monomials $ \{\,\langle\cdot\,,f\rangle^\ell\colon\,f \in C_b (E,\mathbb{R}),\,\ell \in\mathbb{N}\}$ are measure-determining
on $(E,\mathcal{B})$. The dual expectation is a finite sum over terms arising from partition elements that
are coalescing before jumping to the cemetery state. If $\ell$ partition elements remain, then the
$\theta$-dependence is via $\langle \theta,f \rangle^\ell$, which is a continuous function of $\theta$.
\end{proof}
\paragraph{Step 2.}
To deal with a stationary and ergodic random environment, we {\em condition} on the sequence
$(\chi_\alpha)_{\alpha\in\mathbb{N}_0}$. This leads to a sequence of Markov chains in random environment,
indexed by $j$, for which the result in (1) holds, as explained above. After that we argue that
(1)--(3) again imply the claim, because of the stationarity and the fact that we need only consider finite
$\alpha$.
Next, we consider the finite-dimensional laws of the Markov chain in random environment conditional
on $(\chi_\alpha)_{\alpha\in\mathbb{N}_0}$ and we verify the appropriate versions of (1)--(3). To this end,
we extend the duality to a {\em space-time duality} and obtain an expression for the {\em mixed
space-time moments} in terms of triples of parameters
\begin{equation}\label{e3076}
{(c_k, d_k, \chi_k)}_{k = j+1, j, \ldots , j+1-L},
\end{equation}
with $L$ being the order of the marginal distribution we consider.
In the {\em space-time dual}, we work with frozen partition elements which are activated (then
once and forever) at a present time. Namely we add partition elements marked by a label in
$[0,\infty]$, which indicates from which time on the mechanisms of the coalescent are activated.
Before this time, the partition element neither moves nor coalesces. This allows us to characterize
the finite-dimensional marginals of the forward process. Suppose that we want to study the
finite-dimensional distributions associated with times $0 \leq t_1 < t_2 < t_3 \ldots < t_n<t$. Then
we take individuals marked with $0,t-t_n,t-t_{n-1}, \ldots, t-t_1$, consider the test functions in the
duality relations for the time horizon $t_1,t_2, \ldots, t_n,t$, and form the product. The duality
relation holds again. Compare with Greven, Sun and Winter~\cite[Corollary 1.20]{GSW}.
In this setting, (1)--(3) turn into claims about the expectation of the duality expression under the
law of the space-time coalescent, after which the argument proceeds as above.
\end{proof}
\begin{proof}[\bf Proof of cases \textup{(a), (A)}]\label{pr.ofcasesaA}
The limiting transition kernel of the rescaled interaction chain for a given environment degenerates
to a transition kernel concentrated on the traps. We have
\begin{equation}\label{e3086}
K = \infty, \qquad \lim_{k\to\infty} d_k/c_k = 1.
\end{equation}
We must therefore show that
\begin{equation}\label{e3090}
\lim_{K\to\infty} \mathcal{L}\left[\nu_\theta^{1,1,2 K\chi}\right]
= \mathcal{L}\left[\int_E \theta (\mathrm{d} u)\delta_{\delta_u}\right].
\end{equation}
Taking the dual representation, we see that as $K\to\infty$ the rate of the $\Lambda$-coalescence
tends to infinity, implying that the coalescent converges before it jumps, and coalesces into a single
partition element. The duality relation says that the original McKean-Vlasov process is in a
mono-type equilibrium, where the type is chosen at random according to $\theta$. The claim now
follows because for $K(\theta,\cdot) = \int_E \theta (\mathrm{d} u) \delta_{\delta_u}(\cdot)$ the state
$\delta_u$ is a {\em trap}, so that the limiting Markov chain is constant for every $\alpha \neq 0$,
the constant being chosen according to $\theta$ for every realization of the random environment.
\end{proof}
\subsection{Random cluster order}
\label{s.order}
\begin{proof}[\bf Proof of cases \textup{(c), (C2)} and \textup{(d), (C3)[second subcase]}]
In cases (c) and (d), averaging takes place via a law of large numbers and the situation is similar
to the homogeneous environment, for which the results in Theorem~\ref{T.cluform} are of the same
type, and it is only the formula for $d$ that changes.
The claim is that the interaction chain, which is a space-time rescaled Markov chain and a
measure-valued square-integrable martingale, converges to a limit that is a measure-valued
diffusion and a square-integrable martingale. In \cite[Section 6(b)]{DGV95}, it was pointed out
how, for the case of the Fleming-Viot process, this convergence reduces to the study of the
process of conditioned variances along the path, which in turn reduces to showing the following
asymptotic relations for these objects. Pick $\alpha_1,\alpha_2 \in I$ with $\alpha_2<\alpha_1$,
and suppose that $\lim_{j\to\infty} k_\alpha(j)/j=\beta(\alpha)$ with $0 \leq \beta(\alpha_1)
< \beta(\alpha_2) \leq 1$. If the scaled Markov chain is such that
\begin{equation}
\label{Mjscalclustalt}
\lim_{j\to\infty} \var\left(\left\langle f,M^{(j)}_{-k_{\alpha_1}(j)}\right\rangle ~\Big|~
M^{(j)}_{-k_{\alpha_2}(j)}=\theta\right)
= \frac{\beta(\alpha_2) - \beta(\alpha_1)}{\beta(\alpha_2)}\,\var_\theta(f), \qquad f \in C_b(E,\mathbb{R}),
\end{equation}
with $\beta(\alpha)=1-\alpha$, then by applying the transformation $\beta(\alpha) = \mathrm{e}^{-s}$
the right-hand side turns into the expression $(1-e^{-(s_1-s_2)})\,\var_\theta(f)$. Since
this scales like $(s_1-s_2)\,\var_\theta(f)$ for $s_1 \downarrow s_2$, we see that the standard
Fleming-Viot process $Y(s)_{s \geq 0}$ appears as the scaling limit. Since $s=\log(1/(1-\alpha))$,
we get the time-scaled Fleming-Viot process $Y(\log(1/(1-\alpha))_{\alpha \in [0,1)}$ (see
\cite[Section 6]{DGV95}).
With suitable time transformations, we can also handle the other forms of scaling $j \to k_\alpha(j)$
in Definition~\ref{def:clusuniv}. Namely, we have to identify the function $F(\alpha_1,\alpha_2)$
appearing in front of $\var_\theta(f)$ and find the transformation $\alpha = L(s)$ such that
\begin{equation}
\label{delF}
\Delta F(s_2) = \lim_{s_1 \downarrow s_2} \frac{F(L(s_1),L(s_2))}{s_1-s_2} \equiv 1,
\end{equation}
so that again the standard Fleming-Viot process $(Y(s))_{s \geq 0}$ appears as the scaling limit.
Since $s=L^{-1}(\alpha)$, we get the time-scaled Fleming-Viot process $(Y(L^{-1}(\alpha))_{\alpha \in I}$.
It was pointed out in \cite[Section 9.3]{GHKK14} how \eqref{Mjscalclustalt} is established for the
homogeneous hierarchical Cannings process by using the scaling analysis of the coefficients
$\underline{d} = (d_k)_{k \in \mathbb{N}_0}$. In our case, we need to work with a random sequence
$(\mu_k\rho_k(\omega))_{k \in \mathbb{N}_0} $ instead of $(\mu_k)_{k \in \mathbb{N}_0}$, where $\rho_k(\omega)$
arises from the term $\Lambda=\Lambda^{(\eta,k)}((0,1])(\omega)$ in the following variance formula
\begin{align}\label{a3103}
\int_E \mu_\theta^{c,d,\Lambda}(\mathrm{d} x)\big(\langle f,x\rangle^2-\langle f,\theta\rangle^2\big)
= \frac{2c}{2c + \lambda\rho_k(\omega) + 2d}\,\var_\theta(\langle f,x\rangle), \qquad f \in C_b(E,\mathbb{R}).
\end{align}
We thus have to see whether the product (with $\rho_k(\omega) = \rho^{\mathrm{MC}_k(0)}(\omega)$)
\begin{equation}\label{e3108}
\prod_{k=j_1}^{j_2} \frac{2c_k}{2c_k+\lambda_k\rho_k(\omega)+2d_k},
\end{equation}
appearing in the expression for the variance in \eqref{Mjscalclustalt}, does indeed exhibit averaging
based on the tail triviality of the random sequence $(\rho_k(\omega))_{k \in \mathbb{N}_0}$ (see
\cite[Eq.~(8.14)]{GHKK14}).
To that end, we abbreviate
\begin{equation}
\label{e3116}
m_k(\omega)=\frac{\mu_k \rho_k(\omega)+d_k}{c_k},
\end{equation}
consider the relation
\begin{equation}
\label{e2858}
\var\left( \left\langle M^{(j_2)}_{j_1},f\right\rangle \mid M^{(j_2)}_{j_2+1} = \theta\right)
= \left[\sum_{k=j_1}^{j_2} \frac{d_{k+1}}{c_k} \prod^{j_2}_{l=k+1}
\frac{1}{1+m_l(\omega)}\right]\, \var(\langle \theta,f \rangle)
\end{equation}
and analyse its behaviour as $j\to\infty$ for appropriate choices of $j_1=j_1(j)$ and $j_2=j_2(j)$.
We must show that, for $\mathbb{P}$ almost all $\omega$, \eqref{e2858} behave asymptotically like the right-hand
side of \eqref{Mjscalclustalt}, and we must identify the associated $F$, $\Delta F$ and $L$.
In order to decide how the product scales as $j_2-j_1\to\infty$, we take logarithms to turn this
into the question whether the sum
\begin{equation}
\label{e3128}
\sum_{k=j_1}^{j_2} \frac{\mu_k\rho_k(\omega)+d_k}{c_k}
= \sum^{j_2}_{k=j_1} m_k(\omega)
\end{equation}
has a certain scaling behaviour, and we link this to the scaling behaviour of $\mu_k/c_k$ and
$d_k/c_k$ for $k\to\infty$ (which we know from Theorems~\ref{T.scaleFVpol} and \ref{T.scaleFVexp})
to derive the relevant asymptotics. We have to show that this asymptotics does \emph{not}
depend on $\omega$ and is equal to that with $\rho_k(\omega)$ replaced by its mean $1$. To
achieve the latter, we use the stationarity of $(\rho_k(\omega))_{k \in \mathbb{N}_0}$, plus the fact that it
has bounded and decaying covariances (recall \eqref{Aprop}--\eqref{fdh:tail}). The key is the
following lemma.
\begin{lemma}
\label{lem:wlln}
Define $S(j_1,j_2)(\omega) = \sum_{k=j_1}^{j_2} m_k(\omega)$. Then,
\begin{equation}
\lim_{j_2-j_1\to\infty} \frac{S(j_1,j_2)(\omega)}{\mathbb{E}[S(j_1,j_2)(\omega)]} = 1
\quad \text{ in $\mathbb{P}$-probability}.
\end{equation}
\end{lemma}
\begin{proof}
Define
\begin{equation}
\chi_k(j_1,j_2) = \frac{\mu_k/c_k}{\sum_{k=j_1}^{j_2} (\mu_k+d_k)/c_k},
\qquad j_1 \leq k \leq j_2.
\end{equation}
Then
\begin{equation}
\frac{S(j_1,j_2)(\omega)}{\mathbb{E}[S(j_1,j_2)(\omega)]} - 1 = \sum_{k=j_1}^{j_2}
\chi_k(j_1,j_2) [\rho_k(\omega)-1].
\end{equation}
With the help of Chebyshev's inequality we see that it suffices to show that
\begin{equation}
\lim_{j\to\infty} \sum_{k=j_1}^{j_2} \sum_{l=j_1}^{j_2} \chi_k(j_1,j_2) \chi_l(j_1,j_2)\,
\mathbb{C}\mathrm{ov}[\rho_k(\omega),\rho_l(\omega)] = 0.
\end{equation}
We have $\mathbb{C}\mathrm{ov}[\rho_k(\omega),\rho_l(\omega)] = C_{|k-l|}$ with
$\lim_{m\to\infty} C_m = 0$. Since, by our assumptions on $(c_k)_{k\in\mathbb{N}_0}$ and
$(\mu_k)_{k\in\mathbb{N}_0}$, we have
\begin{equation}
\lim_{j_2-j_1\to\infty} \max_{j_1 \leq k \leq j_2} \chi_k(j_1,j_2) = 0, \qquad
\sum_{j_1 \leq k \leq j_2} \chi_k(j_1,j_2) \leq 1,
\end{equation}
the claim follows.
\end{proof}
\begin{remark}
\label{rem:homclus}
{\rm The role of Lemma~\ref{lem:wlln} is to show that the same clustering behaviour occurs
in the random environment as in the homogeneous environment. We are only able to prove
convergence in $\mathbb{P}$-probability and not $\mathbb{P}$-a.s. In the prefactor in the
right-hand side of \eqref{e2858} weighted averages over $j_1,j_2$-dependent sliding windows
of the random environment appear, which would need to be shown to converge $\mathbb{P}$-a.s.
It is unclear how to do this, even for an i.i.d.\ random environment.}
\end{remark}
Lemma~\ref{lem:wlln} implies that the term between square brackets in \eqref{e2858} scales like
\begin{equation}
\label{FdHscal1}
\Delta(j_1,j_2) = \sum_{k=j_1}^{j_2} \frac{d_{k+1}}{c_k}
\exp\left[-\sum_{l=k+1}^{j_2} \frac{\mu_l+d_l}{c_l}\right]
\qquad \text{ in $\mathbb{P}$-probability as $j_2-j_1\to\infty$},
\end{equation}
where we use that $\lim_{l\to\infty} (\mu_l+d_l)/c_l = 0$ in all cases of interest. In the remainder of
the proof, we pick $j_1=k_{\alpha_1}(j)$ and $j_2=k_{\alpha_2}(j)$ with $\alpha_2<\alpha_1$, with
$k_\alpha(j)$ as in Definition~\ref{def:clusuniv}, and compute the limit of \eqref{FdHscal1} as
$j\to\infty$. We omit writing $\lfloor\cdot\rfloor$ at places where labels are obviously integer. We
determine $k_\alpha$ and identify $F$, $L$ (recall the discussion leading up to \eqref{delF}) for the
different cases, in the order (c), (C2), (d), (C3). Recall that $K_k= \frac{\mu_k}{c_k}$ and $\bar{K}_k
= \frac{\bar{\mu}_k}{\bar{c}_k}$.
\medskip\noindent
{\bf Case (c).}
Pick $k_\alpha(j)= j+1-\alpha h(j)$ with $h(j)=1/\sqrt{K_j}$, and insert $d_k \sim \sqrt{c_k\mu_k}
= c_k \sqrt{K_k}$ and $d_{k+1} \sim d_k$, to obtain that \eqref{FdHscal1} scales like
\begin{equation}
\label{FdHscal2}
\Delta(j) = \sum_{k= j+1-\alpha_1/\sqrt{K_j}}^{j+1-\alpha_2/\sqrt{K_j}}
\sqrt{K_k}\,\,\exp\left[-\sum_{l=k+1}^{j+1-\alpha_2/\sqrt{K_j}} \big(K_l + \sqrt{K_l}\big) \right].
\end{equation}
Putting $x = (j+1-k)\sqrt{K_j}$, and using that $\lim_{k\to\infty} K_k = 0$, $\lim_{k\to\infty} k^2K_k
= \infty$ and $K_k \sim K_l \sim K_j$ uniformly in $k,l$ in both sums, we get
\begin{equation}
\lim_{j\to\infty} \Delta(j) = \int_{\alpha_2}^{\alpha_1} \mathrm{d} x\,\,\exp[-(x-\alpha_2)]
= 1 - \exp[-(\alpha_1-\alpha_2)].
\end{equation}
Pick $\alpha = L(s) = s$. Then $\Delta F \equiv 1$. Since $s=L^{-1}(\alpha) = \alpha$, this proves
the claim.
\medskip\noindent
{\bf Case (C2)[subcase $\lim_{k\to\infty} k \bar{K}_k = \infty$].}
Pick $k_\alpha(j)=j+1-\alpha h(j)$ with $h(j)=1/\bar{K}_j$, and insert $d_k \sim \mu_k/(\mu-1)
= \bar{K}_k c_k/(\mu-1)$ and $d_{k+1} \sim \mu d_k$, to obtain that \eqref{FdHscal1} scales like
\begin{equation}
\label{FdHscal3}
\Delta(j) = \frac{\mu}{\mu-1} \sum_{k=j+1-\alpha_1/\bar{K}_j}^{j+1-\alpha_2/\bar{K}_j}
\bar{K}_k\,\,\exp\left[- \frac{\mu}{\mu-1} \sum_{l=k+1}^{j+1-\alpha_2/\bar{K}_j} \bar{K}_l \right].
\end{equation}
Putting $x = (j+1-k)\bar{K}_j$, and using that $\lim_{k\to\infty} \bar{K}_k = 0$, $\lim_{k\to\infty}
k\bar{K}_k = \infty$ and $\bar{K}_k \sim \bar{K}_l \sim \bar{K}_j$ uniformly in $k,l$ in both sums,
we get
\begin{equation}
\label{eq:FdHscal3.5}
\lim_{j\to\infty} \Delta(j) = \frac{\mu}{\mu-1} \int_{\alpha_2}^{\alpha_1} \mathrm{d} x\,\,
\exp\left[- \frac{\mu}{\mu-1}(x-\alpha_2)\right]
= 1 - \exp\left[- \frac{\mu}{\mu-1}(\alpha_1-\alpha_2)\right].
\end{equation}
Pick $\alpha = L(s) = \frac{\mu-1}{\mu}s$. Then $\Delta F \equiv 1$. Since $s=L^{-1}(\alpha)
= \frac{\mu}{\mu-1}\alpha$, this proves the claim.
\medskip\noindent
{\bf Case (d).}
Pick $k_\alpha(j)= (1-\alpha)(j+1)$, and insert $d_k \sim M/\sigma_k$, $\sigma_k c_k \sim k/(1-a)$
and $d_{k+1} \sim d_k$, to obtain that \eqref{FdHscal1} scales like
\begin{equation}
\label{FdHscal4}
\Delta(j) = M(1-a) \sum_{k= (1-\alpha_1)(j+1)}^{(1-\alpha_2)(j+1)}
\frac{1}{k}\,\,\exp\left[-\sum_{l=k+1}^{(1-\alpha_2)(j+1)} \left(K_l + \frac{M(1-a)}{l}\right) \right].
\end{equation}
Putting $x=(j+1-k)/(j+1)$, and using that $\lim_{k\to\infty} k^2K_k = 0$, we get
\begin{equation}
\begin{aligned}
\lim_{j\to\infty} \Delta(j) &= M(1-a) \int_{\alpha_2}^{\alpha_1}
\frac{\mathrm{d} x}{1-x}\,\,\exp\left[-M(1-a) \int_{\alpha_2}^{x} \frac{\mathrm{d} y}{1-y} \right]\\
&= M(1-a)\, (1-\alpha_2)^{-M(1-a)}
\int_{\alpha_2}^{\alpha_1} \mathrm{d} x\,(1-x)^{-1+M(1-a)}
= 1 - \left(\frac{1-\alpha_1}{1-\alpha_2}\right)^{M(1-a)}.
\end{aligned}
\end{equation}
Pick $\alpha = L(s) = 1-e^{-s/R}$ with $R=M(1-a)$. Then $\Delta F \equiv 1$. Since $s= L^{-1}(\alpha)
= \log(1/(1-\alpha)^R)$ we get the claim.
\medskip\noindent
{\bf Case (C2)[subcase $\lim_{k\to\infty} k \bar{K}_k = \bar{N}$].}
This is the same as case (d) with $M(1-a)$ replaced by $\bar{N}\frac{\mu}{\mu-1}$.
\medskip\noindent
{\bf Case (C3)[second subcase].}
This is the same as case (d) with $M$ replaced by 1.
\end{proof}
\addcontentsline{toc}{section}{References}
\bigskip
|
\section{Introduction}
During the last decade, models for the formation and evolution of planetary systems have been developed significantly.
Early stages of planetary system formation have been studied by many authors (e.g. \citealt{Stamatellos09}; \citealt{Meru2010}; \citealt{Fouchet2010}; \citealt{Dodson11}; \citealt{Pinilla12,Pinilla15}; \citealt{Meru2014}; \citealt{Dipierro16}). Observational properties of the early stages of planetary and substellar companion formation and evolution are poorly understood. Authors mainly analyse the possibility to identify the signatures of a planet's formation with the best available techniques, such as ALMA -- Atacama Large Millimeter/submillimeter Array (\citealt{Gonzalez12}; \citealt{Vorobyov13}; \citealt{Zakhozhay13}; \citealt{Dipierro15}; \citealt{Dong16}) and IRAM Plateau de Bure interferometer (e.g. \citealt{Stamatellos11}). Recent ALMA observations of HL~Tau~\citep{ALMA15} and HW~Hya~\citep{Andrews16} very likely show that planetary formation is ongoing in the discs of these systems. The selection of a good target for observations is not a trivial task and it would be very useful to know the indicators that forming companions may leave in other more numerous observational methods, such as spectral energy distributions (SEDs), which are already available for many systems with protoplanetary discs. There are very few publications dedicated to this topic to date. \cite{Vorobyov13} and \cite{Zakhozhay13} show that a presence of hot and massive (32 -- 64~$M_J$) fragments -- proto-brown dwarfs during the first few 10,000 years -- initiate an additional peak in SEDs at 5 -- 10~$\mu$m. \cite{Varniere2006}, considered the example of a 2 Jupiter mass planet on a very close orbit to the star, and found that SEDs from discs with gaps may not only have a reduced emission at $\lambda\sim 5 - 20$~$\mu$m due to the removal of the emission from the gap, but also may have a measurable excess emission at longer wavelengths ($\sim 10 - 100$~$\mu$m), that rises from the heating of the vertical disc wall at the outer edge of the gap. \cite{Pinilla16} show that discs with massive planets lose the near infrared excess after a few Myrs and their SEDs look like true transition discs, while SEDs of the discs with less massive planets look like pre-transition disc SEDs for most of their remaining lifetimes.
\par There is a sufficient number of observational studies that contain large samples of observed SEDs of the systems with protoplanetary discs. For example, a recent paper of \cite{Marel16} contains a large sample of observational SEDs ($\sim$200 systems with transition discs), and $\sim$70$\%$ of these discs, most probably, contain large cavities and holes. These might be cleared by a planetary or a substellar companion or be a result of another physical process, like grain growth (e.g. \citealt{Dullemond05}) or photoevaporative clearing (e.g. \citealt{Alexander06}).
\par In this paper, we concentrate on the gaps cleared by one companion and analyse the possibility to detect signatures of the presence of a substellar or planetary companion in a protoplanetary disc based on SED profiles and also the possibility to determine the companion's parameters and the physical conditions of the disc when the presence of a companion is most evident.
\par This paper is organized as follows. The modeling approach and basic equations are briefly described in Section~\ref{algorithm}. The results and analysis are presented in Section~\ref{results}. Section~\ref{conclusions} contains the conclusions of the paper.
\section{Model assumptions}
\label{algorithm}
To calculate SEDs from the systems with protoplanetary discs and brown dwarf companions, we assume that the total flux from the system ($F$) is the sum of the fluxes from the central star ($F_{\ast}$), the disc ($F_d$) and the companion ($F_{c}$): $F = F_{d} + F_{c} + F_{\ast}$. Fluxes from the star and the companion are calculated using a black-body approximation:
\begin{equation}
F_{\ast} = \frac {\pi R_{\ast}^{2}}{d^2} B_{\nu} (T_{\ast}),
\end{equation}
\begin{equation}
F_{c} = \frac {\pi R_{c}^{2}}{d^2} B_{\nu} (T_{c,ef}),
\end{equation}
where $d$ is a distance to the object, $B_{\nu} (T)$ is the Plank function, $R_{\ast}$ and $T_{\ast}$ are the stellar radius and effective temperature, and $R_{c}$ and $T_{c, ef}$ are the companion's radius and effective temperature. The stellar and companion's physical parameters are taken from \cite{Baraffe15} and \cite{Baraffe98}, respectively, and are summarised in Table~\ref{tlab}.
\subsection{Spectral energy distribution from the disc}
\label{discSED}
\par To calculate the contribution to the SED from the protoplanetary disc, we assume that the companion is moving along a circular orbit and that there is no material inside the gap. The flux from the disc is $F_{d} = F_{disc} - F_{gap}$, where $F_{disc}$ is the flux from the disc, as it would be without the gap opened by the companion, and $F_{gap}$ is a flux from the part of the disc that has been cleared by the companion.
\par We use the solution to the one-dimensional radiative transfer equation to
calculate the protoplanetary disc flux density with the frequency $\nu$:
\begin{equation}
F_{disc} =d^{-2}\int_{R_{in}}^{R_{out}}B_{\nu}(T_r)~Q_{\nu}~2\pi r~{\rm d}r,
\end{equation}
\begin{equation}\label{form3}
Q_{\nu} = 1 - {\rm e}^{-\tau} ; ~~~ \tau = \varSigma_r~\kappa_{\nu},
\end{equation}
where $d$ is a distance to the system from the Sun, $r$ is the radial distance from the central object, $B_{\nu}(T_r)$ is the Planck function for a radius-dependent temperature, $T_r$, of the disc, $R_{in}$ and $R_{out}$ stand for the inner and outer radii of the disc, respectively, $Q_{\nu}$ represents the dust grain emission efficiency, $\tau$ stands for the optical depth of the disc material, which is the product of the wavelength-dependent disc opacity, $\kappa_{\nu}$, and $\varSigma_r$ is the radial surface density distribution of the disc material. To calculate the wavelength-dependent disc opacity, we use the Mie theory and consider compact spherical grains composed of astronomical silicates, a density of 2.5~g$\cdot$cm$^{-3}$, grain sizes between 0.1 and 100~$\mu$m and gas to dust mass ratio of 100.
\par Radial temperature and surface density distributions we assume to be power laws: $T_r \propto r^q$ and $\varSigma_r \propto r^p$. For computations we used $q$ = -0.5 and $p$ = -1.5. The effect of $q$ variation on the resulting SED profile is discussed in details in subsection~\ref{vs_disc}.
\par The disc surface density as a function of the disc mass is described by the equation:
\begin{equation}\label{Sigma_r}
\varSigma_r = \frac {M_d~r^p~(2+p)}{2 \pi~(R_{out}^{2+p} - R_{in}^{2+p})},
\end{equation}
where $M_d$ is a total mass of the disc.
\par The flux from the part of the disc that has been cleared by the companion, $F_{gap}$, is calculated using the same approach as $F_{disc}$, except only in the range $R_{in,gap}$ to $R_{out,gap}$, which are the gap inner and outer radii. These radii are determined by the distance to the star from the companion, $r_c$, and the Hill radius $R_H$:
\begin{equation}
R_H = r_c \Big (\frac{1}{3} \frac{M_c}{M_{\ast} + M_c} \Big)^{\frac{1}{3}},
\end{equation}
where $M_c$ and $M_{\ast}$ are masses of the companion and of the star, respectively.
\par We neglect the additional heating from the companion to the inner and outer edges of the disc gap. As discussed in detail in Appendix~\ref{Additional heat}, the maximum temperature to which the brown dwarf companion could heat the inner and outer sides of the disc gap is comparable, or much smaller, to the disc temperature to which it is been heated by the central star.
\section{Results and analysis}
\label{results}
We simulated synthetic SEDs for a system with a 0.8~$M_{\odot}$ central object and a 30~$M_J$ substellar companion with an age of 5~Myr. The mass of the companion we chose based on the limitation from the previous study of \cite{MaGe2014}, that shows that the maximum mass of the object that can form via gravitational instability in a protostellar disc is $\sim40M_{J}$. Table~\ref{tlab} summarizes stellar and companion physical parameters that were used to compute the SED. The SED for the protoplanetary disc was modeled for $R_{in}$ = $R_{sub}$ (in our particular case $R_{sub}$ = 0.02~AU = 4.7~$R_{\ast}$) and $R_{out}$ = 400~AU. We consider a passive disc and assume there is no material inside the gap, and the companion is moving along circular orbit. We further assume that the width of the gap is one diameter of the Hill sphere, so we consider the minimum possible gap cleared by a companion of a given mass.
\par We don't account for an increase of the emission temperature from the part of the outer edge of the gap that is directly exposed to stellar radiation since its contribution would be very small. For example, if a gap has been cleared by a companion at 1~AU, the total area that is not covered by the inner edge of the gap is $\sim0.07$~AU$^2$ at the distance 1.23~AU. However if the disc has a non-zero inclination, the addition flux from the outer edge of the gap would be more noticeable. In that case, the radiation from this part of the disc would be directly exposed to the observer and then even a small change of the temperature may cause a noticeable change of the flux. This will be studied in a subsequent paper with more detailed disc structure and radiative transfer simulations.
\begin{table*}
\centering
\caption{The physical parameters of the system at 5 Myr$^1$.}
\label{tlab}
\begin{tabular}{@{}cccccccccccccc@{}}
\hline\hline
\multicolumn{4}{c}{Star} && \multicolumn{4}{c}{Companion} &&\\
\hline
$M_{\ast}$ & $T_{\ast}$& $L_{\ast}$& $R_{\ast}$ && $M_{c}$& $T_{c}$& $L_{c}$& $R_{c}$ && $R_{sub}^2$ \\
$(M_{\odot})$ & $(K)$ & $(L_{\ast})$ & $(R_{\odot})$ && $(M_{J})$ & $(K)$ & $(L_{\ast})$ & $(\times 10^{9} cm)$ & & AU \\
\hline
0.8 & 4081 & 0.426 & 1.306 && 30 & 2616 & $5.5 \times 10^{-3}$ & 25.27 && 0.02 \\
\hline\hline
\end{tabular}
\begin{minipage}{150 mm}
$^1$~Stellar parameters are from \citep{Baraffe15}, companion's parameters are from \citep{Baraffe98}
$^2$~$R_{sub}$ is determined using $T_{\ast}$ and $R_{\ast}$ and based on the assumption that disc dust sublimation temperature equals 1500~K, as in \cite{Dullemond01}.\\
\end{minipage}
\end{table*}
\par Figure~\ref{prim} shows SEDs from the system with a companion at 1~AU (black solid line) and without companion (gray solid line). The black dashed and dashed-dotted lines show the fluxes from the inner and outer parts of the disc. The red and green lines show the fluxes from the companion and the star, respectively.
\par Visual examination of Figure~\ref{prim} indicates that the presence of the gap cleared by the substellar companion along its orbital motion in the disc causes an additional minimum at the wavelength interval from 10 to 100~$\mu$m. Table~\ref{dif} contains the difference of the fluxes from the systems with and without the companion, and the absolute values for the fluxes for the each wavelength at the interval 10 -- 160~$\mu$m. The maximum difference $\sim70$~mJy corresponds to 33.8~$\mu$m. In the present study, we have analyzed the fluxes, assuming that the system is located at 250~pc. If the system is located closer to us, then the fluxes would be much more intense and hence the difference would be much more evident. For example, if the system is located at 100~pc, than all the fluxes in Table~\ref{dif} should be multiplied by the factor of 6.25.
\par In the following subsection, we analyse in more detail how the shape of SED profile and the difference between the fluxes with and without the companion depend on some of the companion and disc physical properties.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig1-c.eps}
\end{center}
\caption{
SEDs of the modeled system with protoplanetary disc with an embedded companion (black solid line) that is composed of the inner (black dashed line) and outer (black dashed-dotted line) disc parts, the flux from the star (green line) and the flux of the companion (red line). The gray solid line shows the flux from the system with the same parameters but without a companion. We assume $d = 250$~pc.
}
\label{prim}
\end{figure}
\begin{table}
\centering
\caption{The differences of the fluxes from the system with and without an embedded companion as a function of wavelength.}
\label{dif}
\begin{tabular}{@{}cccc@{}}
\hline\hline
$\lambda$ & $\Delta F$ & $F_{no\ companion}$ & $F_{with\ companion}$ \\
($\mu$m) & (mJy) & (mJy) & (mJy) \\
\hline
10 & 3.3 & 139.1 & 135.8 \\
20 & 46.2 & 288.4 & 242.2 \\
34 $(\Delta F_{max})^1$ & 70.0 & 490.7 & 420.7\\
80 & 36.2 & 1129.2 & 1093.0 \\
100 & 26.7 & 1316.3 & 1289.6 \\
160 & 12.8 & 1411.8 & 1399.0 \\
\hline\hline
\end{tabular}
$^1$ 34 $\mu$m is the wavelength where is the maximum difference
between the fluxes of the system with and without companion.\\
\end{table}
\subsection{Dependence of SED profile from the properties of the protoplanetary discs and the companion.}
\label{protoplanetaryD}
\subsubsection{SED profile shape versus companion's location and mass}
\label{vs_comp}
Figure~\ref{prim} shows the SEDs from the system with an embedded companion in a protoplanetary disc located at 1~AU, because in this case the difference between the fluxes from the systems with and without the companion is the most evident. The distance of 1~AU we determined with the models by varying the distance from the companion to the central star within all possible positions along the disc radius (from 1~AU to 100~AU with steps of 1~AU). Figure~\ref{rp_var} illustrates the dependence of the shape of SED profile from the companion's location and mass, assuming all the parameters, except the one that we are varying, are the same as for the case shown in Figure~\ref{prim}. It is evident that the presence of the companion would be easier to determine if it would be located closer to the star. This is not because of the difference in fluxes, but because the presence of the companion at $<$10~AU causes a noticeable depression in the SED profile. The upper panel of Figure~\ref{rp_var} shows that that the biggest visual difference in intensities between SED profiles of the system with and without the embedded companion is for the system with the companion that is located very close to the inner edge of the disc. While computations show that for the systems with companions located closer to the star, the difference between the systems with and without the companion is smaller. For example, if the companion is located at 0.5~AU, the maximum difference is 49.4~mJy at 23.8~$\mu$m and if it is located at 10~AU, the maximum difference is 221.1~mJy at 106.9~$\mu$m. The largest difference between the fluxes with and without the companion is 266.7~mJy, for a companion at 30~AU from the star. But because this difference is at $\lambda = 154.5$~$\mu$m, where the disc is optically thin, visually it will be impossible to detect the signature of the companion because the SED profile resembling the SED of the system with the similar but shorter disc. Figure~\ref{Mp_short} shows the fluxes from the system with the disc and no companion, $R_{out}$ = 400~AU (gray line), with the same disc that contains a gap cleared by a companion ($M_c = 30$~$M_J$) at 30~AU (black line) and the flux from the shorter disc without companion, $R_{out}$ = 30~AU (red line). In the wavelength interval $\lambda \simeq 6 - 40$~$\mu$m fluxes from the system with the companion at 30~AU and system with shorter disc visually indistinguishable. At longer wavelengths, the flux from the system with a shorter disc is more intense at $\lambda < 100 \mu$m and less intense at $\lambda > 100$~$\mu$m, because the short disc has a significantly smaller amount of cold material comparing to the similar disc of typical size.
\par The lower panel of Figure~\ref{rp_var} illustrates the predictable dependence that more massive companions create deeper minima. The values of $M_c$ are varied from 40~$M_J$ down to 3~$M_J$. For all considered cases, companions are located at 1~AU and that is why the biggest difference between all the fluxes is at 33.8~$\mu$m. The maximum difference is for the system with the companion of 40~$M_J$ is - 76.7~mJy and a smallest difference of 32.8~mJy corresponds to the system with less massive companion of 3~$M_J$.
\par In both panels of figure~\ref{rp_var} at the wavelength $\lambda< 6$~$\mu$m, the spread of the lines is slightly noticeable. This difference is due to the contribution of the flux from the companion and it is non-negligible only for very massive companions with $M_c > 10$~$M_J$. For the companion of 30~$M_J$, it exceeds 1~mJy in the wavelength interval $\lambda \thickapprox 1.43 - 2.76$~$\mu$m, and for the most massive companion considered in this work, 40~$M_J$, in the interval $\lambda \thickapprox 1.04 - 3.95$~$\mu$m. The maximum difference due to the flux from the brown dwarf companions with mass 30~$M_J$ and 40~$M_J$ are 1.14~mJy at $\lambda = 1.94$~$\mu$m and 1.70~mJy at $\lambda = 1.86$~$\mu$m, respectively.
Although, it is important to note that the additional flux from the companion is computed based on the assumption that the companion has the same age as the star and that there is no gas accretion. If the gas from the disc is still accreting onto the companion, the emission from the companion can be noticeably more intense. This will be studied in a future paper with more detailed structure of the disc, the embedded companion, and the gap.
\par The physical parameters for the companions with different masses from \cite{Baraffe98} and $R_H$, that were used to perform the computations are listed in Table~\ref{Mvar} of appendix~\ref{AdditionalSysPar}.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig2-c.eps}
\end{center}
\caption{
SEDs of the systems with companions at different distances from the star ({\it top panel}) and with companions of different masses (bottom panel). In top panel, SEDs of the model system with companion of $M_{c} = 30$~$M_J$ at different distances: 0.5~AU - red line, 1~AU - black line, 3~AU - blue line, 5~AU - green line, 10~AU - orange line, 30~AU - cyan line, 50~AU - magenta line. On bottom panel, SEDs of the modeled system with companion at 1~AU with different masses: 40~$M_J$ - red line, 30~$M_J$ - black line, 20~$M_J$ - blue line, 10~$M_J$ - green line, 7~$M_J$ - orange line, 3~$M_J$ - cyan line. On both panels, modeled SED from the system with the same parameters but without companion is shown with gray line.
}
\label{rp_var}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig3-c.eps}
\end{center}
\caption{
SEDs of the model system with protoplanetary disc with embedded brown dwarf at $r_c =$ 30~AU, $R_{out}$ = 400~AU (black line), SED of the analogous system but without companion (gray line) and from the system without companion and smaller $R_{out}$ = 30~AU (red line).
}
\label{Mp_short}
\end{figure}
\subsubsection{SED profile shape versus disc inner radius and temperature profile}
\label{vs_disc}
In \cite{Woitke16}, the authors showed the effect of different dust and disc parameters on the shape of model SED profile. In the present paper, we consider only two disc parameters: the location of the disc inner radius and a power law of the disc temperature profile. The top panel of Figure~\ref{Rin_var} shows SEDs from the systems with and without companions that have $R_{in}$ = 0.02, 0.1 and 0.5~AU (all another parameters are the same as in Figure~\ref{prim}). The value of $R_{in}$ = 0.02~AU corresponds to the dust sublimation radius \citep{Dullemond01}. The variation of $R_{in}$ doesn't change the difference of the fluxes for the systems with and without companions, or the wavelength of the maximum difference. Although the upper panel of Figure~\ref{Rin_var} indicates that if a companion's orbit is very close to the disc inner radius, like in the case of $R_{in}$ = 0.5~AU (blue lines), then the flux from the inner part of the disc is quite faint and the total SED from the system with a companion doesn't have a double peak profile. That is why the SED from such systems is very similar to the SED from the system with a slightly larger inner radius. On the other hand, for young discs with very small inner holes (with the values close to the sublimation radius), emission from the inner part of the disc (before the gap) will dominate at the wavelength interval $\lambda = 10 - 20$~$\mu$m. In the present paper, we consider quite dense discs that are optically thick at all the wavelengths up to the distances of few 10 of~AU, but if the inner disc would be optically thin then its emission at this wavelength interval would be determined by the chemical composition of the disc.
\par The lower panel of Figure~\ref{Rin_var} illustrates how the SEDs of the systems with and without companions depend on the temperature distribution profile power law $q$ ($T_r \propto r^q$). Obviously SED profiles strongly depend on the $q$ value. Although the difference between the fluxes from the systems with and without companions still corresponds to the wavelength interval $\lambda = 10 - 100$~$\mu$m.
If the temperature profile is slightly shallower ($q$ = -0.47), then the disc would be hotter around the region of the gap and the maximum difference between the fluxes from the systems with and without companions will be larger, 110.6~mJy, and will correspond to a shorter wavelength $\lambda = 28.9$~$\mu$m. If the disc has a steeper temperature profile ($q$ = -0.55), then maximum difference from the fluxes would be 32.6~mJy at $\lambda = 43.3$~$\mu$m. Interestingly, the latter difference is comparable with the maximum difference from the system with a companion of 3~$M_J$ (32.8~mJy at $\lambda = 33.8$~$\mu$m), considered above and graphically presented in the lower panel of Figure~\ref{rp_var}. It additionally confirms the importance of the disc parameter determination, which would require a big observation data sample, corresponding to the wide wavelength interval.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig4-c.eps}
\end{center}
\caption{
SEDs of the model system with an embedded brown dwarf at 1~AU with different $R_{in}$ ({\it top panel}): 0.02~AU (dust sublimation radius) -- black, 0.1~AU -- red and 0.5~AU -- blue dashed lines.
{\it Bottom panel} shows SEDs of the models with protoplanetary discs that have different temperature distribution indices: $q$ = -0.47 (blue), $q$ = -0.5 (black), $q$ = -0.55 (red). On both panels, SEDs from the system with a companion ($M_{c} = 30$~$M_J$, $r_c$ = 1~AU) are shown with dashed lines and solid lines show the fluxes from corresponding discs without companions.
}
\label{Rin_var}
\end{figure}
\subsection{Comparison with another stages of disc evolution}
\label{parameters}
In this subsection, we compare the companion's formation signatures in SED profiles of the systems with protoplanetary and protostellar discs and compare our modelling results to those obtained before in \cite{Vorobyov13}.
These models were based on numerical hydrodynamics simulations of \cite{Vorobyov10}, who studied the formation and evolution of protostellar discs subject to gravitational instability and fragmentation. The basic equations of hydrodynamics were solved on a polar grid in the thin-disc limit. That allowed to follow the gravitational collapse of a pre-stellar condensation (core) into the star plus disc formation stage and further to the T~Tauri stage when most of the parental core has accreted onto the burgeoning disc. The following physical processes were taken into account: disc self-gravity via solution of the Poisson integral and disc viscosity via $\alpha$-parameterization, radiative cooling from the disc suface, stellar and background irradiation, and also viscous and shock heating (for more details see \citealt{Vorobyov10}). In \cite{Vorobyov13}, the authors considered four moments in time: these were 0.09~Myr, 0.1~Myr, 1.1~Myr and 1.3~Myr since the formation of the central protostar. We choose an age of 0.1~Myr based on the similarity with the masses of the companions considered in the present paper.
\par Figure~\ref{age_var} shows SEDs of a system with the protostar at 0.1~Myr (top panel, \citealt{Vorobyov13}) and with the star at 5~Myr (bottom panel) with surrounding discs and forming companions. The hot and most massive (32~$M_{J}$) fragment at 0.1~Myr is located at a radial distance of 108~AU. The embedded companion (30~$M_{J}$) at 5~Myr is located at 1~AU, because at this distance, the difference between the SED profiles of the system with and without the embedded companion is much more evident. In both panels, the total fluxes from the systems are shown with black lines, the fluxes from the star (5~Myr) or protostar (0.1~Myr) are shown with green lines, and the blue line shows the flux from the disc together with the companion. A slightly noticeable peak around 1~$\mu m$ from the disc at 5~Myr is due to the direct emission from the brown dwarf. This flux does not account for possible additional flux due to the gas accreation from the disc.
\par As studied in details in \cite{Vorobyov13}, the formation of a hot and massive fragment in the first 100k years causes an additional peak to be present at 5 -- 10~$\mu$m. The mid-plane temperature of the fragment at that moment in time, presented in Figure~\ref{age_var}, is 1180~K and the surface temperature is 380~K. Such mid-plane temperatures are too cold to form a companion, and that is why it is impossible to make a direct comparison between the different stages of the brown dwarf formation. An important conclusion that we can make by analysing Figure~\ref{age_var} is that SEDs of fragmenting discs and of the older disc with an embedded companion may have a similar double peak profile, caused by different physical processes: an additional peak due to the presence of a very hot clump in the protostellar disc and an addition minima due to the gap cleared by a companion in the protoplanetary disc. Figure~\ref{Tsdens} shows azimuthally averaged surface densities (top panel) and temperatures (bottom panel) as functions of the disc radial distance for the considered protostellar (red lines) and protoplanetary (black line) discs on both panels. The upper panel of Figure~\ref{Tsdens} illustrates the drop of surface density down to 0 from 0.77 to 0.23~AU for protoplanetary disc and an addition peak around 108~AU for protostellar discs - both these features are related to the double peak profiles of SEDs, that we see in figure~\ref{age_var}.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig5-c.eps}
\end{center}
\caption{SED of modeled systems with protostellar ({\it top panel} \citealt{Vorobyov13}) and protoplanetary ({\it bottom panel}) discs. In both panels, the flux from the star (protostar, at 0.1~Myr) is shown with green lines, the fluxes from the disc (including the flux from companion or a fragment, at 0.1~Myr) is shown with the blue line and the black line is a total flux from the system. On the bottom plot, flux from the corresponding system but without a companion is shown by the gray line. All discs have a face-on orientation and are located at 250~pc. The system's physical and geometrical parameters are described in the text and in Table~\ref{tlab}.
}
\label{age_var}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig6-c.eps}
\end{center}
\caption{
Azimuthally averaged surface density ({\it top panel}) and temperature ({\it bottom panel}). The red line shows surface density and temperature of the protostellar disc (as it was previously presented in \citealt{Vorobyov13}) and the black line is for the protoplanetary disc.
}
\label{Tsdens}
\end{figure}
\par The results obtained previously in \citep{Zakhozhay15} for a debris disc with a $40$~$M_{J}$ companion (the closest mass of the object, available in the \cite{Baraffe98} isochrones to the maximum mass of the object that could form in circumstellar disc \cite{MaGe2014}) indicate that the maximum difference between SEDs of the systems with and without a companion at $\approx 75$~$\mu$m is $\sim 10$~mJy (assuming $d$ = 250~pc) which is comparable with the sensitivity of the instrument PACS of the space telescope HERSCHEL in this wavelength regime. Although, in that paper, only one critical case was considered, assuming that the width for the gap is the diameter of one Hill sphere. This assumption is acceptable for the early stages of the disc evolution, like in a present paper, but for older debris discs, the gap cleared by the same companion should be wider. In debris discs, the width of the gap is determined by gravitational perturbations from the companion. Planetesimals entering the region around the companion's orbit (known as the "chaotic zone") are scattered onto highly eccentric orbits, creating an underdensity of the material (\citealt{Chirikov1979}; \citealt{Wisdom1980}). Recently, \citet{Nesvold2015} showed that collisions between planetesimals additionally widen the gap. In this case, the gap width would be at least 2-3 times wider than the one determined by the Hill radius. In a future investigation, we intend to extend our study, performing a similar analysis for debris discs with a wide range of different physical parameters.
\section{Conclusions}
\label{conclusions}
\par In this work, we present a study of SED profiles from a system with a protoplanetary disc that contains an embedded companion. We simulate the SEDs using a simplified flux computation approach and analyzed the effect of the most important disc and companion parameters, that affect the resulting profiles the most. We consider a system with a low mass star (0.8~$M_{\odot}$) and a brown dwarf companion. We find that the gap cleaned by the companion creates an additional depression in the SED profile at $\lambda \sim 10 - 100$~$\mu$m. The drop of the flux intensity strongly depends on the companion's mass and location:
1) a more massive companion initiates the deeper minimum;
2) the difference between the profiles is noticeable only if the companion is close enough (within $\sim$10~AU) to the star.
\par The analysis presented in subsection~\ref{protoplanetaryD} proves the importance of the disc parameters. It is shown that the maximum difference between the fluxes from the disc with and without a companion of 3~$M_J$ is similar to the difference between the systems with and without a companion of 30~$M_J$ but with a slightly steeper temperature profile ($q$ = -0.55). These maximum differences correspond to 33.4~$\mu$m and 43.3~$\mu$m, for 3~$M_J$ ($q$ = -0.50) and 30~$M_J$ ($q$ = -0.55), respectively, while the maximum flux difference is the same $\sim$33~mJy. Computations also indicate that if the disc inner radius is $>$0.5~AU, then the disc SED doesn't have a double peak profile. Thus the results of this paper are mainly related to the early stages of disc evolution when the companion just started to form a gap. On the transitional stage of the disc evolution the inner disc hole, as well as a gap width will be significantly larger and, hence, the evidence of the gap in the disc will be more evident in the SED profile (assuming that the embedded companion is at the same distance). Although at this stage it will be much harder to make a connection between the width of the gap and the mass of the companion, because at longer time intervals more physical processes may affect it.
\par We also made a comparison with SEDs from the protostellar disc with a hot fragment. Hot and massive fragments in a protostellar disc cause additional peaks in SED profile of the system. The intensity and wavelength regimes of the emission from the fragments depend on their temperature, density and spatial sizes. If the fragment is not very hot, like in \cite{Vorobyov13}, it will initiate an additional peak at 5 -- 10~$\mu$m that makes a total SED profile from the disc quite similar to what we have obtained in the present work for a protoplanetary disc with an embedded companion. Future modellings of SEDs from protostellar discs will allow us to determine the criteria that will help to distinguish these two cases. Although it is obvious that the most important is an age determination, because massive fragments have their very hot temperature and huge spacial sizes only at the very beginning of the disc formation.
\par In the subsequent paper we intent to perform a more detailed analysis by using a more precise disc model and computing the disc temperature with the radiative transfer simulations. We will discuss in more detail how the disc chemical composition and vertical structure may affect the SED of the disc with a gap caused by presence of companions.
\begin{acknowledgements}
The author thanks the anonymous referee for useful suggestions that improved the manuscript. The author thanks F.~M\'{e}nard, I.~Kamp, A.~Carmona, D.~Asmus, P.~Pinilla, N.~van~der~Marel, C.~Dullemond, C.~Johnstone and P.~Berczik for useful discussions and Y.~Boehler for providing the dust opacities. Author also thanks the head of the research program $"$Fundamental properties of the material in the wide scale interval of the space and time Branch of Physics and Astronomy, National Academy of Sciences of Ukraine$"$, in the framework of which a part of this work was carried out.
\end{acknowledgements}
\begin{appendix}
\section{Additional heat of the gap rims from the brown dwarf companion}
\label{Additional heat}
\par To estimate the additional heat that gets to the inner and outer edges of the gap of the disc from the substellar companion, let us consider the configuration presented in Figure~\ref{A1}. The temperature of every point of the gap edge is
\begin{equation}
T^4 = T^4_{rim} + T^4_{c,irr},
\end{equation}
where $T_{rim}$ is the temperature of the gap edge, which is heated by energy from the star that is absorbed and re-emitted by disc material, and $T_{c,irr}$ is the temperature to which the companion heats the edge of the gap, which is given by
\begin{equation}
T_{c,irr} = \sqrt [4] {\frac{L_c}{16 \pi \sigma r_{c,g}^2}},
\end{equation}
where $L_c$ is the luminosity of the companion and $r_{c,g}$ is the distance from the companion to the considered point of the gap edge. At the regions where the edge of the gap doesn't get the direct emission from the companion, $T = T_{rim}$.
\par The minimum $r_{c,g}$ is equal to $R_H$, the point where the line connecting the star and the companion is perpendicular to the tangent to the circle which describes the inner and outer edges of the gap. In our particular case ($M_{\ast} = 0.8 M_{\odot}$, $M_c = 30$~$M_J$ and $r_c$ = 1~AU), it is 0.23~AU (at Figure~\ref{A1} points $X_1$ and $Y_1$ are shown for the inner and outer edges, respectively). In Figure~\ref{A1}, $S$ stands for the star and $C$ for the companion. Additionally, let us consider two more points of the inner and outer edges.
\par $X_3$ is the furthermost point at the inner edge of the gap that is heated by the companion. This point is the intersection of the normal drawn from the point C to the circle describing the inner edge of the cavity with this circle. To determine the distance from the companion to this point, let us consider the right triangle $\bigtriangleup SX_3C$. The angle $SX_3C$ is right, $X_3S = r_c - R_H$, $CS = R_H$, and hence
\begin{equation}
CX_3 = \sqrt{2 r_c R_H - R_H^2}.
\end{equation}
In our particular case, $CX_3 =0.64$ AU. $X_2$ is the middle point and the distance to it from the companion is $X_2C = (X_3C+X_1C)/2 = 0.435$ AU. $Y_3$ and $Y_2$ are equidistant points to $X_3$ and $X_2$, respectively, at the outer edge of the gap.
\par In Table~\ref{A1}, we list the resulting temperatures ($T$) at points $X_1$, $X_2$, $X_3$, $Y_1$, $Y_2$, $Y_3$, together with the temperatures at the inner and outer edges without additional heat for the protoplanetary disc ($T_{rim}$), with physical parameters considered in the present paper and the irradiation temperature from the companion ($T_{c,irr}$). As one can see, $T_{c,irr}$ exceeds $T_{rim}$ only in one case at the point $Y_1$, where $r_{c,g}$ has a minimum value, and at further distances it is significantly smaller. The resulting temperatures ($T$) significantly exceed the temperatures of the rims only when the distance to the companion is the minimum: 15\% and 30\% for the inner and outer rims, respectively. And for the further points, the increase in the temperatures due to the heat from the companion at $X_3$ and $Y_3$ is only 2\% and 6\%, respectively. Additional heat to the outer edge of the rim is bigger, because the outer edge is further away from the star and it is always colder. If the companion would be located further out from the star, the disc would be colder and the addition heat would be more noticeable, but the diameter of the companion's Hill sphere will increase as well and hence the resulting temperature increment will be still very small.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{figA1.eps}
\end{center}
\caption{
The schematic view of the star ($S$), the disc and the companion ($C$). $X_{1-3}$ and $Y_{1-3}$ are the points at the inner and outer edges of the rim, respectively (see the text for details).
}
\label{A1}
\end{figure}
\begin{table}
\centering
\caption{Temperatures at the gap rims that account for additional heat from the brown dwarf companion ($T$) and the corresponding temperatures of the rims ($T_{rim}$), irradiation temperatures from the companion ($T_{c,irr}$) and the distances between the companion and corresponding point ($r_{c,g}$).}
\label{rvar}
\begin{tabular}{@{}ccccc@{}}
\hline\hline
point & $T$ & $T_{rim}$ & $T_{c,irr}$& $r_{c,g}$ \\
& (K) & (K)&(K) & (AU)\\
\hline
$X_1$ & 195 & 169 & 158 & 0.23 \\
$X_2$ & 177 & 169 & 115 & 0.44 \\
$X_3$ & 173 & 169 & 95 & 0.64 \\
$Y_1$ & 174 & 132 & 158 & 0.23 \\
$Y_2$ & 148 & 132 & 115 & 0.44 \\
$Y_3$ & 140 & 132 & 95 & 0.64 \\
\hline\hline
\end{tabular}
\end{table}
\end{appendix}
\begin{appendix}
\section{Additional system parameters}
\label{AdditionalSysPar}
\par Table~\ref{Mvar} summarizes the physical parameters for the companions with different masses \citep{Baraffe98} with corresponding Hill radii for the $r_c$=~1~AU, that were used to compute SEDs presented in the bottom panel of Figure~\ref{rp_var}.
\begin{table}
\centering
\caption{The physical parameters of the companions with different masses \citep{Baraffe98} and corresponding Hill radii (for $r_c$ = 1~AU),
that were used for Figure~\ref{rp_var}}
\label{Mvar}
\begin{tabular}{@{}cccc@{}}
\hline\hline
$M_c$ & $T_{c}$ & $R_{c}$ & $R_H$ \\
$(M_J)$ & (K) & $(\times 10^{9} cm)$ & (AU) \\
\hline
3 & 1098 & 10.27 & 0.108\\
7 & 1668 & 11.54 & 0.142\\
10 & 1965 & 12.62 & 0.160\\
20 & 2452 & 18.18 & 0.201\\
30 & 2616 &25.27& 0.229\\
40 & 2754 & 28.52 & 0.251\\
\hline\hline
\end{tabular}
\end{table}
\end{appendix}
|
\section{Introduction}
Given a fixed graph $H$, a graph is called $H$-free if it does not contain a subgraph isomorphic to $H$. Denote by $e(G)$ the size (or number of edges) of graph $G$.
Denote by $ex(n,H)$ the maximum size of $H$-free graphs on $n$ vertices.
In the study history of extremal graph theory, there are two types of important problems: (1) For a given graph $H$, determine or estimate $ex(n,H)$, and describle the (asymptotic) strucure of extremal graphs, as $n\rightarrow \infty$. (2) Determine the typical structure of $H$-free graphs on $n$ vertices, as $n\rightarrow \infty$.
The first problem started in 1941 when Tur\'an determined $ex(K_{r+1},n)\le t_r(n):=e(Tu_r(n))$, where the equality holds only by the {\it Tur\'an graph $Tu_r(n)$} which is formed by partitioning the set of $n$ vertices into $r$-parts of nearly equal size, and connecting two vertices by an edge whenever they belong to two different parts.
In 1946 Erd\"os and Stone \cite{ES} extended the Tur\'an theorem and determined $ex({K_{r+1}^t},n)=t_r(n)+o(n^2)$, where $K_{r+1}^t$ is a $K_{r+1}$ blow-up for some positive integer $t$, i.e., $K_{r+1}^t$ is formed by replacing every vertex $v_i$ of $K_{r+1}$ by an independent set of $t$ vertices and connecting every pair of vertices whenever they belong to different independent sets.
The second problem started in 1976 when Erd\"os, Kleitman and Rothschild \cite{EKR} showed that almost all $K_3$-free graphs are bipartite and asymptotically determined the logarithm of the number of $K_r$-free graphs on $n$ vertices, for every integer $r\ge 3$. This was strengthened by Kolaitis, Pr\"omel and Rothschild \cite{KPR}, who showed that almost all $K_r$-free
graphs are $(r-1)$-partite, for every integer $r \ge 3$. Nowadays there are a vast body of work concerning the maximum number of edges and structure of $H$-free graphs on $n$ vertices (see, e.g. \cite{BBS,BBS2, BBS3, BMS, BMSW, EFR, KPR, OPT, PS2}). And some related results have been proved for hypergraphs recently (see, e.g. \cite{BM, PS}).
However, the corresponding questions for digraphs and oriented graphs are almost all wide open, and are the subject of this paper. We shall give some notions before we start to state some relevant results. Given a digraph $G=(V,E)$, let $f_1(G)$ be the number of pairs $u,v\in V$ such that exactly one of $uv$ and $vu$ is an edge of $G$, and let $f_2(G)$ be the number of pairs $u,v\in V$ such that both $uv$ and $vu$ are edges of $G$ (in this case we call $uv$ as a double edge for convenience).
For a vertex $v$, let $f_1(v)$ be the number of $u\in V$ such that exactly one of $uv$ and $vu$ is an edge of $G$, and let $f_2(v)$ be the number of $u\in V$ such that $uv$ is a double edge. For $a\in \mathbb{R}$ with $a\ge 1$, the {\it weighted size} of $G$ is defined by $e_a(G):= a\cdot f_2(G)+f_1(G)$. For a vertex $v$, its weight is defined by $e_a(v):=a\cdot f_2(v)+f_1(v).$ This definition allows for a unified approach to extremal problems on oriented graphs and digraphs. Because for a digraph $G$, it contains $4^{f_2(G)}2^{f_1(G)}=2^{e_2(G)}$ labelled sub-digraphs and $3^{f_2(G)}2^{f_1(G)}=2^{e_{\log 3}(G)}$ oriented subgraphs if we set $a=2$ and $a=\log 3$, respectively.
Given a digraph $H$, the {\it weighted Tur\'an number $ex_a(n,H)$} is defined as the maximum weighted size $e_a(G)$ among all $H$-free digraphs $G$ on $n$ vertices. Let $DTu_r(n)$ be the digraph obtained from $Tu_r(n)$ by replacing each edge of $Tu_r(n)$ by a double edge. A {\it tournament} is an orientation of a
complete graph. We denote a transitive tournament on $r$ vertices by $T_r$. Note that $DTu_r(n)$ is $T_{r+1}$-free, so $ex_a(n,T_{r+1})\ge e_a(DTu_r(n))=a\cdot t_r(n)$.
For the first problem, Brown and Harary in \cite{BH} determined the extremal digraphs with maximum edges of order $n$ and not containing the transitive tournament $T_{r+1}$. Recently, K\"{u}hn, Osthus, Townsend and Zhao \cite{KOTZ} extended this result to weighted digraphs.
\begin{lem}\cite{KOTZ}\label{LemMaxWeightTFree}
Let $a\in(\frac{3}{2},2]$ be a real number and let $r,n\in \mathbb{N}$. Then $ex_a(n,T_{r+1})=a\cdot t_r(n)$, and $DTu_r(n)$ is the unique extremal $T_{r+1}$-free digraph on $n$ vertices.
\end{lem}
Note that from Lemma \ref{LemMaxWeightTFree} we can see that any $n$-vertex digraph $G$ with $e_a(G)>a\cdot t_r(n)$ contains $T_{r+1}$ for $a\in (\frac{3}{2},2]$. Together with this observation and the Regularity Lemma of digraphs, we establish an analogue Erd\"os-Stone theorem of weighted digraphs as follows:
\begin{thm}\label{Th1}
For all positive integers $r, t$, every real numbers $a\in(\frac{3}{2},2]$ and $\gamma>0$, there exists an integer $n_0$ such that every digraph $G$ with $n\ge n_0$ vertices and
\[ e_a(G) \ge a\cdot t_{r}(n)+\gamma n^2\] contains $T_{r+1}^t$ as a sub-digraph.
\end{thm}
For the second problem, the only results of the above type for oriented graphs were proved by Balogh, Bollob\'as and Morris \cite{BBM, BBM2} who classified the possible `growth speeds' of oriented graphs with a given property.
In 1998 Cherlin \cite{C} gave a classification of countable homogeneous oriented graphs. He remarked that `the striking work of \cite{KPR} does not appear to go over to the directed case' and made the following conjecture.
\begin{con}(Cherlin)\label{con1}
Almost all $T_3$-free oriented graphs are tripartite.
\end{con}
K\"{u}hn, Osthus, Townsend and Zhao \cite{KOTZ} verified this conjecture and showed that almost all $T_{r+1}$-free oriented graphs and almost all $T_{r+1}$-free digraphs are $r$-partite.
The second part of this work is to reconfirm and generalize the Conjecture \ref{con1}, we show that almost all $T_{3}^t$-free oriented graphs and almost all $T_{3}^t$-free digraphs are almost bipartite for every integer $ t\ge 1$.
More pricisely, let $f(n,T_{3}^t)$ and $f^*(n,T_{3}^t)$ denote the number of labelled $T_{3}^t$-free oriented graphs and digraphs on $n$ vertices, respectively. We show that
\begin{thm}\label{Th2}
For every $r,t\in \mathbb{N}$ with $r\ge 2,t\ge 1$ and any $\alpha >0$ there exists $\epsilon >0$ such that the following holds for all sufficiently large $n$.\\
(i) All but at most $f(n,T_{3}^t)2^{-\epsilon n^2}$ $T_{3}^t$-free oriented graphs on $n$ vertices can be made bipartite by changing at most $\alpha n^2$ edges.\\
(ii) All but at most $f^*(n,T_{3}^t)2^{-\epsilon n^2}$ $T_{3}^t$-free digraphs on $n$ vertices can be made bipartite by changing at most $\alpha n^2$ edges.
\end{thm}
The rest of the paper is organized as followed. We lay out some notations and set out some useful tools in Section 2. We introduce the Regularity Lemma of digraphs and give the proof of Theorem \ref{Th1} in Section 3. We establish a stability result of digraphs and give a proof of Theorem \ref{Th2} in Section 4 and give some conclusion remarks in Section 5.
\section{Notations and Tools}
A {\it digraph} is a pair $(V,E)$ where $V$ is a set of vertices and $E$
is a set of ordered pairs of distinct vertices in $V$ (note that this
means we do not allow loops or multiple edges in the same direction in
a digraph). An {\it oriented graph} is a digraph with at most one edge
between two vertices, so may be considered as an orientation of a simple undirected graph.
In some proofs, given $a,b\in \mathbb{R}$ with $0<a,b<1$, we will use the notation $a \ll b$ to mean that we can find an increasing function $f$ for which all of the conditions in the proof are satisfied whenever $a\le f(b)$. We assume all graphs, oriented graphs and digraphs to be labelled unless otherwise stated. We also assume all large numbers to be integers, so that may some times omit floors and ceilings for the sake of clarity.
Let $G=(V,E)$ be a digraph, we write $uv$ for the edge
directed from $u$ to $v$. For a vertex $v\in V$, we define the out-neighborhood
of $v$ in $G$ to be $N_G^+:=\{u\in V: vu \in E\}$, and the in-neighborhood of $v$ to be $N_G^-:=\{u\in V: uv \in E\}$. The out-degree
$d^+_G(v)$ and the in-degree $d^-_G(v)$ of $v$ in $G$ are defined by $|N_G^+|$
and $|N_G^-|$, respectively.
We define the neighborhood of $v$ to be $N_G(v):= N_G^- \bigcup N_G^+$ and the intersection of out-neighborhood and in-neighborhood of $v$ to be $N^{\pm}_G(v):= N_G^- \bigcap N_G^+$. We write $\Delta(G),\Delta^+(G)$ and $\Delta^-(G)$ for the maximum of $|N_G(x)|, |N^+_G(x)|$ and $|N^-_G(x)|$ over all vertices $v\in G$, respectively. Define $\Delta^0(G)$ as the maximum of $d^+(v)$ and $d^-(v)$ among all $v\in V$.
Given a vertex set $A$ of $G$, the sub-digraph of $G$ induced by $A$ is denoted by $G[A]$ which is the digraph obtained from $G$ by deleting vertices not in $A$ and all their incident edges. Given two disjoint subsets $A$ and $B$ of vertices of $G$, an $A-B$ edge is an edge $ab$ where $a\in A$ and $b\in B$.
We write $E(A,B)$ for the set of all these edges and put $e_G(A,B):=|E(A,B)|$.
We denote by $(A,B)_G$ the bipartite oriented subgraph of $G$ whose vertex class are $A$ and $B$ and whose edge set is $E(A,B)$. The density of $(A,B)_G$ is defined to be
\[ d_G(A,B):=\frac{e_G(A,B)}{|A||B|}.
\]
Given $\epsilon > 0$, we call $(A,B)_G$ is an $\epsilon$-{\it regular} pair if for all subsets $X\subseteq A$ and $Y\subseteq B$ with $|X|>\epsilon |A|$ and $|Y|>\epsilon |B|$ we have that $|d(X,Y)-d(A,B)|<\epsilon$. Note that $(B,A)$ may not be an $\epsilon$-regular pair since the order matters.
For a positive integer $k$ we write $[k]:=\{1,\ldots,k\}$. For convenience, we drop the subscripts of all notions if they are unambiguous. For undefined terminology and notations we refer the reader to \cite{D}.
We need the following result of forbidden digraphs container of K\"uhn et al. \cite{KOTZ}, which allows us to reduce an asymptotic counting problem to an extremal problem.
Given an oriented graph $H$ with $e(H)\ge 2$, we let
$$ m(H)=\max\limits_{H'\subset H, e(H')>1} \frac{e(H')-1}{v(H')-2}.$$
\begin{thm}(\cite{KOTZ}, Theorem 3.3)\label{HFreeContainer}
Let $H$ be an oriented graph with $h:=v(H)$ and $e(H)\ge 2$, and let $a\in \mathbb{R}$ with $a\ge 1$. For every $\epsilon>0$, there exists $c>0$ such that for all sufficiently large $N$, there exists a collection $\mathcal{C}$ of digraphs on vertex set $[n]$ with the following properties.\\
(a) For every $H$-free digraph $I$ on $[N]$ there exists $G\in \mathcal{C}$ such that $I\subset G$.\\
(b) Every digraph $G\in \mathcal{C}$ contains at most $\epsilon N^h$ copies of $H$, and $e_a(G)\le ex_a(N,H)+\epsilon N^2$.\\
(c) $\log |\mathcal{C}| \le c N^{2-1/m(H)} \log N$.
\end{thm}
Note that this result is essentially a consequence of a recent and very powerful result of Balogh, Morris and Samotij \cite{BMS} and Saxton and Thomason \cite{ST}, which introduces the notion of hypergraph containers to give an upper bound on the number of independent sets in hypergraphs, and a digraph analogue \cite{KOTZ} of the well-known supersaturation result of Erd\"os and Simonovits \cite{ES}.
\section{The Regularity Lemma and Erd\"os-Stone Theorem of Digraphs}
In this section we give the degree form of the regularity lemma for digraphs. A regularity
lemma for digraphs was proved by Alon and Shapira [3]. The degree form follows
from this in the same way as the undirected version (see [34] for a sketch of the
latter). The interested readers can refer to \cite{KS} for a survey on the Regularity Lemma.
\begin{lem}\cite{AS} (Degree form of the Regularity Lemma of Digraphs).
For all $\epsilon,M'>0$ there exist $M, n_0$
such that if $G$ is a digraph on $n\ge n_0$ vertices and $d\in [0, 1]$, then there exists a
partition of $V(G)$ into $V_0,\ldots,V_k$ and a spanning subdigraph $G'$ of $G$ satisfying the
following conditions:\\
\indent$(1)$ $M'\leq k \leq M$,\\
\indent$(2)$ $|V_0| \leq \epsilon\cdot n$,\\
\indent$(3)$ $|V_1|=\ldots=|V_{k}|=\ell$,\\
\indent$(4)$ $d^+_{G'}(x)>d^+_{G}(x)-(d+\epsilon)n$ for all vertices $x$ of $G$,\\
\indent$(5)$ $d^-_{G'}(x)>d^-_{G}(x)-(d+\epsilon)n$ for all vertices $x$ of $G$,\\
\indent$(6)$ $G'[V_i]$ is empty for all $i=1,\ldots,k$,\\
\indent$(7)$ the bipartite oriented graph $(V_i,V_j)_{G'}$ is $\epsilon$-regular and has density either $0$ or density at least $d$ for all $1\leq i,j \leq k$ and $i\neq j$.
\end{lem}
We call $V_1,\ldots,V_k$ {\it clusters} and $V_0$ the {\it exceptional set}. The last condition of the lemma says that all pairs of clusters are $\epsilon$-regular in both directions (but possibly with different densities). We call the spanning subdigraph $G'\subseteq G$ in the lemma the {\it pure digraph} with parameters $\epsilon, d, \ell$. Given clusters $V_1,\ldots,V_k$ and a digraph $G'$, the reduced digraph $R$ with parameters $\epsilon, d, \ell$ is the digraph whose vertices are $V_1,\ldots,V_k$ and whose edges are all the $V_i-V_j$ edges in $G'$ that is $\epsilon$-regular and has density at least $d$.
Note that a simple consequence of the $\epsilon$-regular pair $(A,B)$: for any subset $Y\subseteq B$ that is not too small, most vertices of $A$ have about the expected number of out-neighbors in $Y$; and similarly for any subset $X\subseteq A$ that is not too small, most vertices of $B$ have about the expected number of in-neighbors in $X$.
\begin{lem}\label{LemExpectNeighbor}
Let $(A,B)$ be an $\epsilon$-regular pair, of density $d$ say, and $X\subseteq A$ has size $|X|\ge \epsilon|A|$ and $Y\subseteq B$ has size $|Y|\ge \epsilon|B|$. Then all but at most $\epsilon|A|$ of vertices in $A$ each of which has at least $(d-\epsilon)|Y|$ out-neighbors in $Y$ and all but at most $\epsilon|B|$ of vertices in $B$ each of which has at least $(d-\epsilon)|X|$ in-neighbors in $X$.
\end{lem}
\noindent {\it Proof.} Let $A'$ be a vertex set with fewer than $(d-\epsilon)|Y|$ out-neighbors in $Y$. Then $e(A',Y)<|A'|(d-\epsilon)|Y|$, so
\[d(A',Y)=\frac{e(A',Y)}{|A'||Y|}<d-\epsilon=d(A,B)-\epsilon.\]
Since $(A,B)$ is $\epsilon$-regular, this implies that $|A'|<\epsilon|A|$.
Similarly, let $B'$ be a vertex set with fewer than $(d-\epsilon)|X|$ in-neighbors in $X$. Then $e(X,B')<|X|(d-\epsilon)|B'|$, so
\[d(X,B')=\frac{e(X,B')}{|X||B'|}<d-\epsilon=d(X,B)-\epsilon.\]
Since $(A,B)$ is $\epsilon$-regular, this implies that $|B'|<\epsilon|B|$.
\hfill \rule{4pt}{7pt}
The following lemma says that the blow-up $R^s$ of the reduced digraph $R$ can be found in $G$, provided that $\epsilon$ is small enough and the $V_i$ are large enough.
\begin{lem}\label{LemBlowUp}
For all $d\in(0,1)$ and $\Delta\ge 1$, there exists an $\epsilon_0>0$ such that if $G$ is any digraph, $s$ is an integer and $R$ is a reduced digraph of $G'$, where $G'$ is the pure digraph of $G$ with parameters $\epsilon\le \epsilon_0$, $\ell\ge s/\epsilon_0$ and $d$. For any digraph $H$ with $\Delta(G')\le \Delta$, then
\[H\subseteq R^s \Rightarrow H\subseteq G'\subseteq G.\]
\end{lem}
\noindent {\it Proof.} The proof is similar with that of Lemma 7.3.2 in \cite{D}.
Given $d$ and $\Delta$, choose $\epsilon_0<d$ small enough that
\begin{align}\label{eq2.1}
\frac{\Delta+1}{(d-\epsilon_0)^\Delta}\epsilon_0 \le 1;
\end{align}
such a choice is possible, since $\frac{\Delta+1}{(d-\epsilon)^\Delta}\epsilon \rightarrow 0$ as $\epsilon \rightarrow 0$. Now let $G,H,s,R$ be given as stated. Let $\{V_0,V_1,\ldots,V_k\}$ be the $\epsilon$-regular partition of $G'$ that give rise to $R$; thus, $\epsilon <\epsilon_0,V(R)=\{V_1,\ldots,V_k\}$ and $|V_1|=\ldots=|V_k|=\ell$. Let us assume that $H$ is actually a sub-digraph of $R^s$, with vertices $u_1,\ldots,u_h$ say. Each vertex $u_i$ lies in one of the $s$-sets $V_j^s$ of $R^s$; this defines a map $\sigma:i\mapsto j$. We aim to define an embedding $u_i \mapsto v_i\in V_{\sigma(i)}$ of $H$ in $G'$; thus, $v_1,\ldots, v_h$ will be distinct, and $v_iv_j$ will be an edge of $G'$ whenever $u_iu_j$ is an edge of $H$.
We choose the vertices $v_1,\ldots,v_h$ inductively. Throughout the induction, we shall have a ``target set" $Y_i \subseteq V_{\sigma(i)}$ assigned to each $i$; this contains the vertices that are still candidates for the choice of $v_i$. Initially, $Y_i$ is the entire set $V_{\sigma(i)}$. As the embedding proceeds, $Y_i$ will get smaller and smaller (until it collapses to $\{v_i\}$): whenever we choose a vertex $v_j$ with $j<i$ and if\\
Case (i): $u_i$ are both out-neighbor and in-neighbor of $u_j$ in $H$, we delete all those vertices from $Y_i$ that are not adjacent to $v_j$ with double edges.\\
Case (ii): $u_i$ is just out-neighbor of $u_j$ in $H$, we delete all those vertices from $Y_i$ that are not the out-neighbor of $v_j$.\\
Case (iii): $u_i$ is just in-neighbor of $u_j$ in $H$, we delete all those vertices from $Y_i$ that are not the in-neighbor of $v_j$.
In order to make this approach work, we have to ensure that the target set $Y_i$ do not get too small. When we come to embed a vertex $u_j$, we consider all the indices $i>j$ such that $u_i$ is adjacent to $u_j$ in $H$; there are at most $\Delta$ such $i$. For each of these $i$, we wish to select $v_j$ so that
\begin{align}\label{eq2.2}
Y_i^j=N^*(v_j)\bigcap Y_i^{j-1}
\end{align}
is large, where
\begin{align*} N^*(v_j)=
\left\{ \begin{array}{ll}
N^{\pm}(v_j) & \mbox{if $u_i$ are both out-neighbor and in-neighbor of $u_j$};\\
N^{+}(v_j) & \mbox{if $u_i$ is out-neighbor of $u_j$};\\
N^{-}(v_j) & \mbox{if $u_i$ is in-neighbor of $u_j$}.\\
\end{array} \right.
\end{align*}
Now this can be done by Lemma \ref{LemExpectNeighbor}: unless $Y^{j-1}_i$ is tiny (of size less than $\epsilon \ell$), all but at most $\epsilon \ell$ choices of $v_j$ will be such that \eqref{eq2.2} implies
\begin{align}\label{eq2.3}
|Y^j_i| \ge (d-\epsilon)|Y_i^{j-1}|
\end{align}
Doing this simultaneously for all of at most $\Delta$ values of $i$ considered, we find that all but at most $\Delta \epsilon \ell$ choices of $v_j$ from $V_{\sigma(j)}$, and in particular from $Y_j^{j-1}\subseteq V_{\sigma(j)}$, satisfy \eqref{eq2.3} for all $i$.
It remains to show that $|Y^{j-1}|-\Delta \epsilon \ell \ge s$ to ensure that a suitable choice for $v_j$ exists: since $\sigma(j')=\sigma(j)$ for at most $s-1$ of the vertices $u_{j'}$ with $j'<j$, a choice between $s$ suitable candidates for $v_j$ will suffice to keep $v_j$ distinct from $v_1,\ldots,v_{j-1}$. But all this follows from our choice of $\epsilon_0$. Indeed, the initial target sets $Y^0_i$ have size $\ell$, and each $Y_i$ has vertices deleted from it only when some $v_j$ with $j<i$ and $u_j$ and $u_i$ are adjacent in $H$, which happens at most $\Delta$ times. Thus,
\[|Y_i^j|-\Delta \epsilon \ell\ge (d-\epsilon)^\Delta-\Delta \epsilon \ell
\ge (d-\epsilon_0)^\Delta-\Delta \epsilon_0 \ell \ge \epsilon_0 \ell \ge s\]
whenever $j<i$, so in particular $|Y_i^j|-\Delta \ge \epsilon_0 \ell\ge \epsilon \ell$ and $|Y_j^{j-1}|-\Delta \ge \epsilon \ell\ge s$. \hfill \rule{4pt}{7pt}
We can now prove Theorem \ref{Th1} using Lemma \ref{LemMaxWeightTFree}, Lemma \ref{LemBlowUp} and the Regularity Lemma of digraphs.\\
{\noindent\bf Proof of Theorem \ref{Th1}.}
Let $d:=\gamma, \Delta=\Delta(K_{r+1}^s)$, then Lemma \ref{LemBlowUp} returns an $\epsilon_0>0$. Assume
\begin{align}\label{ineq3.1}
\epsilon_0 < \gamma/2 <1
\end{align}
Let $M'>1/\gamma$, choose $\epsilon>0$ small enough that $\epsilon \le \epsilon_0$ and $\delta:=(a-1)d-\epsilon-a\epsilon^2/2-a\epsilon >0$.
The Regularity Lemma of digraphs returns an integer $M$.
Assume
\[
n \ge \frac{Ms}{\epsilon_0(1-\epsilon)},
\]
Since $\frac{Ms}{\epsilon_0(1-\epsilon)}\ge M'$. The Regularity Lemma of digraphs provided us with an $\epsilon$-regular partition $\{V_0,V_1,\ldots,V_k\}$ of $G'$, the pure digraph of $G$, with parameters $\epsilon,d,\ell$ and $M'\le k\le M$. That is $|V_1|=\ldots=|V_k|=\ell$ and $|V_0|<\epsilon n$. Then
\begin{align}
n\ge k\ell
\end{align}
\[
\ell=\frac{n-|V_0|}{k}\ge \frac{n-\epsilon n}{M}=n\frac{1-\epsilon}{M}\ge \frac{s}{\epsilon_0}
\]
by the choice of $n$. Let $R$ be the regularity digraph of $G'$ with parameters $\epsilon, \ell, d$ corresponding to the above partition.
Since $\epsilon\le \epsilon_0, \ell \ge s/\epsilon_0$. $R$ satisfies the premise of Lemma \ref{LemBlowUp} and $\Delta(K_{r+1}^s)=\Delta$.
Thus in order to conclude by Lemma \ref{LemBlowUp} that $T^s_{r+1}\subseteq G'$, all that remains to be checked is that $T_{r+1}\subseteq R$.
Our plan was to show $T_{r+1}\subseteq R$ by Lemma \ref{LemMaxWeightTFree}. We thus have to checked that the weight of $R$ is large enough.
First by (4) and (5) of the Regularity Lemma of digraphs, we have
\begin{align}\label{eq4.1}
\parallel G\parallel_a \le \parallel G'\parallel_a +(d+\epsilon)n^2
\end{align}
At most ${|V_0| \choose 2}$ double edges lie inside $V_0$, and at most $|V_0|k\ell\le \epsilon n k\ell$ double edges join $|V_0|$ to other partition sets. The $\epsilon-$regular pairs in $G'$ of $0$ density contribute nothing to the weight of $G'$. Since each edge of $R$ corresponds to at most $\ell^2$ edges of $G'$, we thus have in total
\begin{align*}
\parallel G'\parallel_a \le \frac{1}{2}a\epsilon^2 n^2+a\epsilon n k \ell+\parallel R\parallel_a \ell^2
\end{align*}
This together with \eqref{eq4.1}, for all sufficiently large $n$, we have
\begin{align*}
\parallel R\parallel_a
& \ge k^2 \cdot \frac{a(\frac{r-1}{r}+\gamma)n^2-(d+\epsilon)n^2-\frac{1}{2}a\epsilon^2 n^2-a\epsilon n k \ell}{k^2 \ell^2}\\
&\ge a\frac{r-1}{r}k^2+\delta k^2\\
&= a\cdot t_r(k)+\delta k^2\\
& > a\cdot t_r(k).
\end{align*}
Therefore $T_{r+1}\subseteq R$ by Lemma \ref{LemMaxWeightTFree}, as desired.
\hfill \rule{4pt}{7pt}
Similar with the Erd\"os-Stone theorem of undirected graphs, the Erd\"os-Stone theorem of digraphs is interesting not only in its own right: it also has a most interesting corollary. For an oriented graph $H$, its chromatic number is defined as the chromatic number of its underlying graph. An oriented graph $H$ with chromatic number $\chi(H)$ is called homogeneous if there is an colouring of its vertices by [$\chi(H)$] such that either $E(V_i,V_j)=\emptyset$ or $E(V_j,V_i)=\emptyset$ for every $1\le i\neq j\le \chi(H)$, where $V_i$ is the vertex set with colour $i$.
Given an acyclic homogeneously oriented graph $H$ and an integer $n$, consider the number $h_n:=ex_a(n,H)/(a{n \choose 2})$: the maximum weighted density that an $n-$vertex digraph can have without containing a copy of $H$.
Theorem \ref{Th1} implies that the limit of $h_n$ as $n\rightarrow\infty$ is determined by a very simple function of a natural invariant of $H$--its chromatic number!
\begin{cor}\label{ES}
For every acyclic homogeneously oriented graph $H$ with at least one edge,
\begin{align*}
\lim_{n\rightarrow \infty} \frac{ex_a(n,H)}{a{n \choose 2}}= \frac{\chi(H)-2}{\chi(H)-1}.
\end{align*}
\end{cor}
Before the proof the Corollary \ref{ES} we need the following lemma.
\begin{lem}\cite{D}
\begin{align*}
\lim_{n\rightarrow \infty} \frac{t_{r-1}(n)}{{n \choose 2}}= \frac{r-2}{r-1}.
\end{align*}
\end{lem}
{\bf\noindent Proof of Corollary \ref{ES}}. Let $r:=\chi(H)$. Since $H$ cannot be coloured with $r-1$ colours, we have $H \nsubseteq DTu_{r-1}(n)$ for all $n\in \mathbb{N}$, and hence
\begin{align*}
a t_{r-1}(n)\le ex_a(n,H).
\end{align*}
On the other hand, $H \subseteq T_r^t$
for all sufficiently large $t$, so
\begin{align*}
ex_a(n,H)\le ex_a(n,T_r^t)
\end{align*}
for all those $t$. Let us fix such an $t$. For every $\epsilon>0$, Theorem \ref{Th1} implies that eventually (i.e. for large enough $n$)
\begin{align*}
ex_a(n,T_r^t)<a t_{r-1}(n)+\epsilon n^2.
\end{align*}
Hence for $n$ large,
\begin{align*}
\frac{t_{r-1}(n)}{{n \choose 2}}
&\le \frac{ex_a(n,H)}{a{n \choose 2}}\\
&\le \frac{ex_a(n,T_r^t)}{a{n \choose 2}}\\
&<\frac{t_{r-1}(n)}{{n \choose 2}}+\frac{\epsilon n^2}{a{n \choose 2}}\\
&<\frac{t_{r-1}(n)}{{n \choose 2}}+\frac{2\epsilon}{a(1-1/n)}\\
&\le \frac{t_{r-1}(n)}{{n \choose 2}}+4\epsilon
\end{align*}
Therefore, since $\frac{t_{r-1}(n)}{{n \choose 2}}$ converges to $\frac{r-2}{r-1}$, so does $\frac{ex_a(n,H)}{a{n \choose 2}}$.
\hfill \rule{4pt}{7pt}
\section{Stability Theorem of Digraphs and Proof of Theorem \ref{Th2}}
In this section, we establish a stability of digraphs and give a proof of Theorem \ref{Th2}. Firstly, we give the result of stability of $T_{3}^t$-free digraphs.
\begin{thm}(Stability Theorem)\label{Stability0}
Let $a\in R$
with $3/2 <a \le 2$, and $t$ be positive integer. Then for any $T_{3}^t$-free digraph with
\begin{align*}
e_a(G)= a\bigg{(}\frac{1}{2}+o(1)\bigg{)}\frac{n^2}{2}
\end{align*}
satisfies $G=DTu_{2}(n) \pm o(n^2)$.
\end{thm}
\noindent {\it Proof.}
First of all we can assume that all but $o(n)$ vertices of $G$ have weight at least $\frac{an}{2}\big{(}1+o(1)\big{)}$. For otherwise let $v_1,\ldots, v_k, k=\lfloor \epsilon \cdot n \rfloor$
($\epsilon$ is a small positive number independent of $n$) be the vertices of $G$ each of which has weight less than $\frac{an}{2}\big{(}1-c\big{)}$, where $0<c(\epsilon)<c<1$.
But then we have
\begin{align*}
e_a(G[v_{k+1},\ldots,v_n])
&\ge (\frac{a}{2}+o(1))\frac{n^2}{2} - \frac{ank}{2}(1-c)\\
& =\bigg{(}\frac{a}{4} (n^2-2kn+k^2)-\frac{k^2}{4}+\frac{ckn}{2}+o(1)\frac{n^2}{2}\bigg{)}\\
& > \frac{a }{4}(n-k)^2 \big{(}1+\delta(\epsilon,c) \big{)},
\end{align*}
where $\delta(\epsilon,c)>0$.
By Theorem \ref{Th1} we have that $G[v_{k+1},\ldots,v_n]$ and therefore $G$ contains a $T_{3}^t$ which contradicts our assumption.
Let now $v_1,\ldots,v_p, p=\big{(}1+o(1) \big{)}n$ be the vertices of $G$ each of which has weight not less than $\frac{an}{2}\big{(}1+o(1)\big{)}$. Then the weight of each vertex of $G[v_1,\cdots,v_p]$ in ($G[v_1,\cdots,v_p]$)
is at least $ap\big{(}\frac{1}{2}+o(1)\big{)}=an\big{(}\frac{1}{2}+o(1)\big{)}$.
And $e_a(G[v_1,\cdots,v_p])=\frac{ap^2}{2}\big{(}\frac{1}{2}+o(1)\big{)}=
\frac{an^2}{2}\big{(}\frac{1}{2}+o(1)\big{)}$. Thus to prove our theorem it will suffice to show that $G[v_1,\cdots,v_p]=DTu_{2}(p)\pm o(p^2)$.
Thus it is clear that without loss of generality we can assume that every vertex of our $G$ has weight at least $an\big{(}\frac{1}{2}+o(1)\big{)}$. Note that we now no longer have to use the assumption of $e_a(G)=\frac{an^2}{2}\big{(}\frac{1}{2}+o(1)\big{)}$. Since our assumption
that $e_a(v_i) \ge an\big{(}\frac{1}{2}+o(1)\big{)}, i=1,\ldots,n$ and $G$ is $T_{3}^t-$free already implies that $e_a(G)=\frac{an^2}{2}\big{(}\frac{1}{2}+o(1)\big{)}$.
We shall show that if $G$ is $T_3^t$-free digraph with $e_a(G)=\frac{an^2}{2}\big{(}\frac{1}{2}+o(1)\big{)}$ for some fixed $t$, then $G=DTu_2(n) \pm o(n^2)$.
A pair of adjacent vertices $u$ and $v$ is called {\it bad} if it is contained in only $o(n)$ of $T_3$ of $G$, otherwise it is called {\it good}. We divide the proof according the number of good pairs of vertices.
{\bf Case 1.} If $G$ has at least $\alpha n^2$ good pairs of vertices for some $\alpha>0$.\\
Let $e_1,\ldots,e_s, s\ge\alpha n^2$ be the edges each of which are contained in at least $\beta n$ of $T_3$, where $\alpha,\beta >0$. We now deduce from this assumption that
$G$ contains a $T_3^t$. Let $v_1^{(i)},\ldots,v_{r_i}^{(i)}$ be the vertices which form a $T_3$ with $e_i, r_i\ge \beta n, s \ge i\ge 1$. Since there are $2^{r}$ orientations of a star $S_{r+1}$ of ${r+1}$ vertices.
Therefore there are at least $\beta'n:=\beta n/2^{r}$ vertices of $\{v_{j}^{(i)}, r_i\ge j\ge 1\}$ formed with $e_i$ with homogeneous $T_3$, w.l.o.g., assume $\{v_{j}^{(i)}, r'_i\ge j\ge 1\}, r'_i\ge \beta'n$ connect to both end vertices of $e_i$ in the same way.
Similarly there are at least $\alpha' n^2:=\alpha n^2/2^{r+1}$ edges of $\{e_i,s\ge i\ge 1\}$ each formed with at least $\beta' n$ vertices with homogeneous $T_3$, the addition divisor of two is because there may be two choices of direction of the edges $\{e_i,s\ge i\ge 1\}$.
And all those $T_3$ formed with those at least $\alpha' n^2$ edges $e_i'$
are homogeneous.
Form all possible $t$-tuple from those homogeneous vertices $v_{r_i}^{(i)}$. We get at least
\begin{align*}
\sum\limits_{i=1}^{\alpha' n^2} {r_i' \choose t}\ge \sum\limits_{i=1}^{\alpha' n^2} {\beta'n \choose t} \ge \alpha' n^2 \frac{(\beta'n)^t}{3^t t!}
>\alpha' n^2 (\frac{\beta'}{3})^t {n \choose t}
\end{align*}
$t$-tuples. Since the total number of $t$-tuples formed from $n$ elements is ${n \choose t}$, there is a $t$-tuple say $z_1,\ldots,z_t$ which corresponds to at least $\alpha' n^2 (\frac{\beta'}{3})^t$ edges $e_i$. By Theorem \ref{Th1} these edges determine a $T_2^t$ with vertices $x_1,\ldots,x_t;y_1,\ldots,y_t$. Thus finally $G[x_1,\ldots,x_t;y_1,\ldots,y_t;z_1,\ldots,z_t]$ and thus $G$ contains a $T_3^t$ as stated. But by our assumption our $G$ does not contain a $T_3^t$. This contradiction completes this part of proof.
{\bf Case 2.} If $G$ has $o(n^2)$ good pairs of vertices. Let $G'$ obtained from $G$ by deleting all edges between every good pair of vertices. Since $e_a(G)= a\bigg{(}\frac{1}{2}+o(1)\bigg{)}\frac{n^2}{2}$, we have
\[e_a(G)\ge e_a(G')\ge a\bigg{(}\frac{1}{2}+o(1)\bigg{)}\frac{n^2}{2}-a\cdot o(n^2)= a\bigg{(}\frac{1}{2}+o(1)\bigg{)}\frac{n^2}{2},\]
thus $e_a(G')= a\bigg{(}\frac{1}{2}+o(1)\bigg{)}\frac{n^2}{2}$.
By the same argument as in the beginning of the proof, we may assume that $e_a(v_i) \ge an\big{(}\frac{1}{2}+o(1)\big{)}, (i=1,\ldots,n)$ in $G'$. We divide the proof into two subcases according to whether $G'$ contains double edges or not.\\
{\bf Subcase 2.1.} If $G'$ contains double edge(s). Assume $uv$ is a double edge of $G'$, then $u$ and $v$ connect to $(\frac{1}{2}+o(1))n$ vertices with double edges respectively, such that $N(u)\bigcap N(v)=o(n)$.
For otherwise, $N(u)\bigcap N(v)=\Omega(n)$ since both $u$ and $v$ have weight at least $an\big{(}\frac{1}{2}+o(1)\big{)}$. Therefore $u$ and $v$ would be contained in $\Omega(n)$ of $T_3$'s, contradicting to our assumption that $G'$ contains not any good pairs of vertices. \\
{\bf Claim 1.} Every vertex in $N(u)$ ($N(v)$ resp.) has $o(n)$ neighbors in $N(u)$ ($N(v)$, resp.).
For otherwise, say $w\in N(u)$ has $\Omega(n)$ neighbors in $N(u)$, then $uw$ is contained in $\Omega(n)$ of $T_3$'s and is a good pair of vertices, contradicting our assumption.
Thus each vertex $w\in N(u)$($w\in N(v)$ resp.) connects to $n\big{(}\frac{1}{2}+o(1)\big{)}$ in $N(v)$($N(u)$ resp.) with double edges. And $e_a(G[N(u)])=o(n^2)$ and $e_a(G[N(v)])=o(n^2)$, then a simple computation shows that $G$ differs from $DTu_2(|N(u)|,|N(v)|)$ with the vertex set $\{N(u),N(v)\}$ by $o(n^2)$ edges, and $DTu_2(|N(u)|,|N(v)|)$ differs from $DTu_2(n)$ by $o(n^2)$ edges, which prove our theorem (the remaining $n-|N(u)|-|N(v)|=o(n)$ vertices can be clearly ignored).
{\bf Subcase 2.2.} If $G'$ does not contain any double edges. Let $UG'$ be its underlying undirected graph. Since $a\in (\frac{3}{2},2]$, we assume $a=\frac{3}{2}+\epsilon$ for some $\epsilon>0$. Then
\[e_a(v_i)\ge a\bigg{(}\frac{1}{2}+o(1)\bigg{)}n=(\frac{3}{2}+\epsilon)\bigg{(}\frac{1}{2}+o(1)\bigg{)}n=\frac{3}{4}n+\frac{1}{2}\epsilon n+o(n).\]
Assume $uv\in E(G')$, then
\[|N(u)\bigcap N(v)|\ge 2\big(\frac{3}{4}n+\frac{1}{2}\epsilon n+o(n) \big)-n=\frac{1}{4}n+\epsilon n+o(n).\]
For all vertices but $o(n)$ of $N(u)\bigcap N(v)$, say $w$, we have $wuvw$ is a directed triangle since $uv$ is a bad edge of $G'$. And $w$ only have $o(n)$ neighbors in $N(u)\bigcap N(v)$, for otherwise $uw$ is a good edge. Thus $w$ should have at least $\frac{3}{4}n+\frac{1}{2}\epsilon n+o(n)$ neighbors in the outside of $N(u)\bigcap N(v)$.
But then the number of vertices in $G'$ is at least
\begin{align*}
&\frac{3}{4}n+\frac{1}{2}\epsilon n+o(n)+|N(u)\bigcap N(v)|\\
&\ge \frac{3}{4}n+\frac{1}{2}\epsilon n+o(n)+\big(\frac{1}{4}n+\epsilon n+o(n)\big)\\
&=n+\frac{3}{2}\epsilon n+o(n)\\
&>n,
\end{align*}
which is a contradiction and we thus complete the proof.\hfill \rule{4pt}{7pt}
In order to keep all symbols consistent, we reshape Theorem \ref{Stability0} as follows:\\
{\noindent \bf Theorem of Stability.}\label{Stability}
Let $a\in R$
with $3/2 <a \le 2$, $t$ be positive integer. Then for any $\beta>0$ there exists $\gamma>0$ such that the following holds for all sufficiently large $n$. If a digraph $G$ on $n$ vertices is $T_{3}^t$-free and
\begin{align*}
e_a(G)= a\bigg{(}\frac{1}{2}-\gamma\bigg{)}\frac{n^2}{2},
\end{align*}
then $G=DTu_2(n) \pm \beta n^2$.
We need the Digraph Removal Lemma of Alon and Shapira \cite{AS}.
\begin{lem}(Removal Lemma).\label{RemovalLemma} For any fixed digraph $H$ on $h$ vertices, and any $\gamma >0$ there exists $\epsilon'>0$ such that the following holds for all sufficiently large $n$. If a digraph $G$ on $n$ vertices contains at most $\epsilon'n^h$ copies of $H$, then $G$ can be made $H$-free by deleting at most $\gamma n^2$ edges.
\end{lem}
We now ready to show that almost all $T_{3}^t$-free oriented graphs and almost all $T_{3}^t$-free digraphs are almost bipartite.
{\noindent\bf Proof of Theorem \ref{Th2}.} We only prove (i) here; the proof of (ii) is almost identical. Let $a:=\log 3$. Choose $n_0 \in \mathbb{N}$ and $\epsilon, \gamma, \beta>0$ such that $1/n_0 \ll \epsilon \ll \gamma \ll \beta \ll \alpha,1/r.$ Let $\epsilon':=2\epsilon$
and $n\ge n_0$. By Theorem \ref{HFreeContainer} (with $T_3^t$ and $\epsilon$ taking the roles of $H,N$ and $\epsilon$ respectively) there is a collection $\mathcal{C}$ of digraphs on vertex set $[n]$ satisfying properties $(a)-(c)$. In particular, every
$T_3^t$-free oriented graph on vertex set $[n]$ is contained in some digraph $G\in \mathcal{C}$. Let $\mathcal{C}_1$ be the family of all those $G \in \mathcal{C}$ for which $e_{\log 3}(G)\ge ex_{\log 3}(n,T_3^t)-\epsilon'n^2$.
Then the number of $T_3^t$-free oriented graphs not contained in some $G\in \mathcal{C}_1$ is at most
\begin{align*}
|\mathcal{C}| 2^{ex_{\log 3}(n,T_{3}^t)-\epsilon'n^2} \le 2^{-\epsilon n^2} f(n,T_{3}^t),
\end{align*}
because $|\mathcal{C}| \le 2^{n^{2-\epsilon'}}$ and $f(n,T_{3}^t) \ge 2^{ex_{\log 3}(n,T_{3}^t)}$. Thus it suffices to show that every digraph $G\in \mathcal{C}_1$ satisfies $G=DTu_2(n) \pm \alpha n^2$. By (b), each $G\in \mathcal{C}_1$ contains at most $\epsilon'n^{3t}$ copies of $T^t_3$. Thus by Lemma \ref{RemovalLemma} we obtain a $T^t_{3}$-free digraph $G'$ after deleting at most $\gamma n^2$ edges from $G$. Then $e_{\log 3} (G') \ge ex_{\log 3}(n,T_3^t)-(\epsilon'+\gamma)n^2$. We next apply the Theorem of Stability to $G'$ and derive that $G'=DTu_2(n)\pm \beta n^2$. As a result, the original digraph $G$
satisfies $G=DTu_2(n)\pm (\beta+\gamma)n^2$, hence $G=DTu_2(n)\pm \alpha n^2$ as required.\hfill \rule{4pt}{7pt}
\section{Concluding Remarks}
K\"{u}hn, Osthus, Townsend and Zhao \cite{KOTZ} also gave exactly structures of $T_{r+1}$-free oriented graphs and digraphs, but the exactly structures of $T_{r+1}^t$-free oriented graphs and digraphs
are still out of reach from us. We believe the exact structures are the
same as those of $T_{r+1}$-free oriented graphs and digraphs. Therefore, we ending this paper with the following conjecture:
\begin{con}
Let $r,t\in \mathbb{N}$ with $r\ge 2,t\ge 1$. Then the following hold.\\
(i) Almost all $T^t_{r+1}$-free oriented graph are $r$-partite.\\
(ii) Almost all $T^t_{r+1}$-free digraph are $r$-partite.
\end{con}
|
\section{Introduction}
Quantum groups such as $U_q({\mathfrak g})$ associated to complex semisimple Lie algebras \cite{Dri,Jimbo}, and their finite-dimensional quotients $u_q({\mathfrak g})$ at $q$ a primitive $n$-th root of unity, have been extensively studied since the 1980s and 1990s respectively. The latter are covered in several texts such as \cite{Lus,Jan}, although precise definitions and the relation to Lusztig's celebrated divided-difference versions of $U_q({\mathfrak g})$ are quite subtle and depend on the precise root when $n$ is small. See \cite{Len} for a recent work. There are also corresponding `coordinate algebra' quantum groups $C_q[G]$ and in principle reduced finite-dimensional quotients $c_q[G]$ although again best understood for specific cases \cite{CL}.
In spite of some extensive literature, one problem which we believe to be open till now even for the simplest case of $c_q[SL_2]$ is a description of its dual basis in terms of the generators and relations. Here $u_q(sl_2)$ has generators $F,K,E$ with the relations of $U_q(sl_2)$ and additionally $E^n=F^n=0, K^n=1$, and PBW basis $\{F^i K^jE^k\}$ for $0\le i,j,k<n$. The dual Hopf algebra $c_q[SL_2]$ is a
quotient of ${\Bbb C}_q[SL_2]$ with its standard matrix entry generators $a,b,c,d$, and the additional relations $a^{n}=d^{n}=1, b^{n}=c^{n}=0$ to give a Hopf algebra extension
\[ {\Bbb C}[SL_2]\hookrightarrow {\Bbb C}_q[SL_2]\twoheadrightarrow c_q[SL_2].\]
This $c_q[SL_2]$ has an obvious monomial basis $\{b^ia^jc^k\}$ but its Hopf algebra pairing with the PBW basis of $u_q(sl_2)$ is rather complicated (it can be related to the representation theory of the quantum group) and this does not constitute a dual basis even up to normalisation. Knowing a basis and dual basis is equivalent to knowing the canonical coevaluation element, which has many applications including Hopf algebra Fourier transform. Here we solve the dual basis problem for $c_q[SL_2]$ at $q$ a primitive odd root of unity in Corollary~\ref{dual basis}, finding new generators $X,t,Y$ of $c_q[SL_2]$ such that normalised monomials $\{X^it^jY^k\}$ {\em are} essentially a dual basis in the sense of being dually paired by
\[ \<X^it^jY^k,F^{i'}K^{j'}E^{k'}\>=\delta_{ii'}\delta_{kk'}q^{jj'}[i]_{q^{-1}}![k]_q!,\]
where $[i]_q$ etc. are $q$-integers. An actual dual basis immediately follows. Section~6 similarly computes the dual basis for $c_q[SL_3]$ at certain roots of unity including all $n$ that are prime and congruent to $\pm 1$ mod 12.
This result depends on a general `braided quantum double' or {\em double-bosonisation} construction \cite{db,db2,Primer} which associates to each finite-dimensional braided-Hopf algebra (`braided group') $B$ living in the category of modules over a quasitriangular Hopf algebra $H$, a new quasitriangular Hopf algebra
\[ B^{*\mathrm{\underline{cop}}} {>\!\!\!\triangleleft\kern-.33em\cdot} H {\cdot\kern-.33em\triangleright\!\!\!<} B=:D_H(B),\]
where the second notation has also been used in the literature in line with the view of this in \cite{db2} as the closest one can come to the bosonisation of a `braided double' of $B$ (the latter does not itself exist in the strictly braided case). Thus, for any $n$ we take $H={\Bbb C}[K]/(K^n-1)$ with its natural quasitriangular structure ${\mathcal R}_K$ so that its modules are the category of ${\Bbb Z}$-graded spaces with braiding given by a power of $q$ according to the degrees. We take $B={\Bbb C}[E]/(E^n)$ and $B^{*\mathrm{\underline{cop}}}={\Bbb C}[F]/(F^n)$ and obtain a version
\[ B^{*\mathrm{\underline{cop}}} {>\!\!\!\triangleleft\kern-.33em\cdot} H {\cdot\kern-.33em\triangleright\!\!\!<} B ={\mathfrak u}_q(sl_2)\cong\begin{cases} u_p(sl_2) & n=2m+1,\ p=q^{-m};\ p^2=q,\\ {\rm something\ else} & n\ {\rm even.}\end{cases}\]
To illustrate the even case, ${\mathfrak u}_{-1}(sl_2)$ in Example~\ref{exq2} is an interesting 8-dimensional strictly quasitriangular and self-dual Hopf algebra presumably known elsewhere. Our approach to the dual basis problem is to work out the dual version or {\em co-double bosonisation} and use this to construct the dual of ${\mathfrak u}_q(sl_2)$ in the dual form
\[ B^{\mathrm{\underline{op}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*}=coD_A(B),\]
where each tensor factor pairs with the corresponding factor on the ${\mathfrak u}_q(sl_2)$ side. This dual version of double bosonisation is in Section~3 and is conceptually given by reversing arrows in the original construction, but in practice takes a great deal of care to trace through all the layers of the construction. Moreover, we do not want to be limited to finite-dimensional $A$ and give a self-contained algebraic proof for $A$ any coquasitriangular Hopf algebra and $B$ a finite-dimensional braided group living in the category of its comodules. When $A$ is also finite dimensional, the resulting object has a Hopf algebra duality pairing with the double bosonisation by, schematically, $\langle B^{\mathrm{\underline{op}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*}, B^{*\mathrm{\underline{cop}}}{>\!\!\!\triangleleft\kern-.33em\cdot} H {\cdot\kern-.33em\triangleright\!\!\!<} B \rangle = \langle B^{\mathrm{\underline{op}}}, B^{*\mathrm{\underline{cop}}} \rangle \langle A, H \rangle \langle B^{*},B \rangle$. Unlike double bosonisation, the co-double bosonisation has coalgebra surjections to the constituents $B^{\underline{\rm op}},A,B^*$ making calculations for it harder than the original version.
In Section~\ref{Sec4} we take $B=\mathbb{C}[X]/(X^n)$ and its dual $B^{*}=\mathbb{C}[Y]/(Y^n)$, again braided-lines but this time viewed in the category of $A=\mathbb{C}[t]/(t^n-1)$-comodules with its standard coquasitriangular structure ${\mathcal R}(t,t)=q$ (so that its comodules form the same braided category of ${\Bbb Z}$-graded vector spaces as before). Then the co-double bosonisation gives a coquasitriangular Hopf algebra ${\mathfrak c}_q[SL_2]=c_p[SL_2]$ when $n$ is odd and some other coquasitriangular Hopf algebra when $n$ is even. This is Theorem~\ref{cdbthm} with the dual basis result a corollary of the triangular decomposition. As an application, Hopf algebra Fourier transform ${\mathcal F}:{\mathfrak c}_q[SL_2]\to {\mathfrak u}_q[sl_2]$ is worked out in Section~\ref{Sec5} and shown to behave well with respect to the 3D-calculus of $c_p[SL_2]$. Another application of the canonical element for the pairing, which we do not discuss, is to provide the quasitriangular structure of the Drinfeld double $D(u_p(sl_2))$ of interest in 3D quantum gravity.
Although the details rapidly become complicated, Sections~~\ref{upsl3} and~\ref{cpsl3} similarly study the next iteration, ${\mathfrak c}_q[SL_3]$ as dually paired to ${\mathfrak u}_q(sl_3)$, where $H=\widetilde{{\mathfrak u}_q(sl_2)}$ is a certain central extension and $B={\mathfrak c}_q^2$ denotes the usual quantum-braided plane but reduced at the root of unity by making the generators nilpotent of order $n$. The central extension requires an integer $\beta$ such that $\beta^2=3$ mod $n$, where we assume that $n=2m+1$ is odd and set $p=q^{-m}$. The dual of $B$ has a similar form and double-bosonisation gives us
\[ {\mathfrak u}_q(sl_3)\cong\begin{cases}u_p(sl_3) & m>1,\\ (u_p(sl_3)/\<K_1-K_2\>)\otimes({\Bbb C}[g]/(g^n-1)) & m=1,\end{cases}\]
where the $m=1$ case equates the two Cartan generators of the usual quantum group. This quotient is necessarily quasitriangular by our construction whereas we are not clear if this is the case for $u_p(sl_3)$ itself when $q^3=1$. We then construct the dual by $A=\widetilde{{\mathfrak c}_q[SL_2]}$ and similar quantum-braided planes now as Hopf algebras in its category comodules lead to a dual coquasitriangular Hopf algebra ${\mathfrak c}_q[SL_3]$. For $m>1$ we show that this is isomorphic to the usual $c_p[SL_3]$ while for $m=1$ we obtain a central extension of a sub-Hopf algebra of $c_p[SL_3]$. Clearly, one could go on to analyse other choices of $n$, as well as to look similarly at the next iteration for ${\mathfrak u}_q(sl_4)$ and its duality with ${\mathfrak c}_q[SL_4]$, etc. Even at the second stage of $H=\widetilde{u_q(sl_2)}$, there are other potential choices for braided planes including some that give versions of $u_q(g_2)$ and $u_q(sp_2)$ (details will be given elsewhere) and others that have no classical picture at all. Existence was covered at the semiclassical or Lie bialgebra level in \cite{Ma:blie} as an inductive process that adds one to the rank of the Lie algebra at each iteration, and is also clear for generic $U_q({\mathfrak g})$ in a suitable setting \cite{db}. Section~\ref{fermionic} illustrates a non-classical choice where $A={\Bbb C}_q[GL_2]$ is not finite dimensional, $q$ is generic and $B={\Bbb C}_q^{0|2}$ is the `fermionic quantum-braided plane' in the category of $A$-comodules. This leads to an exotic but still coquasitriangular version of ${\Bbb C}_q[SL_3]$ with some matrix entries `fermionic'. We also note that the inductive approach, even after multiple iterations, preserves a triangular decomposition in which the accumulated central generators form the `Cartan' factor, the accumulated braided groups $B$ form a `positive' braided group on one side and their duals form a `negative' braided group on the other side. For the classical families, this recovers versions of $u_q(n_\pm)$ but now as braided-Hopf algebras with dual bases.
\section{Preliminaries}\label{Sec2}
We recall the notations and facts about Drinfeld's (co)quasitriangular Hopf algebras as can be found in several texts, for example \cite{Foundation,Primer}, braided groups and bosonisation as introduced in \cite{AlgBr, Braid, Bos, Foundation} and double bosonisation \cite{db, db2, Ma:Rem, Primer}. We also establish lemmas needed for a clean presentation of the latter and its dualisation.
\subsection{Quasitriangular Hopf algebras} Recall that a Hopf algebra is $(H, \Delta, \epsilon, S)$ where $H$ is a unital algebra, $\Delta : H \to H \otimes H$ and $\epsilon:H\to k$ form a coalgebra and are algebra maps, and there is an antipode $S : H \to H$ obeying $(Sh\o)h\t=\epsilon(h)=h\o(Sh\t)$ for all $h\in H$. We use Sweedler's notation $\Delta h = h{\o}\otimes h{\t}$ (summation understood) and $k$ is the ground field. Modules/comodules of $H$ have a tensor product defined respectively by pull back/push out along the co/product. Another Hopf algebra $A$ is `dually paired' if $\langle\ ,\ \rangle : A\otimes H \to k$ makes the coalgebra and antipode on one side adjoint to the algebra and antipode on the other (e.g., $\langle ab, h \rangle= \langle a,h{\o} \rangle \langle b,h{\t} \rangle$ for all $a,b\in A$ and $h\in H$). If $H$ is finite dimensional then we can take $A=H^{*}$.
A Hopf algebra is {\em quasitriangular} \cite{Dri} if equipped with invertible $\mathcal{R}=\mathcal{R}^{\o}\otimes \mathcal{R}^{\t}\in H\otimes H$ (summation understood) such that $(\Delta \otimes \mathrm{id})\mathcal{R}=\mathcal{R}_{13}\mathcal{R}_{23}$, $(\mathrm{id}\otimes \Delta)\mathcal{R}=\mathcal{R}_{13}\mathcal{R}_{12}$ and $\mathrm{flip}\circ \Delta h =\mathcal{R}(\Delta h)\mathcal{R}^{-1}$. Here
$\mathcal{R}_{12}=\mathcal{R}\otimes 1$ and so forth. We denote by $\bar{H}$ the quasitriangular Hopf algebra which is the same Hopf algebra as $H$ but with quasitriangular structure $\bar{\mathcal{R}}=\mathcal{R}^{-1}_{21}$. The dual notion, e.g. in \cite{Braid}, of a {\em coquasitriangular} Hopf algebra $A$ is a Hopf algebra with a convolution-invertible map $\mathcal{R}: A \otimes A \to k$ satisfying
\begin{equation}\label{quabic} \mathcal{R}(ab,c)=\mathcal{R}(a,c{\o})\mathcal{R}(b,c{\t}),\quad \mathcal{R}(a,bc)=\mathcal{R}(a{\o},c)\mathcal{R}(a{\t},b),\end{equation}
\begin{equation}\label{quacom} a{\o}b{\o}\mathcal{R}(b{\t},a{\t})=\mathcal{R}(b{\o},a{\o})b{\t}a{\t}\end{equation}
for all $a,b,c\in A$. We define $\bar A$ to be the same Hopf algebra as $A$ but with coquasitriangular structure $\bar{\mathcal{R}}=\mathcal{R}^{-1}_{21}$. Equivalently, $\bar{\mathcal{R}}(a,b)=\mathcal{R}(Sb,a)$ for all $a,b \in \bar{A}$. It is shown in \cite{Braid,Foundation} that the antipode in this context is invertible.
Let $H$ (resp. $A$) be (co)quasitriangular. The monoidal categories of left or right (co)modules are braided in the sense of an isomorphism $\Psi_{V,W}:V\otimes W\to W\otimes V$ obeying axioms similar to the transposition map, but {\em not} $\Psi_{W,V}\Psi_{V,W}={\rm id}$, given by
\[\begin{aligned}
\Psi_{L}(v\otimes w)=&\mathcal{R}^{\t}\rhd w \otimes \mathcal{R}^{\o}\rhd v, & \Psi_{R}(v\otimes w)=& w\lhd \mathcal{R}^{\o}\otimes v \lhd \mathcal{R}^{\t},\\
\Psi^{L}(v\otimes w)=&\mathcal{R}(w^{\bar{\o}},v^{\bar{\o}})w^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes v^{\bar{{}_{\scriptscriptstyle(\infty)}}}, & \Psi^{R}(v\otimes w)=& w^{\bar{{}_{\scriptscriptstyle(0)}}}\otimes v^{\bar{{}_{\scriptscriptstyle(0)}}}\mathcal{R}(v^{\bar{\o}},w^{\bar{\o}})
\end{aligned}\]
where $\Psi_{L}$ is the braiding for the left-modules category ${}_{H}{\mathcal M}$ with action ${\triangleright}$, $\Psi_R$ the same for right modules ${\mathcal M}_H$ and action ${\triangleleft}$, $\Psi^{L}$ for the left-comodule category ${}^{A}\mathcal{M}$ with coaction $\Delta_Lv=v^{\bar{\o}}\otimes v^{\bar{{}_{\scriptscriptstyle(\infty)}}}$ and $\Psi^R$ for right-comodules ${\mathcal M}^A$ with coaction denoted $\Delta_Rv=v^{\bar{{}_{\scriptscriptstyle(0)}}}\otimes v^{\bar{\o}}$ (summations understood).
\subsection{Bosonisation and cobosonisation}
A left $H$-module algebra $B$ means a Hopf algebra $H$ acting from the left on an algebra $B$ such that $h{\triangleright}(bc)=(h\o{\triangleright} b)(h\t {\triangleright} c)$ and $h{\triangleright} 1=\epsilon(h)$ for all $b,c\in B$ and $h\in H$. Equivalently, $B\in {}_H{\mathcal M}$ as an algebra, i.e., an object
and the product and unit maps are morphisms. One has the familiar smash or cross product algebra which we denote $B {>\!\!\!\triangleleft} H$ built on $B\otimes H$ with $(b\otimes h)(c \otimes g)= b(h{\o}\rhd c)\otimes h{\t}g$ for all $b,c\in B$ and $h,g\in H$. Similarly if $B\in {\mathcal M}_H$ as an algebra there is a right cross product algebra $H \rcross B$ built on $H \otimes B$ with $(h\otimes b)(g \otimes c)= hg{\o}\otimes (b\lhd g{\t})c$. Similarly, given a coalgebra $B\in {}^H{\mathcal M}$ (a left-$H$-comodule coalgebra or explicitly
\begin{align}
(\mathrm{id}\otimes \Delta)\Delta_{L}(b)=b{\o}^{\bar{\o}}b{\t}^{\bar{\o}}\otimes b{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes b{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}},\quad ({\rm id}\otimes\epsilon)\Delta_L=1\otimes\epsilon \label{left-comod-coalg}
\end{align}
for all $b\in B$) one has a left cross coproduct coalgebra $B{>\!\!\!\blacktriangleleft} H$ built on $B\otimes H$ with \begin{align*}
\Delta(b\otimes h)=& b{\o}\otimes b{\t}^{\bar{\o}}h{\o}\otimes b{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes h_{\t}.
\end{align*}
Given a coalgebra $B\in {\mathcal M}^H$ (so a right $H$-comodule coalgebra or
\begin{align}
(\Delta \otimes \mathrm{id})\Delta_{R}(b)=b{\o}^{\bar{{}_{\scriptscriptstyle(0)}}}\otimes b{\t}^{\bar{{}_{\scriptscriptstyle(0)}}}\otimes b{\o}^{\bar{\o}}b{\t}^{\bar{\o}},\quad (\epsilon\otimes{\rm id})\Delta_R=\epsilon\otimes 1 \label{right-comod-coalg}
\end{align}
for all $b\in B$) there is a right cross coproduct coalgebra $H{\blacktriangleright\!\!<} B$ built on $H\otimes B$ with \begin{align*}
\Delta (h \otimes b)=& h{\o}\otimes b{\o}^{\bar{{}_{\scriptscriptstyle(0)}}}\otimes h{\t}b{\o}^{\bar{\o}}\otimes b{\t}.
\end{align*}
We refer to \cite{Foundation} for details. When $H$ is quasitriangular there is a braided monoidal functor ${}_H{\mathcal M}\hookrightarrow {}_H^H{\mathcal M}$ in \cite{Ma:dou,Foundation} with a coaction induced by the quasitriangular structure of $H$ so as to form a `crossed' or Yetter-Drinfeld module. Similarly from the right and dually for $A$ coquasitriangular via functors ${}^A{\mathcal M}\hookrightarrow{}_A^A{\mathcal M}$ and ${}^A{\mathcal M}\hookrightarrow{\mathcal M}{}_A^A$. The latter involve an action induced by the given coaction.
We also need the notion of a `braided group' or Hopf algebra $B$ in a braided category ${\mathcal C}$, the basic theory of which was worked out in \cite{Braid,Bos,AlgBr}. The unit element is viewed as a morphism $\eta:\underline 1\to B$ from the unit object which in our case will just be $k$. The product, counit, coproduct and antipode are morphisms and we underline the latter two for clarity. In our concrete setting we write $\underline\Delta b=b\underline{\o}\otimes b_{\underline{\t}}$ (summation understood) and recall that $\underline\Delta$ is an algebra hom to the braided tensor product algebra, so that $\underline{\Delta} \cdot=(\cdot \otimes \cdot)(\mathrm{id}\otimes \Psi \otimes \mathrm{id})(\underline{\Delta}\otimes \underline{\Delta})$ with $\Psi$ the braiding on $B\otimes B$. We have \cite{Braid,AlgBr},
\begin{equation}\label{braidedS}
\underline{S} \circ \cdot = \cdot \circ \Psi \circ (\underline{S}\otimes \underline{S}),\quad
\underline{\Delta}\circ \underline{S}=(\underline{S}\otimes \underline{S}) \circ \Psi \circ \underline{\Delta}. \end{equation}
\begin{lemma}\cite{Foundation,Primer,Bos} \label{boson}
Let $H$ be quasitriangular and $B\in{\mathcal M}_H$ a braided group. Then $H\rcross B$ by the given action and $H{\blacktriangleright\!\!<} B$ by the induced coaction form a Hopf algebra $H {\cdot\kern-.33em\triangleright\!\!\!<} B$ (the bosonisation of $B$).
\end{lemma}
The coproduct here is $\Delta(h\otimes b)=h{\o}\otimes b{\underline{\o}}\lhd \mathcal{R}^{\o}\otimes h{\t}\mathcal{R}^{\t}\otimes b{\underline{\t}}$.
Similarly for $B\in {}_H{\mathcal M}$ to give $B{>\!\!\!\triangleleft\kern-.33em\cdot} H$. If $A$ is coquasitriangular and $B\in{\mathcal M}^A$ then the {\em cobosonisation} is the ordinary Hopf algebra $A{\cdot\kern-.33em\triangleright\!\!\!<} B$ with $A{\blacktriangleright\!\!<} B$ by the given coaction and $A\rcross B$ by the induced action from the above functor. Explicitly, the cross product is $(a \otimes b)(d \otimes c)=ad{\underline{\o}}\otimes b^{\bar{{}_{\scriptscriptstyle(0)}}}c\, \mathcal{R}(b^{\bar{\o}},d{\underline{\t}})$. Both constructions are examples of a more general Radford-Majid biproduct theorem \cite{Rad:str, Ma:skl} (the latter gave the categorical picture) whereby for any Hopf algebra $H$ with invertible antipode, Hopf algebras with split projections to $H$ are of the form $B{>\!\!\!\triangleleft\kern-.33em\cdot} H$ for some $B\in {}_H^H{\mathcal M}$.
Finally, the notion of a dually paired or categorical dual braided group $B^\star$ (when $B$ is a rigid object, e.g. finite-dimensional in our applications) in \cite{Braid,AlgBr} needs a little care to define the pairing $B^\star\otimes B^\star\otimes B\otimes B$ by pairing $B^\star\otimes B$ in the middle first. Pairing maps go to the trivial object. In our context, where objects are built on vector spaces, it is useful to match ordinary Hopf algebra conventions by defining $B^*$ with the adjoint algebra and coalgebra structures in the usual way rather than the above categorical way, which, however, canonically lands $B^*$ in a different category from $B$.
\begin{lemma}\label{lemma dual boson}
Let $H$ be finite dimensional and quasitriangular with dual $A$, and $B$ be a finite-dimensional braided group in $\mathcal{M}_{H}$. Then $B^*\in {\mathcal M}^A$ and $(H{\cdot\kern-.33em\triangleright\!\!\!<} B)^{*}=A{\cdot\kern-.33em\triangleright\!\!\!<} B^{*}$. Similarly, if $B\in {}_{H}\mathcal{M}$ then $B^*\in {}^A{\mathcal M}$ and $(B{>\!\!\!\triangleleft\kern-.33em\cdot} H)^{*}=B^{*}{>\!\!\!\triangleleft\kern-.33em\cdot} A$.
\end{lemma}
\subsection{Double bosonisation}
Another basic fact about braided groups is that if $B\in {\mathcal C}$ with invertible antipode then $B^{\mathrm{\underline{op}}}$ with the same coalgebra structure as $B$ but with braided-opposite product and antipode given by
\begin{align}
\cdot_{\mathrm{\underline{op}}} = \cdot \circ \Psi^{-1}_{B,B},\quad \bar{S}=\underline{S}^{-1}
\end{align}
is a braided group in $\bar {\mathcal C}$, by which we mean ${\mathcal C}$ with the reversed (inverse) braid crossing \cite{Braid}. The same remarks apply for $B^{\mathrm{\underline{cop}}}\in\bar {\mathcal C}$ with $\underline{\Delta}_{\mathrm{\underline{cop}}} = \Psi^{-1}_{B,B} \circ \underline{\Delta}$ and inverted $\underline S$.
If $H$ is quasitriangular and ${\mathcal C}={\mathcal M}_H$ then $\bar{\mathcal C}={\mathcal M}_{\bar H}$. Let $B$ be a braided group in $\mathcal{M}_{H}$. By the theory of bosonisation, we have two Hopf algebras $H {\cdot\kern-.33em\triangleright\!\!\!<} B$ and $B^{*\mathrm{\underline{cop}}}{>\!\!\!\triangleleft\kern-.33em\cdot} \bar{H}$. We can glue them together to get the following theorem.
\begin{theorem}c.f. \cite[Theorem 3.2]{db}\label{Double Bosonisation}
Let $H$ be a quasitriangular Hopf algebra. Let $B$ be a finite-dimensional braided group in $\mathcal{M}_{H}$. There is an ordinary Hopf algebra $B^{*\mathrm{\underline{cop}}}{>\!\!\!\triangleleft\kern-.33em\cdot} H {\cdot\kern-.33em\triangleright\!\!\!<} B$, the double bosonisation, built on $B^{*\mathrm{\underline{cop}}}\otimes H \otimes B$ and containing $B^{*\mathrm{\underline{cop}}}{>\!\!\!\triangleleft\kern-.33em\cdot} \bar{H}$ and $H {\cdot\kern-.33em\triangleright\!\!\!<} B$ as sub-Hopf algebras with cross relation
\begin{equation*}
\begin{aligned}
bc=&(\mathcal{R}_{1}^{\t}\rhd c{\overline{\t}})\mathcal{R}_{2}^{\t}\mathcal{R}_{1}^{{\scriptscriptstyle -(1)}}(b{\underline{\t}}\lhd \mathcal{R}_{2}^{{\scriptscriptstyle -(1)}})\langle \mathcal{R}_{1}^{\o} \rhd c{\overline{\o}}, b{\underline{\o}} \lhd \mathcal{R}_{2}^{\o} \rangle \langle \mathcal{R}_{1}^{{\scriptscriptstyle -(2)}} \rhd \bar{S}c{\overline{\th}}, b{\underline{\th}} \lhd \mathcal{R}_{2}^{{\scriptscriptstyle -(2)}} \rangle.
\end{aligned}
\end{equation*}
Furthermore, $B^{*\mathrm{\underline{cop}}}{>\!\!\!\triangleleft\kern-.33em\cdot} H {\cdot\kern-.33em\triangleright\!\!\!<} B$ has a quasitriangular structure ${\mathcal R}^{\rm new}=\overline{\mathrm{exp}}\cdot\mathcal{R}$,
where $\overline{\mathrm{exp}}=\sum f^{a}\otimes \underline{S}e_{a}$, $\{e_{a}\}$ is a basis of $B$ and $\{f^{a}\}$ is a dual basis of $B^{*}$.
\end{theorem}
In fact, $B$ in \cite{db} is not required to be finite dimensional but we have restricted to the finite-dimensional case for simplicity. Our goal is a dual version of this theorem with $A$ coquasitriangular and $B\in {}^A{\mathcal M}$, in which case the category with reversed braiding is ${}^{\bar A}{\mathcal M}$ and
\begin{align}
a \cdot_{\mathrm{\underline{op}}} b =& \mathcal{R}(Sa^{\bar{\o}},b^{\bar{\o}}) b^{\bar{{}_{\scriptscriptstyle(\infty)}}}a^{\bar{{}_{\scriptscriptstyle(\infty)}}}
\end{align}
for all $a,b \in B^{\mathrm{\underline{op}}}$. As in Lemma~\ref{lemma dual boson}, we think of ${}^A{\mathcal M}$ as ${\mathcal M}_H$ in the finite-dimensional Hopf algebra case by evaluating against a coaction of $A$ to get an action of $H$.
\begin{lemma} \label{lemma about B op braid} If $H$ is finite dimensional and quasitriangular with dual $A$ and $B\in {}^A{\mathcal M}$ is finite dimensional then
$(B^{\underline{\mathrm{op}}})^{*}=B^{*\mathrm{\underline{cop}}}\in {}_{\bar{H}}\mathcal{M}$.
\begin{proof} Here $B^{\underline{\mathrm{op}}}\in {}^{\bar A}{\mathcal M}$ or ${\mathcal M}_{\bar H}$ and $(B^{\underline{\mathrm{op}}})^*\in {}_{\bar H}{\mathcal M}$ where $B^{*\underline{\mathrm{cop}}}$ lives. It is clear that the coproduct of $B^{\underline{\mathrm{op}}}$ corresponds to the product of $B^{*\underline{\mathrm{cop}}}$. For the other half,
\begin{align*}
\langle x, b\cdot_{\mathrm{\underline{op}}} c \rangle&=\langle x, c^{\bar{{}_{\scriptscriptstyle(\infty)}}}b^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \mathcal{R}(Sb^{\bar{\o}},c^{\bar{\o}})=\langle x{\underline{\o}}, c^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle x{\underline{\t}}, b^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle b^{\bar{\o}}, \mathcal{R}^{{\scriptscriptstyle -(2)}} \rangle \langle c^{\bar{\o}}, \mathcal{R}^{{\scriptscriptstyle -(1)}} \rangle\\
&=\langle x{\underline{\o}}, c \lhd \mathcal{R}^{{\scriptscriptstyle -(1)}} \rangle \langle x_{\underline{\t}}, b \lhd \mathcal{R}^{{\scriptscriptstyle -(2)}} \rangle=\langle \mathcal{R}^{{\scriptscriptstyle -(2)}} \rhd x_{\underline{\t}} \otimes \mathcal{R}^{{\scriptscriptstyle -(1)}} \rhd x_{\underline{\o}},b\otimes c \rangle
\end{align*}
which is $\langle \underline{\Delta}_{\mathrm{\underline{cop}}} x, b\otimes c \rangle$ as required.
\end{proof}
\end{lemma}
\section{Co-double Bosonisation} \label{Sec3}
The dual version of Theorem~\ref{Double Bosonisation} can in principle now be deduced using the lemmas in the preceding section, at least when $A$ is finite dimensional. However, we do not want to be limited to this case and give a direct proof of the resulting formulae.
\begin{theorem}[Co-double bosonisation]\label{codbos}
Let $B$ be a finite-dimensional braided group in ${}^{A}\mathcal{M}$ with basis $\{e_{a}\}$. Denote its dual by $B^{*} \in \mathcal{M}^{A}$ with dual basis $\{f^{a}\}$. Then there is an ordinary Hopf algebra $B^{\mathrm{\underline{op}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*}$, the {\em co-double bosonisation}, built on the vector space $B^{\mathrm{\underline{op}}} \otimes A \otimes B^{*}$ with \begin{align*}
(x\otimes & k \otimes y)(w \otimes \ell \otimes z)= x\cdot_{\mathrm{\underline{op}}}w^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes k\t \ell\o \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z \ \mathcal{R}(y^{\bar{\o}}, \ell\t)\mathcal{R}(Sk\o, w^{\bar{\o}}),\\
\Delta&(x\otimes k \otimes y)\\
=& \sum\limits_{a} x\underline{\o}\otimes x\underline{\t}^{\bar{\o}}\o k\o \otimes f^{a}\otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes k{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\underline{\t}}\\
&\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\t}k{\t})\mathcal{R}(S(k{\th}y{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{\th}^{\bar{\o}}) \ \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}{\underline{\t}} \rangle
\end{align*}
for all $x,w \in B^{\mathrm{\underline{op}}}$, $k,\ell \in A$, and $y,z \in B^{*}$.
\end{theorem}
Here $B^{\mathrm{\underline{op}}}$, $A$ and $B^{*}$ are subalgebras of $B^{\mathrm{\underline{op}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*}$ and identifying $x=x\otimes 1\otimes 1$, $k=1\otimes k\otimes 1$ and $y=1\otimes 1\otimes y$ we have $xky \equiv x\otimes k \otimes y$. We also have algebra maps
\[ B^{*} \hookrightarrow B^{\mathrm{\underline{op}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*} \twoheadrightarrow B^{\mathrm{\underline{op}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A,\quad B^{\mathrm{\underline{op}}} \hookrightarrow B^{\mathrm{\underline{op}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*} \twoheadrightarrow A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*}\]
where the surjections are $\mathrm{id}\otimes \underline{\epsilon}$ and $\underline{\epsilon} \otimes \mathrm{id}$ respectively. It remains to prove Theorem~\ref{codbos}.
\begin{lemma}
The product stated in Theorem \ref{codbos} is associative.
\end{lemma}
\begin{proof} We expand the definition of the product to find
\begin{align*}
\Big((x\otimes &k \otimes y)(w \otimes \ell \otimes z)\Big)(m \otimes j \otimes v)\\
=& (x\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes k{\t}\ell{\o} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z)(m \otimes j \otimes v) \ \mathcal{R}(y^{\bar{\o}}, \ell{\t})\mathcal{R}(Sk{\o}, w^{\bar{\o}})\\
=& x\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} m^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\th}\ell{\t}j{\o} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}z^{\bar{{}_{\scriptscriptstyle(0)}}}v \ \mathcal{R}(y^{\bar{\o}}, \ell{\th})\mathcal{R}(Sk{\o}, w^{\bar{\o}})\\
&\quad\mathcal{R}(y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}z^{\bar{\o}},j{\t})\mathcal{R}(S(k{\t}\ell{\o}), m^{\bar{\o}})\\
=&x\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} m^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\th}\ell{\t}j{\o} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z^{\bar{{}_{\scriptscriptstyle(0)}}}v \ \mathcal{R}(y^{\bar{\o}}{\t}, \ell{\th})\mathcal{R}(Sk{\o}, w^{\bar{\o}})\\
&\quad\mathcal{R}(y^{\bar{\o}}{\o}, j{\t})\mathcal{R}(z^{\bar{\o}}, j{\th})\mathcal{R}(S\ell{\o}, m^{\bar{\o}}{\o})\mathcal{R}(Sk{\t}, m^{\bar{\o}}{\t}),
\end{align*}
where the last equality uses the right-coaction property on $y$. Similarly,
\begin{align*}
(x\otimes & k \otimes y)\Big((w \otimes \ell \otimes z)(m \otimes j \otimes v)\Big)\\
=&(x \otimes k \otimes y)(w\cdot_{\mathrm{\underline{op}}} m^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes \ell{\t}j{\o} \otimes z^{\bar{{}_{\scriptscriptstyle(0)}}}v) \ \mathcal{R}(z^{\bar{\o}}, j{\t})\mathcal{R}(S\ell{\o}, m^{\bar{\o}})\\
=&x\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} m^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\t}\ell{\t}j{\o} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z^{\bar{{}_{\scriptscriptstyle(0)}}}v \ \mathcal{R}(z^{\bar{\o}}, j{\th})\mathcal{R}(S\ell{\o}, m^{\bar{\o}})\\
&\quad\mathcal{R}(y^{\bar{\o}}, \ell{\th}j{\t})\mathcal{R}(Sk{\o}, w^{\o}m^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}),
\end{align*}
which by the left-coaction property on $m$ agrees with our first calculation.
\end{proof}
\begin{lemma}
The coproduct $\Delta$ stated in Theorem \ref{codbos} is an algebra map.
\end{lemma}
\begin{proof}
Expanding the product and then the coproduct, we have
\begin{align*}
\Delta&\Big((x\otimes k \otimes y)(w \otimes \ell \otimes z)\Big)\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}}{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}{\o}w^{\bar{{}_{\scriptscriptstyle(\infty)}}}{\underline{\t}}^{\bar{\o}}{{\o}}k{\t}\ell{\o}\otimes f^{a}\\
&\otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}}{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{{}_{\scriptscriptstyle(5)}}\ell{{}_{\scriptscriptstyle(4)}}y^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\o}}^{\bar{\o}}{\t}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\t}\\
&\otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}},x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}{\t}w^{\bar{{}_{\scriptscriptstyle(\infty)}}}{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar\o}{}\t k{\th}\ell{\t})\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}}\ell{\th}y^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\o}}^{\bar{\o}}{\o}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\o}),e_{a}\underline{\th}^{\bar{\o}}) \\
&\mathcal{R}(Sk{\o},w^{\bar{\o}})\mathcal{R}(y^{\bar{\o}},\ell{{}_{\scriptscriptstyle(5)}})\mathcal{R}(Sx{\underline{\t}}^{\bar{\o}},w^{\bar{{}_{\scriptscriptstyle(\infty)}}}{\underline{\o}}^{\bar{\o}})\mathcal{R}(y^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\t}}^{\bar{\o}},z{\underline{\o}}^{\bar{\o}}) \ \langle y^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\t} \rangle\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} w{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\t}}^{\bar{\o}}{\t}w{\underline{\t}}^{\bar{\o}}{\t}k{\t}\ell{\o}\otimes f^{a} \otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes k{{}_{\scriptscriptstyle(5)}}\ell{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\t}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \ \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\th}w{\underline{\t}}^{\bar{\o}}{\th}k{\th}\ell{\t})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}}\ell{\th}y{\underline{\o}}^{\bar{\o}}{\o}z{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}) \mathcal{R}(Sk{\o}, w{\underline{\o}}^{\bar{\o}}{\o}w{\underline{\t}}^{\bar{\o}}{\o})\mathcal{R}(y{\underline{\o}}^{\bar{\o}}{\th}y{\underline{\t}}^{\bar{\o}}{\t},\ell{{}_{\scriptscriptstyle(5)}})\\
&~~~~\mathcal{R}(Sx{\underline{\t}}^{\bar{\o}}{\o},w{\underline{\o}}^{\bar{\o}}{\t})\mathcal{R}(y{\underline{\t}}^{\bar{\o}}{\o},z{\underline{\o}}^{\bar{\o}}) \ \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\t} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\th} \rangle\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} w{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\t}}^{\bar{\o}}{\t}w{\underline{\t}}^{\bar{\o}}{\t}k{\th}\ell{\t}\otimes f^a \otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes k{{}_{\scriptscriptstyle(6)}}\ell{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\t}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \ \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\th}w{\underline{\t}}^{\bar{\o}}{\th}k{{}_{\scriptscriptstyle(4)}}\ell{\t})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(5)}}\ell{\th}y{\underline{\o}}^{\bar{\o}}{\o}z{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}) \mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{\o}k{\o}),w{\underline{\o}}^{\bar{\o}}) \mathcal{R}(Sk{\t}, w{\underline{\t}}^{\bar{\o}}{\o})\\
&\mathcal{R}(y{\underline{\o}}^{\bar{\o}}{\th},\ell{{}_{\scriptscriptstyle(5)}}) \mathcal{R}(y{\underline{\t}}^{\bar{\o}},\ell{{}_{\scriptscriptstyle(6)}}z{\underline{\o}}^{\bar{\o}})\ \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\t} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\th} \rangle
\end{align*}
for all $x,w \in B^{\mathrm{op}}$, $k,\ell \in A$, $y,z \in B^{*}$. The second equality uses the comodule coalgebra property (\ref{left-comod-coalg}) on $w$ and coassociativity. The last expression uses coquasitriangularity (\ref{quabic}) to gather the parts of $w{\underline{\o}}^{\bar{\o}}$ and $y{\underline{\t}}^{\bar{\o}}$ inside $\mathcal{R}$. On the other side,
\begin{align*}
\Delta&(x\otimes k \otimes y)\Delta(w \otimes \ell \otimes z)\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} w{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\t}}^{\bar{\o}}{\o\t}k{\o\t}w{\underline{\t}}^{\bar{\o}}{\o\o}\ell{\o\o}\otimes f^{a\bar{{}_{\scriptscriptstyle(0)}}}f^{b}\\
&\otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes k{{}_{\scriptscriptstyle(4)}\t}y{\underline{\o}}^{\bar{\o}}{\t\t}\ell{{}_{\scriptscriptstyle(4)}\o}z{\underline{\o}}^{\bar{\o}}{\t\o}\otimes y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \ \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}},x{\underline{\t}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(S(k{\th}y{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{\th}^{\o})\mathcal{R}(e_{b}\underline{\o}^{\bar{\o}},w{\underline{\t}}^{\bar{\o}}{\t}\ell{\t}) \mathcal{R}(S(\ell{\th}z{\underline{\o}}^{\bar{\o}}{\o}),e_{b}\underline{\th}^{\bar{\o}}) \\
&\mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{\o\o}k{\o\o}),w{\underline{\o}}^{\bar{\o}})\mathcal{R}(f^{a\bar{\o}},w{\underline{\t}}^{\bar{\o}}{\o\t}\ell{\o\t})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}\o}z{\underline{\o}}^{\bar{\o}}{\t\o}),e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}})\mathcal{R}(y{\underline{\t}}^{\bar{\o}},\ell{{}_{\scriptscriptstyle(4)}\t}z{\underline{\o}}^{\bar{\o}}{\t\t})\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\t} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{b}\underline{\t} \rangle\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} w{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes x{\underline{\t}}^{\bar{\o}}{\t}k{\t}w{\underline{\t}}^{\bar{\o}}{\o}\ell{\o}\otimes f^{a}f^{b}\\
&\otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(6)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{\th}\ell{{}_{\scriptscriptstyle(5)}}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \ \mathcal{R}(e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\th}k{\th})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\o}), e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}})\mathcal{R}(e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}},w{\underline{\t}}^{\bar{\o}}{\th}\ell{\th})\mathcal{R}(S(\ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{{}_{\scriptscriptstyle(6)}}^{\bar{\o}})\\
&\mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{\o}k{\o}),w{\underline{\o}}^{\bar{\o}})\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}e_{a}\underline{\t}^{\bar{\o}}e_{a}\underline{\th}^{\bar{\o}},w{\underline{\t}}^{\bar{\o}}{\t}\ell{\t})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(5)}}y{\underline{\o}}^{\bar{\o}}{\t}),e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}e_{a}{\underline{{}_{\scriptscriptstyle(6)}}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}})\mathcal{R}(y{\underline{\t}}^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(6)}}z{\underline{\o}}^{\bar{\o}}{\th})\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{{}_{\scriptscriptstyle(5)}} \rangle\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} w{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\t}}^{\bar{\o}}{\t}k{\t}w{\underline{\t}}^{\bar{\o}}{\o}\ell{\o}\otimes f^{a} \otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes k{{}_{\scriptscriptstyle(5)}}y{\underline{\o}}^{\bar{\o}}{\th}\ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \ \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}{\t}, x{\underline{\t}}^{\bar{\o}}{\th}k{\th})\mathcal{R}(S(\ell{\th}z{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o})\\
&\mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{\o}k{\o}),w{\underline{\o}}^{\bar{\o}}) \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}{\o}y{\underline{\o}}^{\bar{\o}}{\o}, w{\underline{\t}}^{\bar{\o}}{\t}\ell{\t})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\t}), w{\underline{\t}}^{\bar{\o}}{\th}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\t})\mathcal{R}(y{\underline{\t}}^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(5)}}z{\underline{\o}}^{\bar{\o}}{\th}) \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\t} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\th} \rangle\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} w{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes x{\underline{\t}}^{\bar{\o}}{\t}k{\t}w{\underline{\t}}^{\bar{\o}}{\o}\ell{\o}\otimes f^{a}\otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{\th}\ell{{}_{\scriptscriptstyle(5)}}z{\underline{\o}}^{\bar{\o}}{\t} \otimes y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}z{\t}\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\th}k{\th}w{\underline{\t}}^{\bar{\o}}{\t}\ell{\t}) \\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(5)}}y{\underline{\o}}^{\bar{\o}}{\t}\ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\o}), e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}})\mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{\o}k{\o}),w{\underline{\o}}^{\bar{\o}})\mathcal{R}(y{\underline{\t}}^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(6)}}z{\underline{\o}}^{\bar{\o}}{\th}) \\
&\mathcal{R}(y{\underline{\o}}^{\bar{\o}}{\o}, \ell{\th})\mathcal{R}(Sk{{}_{\scriptscriptstyle(4)}}, w{\underline{\t}}^{\bar{\o}}{\th})\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\t} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\th} \rangle,
\end{align*}
where the second equality uses duality $\langle f^{a\bar{{}_{\scriptscriptstyle(0)}}}, e_{a} \rangle f^{a\bar{\o}}=\langle f^{a}, e_{a}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle e_{a}^{\bar{\o}}$ followed by the comodule coalgebra property (\ref{left-comod-coalg}) on $e_{a}$. The third equality cancels $(\bar{S}e_{a}\underline{\th}\cdot_{\mathrm{\underline{op}}} e_{a}\underline{{}_{\scriptscriptstyle(4)}})^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}$ making all subsequent coactions trivial. The fourth equality uses (\ref{quabic}) to gather the parts of $e_{a}\underline{\o}^{\bar{\o}}$ and $e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}$ inside ${\mathcal R}$, and cancels some ${\mathcal R}$s. In the final expression, one can use quasicommutativity (\ref{quacom}) to reorder the second tensor factor so as to coincide with the result of the first calculation.
\end{proof}
\begin{lemma}
The coproduct $\Delta$ stated in Theorem \ref{codbos} is coassociative.
\end{lemma}
\begin{proof} We expand the definition of the coproduct to find
\begin{align*}
(\mathrm{id}&\otimes \Delta)\Delta(x\otimes k \otimes y)\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{a} \otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes e_{a}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\th}k{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\t}\otimes f^{b}\\
&\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{{}_{\scriptscriptstyle(7)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}y{\underline{\t}}^{\bar{\o}}{\t} \otimes y{\underline{\th}}\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\th}\rangle \langle y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{b}\underline{\t} \rangle \ \mathcal{R}(e_{b}\underline{\o}^{\bar{\o}},e_{a}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(5)}}y{\underline{\o}}^{\bar{\o}}{\th}) \\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}y{\underline{\t}}^{\bar{\o}}{\o}),e_{b}\underline{\th}^{\bar{\o}})\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}e_{a}\underline{\t}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(S(k{\th}y{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o}e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}}{\o})\mathcal{R}(S(e_{a}\underline{\t}^{\bar{\o}}{\th}x{\underline{\th}}^{\bar{\o}}{\th}),e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}}{\th})\\
&\mathcal{R}(Se_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\t},e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}}{\t})\mathcal{R}(Se_{a}\underline{\t}^{\bar{\o}}{\t},x{\underline{\t}}^{\bar{\o}}{\th})\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{a} \otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes e_{a}\underline{\t}^{\bar{\o}}{\th}x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\th}k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{\th}\otimes f^{b}\\
&\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{{}_{\scriptscriptstyle(9)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}y{\underline{\t}}^{\bar{\o}}{\t}\otimes y{\underline{\th}}\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\th} \rangle \langle y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{b}\underline{\t} \rangle \ \mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}, e_{a}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(7)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(8)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}y{\underline{\t}}^{\bar{\o}}{\o}),e_{b}\underline{\th}^{\bar{\o}})\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(e_{a}\underline{\t}^{\bar{\o}}{\o}, x{\underline{\th}}^{\bar{\o}}{\th}k{\th})\mathcal{R}(S(e_{a}\underline{\t}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\t}k{{}_{\scriptscriptstyle(5)}}y{\underline{\o}}^{\bar{\o}}{\t}),e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o})\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{a} \otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}e_{a}\underline{\t}^{\bar{\o}}{\th}k{{}_{\scriptscriptstyle(5)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\th}y{\underline{\o}}^{\bar{\o}}{\th} \otimes f^{b}\\
&\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{{}_{\scriptscriptstyle(9)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}y{\underline{\t}}^{\bar{\o}}{\t}\otimes y{\underline{\th}}\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\th} \rangle \langle y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\t} \rangle \ \mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}e_{a}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(6)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(8)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}y{\underline{\t}}^{\bar{\o}}{\o}),e_{b}\underline{\th}^{\bar{\o}}) \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}},x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(S(x{\underline{\th}}^{\bar{\o}}{\th}e_{a}\underline{\t}^{\bar{\o}}{\t}k{{}_{\scriptscriptstyle(4)}}e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\t}y{\underline{\o}}^{\bar{\o}}{\t}),e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}})\mathcal{R}(e_{a}\underline{\t}^{\bar{\o}}{\o},k{\th})\\
&\mathcal{R}(e_{a}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}})\mathcal{R}(Sy{\underline{\o}}^{\bar{\o}}{\o},e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o})\mathcal{R}(Sk{{}_{\scriptscriptstyle(7)}},e_{a}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}),
\end{align*}
where the second equality uses (\ref{quabic}) to gather the parts of $e_{a}\underline{\o}^{\bar{\o}}$ and $e_{a}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}}$, cancelling some of the ${\mathcal R}$s. We lastly use (\ref{quacom}) to change the order in the fifth tensor factor and in a similar term inside $\mathcal{R}$, again cancelling some of the ${\mathcal R}$s. On the other side,
\begin{align*}
(\Delta &\otimes \mathrm{id})\Delta(x\otimes k \otimes y)\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{b}\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(4)}}f^{a}{\underline{\o}}^{\bar{\o}}{\t}\\
&\otimes f^{a}{\underline{\t}}\otimes e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{{}_{\scriptscriptstyle(7)}}y{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\underline{\t}} \ \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\t} \rangle \langle f^{a}{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\t} \rangle\\
&\mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\mathcal{R}(S(x{\underline{\th}}^{\bar{\o}}{\th}k{\th}f^{a}{\underline{\o}}^{\bar{\o}}{\o}),e_{b}\underline{\th}^{\bar{\o}})\\
&\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}k{{}_{\scriptscriptstyle(5)}})\mathcal{R}(S(k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{}_{\o}),e_{a}\underline{\th}^{\bar{\o}})\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{b}\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(4)}}e_{b}\underline{\t}^{\bar{\o}}{\t}\\
&\otimes f^{c} \otimes (e_{c}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}^{\bar{{}_{\scriptscriptstyle(\infty)}}}){\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\bar{S}(e_{c}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}^{\bar{{}_{\scriptscriptstyle(\infty)}}}){\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{{}_{\scriptscriptstyle(7)}}y{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\underline{\t}}\\
&\langle f^{a},e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, (e_{c}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}^{\bar{{}_{\scriptscriptstyle(\infty)}}}){\underline{\t}} \rangle\ \mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(S(x{\underline{\th}}^{\bar{\o}}{\th}k{\th}e_{b}\underline{\t}^{\bar{\o}}{\o}),e_{b}\underline{\th}^{\bar{\o}})\mathcal{R}((e_{c}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}^{\bar{{}_{\scriptscriptstyle(\infty)}}}){\underline{\o}}^{\bar{\o}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}k{{}_{\scriptscriptstyle(5)}})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{\o}),(e_{c}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}^{\bar{{}_{\scriptscriptstyle(\infty)}}}){\underline{\th}}^{\bar{\o}}){\mathcal R}(e_c ^{\bar{\o}}, e_a ^{\bar{\o}})\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{b}\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(4)}}e_{b}\underline{\t}^{\bar{\o}}{\t}e_{b}\underline{\th}^{\bar{\o}}{\t}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\t}\otimes f^{c}\\
&\otimes e_{c}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}(e_{c}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}})\otimes k{{}_{\scriptscriptstyle(7)}}y{\underline{\o}}^{\bar{\o}}{\t}\otimes y{\t}\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{c}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \ \mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(S(x{\underline{\th}}^{\bar{\o}}{\th}k{\th}e_{b}\underline{\t}^{\bar{\o}}{\o}e_{b}\underline{\th}^{\bar{\o}}{\o}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o}), e_{b}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}})\mathcal{R}(e_{c}\underline{\o}^{\bar{\o}}{\t}e_{b}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}k{{}_{\scriptscriptstyle(5)}})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{\o}),e_{c}\underline{\th}^{\bar{\o}}{\th}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(e_{c}\underline{\o}^{\bar{\o}}{\o}e_{c}\underline{\t}^{\bar{\o}}{\o}e_{c}\underline{\th}^{\bar{\o}}{\o},e_{b}\underline{\t}^{\bar{\o}}{\th}e_{b}\underline{\th}^{\bar{\o}}{\th}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\th})\\
&\mathcal{R}(Se_{c}\underline{\th}^{\bar{\o}}{\t} e_{b}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}e_{b}\underline{\th}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(Se_{c}\underline{\t}^{\bar{\o}}{\t},e_{b}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}).
\end{align*}
For the second equality we use duality $\langle f^{a}{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{b}\underline{\t} \rangle f^{a}{\underline{\o}}^{\bar{\o}} = \langle f^{a}{\underline{\o}}, e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle e_{b}\underline{\t}^{\bar{\o}}$ to replace $f^{a}{\underline{\o}}^{\bar{\o}}$ by $e_{b}\underline{\t}^{\bar{\o}}$, followed by
\[ e_{a}\otimes f^{a}{\underline{\o}}\otimes f^{a}{\underline{\t}} = e_{c}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{a}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes f^{a}\otimes f^{c} \ {\mathcal R}(e_c ^{\bar{\o}}, e_a ^{\bar{\o}})\]
to replace $f^{a}{\underline{\o}}\otimes f^{a}{\underline{\t}}$ by $f^{a}\otimes f^{c}$. For the third equality, we use $\langle f^{a},e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle$ to replace $e_{a}$ by $e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}$, after which we expand $(e_{c}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}})\underline{\o}$ etc. using $\underline\Delta$ a braided-homomorphism. In the last expression, we expand $\bar{S}$ of a $\cdot_{\mathrm{\underline{op}}}$ product and use
\begin{align*}
\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{c}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle=\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\o}}, e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\t}}, e_{c}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle {\mathcal R}(Se_{c}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}, e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}}).
\end{align*}
By the comodule coalgebra property (\ref{right-comod-coalg}), the first pairing on the right becomes $\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\langle y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{c}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle$ and duality $\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle e_{b}\underline{\th}^{\bar{\o}}=\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\th} \rangle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}$ replaces $e_{b}\underline{\th}^{\bar{\o}}$ by $y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}$. The other pairing similarly replaces $e_{c}\underline{\t}^{\bar{\o}}$ by $y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}$, so
\begin{align*}
(\Delta &\otimes \mathrm{id})\Delta(x\otimes k \otimes y)\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{b}\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(4)}}e_{b}\underline{\t}^{\bar{\o}}{\t}y{\underline{\o}}^{\bar{\o}}{\t}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\t}\otimes f^{c}\\
&\otimes e_{c}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}}\bar{S}e_{c}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes k{{}_{\scriptscriptstyle(7)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(7)}}y{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}\otimes y{\th}\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\th} \rangle \langle y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{c}\underline{\t} \rangle \ \mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(S(x{\underline{\th}}^{\bar{\o}}{\th}k{\th}e_{b}\underline{\t}^{\bar{\o}}{\o}y{\underline{\o}}^{\bar{\o}}{\o}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o}),e_{b}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}})\mathcal{R}(e_{c}\underline{\o}^{\bar{\o}}{\t}e_{b}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}k{{}_{\scriptscriptstyle(5)}})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(6)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}y{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}),e_{c}\underline{\th}^{\bar{\o}}{\th}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(e_{c}\underline{\o}^{\bar{\o}}{\o}y{\underline{\t}}^{\bar{\o}}{\o}e_{c}\underline{\t}^{\bar{\o}}{\o}, e_{b}\underline{\t}^{\bar{\o}}{\th}y{\underline{\o}}^{\bar{\o}}{\th}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o})\\
&\mathcal{R}(Se_{c}\underline{\th}^{\bar{\o}}{\t},e_{b}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(Sy{\underline{\t}}^{\bar{\o}}{\t},e_{b}\underline{\t}^{\bar{\o}}{}{{}_{\scriptscriptstyle(5)}})\mathcal{R}(Sy{\underline{\t}}^{\bar{\o}}{\th},y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}})\mathcal{R}(Se_{c}\underline{\th}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}},e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}})\\
=&x{\underline{\o}}\otimes x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o}k{\o}\otimes f^{b}\otimes e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
&\otimes x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}k{{}_{\scriptscriptstyle(4)}}e_{b}\underline{\t}^{\bar{\o}}{\t}y{\underline{\o}}^{\bar{\o}}{\t}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\t}\otimes f^{c}\\
&\otimes e_{c}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}}\bar{S}e_{c}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes k{{}_{\scriptscriptstyle(9)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}y{\underline{\t}}^{\bar{\o}}{\t}\otimes y{\th}\\
&\langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\th} \rangle \langle y{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{c}\underline{\t} \rangle \ \mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(S(x{\underline{\th}}^{\bar{\o}}{\th}k{\th}e_{b}\underline{\t}^{\bar{\o}}{\o}y{\underline{\o}}^{\bar{\o}}{\o}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\o}),e_{b}\underline{{}_{\scriptscriptstyle(5)}}^{\bar{\o}})\mathcal{R}(S(k{{}_{\scriptscriptstyle(8)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}y{\underline{\t}}^{\bar{\o}}{\o}),e_{c}\underline{\th}^{\bar{\o}})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(7)}}y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}),e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(e_{c}\underline{\o}^{\bar{\o}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}k{{}_{\scriptscriptstyle(5)}}e_{b}\underline{\t}^{\bar{\o}}{\th}y{\underline{\o}}^{\bar{\o}}{\th}e_{b}\underline{{}_{\scriptscriptstyle(4)}}^{\bar{\o}}{\th})\\
&\mathcal{R}(e_{b}\underline{\t}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}, x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}k{{}_{\scriptscriptstyle(6)}}),
\end{align*}
using (\ref{quabic}) to gather $e_{c}\underline{\o}^{\bar{\o}}$ and $e_{c}\underline{\th}^{\bar{\o}}$, and cancelling some $\mathcal{R}$s. In the final expression one can use (\ref{quacom}) to change the order in the fifth tensor factor as well as inside $\mathcal{R}$, to recover our calculation of $(\mathrm{id}\otimes \Delta)\Delta(x\otimes k \otimes y)$ up to a change of basis labels. \end{proof}
\begin{lemma}\label{antipode}
The antipode of $B^{\mathrm{op}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*}$ in Theorem \ref{codbos} is given by
\begin{align*}
S(x\otimes k \otimes y)=&\bar{S}(e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}) \otimes S(x^{\bar{\o}}{\t}k{\t}f^{a\bar{\o}}{\th})\otimes \underline{S}f^{a\bar{{}_{\scriptscriptstyle(0)}}}\\
&\mathcal{R}(f^{a\bar{\o}}{\o},S(x^{\bar{\o}}{\o}k{\o})) \mathcal{R}(S^{2}(k{{}_{\scriptscriptstyle(5)}}y^{\bar{\o}}{\t}),e_{a}\underline{\o}^{\bar{\o}}{\th}x^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}e_{a}\underline{\th}^{\bar{\o}}{\th})\\
&\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}{\o}, x^{\bar{\o}}{\th}k{\th})\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}}y^{\bar{\o}}{\o}), e_{a}\underline{\th}^{\bar{\o}}{\o}) \ \langle y^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\t} \rangle\\
&v(f^{a\bar{\o}}{\t})\bar{u}(e_{a}\underline{\o}^{\bar{\o}}{\t}x^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}e_{a}\underline{\th}^{\bar{\o}}{\t}),
\end{align*}
where $v(k)=\mathcal{R}(k{\o},Sk{\t})$ and $\bar{u}(k)=\mathcal{R}(S^{2}k{\o},k{\t})$.
\end{lemma}
\begin{proof}
We first compute $(S(x\otimes k \otimes y){\o})(x\otimes k \otimes y){\t}$, which on expanding out the product has in the first tensor factor
\begin{align*}
\bar{S}(e_{b}&\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}})\cdot_{\mathrm{\underline{op}}} e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\\
=&(\underline{\epsilon}(e_{a}\underline{\o})\underline{\epsilon}(e_{a}\underline{\th})\underline{\epsilon}(x)\underline{\epsilon}(e_{b}\underline{\o})\underline{\epsilon}(e_{b}\underline{\th}))^{\bar{{}_{\scriptscriptstyle(\infty)}}},
\end{align*}
which further collapses the full expression to give
\begin{align*}
(S(x\otimes & k \otimes y){\o})(x\otimes k \otimes y){\t}
=\underline{\epsilon}(x)\otimes S(k{\t}f^{b\bar{\o}}{{}_{\scriptscriptstyle(4)}})k{\th}y{\underline{\o}}^{\bar{\o}}{\o}\otimes (\underline{S}f^{b\bar{{}_{\scriptscriptstyle(0)}}})y{\underline{\t}}\\
&\quad\quad\langle f^{a},e_{b} \rangle \langle y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a} \rangle \ \mathcal{R}(f^{b\bar{\o}}{\t},Sk{\o})\mathcal{R}(f^{b\bar{\o}}{\o},k{{}_{\scriptscriptstyle(4)}}y{\underline{\o}}^{\bar{\o}}{\t})v(f^{b\bar{\o}}{\th})\\
=&\underline{\epsilon}(x) \otimes (Sy{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{{}_{\scriptscriptstyle(5)}})(Sk{\t})k{\th}y{\underline{\o}}^{\bar{\o}}{\o}\otimes (\underline{S}y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}})y{\underline{\t}}\\
&\mathcal{R}(y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\t}, Sk{\o})\mathcal{R}(y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\o}, y{\underline{\o}}^{\bar{\o}}{\t})\mathcal{R}(y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\t},k{{}_{\scriptscriptstyle(4)}})v(y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{{}_{\scriptscriptstyle(4)}})\\
=&\underline{\epsilon}(x) \otimes \epsilon(k)(Sy{\underline{\o}}^{\bar{\o}}{\th})y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}\otimes (\underline{S}y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}})y{\underline{\t}} \ \mathcal{R}(y{\underline{\o}}^{\bar{\o}}{\o}, y{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}})v(y{\underline{\o}}^{\bar{\o}}{\t})\\
=&\underline{\epsilon}(x)\otimes \epsilon(k) \otimes (\underline{S}y{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}})y{\underline{\t}} \ \mathcal{R}(y{\underline{\o}}^{\bar{\o}}{\o}, y{\underline{\o}}^{\bar{\o}}{\th})v(y{\underline{\o}}^{\bar{\o}}{\t})\\
=&\underline{\epsilon}(x)\otimes \epsilon(k)\otimes \underline{\epsilon}(y)=\epsilon(x\otimes k \otimes y).
\end{align*}
Similarly, on computing $(x\otimes k \otimes y){\o}(S(x\otimes k \otimes y){\t})$ we have $f^{a\bar{{}_{\scriptscriptstyle(0)}}}\underline{S}f^{b\bar{{}_{\scriptscriptstyle(0)}}}=(\underline{\epsilon}f^{a}\underline{\epsilon}f^{b})^{\bar{{}_{\scriptscriptstyle(0)}}}$ in the third tensor factor which collapses the expressions to give \begin{align*}
(x\otimes &k \otimes y){\o}(S(x\otimes k \otimes y){\t})
=x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} \bar{S}x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes x{\underline{\t}}^{\bar{\o}}{\t}k{\t}S(x{\underline{\t}}^{\bar{\o}}{\th}k{\th}) \otimes \underline{\epsilon}y\\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\mathcal{R}(S^{2}k{{}_{\scriptscriptstyle(4)}}, x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}})\mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{\o}k{\o}), x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}})\bar{u}(x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} \bar{S}x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes x{\underline{\t}}^{\bar{\o}}{\t}k{\t}Sk{\th}Sx{\underline{\t}}^{\bar{\o}}{\th}\otimes \underline{\epsilon}y \ \mathcal{R}(S^{2}k_{\t}, x{\underline{\t}}^{\bar{\o}}{\th})\\
&\mathcal{R}(Sk{\o}Sx{\underline{\t}}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(6)}}) \bar{u}(x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} \bar{S}x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes 1 \otimes \underline{\epsilon}y \ \mathcal{R}(S^{2}k{\t}, x{\underline{\t}}^{\bar{\o}}{\th})\\
&\mathcal{R}(Sk{\o},x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(Sx{\underline{\t}}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}})\bar{u}(x{\underline{\t}}^{\bar{\o}}{\t})\\
=&x{\underline{\o}}\cdot_{\mathrm{\underline{op}}} \bar{S}x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes \epsilon k \otimes \underline{\epsilon}y \ \mathcal{R}(Sx{\underline{\t}}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{\th})\bar{u}(x{\underline{\t}}^{\bar{\o}}{\t})
=\epsilon(x\otimes k \otimes y).
\end{align*}
\end{proof}
Finally, we show that the co-double bosonisation is coquasitriangular so as to have an inductive construction of such Hopf algebras.
\begin{proposition}\label{codouble coquasitriangularity}
The co-double bosonisation $B^{\underline{\mathrm{op}}} {>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^{*}$ is coquasitriangular with
\begin{align*}
\mathcal{R}(x\otimes k \otimes y, w \otimes \ell \otimes z)=& \langle \underline{S}z^{\bar{{}_{\scriptscriptstyle(0)}}}, x \rangle \mathcal{R}(k, \ell z^{\bar{\o}})\underline{\epsilon}(y)\underline{\epsilon}(w)
\end{align*}
for all $x,w \otimes B^{\underline{\mathrm{op}}}$, $k,\ell \in A$, and $y,z \in B^{*}$.
\end{proposition}
\begin{proof} (i) Expanding the definitions of the product and the coquasitriangular structure,
\begin{align*}
\mathcal{R}\Big((m\otimes j &\otimes v), (x\otimes k \otimes y)(w \otimes \ell \otimes z)\Big)\\
=&\langle \underline{S}(y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}z^{\bar{{}_{\scriptscriptstyle(0)}}}), m \rangle \mathcal{R}(j, k\ell{\o}y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}z^{\bar{\o}})\mathcal{R}(y^{\bar{\o}}, \ell{\t})\\
=&\langle \underline{S}(y^{\bar{{}_{\scriptscriptstyle(0)}}}z^{\bar{{}_{\scriptscriptstyle(0)}}}),m\rangle \mathcal{R}(j,k\ell{\o}y^{\bar{\o}}{\o}z^{\bar{{}_{\scriptscriptstyle(0)}}}) \mathcal{R}(y^{\bar{\o}}{\t},\ell_{\t})\\
=&\langle \underline{S}z^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}\underline{S}y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, m \rangle \mathcal{R}(j,k\ell{\o}y^{\bar{\o}}{\o}z^{\bar{{}_{\scriptscriptstyle(0)}}}) \mathcal{R}(y^{\bar{\o}}{\t},\ell{\t}) \mathcal{R}(y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}, z^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}})\\
=&\langle \underline{S}z^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\o}} \rangle \langle \underline{S}y^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\t}} \rangle \mathcal{R}(j,k\ell{\o}y^{\bar{\o}}{\t}z^{\bar{\o}}{\t})\mathcal{R}(y^{\bar{\o}}{\th},\ell{\t})\mathcal{R}(y^{\bar{\o}}{\o}, z^{\bar{\o}}{\o})\\
=&\langle \underline{S}z^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\o}} \rangle \langle \underline{S}y^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\t}} \rangle \mathcal{R}(j, ky^{\bar{\o}}{\th}\ell{\t}z^{\bar{\o}}{\t})\mathcal{R}(y^{\bar{\o}}, z^{\bar{\o}}{\o})\mathcal{R}(y^{\bar{\o}}{\t}, \ell{\o}).
\end{align*}
The second equality uses the right coaction on $y$. The third equality expands the braided-antipode $\underline{S}(y^{\bar{{}_{\scriptscriptstyle(0)}}}z^{\bar{{}_{\scriptscriptstyle(0)}}})$. The fourth equality uses the right-coaction on $y$ and $z$, and evaluation. The last equality is quasicommutativity to change the order of product inside the first ${\mathcal R}$. On the other side,
\begin{align*}
\mathcal{R}\Big((m\otimes & j \otimes v){\o}, w \otimes \ell \otimes z\Big)\mathcal{R}\Big((m\otimes j \otimes v){\t}, x \otimes k \otimes y\Big)\\
=&\mathcal{R}(m{\underline{\o}}\otimes m{\underline{\t}}^{\bar{\o}}{\o}j{\o} \otimes f^{a}, w \otimes \ell \otimes z)\\
&\mathcal{R}(e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} m{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes j{{}_{\scriptscriptstyle(4)}}v{\underline{\o}}^{\bar{\o}}{\t}\otimes v{\t}, x \otimes k \otimes y)\\
& \mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, m{\underline{\t}}^{\bar{\o}}{\t}j{\t})\mathcal{R}(S(j{\th}v{\underline{\o}}^{\bar{\o}}{\o}),e_{a}\underline{\th}^{\bar{\o}}) \langle v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\t} \rangle\\
=&\langle \underline{S}z^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\o}} \rangle \langle \underline{S}y^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \mathcal{R}(m{\underline{\t}}^{\bar{\o}}j{\o}, \ell z^{\bar{\o}})\mathcal{R}(j{\t}, k y^{\bar{\o}})\\
=&\langle \underline{S}z^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\o}} \rangle \langle \underline{S}y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\t}} \rangle \mathcal{R}(y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}j{\o}, \ell z^{\bar{\o}})\mathcal{R}(j{\t}, k y^{\bar{\o}})\\
=&\langle \underline{S}z^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\o}} \rangle \langle \underline{S}y^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\t}} \rangle\mathcal{R}(y^{\bar{\o}}{\o}j{\o}, \ell z^{\bar{\o}})\mathcal{R}(j{\t}, k y^{\bar{\o}}{\t}).
\end{align*}
The third equality uses $
\langle y^{\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\rangle m{\underline{\t}}^{\bar{\o}} = \langle y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, m{\underline{\t}} \rangle y^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}$ and the fourth uses the right coaction on $y$. We can then use (\ref{quacom}) to gather the parts of $j$ and obtain the same expression as on the first side. (ii) Similarly expanding the definitions,
\begin{align*}
\mathcal{R}\Big((x\otimes & k \otimes y)(w\otimes \ell \otimes z), (m \otimes j \otimes v)\Big)\\
=&\langle \underline{S}v^{\bar{{}_{\scriptscriptstyle(0)}}}, x\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \mathcal{R}(k{\t}\ell, jv^{\bar{\o}})\mathcal{R}(Sk{\o}, w^{\bar{\o}})\\
=&\langle \underline{S}v^{\bar{{}_{\scriptscriptstyle(0)}}}, w^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{{}_{\scriptscriptstyle(\infty)}}}x^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \mathcal{R}(k{\t}\ell,jv^{\bar{\o}})\mathcal{R}(Sk{\o},w^{\bar{\o}})\mathcal{R}(Sx^{\bar{\o}},w^{\bar{{}_{\scriptscriptstyle(\infty)}}\bar{\o}})\\
=&\langle \underline{S}v^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle \underline{S}v^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, w^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \mathcal{R}(k{\t}\ell, jv^{\bar{\o}}) \mathcal{R}(Sk{\o}, w^{\bar{\o}}{\o})\\
&\mathcal{R}(Sx^{\bar{\o}}, w^{\bar{\o}}{\t}) \mathcal{R}(v^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\o}}^{\bar{\o}},v^{\bar{{}_{\scriptscriptstyle(0)}}}{\underline{\t}}^{\bar{\o}})\\
=&\langle \underline{S}v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle \underline{S}v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, w^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \mathcal{R}(k{\t}\ell, jv{\underline{\o}}^{\bar{\o}}{\t}v{\underline{\t}}^{\bar{\o}}{\t})\mathcal{R}(Sk{\o}, w^{\bar{\o}}{\o})\\
&\mathcal{R}(Sx^{\bar{\o}}, w^{\bar{\o}}{\t})\mathcal{R}(v{\underline{\o}}^{\bar{\o}}{\o},v{\underline{\t}}^{\bar{\o}}{\o})\\
=&\langle \underline{S}v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, x \rangle \langle \underline{S}v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, w \rangle \mathcal{R}(k{\t}\ell, jv{\underline{\o}}^{\bar{\o}}{\t}v{\underline{\t}}^{\bar{\o}}{\t})\mathcal{R}(Sk{\o}, v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\o}) \\
&\mathcal{R}(Sv{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}, v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\t})\mathcal{R}(v{\underline{\o}}^{\bar{\o}}{\o},v{\underline{\t}}^{\bar{\o}}{\o})\\
=&\langle \underline{S}v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x \rangle \langle \underline{S}v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, w \rangle\mathcal{R}(k{\t}\ell, jv{\underline{\o}}^{\bar{\o}}{\th}v{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}) \mathcal{R}(Sk{\o}, v{\underline{\t}}^{\bar{\o}}{\o})\\
&\mathcal{R}(Sv{\underline{\o}}^{\bar{\o}}{\o}, v{\underline{\t}}^{\bar{\o}}{\t})\mathcal{R}(v{\underline{\o}}^{\bar{\o}}{\t}, v{\underline{\t}}^{\bar{\o}}{\th})\\
=&\langle \underline{S}v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}},x \rangle \langle \underline{S}v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, w \rangle \mathcal{R}(k, j{\o}v{\underline{\o}}^{\bar{\o}}{\o})\mathcal{R}(\ell, j{\t}v{\underline{\o}}^{\bar{\o}}{\t}v{\underline{\t}}^{\bar{\o}}).
\end{align*}
The second equality expands the braided-product $\cdot_{\mathrm{\underline{op}}}$. The third equality uses the left-coaction on $w$, followed by the duality pairing and taking $\underline S$ to the left in $\underline{\Delta}(\underline{S}v^{\bar{{}_{\scriptscriptstyle(0)}}})$. The fourth equality uses the comodule coalgebra property (\ref{right-comod-coalg}) on $v$ and the right coaction axioms. The fifth equality moves the coactions onto $x,w$ by duality. The sixth equality is similar to the fourth. For the last equality we cancel the last two $\mathcal{R}$s and use (\ref{quabic}) to gather $k$ inside ${\mathcal R}$ and cancel further. On the other side,
\begin{align*}
\mathcal{R}\Big((x&\otimes k \otimes y), (m\otimes j \otimes v){\o}\Big)\mathcal{R}\Big((w \otimes \ell \otimes z),(m\otimes j \otimes v){\t} \Big)\\
=& \langle \underline{S}f^{a\bar{{}_{\scriptscriptstyle(0)}}}, x \rangle \langle \underline{S}v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, w \rangle \langle v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a} \rangle \mathcal{R}(k, j{\underline{\o}}f^{a\underline{\o}}) \mathcal{R}(\ell, j{\t}v{\underline{\o}}^{\bar{\o}}v{\underline{\t}}^{\bar{\o}})\\
=&\langle \underline{S}v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, x \rangle \langle \underline{S}v{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, w \rangle \mathcal{R}(k,j{\o}v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}) \mathcal{R}(\ell, j{\t}v{\underline{\o}}^{\bar{\o}}v{\underline{\t}}^{\bar{\o}})
\end{align*}
on substituting $f^{a}=v{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}$. We can then use the right coaction property on $v{\underline{\o}}$ to recover the result of our first calculation. (iii) We expand the definitions to compute
\begin{align*}
(x\otimes& k \otimes y){\t}(w \otimes \ell \otimes z){\t} \mathcal{R}\Big((x \otimes k \otimes y){\o},(w \otimes \ell \otimes z){\t}\Big)\\
=& x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_{b}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{b}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes k{\th}\ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\t} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}} z{\underline{\t}}\\
&\mathcal{R}(Sk{\t}, e_{b}\underline{\o}^{\bar{\o}}{\t}w^{\bar{\o}}{\th}e_{b}\underline{\th}^{\bar{\o}}{\t}) \mathcal{R}(y^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(5)}}z{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(e_{b}\underline{\o}^{\bar{\o}}{\o}, w^{\bar{\o}}{\t}\ell{\t})\\
&\mathcal{R}(S(\ell{\th}z{\underline{\o}}^{\bar{\o}}{\o}), e_{b}\underline{\th}^{\bar{\o}}{\o}) \mathcal{R}(x{\underline{\t}}^{\bar{\o}}k{\o}, w^{\bar{\o}}{\o}\ell{\o}f^{b\bar{\o}}) \langle \underline{S}f^{b\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\o}} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b}\underline{\t} \rangle\\
=&x{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}^{-1}x{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} w^{\bar{{}_{\scriptscriptstyle(\infty)}}} \cdot_{\mathrm{\underline{op}}} \bar{S}\bar{S}^{-1}x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\underline{\th}}\ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}}\\
&\mathcal{R}(Sk{\t}, x{\th}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}w^{\bar{\o}}{\th}x{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(y^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(5)}}z{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(x{\underline{\th}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}, w^{\bar{\o}}{\t}\ell{\t})\\
&\mathcal{R}(S(\ell{\th}z{\underline{\o}}^{\bar{\o}}{\o}), x{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(x{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{\o}}k{\o}, w^{\bar{\o}}{\o}\ell{\o}x{\underline{\o}}^{\bar{\o}}{\o}x{\underline{\t}}^{\bar{\o}}{\o}x{\underline{\th}}^{\bar{\o}}{\o})\\
&\mathcal{R}(x{\underline{\t}}^{\bar{\o}}{\t}x{\underline{\th}}^{\bar{\o}}{\t}, x{\underline{\o}}^{\bar{\o}}{\t})\mathcal{R}(x{\underline{\th}}^{\bar{\o}}{\th}, x{\underline{\t}}^{\bar{\o}}{\th}) \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, \bar{S}^{-1} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\\
=&w^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\th}\ell{\th}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \\
&\mathcal{R}(Sk{\t}, w^{\bar{\o}}{\t}x{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(y^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(S(\ell{\t}z{\underline{\o}}^{\bar{\o}}{\o}), x{\underline{\o}}^{\bar{\o}}{\th})\\
&\mathcal{R}(k{\o},w^{\bar{\o}}{\o}\ell{\o}x{\underline{\o}}^{\bar{\o}}{\o}x{\underline{\t}}^{\bar{\o}}{\o})\mathcal{R}(x{\underline{\t}}^{\bar{\o}}{\t}, x{\underline{\o}}^{\bar{\o}}{\t}) \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, \bar{S}^{-1}x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\\
=&w^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\th}\ell{\th}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}}\\
&\mathcal{R}(Sk{\t}, w^{\bar{\o}}{\t}x{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(y^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(S(\ell{\t}z{\underline{\o}}^{\bar{\o}}{\o})x{\underline{\o}}^{\bar{\o}}{\th})\\
&\mathcal{R}(k{\o}, w^{\bar{\o}}{\o}\ell{\o}x{\underline{\o}}^{\bar{\o}}{\o}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\o})\mathcal{R}(z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}{\t}, x{\underline{\o}}^{\bar{\o}}{\t})\langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}} \rangle\\
=&w\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\t}\ell{\th}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}} \langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}} \rangle\\
&\mathcal{R}(Sk{\t}, x{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(y^{\bar{\o}},\ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(S\ell{\t}, x{\underline{\o}}^{\bar{\o}}{\t}) \mathcal{R}(k{\o}, \ell{\o}x{\underline{\o}}^{\bar{\o}}{\o}z{\underline{\o}}^{\bar{\o}}{\o})\\
=&w\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes k{\t}\ell{\th}z{\underline{\o}}^{\bar{\o}}{\t}\otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}}\ \langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}} \rangle\\
&\mathcal{R}(y^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(S\ell{\o}, x{\underline{\o}}^{\bar{\o}})\mathcal{R}(k{\o}, \ell{\t}z{\underline{\o}}^{\bar{\o}}{\o})\\
=&w\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes \ell{\t}z{\underline{\o}}^{\bar{\o}}{\o}k{\o} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z \ \langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}} \rangle\\
&\mathcal{R}(y^{\bar{\o}}, \ell{{}_{\scriptscriptstyle(4)}}z{\underline{\o}}^{\bar{\o}}{\th})\mathcal{R}(S\ell{\o},x{\underline{\o}}^{\bar{\o}})\mathcal{R}(k{\t}, \ell{\th}z{\underline{\o}}^{\bar{\o}}{\t}).
\end{align*}
The second equality uses the duality $\langle \underline{S}f^{b\bar{{}_{\scriptscriptstyle(0)}}}, x_{\underline{\o}} \rangle f^{b\bar{\o}} = \langle f^{b}, \bar{S}^{-1}x_{\underline{\o}}^{~~\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle x_{\underline{\o}}^{~~\bar{\o}}$ to substitute $e_{b}=\bar{S}^{-1}x_{\underline{\o}}^{~~\bar{{}_{\scriptscriptstyle(\infty)}}}$ in all the places where it occurs, and the comodule algebra property (\ref{left-comod-coalg}). The third equality cancels $(x_{\underline{{}_{\scriptscriptstyle(4)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}^{-1}x_{\underline{\th}})^{\bar{{}_{\scriptscriptstyle(\infty)}}}$ resulting in trivial coactions. We use the duality $
\langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, \bar{S}^{-1}x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle x{\underline{\t}}^{\bar{\o}} = \langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}} \rangle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}$ for the fourth equality and gather $w^{\bar{\o}}$ inside $\mathcal{R}$ to cancel it for the fifth. The sixth equality uses (\ref{quacom}) in \[ \ell{\o}x{\underline{\o}}^{\bar{\o}}{\o}\bar{\mathcal{R}}(x{\underline{\o}}^{\bar{\o}}{\t},\ell{\t})= x{\underline{\o}}^{\bar{\o}}{\t}\ell{\t}\bar{\mathcal{R}}(x{\underline{\o}}^{\bar{\o}}{\o}, \ell{\t})\]
and then gathers the parts of $x{\underline{\o}}^{\bar{\o}}$, and cancels some ${\mathcal R}$s. We finally use (\ref{quacom}) to change the order of products in the third tensor factor. On the other side,
\begin{align*}
(w&\otimes \ell \otimes z){\o}(x\otimes k \otimes y){\o}\mathcal{R}\Big((x\otimes k \otimes y){\t},(w\otimes \ell \otimes z){\t}\Big)\\
=&w\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes \ell{\t}x{\underline{\t}}^{\bar{\o}}{\o}k{\o}\otimes f^{b\bar{{}_{\scriptscriptstyle(0)}}}f^{a} \ \langle \underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\\
&\mathcal{R}(k{{}_{\scriptscriptstyle(5)}}y^{\bar{\o}}{\t}, \ell{\th}z{\underline{\o}}^{\bar{\o}}z{\underline{\t}}^{\bar{\o}})\mathcal{R}(S\ell{\o}, x{\underline{\o}}^{\bar{\o}}) \mathcal{R}(f^{b\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\t}k{\t})\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}, x{\underline{\t}}^{\bar{\o}}{\th}k{\th})\\
&\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}}y^{\bar{\o}}{\bar{\o}}),e_{a}\underline{\th}^{\bar{\o}}) \ \langle y^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\t} \rangle \langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b} \rangle\\
=&w\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes \ell{\t}x{\underline{\t}}^{\bar{\o}}{\o}k{\o}\otimes z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}f^{a} \ \langle \underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, (\underline{S}^{-1}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}})x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\\
&\mathcal{R}(k{{}_{\scriptscriptstyle(5)}}y^{\bar{\o}}{\t}, \ell{\th}z{\underline{\o}}^{\bar{\o}}{\t}z{\underline{\t}}^{\bar{\o}})\mathcal{R}(S\ell{\o},x{\underline{\o}}^{\bar{\o}})\mathcal{R}(z{\underline{\o}}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{\t}k{\t})\\
&\mathcal{R}(e_{a}\underline{\o}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{\th}k{\th})\mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}} y^{\bar{\o}}{\o}),e_{a}\underline{\th}^{\bar{\o}}{\o})\mathcal{R}(Se_{a}\underline{\o}^{\bar{\o}}{\t},x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\\
&\mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}e_{a}\underline{\o}^{\bar{\o}}{\th})e_{a}\underline{\th}^{\bar{\o}}{\t}) \ \langle y^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\t} \rangle\\
=&w\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes \ell{\t}x{\underline{\t}}^{\bar{\o}}{\o}k{\o}\otimes z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}f^{a} \ \mathcal{R}(S\ell{\o}, x{\underline{\o}}^{\bar{\o}}) \mathcal{R}(z{\underline{\o}}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{\t}k{\t}) \\
&\mathcal{R}(z{\underline{\t}}^{\bar{\o}}{\o}, x{\underline{\t}}^{\bar{\o}}{\th}k{\th}) \mathcal{R}(S(k{{}_{\scriptscriptstyle(4)}} y^{\bar{\o}}{\o}), z{\underline{\o}}^{\bar{\o}}{\o})\mathcal{R}(z{\underline{\o\t}}^{\bar{\o}}{\t},x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\\
&\mathcal{R}(S(x{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}z{\underline{\t}}^{\bar{\o}}{\th}),z{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{\o}}{\t}) \mathcal{R}(k{{}_{\scriptscriptstyle(5)}}y^{\bar{\o}}{\t}, \ell{\th}z{\underline{\o}}^{\bar{\o}}{\t}z{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}}z{\underline{\th}}^{\bar{\o}}{\th}z{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{\o}}{{}_{\scriptscriptstyle(5)}})\\
&\mathcal{R}(z{\underline{\th}}^{\bar{\o}}{\o}, z{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{\o}}{\th}) \mathcal{R}(z{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}z{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}) \langle \underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle y^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a}\underline{\t} \rangle \langle \underline{S}z{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{{}_{\scriptscriptstyle(0)}}},\underline{S}^{-1}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \\
&\langle \underline{S}z{\underline{\th}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\\
=&w\cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes \ell{\t}x{\underline{\t}}^{\bar{\o}}{\o}k_{\o} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}} \ \mathcal{R}(S\ell{\o}, x{\underline{\o}}^{\bar{\o}}) \mathcal{R}(S(k{\underline{\t}}y^{\bar{\o}}{\o}),z{\underline{\t}}^{\bar{\o}}{\o})\\
&\mathcal{R}(Sx{\underline{\t}}^{\bar{\o}}{\t}, z{\underline{\t}}^{\bar{\o}}{\t}) \mathcal{R}(k{\th}y^{\bar{\o}}{\t}z{\underline{\o}}^{\bar{\o}}{\t}z{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}})\mathcal{R}(z{\underline{\o}}^{\bar{\o}}{\o},z{\underline{\t}}^{\bar{\o}}{\th}) \langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\\
=&w \cdot_{\mathrm{\underline{op}}} x{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes \ell{\t}z{\underline{\o}}^{\bar{\o}}{\o}k{\o} \otimes y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}} \ \mathcal{R}(S\ell{\o},x{\underline{\o}}^{\bar{\o}}{\o})\mathcal{R}(S(k{\t}y^{\bar{\o}}{\o}),z{\underline{\t}}^{\bar{\o}}{\o})\\
&\mathcal{R}(k{\th}y^{\bar{\o}}{\t}, \ell{\th}z{\underline{\o}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}z{\underline{\t}}^{\bar{\o}}{{}_{\scriptscriptstyle(4)}}) \mathcal{R}(Sz{\underline{\o}}^{\bar{\o}}{\t}, z{\underline{\t}}^{\bar{\o}}{\t})\mathcal{R}(z{\underline{\o}}^{\bar{\o}}{\th}, z{\underline{\t}}^{\bar{\o}}{\th}) \langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\o}}^{\bar{\o}} \rangle.
\end{align*}
The second equality uses $\langle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, e_{b} \rangle$ to substitute $f^{b}=z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}$ and we then expand $\cdot_{\mathrm{\underline{op}}}$ inside the pairing. For the third equality, we use
\begin{align*}
\langle \underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}&, (\underline{S}^{-1}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}})x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}}e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle\\
=&\langle (\underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}){\underline{\o}}, \underline{S}^{-1}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle (\underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}}){\underline{\t}}, x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle \langle (\underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}})_{\underline{\th}}, e_{a}\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle
\end{align*}
and move $\underline S$ to the left in $\underline{\Delta}^{2}(\underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}})$. For the fourth equality we gather the coproducts of $e_{a}$ to give $\langle (\underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}})y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{{}_{\scriptscriptstyle(0)}}},e_{a} \rangle$ so that we can set $f^{a}= (\underline{S}z{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(0)}}})y^{\bar{{}_{\scriptscriptstyle(0)}}}z{\underline{{}_{\scriptscriptstyle(4)}}}^{\bar{{}_{\scriptscriptstyle(0)}}}$, allowing us to cancel $(z{\underline{\o}}\underline{S}z{\underline{\t}})^{\bar{{}_{\scriptscriptstyle(0)}}}$ and drop out following coactions. For the fifth equality, we use the duality pairing $
\langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}}^{\bar{{}_{\scriptscriptstyle(\infty)}}} \rangle x{\underline{\t}}^{\bar{\o}} = \langle \underline{S}z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{{}_{\scriptscriptstyle(0)}}}, x{\underline{\t}} \rangle z{\underline{\o}}^{\bar{{}_{\scriptscriptstyle(0)}}\bar{\o}}$ and then gather $z{\underline{\t}}^{\bar{\o}}$ inside $\mathcal{R}$ so as to cancel it and recover the result of our first calculation.
\end{proof}
\section{Construction of ${\mathfrak c}_{q}[SL_{2}]$ by co-double bosonisation} \label{Sec4}
The coquasitriangular Hopf algebra $\mathbb{C}_{q}[SL_{2}]$ in some standard conventions is generated by $a,b,c,d$ with the relations,
\[ba=qab, \quad ca=qac, \quad db=qbd, \quad dc=qcd, \quad cb=bc,\]
\[ad-q^{-1}bc=1, \quad da-ad=(q-q^{-1})bc,\]
a `matrix' form of coproduct (so $\Delta a=a\otimes a+b\otimes c$ etc.), $\epsilon(a)=\epsilon(d)=1$, $\epsilon(b)=\epsilon(c)=0$ and antipode $Sa=d, Sd=a, Sb=-qb, Sc=-q^{-1}$. The reduced version $c_{q}[SL_{2}]$ has
\[ a^{n}=1=d^{n},\quad b^{n}=0=c^{n}\] as additional relations when $q$ is a primitive $n$-th root of unity. We will show how some version of this
is obtained by co-double bosonisation. Let $A=\mathbb{C}_{q}[t]/(t^{n}-1)$ be a coquasitriangular Hopf algebra with $t$ grouplike and $\mathcal{R}(t^{r},t^{s})= q^{rs}$. Also let $B=\mathbb{C}[X]/(X^{n})$ be a braided group in ${}^{A}\mathcal{M}$ with
\[\Delta_{L} X= t \otimes X, \quad \underline{\Delta}X=1\otimes X + X \otimes 1, \quad \underline{\epsilon}X=0, \quad \underline{S}X=-X,\quad \Psi(X^{r}\otimes X^{s})=q^{rs}X^{s}\otimes X^{r}.\]
The dual $B^{*}=\mathbb{C}[Y]/(Y^{n})$ lives in $\mathcal{M}^{A}$ with the same form of coproduct, etc., as for $B$, but with right-action $\Delta_R Y= Y\otimes t $. We choose pairing $\langle X, Y \rangle =1$ and take a basis of $B$ and a dual basis of $B^*$ respectively as
\[ \{e_{a}\}=\{X^a\}_{0\leq a < n},\quad \{f^{a}\}=\Big\{\dfrac{Y^a}{[a]_q!}\Big\}_{0\leq a < n},\]
where $[a]_q$ is a $q$-integers defined by $[a]_{q}=(1-q^{a})/(1-q)$ and $[a]_{q}!=[a]_{q}[a-1]_{q}\cdots [1]_{q}!$ with $[0]_{q}!=1$. We also write $\left[ a \atop r\right]_{q}=\frac{[a]_{q}!}{[r]_{q}![a-r]_{q}!}$. We write $X^{a(\underline{\mathrm{op}})} = X\cdot_{\mathrm{\underline{op}}} X \cdot_{\mathrm{\underline{op}}} \cdots \cdot_{\mathrm{\underline{op}}} X$ with $a$-many $X$, and find inductively that
\begin{align}\label{Xop}
X^{a}=&q^{\frac{a(a-1)}{2}}X^{a(\underline{\mathrm{op}})}, \quad \bar{S}(X^{a})=(-1)^{a}X^{a(\underline{\mathrm{op}})}.
\end{align}
\begin{theorem}\label{cdbthm} Let $q$ be a primitive $n$-th root of unity and $A,B,B^{*}$ be as above.
\begin{enumerate}
\item The co-double bosonisation of $B$, denoted ${\mathfrak c}_q[SL_2]$, has generators $X,t,Y$ and
\begin{align*}
&X^{n}=Y^{n}=0, \quad t^{n}=1, \quad YX=XY,\quad Xt=qtX,\quad Yt=qtY,\\
\Delta t&= q\sum\limits_{a=0}^{n-2}(q-1)^{a-1}(1-q^{-a-1}) tY^{a}\otimes X^{a}t\\
&={t\over q-1}\left(q{1-((q-1)Y\otimes X)^{n-1}\over 1-(q-1)Y\otimes X}- {1-((1-q^{-1})Y\otimes X)^{n-1}\over 1-(1-q^{-1})Y\otimes X} \right)t, \\
\Delta X&=X\otimes 1 + \sum\limits_{a=0}^{n-2} (q-1)^a tY^{a}\otimes X^{a+1}=X\otimes 1+t\left({1-((q-1)Y\otimes X)^{n-1}\over1-(q-1)Y\otimes X}\right)X,\\
\Delta Y&=1\otimes Y +\sum\limits_{a=0}^{n-2}(1-q^{-1})^a Y^{a+1}\otimes X^{a}t=1\otimes Y+Y\left({1-((1-q^{-1})Y\otimes X)^{n-1}\over1-(1-q^{-1})Y\otimes X}\right) t. \end{align*}
\item If $n=2m+1$, there is an isomorphism $\phi : {\mathfrak c}_q[SL_2] \to c_{q^{-m}}[SL_2]$ defined by
\[ \phi(X)= bd^{-1},\quad \phi(t)=d^{-2}, \quad \phi(Y)={d^{-1}c\over q^m-q^{-m}}.\]
\end{enumerate}
\end{theorem}
\begin{proof}
(1) First we determine the products
\begin{align*}
(1\otimes 1 \otimes Y)(X \otimes 1 \otimes 1)=& X^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes 1 \otimes Y^{\bar{{}_{\scriptscriptstyle(0)}}} \ \mathcal{R}(Y^{\bar{\o}},1)\mathcal{R}(S1,X^{\bar{\o}})=X\otimes 1 \otimes Y,\\
(1\otimes t \otimes 1)(X \otimes 1 \otimes 1)=& X^{\bar{{}_{\scriptscriptstyle(\infty)}}} \otimes t \otimes 1 \ \mathcal{R}(St, X^{\bar{\o}})=q^{-1} X\otimes t \otimes 1,\\
(1 \otimes 1 \otimes Y)(1 \otimes t \otimes 1)=& 1 \otimes t \otimes Y^{\bar{{}_{\scriptscriptstyle(0)}}} \ \mathcal{R}(Y^{\bar{\o}},t)=q 1 \otimes t \otimes Y
\end{align*}
as stated. The algebra generated by $X,Y,t$ with these relations is $n^3$ dimensional, hence these are all the relations we need. Before go further, we note the $q$-identities
\begin{align}\label{q-identity1}
\sum\limits_{r=0}^{a} (-1)^{r}{q^{\frac{r(r+1)}{2}}\over [r]_q! [a-r]_q!}=(1-q)^a, \ \sum\limits_{r=0}^{a} q^r(-1)^{r}{q^{\frac{r(r+1)}{2}}\over [r]_q! [a-r]_q!}=(1-q)^a[a+1]_q.
\end{align}
Then, using (\ref{Xop}), we compute
\begin{align*}
\Delta&(1\otimes t \otimes 1)=\sum\limits_{a=0}^{n-1}1 \otimes t \otimes \frac{Y^{a}}{[a]_{q}!}\otimes \big(\sum\limits_{r=0}^{a} \left[a\atop r \right]_{q} X^{r}\cdot_{\mathrm{\underline{op}}} \bar{S}X^{a-r}\otimes t \otimes 1 \ \mathcal{R}(t^{r},t)\mathcal{R}(St,t^{a-r})\big)\\
=&\sum\limits_{a=0}^{n-1}1 \otimes t \otimes \frac{Y^{a}}{[a]_{q}!}\otimes \big(\sum\limits_{r=0}^{a} \left[a\atop r \right]_{q}(-1)^{a-r} q^{\frac{r(r-1)}{2}+2r-a}X^{r(\underline{\mathrm{op}})}\cdot_{\mathrm{\underline{op}}} X^{(a-r)(\underline{\mathrm{op}})}\otimes t \otimes 1\big)\\
=& \sum\limits_{a=0}^{n-2}\sum\limits_{r=0}^{a} \frac{(-1)^{a-r}\left[ a \atop r \right]_{q} q^{\frac{r(r+1)}{2}}q^{-a+r}}{[a]_{q}!} tY^{a}\otimes X^{a}t \label{Delta t}
\end{align*}
since there is no contribution when $a=n-1$. We then use (\ref{q-identity1}). Similarly,
\begin{align*}
\Delta(X \otimes &1 \otimes 1)=X \otimes 1 \otimes 1 \otimes 1 \otimes 1 \otimes 1 \\
&+ \sum\limits_{a=0}^{n-1} 1 \otimes t \otimes \frac{Y^{a}}{[a]_{q}!}\otimes\big(\sum\limits_{r=0}^{a}\left[a \atop r\right]_{q} X^{r}\cdot_{\mathrm{\underline{op}}} X \cdot_{\mathrm{\underline{op}}} \bar{S}X^{a-r}\otimes 1 \otimes 1 \ \mathcal{R}(t^{r}, t)\big)\\
&=X \otimes 1 \otimes 1 \otimes 1 \otimes 1 \otimes 1\\
&+ \sum\limits_{a=0}^{n-1} 1 \otimes t \otimes \frac{Y^{a}}{[a]_{q}!}\otimes \big(\sum\limits_{r=0}^{a}\left[a \atop r\right]_{q} q^{\frac{r(r+1)}{2}} X^{r(\underline{\mathrm{op}})}\cdot_{\mathrm{\underline{op}}} X \cdot_{\mathrm{\underline{op}}} (-1)^{a-r}X^{(a-r)(\underline{\mathrm{op}})}\otimes 1 \otimes 1\big)\\
&= X\otimes 1 + \sum\limits_{a=0}^{n-2}\sum\limits_{r=0}^{a}\frac{(-1)^{a-r}\left[a \atop r\right]_{q} q^{\frac{r(r+1)}{2}}}{[a]_{q}!}tY^{a}\otimes X^{a+1},
\end{align*}
where for $a=n-1$, we will have the term $tY^{n-1}\otimes X^{n}=0$.
We again use (\ref{q-identity1}). Finally, we use $\underline{\Delta}^{2}(e_{a})=\underline{\Delta}^{2}(X^{a})=\sum\limits_{r=0}^{a}\sum\limits_{s=0}^{r}\left[ a \atop r\right]_{q}\left[r \atop s \right]_{q} X^{s}\otimes X^{r-s}\otimes X^{a-r}$ to find
\begin{align*}
&\Delta(1\otimes 1 \otimes Y)= 1 \otimes 1 \otimes 1 \otimes 1 \otimes 1 \otimes Y \\
&\ + \sum\limits_{a=0}^{n-1}1\otimes 1 \otimes f^{a}\otimes e_{a}\underline{\o}\cdot_{\mathrm{\underline{op}}}\bar{S}e_{a}\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes t \otimes 1 \ \mathcal{R}(St, e_{a}\underline{\th}^{\bar{\o}})\langle Y, e_{a}\underline{\t} \rangle.\\
&= 1 \otimes 1 \otimes 1 \otimes 1 \otimes 1 \otimes Y \\
&\ +\sum\limits_{a=0}^{n-1}1\otimes 1 \otimes \frac{Y^{a}}{[a]_{q}!}\otimes \big(\sum\limits_{r=0}^{a}\sum\limits_{s=0}^{r}\left[ a \atop r\right]_{q}\left[r \atop s \right]_{q} X^{s}\cdot_{\mathrm{\underline{op}}}\bar{S}X^{a-r}\otimes t \otimes 1 \ \mathcal{R}(St, t^{a-r})\langle Y, X^{r-s} \rangle\big)\\
&=1 \otimes 1 \otimes 1 \otimes 1 \otimes 1 \otimes Y \\
&\ + \sum\limits_{a=0}^{n-1}1\otimes 1 \otimes \frac{Y^{a}}{[a]_{q}!}\otimes \big(\sum\limits_{r=0}^{a}\sum\limits_{s=0}^{r}\left[ a \atop r\right]_{q}\left[r \atop s \right]_{q} \delta_{1, r-s}q^{\frac{s(s-1)}{2}}q^{-a+r}X^{s(\underline{\mathrm{op}})}\cdot_{\mathrm{\underline{op}}}(-1)^{a-r}X^{(a-r)(\underline{\mathrm{op}})}\otimes t \otimes 1\big)\\
&=1\otimes Y + \sum\limits_{a=0}^{n-1}\sum\limits_{r=0}^{a}\sum\limits_{s=0}^{r} \frac{(-1)^{a-r}\delta_{1, r-s}\left[ a \atop r\right]_{q}\left[r \atop s \right]_{q}q^{\frac{s(s-1)}{2}}q^{-a+r}}{[a]_{q}!}Y^{a}\otimes X^{a-r+s}t\\
&=1\otimes Y +\sum\limits_{a=0}^{n-2}\sum\limits_{r=0}^{a}\frac{(-1)^{a-r}\left[a \atop r \right]_{q}q^{\frac{r(r+1)}{2}}q^{-a}}{[a]_{q}!} Y^{a+1}\otimes X^{a}t.
\end{align*}
There was no contribution from $a=0$ and for $a>0$ we needed $s=r-1$ for a contribution. We then use (\ref{q-identity1}). The general theory in Section~\ref{Sec3} ensures that the Hopf algebra is coquasitriangular.
(2) If $n=2m+1$ then $\varphi : c_{q^{-m}}[SL_2] \to {\mathfrak c}_q[SL_2]$ defined by
\[
\varphi(a)=t^{m+1}+ (q^{m}-q^{-m}) Xt^{m}Y,\quad
\varphi(b)=Xt^{m},\quad
\varphi(c)=(q^{m}-q^{-m})t^{m}Y,\quad
\varphi(d)=t^{m}. \]
is an algebra map and inverse to $\phi$. Tedious but straightforward calculation gives
\begin{align*}
\Delta(\varphi(d))=\Delta t^{m}=&t^{m}\otimes t^{m} + (q^{2m}-1)t^{m}Y \otimes t^{m}X=t^m\otimes t^m+(q^{m}-q^{-m})t^{m}Y \otimes Xt^{m},
\end{align*}
to prove that $\Delta(\varphi(d))=(\varphi\otimes \varphi)\Delta d$. The coalgebra map property on the other generators then follows using this formula for $\Delta t^m$. Furthermore, the coquasitriangular structure from Lemma \ref{codouble coquasitriangularity} computed on $\varphi(a),\varphi(b),\varphi(c),\varphi(d)$ as a matrix $\varphi(t^{i}_{~j})$ is
\begin{align}\label{RSL2}
R^{I}{}_J=q^{m(m+1)}\begin{pmatrix}
q^{-m} & 0 & 0 & 0\\
0 & 1 & q^{-m}-q^{m} & 0\\
0 & 0 & 1 & 0\\
0 & 0 & 0 & q^{-m}
\end{pmatrix}.
\end{align}
for the values of ${\mathcal R}(\varphi(t^{i}_{~j}), \varphi(t^{k}_{~l}))$ where $I=(i,k)$ is (1,1),(1,2),(2,1), or (2,2) and similarly for $J=(j,l)$. If we set $p=q^{-m}$ then any power of $p$ is also a $2m+1$-th root of unity and $q=q^{-2m}=p^{2}$ so that our Hopf algebra is $c_{p}[SL_{2}]$ with its standard coquasitriangular structure with the correct factor $q^{m(m+1)}=p^{-m-1}=p^m=p^{-{1\over 2}}$.
\end{proof}
We now recall explicitly that for $q$ a primitive $n$-th root of unity and $q^2\ne 1$, $u_{q}(sl_{2})$ is generated by $E,F,K$, with relations, coproducts and coquasitriangular structure
\[ E^n=F^n=0, \quad K^n=1, \quad KEK^{-1}=q^{-2}E, \quad KFK^{-1}=q^{2}F, \quad [E,F]=K-K^{-1},\]
\[\Delta K = K\otimes K, \quad \Delta F= F\otimes 1 + K^{-1}\otimes F , \quad \Delta E= E\otimes K + 1\otimes E,\]
\[{\mathcal R}=\dfrac{1}{n}\sum\limits_{r,a,b=0}^{n-1}\dfrac{(-1)^r q^{-2ab}}{[r]_{q^{-2}}}F^rK^a \otimes E^r K^b, \]
where in our conventions we do not divide by the usual $q-q^{-1}$ in the $[E,F]$-relation (and where we use $q^{-2}$ rather than $q^2$ in the remaining relations compared with \cite{Primer}). One can consider this as an unconventional normalisation of $E$ which is cleaner when we are not interested in a classical limit. It gives a commutative Hopf algebra $u_{-1}(sl_2)$ when $q=-1$. We first show that double bosonisation gives us some version of such reduced quantum groups, agreeing for primitive odd roots. This was outlined in \cite[Example 17.6]{Primer} in the odd root case but we give a short derivation for all roots.
\begin{lemma}\label{dboslem}\cite{Primer} Let $q$ be a primitive $n$-th root of unity and let $H={\Bbb C}_{q}{\Bbb Z}_{n}={\Bbb C}_q[K]/(K^n-1)$ be a quasitriangular Hopf algebra by ${\mathcal R}_K = \frac{1}{n}\sum\limits_{a,b=0}^{n-1}q^{-ab}K^{a}\otimes K^{b}$ as in \cite{Foundation}. Let $B= {\Bbb C}[E]/(E^{n})$ be a braided group in ${\mathcal M}_{H}$ and dual $B^{*}={\Bbb C}[F]/(F^{n})$ in ${}_{H}{\mathcal M}$ with actions $E\lhd K =qE$ and $K \rhd F =qF$.
\begin{enumerate}\item The double bosonisation $B^{*{\mathrm{\underline{cop}}}}{>\!\!\!\triangleleft\kern-.33em\cdot} H {\cdot\kern-.33em\triangleright\!\!\!<} B$ is a quasitriangular Hopf algebra, which we denote ${\mathfrak u}_q(sl_2)$, with the same coalgebra structure as above but with
\[E^n=F^n=0, \quad K^n=1, \quad KEK^{-1}=q^{-1}E, \quad KFK^{-1}=qF, \quad [E,F]=K-K^{-1},\]
\[{\mathcal R}_{{\mathfrak u}_q(sl_2)}=\dfrac{1}{n}\sum\limits_{r,a,b=0}^{n-1}\dfrac{(-1)^r q^{-ab}}{[r]_{q^{-1}}!}F^r K^a\otimes E^r K^b.\]
\item If $n=2m+1$ then ${\mathfrak u}_q(sl_2)$ is isomorphic to $u_{q^{-m}}(sl_{2})$ with its standard quasitriangular structure.
\end{enumerate}
\end{lemma}
\begin{proof}
Here $EK\equiv (1\otimes E)(K \otimes 1)= K \otimes E \lhd K=K \otimes qE \equiv qKE$ and $KF\equiv (1\otimes K)(F\otimes 1)=K\rhd F \otimes K=qF\otimes K\equiv qFK$. From the cross relations stated in Theorem \ref{Double Bosonisation}, we also have
\begin{align*}
EF=&FE+\dfrac{1}{n}\sum\limits_{a,b=0}^{n-1}q^{-ab}K^{b}\langle F, E\lhd K^{a} \rangle - \dfrac{1}{n}\sum\limits_{a,b=0}^{n-1}q^{ab}K^{a}\langle K^{b} \rhd F,E \rangle\\
=&FE+\dfrac{1}{n}\sum\limits_{b=0}^{n-1}\left(\frac{1-q^{-n(b-1)}}{1-q^{-(b-1)}}\right)K^{b}\langle F,E \rangle-\dfrac{1}{n}\sum\limits_{a=0}^{n-1}\left(\frac{1-q^{n(a+1)}}{1-q^{a+1}}\right)K^{a}\langle F,E \rangle\\
=&FE+K-K^{-1},
\end{align*}
where we choose $\langle F,E \rangle = 1$. This is the same choice of normalisation for the braided line duality as in the calculation in Theorem \ref{cdbthm}. For the coproduct, clearly $\Delta K = K \otimes K$ while $\Delta E \equiv \Delta(1\otimes E)=1 \otimes 1 \otimes 1 \otimes E + 1 \otimes E \lhd {\mathcal R}_K^{\o} \otimes {\mathcal R}_K^{\t}\otimes 1 = 1\otimes 1 \otimes 1 \otimes E + 1 \otimes E \otimes K \otimes 1 \equiv 1 \otimes E + E \otimes K$ and $\Delta F \equiv \Delta(F \otimes 1)= F \otimes 1 \otimes 1 \otimes 1 + 1 \otimes {\mathcal R}_K^{{\scriptscriptstyle -(1)}}\otimes {\mathcal R}_K^{{\scriptscriptstyle -(2)}}\rhd F \otimes 1 = F \otimes 1 \otimes 1 \otimes 1 + 1 \otimes K^{-1}\otimes F \otimes 1 \equiv F\otimes 1 + K^{-1}\otimes F$. Hence we have the relations and coalgebra as stated. Also from Theorem \ref{Double Bosonisation},
\[{\mathcal R}_{{\mathfrak u}_q(sl_2)}=\sum\limits_{r=0}^{n-1}\left(\dfrac{F^r}{[r]_{q}!}\otimes\underline{S}E^r\right){\mathcal R}_K=\dfrac{1}{n}\sum\limits_{r,a,b=0}^{n-1}\dfrac{(-1)^r q^{\frac{r(r-1)}{2}}q^{-ab}}{[r]_q!}F^r K^a\otimes E^r K^b,
\]
which we write as stated. When $n=2m+1$, it is easy to see that the relations and quasitriangular structure become those of $u_{p}(sl_2)$ with $p=q^{-m}$, which are the same as in \cite{Primer} after allowing for the normalisation of the generators. Note that if $q$ is an even root of unity then ${\mathcal R}_{{\mathfrak u}_q(sl_2)}$ need not be factorisable, see Example \ref{exq2}. In fact, ${\mathcal R}_{{\mathfrak u}_q(sl_2)}$ is factorisable iff $n$ is odd, which can be proven in a similar way to the proof in \cite{Lyu}. \end{proof}
We see that the double bosonisation ${\mathfrak u}_q(sl_2)$ recovers $u_{p}(sl_{2})$ in the odd root of unity case with $p=q^{1\over 2}$, in line with the generic $q$ case in \cite{db}. Clearly ${\mathfrak u}_q(sl_2)$ has a PBW-type basis $\{F^{i}K^{j}E^{k}\}_{0 \leq i,j,k \leq n-1}$ as familiar in the odd case for $u_p(sl_{2})$.
\begin{corollary}\label{dual basis}
The basis $\{X^{i}t^{j}Y^{k}\}_{0\leq i,j,k \leq n-1}$ of ${\mathfrak c}_q[SL_2]$ is, up to normalisation, dual to the PBW basis of ${\mathfrak u}_q(sl_{2})$ in the sense
\begin{align*}
\langle X^{i}t^{j}Y^{k}, F^{i'}K^{j'}E^{k'} \rangle = \delta_{i,i'}\delta_{k,k'} q^{jj'}[i]_{q^{-1}}![k]_{q}!.
\end{align*}
More precisely, $\Big\{\dfrac{X^i \delta_j(t) Y^k}{[i]_{q^{-1}}![k]_{q}!}\Big\}_{0\leq i,j,k<n}$ is a dual basis to $\{F^i K^j E^k\}_{0\leq i,j,k<n}$, where $\delta_j(t)={1\over n}\sum\limits_{l=0}^{n-1}q^{-jl}t^l$.
\end{corollary}
\begin{proof}
The duality pairing between the double and co-double bosonisations is
\begin{align*}
\langle X^{i}t^{j}Y^{k}, F^{i'}K^{j'}E^{n'} \rangle = \langle X^{i(\underline{\mathrm{op}})}, F^{i'} \rangle \langle t^{j}, K^{j'} \rangle \langle Y^{k}, E^{k'} \rangle,
\end{align*}
where the pairing between $(\mathbb{C}[X]/(X^{n}))^{\mathrm{\underline{op}}}$ and $(\mathbb{C}[F]/(F^{n}))^{\mathrm{\underline{cop}}}$ implied by $\langle X, F \rangle =1$ is $\langle X^{i(\underline{\mathrm{op}})}, F^{i'} \rangle =\delta_{i,i'} [i]_{q^{-1}}!$ while $\langle t^{j}, K^{j'} \rangle =q^{jj'}$ is implied by $\langle t,K \rangle =q$. The latter is the duality pairing in the Pontryagin sense in which ${\Bbb Z}_n$ is self-dual, and can be written as a usual dual pairing with the $\delta_j$. Equally well, $\Big\{ \dfrac{F^i \delta_j(K)E^k}{[i]_{q^{-1}}![k]_{q}!} \Big\}_{0\leq i,j,k < n}$ is a dual basis to $\{X^i t^j Y^k\}_{0\leq i,j,k<n}$.
\end{proof}
This applies even when $q=-1$, in that case as a self-duality pairing.
\begin{example}\label{exq2}
If $q=-1$ then the double bosonisation $\mathbf{{\mathfrak u}}_{-1}(sl_2)=B^{*\mathrm{\underline{cop}}}{>\!\!\!\triangleleft\kern-.33em\cdot} H {\cdot\kern-.33em\triangleright\!\!\!<} B$ from Lemma~\ref{dboslem} has relations and coalgebra
\begin{align*}
&E^2=F^2=0,\quad K^2=1,\quad EF=FE, \quad KE=-EK, \quad KF=-FK,\\
&\Delta K=K\otimes K, \quad \Delta F= F\otimes 1 + K\otimes F,\quad \Delta E= E \otimes K+1\otimes E
\end{align*}
and is self-dual and strictly quasitriangular with
\[ {\mathcal R}=(1\otimes 1 - F\otimes E){\mathcal R}_{K},\quad {\mathcal R}_K={1\over 2}(1\tens1+1\otimes K+K\otimes 1-K\otimes K).\]
It is easy to check that this is not triangular, i.e $Q:={\mathcal R}_{21}{\mathcal R}=1 \otimes 1-E\otimes F-KF\otimes EK-EKF\otimes FKE \ne 1\otimes 1$, and also not factorisable in the sense that the map $\mathbf{{\mathfrak u}}_{-1}(sl_2)^*\to \mathbf{{\mathfrak u}}_{-1}(sl_2)$ which sends $\phi \mapsto (\phi\otimes \mathrm{id})Q$ is not surjective (the element $FK \in \mathbf{{\mathfrak u}}_{-1}(sl_2)$ is not in the image). On the other hand, Theorem~\ref{cdbthm} (1) gives us an isomorphic Hopf algebra by $X\mapsto F, Y\mapsto E$ and $t\mapsto K$, so our Hopf algebra is self-dual, i.e., ${\mathfrak u}_{-1}(sl_2)\cong {\mathfrak c}_{-1}[SL_2].$ Note that ${\mathfrak u}_{-1}(sl_2)$ has the same dimension and coalgebra as $u_{-1}(sl_2)$ but cannot be isomorphic, being noncommutative. One can also check that ${\mathfrak c}_{-1}[SL_2]$ is not isomorphic as a Hopf algebra to $c_{-1}[SL_{2}]$ and the latter, being noncocommutative, cannot be dual to $u_{-1}(sl_2)$. \end{example}
\section{Application to Hopf Algebra Fourier Transform} \label{Sec5}
As a corollary of the above results, we briefly consider Hopf algebra Fourier transform between our double and co-double bosonisations. Recall from standard Hopf algebra theory, e.g. \cite{Foundation}, that for a finite-dimensional Hopf algebra $H$ there is, up to scale, a unique right integral structure $\int : H \to k$ satisfying
\begin{align*}
\left(\int \otimes \mathrm{id}\right)\Delta h = \left(\int h\right)1
\end{align*}
for all $h\in H$. Such a right integral is the main ingredient for Fourier transform $\mathcal{F}: H \to H^{*}$. The following preliminary lemma is essentially well-known (see \cite[Proposition, 1.7.7]{Foundation}), but for completeness we give the easier part that we need.
\begin{lemma} \label{Fourier transform}
Let $\int, \int^{*}$ be right integrals on finite-dimensional Hopf algebras $H, H^{*}$ respectively and $\mu = {\int(\int^{*})}$. The Fourier transform $\mathcal{F}: H\to H^{*}$ and adjunct ${\mathcal F}^*$ obey
\[\mathcal{F}(h)=\sum\limits_{a}\left(\int e_{a} h\right)f^{a}, \quad \mathcal{F}^{*}(\phi)=\sum\limits_{a} e_{a}\Big(\int^{*}\phi f^{a}\Big), \quad {\mathcal F}^{*}\circ {\mathcal F} = \mu S,
\]
where $\{e_{a}\}$ is basis of $H$, $\{f^{a}\}$ is the dual basis of $H^{*}$. Hence ${\mathcal F}$ is invertible if $\mu\ne 0$.
\end{lemma}
\begin{proof}
We write $\int^{*}=\Lambda^{*}$ when regarded as element in $H$. Then
\begin{align*}
{\mathcal F}^{*}\circ {\mathcal F}(h)=& e_{a}\Big(\int^{*}(\int e_{b} h)f^{b} f^{a}\Big) = (\int e_{a_{\o}} h) e_{a_{\t}}(\int ^{*}f^{a}) \\
=&(\int \Lambda^{*}{\o}h_{\o})\Lambda^{*}\t h_{\t}Sh_{\th} = (\int \Lambda^{*} h_{\o})Sh_{\t}=(\int \Lambda^{*}) Sh = \mu Sh.
\end{align*}
If $\mu\ne 0$ then this implies that ${\mathcal F}$ is injective and hence in our case invertible (with a bit more work \cite{Foundation} one can show that the inverse is $\mu^{-1}S^{-1}{\mathcal F}^*$).
\end{proof}
\begin{proposition} \label{Fourier c_p[SL2]}
Let $q$ be a primitive $n$-th root of unity. The Fourier transform $\mathcal{F} : {\mathfrak c}_{q}[SL_{2}]\to {\mathfrak u}_{q}(sl_{2})$ is invertible and given by
\[ \mathcal{F}(X^{\alpha}t^{\beta}Y^{\gamma}) = \sum\limits_{l=0}^{n-1}\dfrac{q^{-(l+\alpha)(1-\beta)+\beta(n-1-\gamma)}}{n[n-1-\alpha]_{q^{-1}}![n-1-\gamma]_{q}!}F^{n-1-\alpha}K^{l}E^{n-1-\gamma}.\]
\end{proposition}
\begin{proof}
The right integral for ${\mathfrak c}_{q}[SL_{2}]$ is given by.
\[\int X^\alpha t^\beta Y^\gamma = \begin{cases}
1, \text{ if $\alpha=\gamma=n-1, \beta=1$}\\
0, \text{ otherwise.}
\end{cases}\]
This integral is equivalent in usual generators to $\int b^{n-1}c^{n-1}=1$ and zero otherwise. We use Corollary \ref{dual basis} to give us the basis $\{ e_{a}\}=\{X^{i}t^{j}Y^{k}\}_{0\leq i,j,k\leq n-1}$ of ${\mathfrak c}_{q}[SL_2]$ and the dual basis $\{f^{a}\}=\{\frac{F^{i}\delta_{j}(K)E^{k}}{[i]_{q^{-1}}!q^{j^{2}}[k]_{q}!}\}_{0\leq i,j,k\leq n-1}$ of ${\mathfrak u}_{q}(sl_{2})$. Then
\begin{align*}
{\mathcal F}(X^\alpha t^\beta Y^\gamma)=&\sum\limits_{i,j,k=0}^{n-1}(\int X^i t^j Y^k X^\alpha t^\beta Y^\gamma)\dfrac{F^i \delta_{j}(K)E^k}{[i]_{q^{-1}}![k]_{q}!}\\
=&\sum\limits_{i,j,k=0}^{n-1}q^{-\alpha j+\beta k}(\int X^{i+\alpha}t^{j+\beta}Y^{k+\gamma})\dfrac{F^i \delta_{j}(K)E^k}{[i]_{q^{-1}}![k]_{q}!}\\
=&q^{-\alpha(1-\beta)+\beta(n-1-\gamma)}\dfrac{F^{2-\alpha} \delta_{1-\beta}(K)E^{n-1-\gamma}}{[n-1-\alpha]_{q^{-1}}![n-1-\gamma]_{q}!}\\
=&\sum\limits_{l=0}^{n-1}\dfrac{q^{-(l+\alpha)(1-\beta)+\beta(n-1-\gamma)}}{(n-1+1)[n-1-\alpha]_{q^{-1}}![n-1-\gamma]_{q}!}F^{n-1-\alpha}K^{l}E^{n-1-\gamma}.
\end{align*}
The similar right integral of ${\mathfrak u}_{q}(sl_{2})$ and resulting $\mu$ are
\[\int^* F^\alpha K^\beta E^\gamma = \begin{cases}
1 & \text{ if $\alpha=\gamma=n-1, \beta=1$}\\
0 & \text{ otherwise},
\end{cases} \quad \mu =\dfrac{q^{-1}}{n[n-1]_{q^{-1}}![n-1]_{q}!},\]
which is nonzero.
\end{proof}
It appears to be a hard computational problem to give the general formula of the inverse Fourier transform, but one can compute it for specific cases.
\begin{example}
Let $q$ be a primitive cube root of unity. First, observe that for $\alpha,\beta=0,1,2$, we have
\begin{align*}[E^{\alpha},F^{\beta}]=F^{\beta-1}&([\alpha]_q [\beta]_q K - [\alpha]_{q^{-1}}[\beta]_{q^{-1}}K^{-1})E^{\alpha-1}\\ & + F^{\beta-2}([2]_q K - [2]_{q^{-1}}K^{-1})(K-K^{-1})E^{\alpha-2}\end{align*}
in ${\mathfrak u}_q(sl_2)$. Using this commutation relation, we obtain
\begin{align*}
{\mathcal F}^{*}(F^\alpha K^\beta E^\gamma)=&\sum\limits_{l=0}^{2}\dfrac{q^{\beta(2-\alpha)+(\gamma
-l)(1-\beta)}}{3[2-\alpha]_{q^{-1}}![2-\gamma]_{q}!}X^{2-\alpha}t^l Y^{2-\gamma}\\
&~+\sum\limits_{l=0}^{2}\dfrac{q^{\beta(l-\alpha-\gamma)}([\gamma]_q [3-\alpha]_q - q^{2(\gamma-l)+1}[\gamma]_{q^{-1}}[3-\alpha]_{q^{-1}})}{3[3-\alpha]_{q^{-1}}![3-\gamma]_{q}!}F^{3-\alpha}t^l Y^{3-\gamma}\\
&~-\dfrac{q^{2\beta+1}}{3[4-\alpha]_{q^{-1}}![4-\gamma]_{q}!}X^{4-\alpha}t^2 Y^{4-\gamma}.
\end{align*}
One can check that ${\mathcal F}^{*}{\mathcal F}(X^\alpha t^\beta Y^\gamma)=\mu S(X^\alpha t^\beta Y^\gamma)$, where $\mu = \frac{q^{-1}}{3 [2]_{q^{-1}}![2]_{q}!}={q^2 \over 3}$ and
\begin{align*}
S(X^\alpha t^\beta Y^\gamma) =& \dfrac{q^{\alpha \beta-\beta\gamma} }{[2-\alpha]_{q^{-1}}![\alpha]_{q^{-1}}![2-\gamma]_{q}![\gamma]_{q}!}X^\alpha t^{-\delta} Y^\gamma\\
&+\dfrac{q^{\alpha \beta-\beta \gamma}(q^{2\beta-2}-q^{\delta-2})}{[2-\alpha]_{q^{-1}}![\alpha]_{q^{-1}}![1-\gamma]_q ![1+\gamma]_q!}X^{\alpha+1}t^{-\delta-1}Y^{\gamma+1}\\
&-\dfrac{1+q^{\beta+1}+q^{2\beta+2}}{[2-\alpha]_{q^{-1}}![2+\alpha]_{q^{-1}}![2-\gamma]_{q}![2+\gamma]_{q}!} X^{2+\alpha}t^2Y^{2+\gamma},
\end{align*}
where $\delta=\alpha+\beta+\gamma$.
\end{example}
\begin{example}
At $q=-1$, the Fourier transform in Proposition \ref{Fourier c_p[SL2]} combined with the self-duality in Example~\ref{exq2} becomes a Fourier transform operator ${\mathfrak c}_{-1}[SL_2]\to {\mathfrak c}_{-1}[SL_2]$. This has eigenvalues $\pm \frac{\imath}{\sqrt{2}}$ with multiplicity 2, $\pm \frac{(-1)^{1/4}}{\sqrt{2}}$ and $\pm \frac{(-1)^{3/4}}{\sqrt{2}}$ with multiplicity 1, and characteristic polynomial $f(x)=\frac{1}{16} +\frac{x^2}{4}+\frac{x^4}{2} + x^6 + x^8$. We also have
\[{\mathcal F}^*(F^a K^b E^c)=\dfrac{1}{2}\sum\limits_{l=0}^{1} (-1)^{(1-b)(c-l)+b(1-a)}X^{1-a}t^l Y^{1-c} \]
and one can check that ${\mathcal F}^{-1}=\mu^{-1}S^{-1}{\mathcal F}^*$ as in Lemma \ref{Fourier transform}.
\end{example}
It is known that Fourier transform behaves well with respect to the coregular representation. This implies that it behaves well with respect to any covariant calculus. Thus, let $(\Omega^1,{\rm d})$ be a left-covariant calculus on $H$. By definition, a differential calculus means an $H$-$H$-bimodule $\Omega^1$ together with a derivation ${\rm d}:H\to \Omega^1$ such that the map $H\otimes H\to \Omega^1$ sending $h\otimes g \mapsto h{\rm d} g$ is surjective. This is left covariant if the map
\[ \Delta_L(h{\rm d} g)=h\o g\o \otimes h\t {\rm d} g\t,\quad \Delta_L:\Omega^1\to H\otimes\Omega^1\]
is well-defined. In this case it is a left coaction and ${\rm d}$ is a comodule map with respect to the left coproduct on $H$. By the Hopf-module lemma, such $\Omega^1$ are free modules over their space $\Lambda^1$ of invariant 1-forms while
${\rm d} h = h\o \varpi\pi_\epsilon h\t$
for all $h\in H$, where $\pi_\epsilon={\rm id}-1\epsilon:H\to H^+$ and $\varpi:H^+\twoheadrightarrow \Lambda^1$ is the Maurer-Cartan form $\varpi(h)=Sh\o{\rm d} h\t$ for all $h\in H^+$. We refer to \cite{Woronowicz89,Primer} for details. The following is known, see e.g. \cite{Ma:Hod}, but we include a short derivation in our conventions. In our case $H$ is finite-dimensional.
\begin{lemma} Let $\{e_a\}$ be a basis of $\Lambda^1$, $\{f^a\}$ a dual basis and define partial derivatives $\partial^a:H\to H$ by ${\rm d} h=\sum_a (\partial^a h)e_a$ and $\chi_a\in H^*{}$ by $\chi_a(h)=\<f^a,\varpi\pi_\epsilon S^{-1} h\>$ for all $h\in H$. Then ${\mathcal F}(\partial^a h)=({\mathcal F} h)\chi_a$ for all $h\in H$.
\end{lemma}\begin{proof}Using the right-integral property, we have
\begin{align*}{\mathcal F}&(\partial^a h)={\mathcal F}(h\o) \<f^a,\varpi\pi_\epsilon h\t\>=\sum_b( \int e_b\o h\o)f^b \<f^a,\varpi\pi_\epsilon ((S^{-1}e_b\th)e_b\t h\t\>\\
=& \sum_b (\int e_b\o h)f^b \<f^a,\varpi\pi_\epsilon (S^{-1}e_b\t\>= \sum_{b,c}( \int e_b h)f^b f^c \<f^a,\varpi\pi_\epsilon (S^{-1}e_c\>=({\mathcal F} h)\chi_a.
\end{align*} \end{proof}
\begin{example}
The 3D calculus c.f. \cite{Woronowicz89} has left-invariant basic 1-forms $e_\pm, e_0$ with $e_\pm h =p^{|h|}h e_\pm$ and $e_0 h=p^{2|h|}h e_0$ where $p=q^{-m}$ and $|\ |$ denotes a grading with $a,c$ grade 1 and $b,d$ grade -1 as a ${\Bbb Z}_n$-grading of $c_p[SL_2]$. Correspondingly for ${\mathfrak c}_q[SL_2]$, we have a calculus with $|X|=0, |t|=|Y|=2$
and one can compute
\[
{\rm d} X =q^{-m} te_-, \quad {\rm d} t= (1+q)\big(q(q^{-m}-q^m)tYe_- + te_0 \big),\]
\[{\rm d} Y =(q^{-1}-1)^{-1}e_+ + (1+q)Ye_0 +q(q^{-m}-q^m)Y^{2}e_-,
\]
which implies on a general monomial basis element that
\begin{align*}
{\rm d} (X^i t^j Y^k)=&(q^{-1}-1)^{-1}[k]_{q}X^{i}t^jY^{k-1}e_+ + (1+q)[j+k]_{q^2}X^i t^j Y^k e_0\\
&~+\Big(q(q^{-m}-q^m)[2j+k]_{q}X^i t^j Y^{k+1} + [i]_{q^{-1}}q^{-m+j+k}X^{i-1}t^{j+1}Y^k\Big)e_-.
\end{align*}
We determine $\chi_a\in u_{p}(sl_2)$ from $\langle X^i t^j Y^k, \chi_a \rangle=\epsilon(\partial^a (X^i t^j Y^k))$ with the result
\[\chi_+=\sum\limits_{j=0}^{n-1}\delta_{j}(K)E = \sum\limits_{i,j=0}^{n-1}\dfrac{(q^{-1}-1)^{-1} q^{-ij}}{n}K^i E = \dfrac{E}{(q^{-1}-1)},\]
\[\chi_0=\sum\limits_{j=0}^{n-1}(1+q)[j]_{q^{2}}\delta_{j}(K)=\sum\limits_{i,j=0}^{n-1}\dfrac{q^{-ij}(1-q^{2j})}{n(1-q)}K^{i}=\dfrac{1-K^2}{1-q}, \]
\[\chi_-= \sum\limits_{j=0}^{n-1}q^{-m+j}F \delta_{j}(K)=\sum\limits_{i,j=0}^{n-1}\frac{q^{-m}q^{j-ij}}{n} FK^i =q^{-m}FK.\]
These are versions of similar elements found for ${\Bbb C}_q[SU_2]$ with real $q$ in \cite{Woronowicz89}.
\end{example}
\section{Construction of ${\mathfrak u}_q(sl_3)$ and ${\mathfrak c}_q[SL_3]$ by (co)double bosonisation}\label{SecSL3}
As mentioned in the introduction, double bosonisation can in principle be used iteratively to construct all the $u_q(\mathfrak{g})$ \cite{db2,Primer} and hence now, by making the corresponding co-double bosonisation at each step, an appropriate dual $c_q[G]$. The quantum-braided planes and their duals adjoined at each step generally have a more straightforward duality pairing given by braided factorial operators, see \cite{Ma:Hod}. Here we find
\[ {\mathfrak u}_q(sl_3)={\mathfrak c}_q^{2}{>\!\!\!\triangleleft\kern-.33em\cdot} \widetilde{{\mathfrak u}_q(sl_{2})}{\cdot\kern-.33em\triangleright\!\!\!<} {\mathfrak c}_q^{2}={\mathfrak c}_q^{2}{>\!\!\!\triangleleft\kern-.33em\cdot} ({\mathfrak c}_q^{1}{>\!\!\!\triangleleft\kern-.33em\cdot} {\Bbb C}{\Bbb Z}^2_n {\cdot\kern-.33em\triangleright\!\!\!<} {\mathfrak c}_q^{1}){\cdot\kern-.33em\triangleright\!\!\!<} {\mathfrak c}_q^{2}\]
for certain $n$-th roots of unity, and a parallel result for ${\mathfrak c}_q[SL_3]$. The former was explained for generic $q$ in \cite{db} but at roots of unity we need to be much more careful.
\subsection{Construction of ${\mathfrak u}_q(sl_3)$ from ${\mathfrak u}_q(sl_2)$}\label{upsl3}
The quantum group $u_q(sl_3)$ in more or less standard conventions is generated by $E_i,F_i,K_i$ for $i=1,2$, with, c.f. \cite{Jan},
\[E_i^n=F_i^n=0, \quad K_i^n=1,\]
\[K_iK_j=K_j K_i, \quad E_i K_j = q^{a_{ij}}K_i E_j, \quad K_i F_j = q^{a_{ij}}F_j K_i, \quad [E_i, F_j]=\delta_{ij}(K_i-K_i^{-1}),\]
\[\Delta K_i=K_i\otimes K_i, \quad \Delta E_i=E_i\otimes K_i + 1\otimes E_i, \quad \Delta F_i = F_i\otimes 1 + K_i^{-1}\otimes F_i,\]
where $a_{11}=a_{22}=2$ and $a_{12}=a_{21}=-1$. As before, we absorbed a factor $q-q^{-1}$ in the cross relation as a normalisation of $E_i$. We also require the $q$-Serre relations
\[ E_i^2E_j-(q+q^{-1})E_iE_jE_i+E_jE_i^2=0, \quad F_i^2F_j-(q+q^{-1})F_iF_jF_i+F_jF_i^2=0\]
for $i\ne j$. Note that $u_q(sl_2)$ appears as a sub-Hopf algebra generated by $E_1,F_1,K_1$.
Let $q$ be a primitive $n$-th root of unity with $n=2m+1$ and $p=q^{-m}=q^{1\over 2}$. Let $B={\mathfrak c}_q^2$ be the algebra generated by $e_1,e_2$ with relation $e_2e_1=q^{-m}e_1e_2$ in the category of right ${\mathfrak u}_q(sl_2)$-modules. The canonical left-action of ${\mathfrak u}_q(sl_2)$ on $B$ is given by
\begin{align}\label{u_q(sl_2) right-action on c_q^2}
\begin{split}
\left( e_1 \atop e_2 \right) \lhd K &= \langle K, \begin{pmatrix}
a & b \\ c & d
\end{pmatrix} \rangle\left( e_1 \atop e_2 \right) = \begin{pmatrix}
q^{-m} & 0\\ 0 & q^{m}
\end{pmatrix}\left( e_1 \atop e_2 \right),\\
\left( e_1 \atop e_2 \right) \lhd E &= \langle E, \begin{pmatrix}
a & b \\ c & d
\end{pmatrix} \rangle\left( e_1 \atop e_2 \right) = \begin{pmatrix}
0 & 0\\ \lambda & 0
\end{pmatrix}\left( e_1 \atop e_2 \right), \\
\left( e_1 \atop e_2 \right) \lhd F &= \langle F, \begin{pmatrix}
a & b \\ c & d
\end{pmatrix} \rangle\left( e_1 \atop e_2 \right) = \begin{pmatrix}
0 & 1\\ 0 & 0
\end{pmatrix}\left( e_1 \atop e_2 \right),
\end{split}
\end{align}
where $\lambda = q^{m}-q^{-m}$. The duality between ${\mathfrak u}_q(sl_2)$ and $c_{q^{-m}}[SL_2]$ is the standard one when the former is identified with $u_{q^{-m}}(sl_2)$, or can be obtained from Corollary~\ref{dual basis}.
\begin{lemma}\label{braided-plane c_q^2}
Let $q$ be a primitive $n$-th root of 1 with $n=2m+1$ such that $\beta^2=3$ has a solution mod $n$. Let $H=\widetilde{{\mathfrak u}_{q}(sl_2)}={\mathfrak u}_q(sl_2)\otimes {\Bbb C}_q[g]/(g^n-1)$, and let $g$ act on $e_i$ by
\[e_i\lhd g=q^{m\beta}e_i. \]
Then ${\mathfrak c}_q^2$ is a braided-Hopf algebra in the braided category of right $H$-modules with
\[e_1^n=e_2^n=0,\quad e_2e_1=q^{-m}e_1e_2, \quad \underline{\Delta}(e_i)=e_i\otimes 1 + 1 \otimes e_i, \quad \underline{\epsilon}(e_i)=0, \quad \underline{S}(e_i)=-e_i,\]
\[\Psi(e_i\otimes e_i)=q e_i \otimes e_i, \quad\Psi(e_1 \otimes e_2)=q^{-m}e_2\otimes e_1, \quad \Psi(e_2\otimes e_1)=q^{-m}e_1\otimes e_2 + (q-1) e_2 \otimes e_1.\]
\end{lemma}
\begin{proof} The quasitriangular structure of $\widetilde{{\mathfrak u}_q(sl_2)}$ is given by ${\mathcal R}_{{\mathfrak u}_q(sl_2)}{\mathcal R}_g$, where ${\mathcal R}_g=\frac{1}{n}\sum\limits_{s,t=0}^{n-1}q^{-st}g^s\otimes g^t$ and ${\mathcal R}_{{\mathfrak u}_q(sl_2)}$ is given in Lemma \ref{dboslem}. Thus, we can compute that
\[\Psi(e_i\otimes e_j)=q^{m^2\beta^2}(e_i\otimes e_j)\lhd {\mathcal R}_{{\mathfrak u}_q(sl_2)}.\]
This braiding is equal to the correctly normalised braiding in the statement (as needed for $\underline\Delta$ to extend as a homomorphism to the braided tensor product algebra) iff $m^2\beta^2=m(m-1)$ mod $n$, or $m\beta^2=m-1$ since any $m>0$ is invertible mod $n$ (this is true for $m=1$ and if $m>1$ then $m$ and $2m+1$ are coprime). Thus the condition for ${\mathfrak c}_q^2$ to form a braided-Hopf algebra in the category of $\widetilde{{\mathfrak u}_q(sl_2)}$-modules by an action of the stated form is $m(\beta^2-1)=-1=2m \mod n$, or $\beta^2= 3 \mod n$. Some version of this lemma was largely in \cite{EJB : bar}, working directly with $p=q^{-m}$.
\end{proof}
Here $\beta=0$ is only possible for $m=1$, i.e., $n=3$. In this case ${\mathfrak c}_q^2$ is already a braided-Hopf algebra in the category of ${\mathfrak u}_q(sl_2)$-modules without a central extension being needed. Otherwise, the least $n$ satisfying the condition is $n=11$ with $\beta=5$. For $n$ prime, $\beta$ exists if and only if $n=\pm 1 \mod 12$, see \cite{NumberTheory}.
The dual $B^*=({\mathfrak c}_q^2)^* \in {}_{H}M$ is generated by $f_1,f_2$ satisfying the same relations $f_2f_1=q^{-m}f_1f_2$ and additive braided coproduct as $B$ but with the left action
\begin{align}\label{u_q(sl_2) left-action on c_q^2}
\begin{split}
K\rhd \begin{pmatrix}
f_1 & f_2
\end{pmatrix} &= \begin{pmatrix}
f_1 & f_2
\end{pmatrix} \begin{pmatrix}
q^{-m} & 0\\ 0& q^m
\end{pmatrix}, \quad g \rhd f_i = q^{m\beta} f_i,\\
E\rhd \begin{pmatrix}
f_1 & f_2
\end{pmatrix} &= \begin{pmatrix}
f_1 & f_2
\end{pmatrix}\begin{pmatrix}
0 & 0 \\ \lambda & 0
\end{pmatrix}, \quad F\rhd \begin{pmatrix}
f_1 & f_2
\end{pmatrix} = \begin{pmatrix}
f_1 & f_2
\end{pmatrix}\begin{pmatrix}
0 & 1 \\ 0 & 0
\end{pmatrix}.
\end{split}
\end{align}
\begin{lemma}\label{exp of braided-plane} The quantum-braided planes ${\mathfrak c}_q^2$ and $({\mathfrak c}_q^2)^*$ in Lemma~\ref{braided-plane c_q^2} are dually paired by $\langle e_1^r e_2^s, f_1^{r'} f_2^{s'}\rangle = \delta_{r,r'}\delta_{s,s'}[r]_{q}![s]_{q}! $. \end{lemma}
\begin{proof}
It is not hard to see that $\langle e_i^r, f_i^{r'} \rangle = \delta_{r,r'}[r]_{q}! $ and this implies that
\begin{align*}
\langle e_1^r e_2^s, f_1^{r'} f_2^{s'}\rangle =& \langle e_1^r \otimes e_2 ^s , \sum\limits_{r_1=0}^{r'} \sum\limits_{s_1=0}^{s'}\begin{bmatrix}
r' \\ r_1 \end{bmatrix}_{q}\begin{bmatrix}
s' \\ s_1 \end{bmatrix}_{q} q^{-ms_1(r'-r_1)}f_1^{r_1} f_2^{s_1} \otimes f_1^{r'-r_1}f_2^{s'-s_1} \rangle\\
&=\langle e_1^r \otimes e_2 ^s, f_1^{r'} \otimes f_2^{s'} \rangle = \delta_{r,r'}\delta_{s,s'}[r]_{q}![s]_{q}!.
\end{align*}
\end{proof}
In the double bosonisation, we read the generators $e_1,e_2$ of the quantum-braided plane $B={\mathfrak c}_q^2$ as $E_{12}$ and $E_2$ respectively. Similarly, the generators $f_1,f_2$ of its dual quantum-braided plane $({\mathfrak c}_q^2)^*$ are read as $F_{12},F_2$ respectively. Also, we read the generators $E,F,K$ of ${\mathfrak u}_q(sl_2)$ as $E_1, F_1$ and $K_1$ so that
\[E_1^n=F_1^n=0, \quad K_1^n=1, \quad K_1E_1K_1^{-1}=q^{-1}E_1, \quad K_1F_1K_1^{-1}=qF_1, \quad [E_1,F_1]=K_1-K_1^{-1}.\]
\begin{lemma}\label{dbos u_q(sl_3)} Suppose the setting of Lemma~\ref{braided-plane c_q^2} with $n$ odd and $\beta^2=3$ solved mod $n$.
\begin{enumerate}
\item The double bosonisation of ${\mathfrak c}_q^2$, which we denote ${\mathfrak u}_q(sl_3)$, is generated by $E_i,F_i,K_1,g$ for $i=1,2$, with $E_1,F_1,K_1$ generating ${\mathfrak u}_q(sl_2)$ as a sub-Hopf algebra, and has
\[E_2K_1=q^{m} K_1E_2, \quad E_2g=q^{m\beta}gE_2,\quad K_1F_2=q^{m}F_2 K_1, \quad gF_2=q^{m\beta}F_2g, \]
\[[E_1,F_2]=[E_2,F_1]=0, \quad [E_2,F_2]=K_1^{m}g^{m\beta}-K_1^{-m}g^{-m\beta}, \]
\[ \{E_i^2,E_j\}=(q^m+q^{-m})E_iE_jE_i,\quad \{F_i^2,F_j\}=(q^m+q^{-m})F_iF_jF_i;\quad i\ne j,\]
\[\Delta E_2 = 1\otimes E_2 + E_2\otimes K_1^{m}g^{m\beta}, \quad \Delta F_2 = F_2 \otimes 1 + g^{-m\beta}K_1^{-m}\otimes F_2,\]
\[{\mathcal R}_{{\mathfrak u}_q(sl_3)}=\dfrac{1}{n^2}\sum\limits \dfrac{(-1)^{r+v+w}q^{vw-st-ab}}{[r]_{q^{-1}}![v]_{q^{-1}}![w]_{q^{-1}}!} F_{12}^v F_2^w F_1^r K_1^s g^{a}\otimes E_{12}^v E_2^w E_1^r K_1^t g^{b},\]
where we sum over $r,s,a,t,b,v,w$ from $0$ to $n-1$ and
\[ E_2E_1=q^m E_1E_2+ \lambda E_{12},\quad F_1F_2=q^{-m}F_2F_1 + F_{12};\quad \lambda=q^m-q^{-m}.\]
\item If $n>3$ and is not divisible by $3$ then ${\mathfrak u}_q(sl_3)$ is isomorphic to $u_{q^{-m}}(sl_3)$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) This is a direct computation using Theorem \ref{Double Bosonisation}. First, we have that $E_2h=h\o(E_2\lhd h\t)$ and $hF_2=(h\o \rhd F_2)h\t$ for all $h\in \widetilde{{\mathfrak u}_q(sl_2)}$ and using the correct actions mentioned above. Those not involving $E_{12},F_{12}$ are as listed, while two more are regarded in the statement as definitions of $E_{12},F_{12}$ in terms of the other generators. In this case the remaining cross relations
\[E_{12}K_1 = q^{-m} K_1E_{12},\quad K_1F_{12}=q^{-m}F_{12}K_1,\quad E_{12}g=q^{m\beta}gE_{12},\quad gF_{12}=q^{m\beta}F_{12}g\]
are all empty and can be dropped. Similarly, the first two of
\[ [ E_{12},F_1]=K_1^{-1}E_2,\quad [E_1,F_{12}]=\lambda F_2K_1,\quad E_{12}E_1=q^{-m}E_1E_{12},\quad F_1F_{12}=q^{m}F_{12}F_1\]
are empty and can be dropped. The remaining two and the original quantum-braided plane relations $E_{12}E_2=q^mE_2E_{12}, F_{12}F_2=q^mF_2F_{12}$ are the four $q$-Serre relations stated for $i\ne i$.
We next look at the cross relations between the two quantum-braided planes. For example,
\[[E_2,F_2]={\mathcal R}^{\t}\langle F_2, E_2 \lhd {\mathcal R}^{\o} \rangle - {\mathcal R}^{{\scriptscriptstyle -(1)}}\langle {\mathcal R}^{{\scriptscriptstyle -(2)}}\rhd F_2, E_2 \rangle \]
Putting in the form of ${\mathcal R}$ and ${\mathcal R}^{-1}$ gives the stated cross relation. One similarly has
\[ [E_{12}, F_2]=-E_1K_1^{m}g^{m\beta}, \quad [E_2, F_{12}]=\lambda g^{-m\beta}K_1^{-m}F_1,\quad [E_{12},F_{12}]= K_1^{-m}g^{m\beta}-K_1^{m} g^{-m\beta}\]
of which the first two are empty by a similar computation to the one above and the last is also empty by a more complicated calculation. In fact all these identities can be useful even though we do not include them in the defining relations. We also have
\begin{align*}
\Delta E_{2}&= 1\otimes E_{2} + \frac{1}{n^2}\sum\frac{(-1)^r}{[r]_{q^{-1}}!} q^{-st-ab}(E_{2} \lhd F_1^r K_1^s g^{a})\otimes E_1^r K_1^t g^{b}\\
&=1\otimes E_{2} + \frac{1}{n^2}\sum q^{-s(t-m)- a(b-m\beta)} E_{2}\otimes K_1^t g^{b} =1\otimes E_{2} + E_{2} \otimes K_1^{m}g^{m\beta},
\end{align*}
where we sum over $r,s,a,t,b$ from 0 to $n-1$. To compute $\Delta F_2$, we need
\[ {\mathcal R}^{-1}=S{\mathcal R}^{\o}\otimes {\mathcal R}^{\t}=\frac{1}{n^2}\sum\frac{q^{-st- ab}}{[r]_{q^{-1}}!}g^{-a}K_1^{r-s}F_1^r \otimes E_1^r K_1^tg^{b}.\]
Only the first term contributes when acting on $F_2$,
\begin{align*}
\Delta F_{2}&= F_{2}\otimes 1 + \frac{1}{n^2}\sum\frac{(-1)^r}{[r]_{q^{-1}}!}q^{-st-ab}g^{-a}K_1^{r-s}F_1^r\otimes E_1^rK_1^tg^{b}\rhd F_{2}\\
&= F_{2} \otimes 1 + \frac{1}{n^2}\sum q^{-t(s-m)-b(a-m\beta)}g^{-a}K_1^{-s}\otimes F_{2} =F_{2}\otimes 1 + g^{-m\beta}K_1^{-m}\otimes F_{2}
\end{align*}
and similarly for $\Delta E_2$. One also has
\[\Delta E_{12} = 1\otimes E_{12} + E_{12}\otimes K_1^{-m}g^{m\beta} - E_2\otimes E_1K_1^{m}g^{m\beta},\]
\[\Delta F_{12} = F_{12} \otimes 1 + g^{-m\beta}K_1^{m}\otimes F_{12} + \lambda g^{-m\beta}K_1^{-m}F\otimes F_2\]
which we did not state as $E_{12},F_{12}$ are not generators. By Theorem \ref{Double Bosonisation} and Lemma \ref{exp of braided-plane}, the quasitriangular structure of ${\mathfrak u}_q(sl_3)$ is
\begin{align}\label{quasitriangular uq(sl3)}
{\mathcal R}_{{\mathfrak u}_q(sl_3)}=(\sum\limits^{n-1}_{v,w=0} \dfrac{F_{12}^v F_2^w}{[v]_q![w]_q!}\otimes \underline{S}(E_{12}^v E_2^w)){\mathcal R}_{{\mathfrak u}_q(sl_2)}{\mathcal R}_{g},
\end{align}
where ${\mathcal R}_{{\mathfrak u}_q(sl_2)}{\mathcal R}_{g}$ is explained in the proof of Lemma \ref{braided-plane c_q^2}. By (\ref{braidedS}) for the braided-antipode, we find
\[\underline{S}(E_{12}^v E_2^w)=(-1)^{v+w}q^{\frac{v(v-1)+w(w-1)}{2}+vw}E_{12}^w E_{2}^v,\]
so that (\ref{quasitriangular uq(sl3)}) becomes the expression stated.
(2) If $m>1$, we define $\varphi : u_{q^{-m}}(sl_3)\to {\mathfrak u}_q(sl_3)$ by
\[ \varphi(E_i)=E_i,\quad \varphi(F_i)=F_i, \quad \varphi(K_1)=K_1,\quad \varphi(K_2) = K_1^{m}g^{m\beta}.\]
It is easy to see that $\varphi$ is an algebra and coalgebra map. In the other direction, when $m>1$, $\beta$ is invertible mod $n$ iff $3$ is. We then define $\phi : {\mathfrak u}_q(sl_3) \to u_{q^{-m}}(sl_3)$ by
\[ \phi(E_i)=E_i,\quad \phi(F_i)=F_i, \quad \phi(K_1)=K_1,\quad \varphi(g)=(K_1^{-m}K_2)^{1\over m \beta},\]
which is clearly inverse to $\varphi$.
\end{proof}
We again write $p=q^{-m}$ so that ${\mathfrak u}_{q}(sl_3)$ is isomorphic to $u_p(sl_3)$ under our assumptions, where $n=33$ and $\beta=6$ is the first case excluded. The double bosonisation construction also gives $\{F_{12}^{i_1}F_2^{i_2}F_1^{i_3}K_1^{i_4}g^{i_5}E_1^{i_6}E_{12}^{i_7}E_2^{i_8}\}$ as a basis of ${\mathfrak u}_{q}(sl_3)$.
\begin{example}\label{dbos u_q(sl_3) at q^3=1}
As mentioned before, when $q$ is a primitive cubic root of unity i.e., when $\beta=0$, ${\mathfrak c}_q^2$ is already a braided-Hopf algebra in the category of ${\mathfrak u}_q(sl_2)$-modules without an extension needed. Then Theorem \ref{Double Bosonisation} gives us a quasitriangular Hopf algebra, which we denote ${\mathfrak u}_q'(sl_3)$, generated by $E_{i},F_{i},K_1$ with $i=1,2$ with the relations and coproducts
\[E_iK_1=qK_1E_i, \quad K_1F_i=qF_iK_1,\quad [E_i,F_j]=\delta_{i,j} (K_1-K_1^{-1}),\]
\[ \{E_i^2,E_j\}=(q+q^{-1})E_iE_jE_i,\quad \{F_i^2,F_j\}=(q+q^{-1})F_iF_jF_i;\quad i\ne j,\]
\[\Delta E_i=1\otimes E_i+E_i\otimes K_1 , \quad \Delta F_i=F_i\otimes 1+ K_1^{-1} \otimes F_i,\]
\[{\mathcal R}_H=\dfrac{1}{9} \sum \dfrac{(-1)^{r+v+w} q^{vw-st}}{[r]_{q^{-1}}![v]_{q^{-1}}![w]_{q^{-1}}!} F_{12}^v F_2^w F_1^r K_1^s \otimes E_{12}^v E_2^w E_1^r K_1^t,\]
where the sum is over $r,s,t,v,w$ from 0 to 2. This ${\mathfrak u}_q'(sl_3)$ is not isomorphic to $u_{q^{-1}}(sl_3)$ since we do not have the generator $K_2$. However, the element $K_1^{-1}K_2$ is central and group-like in $u_{q^{-1}}(sl_3)$ and ${\mathfrak u}_q'(sl_3)\cong u_{q^{-1}}(sl_3)/\<K_1^{-1}K_2-1\>$. In addition, Lemma \ref{dbos u_q(sl_3)} still applies and $g$ is already group-like, and central when $\beta=0$. Therefore we have ${\mathfrak u}_{q}(sl_3)= {\mathfrak u}_q'(sl_3)\otimes {\Bbb C}_q[g]/(g^3-1)$ for $m=1$.
\end{example}
\subsection{Construction of ${\mathfrak c}_q[SL_3]$ from ${\mathfrak c}_q[SL_2]$}\label{cpsl3} Recall, see e.g. \cite{Foundation}, that the coquasitriangular Hopf algebra ${\Bbb C}_{q}[SL_3]$ is generated by $\mathbf{t}=(t^i{}_j)$ for $i,j=1,2,3$, with matrix-form of coproduct $\Delta \mathbf{t}=\mathbf{t}\otimes \mathbf{t}$, and for $i<k$, $j<l$, the relations
\[[t^i{}_l, t^i{}_j]_q=0,\quad [t^k{}_j,t^i{}_j]_q=0, \quad [t^i{}_l, t^k{}_j]=0, \quad [t^k{}_l, t^i{}_l]=\lambda t^i{}_l t^k{}_j,\]
\[\mathrm{det}_q(\mathbf{t}):=t^1{}_1(t^2{}_2t^3{}_3-q^{-1}t^2{}_3t^3{}_2)-q^{-1}t^1{}_2(t^2{}_1t^3{}_3-q^{-1}t^2{}_3t^3{}_1)+q^{-2}t^1{}_3(t^2{}_1t^3{}_2-q^{-1}t^2{}_2t^3{}_1)=1,\]
where $[a,b]_q:=ba-qab$ and $\lambda = q-q^{-1}$. The reduced version is denoted by $c_q[SL_3]$ and has the additional relations
\[(t^i{}_j)^n=\delta_{ij}.\]
Throughout this section we limit ourselves to $q$ a primitive $n=2m+1$-th root of unity so that ${\mathfrak c}_q[SL_2]\cong c_{q^{-m}}[SL_2]$ according to Theorem~\ref{cdbthm}. Since we only consider this case, it will be convenient to use the isomorphism to define new generators $a,b,c,d$ of ${\mathfrak c}_q[SL_2]$ related to our previous ones by $X=bd^{-1}, t=d^{-2}$ and $Y=d^{-1}c/(q^m-q^{-m})$. Then we can benefit from both the matrix form of coproduct on the new set and the dual basis feature of the original set. We let $A =\widetilde{{\mathfrak c}_q[SL_2]}= {\mathfrak c}_q[SL_2]\otimes {\Bbb C}_{q}[\varsigma]/(\varsigma^n-1)$ be the central extension dual to $\widetilde{{\mathfrak u}_q(sl_2)}={\mathfrak u}_q(sl_2)\otimes {\Bbb C}_{q}[g]/(g^n-1)$. Here $\langle \varsigma, g \rangle=q$ and ${\mathcal R}(\varsigma,\varsigma)=q$ is the coquasitriangular structure on the central extension factor. Let $B$ be a quantum-braided plane ${\mathfrak c}_q^2$ as in Lemma~\ref{braided-plane c_q^2} but viewed in the category of left comodules over $A$ with left coaction
\begin{align}\label{braided-plane left coaction}
\Delta_L\begin{pmatrix}X_1 \\ X_2\end{pmatrix}=\begin{pmatrix}
\tilde{a} & \tilde{b}\\
\tilde{c} & \tilde{d}
\end{pmatrix}\otimes \begin{pmatrix}X_1 \\ X_2\end{pmatrix};\quad \begin{pmatrix}
\tilde{a} & \tilde{b}\\
\tilde{c} & \tilde{d}
\end{pmatrix}=\begin{pmatrix}
a& b \\ c& d
\end{pmatrix}\varsigma^{m\beta},
\end{align}
where we now denote the generators $X_1,X_2$. In this case we will have
\[\Psi(X_i\otimes X_j)=q^{m^2\beta^2}R^j{}_k{}^i{}_l X_k \otimes X_l,\]
where $R$ was given in (\ref{RSL2}). We again require that $\beta^2=3 \mod n$ so that
$q^{3m^2}R$ has the correct normalisation factor $q^{3m^2+m(m+1)}=q^{-m}$ in front of the
matrix in (\ref{RSL2}), as needed to obtain a braided-Hopf algebra. One also has, c.f. \cite{Foundation},
\[\underline{\Delta}(X_1^r X_2^s) = \sum\limits_{r_1=0}^{r} \sum\limits_{s_1=0}^{s}\begin{bmatrix}
r \\ r_1 \end{bmatrix}_{q}\begin{bmatrix}
s \\ s_1 \end{bmatrix}_{q} q^{-ms_1(r-r_1)}X_1^{r_1} X_2^{s_1} \otimes X_1^{r-r_1}X_2^{s-s_1}.\]
The dual $B^*$ was likewise explained in the previous section and is now taken with generators $Y_i$ and regarded in the category of right comodules over $A$ with
\begin{align}\label{braided-plane rigt coaction}
\Delta_R \begin{pmatrix}
Y_1 & Y_2
\end{pmatrix}=\begin{pmatrix}
Y_1 & Y_2
\end{pmatrix}\otimes\begin{pmatrix}
\tilde{a} & \tilde{b}\\
\tilde{c} & \tilde{d}\\
\end{pmatrix} .
\end{align}
\begin{theorem}\label{codbos cq[SL_3]} Let $n=2m+1$ such that $\beta^2=3$ is solved mod $n$. Let $A=\widetilde{{\mathfrak c}_q[SL_2]}={\mathfrak c}_q[SL_2]\otimes {\Bbb C}_q[\varsigma]/(\varsigma^n-1)$ regarded with generators $\varsigma,\tilde{a},\tilde{b},\tilde{c},\tilde{d}$. Let $B, B^*$ be quantum-braided planes with generators $X_i,Y_i$ for $i=1,2$ as above.
\begin{enumerate}
\item The co-double bosonisation, denoted ${\mathfrak c}_q[SL_3]$, has cross relations and coproducts
\[X_iY_j=Y_jX_i, \quad X_i \varsigma = q^{m\beta}\varsigma X_i, \quad Y_i\varsigma=q^{m\beta}\varsigma Y_i, \quad \begin{pmatrix}
\tilde{a} & \tilde{b}\\ \tilde{c} & \tilde{d}
\end{pmatrix} X_1 = \begin{pmatrix}
q^{-1} X_1 \tilde{a} & q^{-1}X_1 \tilde{b} \\ q^{m}X_1 \tilde{c} & q^{m} X_1 \tilde{d}
\end{pmatrix},\]
\[\begin{pmatrix}
\tilde{a} & \tilde{b}\\ \tilde{c} & \tilde{d}
\end{pmatrix} X_2 = \begin{pmatrix}
q^{m} X_2 \tilde{a} + (q^{-1}-1)X_1 \tilde{c} & q^{m}X_2 \tilde{b} + (q^{-1}-1)X_1 \tilde{d} \\ q^{-1}X_2 \tilde{c} & q^{-1} X_2 \tilde{d}
\end{pmatrix},\]
\[Y_1 \begin{pmatrix}
\tilde{a} & \tilde{b}\\ \tilde{c} & \tilde{d}
\end{pmatrix} = \begin{pmatrix}
q \tilde{a} Y_1 & q^{-m}\tilde{b} Y_1\\ q \tilde{c} Y_1 & q^{-m}\tilde{d} Y_1
\end{pmatrix}, \quad Y_2 \begin{pmatrix}
\tilde{a} & \tilde{b}\\ \tilde{c} & \tilde{d}
\end{pmatrix} = \begin{pmatrix}
q^{-m} \tilde{a} Y_2 + (q-1)\tilde{b} Y_1 & q\tilde{b} Y_2\\ q^{-m} \tilde{c} Y_2+(q-1)\tilde{d} Y_1 & q\tilde{d} Y_2
\end{pmatrix},\]
\begin{align*}
\Delta X_1=&X_1\otimes 1 +\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \big(\tilde{a} Y_1^r Y_2^s \otimes X_1^{r+1}X_2^s +q^{-mr} \tilde{b} Y_1^r Y_2^s \otimes X_1^{r}X_2^{s+1} \big),\\
\Delta X_2 =&X_2\otimes 1+ \sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \big(\tilde{c} Y_1^r Y_2^s \otimes X_1^{r+1}X_2^s +q^{-mr} \tilde{d} Y_1^r Y_2^s \otimes X_1^{r}X_2^{s+1} \big),\\
\Delta Y_1 =&1\otimes Y_1 +\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s-1} q^{-r+ms+1} \begin{bmatrix} r+s-1\\ s \end{bmatrix}_{q} Y_1^r Y_2^s \otimes X_1^{r-1}X_2^{s}\tilde{a} \\
&+\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s-1} q^{-r-s+1} \begin{bmatrix} r+s-1\\ s-1 \end{bmatrix}_{q} Y_1^r Y_2^s \otimes X_1^{r}X_2^{s-1}\tilde{c}, \\
\Delta Y_2 =&1\otimes Y_2 +\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s-1} q^{-r+ms+1} \begin{bmatrix} r+s-1\\ s \end{bmatrix}_{q} Y_1^r Y_2^s \otimes X_1^{r-1}X_2^{s}\tilde{b} \\
&+\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s-1} q^{-r-s+1} \begin{bmatrix} r+s-1\\ s-1 \end{bmatrix}_{q} Y_1^r Y_2^s \otimes X_1^{r}X_2^{s-1}\tilde{d},
\end{align*}
\[\Delta \varsigma=\sum\limits_{r,s=0}^{n-1}q^{-m\beta(r+s)}(q-1)^{r+s}\begin{bmatrix}
r+2m\beta-1\\r
\end{bmatrix}_q \begin{bmatrix}
r+s+2m\beta-1\\s
\end{bmatrix}_q \varsigma Y_1^r Y_2^s \otimes X_1^r X_2^s \varsigma, \]
\begin{align*}
\Delta \tilde{a}=& \sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{-r-s} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{a} Y_1^r Y_2^s \otimes \big( q^{-ms}[r+1]_{q} X_1^r X_2^s\tilde{a} + [s]_{q} X_1^{r+1}X_2^{s-1}\tilde{c} \big)\\
&+\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{m(r+s)} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{b} Y_1^r Y_2^s \otimes \big(q^{-m}[r]_{q} X_1^{r-1}X_2^{s+1}\tilde{a} + \otimes q^s X_1^{r}X_2^{s}\tilde{c} \big),\\
\Delta \tilde{b}=& \sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{-r-s} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{a} Y_1^r Y_2^s \otimes \big( q^{-ms}[r+1]_{q} X_1^r X_2^s\tilde{b} + [s]_{q} X_1^{r+1}X_2^{s-1}\tilde{d} \big)\\
&+\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{m(r+s)} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{b} Y_1^r Y_2^s \otimes \big(q^{-m}[r]_{q} X_1^{r-1}X_2^{s+1}\tilde{b} + \otimes q^s X_1^{r}X_2^{s}\tilde{d} \big),\\
\Delta \tilde{c}=& \sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{-r-s} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{c} Y_1^r Y_2^s \otimes \big( q^{-ms}[r+1]_{q} X_1^r X_2^s\tilde{a} + [s]_{q} X_1^{r+1}X_2^{s-1}\tilde{c} \big)\\
&+\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{m(r+s)} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{d} Y_1^r Y_2^s \otimes \big(q^{-m}[r]_{q} X_1^{r-1}X_2^{s+1}\tilde{a} + \otimes q^s X_1^{r}X_2^{s}\tilde{c} \big),\\
\Delta \tilde{d}=& \sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{-r-s} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{c} Y_1^r Y_2^s \otimes \big( q^{-ms}[r+1]_{q} X_1^r X_2^s\tilde{b} + [s]_{q} X_1^{r+1}X_2^{s-1}\tilde{d} \big)\\
&+\sum\limits_{r,s=0}^{n-1} (q-1)^{r+s} q^{m(r+s)} \begin{bmatrix} r+s\\ s \end{bmatrix}_{q} \tilde{d} Y_1^r Y_2^s \otimes \big(q^{-m}[r]_{q} X_1^{r-1}X_2^{s+1}\tilde{b} + \otimes q^s X_1^{r}X_2^{s}\tilde{d} \big).
\end{align*}
\item If $n>3$ and is not divisible by $3$ then ${\mathfrak c}_q[SL_3]$ is isomorphic to $c_{q^{-m}}[SL_3]$ with its standard coquasitriangular structure.
\end{enumerate}
\end{theorem}
\begin{proof}
(1) From $X_2 \cdot_{\mathrm{\underline{op}}} X_1 = q X_1 \cdot_{\mathrm{\underline{op}}} X_2$, we work inductively and find that
\[X_1^{r({\mathrm{\underline{op}}})}\cdot_{\mathrm{\underline{op}}} X_2^{s({\mathrm{\underline{op}}})} = q^{\frac{-(r+s)(r+s-1)}{2}}X_1^rX_2^s, \quad \bar{S}(X_1^{r({\mathrm{\underline{op}}})}\cdot_{\mathrm{\underline{op}}} X_2^{s({\mathrm{\underline{op}}})})=(-1)^{r+s}q^{\frac{-(r+s)(r+s-1)}{2}}X_1X_2\]
where $X_1^{r({\mathrm{\underline{op}}})}$ means $X_1 \cdot_{\mathrm{\underline{op}}} X_1 \cdot_{\mathrm{\underline{op}}} \cdots $ $r$-times. We also need that
\[ \Delta_L(X_1^r) = \sum\limits_{r_1=0}^{r}\begin{bmatrix}
r\\ r_1
\end{bmatrix}_{q} \tilde{a}^{r_1}\tilde{b}^{r-r_1}\otimes X_1^{r_1}X_2^{r-r_1}, \quad \Delta_L(X_2^s) = \sum\limits_{s_1=0}^{s}\begin{bmatrix}
s\\ s_1
\end{bmatrix}_{q} \tilde{c}^{r_1}\tilde{d}^{s-s_1}\otimes X_1^{s_1}X_2^{s-s_1}\]
and that $\zeta$ commutes with $\tilde{a},\tilde{b},\tilde{c},\tilde{d}$. Then computation from Theorem~\ref{codbos} gives
\[X_iY_j=Y_jX_i, \quad X_i \varsigma = q^{m\beta}\varsigma X_i, \quad Y_i\varsigma=q^{m\beta}\varsigma Y_i, \quad \begin{pmatrix}
a & b\\ c & d
\end{pmatrix} X_1 = \begin{pmatrix}
q^{-m^2} X_1 a & q^{-m^2}X_1 b \\ q^{m^2}X_1 c & q^{m^2} X_1 d
\end{pmatrix},\]
\[\begin{pmatrix}
a & b\\ c & d
\end{pmatrix} X_2 = \begin{pmatrix}
q^{m^2} (X_2 a + (q^{m}-q^{-m})X_1 c) & q^{m^2}(X_2 b + (q^m-q^{-m})X_1 d) \\ q^{-m^2}X_2 c & q^{-m^2} X_2 d
\end{pmatrix},\]
\[Y_1 \begin{pmatrix}
a & b\\ c & d
\end{pmatrix} = \begin{pmatrix}
q^{m^2} a Y_1 & q^{-m^2}b Y_1\\ q^{m^2} c Y_1 & q^{-m^2}d Y_1
\end{pmatrix}, \ Y_2 \begin{pmatrix}
a & b\\ c & d
\end{pmatrix} = \begin{pmatrix}
q^{-m^2}(a Y_2 - (q^m-q^{-m})b Y_1) & q^{m^2}b Y_2\\ q^{-m^2} (c Y_2-(q^m-q^{-m})d Y_1) & q^{m^2}d Y_2
\end{pmatrix}\]
and hence the relations stated. The algebra generated by $X_i, Y_i, \varsigma, \tilde{a}, \tilde{b}, \tilde{c}, \tilde{d}$ is $n^8$ dimensional as required for these to be all the relations. For the coproduct, we use Lemma~\ref{exp of braided-plane} to provide a basis and dual basis of ${\mathfrak c}_q^2$ and $({\mathfrak c}_q^2)^*$. Then
\begin{align*}
\Delta& X_1 =X_1\otimes 1\\
&+ \sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q} (-1)^{r+s-r_1-s_1}q^{ s_1(r-r_1)+\frac{(r_1+s_1)(r_1+s_1+1)}{2}+r_1+s_1} \dfrac{\tilde{a} Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^{r+1} X_2^s\\
&+\sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q} (-1)^{r+s-r_1-s_1}q^{s_1(r-r_1)+\frac{(r_1+s_1)(r_1+s_1-1)}{2}+s_1-mr}(1+[r_1]_q(q-1))\\
&\kern11cm \dfrac{\tilde{b} Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^{r} X_2^{s+1}
\end{align*}
and similarly for $\Delta X_2$. Likewise,
\begin{align*}
\Delta Y_1 =&1\otimes Y_1+ \sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q}[r_1]_q (-1)^{r+s-r_1-s_1}q^{s_1(r-r_1)+\frac{(r_1+s_1-1)(r_1+s_1-2)}{2}-r+ms+r_1+s_1}\\
&\kern10cm \dfrac{Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^{r-1}X_2^s \tilde{a}\\
&+\sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q}q^{r_1+s_1-r-s+\frac{(r_1+s_1)(r_1+s_1-1)}{2}}(-1)^{r+s-r_1-s_1} ([s_1]_q+[r_1]_q [s-s_1]_q)\\
&\kern10cm\dfrac{Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^{r}X_2^{s-1} \tilde{c}
\end{align*}
and similarly for $\Delta Y_2$. Next, we have
\begin{align*}
\Delta \varsigma=& \sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1\end{bmatrix}_q \begin{bmatrix}s\\s_1 \end{bmatrix}_q q^{s_1(r-r1)+m\beta(2r_1+2s_1-r-s)+\frac{(r_1+s_1)(r_1+s_1-1)}{2}}(-1)^{r+s-r_1-s_1}\dfrac{\varsigma Y_1^r Y_2^s}{[r]_q![s]_q!}\otimes X_1^r X_2^s \varsigma.
\end{align*}
\begin{align*}
\Delta \tilde{a}&=\sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q}(-1)^{r+s-r_1-s_1} q^{s_1(r-r_1) +2r_1+s_1-r+ms+\frac{(r_1+s1)(r_1+s_1-1)}{2}}\\
&\kern10cm\dfrac{\tilde{a} Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^r X_2^s \tilde{a}\\
&+\sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q}[s-s_1]_{q}(-1)^{r+s-r_1-s_1}q^{s_1(r-r_1)+2r_1+2s_1-r-s+\frac{(r_1+s1)(r_1+s_1-1)}{2}}\\
&\kern9cm(1-q) \dfrac{\tilde{a} Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^{r+1}X_2^{s-1}\tilde{c}\\
&+\sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q}[r_1]_{q}(-1)^{r+s-r_1-s_1}q^{s_1(r-r_1)+r1+s1+mr+ms+\frac{(r_1+s1)(r_1+s_1-1)}{2}}\\
&\kern8cm(q^{-m}-q^{m})\dfrac{\tilde{b} Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^{r-1}X_2^{s+1}\tilde{a}\\
&+\sum\limits_{r,s=0}^{n-1}\sum\limits_{r_1=0}^{r}\sum\limits_{s_1=0}^{s}\begin{bmatrix}
r\\r_1 \end{bmatrix}_{q}\begin{bmatrix} s\\s_1 \end{bmatrix}_{q}(-1)^{r+s-r_1-s_1}q^{s_1(r-r_1)+\frac{(r_1+s1)(r_1+s_1-1)}{2}+r_1+2s_1+mr-s}\\
&\kern6cm(1-[r_1]_{q}[s-s_1]_{q}(q-1)^2)\dfrac{\tilde{b} Y_1^r Y_2^s}{[r]_{q}![s]_{q}!}\otimes X_1^{r}X_2^{s}\tilde{c}
\end{align*}
and similarly for $\Delta \tilde{b}, \Delta \tilde{c}, \Delta \tilde{d}$. The stated coproducts follow from the $q$-identity
\begin{align}\label{q-identity2}
\sum\limits_{r_1=0}^{r}\frac{(-1)^{r-r_1}q^{\frac{r_1(r_1+1)}{2}}q^{s r_1}}{[r_1]_{q}![r-r_1]_q!}=(q-1)^{r}\begin{bmatrix}
r+s\\r
\end{bmatrix}_q
\end{align}
for all $r,s$ (of which (\ref{q-identity1}) are specials cases) and further calculation.
(2) If $n>3$ and is not divisible by $3$ then $\beta$ is invertible mod $n$. We define $\varphi:c_{q^{-m}}[SL_3]\to {\mathfrak c}_q[SL_3]$ by
\[\varphi(t^1{}_1)=a\varsigma^{\frac{m}{\beta}}+ \lambda X_1, \quad \varphi(t^1{}_2)=b\varsigma^{\frac{m}{\beta}}+ \lambda X_1, \quad \varphi(t^1{}_3)=X_1\varsigma^{\frac{1}{\beta}},\]
\[\varphi(t^2{}_1)=c\varsigma^{\frac{m}{\beta}}+ \lambda X_2, \quad \varphi(t^2{}_2)=d\varsigma^{\frac{m}{\beta}}+ \lambda X_2 \varsigma^{\frac{1}{\beta}}Y_2, \quad \varphi(t^2{}_3)=X_2\varsigma^{\frac{1}{\beta}},\]
\[\varphi(t^3{}_1)=\lambda \varsigma^{\frac{1}{\beta}}Y_1, \quad \varphi(t^3{}_2)=(q^m-q^{-m})\varsigma^{\frac{1}{\beta}}Y_2, \quad \varphi(t^3{}_3)= \varsigma^{\frac{1}{\beta}},
\]
where $\lambda = q^m-q^{-m}$. A tedious calculation shows that this extends as an algebra map and is a coalgebra map. In the other direction, we define $\phi : {\mathfrak c}_q[SL_3] \to c_{q^{-m}}[SL_3]$ by
\[\phi(\varsigma)=(t^3{}_3)^{\beta}, \quad \phi(X_1)=t^1{}_3 (t^3{}_3)^{-1}, \quad \phi(X_2)=t^2{}_3(t^3{}_3)^{-1},\]
\[\phi(Y_1)=(q^m-q^{-m})^{-1} (t^3{}_3)^{-1} t^3{}_1, \quad \phi(Y_1)=(q^m-q^{-m})^{-1} (t^3{}_3)^{-1} t^3{}_2,\]
\[\phi(a)=t^1{}_1(t^3{}_3)^{-m}-q^m t^1{}_3 t^3{}_1(t^3{}_3)^m, \quad \phi(b)=t^1{}_2(t^3{}_3)^{-m}-q^m t^1{}_3 t^3{}_2(t^3{}_3)^m,\]
\[\phi(c)=t^2{}_1(t^3{}_3)^{-m}-q^m t^2{}_3 t^3{}_1(t^3{}_3)^m, \quad \phi(d)=t^2{}_2(t^3{}_3)^{-m}-q^m t^2{}_3 t^3{}_2(t^3{}_3)^m\]
as inverse to $\varphi$. Although one can verify these matters directly, the map $\varphi$ was obtained as adjoint to the isomorphism $u_{q^{-m}}(sl_3)\to {\mathfrak u}_q(sl_3)$ in part (2) of Lemma~\ref{dbos u_q(sl_3)} as follows. The standard duality between $u_{q^{-m}}(sl_3)$ and $c_{q^{-m}}[SL_3]$ is by
\[ \langle \mathbf{t},F_1\rangle = \mathrm{\mathbf{e}}_{12}, \quad \langle \mathbf{t},F_2\rangle = \mathrm{\mathbf{e}}_{23}, \quad \langle \mathbf{t},F_{12}\rangle = \mathrm{\mathbf{e}}_{13}, \quad \langle \mathbf{t},E_1\rangle = \lambda \mathrm{\mathbf{e}}_{21}, \quad \langle \mathbf{t},E_2\rangle = \lambda \mathrm{\mathbf{e}}_{32},\]\[ \langle \mathbf{t}, E_{12} \rangle= \lambda \mathrm{\mathbf{e}}_{31},\quad
\langle \mathbf{t},K_1\rangle = \begin{pmatrix}
q^{-m} & 0 & 0 \\
0 & q^m & 0 \\
0 & 0 & 1
\end{pmatrix}, \quad \langle \mathbf{t},K_2\rangle = \begin{pmatrix}
1 & 0 & 0 \\
0 & q^{-m} & 0 \\
0 & 0 & q^m
\end{pmatrix},
\]
where $\mathrm{\mathbf{e}}_{ij}$ is an elementary matrix with entry 1 in $(i,j)$-position and 0 elsewhere. The duality between ${\mathfrak u}_q(sl_3)$ and ${\mathfrak c}_q[SL_3]$ is part of our construction with a natural basis of ${\mathfrak c}_q[SL_3]$ built from bases of ${\mathfrak c}_q^2, ({\mathfrak c}_q^2)^*$ and $\widetilde{{\mathfrak c}_q[SL_2]}={\mathfrak c}_q[SL_2]\otimes {\Bbb C}_{q}[\varsigma]/(\varsigma^n-1)$. The first tensor factor here has a basis of monomials in $X,t,Y$ by Theorem~\ref{cdbthm}. Therefore we have $\{X_1^{i_1}X_2^{i_2}X^{i_3}t^{j_1}\varsigma^{j_2}Y^{k_1}Y_1^{k_2}Y_2^{k_3}\}$ as a basis of ${\mathfrak c}_q[SL_3]$ essentially dual to the PBW basis of ${\mathfrak u}_q(sl_3)$ in the sense
\begin{align*}
\langle X_1^{i_1}X_2^{i_2}&X^{i_3}t^{j_1}\varsigma^{j_2}Y^{k_1}Y_1^{k_2}Y_2^{k_3},F_{12}^{i'_1}F_2^{i'_2}F_1^{i'_3}K_1^{j'_1}g^{j'_2}E_1^{k'_1}E_{12}^{k'_2}E_2^{k'_3}\rangle\\
&=\delta_{i_1i'_1}\delta_{i_2i'_2}\delta_{i_3i'_3}\delta_{k_1k'_1}\delta_{k_2k'_2}\delta_{k_3k'_3}[i_1]_{q^{-1}}![i_2]_{q^{-1}}![i_3]_{q^{-1}}!q^{j_1j'_1+j_2j'_2}[k_1]_{q}![k_2]_{q}![k_3]_{q}!.
\end{align*}
This is the dual basis result for ${\mathfrak u}_q(sl_3)$ and ${\mathfrak c}_q[SL_3]$ analogous to Corollary \ref{dual basis} in the $sl_2$ case. Hence the coefficients of $\varphi(t^i{}_j)$ in this basis of ${\mathfrak c}_q[SL_3]$ will be picked out by evaluation against the dual basis $F_{12}^{i_1}F_2^{i_2}F_1^{i_3}\delta_{j_1}(K_1)\delta_{j_2}(g)E_1^{k_1}E_{12}^{k_2}E_2^{k_3}$, where $\delta_j(K_1),\delta_j(g)$ are defined as in Corollary~\ref{dual basis}. These values are given by the matrix representation as above except that we still need the matrix representation of $g$. From Lemma \ref{dbos u_q(sl_3)} we recall that that ${\mathfrak u}_q(sl_3) \cong u_{q^{-m}}(sl_3)$ with $g\mapsto(K^{-m}K_2)^{\frac{1}{m\beta}}$, hence we have $\langle \mathbf{t},g \rangle= {\rm diag}(q^{\frac{ m}{\beta}}, q^{\frac{ m}{\beta}}, q^{\frac{ 1}{\beta}})$. This gives, for example,
\begin{align*} \langle \varphi(t^1{}_1), F_{12}^{i_1}F_2^{i_2}F_1^{i_3}\delta_{j_1}(K)\delta_{j_2}(g)E_1^{k_1}&E_{12}^{k_2}E_2^{k_3} \rangle=\delta_{i_1,0}\delta_{i_2,0}\delta_{i_3,0}\delta_{j_1,-m}\delta_{j_2,\frac{m}{\beta}}\delta_{k_1,0}\delta_{k_2,0}\delta_{k_3,0}\\
&+\delta_{i_1,0}\delta_{i_2,0}\delta_{i_3,1}\delta_{j_1,m}\delta_{j_2,\frac{m}{\beta}}\delta_{k_1,1}\delta_{k_2,0}\delta_{k_3,0}(q^m-q^{-m})\\
&+\delta_{i_1,0}\delta_{i_2,1}\delta_{i_3,0}\delta_{j_1,0}\delta_{j_2,\frac{1}{\beta}}\delta_{k_1,0}\delta_{k_2,0}\delta_{k_3,1}(q^m-q^{-m}),
\end{align*}
which by summing against the dual basis implies that
\[\varphi(t^1{}_1)=t^{-m}\varsigma^{\frac{m}{\beta}}+(q^m-q^{-m})Xt^m\varsigma^{\frac{m}{\beta}}Y+(q^m-q^{-m})X_1\varsigma^{\frac{1}{\beta}}Y_1.\]
We then convert over to the $a,b,c,d$ generators as discussed.
Finally, the coquasitriangular structure of ${\mathfrak c}_q[SL_3]$ computed using Lemma~\ref{codouble coquasitriangularity} and pulled back to the $c_{q^{-m}}[SL_3]$ generators is ${\mathcal R}(\varphi(t^i{}_j),\varphi(t^k{}_l))=R^I{}_J$, where $I=(i,k)$, $J=(j,l)$ are taken in lexicographic order $(1,1),(1,2),\cdots,(3,3)$ and
\[R^I{}_J= q^{\frac{m}{3}}\begin{pmatrix}
q^{-m} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 1& 0 & q^{-m}-q^m & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 1 & 0 & 0 & 0 & q^{-m}-q^m &0 &0\\
0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & q^{-m} & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 1 & 0 & q^{-m}-q^m &0\\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & q^{-m}
\end{pmatrix}, \]
which is the standard coquasitriangular structure on the generators of $c_{p}[SL_3]$ given in \cite{Primer} when specialised to the root of unity $p=q^{-m}$. \end{proof}
\begin{remark}\label{work with Cq[GL_2]}
In the case (2) of the theorem above, we can identify $\widetilde{{\mathfrak c}_q[SL_2]}\cong c_{q^{-m}}[GL_2]$ by sending the four matrix generators of the latter to $\tilde{a}=a\varsigma^{m\beta},\tilde{b}=b\varsigma^{m\beta},\tilde{c}=c\varsigma^{m\beta},\tilde{d}=d\varsigma^{m\beta}$. The $q$-determinant $D$ maps to $\varsigma^{2m\beta}$. The converse direction is clear since $\beta$ is invertible mod $n$ when $n>3$ and not divisible by 3, so we can write $\varsigma =D^{\frac{1}{2m\beta}}$. \end{remark}
\begin{example}
At $q^3=1$, ${\mathfrak c}_q^2$ is already a braided-Hopf algebra in the category of ${\mathfrak c}_q[SL_2]$-comodules without a central extension. Therefore we can apply Theorem~\ref{codbos} and obtain a Hopf algebra, which we denote ${\mathfrak c}_q'[SL_3]$, generated by $X_i,Y_i,a,b,c,d$ with the additional cross relations and coproducts
\[\begin{pmatrix}
a & b\\ c & d
\end{pmatrix} X_1 = \begin{pmatrix}
q^{-1} X_1 a & q^{-1}X_1 b \\ qX_1 c & q X_1 d
\end{pmatrix}, \quad \begin{pmatrix}
a & b\\ c & d
\end{pmatrix} X_2 = \begin{pmatrix}
q X_2 a + (q^{-1}-1)X_1 c & qX_2 b + (q^{-1}-1)X_1 d \\ q^{-1}X_2 c & q^{-1} X_2 d
\end{pmatrix},\]
\[Y_1 \begin{pmatrix}
a & b\\ c & d
\end{pmatrix} = \begin{pmatrix}
q a Y_1 & q^{-1}b Y_1\\ q c Y_1 & q^{-1}d Y_1
\end{pmatrix}, \quad Y_2 \begin{pmatrix}
a & b\\ c & d
\end{pmatrix} = \begin{pmatrix}
q^{-1} a Y_2 + (q-1)b Y_1 & qb Y_2\\ q^{-1} c Y_2+(q-1)d Y_1 & qd Y_2
\end{pmatrix},\]
\begin{align*}
\Delta X_1&= X_1\otimes 1 + a\otimes X_1 + b\otimes X_2 + \lambda aY_1\otimes X_1^2 + \lambda bY_2 \otimes X_2^2 + \lambda aY_2\otimes X_1X_2 + \lambda q^2 bY_1\otimes X_1X_2\\
&+\lambda^2 aY_2^2 \otimes X_1X_2^2 + \lambda^2 q bY_1^2\otimes X_1^2X_2 + \lambda^2 [2]_{q}aY_1Y_2\otimes X_1^2X_2 + \lambda^2 [2]_{q} q^2 bY_1Y_2 \otimes X_1X_2^2,\\
\Delta Y_1&=1\otimes Y_1 + Y_1\otimes a + Y_2\otimes c + \lambda q^2 Y_1^2\otimes X_1a + \lambda q^2 Y_2^2\otimes X_2c + \lambda q Y_1Y_2\otimes X_2a+ \lambda q^2 Y_1Y_2\otimes X_1c\\
&+\lambda^2 q^2 Y_1Y_2^2\otimes X_2^2a + \lambda^2 [2]_{q}q Y_1Y_2^2\otimes X_1X_2c + \lambda^2 [2]_{q}Y_1^2Y_2\otimes X_1X_2a + \lambda^2 q Y_1^2 Y_2 \otimes X_1^2c,\\
\Delta a &=a\otimes a + b\otimes c +\lambda [2]_{q}q^2 aY_1\otimes X_1a + \lambda q aY_2\otimes X_2a + \lambda q^2 aY_2\otimes X_1c + \lambda [2]_{q} q^{-1} bY_2\otimes X_2c \\
&+\lambda bY_1\otimes X_2a + \lambda q bY_1\otimes X_1c + \lambda^2 q^2 aY_2^2 \otimes X_2^2a + \lambda^2 q^2 bY_1^2 \otimes X_1^2c + \lambda^2 [2]_{q} q aY_2^2\otimes X_1X_2c,\\
&+\lambda^2 [2]_{q} q bY_1^2\otimes X_1X_2a + \lambda^2 [2]_q^2 aY_1Y_2 \otimes X_1X_2a + \lambda^2[2]_q^2 bY_1Y_2\otimes X_1X_2c \\
&+ \lambda^2 [2]_{q} q aY_1Y_2\otimes X_1^2c + \lambda^2 [2]_{q}q bY_1Y_2\otimes X_2^2a,
\end{align*}
and similarly for the remaining coproducts. Here $\lambda=q-1$. This ${\mathfrak c}_q'[SL_3]$ is dual to ${\mathfrak u}_q'(sl_3)$ in Example~\ref{dbos u_q(sl_3) at q^3=1} and it is not isomorphic to $c_{q^{-1}}[SL_3]$, but rather to a sub-Hopf algebra by $\phi:{\mathfrak c}_q'[SL_3]\hookrightarrow c_{q^{-1}}[SL_3]$ with
\[\phi(X_1)=t^1{}_3(t^3{}_3)^{-1}, \quad \phi(X_2)=t^2{}_3(t^3{}_3)^{-1}, \quad \phi(Y_1)=-\dfrac{t^3{}_1(t^3{}_3)^{-1}}{\lambda}, \quad \phi(Y_2)=-\dfrac{t^3{}_2(t^3{}_3)^{-1}}{\lambda},\]
\[\phi(a) = t^1{}_1(t^3{}_3)^{-1}-t^1{}_3(t^3{}_3)^{-1}t^3{}_1(t^3{}_3)^{-1}, \quad \phi(b) = t^1{}_2(t^3{}_3)^{-1}-t^1{}_3(t^3{}_3)^{-1}t^3{}_2(t^3{}_3)^{-1}, \]
\[\phi(c) = t^2{}_1(t^3{}_3)^{-1}-t^2{}_3(t^3{}_3)^{-1}t^3{}_1(t^3{}_3)^{-1}, \quad \phi(d) = t^2{}_2(t^3{}_3)^{-1}-t^2{}_3(t^3{}_3)^{-1}t^3{}_2(t^3{}_3)^{-1}. \]
Moreover, ${\mathfrak c}_q'[SL_3]$ is a coquasitriangular Hopf algebra by Lemma~\ref{codouble coquasitriangularity}. Writing $s^1{}_1=a, s^1{}_2=b, s^2{}_1=c,s^2{}_2=d$ for the matrix form of the generators of ${\mathfrak c}_q[SL_2]$, the coquasitriangular structure of ${\mathfrak c}_q'[SL_3]$ comes out as
\[{\mathcal R}(s^i{}_j,s^k{}_l)=R^i{}_j{}^k{}_l,\quad {\mathcal R}(X_i,Y_j)=-\delta_{i,j}, \quad {\mathcal R}(X_i,X_j)={\mathcal R}(Y_i,Y_j)={\mathcal R}(Y_i,X_j)=0,\]
\[{\mathcal R}(X_i,s^j{}_k)={\mathcal R}(Y_i,s^j{}_k)={\mathcal R}(s^i{}_j,X_k)={\mathcal R}(s^i{}_j,Y_k)=0,\]
\[{\mathcal R}(X_is^j{}_k,s^u{}_vY_w)=-\delta_{w_1,i}R^j{}_{j_1}{}^{w_1}{}_w R^{j_1}{}_k{}^u{}_v, \quad {\mathcal R}(s^i{}_jY_k, X_u s^v{}_w)=0,\]
where $R$ is as in (\ref{RSL2}) with $m=1$.
Theorem~\ref{codbos cq[SL_3]} (1) still applies at $q^3=1$ with $\beta=0$ giving that $\varsigma$ is central and group-like in ${\mathfrak c}_q[SL_3]$ and that ${\mathfrak c}_{q}[SL_3]\cong {\mathfrak c}_q'[SL_3]\otimes {\Bbb C}_q[\varsigma]/(\varsigma^3-1)$.
\end{example}
\subsection{Fermionic version of ${\Bbb C}_q[SL_3]$}\label{fermionic}
Here we similarly apply co-double bosonisation but this time to obtain a part-fermionic version of ${\Bbb C}_q[SL_3]$ by using the fermionic quantum-braided plane. We no longer work at roots of unity but rather with $q$ generic and also, in the spirit of Remark~\ref{work with Cq[GL_2]}, we take as our middle Hopf algebra $A = {\Bbb C}_q[GL_2]$, the coquasitriangular Hopf algebra generated by $\tilde{a}, \tilde{b}, \tilde{c}, \tilde{d}$ with the same $q$-commutation relations and coalgebra structure as ${\Bbb C}_{q}[SL_2]$, but with $D=\tilde{a}\tilde{d}-q^{-1}\tilde{b}\tilde{c}=\tilde{d}\tilde{a}-q\tilde{b}\tilde{c}$ inverted. The antipode and coquasitriangular structure are given in matrix form by
\[S\begin{pmatrix}
\tilde{a} & \tilde{b} \\
\tilde{c} & \tilde{d}
\end{pmatrix} = D^{-1}\begin{pmatrix}
\tilde{d} & -q\tilde{b}\\
-q^{-1}\tilde{c} & \tilde{a}
\end{pmatrix}, \quad R = -q^{-1} \begin{pmatrix}
q & 0 & 0 & 0\\ 0 & 1 & q-q^{-1} & 0\\ 0 & 0 & 1 & 0\\ 0& 0 & 0 & q
\end{pmatrix}.\]
In fact the normalisation of $R$ here can be chosen freely (there is a 1-parameter family of such quasitriangular structures on this Hopf algbra) which we have fixed so that we have $B={\Bbb C}_q^{0|2} \in {}^{A}{\mathcal M}$ as a fermionic quantum-braided plane generated by $e_1, e_2$ with the relations and coproduct and braiding
\[e_i^2=0, \quad e_2e_1+q^{-1}e_1e_2=0, \quad \underline{\Delta} e_i = e_i\otimes 1 + 1\otimes e_i,\quad \underline{\epsilon}e_i=0, \quad \underline{S}e_i=-e_i,\]
\[\Psi(e_i\otimes e_i)=-e_i\otimes e_i, \quad \Psi(e_1\otimes e_2)=-q^{-1}e_2\otimes e_1, \quad \Psi(e_2\otimes e_1)=-q^{-1}e_1\otimes e_2 - (1-q^{-2})e_2\otimes e_1.\]
This has a left ${\Bbb C}_{q}[GL_2]$-coaction as in (\ref{braided-plane left coaction}). Similarly, its dual $B^* =
({\Bbb C}_q^{0|2})^*$ lives in the category of right ${\Bbb C}_q[GL_2]$-comodules with coaction as in (\ref{braided-plane rigt coaction}).
\begin{proposition}
Let $q\in {\Bbb C}^*$ not be a root of unity. The co-double bosonisation $B^{{\mathrm{\underline{op}}}}{>\!\!\!\triangleleft\kern-.33em\cdot} A {\cdot\kern-.33em\triangleright\!\!\!<} B^*$ with the above $B,A,B^{*}$ is a coquasitriangular Hopf algebra ${\Bbb C}_q^{fer}[SL_3]$ generated by $e_i, f_i$ for $i=1,2$ and $\tilde{a},\tilde{b}, \tilde{c}, \tilde{d},D,D^{-1}$, with cross relations and coproducts
\[f_i e_j = e_j f_i, \quad \begin{pmatrix} \tilde{a} & \tilde{b} \\ \tilde{c} & \tilde{d} \end{pmatrix}e_1 = \begin{pmatrix} -e_1 \tilde{a} & -e_1 \tilde{b} \\ -qe_1\tilde{c} & -qe_1\tilde{d} \end{pmatrix},\]
\[ \begin{pmatrix} \tilde{a} & \tilde{b} \\ \tilde{c} & \tilde{d}\end{pmatrix}e_2 = \begin{pmatrix} -qe_2\tilde{a}-(1-q^2)e_1 \tilde{c} & -q e_2\tilde{b} - (1-q^2)e_1\tilde{d} \\ -e_2\tilde{c} & -e_2\tilde{d} \end{pmatrix}, \]
\[ f_1 \begin{pmatrix} \tilde{a} & \tilde{b} \\ \tilde{c} & \tilde{d}\end{pmatrix} = \begin{pmatrix} -\tilde{a} f_1 & -q^{-1}\tilde{b} f_1 \\ -\tilde{c} f_1 & -q^{-1}\tilde{d} f_1 \end{pmatrix}, \quad f_2 \begin{pmatrix} \tilde{a} & \tilde{b} \\ \tilde{c} & \tilde{d} \end{pmatrix} = \begin{pmatrix} -q^{-1}\tilde{a} f_2 - (1-q^{-2})\tilde{b} f_1 & -\tilde{b} f_2\\ -q^{-1}\tilde{c} f_2 - (1-q^{-2})\tilde{d} f_1 & -\tilde{d} f_2 \end{pmatrix}, \]
\[\Delta e_1 = e_1 \otimes 1 + \tilde{a} \otimes e_1 + \tilde{b}\otimes e_2 +(1-q^{-2})(q^{-1}\tilde{b} f_1 - \tilde{a} f_2)\otimes e_1 e_2,\]
\[\Delta e_2 = e_2 \otimes 1 + \tilde{c}\otimes e_1 + \tilde{d} \otimes e_2 +(1-q^{-2})(q^{-1}\tilde{d} f_1 - \tilde{c} f_2)\otimes e_1 e_2,\]
\[\Delta f_1 = 1 \otimes f_1 + f_1 \otimes \tilde{a} + f_2 \otimes \tilde{c} +(1-q^2)f_1f_2\otimes (e_1\tilde{c}-q^{-1}e_2\tilde{a}),\]
\[\Delta f_2 = 1 \otimes f_2 + f_1 \otimes \tilde{b} + f_2 \otimes \tilde{d} +(1-q^2)f_1f_2\otimes (e_1\tilde{d}-q^{-1}e_2\tilde{b}),\]
\[\Delta \tilde{a} = \tilde{a}\otimes \tilde{a} + \tilde{b}\otimes \tilde{c} + (q-q^{-1})(\tilde{b} f_1-q\tilde{a} f_2)\otimes(e_1\tilde{c}-q^{-1}e_2\tilde{a}),\]
\[\Delta \tilde{b} = \tilde{a}\otimes \tilde{b} + \tilde{b} \otimes \tilde{d} + (q-q^{-1})(\tilde{b} f_1-q\tilde{a} f_2)\otimes(e_1\tilde{d}-q^{-1}e_2\tilde{b}),\]
\[\Delta \tilde{c} = \tilde{c}\otimes \tilde{a} + \tilde{d}\otimes \tilde{c} + (q-q^{-1})(\tilde{d} f_1-q\tilde{c} f_2)\otimes(e_1\tilde{c} -q^{-1}e_2\tilde{a}),\]
\[\Delta \tilde{d} = \tilde{c} \otimes \tilde{b} + \tilde{d}\otimes \tilde{d} + (q-q^{-1})(\tilde{d} f_1-q\tilde{c} f_2)\otimes(e_1\tilde{d}-q^{-1}e_2\tilde{b}).\]
\end{proposition}
\begin{proof}
First note that
\[{\mathcal R}(S\tilde{a},\tilde{a})= {\mathcal R}(S\tilde{d},\tilde{d})=-1, \quad {\mathcal R}(S\tilde{a},\tilde{d})={\mathcal R}(S\tilde{d},\tilde{a})=-q, \quad {\mathcal R}(S\tilde{b},\tilde{c})=-(1-q^{2})\]
and zero on other cases of this form. Then the inverse braiding is
\[\Psi^{-1}(e_1 \otimes e_2) = {\mathcal R}(Se_{1}^{\bar{\o}},e_2^{\bar{\o}})e_2^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes e_1^{\bar{{}_{\scriptscriptstyle(\infty)}}} = -q e_2 \otimes e_1 - (1-q^{2})e_1 \otimes e_2,\]
\[\Psi^{-1}(e_2\otimes e_1)= {\mathcal R}(Se_2^{\bar{\o}}, e_1^{\bar{\o}})e_1^{\bar{{}_{\scriptscriptstyle(\infty)}}}\otimes e_2^{\bar{{}_{\scriptscriptstyle(\infty)}}} = -qe_1 \otimes e_2,\]
with the result that $\bar{S}(e_1\cdot_{\mathrm{\underline{op}}} e_2)=q^{2} e_1\cdot_{\mathrm{\underline{op}}} e_2$ and $e_2 \cdot_{\mathrm{\underline{op}}} e_1 + q^{-1} e_1 \cdot_{\mathrm{\underline{op}}} e_2 = 0$ in $B^{{\mathrm{\underline{op}}}}$. We now apply the co-double bosonisation theorem. It is easy to see that $f_ie_j \equiv (1\otimes 1 \otimes f_i)(e_j \otimes 1 \otimes 1)=e_j\otimes 1 \otimes 1 \equiv e_jf_i$. Next, we compute that for any $s^i{}_j \in {\Bbb C}_q[GL_2]$, where $s^1{}_1=\tilde{a}, s^1{}_2=\tilde{b}, s^2{}_1=\tilde{c}, s^2{}_2=\tilde{d}$,
\[s^i{}_j e_k = e_k^{\bar{{}_{\scriptscriptstyle(\infty)}}}(s^i{}_j)\t {\mathcal R}(S(s^i{}_j)\o ,e_k^{\bar{\o}}) = \sum\limits_{l=1}^{2}e_k^{\bar{{}_{\scriptscriptstyle(\infty)}}}s^l{}_j {\mathcal R}(Ss^i{}_l,e_k^{\bar{\o}}),\]
\[f_k s^i{}_j = (s^i{}_j)\o f_{k}^{\bar{{}_{\scriptscriptstyle(0)}}}{\mathcal R}(f_{k}^{\bar{\o}}, (s^i{}_j)\t) = \sum\limits_{l=1}^{2} s^i{}_l f_{k}^{\bar{{}_{\scriptscriptstyle(0)}}} {\mathcal R}(f_{k}^{\bar{\o}}, s^l{}_j),\]
which comes out as the stated cross relations. Now let
\[\{ e_a\}=\{1,e_1,e_2,e_1e_2\}, \quad \{f^a \}= \{1,f_1,f_2, f_1f_2 \}\]
be a basis and dual basis of $B,B^*$ respectively. Then
\[\Delta e_i = e_i \otimes 1 +\sum\limits_{a=1}^{2} e_i^{\bar{\o}}\o f^a\otimes (e_a\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} e_i^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_a\underline{\t}) \mathcal{R}(e_a\underline{\o}^{\bar{\o}}, e_i^{\bar{\o}}\t),\]
\[\Delta f_i = 1 \otimes f_i +\sum\limits_{a=1}^{2} f^a\otimes(e_a\underline{\o}\cdot_{\mathrm{\underline{op}}} \bar{S}e_a\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}})f_{i}^{\bar{\o}}\t {\mathcal R}(Sf_{i}^{\bar{\o}}\o, e_a\underline{\th}^{\bar{{}_{\scriptscriptstyle(\infty)}}}) \langle f_i, e_a\underline{\t} \rangle,\]
\[\Delta s^i{}_j = \sum\limits_{a,k,l,r=1}^{2}s^i{}_k f^{a}\otimes (e_a\underline{\o}^{\bar{{}_{\scriptscriptstyle(\infty)}}}\cdot_{\mathrm{\underline{op}}} \bar{S}e_a\underline{\t}^{\bar{{}_{\scriptscriptstyle(\infty)}}})s^r{}_j {\mathcal R}(e_a\underline{\o}^{\o}, s^k{}_l){\mathcal R}(Ss^l{}_r, e_a\underline{\t}^{\o}),\]
which comes out as stated for all $i,j\in\{1,2\}$. Finally, we let
\[s_1 = (q-q^{-1})(\tilde{b} f_1-q\tilde{a} f_2), \quad s_2 = (q-q^{-1})(\tilde{d} f_1-q\tilde{c} f_2),\]
\[t_1 = e_1\tilde{c}-q^{-1}e_2\tilde{a}, \quad t_2 = e_1 \tilde{d} -q^{-1}e_2\tilde{b}, \]
and write ${\Bbb C}_q^{fer}[SL_3]$ as having a matrix of generators $t^i{}_j$, where now $i,j\in \{1,2,3\}$, by
\[\mathbf{t}=\begin{pmatrix}
t^1{}_1 & t^1{}_2 & t^1{}_3\\
t^2{}_1 & t^2{}_2 & t^2{}_3\\
t^3{}_1 & t^3{}_2 & t^3{}_3
\end{pmatrix} = \begin{pmatrix}
X & t_1 & t_2\\
s_1 & \tilde{a} & \tilde{b}\\
s_2 & \tilde{c} & \tilde{d}
\end{pmatrix}; \]
\begin{align}\label{elementX}
X=D + (t_1D^{-1}(\tilde{d} s_1-q \tilde{b} s_2)-q^{-1} t_2D^{-1}(\tilde{c} s_1-q \tilde{a} s_2)).
\end{align}
Here $D$ obeys $Dt_i=qt_iD$ and $Ds_i=qs_iD$ for $i=1,2$. The coproduct now has the standard matrix form $\Delta \mathbf{t}=\mathbf{t}\otimes \mathbf{t}$ and in these terms the quadratic relations are
\[(t^1{}_2)^2=(t^1{}_3)^2=(t^2{}_1)^2=(t^3{}_1)^2=0,\]
\[[t^1{}_2,t^1{}_1]_{q^{-1}}=[t^1{}_3,t^1{}_1]_{q^{-1}}=[t^2{}_1,t^1{}_1]_{q^{-1}}= [t^3{}_1,t^1{}_1]_{q^{-1}}=[t^2{}_3,t^2{}_2]_{q}=[t^3{}_2,t^2{}_2]_{q}=0,\]
\[[t^3{}_3,t^2{}_3]_{q}=[t^3{}_3,t^3{}_2]_{q}=[t^2{}_1,t^1{}_2]=[t^2{}_1,t^1{}_3]=[t^3{}_1,t^1{}_2]=[t^3{}_1,t^1{}_3]=[t^3{}_2, t^2{}_3]=0,\]
\[[t^2{}_2,t^1{}_1]=-\lambda t^1{}_2 t^2{}_1, \quad [t^2{}_3,t^1{}_1]=-\lambda t^1{}_3 t^2{}_1, \quad [t^3{}_2,t^1{}_1]=-\lambda t^1{}_2 t^3{}_1,\]
\[[t^3{}_3,t^1{}_1]=-\lambda t^1{}_3 t^3{}_1, \quad [t^3{}_3,t^2{}_2]= \lambda t^2{}_3 t^3{}_2, \quad \{t^1{}_3,t^1{}_2 \}_{q^{-1}} = \{t^3{}_1,t^2{}_1 \}_{q^{-1}}=0,\]
\[\{t^2{}_2,t^1{}_2\}_{q}=\{t^2{}_2,t^2{}_1\}_{q}=\{t^2{}_3,t^1{}_3\}_{q}=\{t^2{}_3, t^2{}_1\}_{q}= \{t^3{}_2,t^1{}_2\}_{q}=\{t^3{}_2,t^3{}_1\}_{q}=0,\]
\[\{t^3{}_3,t^1{}_3\}_{q}=\{t^3{}_3, t^3{}_1\}_{q}=\{t^2{}_2,t^1{}_3\}=\{t^3{}_1,t^2{}_2\}=\{t^3{}_1,t^2{}_3\}=\{t^3{}_2, t^1{}_3\}=0,\]
\[\{t^2{}_3,t^1{}_2\}=-\lambda t^1{}_3 t^2{}_2, \quad \{t^3{}_3,t^1{}_2\}=-\lambda t^1{}_3 t^3{}_2, \quad \{t^3{}_2,t^2{}_1\}= \lambda t^2{}_2 t^3{}_1,\quad \{t^3{}_3,t^2{}_1\}=\lambda t^2{}_3 t^3{}_1,\]
where $[\ ,\ ]_{q}$ is as before, similarly $\{a,b\}_{q}=a b+q b a$ for any $a,b$ is the $q$-anti-commutator, and $\lambda =q-q^{-1}$. Using Lemma \ref{codouble coquasitriangularity}, the values ${\mathcal R}(t^i{}_j,t^k{}_l)$ of the coquasitriangular structure of ${\Bbb C}_q^{fer}[SL_3]$ come out, in the same conventions as in the proof of part (2) of Theorem~\ref{codbos cq[SL_3]}, as
\[ R^I{}_J=\begin{pmatrix}
q^{-2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
0 & q^{-1} & 0 & -q^{-1}\lambda & 0 & 0 & 0 & 0 & 0\\
0 & 0 & q^{-1} & 0 & 0 & 0 & -q^{-1}\lambda & 0 & 0 \\
0 & 0 & 0 & q^{-1} & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & -1 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & -q^{-1} & 0 & -q^{-1}\lambda &0\\
0 & 0 & 0 & 0 & 0 & 0 & q^{-1} & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & -q^{-1} & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -1
\end{pmatrix}.\]
Note that since $t^1{}_1$ was defined in terms of the other generators including the $q$-sub-determinant $D=t^2{}_2 t^3{}_3-q^{-1}t^2{}_3 t^3{}_2$, there are in fact only $8$ algebra generators and 28 $q$-(anti)commutation relations other than the nilpotency ones and those involving $t^1{}_1$, putting this conceptually on a par with ${\Bbb C}_q[SL_3]$. Instead of a cubic $q$-determinant relation, we can regard (\ref{elementX}) as the cubic-quartic relation
\[Dt^1{}_1-q t^1{}_2(t^3{}_3t^2{}_1-qt^2{}_3t^3{}_1)+t^1{}_3(t^3{}_2t^2{}_1-qt^2{}_2t^3{}_1)=D^2.\]
Also note that (\ref{quacom}) in the `R-matrix' form $R^i{}_m{}^k{}_n t^m{}_j t^n{}_l=t^k{}_n t^i{}_m R^m {}_j{}^n{}_l$ (sum over repeated indices) encodes exactly the quadratic relations above for ${\Bbb C}_q^{fer}[SL_3]$ including the nilpotent ones.
\end{proof}
|
\section{Introduction}
\label{intro}
Sgr~A$^\star$, the supermassive black hole (BH) at the Milky Way's center, with
a bolometric luminosity of $L\sim10^{36}$~erg~s$^{-1}$ is currently characterised
by an exceptionally low Eddington ratio ($\sim10^{-8}$; Genzel et al. 2010a),
despite indications that Sgr~A$^\star$\ might have been brighter
in the past (see Ponti et al. 2013 for a review).
Therefore, Sgr~A$^\star$\ provides us with the best chance to get a glimpse of the
physical processes at work in quiescent BH.
Sgr~A$^\star$\ has been intensively studied over the past several decades
at various wavelengths. The black points
(upper-lower limits) in Fig. \ref{SED} show a compilation
of measurements of Sgr~A$^\star$'s quiescent emission from radio to mid-IR
(values are taken from Falcke et al. 1998; Markoff et al. 2001; An et al. 2005;
Marrone et al. 2006; Sch{\"o}del et al. 2007; 2011; Dodds-Eden et al. 2009;
Bower et al. 2015; Brinkerink et al. 2015; Liu et al. 2016; Stone et al. 2016)
as well as the radiatively inefficient accretion flow model proposed
by Yuan et al. (2003). The bulk of Sgr~A$^\star$'s steady
radiation is emitted at sub-mm frequencies, forming the so called "sub-mm
bump" (dot-dashed line in Fig. \ref{SED}). This emission is linearly polarised
(2-9~\%; Marrone et al. 2006; 2007), slowly variable and decreases rapidly
with frequency (with stringent upper limits in the mid-IR band; Sch\"{o}del et
al. 2007; Trap et al. 2011). This indicates that the sub-mm radiation is
primarily due to optically thick synchrotron radiation originating in the central
$\sim10$~$R_S$\footnote{$R_S$ is the Schwarzschild radius
$R_S=2GM_{BH}/c^2$, where $M_{BH}$ is the BH mass, $G$ the
gravitational constant and $c$ the speed of light.} and produced by
relativistic ($\gamma_e\sim10$, where $\gamma_e$ is the electron
Lorentz factor) thermal electrons with temperature
and densities of $T_e\sim$ few $10^{10}~K$ and $n_e\sim10^6$~cm$^{-3}$,
embedded in a magnetic field with a strength of $\sim10-50$~G (Loeb \& Waxman
2007; Genzel et al. 2010a; in Fig. \ref{SED} a possible inverse Compton
component is also shown). Moreover, Faraday rotation measurements
constraint the accretion rate at those scales to be within
$2\times10^{-9}$ and $2\times10^{-7}$~M$_\odot$ yr$^{-1}$ (Marrone
et al. 2006; 2007; Genzel et al. 2010a).
At low frequency ($\nu<10^{11}$~Hz) Sgr~A$^\star$'s SED changes slope
($F_\nu\propto\nu^{0.2}$) showing excess emission above
the extrapolation of the thermal synchrotron radiation and variability
on time-scales of hours to years (Falcke et al. 1998; Zhao et al. 2003; 2004;
Herrnstein et al. 2004). This suggests either the presence of a non-thermal tail
in the electron population, taking $\sim1$~\% of the steady state electron
energy (\"{O}zel et al. 2000; see dashed line in Fig. \ref{SED}) or a compact
radio jet (Falcke et al. 1998; Mo\'scibrodzka et al. 2009; 2013; 2014).
The presence of this non-thermal tail is well constrained at low radio
frequencies, while its extrapolation in the mid and near infra-red band
is rather uncertain (see dashed line in Fig. \ref{SED}).
Sgr~A$^\star$\ also appears as a faint ($L_{2-10~keV}\sim2\times10^{33}$~erg~s$^{-1}$)
X-ray source (Baganoff et al. 2003; Xu et al. 2006) observed to be extended
with a size of about $\sim1^{\prime\prime}$. The observed size is comparable
to the Bondi radius and the quiescent X-ray emission is thought to be the
consequence of material that is captured at a rate of $10^{-6}$~M$_\odot$ yr$^{-1}$
from the wind of nearby stars (Melia 1992; Quataert 2002; Cuadra et al. 2005;
2006; 2008). Indeed, this emission is thought to be produced via
bremsstrahlung emission from a hot plasma with $T\sim7\times10^7$~K, density
$n_e\sim100$~cm$^{-3}$ emitted from a region $\sim10^5$~R$_S$
(Quataert 2002; see dotted line in Fig. \ref{SED}).
\begin{figure}
\includegraphics[height=0.49\textwidth,angle=90]{X-rayNIR_r2.ps}
\caption{Multi-wavelength emission from Sgr~A$^\star$. The radio to mid-IR data
points (open circles as well as upper-lower limits) constrain Sgr~A$^\star$'s quiescent
emission (these constraints are taken from the literature, see text). The black
lines show the radiatively inefficient accretion disc model proposed by Yuan
et al. (2003). The dash-dotted black line shows the contribution from the thermal
electrons with $T_e\sim few~10^{10}~K$ and $n_e\sim10^6$~cm$^{-3}$,
embedded in a magnetic field with strength of $\sim10-50$~G producing the
sub-mm peak and (possibly) inverse Compton emission at higher energies.
The dashed line shows the contribution from a non-thermal tail in the electron
population, while the dotted line shows the bremsstrahlung
emission from hot plasma at the Bondi radius (Quataert 2002).
The blue filled data points shows the mean NIR and X-ray spectra of the very
bright flare VB3 and the blue solid line shows the best fit power-law
with cooling break (see also Fig. 6). }
\label{SED}
\end{figure}
For more than a decade it has been known that Sgr~A$^\star$\ also shows flaring
activity both in X-rays and IR (Baganoff et al. 2001; Goldwurm et al. 2003;
Genzel et al. 2003; Ghez et al. 2004; Porquet et al. 2003; 2008;
B\'elanger et al. 2005; Eckart et al. 2004; 2006; Marrone et al. 2008;
Nowak et al. 2012; Haubois et al. 2012; Neilsen et al. 2013; 2015;
Degenaar et al. 2013; Barri\`ere et al. 2014; Moussoux et al. 2015;
Ponti et al. 2015a; Yuan \& Wang 2016).
X-ray flares appear as clear enhancements above the
constant quiescent emission, with peak luminosities occasionally exceeding
the quiescent luminosity by up to two orders of magnitude
(see the blue points in Fig. \ref{SED} for an example of a very bright flare;
Baganoff et al. 2001; Porquet et al. 2003; 2008; Nowak et al. 2012).
X-ray flare durations, fluences, and peak luminosities are correlated
(Neilsen et al. 2013). Moreover, weak X-ray flares are more common
than strong ones (Neilsen et al. 2013; Ponti et al. 2015a).
The appearance of the X-ray light curves suggests that flares are individual
and distinct events, randomly punctuating an otherwise quiescent source
(Neilsen et al. 2013; Ponti et al. 2015a).
Typically X-ray flares coincide with clear peaks in the near infrared (NIR) light
curves (e.g. Genzel et al. 2003; Clenet et al. 2004; Ghez et al. 2004;
Eckart et al. 2006; 2008; Meyer et al. 2006; 2007; 2008;
Yusef-Zadeh et al. 2006; 2009;
Hornstein et al. 2007; Do et al. 2009). However, the appearance of the NIR
light curves is significantly different from the X-ray ones. Indeed, the NIR and
sub-mm emission of Sgr~A$^\star$\ are continuously varying and they can be described
by a red noise process at high frequencies, breaking at time-scales longer
than a fraction of a day\footnote{Interestingly, AGN of similar BH mass,
but clearly much higher accretion rate, show power density spectra (PDS)
of their X-ray light curves consistent with the NIR PDS of Sgr~A$^\star$\
(Meyer et al. 2009). } (Do et al. 2009; Meyer et al. 2009;
Dodds-Eden et al. 2011; Witzel et al. 2012; Dexter et al. 2014; Hora et al. 2014).
Therefore, the NIR light curves do not support the notion of flares
as individual events; they would alternatively corroborate the concept that flares
are simply peaks of emission on a continuous red noise process.
Despite the fact that the notion of NIR-flares is still unsettled, {\it we
will refer to the X-ray flares, and by extension, to the NIR peaks in emission,
as flares throughout this paper}.
The origin and radiative mechanism of the flares of Sgr~A$^\star$\ are still
not completely understood. Several multi-wavelength campaigns have been
performed, but the radiative mechanisms at work during the flares is still highly
debated (Eckart et al. 2004, 2006, 2008, 2009, 2012; Yusef-Zadeh et al. 2006, 2008,
2009; Hornstein et al. 2007; Marrone et al. 2008; Dodds-Eden et al. 2009;
Trap et al. 2011; Barriere et al. 2014).
Indeed, even though 15 years have passed since
the launch of {\it XMM-Newton}\ and {\it Chandra}, simultaneous X-ray and NIR spectra
of a bright flare (as these allow a precise determination of the spectral index
in the two bands) has not yet been published\footnote{Indeed, the few
multi-wavelength campaigns that caught a very bright X-ray flare were
missing spectroscopic information in the NIR band (e.g., Dodds-Eden et
al.\ 2009), while the few campaigns with spectroscopic information in
both bands, failed to detect a bright X-ray flare with a NIR counterpart
(e.g., flare A and B of Trap et al.\ 2011). }.
Polarisation in the sub-mm and NIR bands suggests that the NIR radiation
is produced by synchrotron emission. The origin of the X-ray emission
is still debated. Indeed, the X-ray radiation could be
produced by synchrotron itself or inverse Compton processes such as
synchrotron self-Compton or external Compton (see Genzel et al. 2010a
for a review).
Different models explain the data with a large range of physical parameters,
however, models with synchrotron emission extending with a break
from NIR to the X-ray seem to be best able to account for the X-ray
data with reasonable physical parameters (Dodds-Eden et al. 2009;
Trap et al. 2010; Dibi et al. 2014; 2016; Barriere et al. 2014).
We report here the first simultaneous observation of the X-ray ({\it XMM-Newton}),
hard X-ray ({\it NuSTAR}), and NIR ({\it SINFONI}) spectra of a very bright flare of
Sgr~A$^\star$, which occurred between 2014 August $30^{th}$ and $31^{st}$
(Ponti et al. 2015a; 2015c), and an analysis of flare models which could explain
the emission. The remainder of this paper is organised as follows:
Section \ref{datared} details the reduction of the X-ray and NIR data.
In section \ref{brightxmm} we present a characterisation of the
obscuration and mean spectral properties of the very bright flares
observed by {\it XMM-Newton}.
In section \ref{multiwave} we investigate the mean properties
of the VB3 flare, in particular we constrain the radiative mechanism
through the study of the mean multi-wavelength spectrum.
In section \ref{SecSpEvol}, we follow the evolution of the flare
emission to determine time-dependent parameters of
the emission models. Section \ref{quie} scrutinises a "quiescent"
interval after the very bright flare.
In Section \ref{SpEvol} we focus the analysis on the X-ray band
only and we study the evolution of the X-ray spectral shape
throughout all bright and very bright X-ray flares.
We discuss the results of the model fits in section \ref{discussion}
and conclude in section \ref{conclusions}.
\section{Data reduction}
\label{datared}
We consider two sets of data in this paper.
The first set comprises simultaneous X-ray ({\it XMM-Newton}\ and {\it NuSTAR}) and
NIR data of one very bright flare, called VB3 (see Ponti et al. 2015 for the
definition of the naming scheme). The analyses of the
{\it XMM-Newton}, {\it NuSTAR}\ and {\it SINFONI}\ data on flare VB3 are discussed
in sections \ref{sub:xmm}, \ref{sub:nustar} and \ref{sub:sinfoni}.
The second set of data consists of all of Sgr~A$^\star$'s bright or very bright
X-ray flares as detected with {\it XMM-Newton}. The reduction of this set of data
is discussed in section \ref{sub:xmm} along with the description of
the flare VB3.
\subsection{Basic assumptions}
\label{Assumptions}
Throughout the paper we assume a distance to Sgr~A$^\star$\ of 8.2~kpc
and a mass of M$ = 4.4\times10^6$~M$_{\odot}$ (Genzel et al. 2010a).
The errors and upper limits quoted on spectral fit results correspond
to 90 \% confidence level for the derived parameters (unless
otherwise specified), while uncertainties associated with measurements
reported in plots are displayed at the $1~\sigma$ confidence level.
The neutral absorption affecting the X-ray spectra is fitted with the model
{\sc TBnew}\footnote{http://pulsar.sternwarte.uni-erlangen.de/wilms/research/tbabs/}
(see Wilms et al. 2000) with the cross sections of Verner et al. (1996) and
abundances of Wilms et al. (2000).
The dust scattering halo is fitted with the model {\sc fgcdust} in
{\sc XSpec} (Jin et al. 2016; see \S \ref{brightxmm}) and
it is assumed to be the same as the "foreground" component along
the line of sight towards AX~J1745.6-2901\ (Jin et al. 2016). More details on the
implications of this assumptions are included in \S \ref{brightxmm}
and Appendix \ref{SecDSH}.
In \S \ref{brightxmm} we justify the assumption of a column density
of neutral absorbing material of $N_H=1.60\times10^{23}$~cm$^{-2}$
and we apply it consistently thereafter. Throughout our discussion
we assume that the effects of beaming are negligible, as well as
a single zone emitting model for the source.
Unless otherwise stated, we follow Dodds-Eden et al. (2009)
and we assume a constant escape time of the synchrotron
emitting electrons equal to $t_{esc}=300$~s. Under this
assumption, the frequency of the synchrotron cooling break
can be used to derive the amplitude of the source magnetic field.
\subsection{{\it XMM-Newton}}
\label{sub:xmm}
In this work, we considered all of the {\it XMM-Newton}\ observations during which
either a bright or very bright flare has been detected through the Bayesian
block analysis performed by Ponti et al. (2015a).
Full details about the observation identification (obsID) are reported
in Tab. \ref{obsid}.
Starting from the {\it XMM-Newton}\ observation data files, we reprocessed all of the data sets
with the latest version (15.0.0) of the {\it XMM-Newton}\ Science Analysis System (SAS),
applying the most recent (as of 2016 April 27, valid for the observing day) calibrations.
Whenever present, we eliminated strong soft proton background flares,
typically occurring at the start or end of an observation, by cutting the exposure time
as done in Ponti et al. (2015a; see Tab. 7).
To compare data taken from different satellites and from the ground, we performed
barycentric correction by applying the {\sc barycen} task of {\sc sas}.
The errors quoted on the analysis of the light curves correspond
to the 1 $\sigma$ confidence level (unless otherwise specified).
{\sc XSpec} v12.8.2 and {\sc matlab} are used for the spectral analysis
and the determination of the uncertainties on the model parameters.
\begin{table}
\begin{center}
\small
\begin{tabular}{ c c c c c c }
\hline
\hline
{\it XMM-Newton}\ & Date &t$_{\rm start}$ &t$_{\rm stop}$ & NAME \\
obsID & &TT$_{\rm TBD}$&TT$_{\rm TBD}$& \\
\hline
\multicolumn{5}{l}{\it Archival data} \\
\hdashline
0111350301 & 2002-10-03 & 150026757 & 150029675 & {\bf VB1} \\
& & 150026757 & 150027730 & VB1-Rise \\
& & 150027730 & 150028702 & VB1-Peak \\
& & 150028702 & 150029675 & VB1-Dec \\
\hdashline
0402430401 & 2007-04-04 & 292050970 & 292054635 & {\bf VB2} \\
& & 292050970 & 292052192 & VB2-Rise \\
& & 292052192 & 292053413 & VB2-Peak \\
& & 292053413 & 292054635 & VB2-Dec \\
& & 292084140 & 292087981 & {\bf B1} \\
\hdashline
0604300701 & 2011-03-30 & 417894177 & 417896560 & {\bf B2} \\
\hline
\multicolumn{5}{l}{\it This campaign} \\
\hdashline
0743630201 & 2014-08-30 & 525829293 & 525832367 & {\bf VB3} \\
& & 525827793 & 525829193 & VB3-Pre \\
& & 525829193 & 525830593 & VB3-Rise \\
& & 525830593 & 525831843 & VB3-Peak \\
& & 525831843 & 525832743 & VB3-Dec \\
& & 525832743 & 525834893 & VB3-Post \\
& 2014-08-31 & 525846661 & 525848532 & {\bf B3} \\
\hdashline
0743630301 & 2014-09-01 & 525919377 & 525924133 & {\bf B4} \\
\hdashline
0743630501 & 2014-09-29 & 528357937 & 528365793 & {\bf B5} \\
\hline
\hline
\end{tabular}
\caption{List of {\it XMM-Newton}\ observations and flares considered in this work.
The flares are divided into two categories, bright (B) and very bright (VB),
classified according to their total fluence (see Ponti et al. 2015a).
The different columns show the {\it XMM-Newton}\ obsID, flare start and end times in
Terrestrial Time (TT$\ddag$; see Appendix A) units and flare name,
respectively. The flare start and end times are barycentric
corrected (for comparison with multi-wavelength data) and correspond to the
flare start time (minus 200 s) and flare stop time (plus 200 s) obtained through
a Bayesian block decomposition (Ponti et al 2015a; please note that
the time stamps in Ponti et al 2015a are not barycentric corrected).
Neither a moderate nor weak flare is detected during these {\it XMM-Newton}\ observations.
To investigate the presence of any spectral variability within each
very bright flare, we extracted three equal duration spectra catching
the flare rise, peak and decay, with the exception of VB3. In the latter case,
we optimised the duration of these time intervals according
to the presence of simultaneous NIR observations (see Fig. \ref{LCp}). }
\label{obsid}
\end{center}
\end{table}
We extracted the source photons from a circular region with
$10~^{\prime\prime}$ radius,
corresponding to $\sim5.1\times10^4$~AU, or $\sim6.5\times10^5$~$R_S$
(Goldwurm et al. 2003; B\'elanger et al.\ 2005; Porquet et al.\ 2008;
Trap et al.\ 2011; Mossoux et al.\ 2015).
For each flare we extracted source photons during the time window
defined by the Bayesian block routine (applied on the EPIC-pn light curve,
such as in Ponti et al. 2015a), adding 200~s before and after the flare (see
Tab. \ref{obsid}).
Background photons have been extracted from the same source regions by
selecting only quiescent periods. The latter are defined as moments
during which no flare of Sgr~A$^\star$\ is detected by the Bayesian block procedure
(Ponti et al. 2015a) and additionally leaving a 2~ks gap before the start and
after the end of each flare.
Given that all of the observations were taken in Full frame mode, pile-up
is expected to be an issue only when the count rate exceeds
$\sim2$ cts/s\footnote{{\it XMM-Newton}\
User Handbook Issue 2.12, Longinotti et al.\ 2014}. This threshold is above
the peak count rate registered even during the brightest flares of Sgr~A$^\star$.
This provides {\it XMM-Newton}\ with the key advantage of being able to collect pile-up
free, and therefore unbiased, spectral information even for the brightest flares.
For each spectrum, the response matrix and effective area have been
computed with the XMM-SAS tasks {\sc rmfgen} and {\sc arfgen}.
See Appendix A for further details on the {\it XMM-Newton}\ data reduction.
\subsection{{\it NuSTAR}}
\label{sub:nustar}
To study the flare characteristics in the broad X-ray band,
we analyzed the two {\it NuSTAR} observations (obsID: 30002002002,
30002002004) taken in fall 2014 in coordination with {\it XMM-Newton}.
We processed the data using the {\it NuSTAR}\ {\it Data Analysis Software
NuSTARDAS} v.1.3.1. and HEASOFT v. 6.13, filtered for periods of high
instrumental background due to SAA passages and known bad
detector pixels.
Photon arrival times were corrected for on-board clock drift and
precessed to the Solar System barycenter using the JPL-DE200 ephemeris.
For each observation, we registered the images with the brightest point
sources available in individual observations, improving the astrometry
to $\sim4^{\prime\prime}$.
We made use of the date obtained by both focal plane modules
FPMA and FPMB.
Four {\it XMM-Newton} flares were captured in the coordinated
{\it NuSTAR} observations: VB3, B3, B4 and B5.
We extract the {\it NuSTAR} flare spectra using the same flaring intervals
as determined from the {\it XMM-Newton} data (see Table 1).
The flare times are barycentric corrected for comparison between
different instruments. Due to interruption caused by earth occultation,
{\it NuSTAR} good time intervals (GTIs) detected only a portion of the flares.
For flare VB3, {\it NuSTAR} captured the first $\sim1215~s$ of the
full flare, corresponding to pre-, rising- and part of the peak-flare,
while the dec- and post-flare intervals were missed.
Similarly, part of the rising-flare stage of flare B3 and the middle half
of flare B4 were captured in the {\it NuSTAR} GTIs.
Flare B5 was not significantly detected with {\it NuSTAR}, resulting
in $\sim2\sigma$ detection in the {\it NuSTAR} energy band.
To derive the flare spectra, we used a source extraction region with
$30^{\prime\prime}$ radius centered on the position of {Sgr~A$^{\star}$}.
While the source spectra were extracted from the flaring intervals,
the background spectra were extracted from the same region in
the off-flare intervals within the same observation.
The spectra obtained by FPMA and FPMB are combined and then
grouped with a minimum of $3\sigma$ signal-to-noise significance
per data bin, except the last bin at the high-energy end for which
we require a minimum of $2\sigma$ significance.
\begin{table*}
\begin{tabular}{lcccc}
\hline
\multicolumn{5}{c}{Coordinated {\it NuSTAR} and {\it XMM-Newton} observations and flares detected} \\
{\it NuSTAR} & $t_{start}$ & Exposure & Joint {\it XMM-Newton} & NAME \\
obsID & & (ks) & obsID & \\
\hline
30002002002 & 2014-08-30 19:45:07 &59.79~ks & 0743630201 & VB3, B3 \\
& & & 0743630301 & B4 \\
30002002004 & 2014-09-27 17:31:07 &67.24~ks & 0743630501 & B5 \\
\hline
\end{tabular}
\caption{The different columns show the {\it NuSTAR} obsID, observation
start time, total exposure, coordinated {\it XMM-Newton} obsID and
the flares detected in the observation.}
\end{table*}
\subsection{{\it SINFONI}}
\label{sub:sinfoni}
\subsubsection{Observations and data reduction}
We observed Sgr~A$^\star$\ with {\it SINFONI}\ (Eisenhauer et al. 2003; Bonnet et al. 2004)
at VLT between 30-08-2014 23:19:38 UTC and 31-08-2014 01:31:14 UTC.
{\it SINFONI}\ is an adaptive optics (AO) assisted integral-field spectrometer mounted
at the Cassegrain focus of Unit Telescope 4 (Yepun) of the ESO Very Large
Telescope.
The field of view used for this observation was $0.8^{\prime\prime} \times
0.8^{\prime\prime}$, which is divided
into $64\times32$ spatial pixels by the reconstruction of the pseudoslit into a 3D
image cube. We observed in H$+$K bands with a spectral resolution of $\sim1800$.
We accumulated seven spectra (see observation log in Tab. \ref{IRtime}) of 600~s
each using an object-sky-object observing pattern.
There are gaps between observations for the 600~s sky exposures, as well as
a longer gap due to a brief telescope failure during what would have been
an additional object frame at the peak of the X-ray flare.
In total, we accumulated four sky frames on the sky field (712$^{\prime\prime}$
west, $406^{\prime\prime}$ north of Sgr~A$^\star$).
During our observations, the seeing was $\sim$ $0.7^{\prime\prime}$ and
the optical coherence
time was $\sim$ 2.5 ms. The AO loop was closed on the closest optical guide star
($m_R$ = 14.65; $10.8^{\prime\prime}$ east, $18.8^{\prime\prime}$ north
of Sgr~A$^\star$), yielding a spatial resolution
of $\sim 90 $ mas FWHM at 2.2 $\mu $m, which is $\sim$1.5 times
the diffraction limit of UT4 in K band.
The reduction of the {\it SINFONI}\ data followed the standard steps. The object
frames were sky subtracted using the nearest-in-time sky frame to correct
for instrumental and atmospheric backgrounds. We applied bad pixel correction,
flat-fielding, and distortion correction to remove the intrinsic distortion
in the spectrograph. We performed an initial wavelength calibration with calibration
source arc lamps, and then fine-tuned the wavelength calibration using
the atmospheric OH lines of the raw frames. Finally, we assembled the data
into cubes with a spatial grid of 12.5 mas per pixel.
\subsubsection{Sgr~A$^\star$\ spectrum extraction}
The source spectrum extraction uses a procedure to extract a noisy spectrum
from Sgr~A$^\star$. This can then be binned and the scatter within a bin used
as an estimate of the error on the flux in that bin.
In each of the seven data cubes, we use a rectangular region of the spatial
dimensions of size ($0.31^{\prime\prime}$ $\times$ $0.36^{\prime\prime}$)
centered roughly between Sgr~A$^\star$\ and the bright star S2 (0.05" south of S2).
Within this region are four known S-stars, S2, S17, S19, and S31.
Figure \ref{ImaNIR} shows a combined data cube assembled from
the seven observations,
as well as a simulated image of the Galactic centre S-stars.
The four known stars used in the fitting procedure are labelled in both images.
We extract $\sim$100 noisy images from the data cube by collapsing
the cube along the spectral direction (median in the spectral direction)
in bins with 3.5 nm width (seven spectral channels per bin) in the spectral
range 2.03-2.39 $\mu$m.
This initial binning is necessary, as the signal to noise of a single spectral
channel is not high enough on its own to perform the next step.
In each noisy image, we determine the flux of Sgr~A$^\star$\ from a fit with 6 Gaussians
to the image. Five Gaussians with a common (variable) width describe
the five sources in each image.
The sixth Gaussian has a width of 3.5 times wider than the sources, and
describes the AO seeing halo of the brightest star, S2, which has a K magnitude
of $\sim$14.
The seeing halos of the dimmer stars (K magnitude $<$15) are neglected in the fit.
The positions of the four stars relative to one another and to Sgr~A$^\star$\ are fixed
based on the known positions of the stars.
The flux ratios of the four stars are fixed based on previous photometric
measurements of the stars.
Note that fixing the flux ratios assumes that the spectral indices of the
various S-stars are not significantly different, an excellent assumption given
the strong extinction toward the Galactic center (GC).
The final fit has five free parameters: The overall amplitude of the S-stars,
the background, the Gaussian width of the sources, the flux ratio of the seeing
halo/S2, and the flux ratio of Sgr~A$^\star$/S2.
This fitting procedure allows a measurement of the variability of Sgr~A$^\star$\ in
the presence of variations (in time and wavelength) in the background, Strehl
ratio, and seeing.
The result of this procedure is a flux ratio of Sgr~A$^\star$/S2 in each of the $\sim$ 100 spectral bins.
We obtain a noisy, color-corrected spectrum of Sgr~A$^\star$\ by multiplying a calculated
spectrum of S2 by the flux ratio Sgr~A$^\star$/S2 obtained from the fit in each
extracted image.
The calculated S2 spectrum used is $\nu S_{\nu}$ for a blackbody with a
temperature of 25,000K, and a stellar radius of 9.3 $R_\odot$, the best fit
temperature and radius for S2 found in Martins et al. (2009).
The source is placed at 8.2 kpc (Genzel et al. 2010a) from the Earth.
This spectrum is normalized to a value of 20 mJy at 2.2 $\mu$m wavelength
to match previous photometric measurements of S2.
This procedure corrects for the effects of interstellar extinction.
Note that by normalizing the spectrum of S2 to a value of 20~mJy at 2.2~$\mu$m,
we do not take into account the error on the previous measurements
of the flux of S2. Since errors on this value result only in an overall error
on the amplitude and not in the spectral shape, this additional uncertainty
in the normalization of the spectra is taken into account in the later model fits
by allowing the overall amplitude of the NIR spectrum to vary and determining
the effect of this variation on the fit parameters.
To obtain the final NIR data points used for the model fitting in this paper,
the noisy spectrum is binned into 10 spectral bins (median of the values
in each bin) of width 35 nm. The error on each point is the standard deviation
of the sample, or $\sigma/\sqrt{N}$.
We have tested varying the number of initial spectral samples used to create
the extracted images used for fitting Sgr~A$^\star$/S2, and find that it has almost
no effect on the final data values and only a small effect on the derived error bars.
We have also tried fixing parameters in the fits to determine their effects
on the final spectra.
We tried fixing the FWHM of the Gaussians and the background level (which
both naturally vary with wavelength) to their median values, and found that
this affects the spectral index of the final data points by at most a few percents.
\begin{figure*}
\includegraphics[height=0.45\textwidth,angle=0]{flare_image2.ps}
\caption{The GC as seen with {\it SINFONI}\ (each image is 0.51"
$\times$ 0.61").
The image on the left is the collapsed image from the spectral range
2.25-2.35 $\mu$m for the seven data cubes combined.
The image on the right is a diffraction-limited simulated image of
the locations of the S-stars in the GC. In both images
the location of Sgr~A$^\star$\ is indicated with a cross, and the four stars
included in our Gaussian fits for spectrum extraction are labeled
with circles. The flare is clearly visible above the background
in the left image.}
\label{ImaNIR}
\end{figure*}
\begin{table*}
\begin{center}
\small
\begin{tabular}{ c c c c c c }
\hline
Spec & t$_{\rm start}$ & t$_{\rm stop}$ &$\Gamma$ & $F_{2.2\mu m}\dag$ \\
& BJD$_{\rm TBD}$ & BJD$_{\rm TDB}$& & (mJy) \\
\hline
IR1 & 2456900.47470 & 2456900.48164 & $1.48\pm0.23$ & $8.87\pm0.10$ \\
IR2 & 2456900.48971 & 2456900.49665 & $1.37\pm0.19$ & $7.91\pm0.07$ \\
IR3 & 2456900.49694 & 2456900.50388 & $1.80\pm0.13$ & $10.09\pm0.07$ \\
IR4 & 2456900.52160 & 2456900.52855 & $1.76\pm0.21$ & $7.52\pm0.06$ \\
IR5 & 2456900.53675 & 2456900.54370 & $3.71\pm0.48$ & $3.12\pm0.07$ \\
IR6 & 2456900.54398 & 2456900.55093 & $2.72\pm0.23$ & $4.23\pm0.04$ \\
IR7 & 2456900.55914 & 2456900.56608 & $2.56\pm0.21$ & $7.07\pm0.06$ \\
\hline
\end{tabular}
\caption{The second and third column show the Barycentre corrected start
and end times of each of the seven {\it SINFONI}\ spectra. Times are Barycentric
Julian Dates in Barycentric Dynamical Time. The last two columns show the
test fit photon index ($\Gamma$) and flux once the {\it SINFONI}\ data are fitted
with a simple power-law model. \dag The error bars are statistical only and
they do not include systematic effects due to spectral extraction. The effect
of systematics is treated in \S 5.3.4.}
\label{IRtime}
\end{center}
\end{table*}
\section{X-ray obscuration and mean X-ray properties of bright flares}
\label{brightxmm}
We started the study of Sgr~A$^\star$'s emission by investigating the properties
of the absorption-scattering layers that distort its spectrum.
\subsection{Dust scattering}
\label{IntroDSH}
Scattering on dust grains along the line of sight can have a significant impact
on the observed X-ray spectra (Predehl \& Schmitt 1995; Smith et al. 2016).
The main effect of dust scattering is to create a halo around the source,
by removing flux from the line of sight. Both the flux in the halo and its size
decrease with energy (with a dependence of $\propto E^{-2}$ and
$\propto E^{-1}$, respectively) as a consequence of the probability
of scattering that drops steeply with energy. If the events used to
extract the source photons are selected from a small region containing only
a small part of the halo, such as typically the case for X-ray observations
of Sgr~A$^\star$, then the spectral shape will be distorted by the effects of
dust scattering.
Whenever the distortions are not accurately accounted for, this will
cause significant biases in the measured absorption column densities,
source brightness and spectral slopes (see appendix \ref{SecDSH}).
Frequently used models, aimed at mitigating the effects of dust
scattering, are: {\sc dust}
(Predehl \& Schmitt 1995); {\sc scatter} (Porquet et al. 2003; 2008;
Dodds-Eden et al. 2009) and; {\sc dustscat} (Baganoff et al. 2003;
Nowak et al. 2012). In all these models the dust optical depth and therefore
the magnitude of the correction, is assumed to be proportional to the
X-ray absorbing column density, with a factor derived from Predehl
\& Schmitt (1995). The underlying assumption is that the dust properties
(e.g., dust to gas ratio, size distribution, composition, etc.) towards
Sgr~A$^\star$\ are equal to the average estimate derived from the study
of all the Galactic sources considered in the work of Predehl \& Schmitt (1995).
After considering the limitations of this approach, we decided to use
a completely different method.
Thanks to the analysis of all the {\it XMM-Newton}\ and {\it Chandra}\ observations
of the GC, Jin et al. (2016) just completed the accurate characterisation
of the dust scattering halo towards AX~J1745.6-2901, a bright X-ray binary located
only $\sim1.45^{\prime}$ from Sgr~A$^\star$. The authors deduced that
$74\pm7$~\% of the dust towards AX~J1745.6-2901\ resides in front of the GC
(e.g. in the spiral arms of the Galaxy).
Moreover, the detailed modelling of the dust scattering halo allowed
Jin et al. (2016) to provide an improved model of the spectral distortions
generated by the dust scattering ({\sc fgcdust}), without the
requirement to assume fudge scaling factors. We therefore decided
to fit Sgr~A$^\star$'s spectrum with the {\sc fgcdust} model, implicitly
assuming that the dust has similar properties along the line of sight
towards Sgr~A$^\star$\ and the foreground component in the direction of AX~J1745.6-2901.
This is corroborated by the study of the radial and azimuthal
dependence of the halo. In fact, the smoothness of the profile indicates
that the foreground absorption has no major column density variations
within $\sim100-150^{\prime\prime}$ from AX~J1745.6-2901\ (Jin et al. 2016).
Further details on the spectral distortions (and their correction) introduced
by dust scattering are discussed in Appendix \ref{SecDSH}.
\subsection{Foreground absorption towards the bright sources within
the central arcmin}
\label{ISMsources}
We review here the measurements of the X-ray column density
of neutral/low-ionised material along the line of sights towards
compact sources located close to Sgr~A$^\star$.
Due to the variety of assumptions performed in the different works
(e.g., absorption models, abundances, cross sections, dust scattering
modelling, etc.), we decided to refit the spectra to make all measurements
comparable with the abundances, cross sections and absorption models
assumed in this work (see \S 2.1; Fig. 5 and Tab. 2 of Ponti et al. 2015
and Ponti et al. 2016). \\
{\bf AX~J1745.6-2901\ located at $\sim1.45^{\prime}$ from Sgr~A$^\star$}\\
AX~J1745.6-2901\ is a dipping and eclipsing neutron star low mass X-ray binary (Ponti
et al. 2016). Such as typically observed in high inclination low mass
X-ray binaries, AX~J1745.6-2901\ shows both variable ionised and neutral local
absorption (Ponti et al. 2015).
The total neutral absorption column density towards AX~J1745.6-2901\ has been
measured by Ponti et al. (2015b). We re-fitted those spectra of AX~J1745.6-2901\ with
the improved correction for the dust scattering distortions. By considering
that only $74\pm7$~\% of the dust towards AX~J1745.6-2901\ resides in front of
the GC (Jin et al. 2016), we measured a total column density in the
foreground component\footnote{The large
uncertainty in this measurement is driven by the uncertainty in
the determination of the fraction of column density in the foreground
component. } of $N_H=(1.7\pm0.2)\times10^{23}$~cm$^{-2}$.
We note that the halo associated to the foreground component
is still detected at radii larger than $r>100^{\prime\prime}$
(Jin et al.\ 2016), therefore from a radius more than ten times
larger than the one chosen to extract Sgr~A$^\star$'s photons,
indicating that a careful treatment of the distortions introduced
by dust scattering is essential.
\\
{\bf SWIFT~J174540.7-290015 located at $\sim16^{\prime\prime}$
from Sgr~A$^\star$}\\
A deep {\it XMM-Newton}\ observation performed during the recent outburst of Swift~J174540.7-290015\
(Reynolds et al. 2016), allowed Ponti et al. (2016) to measure the
column density along this line of sight
and to find $N_H=(1.70\pm0.03)\times10^{23}$ cm$^{-2}$, by fitting the spectrum
with the sum of a black body plus a Comptonisation component\footnote{The
authors find marginal evidence for sub-Solar iron abundance,
suggesting that iron is depleted into dust grains. The detailed investigation
of the metal abundances is beyond the scope of this paper. }.
By applying the improved modelling of the dust scattering halo to the same
data, we measured a column density of $N_H=(1.60\pm0.03)\times10^{23}$
cm$^{-2}$.\\
{\bf SGR~J1745-2900 located at $\sim2.4^{\prime\prime}$ from Sgr~A$^\star$}\\
SGR J1745-2900\ is a magnetar located at a small projected distance from Sgr~A$^\star$\
(Mori et al. 2013; Kennea et al. 2013),
and it is most likely in orbit around the supermassive BH (Rea et al. 2013).
Coti-Zelati et al. (2015) fitted the full {\it XMM-Newton}\ and {\it Chandra}\ dataset available
on SGR J1745-2900, without considering the effects of the dust scattering halo, and
found $N_H=(1.90\pm0.02)\times10^{23}$ cm$^{-2}$ for {\it Chandra}\ and
$N_H=(1.86^{+0.05}_{-0.03})\times10^{23}$ cm$^{-2}$ for {\it XMM-Newton}.
We refitted the {\it XMM-Newton}\ dataset at the peak of emission (obsid: 0724210201),
using as background the same location when the magnetar was
in quiescence and considering the improved dust model, and
we obtained $N_H=(1.69^{+0.17}_{-0.10})\times10^{23}$ cm$^{-2}$.
Radio observations of the pulsed emission from SGR J1745-2900\ allowed Bower et al.
(2014) to provide a full characterisation of the scattering properties of
the absorption. The authors found the obscuring-scattering layer
to be located in the spiral arms of the Milky Way, most likely at
a distance $\Delta=5.8\pm0.3$~kpc from the GC (however a uniform scattering
medium was also possible). Moreover, the source sizes at different
frequencies are indistinguishable
from those of Sgr~A$^\star$, demonstrating that SGR J1745-2900\ is located behind
the same scattering medium of Sgr~A$^\star$.
We note that the column densities of absorbing material along
the line of sights towards SWIFT~J174540.7-290015, SGR J1745-2900\ and
the foreground component towards AX~J1745.6-2901\ are consistent (to an
uncertainty of $\sim2-10$~\%) within each other. Of course,
the neutral absorption towards these accreting sources might even,
in theory, be local and variable (e.g. Diaz-Trigo et al. 2006;
Ponti et al. 2012; 2016b), however the similar values observed
in nearby sources indicate a dominant ISM origin. We note
that the location of the scattering medium towards Sgr~A$^\star$\ and
the foreground component of AX~J1745.6-2901\ are also cospatial.
This suggests that all these sources are absorbed by a common,
rather uniform, absorbing layer located in the spiral arms of
the Milky Way (Bower et al. 2014; Jin et al. 2016). This result is also
in line with the small spread, of the order of $\sim10$~\%, in the extinction
observed in NIR towards the central $\sim20^{\prime\prime}$ of the
Galaxy (Sch\"odel et al. 2010; Fritz et al. 2011).
Indeed, for a constant dust to gas ratio, this would induce a spread in
the X-ray determined $N_H$ of a similarly small amplitude.
In addition to this layer, AX~J1745.6-2901\ also shows another absorbing
component, located closer to the source, possibly associated
either with the clouds of the central molecular clouds or a
local absorption (Jin et al. 2016).
Studies of the scattering sizes from large scale ($\sim2^\circ$)
low frequency radio maps also agree with the idea that
the intervening scattering in the GC direction is composed
of two main absorption components, one uniform on a large
scale and one patchy at an angular scale of $\sim10^{\prime}$
and with a distribution following the clouds of the central
molecular zone (Roy 2013).
\subsection{The mean spectra of the {\it XMM-Newton}\ very bright flares}
We extracted an EPIC-pn and -MOS spectra for each of the bright and
very bright flares detected by {\it XMM-Newton}. We used a Bayesian block decomposition
of the EPIC-pn light curve to define the start and end flare times (see Tab.
\ref{obsid}).
We fitted each spectrum with a power-law model modified by neutral
absorption (see \S \ref{Assumptions}) and by the contribution from the
dust scattering halo ({\sc fgcdust * TBnew * power-law} in {\sc XSpec}).
Each spectrum is well fitted by this simple model (see Tab. \ref{bestfitFlares}).
In particular, the column density of absorption material and the photon
index are consistent with being the same between the different flares
and consistent with the values observed in nearby sources (see \S
\ref{ISMsources}). This agrees with the idea that most of the neutral absorption
column density observed towards Sgr~A$^\star$\ is due to the interstellar medium
(ISM). If so, the absorption should not vary significantly over time (see
Tab. \ref{bestfitFlares}).
Therefore, we repeat the fit of the spectra assuming that the three very
bright flares are absorbed by the same column density of neutral material.
The three spectra are well described by this simple model
(see Tab. \ref{bestfitFlares}), significantly reducing the uncertainties.
The best fit spectral index is $\Gamma_{VB123}=2.20\pm0.15$, while the
column density is: $N_H=(1.59\pm0.15)\times10^{23}$~cm$^{-2}$.
This value is fully consistent with the one observed towards the foreground
component towards bright nearby X-ray sources, reinforcing the suggestion
that the column density is mainly due to the ISM absorption.
For this reason hereafter we will fix it to the most precisely constrained
value $N_H=1.60\pm0.03\times10^{23}$~cm$^{-2}$ (Ponti et al. 2016b).
The resulting best fit photon index with this value of $N_H$ is
$\Gamma_{VB123_{FixN_{H}}}=2.21\pm0.09$.
\begin{table}
\begin{center}
\scriptsize
\begin{tabular}{ c c c c c c }
\hline
\multicolumn{5}{c}{\bf Absorbed power-law fit to X-ray spectra} \\
\hline
Name & $N_H$ & $\Gamma_X$ & Flux$_{2-10}$ & $\chi^2$/dof \\
\hline
VB1 & $1.6\pm0.2$ & $2.2\pm0.3$ & $9.6^{+7}_{-4}$ & 89.9/114 \\
VB2 & $1.6\pm0.3$ & $2.3\pm0.4$ & $5.0^{+5}_{-2.4}$ & 89.3/98 \\
VB3 & $1.6\pm0.3$ & $2.3\pm0.3$ & $7.6^{+7.1}_{-3.4}$ &127.2/117 \\
VB123\dag & $1.59\pm0.15$ & $2.20\pm0.15$ & & 302.8/331 \\
VB123$_{FixN_{H}}$\ddag & $1.6$ & $2.21\pm0.09$ && 302.8/332 \\
VB3$_{XMM+Nu}$\ddag & $1.6$ & $2.27\pm0.12$ &$7.5\pm1.5$&141.4/133 \\
\hline
\end{tabular}
\caption{Best fit parameters of the fit of the very bright flares of Sgr~A$^\star$.
Column densities are given in units of $10^{23}$~cm$^{-2}$ and
the absorbed fluxes are in units of $10^{-12}$ erg cm$^{-2}$ s$^{-1}$.
\dag The VB123 flare indicates the average of VB1+VB2+VB3.
\ddag The VB123$_{FixN_{H}}$ shows the best fit results of flare
VB123, once the column density of neutral absorbing material
has been fixed. The VB3$_{XMM+Nu}$ shows the best fit results of
flare VB3 (by fitting both {\it XMM-Newton}\ and {\it NuSTAR}\ data), once the column
density has been fixed. }
\label{bestfitFlares}
\end{center}
\end{table}
\section{Mean properties of VB3}
\label{multiwave}
We investigate here the mean properties of a very bright flare (VB3, see Ponti
et al. 2015a) during which, for the first time, simultaneous time-resolved
spectroscopy in NIR and X-rays has been measured.
\subsection{X-ray ({\it XMM-Newton}\ and {\it NuSTAR}) mean spectra of VB3}
We first simultaneously fitted the {\it XMM-Newton}\ (pn and both MOS) and {\it NuSTAR}\
mean spectra of VB3 (see Fig. \ref{MSpec}, Tab. \ref{obsid}) with
an absorbed power law model.
The {\it NuSTAR}\ data cover only part of the flare, missing the decaying flank
of the flare, therefore probing different stages of a variable phenomenon.
We accounted for this by allowing the fit to have different power-law
normalisation between the {\it NuSTAR}\ and {\it XMM-Newton}\ spectra\footnote{The
normalisations are, however, consistent between the two instruments.}.
The best fit photon index is: $\Gamma_{XMM+Nu}=(2.27\pm0.12)$
and absorbed 2-10 keV flux $F_{2-10}=7.5\pm1.5\times10^{-12}$ erg
cm$^{-2}$ s$^{-1}$. The spectra were very well fit by this simple model
with $\chi^2=141.4$ for 133 dof.
We investigated for the presence of a possible high-energy cut-off
(or high-energy spectral break) by fitting the spectra with an absorbed
broken power-law model.
We fixed the photon index of the lower energy power-law slope
to $\Gamma_{VB123_{FixN_{H}}}=2.20$ (the best fit value of the
simultaneous fit of VB1+VB2+VB3, see Tab. \ref{bestfitFlares}).
No significant improvement was observed ($\chi^2=141.2$ for 132 dof).
\begin{figure}
\hspace{-0.5cm}
\includegraphics[height=0.47\textwidth,angle=-90]{MeanSpecVB3.ps}
\caption{Mean X-ray spectrum of VB3. The black squares, red circles and green
stars show the EPIC-pn, combined EPIC-MOS and combined {\it NuSTAR}\
spectra, respectively. The combined {\it XMM-Newton}\ and {\it NuSTAR}\ spectra greatly
improve the determination of the X-ray slope. The data are fitted with
an absorbed power law model, which takes into account the distortions
induced by the dust scattering (see text for more details). }
\label{MSpec}
\end{figure}
\subsection{Multi-wavelength mean spectra of VB3}
\label{evolution}
We then extended our investigation by adding the NIR spectra.
Multiple {\it SINFONI}\ spectra (e.g., IR2, IR3 and IR4) have been accumulated
during the duration of the X-ray emission of the VB3
flare\footnote{Defined on the basis of the X-ray light curve,
therefore it represents the full duration of the X-ray flare. } (see Fig. \ref{LCp}).
Therefore, we created the mean spectrum from these NIR spectra
and fitted these simultaneously with the mean X-ray spectra of VB3
(see Tab. \ref{IRtime} and Fig. \ref{LCp}).
\subsubsection{Single power-law (plain Synchrotron)}
\label{PureSync}
We started fitting the mean spectrum from NIR to hard X-ray with a simple
power-law model, as expected in the case of plain Synchrotron emission
(Dodds-Eden et al. 2009).
The best fit photon index was $\Gamma=2.001\pm0.005$ (see Tab.
\ref{meanSp} and Fig. \ref{PLCool}).
However, this very simple model provided us with an unsatisfactory
result ($\chi^2=189.7$ for 142 dof; Tab. \ref{meanSp}). This is mainly
driven by the different slopes observed in the NIR and X-ray bands
(see Fig. \ref{SED} and \ref{PLCool}).
We therefore concluded that a plain Synchrotron model is ruled out.
\subsubsection{Broken power-law model (BPL, phenomenological model)}
\label{BPL}
We then performed a phenomenological description of the data
with a broken power-law (BPL) model. We observed a significant improvement
and an acceptable description of the spectrum by fitting the data with
this model, where the NIR and X-ray slopes were
free to vary ($\chi^2=154.9$ for 142 dof, $\Delta\chi^2=34.8$ for the
addition of 2 dof, corresponding to an F-test probability of
$\sim7\times10^{-7}$; Tab. \ref{meanSp}, Fig. \ref{PLCool}).
The resulting best fit NIR and X-ray photon indexes are
$\Gamma_{NIR}=1.7\pm0.1$ and $\Gamma_X=2.27\pm0.12$,
respectively (Tab. \ref{meanSp}).
The spectral steepening $\Delta\Gamma=0.57\pm0.15$
($\pm0.09$ at 1-$\sigma$)
is slightly steeper, but fully consistent with the value expected in the
Synchrotron scenario in the presence of a cooling break
($\Delta\Gamma=0.5$), strongly suggesting this latter scenario as
the correct radiative mechanism.
\begin{table}
\begin{center}
\footnotesize
\begin{tabular}{ c c c c c c c }
\hline
\multicolumn{6}{c}{\bf VB3 mean spectrum} \\
\hline
& Single PL & BPL & TSSC & PLCool \\
\hline
$\Gamma_{NIR}$& $2.001\pm0.005$ & $1.7\pm0.1$ & & $1.74\pm0.08$ \\
$\Gamma_X$ & & $2.27\pm0.12$ & & \\
$\Delta\Gamma$& & $0.57\pm0.15$ & & 0.5 \\
Log(B) & & $0.94\pm0.16$ & $4.0\pm0.4$ & $0.94\pm0.16$ \\
$\Theta_e$ & & & $9\pm4$ & \\
Log($N_e$) & & & $39.5\pm0.5$ & \\
Log(R$_F$) & & & $-3.5\pm0.5$ & \\
\hline
$\chi^2$/dof & 189.7/142 & 154.9/140 & 162.7/139 & 156.8/141 \\
\hline
\end{tabular}
\caption{Best fit parameters of the mean spectrum of VB3
with the Single PL (plain synchrotron), BPL (broken power-law),
TSSC (thermal synchrotron self Compton) and PLCool (power-law
cool) models.
See \S \ref{multiwave} for a description of the parameters. }
\label{meanSp}
\end{center}
\end{table}
\subsubsection{Thermal Synchrotron Self Compton (TSSC)}
\label{TSSC}
Before fitting the VB3 mean spectrum with a Synchrotron model with
a cooling break, we considered an alternative interpretation, where
the NIR band is produced via synchrotron radiation by a thermal
distribution of electrons. Moreover, the same population of electrons
generates via inverse Compton the high-energy (e.g. X-ray) emission
(see e.g. Dodds-Eden et al. 2011). We called this model thermal
Synchrotron self Compton (TSSC).
The free parameters in this model are: $B$, the strength of the magnetic
field; $\theta_E$, the dimensionless electron temperature (defined
as $\theta_E=\frac{kT_e}{m_e c^2}$, where $k$ is the Boltzman constant,
$T_e$ is the temperature of the thermal electrons, $m_e$ is the
electron mass and $c$ is the speed of light); $N$, the total
number of NIR synchrotron emitting electrons; and $R_F$, the size of
the region containing the flaring electrons, controlling the photon density
of the seed photons. The very short variability time-scale (of the order
of $10^2$~s) suggests a very compact source with a size of the order
of (or smaller than) a few Schwarzschild radii, likely located within
or in the proximity of the hot accretion flow of Sgr~A$^\star$.
Radio and sub-mm observations constrain the physical
parameters of the steady emission from the inner hot accretion flow
(within the central $\sim10$~R$_S$) to be $B\sim10-50$~G,
$T_e\sim10^{10}$~K, $\gamma_e\sim10$; $n_e\sim10^6$~cm$^{-3}$
(see \S~\ref{intro}; Loeb \& Waxman 2007; Genzel et al. 2010).
These are likely the pre-flare plasma conditions.
The TSSC model provides an acceptable fit to the data
($\chi^2=162.7$ for 139 dof; see Fig. \ref{PLCool})\footnote{The fits
have been performed in {\sc matlab},
implementing the equations reported in Dodds-Eden et al. (2009)
and references therein. The best fit was computed through a $\chi^2$
minimisation technique. The parameter space to determine the
uncertainties on the best fit parameters has been explored through
a Bayesian Markov Chain Monte Carlo approach. }.
However, as observed in previous very bright flares (Dodds-Eden et
al. 2009), the best fit parameters of this model are very different
from the reasonable range expected to be present in the accretion
flow of Sgr~A$^\star$. Indeed, this model produces the flare via a magnetic
field with a staggering intensity of $Log(B)=4.0\pm0.4$~G, about
three orders of magnitude larger than the magnetic field intensity
within the steady hot accretion flow of Sgr~A$^\star$, on a population
of ``not-so-energetic" ($\theta_E=9\pm4$) electrons.
Moreover, in order to make the inverse Compton process
efficient enough to be competitive to synchrotron, the electron
density has to be as high as $n_e=10^{13}$~cm$^{-3}$, about
seven orders of magnitude higher than in the accretion flow.
This appears unlikely.
The total number of TSSC emitting electrons is constrained
by the model to be $Log(N_e)=39.5\pm0.5$, therefore to reach
such an excessively high electron density, the size of
the emitting region has to be uncomfortably small,
$Log(R_F/R_S)=-3.5\pm0.5$. Such a source would be
characterised by a light crossing time of the order of only
$\sim10$~ms. Indeed, variability on such time-scales are typically
observed in accreting X-ray binaries (e.g., Belloni et al. 2002;
De Marco et al. 2015), where the system is $\sim10^6$ times
more compact than in Sgr~A$^\star$\ (Czerny et al. 2001; Gierlinski et
al. 2008; Ponti et al. 2012b),
while Sgr~A$^\star$'s power spectral density appears dominated by
variability at much larger time-scales (Do et al. 2009; Meyer et al.
2009; Witzel et al. 2012; Hora et al. 2014)\footnote{
An excessively compact source with such high densities
and magnetic field is hardly achievable event through compression
of a fraction of the quiescent electrons. The quiescent density of
electrons ($n_e=10^6$~cm$^{-3}$) dictates that $N_e=10^{39.5}$
electrons are contained within a sphere of $\sim0.07$~R$_S$,
that therefore would need to be compressed by 2.3 orders of
magnitude to reach the required TSSC source size and density.
We note that, assuming conservation of the magnetic flux,
the magnetic field strength would rise to $Log(B)\sim6.2$,
two orders of magnitude higher than the best fit value.
The magnetic energy would also rise, but would still be 2-4 orders
of magnitude smaller than the one required to power the flare. }.
Again, this appears as a weakness of this model.
As already discussed in Dodds-Eden et al. (2009; 2010; 2011) and
Dibi et al. (2014; 2016), these physical values are different by
several orders of magnitude from the ones observed in quiescence
and, therefore they appear unlikely. With this study we show that the same
"unlikely" physical parameters are not only observed during VB2
(the flare analysed by Dodds-Eden et al. 2009), but also during the very
bright flare considered here (VB3). This confirms that this model
produces unreasonable parameters for part (if not all) of the bright flares.
\subsubsection{Synchrotron emission with cooling break (PLCool)}
\label{SecPLCool}
In this scenario, the synchrotron
emission is produced by a non-thermal distribution of relativistic electrons,
embedded in a magnetic field with strength $B$, therefore they radiate
synchrotron emission. At the acceleration site, the injected electrons are
assumed to have a power-law distribution in $\gamma_e$ with index
$p$ (i.e. $N(\gamma_e)\propto\gamma_e^{-p}$),
defined between $\gamma_{min}$ and $\gamma_{max}$ ($\gamma_e$ being
the electron Lorentz factor). We assume as lower boundary
$\gamma_{min}=10$, supposing that the radiating electrons are accelerated
from the thermal pool producing the sub-mm peak in quiescence (Narayan et al.
1998; Yuan et al. 2003). We also assume that at any point during the flare,
the engine is capable of accelerating electrons to $\gamma_{max}>10^6$,
so that they can produce X-ray emission via synchrotron radiation
(this appears as a less reliable assumption and indeed an alternative to
this scenario will be discussed in \S~\ref{SecDiscPLCoolEv}).
A well known property of high-energy electrons radiating via synchrotron
emission is that they cool rapidly, quickly radiating their energy on
a time-scale $t_{cool}=220 (B/50~G)^{-3/2} (\nu/10^{14}~Hz)^{-1/2}$~s
(where $\nu$ is the frequency of the synchrotron emitted radiation;
see Pacholczyk 1970). In particular, higher energy electrons
cool faster than the NIR ones. The competition between synchrotron
cooling and particle escape from the acceleration zone then generates
a break in the synchrotron spectrum at a frequency:
$\nu_{br}=2.56(B/30~G)^{-3}(t_{esc}/300~s)^{-2}\times10^{14}$~Hz.
Furthermore, in case of continuous acceleration, a steady solution exists
where the slope of the power-law above the break is steeper by
$\Delta\Gamma=0.5$ (Kardashev et al. 1962) than the lower energy
power-law\footnote{Synchrotron radiation from the cooling power-law
distribution of electrons (with electron power-law index $p$) generates
a spectrum $\nu F_\nu \propto \nu^{(3-p)/2}$ at frequencies lower
than the cooling break and $\nu F_\nu \propto \nu^{(2-p)/2}$ above.
This implies that $p$ relates to the photon index $\Gamma$ such as:
$\Gamma=\frac{p+1}{2}$ below and $\Gamma=\frac{p+2}{2}$ at
frequencies above the cooling break. }.
Following the nomenclature of Dodds-Eden et al. (2009),
we call this model "PLCool". The free parameters of the PLCool model
are: $B$; $p$; and the normalisation.
As described in \S \ref{BPL} a broken power-law model provides
an excellent fit to the mean VB3 multi-wavelength spectrum
($\chi^2=154.9$ for 140 dof).
In particular, we note that the difference between the NIR and X-ray
photon indices $\Delta\Gamma=0.57\pm0.15$ ($\pm0.09$ at 1-$\sigma$)
is consistent with the value expected by the PLCool model ($\Delta\Gamma=0.5$;
due to synchrotron emission with continuous acceleration).
Indeed, imposing such spectral break ($\Gamma_{X}=\Gamma_{NIR}+0.5$),
the fit does not change significantly ($\chi^2=156.8$ for 141 dof), with the
photon index $\Gamma_{NIR}=1.74\pm0.08$ and the break at
$0.04^{+0.12}_{-0.03}$~keV ($B=8.8^{+5.0}_{-3.0}$~G; Fig. \ref{PLCool}).
We note that the PLCool model provides a significantly better fit
($\chi^2=156.8$) than the TSSC model ($\chi^2=161.3$) despite having
two fewer free parameters (Tab. \ref{meanSp}).
\begin{figure}
\includegraphics[height=0.49\textwidth,angle=90]{X-rayNIR_x2.ps}
\caption{The red and black points show the mean NIR ({\it SINFONI})
and X-ray ({\it XMM-Newton}\ and {\it NuSTAR}) emission during the VB3 flare.
The dotted red and black straight lines show the uncertainties
on the determination of the NIR and X-ray power-law slope
(with model BPL), respectively. The solid line shows the best fit
PLCool model that imposes $\Gamma_{X}=\Gamma_{NIR}+0.5$.
The X-ray slope is slightly steeper ($\Delta\Gamma=0.57\pm0.09$,
1$\sigma$), although consistent with the predictions of the PLCool
model. Both X-ray and NIR data and models have been corrected
for absorption and the effects of dust scattering halo.
The blue solid line shows the best fit TSSC model (\S 5.3.3).
For a description of the other lines see Fig. \ref{SED}. }
\label{PLCool}
\end{figure}
To investigate the effects of potential uncertainties on the normalisation
of the NIR emission, we artificially increased (and decreased)
the {\it SINFONI}\ spectrum by a factor 1.25 (and 0.75). The statistical quality
of the fit does not change ($\chi^2=156.8$ for 141 dof, in all
cases), and neither does the best fit photon index, $\Gamma_{NIR}=1.74\pm0.08$.
As expected, the main effect of the higher (lower) NIR normalisation is
to shift the break towards lower (higher) energies, i.e.
$E_{br}=0.018^{+0.071}_{-0.014}$~keV ($E_{br}=0.046^{+0.20}_{-0.037}$~keV),
corresponding to $B=11.7^{+8.5}_{-3.3}$~G ($B=8.5^{+5.5}_{-3.0}$~G).
\section{Evolution during VB3}
\label{SecSpEvol}
\subsection{Light curves of VB3}
The black and red points in Fig. \ref{LCp} show {\it XMM-Newton}\ and {\it NuSTAR}\
light-curves of VB3 in the 2-10~keV and 3-20 keV bands, respectively.
The black dashed line indicates the level of diffuse and quiescent
emission observed by {\it XMM-Newton}. For display purposes, we subtracted a
constant rate of 0.13~cts~s$^{-1}$ from the {\it NuSTAR}\ light curve, to take
into account the different contribution of the diffuse and quiescent
emission. The squares in Fig. \ref{LCp} show the NIR light-curve as
observed with {\it SINFONI}. Despite the sparse sampling of the light curve
allowed by the {\it SINFONI}\ integral field unit, it is clear from Fig. \ref{LCp}
that for VB3 the NIR flare lasts longer than the X-ray one.
In particular, the NIR flare is already in progress during our first {\it SINFONI}\
integration, $\sim10^3$~s before the start of the X-ray flare and it is
still in progress at the end of IR4, with a duration longer than $3.4$~ks
(see Fig. \ref{LCp}, Tab. \ref{obsid} and \ref{IRtime}).
This is not surprising, indeed, a similar trend has already been observed
by Dodds-Eden et al. (2009) and Trap et al. (2011) in the only other very
bright flare with simultaneous NIR coverage (see Fig. 3 of Dodds-Eden et al. 2009).
The pink dashed lines in Fig. \ref{LCp} indicate the start and end
of the {\it XMM-Newton}\ VB3 flare as determined by the Bayesian block
decomposition (see Ponti et al. 2015a).
The dotted green lines indicate the periods during which VB3-Pre, VB3-Rise,
VB3-Peak, VB3-Dec and VB3-Post, have been integrated (Tab. \ref{obsid}).
\begin{figure}
\includegraphics[height=0.37\textwidth,angle=0]{cluceFlare3.eps}
\caption{The black and red points show the {\it XMM-Newton}\ 2-10 keV (sum of the three
EPIC cameras) and {\it NuSTAR}\ 3-20~keV light curve of Sgr~A$^\star$'s flare VB3
(Ponti et al. 2015), respectively. A constant rate of 0.13 cts s$^{-1}$
has been subtracted from the {\it NuSTAR}\ light curve for display purposes,
to take into account the different contribution of the diffuse and quiescent
emission (gaps in the {\it NuSTAR}\ light curve are due to Earth occultation).
The squares show the extinction corrected {\it SINFONI}\ light curve
of Sgr~A$^\star$\ during the VB3 flare. Each point corresponds to a NIR
spectrum integration time of 600~s. The y-axis reports the observed
renormalised (divided by 4 for display purposes) flux density at
$2.2~\mu$m in mJy units. The black dashed line indicates the level
of the ``non-flare" (quiescent in X-rays) emission. The pink dashed lines indicate
the start and end of the {\it XMM-Newton}\ VB3 flare, as indicated by the Bayesian
block decomposition (see Ponti et al. 2015a). Excess X-ray emission
is observed $\sim2000$ and $\sim4000$~s after the X-ray flare peak.
The dotted green lines show the intervals
for the integration of the pre-, rise, peak, decrease and post-flare
spectra during VB3. The zero point of the abscissa corresponds
to $525831144.7$~s (TT$_{\rm TBD}$) and 2456900.50784~day
(BJD$_{\rm TBD}$), respectively. }
\label{LCp}
\end{figure}
\subsection{NIR spectral evolution during VB3}
\label{NIR}
We fit all the seven high quality {\it SINFONI}\ spectra (see top panel of
Fig. \ref{GammaIR}) with a simple power-law model, normalised
at $2.2~\mu$m ({\sc pegpwrlw}).
The fit with this simple model provides
a $\chi^2=96.8$ for 56 dof. The bottom panel of Fig. \ref{GammaIR}
shows the best fit photon index ($\Gamma_{NIR}$, where $\Gamma=1-\alpha$
and $\alpha$ is the spectral index $F_\nu\propto\nu^\alpha$) as a function of the
flux density (in mJy) at 2.2~$\mu$m.
During the {\it SINFONI}\ observations the $2.2~\mu$m flux density ranges
from $\sim3$ to $\sim10$~mJy, spanning the range between
a classical dim and bright NIR period (Bremer et al. 2011).
This suggests that this very bright X-ray flare is associated with a very
bright NIR flux excursion. In agreement with previous results, we observe
a photon index consistent with $\Gamma_{NIR}=1.6$ above $\sim7$~mJy
(solid line in Fig. \ref{GammaIR}; Hornstein et al. 2007; Witzel et al. 2014).
On the other hand, Fig. \ref{GammaIR} also shows steeper NIR spectral
slopes at low fluxes. We note that steep NIR slopes at low fluxes have
been already reported (Eisenhauer et al. 2005; Gillessen et al. 2006;
Bremer et al. 2011), however recent observations by Witzel et al. (2014)
indicate no spectral steepening at low fluxes. The results of our work appear
to suggest an evolution of the spectral slope at low fluxes during and after
this very bright X-ray flare, however higher quality data are necessary
to finally clarify this trend.
\begin{figure}
\includegraphics[height=0.35\textwidth,angle=0]{IRGammaFluxLast_eufLin.ps}
\includegraphics[height=0.34\textwidth,angle=0]{IRGammaFluxLast.ps}
\caption{{\it (Top panel)} {\it SINFONI}\ spectra fitted with a power-law model
in the energy band $E\sim0.525-0.608$~eV.
The color (black, red, green, blue, cyan, magenta and yellow)
indicates the chronological sequence of the spectra.
{\it (Bottom panel)} Best fit photon index ($\Gamma_{NIR}$) as a function of
the $2.2~\mu$m flux density (in mJy units). The NIR photon indexes
are shown with filled squares, with the same color code as before.
The empty dark grey circles show the spectral indexes in the 2-10 keV band,
during the rise, decay and peak of very bright flares. For these points,
we associate to the flare rise and decay X-ray photon indexes the simultaneous
NIR fluxes. For the flare peak, we assume a value of 11.5~mJy.
The dotted lines show the best fit of the NIR photon indexes with a linear relation.
The solid line shows the constant photon index typically
observed at medium-high fluxes (flux density $>7$~mJy; $\Gamma_{NIR}=1.6$;
Hornstein et al. 2007). The dashed line shows the associated X-ray
slope, if the spectrum is dominated by synchrotron emission with a cooling
break $\Gamma_X=\Gamma_{NIR}+0.5$. }
\label{GammaIR}
\end{figure}
\subsection{Multi-wavelength spectral evolution during VB3}
We extracted strictly simultaneous {\it XMM-Newton}\ and {\it NuSTAR}\ spectra for each
of the 7 NIR {\it SINFONI}\ spectra (see Fig. \ref{LC}). All are covered by {\it XMM-Newton}\
and {\it NuSTAR}, apart from spectrum IR4, for which Sgr~A$^\star$\ was not visible by
{\it NuSTAR}\ at that time (due to Earth occultation). The first four (from IR1
to IR4) of these spectra have been accumulated either when the X-ray
counterpart of VB3 was visible or in its close proximity and they
all show bright NIR emission, therefore we present the results of
the analysis of those "flaring spectra" here. The remaining three,
associated with faint NIR and X-ray quiescent emission, are investigated
in the next section (\S \ref{quie}).
We stress again that during the IR1 spectrum the flare was already
very bright ($F_{2.2\mu m}\sim9$~mJy) in the NIR band, while only
upper limits were observed in X-rays (see Fig. \ref{LCp} and \ref{SEDev}).
Indeed, the X-ray flare started roughly 20 min later, during IR2, and
peaked just after IR3.
Bright NIR emission with no X-ray counterpart in the early
phases of the flare places tight constraints on the PLCool model
(see \S \ref{discussion}). During IR4 the NIR flux was still high
($F_{2.2\mu m}\sim7.5$~mJy), while the X-ray flare was about
to end (Fig. \ref{LCp}). After IR4 the NIR droped significantly and
the X-ray emission returned to the quiescent level.
\begin{table}
\begin{center}
\begin{tabular}{ c c c c c }
\hline
\multicolumn{5}{c}{\bf Simultaneous (600 s) NIR to X-ray spectra during VB3} \\
\hline
\multicolumn{5}{c}{\bf BPL} \\
& $\Gamma_{NIR}$ & $\Gamma_{X}$ & $E_{br}$ & $\chi^2/dof$ \\
& & & (eV) & \\
IR1 & $1.5\pm0.2$ & $>2.2$ & 1\ddag & 20.8/15 \\
IR2 & $1.4\pm0.2$ & $3.2\pm0.4$ & $0.16^{+0.20}_{-0.11}$ & 32.8/24 \\
IR3 & $1.8\pm0.2$ & $2.57\pm0.16$ & $420^{+980}_{-210}$ & 43.8/43 \\
IR4 & $1.8\pm0.2$ & $2.14\pm0.02$ & 1\ddag & 11.4/10 \\
\hline
\multicolumn{5}{c}{\bf PLCool} \\
& $\Gamma_{NIR}$&&$E_{br}$ & $\chi^2/dof$ \\
& && (eV) & \\
IR1 & $1.72\pm0.04$&& $0.60\pm0.03$\dag & 23.7/15 \\
IR2 & $1.58\pm0.01$&& $0.61\pm0.03$\dag & 43.8/25 \\
IR3 & $1.87\pm0.07$&& $530^{+1400}_{-380}$ & 46.0/44 \\
IR4 & $1.77^{+0.06}_{-0.02}$ && $9.7^{+77}_{-8.0}$ & 11.2/10 \\
\hline
\multicolumn{5}{c}{\bf PLCoolEv} \\
& $\Gamma_{NIR}$ & $E_c$ &$E_{br}$ & $\chi^2/dof$ \\
& & (keV) & (eV) & \\
IR1 &$1.48^{+0.25}_{-0.05}$ & $6\times10^{-4}-1$& $0.6-1000$ & 17.4/14 \\
IR2 &$1.5\pm0.2$ & $3.5^{+0.3}_{-1.1}$& $1^{+2}_{-0.5}$ & 35.6/24 \\
IR3 &$1.82\pm0.07$ & $>9$ & $250^{+780}_{-150}$ & 45.3/43 \\
IR4 &$1.77^{+0.06}_{-0.02}$ & $>10$ & $9.8^{+77}_{-8.3}$ & 11.0/9 \\
\hline
\end{tabular}
\caption{Best fit parameters of Sgr~A$^\star$'s emission as fitted during
each of the (600~s) {\it SINFONI}\ and strictly simultaneous X-ray spectra
accumulated during the flare VB3. The spectra are fit with both
the BPL, the PLCool and the PLCoolEv models.
$\Gamma_{NIR}$ and $\Gamma_X$ indicates the power
law photon indexes fitting the NIR, the X-ray band, respectively.
For the PLCool and PLCoolEv models the $\Gamma_{NIR}$
indicates the best fit NIR slope, once the total band is fitted with
the assumption that $\Gamma_X=\Gamma_{NIR}+0.5$.
$E_{br}$ indicates the energy of the cooling break.
$E_{c}$ indicates the energy of the high-energy cut-off (induced
by $\gamma_{max}$). \dag The best fit energy of the break falls
right at the higher edge of the {\it SINFONI}\ energy band. \ddag Unconstrained
value, therefore fixed to 1~eV.}
\label{TabSEDev}
\end{center}
\end{table}
\subsection{Is the spectral evolution required?}
We started the time-resolved spectral analysis by testing whether
the data require any spectral evolution during VB3.
We therefore simultaneously fitted the multi-wavelength flaring
spectra (IR1, IR2, IR3 and IR4) with a broken power law model,
forcing the NIR and X-ray photon indexes and the break energy
to be constant over time.
This provides an unacceptable fit ($\chi^2=237.6$ for $104$ dof),
demonstrating that significant spectral variability is required
during the flare. The best fit photon indexes are
$\Gamma_{NIR}=1.71\pm0.09$, $\Gamma_{X}=2.21\pm0.10$,
with the break at $E_{br}=10^{+150}_{-3}$~eV. We note that, similar
to what has been found in the analysis of the mean spectrum, the spectral
steepening is $\Delta\Gamma=0.50\pm0.13$, therefore perfectly
consistent with $\Delta\Gamma=0.5$.
We then refitted the spectra with the same model, allowing
the NIR photon index and the break energy to evolve with time,
while imposing the X-ray photon index to be
$\Gamma_{X}=\Gamma_{NIR}+0.5$.
This provided a significant improvement to the fit ($\Delta\chi^2=97.6$
for the addition of 5 new parameters), demonstrating that Sgr~A$^\star$'s
spectrum changed shape during VB3. Indeed, we observe best fit
photon indexes of: $\Gamma_{NIR1}=1.70\pm0.05$,
$\Gamma_{NIR2}=1.60\pm0.08$, $\Gamma_{NIR3}=1.91\pm0.07$ and
$\Gamma_{NIR4}=1.81\pm0.13$, while the break is at
$E_{br1}=0.6\pm0.03$, $E_{br2}=0.9\pm0.03$,
$E_{br3}=1150^{+1800}_{-800}$ and $E_{br4}=14^{+400}_{-11}$~eV.
We note that this model can acceptably reproduce the
data ($\chi^2=140.0$ for 99 dof).
\subsection{Evolution of the BPL model during VB3}
Before considering the PLCool model, where the slopes in NIR
and X-rays are tied by the relation $\Gamma_{X}=\Gamma_{NIR}+0.5$,
we fitted each time-resolved multi-wavelength spectrum with the
phenomenological BPL model (\S \ref{BPL}), where the slopes
in the NIR and X-ray bands are free to vary (Tab. \ref{TabSEDev}).
The NIR slope was always well determined ($\Delta\Gamma_{NIR}\sim0.2$;
see Tab. \ref{TabSEDev} and \ref{IRtime}).
On the other hand, the presence of either upper limits or low statistics
prevented us from determining $\Gamma_X$ at the same time of $E_{br}$
in spectra IR1 and IR4 (Tab. \ref{TabSEDev}). We, therefore, "a priori" assumed
that the breaks in IR1 and IR4 occur at 1~eV (which corresponds to
$B=30$~G, if interpreted as a cooling break). We then constrained
the power-law slopes in the X-ray band under this assumption
($\Gamma_X>2.2$ and $\Gamma_X=2.14\pm0.02$ for IR1 and
IR4, respectively). Moving the break to 25~eV (corresponding to $B=10$~G)
the slope would steepen to $\Gamma_X>2.6$ and
$\Gamma_X=2.4\pm0.1$, respectively.
For IR2 and IR3, the X-ray data are of good enough quality to have a
good constraint on the X-ray slope (see Tab. \ref{TabSEDev}).
The dotted lines in the corresponding panels of Fig. \ref{SEDev} show
the uncertainties on the X-ray and NIR slopes.
The BPL model produced an acceptable description of the spectra
(the surviving residuals are due to intrinsic scatter in the NIR band;
see Tab. \ref{TabSEDev}).
\subsection{Evolution of the PLCool model during VB3}
\label{EvPLCool}
We then fitted the spectra with the PLCool model (Fig. \ref{SEDev};
Tab. \ref{TabSEDev}). This model reproduces Synchrotron emission
with a cooling break under the assumption that, at any time,
$\gamma_{max}>10^6$.
\subsubsection{Successes of the PLCool model}
For IR1, IR3 and IR4 the PLCool model provides a good fit to the data
of indistinguishable (at 90~\% confidence) quality compared to
the phenomenological BPL model (Tab. \ref{TabSEDev}).
The advantage over BPL is that the PLCool model is physically motivated.
We observed that for all spectra (from IR1 to IR4) the NIR spectra
are flat and consistent with being
constant (e.g. $\Gamma_{NIR}\sim1.6$, see Fig. \ref{GammaIR})
before and during the full duration of the X-ray flare. This is in line
with the values typically observed during bright NIR flux excursions
(Hornstein et al. 2007)\footnote{More observations are needed to
confirm the tentative hint for a steeper NIR slope in the early phase
of the NIR flare (IR1 and IR2 compared to IR3 and IR4; see
\ref{TabSEDev}).}. We also noted that at the peak of the X-ray flare,
when the constraints are best, the X-ray slope is steeper than
the simultaneous NIR one by $\Delta\Gamma\sim0.5$,
consistent with the one expected by the PLCool model (see
Fig. \ref{SEDev} and Tab. \ref{TabSEDev}).
\begin{figure*}
\includegraphics[height=0.49\textwidth,angle=90]{IR1x.ps}
\includegraphics[height=0.49\textwidth,angle=90]{IR2x.ps}
\includegraphics[height=0.49\textwidth,angle=90]{IR3x_v2.ps}
\includegraphics[height=0.49\textwidth,angle=90]{IR4x.ps}
\caption{Evolution of Sgr~A$^\star$'s SED during the very bright flare VB3.
Each panel shows the {\it SINFONI}\ and simultaneous X-ray spectra fitted with
the PLCool model, during each of the 4 {\it SINFONI}\ spectra integrated
during the VB3 flare (see Fig. \ref{LCp}). The colour code is the
same as in Fig. \ref{LCp}, with the temporal sequence: black; red; green and; blue.
The red and black dotted lines show the uncertainties in the determination
of the NIR and X-ray power-law slopes, respectively. The X-ray slopes
are well determined only for the IR2 and IR3 spectra.
The black dashed lines show the best fit PLCool models, where
$\Gamma_X=\Gamma_{NIR}+0.5$ is imposed. For IR2 the observed X-ray
slope is inconsistent with the predictions of the PLCool model.
For both IR1 and IR2 the cooling break is suspiciously pegged in
the NIR band. The black solid lines show the best fit PLCoolEv models.
During IR1 both the cooling break and the cut-off have large uncertainties,
but are constrained to lie within few $10^{14}<\nu<10^{18}$~Hz. During IR2
the cut-off is in the X-ray band.
From IR2 to IR3 the cooling break evolves to higher energies and then
back to lower energies during IR4.
Sgr~A$^\star$\ is undetected in X-rays during observation IR1.
Such as in Fig. \ref{PLCool} both data and models are de-absorbed and
corrected for the effects of the dust scattering halo.
For a description of the other lines see Fig. \ref{SED}.}
\label{SEDev}
\end{figure*}
Blindly applying the PLCool model to all time resolved spectra
(it might be incorrect to apply the PLCool model when no X-ray
emission is detected), we observed a significant evolution
of the energy of the cooling break, that implies
(under the assumption of a constant escape time)
a variation of the strength of the magnetic field (Tab. \ref{TabSEDev}).
The black, red, green, blue and grey dotted and solid lines
in Fig. \ref{Cont} show the 68 and 90~\% confidence contours
of the uncertainty on $E_{br}$ and $\Gamma_{IR}$ for IR1, IR2, IR3,
IR4 and the mean spectrum, respectively. We note that a highly
significant evolution of the cooling break is observed.
Indeed, during both IR1 and IR2, the break appears to be at very
low energy, corresponding to a magnetic field of the order of
$B\sim35$~G. While the energy of the break is significantly
higher during IR3, indicating that the magnetic field had significantly
reduced around the peak of the X-ray flare ($B=3.8^{+2.0}_{-1.3}$~G).
The energy of the break then drops again in the decreasing flank of
the X-ray flare to a value of $E_{br}=9.7^{+77}_{-8.0}$~eV,
corresponding to an increase in the strength of the magnetic
field ($B=14.3^{+11.3}_{-7.4}$~G).
\subsubsection{Difficulties of the PLCool model}
\label{PLCoolEv}
As we have outlined, the PLCool model (which assumes
$\gamma_{max}>10^6$ at all times) presented many successes.
However, we also point out here three severe weaknesses
that will be discussed further in the discussion section:
i) twice out of four times the cooling break is observed to peg
just above the NIR band ($E\sim0.6$~eV). This appears as a rather
unlikely possibility; ii) during IR1, bright and flat
($\Gamma=1.48\pm0.2$) NIR emission is associated to no
enhanced X-ray emission
($F_{3-10~keV}<2.3\times10^{-12}$~erg~cm$^{-2}$~s$^{-1}$).
This is hard to reconcile with the PLCool model that, in order to
fit this spectrum, pushes the best fit NIR photon index
to $\Gamma_{NIR}=1.72\pm0.04$; iii) the very steep
X-ray spectrum during IR2 ($\Gamma_X=3.2\pm0.4$) implies a
spectral steepening incompatible with the PLCool model
($\Delta\Gamma=1.8\pm0.4$, instead of $\Delta\Gamma=0.5$).
\begin{figure}
\includegraphics[height=0.49\textwidth,angle=-90]{ContPLAll.ps}
\caption{The black, red, green, blue and grey lines show confidence
contours of the uncertainty on $E_{br}$ and $\Gamma_{IR}$ for IR1,
IR2, IR3, IR4 and the mean spectrum once fitted with the PLCool model,
respectively. The dotted and solid lines show the 68 and 90~\% confidence
contours, respectively. During IR1 and IR2 the cooling break pegs
at its lowest value, being located just above the NIR band.
Once fitted with the PLCoolEv model, the confidence contours remain
unchanged for IR3, IR4 and the mean spectrum, because during
these intervals the high energy cut off is at very high energy.
The magenta lines show the confidence contours during IR2,
when the cut off is observed in the X-ray band. During IR1, the cut off
falls between the NIR and X-ray band, therefore the location of
the cooling break is unconstrained. }
\label{Cont}
\end{figure}
\subsection{Can the TSSC model fit the IR2 spectrum?}
In theory SSC models, with a thermal distribution, can produce fairly
steep spectral shapes at high energies. Therefore, although the TSSC
model produced unreasonable parameter values when applied to the mean VB3
spectrum, we checked whether TSSC might be the dominant radiative
mechanism during peculiar and short duration intervals, such as IR2.
The best fit TSSC model was significantly worse than the PLCool model,
despite having
two more free parameters ($\chi^2=60.1$ for 23 dof). Indeed, the model
failed to produce a better fit because it was mainly constrained by the flat
X-ray photon index produced by the TSSC model.
In addition, we noted that
the fit of IR2 leaded to unreasonable best fit parameters, similar to the
ones fitting the mean VB3 spectrum.
Indeed, we observed: $Log(B)=3.55$, $\theta_E=32.6$, $Log(N_e)=39.3$
and $Log(R_F/R_S)=-3.0$. Once again the magnetic field strength
appears, unreasonably large, the size of the source unreasonably
small, and the source density many orders of magnitude higher than
expected.
\subsection{Synchrotron emission with cooling break and evolving
$\gamma_{max}$ (PLCoolEv)}
\label{Evgamma}
When we considered the PLCool model, we "a priori" used
the assumption that at any time the source can accelerate particles
to very high energies $\gamma_{max}>10^6$. This implies
that an engine, able to accelerate electrons to $\gamma_{max}>10^6$,
is created on a negligible time-scale, at the start of the NIR flare.
If so, the PLCool model can be applied to the entire duration of
the flare (as we performed in \S \ref{PLCoolEv}).
On the other hand, if the engine has a size of a few Schwarzschild
radii, its light (or Alfven speed) crossing time would be of the order
of a hundred seconds, comparable to the time-scales of the flares
under investigation here.
Thus, it might be possible that the creation and destruction of the engine
occurs on a similar time-scale to the flares and that the engine is
initially not powerful enough to accelerate
particles to $\gamma_{max}>10^6$. Based on these considerations,
we introduce a new phenomenological model dubbed PLCoolEv,
by adding to the PLCool model the freedom of having a variable value of
$\gamma_{max}$. We performed this by adding a high energy
cut off to the PLCool model (reproduced by the {\sc highecut} component
in {\sc XSpec}). Indeed, the PLCoolEv model assumes that
$\gamma_{max}$ evolves during the flare. A low value of
$\gamma_{max}$ (e.g., $\gamma_{max}<<10^6$) would imply
that no electrons are accelerated to such high energies to produce
X-ray photons\footnote{Assuming that all emission is radiated at
0.29 times the critical frequency, it follows that:
$\nu_c\sim2.5\times10^{19}$~Hz $(\gamma_{max}/10^6)^2$ $(B/20$~G$)$
(Longair 2011, equation 8.127), where $\nu_c$ is the frequency
associated to the high energy cut off.}. As a result, the emitted spectrum
would show a high energy cut off ($E_c$) at an energy related to
$\gamma_{max}$ and lower than the X-ray band.
As in the PLCool model (with free parameters: $B$, $p$ and
normalisation), this model is characterised by the cooling
break $E_{br}$, linked to the strength of the magnetic field,
plus a cut-off at high-energy $E_c$ (induced
by $\gamma_{max}$).
We assumed an exponential shape above the cut off energy.
We noted that such shape is constrained only by the IR2 spectrum
and it appears steeper ($\Gamma_X=3.2\pm0.4$)
than the simultaneous and cooled NIR slope (see Tab. 6).
Therefore any high energy slope steeper than $\Gamma_X\sim3$ could
reproduce the data. We point out that either an exponential or sub-exponential
slope can equally fit the data. We also note that most likely the electron
distribution will not cut abruptly at $\gamma_{max}$, therefore it is expected
that the cut off will be further broadened. Despite we could not constrain
whether the break is broad, we fixed the shape of the high energy cut off
such that the e-folding energy of the exponential cut off is equal to the cut
off energy.
The PLCoolEv model provides an excellent representation
of the multi-wavelength spectrum at all times during the flare
(see Tab. \ref{TabSEDev} and Fig. \ref{SEDev}).
It produces either superior fits compared to PLCool model
(in particular for IR2), or of comparable statistical quality
to the BPL parametrisation and it is physically motivated.
Figure \ref{Cont} shows the confidence contours projected
over the $E_{br}$ versus $\Gamma_{NIR}$ plane for
the PLCoolEv model. The high-energy cut-off is constrained to be
at $7\times10^{-4}<E_c<1$~keV in the IR1 spectrum, before the start
of the X-ray flare (not shown in Fig. \ref{Cont}). In the PLCool model,
the steep photon index observed in X-ray during IR2 is the result of
the evolution of the high energy cut-off, which at that time was detected
in the X-ray band at $E_c=3.5^{+0.3}_{-1.1}$~keV. As a consequence
of this, the cooling break is not pegged anymore at $E_{br}=0.6$~eV,
instead it spans a larger range of reasonable cooling break energies.
Additionally, a flatter NIR slope is allowed.
During IR3, IR4 and the mean spectrum, the high energy cut-off
was at energies higher than the observed X-ray band ($E_c\gg10$~keV),
consistent with the assumptions of the PLCool model (indeed, we obtained
similar results). According to the PLCoolEv model,
during the early phase of VB3, the high energy cut-off was evolving
and it was located between the NIR and X-ray band. It was caught
within the X-ray band during IR2 and it was at very high energy
at the X-ray peak (and during IR4).
\subsubsection{Evolution of the magnetic field (assuming a constant escape time)}
\label{SecEvB}
In this section we interpret the derived evolution of the energy of
the cooling break, as being uniquely due to the variation of
the magnetic field of the source (e.g., assuming no variation of the
escape time).
Figure \ref{EvolPLCool} shows the light curve of the evolution
of the magnetic field intensity during the flare.
The flat NIR slope observed at all times indicated
that the break has to be, at higher frequency compared
to the {\it SINFONI}\ band, corresponding to $B<36$~G.
In particular, during IR2 the cooling break is observed at
$E_{br}=1^{+2}_{-0.5}$~eV, corresponding to $B=30\pm8$~G.
The values derived by fitting the IR1 spectrum are consistent
with this value, however the degeneracy between the energy of
the cooling break and of the high energy cut-off led to large
uncertainties on the magnetic field strength.
We note that a value of $B=30\pm8$~G is fully consistent with
the magnetic field present within the central ten Schwarzschild
radii and generating the steady emission of Sgr~A$^\star$.
During IR3, close to the peak of the X-ray flare, the magnetic
field is observed to be $B=4.8\pm1.7$~G. Interestingly,
the magnetic field varied by a factor of $>6$ in less
than $\sim650$~s. During IR4, we assumed that the cut off
is located at energies higher than the X-ray band (indeed no evidence
for a cut off at or below the X-ray band is observed). Under this
assumption\footnote{ Would, during IR4, the cut off be located
in the X-ray band or below, then the current constraints on the
energy of the cooling break should be considered only as upper limits.
If this is indeed the case during IR4, then weaker magnetic fields would
be allowed and the data point in Fig. \ref{EvolPLCool} should be
considered as an upper limit. }, we observe that after the X-ray peak
and towards
the end of the X-ray flare, the magnetic field was measured
to rise again to values $B=14.3^{+12.3}_{-7.0}$~G. The red point in Fig.
\ref{EvolPLCool} shows the magnetic field strength derived from
the fit of the mean spectrum of VB3. As expected, the average
magnetic field value during the flare ($B=8.8^{+5.0}_{-3.0}$~G)
was intermediate between IR2, IR3 and IR4 and it was significantly
smaller than the one derived during quiescence.
\begin{figure}
\includegraphics[height=0.35\textwidth,angle=0]{EvoPLCoolEvG-B2.ps}
\caption{Evolution of the strength of the magnetic
field (in Gauss) during and after the flare VB3, in the
PLCoolEv model. The black square show the
magnetic field strength during IR1 to IR4. The red circle shows
the measurement of the average magnetic field during the entire
duration of the X-ray flare VB3 (under the assumption of a constant
escape time of $t_{esc}=300$~s). The dotted lines indicate the typical
range of magnetic field strengths during quiescence ($B\sim20-50$~G).
The dashed line shows the magnetic field strength ($B\sim36$~G)
corresponding to a cooling break within the narrow {\it SINFONI}\ band.
The flat NIR slopes observed at all times during the flare suggest
$B<36$~G, while the steep NIR slope observed after the VB3 flare
(during IR5, IR6 and IR7) suggest $B>36$~G. The error bars
correspond to the 1-sigma uncertainties as derived from the confidence
contours shown in Fig. \ref{Cont}. No evidence for a cut off at or
below the X-ray band is observed during IR4, therefore the associated
measurement is valid under the assumption that the cut off is at energies
higher then the X-ray band. Would this assumption be invalid,
such constraint should be considered as a upper limit. }
\label{EvolPLCool}
\end{figure}
\subsubsection{Evolution of the escape velocity (constant magnetic field)}
The results obtained in section \ref{SecEvB} are valid under
the assumption that the synchrotron escape time is constant over the
entire duration of the flare. However, it is not a priori set that the escape
time has to remain constant over time. Therefore, we now investigate
the hypothesis that the escape time evolves, while assuming
a constant magnetic field ($B=30$~G).
If so, the energies of the break frequencies observed during IR2,
IR3 and IR4 ($E_{br}=1^{+2}_{-0.5}$, $250^{+780}_{-150}$
and $9.8^{+77}_{-8.3}$~eV) correspond to an escape velocity of
$t_{esc1}=310\pm130$~s, $t_{esc2}=20\pm10$~s and
$t_{esc3}=100^{+150}_{-70}$~s. Therefore, the escape time would
drop by a factor of $\sim16$ in $\sim600$~s, to then increase again.
The escape time is likely related to the source size and/or to the position
of the source onto the accretion disc. For example, Dodds-Eden et al. (2009)
assume that the escape time is comparable to the dynamical time at a given
radius in an accretion disc ($t_{dyn}$):
$t_{esc}\sim t_{dyn} = \sqrt{R^3/2GM_{BH}}$, where $R$ is the source
position within the accretion disc, $G$ is the gravitational constant and
$M_{BH}$ is the black hole mass.
We note that, under this assumption, the escape time assumed throughout
the paper ($t_{esc}=300$~s) corresponds to a reasonable radial position of
$\sim3.5$~R$_S$ from the BH.
\
In summary, the PLCoolEv can adequately fit not only the mean properties
of the VB3 flare but also its evolution. The major weaknesses of the PLCool
model are solved by allowing the high energy cut-off ($\gamma_{max}$)
to evolve during the flare.
We note that the evolution of the cooling break appears more
likely induced by a variation of the magnetic field that drops its intensity
by discharging magnetic energy density into particle acceleration
and then rises again to its average value, compared to a variation
of the escape velocity. Indeed, in the latter scenario it would naively
be expected that the energy release produced by the source would
make the source size expand with time, instead of contracting.
However, we point out that these considerations are not conclusive.
Indeed, because of the limitations of our simplified single
zone models, we can not discriminate between a pure magnetic field
evolution or a pure escape time evolution (or a combination of both).
\section{Emission after VB3 (X-ray quiescence)}
\label{quie}
IR5 and IR6 have been accumulated after the end of VB3 when
only upper limits are observed in X-rays and the NIR flux
($F_{2.2\mu m}\leq4.5$~mJy) corresponds to the faintest fluxes
of Sgr~A$^\star$, detected so far (e.g. Dodds-Eden et al. 2010).
These time intervals appear similar to classical quiescent periods.
During IR7 a re-brightening is observed in NIR, associated with a hint
for an excess in the X-ray band (Fig. \ref{LCp}). Though the NIR
flux is relatively high ($F_{2.2\mu m}\sim7$~mJy), the NIR spectral
slope appears steeper than during the flare and fully consistent with
the value observed during NIR quiescence. The sparse NIR light curve
as well as the low significance of the X-ray excess do not allow us
to clarify whether the emission during IR7 is produced by a faint flare,
with associated feeble X-ray emission or it is just a NIR fluctuation
characteristic of a red noise process, commonly occurring during
X-ray quiescence.
We observed that the NIR spectra steepened $\sim15$~min
after the end of the X-ray flare (see Tab. \ref{IRtime}).
The steepening was so large
($\Delta\Gamma_{NIR}\sim1-2$) that in all cases (from IR5 to IR7)
the extrapolation of the steep power-law observed in the
{\it SINFONI}\ band (Tab. \ref{IRtime}) was consistent with the X-ray
upper limits. Figure \ref{GammaIR} shows
that at medium-high NIR fluxes ($F_{2.2~\mu m}>7$~mJy) the photon
index is consistent with a constant value of $\Gamma_{NIR}=1.6$.
This implies an electron distribution index of $p\sim2.2$.
We note that in the PLCool (and PLCoolEv) model, as long as
$B>35-40$~G, the cooling break would move to frequencies lower
than the {\it SINFONI}\ band, inducing a steepening of the observed photon
index by $\Delta\Gamma_{NIR}=0.5$.
However, this steepening appears too small to reproduce the full
extent of the observed photon index variation.
We remind the reader that accurate photon index determination
at low NIR fluxes are challenging. Therefore, we leave to future dedicated
studies to establish the full extent and the reliability of the NIR spectral
steepening at low NIR fluxes. If future data confirms the presence
of such steep NIR slopes after very bright flares, then this radiation
might be associated with thermal Synchrotron emission from electrons
transiently heated during VB3.
Indeed, we note that, for a magnetic field strength of $B\sim30$~G,
the cooling time of NIR synchrotron electrons is of the order of
$\sim500$~s, shorter, but comparable, to the time interval between
the end of IR4 and the start of IR5 (see Fig. \ref{LCp} and Tab. \ref{IRtime}).
\section{Does a slow evolvution of $\gamma_{max}$ agree with
the evolution of bright X-ray flares?}
\label{SpEvol}
The detailed investigation of the X-ray and NIR emission during VB3
indicates the PLCoolEv as the favourite model (see \S 5 and 6).
The essential component that distinguish the PLCoolEv model from
the simpler PLCool model is the evolution of the cut-off ($\gamma_{max}$).
In particular, we suggest that the evolution of $\gamma_{max}$ might
be relatively slow, spanning the range from optical-NIR to X-rays and
beyond on macroscopic time-scales ($\sim10^2-10^3$~s).
We also note that the X-ray band has significant extensions in
frequency, spanning over a decade in frequency.
Therefore, should the PLCoolEv model be correct and should the
behaviour observed during the very bright flare VB3 be universal,
then this model would predict an energy dependent evolution of
X-ray flares that might be tested with archival data of other bright
X-ray flares.
Indeed, it is expected that the passage of the cut-off (induced
by $\gamma_{max}$) through the X-ray band would induce slightly
shorter flares at higher energies as well as steeper spectral slopes
at the start and end of the X-ray flare.
Clearly the full extent of these effects can not be predicted, because
it depends on how rapidly the cut-off spans the X-ray band, but
we investigated whether we can exclude that such evolution is present
during bright X-ray flares.
Indeed, despite the fact that at present there are only few bright and
very bright flares with multi-wavelength coverage and only one (VB3)
with simultaneous NIR and X-ray spectra, the {\it XMM-Newton}\ archive contains
several bright flares suitable for studying the spectral evolution
in the X-ray band\footnote{We do not consider bright flares
detected by {\it Chandra}, because the vast majority of those are affected
by strong pile-up, significantly distorting the spectral shape. }.
\subsection{Time dependence of X-ray spectra of very bright flares}
To follow the evolution of Sgr~A$^\star$'s X-ray emission during very bright
X-ray flares, we consider here all bright and very bright flares
observed by {\it XMM-Newton}\ (see Tab. \ref{obsid}; Ponti et al. 2015a).
For each of these flares we extract three spectra, one during the rise,
one at the peak and one during the decay (see Tab.
\ref{obsid})\footnote{For flares VB1 and VB2 we chose three intervals
of equal duration, while for flare VB3, the extraction of the spectra
during flare rise, peak and decay are chosen in order to optimise the
coverage of the {\it SINFONI}\ spectra (see Tab. \ref{obsid}, \ref{IRtime} and
Fig. \ref{LCp}).}.
We fit the spectra from both the EPIC-pn and MOS data during the peak
of VB1, VB2 and VB3 with the absorbed power-law model.
The observed spectral indices
at peak are consistent between the different flares; therefore we assume
the same value and redo the fit to obtain a best fit value of $\Gamma=2.08\pm0.11$
($\pm0.07$ at 1 $\sigma$). We then repeat this exercise fitting the spectra
during both the flare rise and decay, obtaining $\Gamma=2.33\pm0.23$
and $\Gamma=2.45\pm0.25$, respectively. Again, we observe consistent
values for the rises and decays of different flares. Therefore we assume
the same spectral index during both flanks of the flares, obtaining a
best fit value of $\Gamma=2.36\pm0.15$ ($\pm0.09$ at 1 $\sigma$),
slightly steeper ($\sim2.4 \sigma$ significance) than the spectral index
observed at peak. We conclude that the spectra of very bright flares
provide hints for (or at least are not in disagreement with) an evolution
of the order of $\Delta\Gamma\sim0.3$ between the peak and
the flanks of the flares. This behaviour is reminiscent of what was
observed during the evolution of VB3. Indeed, during the early phases
of the X-ray emission (IR2), the X-ray spectrum was steeper than at
the peak of the X-ray emission (IR3; \S \ref{SecSpEvol}), most
likely because of the evolution of $\gamma_{max}$.
\subsection{Colour dependence of bright X-ray flares}
Figure \ref{LC} shows the 2-10 keV band light curves of the three
very bright flares observed by {\it XMM-Newton}\ (Ponti et al. 2015a). We combined
the light curves from the three EPIC cameras.
Figure \ref{LC} shows a remarkable similarity in the evolution of these
flares, suggesting an analogous origin.
We fit each of the light curves with a model composed by a constant
plus a Gaussian profile (to fit the flare)\footnote{The variation of the non-flare
emission during both observations taken on 2007-04-03 and
2014-08-31 was produced by the contribution from the magnetar SGR J1745-2900\
(at the level of $\sim50$ \% to the total observed quiescent
flux on 2014-08-30; see also Ponti et al. 2015a) and from the very
bright source AX~J1745.6-2901\ (Ponti et al. 2015b). }.
\begin{figure}
\includegraphics[height=0.5\textwidth,angle=90]{xmm_obscompare_pnmos_2_10.ps}
\caption{Light curves, in the 2-10~keV band, of the three brightest {\it XMM-Newton}\
flares, VB1, VB2 and VB3 are shown with back squares, red circles and
blue stars, respectively.
The light curves of the three very bright flares observed by {\it XMM-Newton}\ show
very similar time evolution and comparable duration. These light curves
are the result of the sum of the data from EPIC-pn and MOS.
For display purposes, we shifted the time axis aligning the peak of
the Gaussian best fitting the flare profile (see Tab. \ref{modLC}). }
\label{LC}
\end{figure}
\begin{figure}
\includegraphics[height=0.47\textwidth,angle=-90]{length4.ps}
\includegraphics[height=0.33\textwidth,angle=0]{TimeDelay2.ps}
\caption{{\it (Top panel)} Flare duration (FWHM) at various energies
of the three brightest {\it XMM-Newton}\ flares, VB1, VB2, VB3 and of the combined
{\it XMM-Newton}\ flare light curve are shown with back squares, red circles, blue stars
and green triangles, respectively. The flare duration are computed in the
2-4, 4-6 and 6-10~keV energy bands (the points are slightly shifted for
display purposes). Flares are typically shorter at higher energies.
We also show the duration of VB2 (as observed in the L' band,
see \S \ref{NIR}) and the lower limit on the duration of VB3 (as observed
with {\it SINFONI}). The flare durations in the NIR band connect with the extension
of the trend observed in X-rays.
{\it (Bottom panel)} Peak occurrence delay, between different energies
(same energy bands as above) for the three very bright flares observed
by {\it XMM-Newton}. The delays are computed as the best fit peak value of the
Gaussian at each energy minus the same value observed in the 4-6 keV
band. Colour code as before. }
\label{LCEne}
\end{figure}
To investigate possible dependences of the X-ray flares on energy, we
extracted the X-ray light curves of the three very bright {\it XMM-Newton}\ flares
in the 2-4, 4-6 and 6-10 keV energy bands. For each flare, we fit
the light curves with a constant plus a Gaussian profile, to characterise
the flare shape (see Table \ref{modLC}).
The top and bottom panels of Fig. \ref{LCEne} show the best fit
flare duration ($FWHM$) and delay as a function of energy.
For each energy we report with black squares, red circles and blue stars
the values obtained for the flares VB1, VB2 and VB3, respectively.
In particular, we show the width of the best fit Gaussians as a proxy
for the flare duration and the delay is defined as the peak time of the
Gaussian at a given energy minus the peak time in the 4-6~keV band.
The top panel of Fig. \ref{LCEne} suggests that Sgr~A$^\star$'s flares shorten
with energy, typically lasting $\sim5~$\% less time in the hard band (6-10~keV)
compared to the soft one (2-4~keV). The top panel of Fig. \ref{LCEneComp}
shows the combination of all three very bright flares. With the solid red,
dotted orange and dashed blue lines the light curves in the 2-4 keV,
4-6 keV and 6-10 keV energy bands are shown. The light curves are
shifted by the center of their best-fit Gaussian profile, subtracted by the
best-fit local underlying continuum, and normalised by the peak of their
best-fit Gaussian (see Tab. \ref{modLC}).
The flare profile is tighter at higher energies
(Fig. \ref{LCEneComp}, Tab. \ref{modLC}).
Indeed, the width of the Gaussian fitting the
6-10~keV band appears to be significantly smaller (at $\sim4.4\sigma$
significance) by $\sim360$~s compared to the 2-4~keV band one
(see Tab. \ref{modLC}).
To test whether this is a common property of all X-ray flares or whether it
is a peculiarity of very bright flares, we combined all {\it Chandra}\ bright
and very bright flares (i.e. with fluence larger than
$5\times10^{-9}$~erg~cm$^{-2}$; see Ponti et al. 2015a for the definition).
To avoid flares significantly affected by pile up, we excluded the ones
observed in either ACIS-I or ACIS-S with no subarray mode and reaching a
block count rate equal or higher than 0.1~ph~s$^{-1}$ (see
Ponti et al. 2015a for details). We also excluded the flares observed
in ACIS-S 1/8 subarray mode and reaching a block count rate equal
or higher than 0.8~ph~s$^{-1}$ (Ponti et al. 2015a).
The bottom panel of Fig. \ref{LCEneComp} shows the combined
{\it Chandra}\ flare light curve in the 2-4.5 and 6-9 keV energy bands with
red circles and blue squares, respectively. We find that bright {\it Chandra}\ flares also
last longer in the soft energy band compared to the hard one, with
a difference in $FWHM$ of $\sim300$~s (Fig. \ref{LCEneComp} and
Tab. \ref{modLC}).
No significant time shift with energy is apparent, with upper limits as tight
as $\sim100-200$~s (see bottom panel of Fig. \ref{LCEne}).
\begin{figure}
\includegraphics[width=0.23\textwidth,angle=90]{xmm_composite_lc.ps}
\includegraphics[width=0.23\textwidth,angle=90]{cxo_composite_lc.ps}
\caption{
{\it (Upper panel)} The combination of the light curves of VB1, VB2 and VB3
in the 2-4 keV (red squares), 4-6 keV (orange stars), 6-10 keV (blue squares),
light curves are shifted by the center of their best-fit Gaussian profile,
subtracted by the best-fit local underlying continuum, and normalised by
the peak of their best-fit Gaussian. Time bins of 150~s are used.
Flares are shorter in the hard band.
{\it (Bottom panel)} Chandra composite light curve for 2-4.5 keV (red circles)
and 6-9 keV (blue squares). All bright flares not significantly affected by pile-up
(peak count rate $<0.1$~ph~s$^{-1}$) are considered here. The light curves
are shifted with the same method described above. Time bins of 250~s are used.
Flares last longer in the
soft band. }
\label{LCEneComp}
\end{figure}
\begin{table}
\centering
\caption{Results of single Gaussian fitting to the 200 second binned flare light
curves without background subtraction, but with {\tt epiclccorr} applied.
Note that using the background subtracted light curves only changes these
values very slightly.
$^{1}$best-fit local continuum under the flare.
$^{2}$delay time related to the 4-6 keV band. The errors are all 1 $\sigma$.
The composite flare from {\it XMM-Newton}\ is the combination of VB1, VB2 and VB3.
The composite flare from {\it Chandra}\ is the combination of a set of unpiled-up
flares observed by {\it Chandra}.}
\centering
\begin{tabular}{llllll}
\hline
Flare ID & Band & Contiuum$^{1}$ & $FWHM$ & Delay$^{2}$\\
& {\it keV} & \multicolumn{1}{c}{{\it $cts~s^{-1}$}} & \multicolumn{1}{l}{{\it s}} & \multicolumn{1}{l}{{\it s}}\\
\hline
VB1 &2 - 4 & $0.0446\pm0.0017$ & $1500\pm85$ & $+8\pm$49\\
&4 - 6 & $0.0310\pm0.0014$ & $1550\pm57$ & 0\\
&6 - 10 & $0.0128\pm0.0009$ & $1385\pm61$ & $+17\pm$42\\
\hline
VB2 &2 - 4 & $0.0485\pm0.0008$ & $1860\pm130$ & $-25\pm$73\\
&4 - 6 & $0.0433\pm0.0008$ & $1570\pm80$ & 0\\
&6 - 10 & $0.0198\pm0.0005$ & $1610\pm100$ & $-16\pm$62\\
\hline
VB3 &2 - 4 & $0.0956\pm0.0016$ & $1840\pm153$ & $-48\pm$79\\
&4 - 6 & $0.0893\pm0.0015$ & $1455\pm80$ & 0 \\
&6 - 10 & $0.0273\pm0.0008$ & $1280\pm75$ & $-100\pm$53\\
\hline
Composite &2 - 4 & $0.0610\pm0.0007$ & $1813\pm68$ & --\\
(XMM) &4 - 6 & $0.0559\pm0.0007$ & $1554\pm42$ & --\\
&6 - 10 & $0.0242\pm0.0004$ & $1450\pm47$ & --\\
\hline
Composite &2 - 4.5 & $7.9\pm0.3\times10^{-4}$ & $1730\pm56$ & --\\
({\it Chandra}) &4.5 - 6 & $4.1\pm0.2\times10^{-4}$ & $1693\pm59$ & --\\
&6 - 9 & $6.0\pm0.4\times10^{-4}$ & $1424\pm89$ & --\\
\hline
\end{tabular}
\label{modLC}
\end{table}
We conclude that, at present, the X-ray data of the bright flares
are not in contradiction with a slow variation of $\gamma_{max}$.
\section{Discussion}
\label{discussion}
Simultaneous {\it XMM-Newton}, {\it NuSTAR}\ and {\it SINFONI}\ observations of Sgr~A$^\star$\
allowed us to determine, for the first time, the spectral shape
and the evolution of the radiation of a very bright flare.
This enabled us to pin down the radiative mechanism during bright
flares of Sgr~A$^\star$\ and its evolution during the flare.
We can rule out that a simple power-law model, representing plain
Synchrotron emission, can reproduce the flare emission.
\subsection{TSSC}
A TSSC model provides
an acceptable fit to the data, from a statistical point of view. However:
i) the fit is worse than the PLCool and PLCoolEv
models, despite the larger number of free parameters;
ii) in this framework the observed spectral steepening
($\Delta\Gamma=0.57\pm0.09$ at 1 $\sigma$) between NIR and X-ray
would be just a coincidence, moreover;
iii) the best fit parameters appear implausible.
Indeed, the best fit magnetic field appears unreasonably
high ($B\sim10^4$), the source size unphysically small
($R_F\sim10^{-2}-10^{-4}$~R$_S$) and the required source
density about seven orders of magnitude higher than what is
estimated to be present in the accretion flow around Sgr~A$^\star$\
(Loeb \& Waxman 2007; Genzel et al. 2010).
As discussed in Dodds-Eden et al. (2009), these unreasonable best fit
parameters are consequences of the assumptions intrinsic
to the TSSC model considered here. In particular, to fit the
soft X-ray emission via inverse Compton up-scattering of NIR or
sub-mm radiation, the energies of the electrons involved in the flare
is restricted to be lower than $\gamma_e<100$. On the other hand
the requirement of the observed hard NIR slope constrains the magnetic
field to be larger than $B>10^3$~G. Finally, the ratio of the synchrotron
to inverse Compton luminosity requires that the size of the source has to
be $R_F<10^{-2}$~$R_S$. Therefore, it appears that this simplistic
TSSC model, cannot adequately explain the flare emission.
The observed spectral steepening ($\Delta\Gamma=0.57\pm0.09$,
1 $\sigma$) between NIR and X-rays of the mean spectrum suggests that
the radiative process during bright flares might be synchrotron
radiation with a cooling break. Therefore, we explored in more
details this scenario, instead of considering more complex
synchrotron self Compton (SSC) models.
Nonetheless, this does not rule out that more complex TSSC and
non-thermal SSC models might be invoked to explain the X-ray
radiation.
\subsection{PLCool}
We observed that the synchrotron model with a cooling break can
reproduce both the mean spectrum and the evolution of the SED
of Sgr~A$^\star$\ during the entire duration of a very bright flare (apart from IR2).
In particular, the observed spectral steepening of both the mean
spectrum and of the emission at the X-ray peak (IR3) is a strong
indication that synchrotron with cooling break might be the dominant
radiative mechanism.
In this simplistic model, the ``unknown" motor powers the continuous
acceleration of energetic electrons with a power-law distribution.
An important difference of the PLCool model compared to the TSSC model,
is that the motor is assumed to accelerate electrons
into a power-law distribution up to $\gamma_{max}\geq10^6$,
therefore the synchrotron radiation is not limited to the NIR band,
instead it extends to X-ray and higher energies.
\subsubsection*{Limitations of the model}
We point out that, for simplicity, we reproduce the synchrotron
emission with a simple broken power law model. We note that
the cooling break typically occurs in the unobserved optical-UV band,
therefore we can not currently constrain whether the cooling break
is a sharp feature or it is significantly extended in energy.
Indeed, we do not observe any significant curvature in either
the NIR or X-ray band, however this has to be attributed to the
small frequency windows sampled by our data.
Therefore, for simplicity, we assume a sharp break, though realistic
synchrotron models can be significantly broadened (by up to more
than a decade in energy; Dibi et al. 2014). Indeed, it is beyond
the scope of this paper to employ more complex synchrotron models.
We also note that significantly different statistics characterise
the time resolved spectra in the NIR and X-ray bands. Therefore,
the broad band fit of the time resolved spectra are primarily driven
by the NIR photon index. It is important to point out that at
the beginning of the flare, the cooling break is observed to be located
either within or very close to the NIR band. Therefore, should
a broadened break be present at that time, it might potentially affect
some of the model parameters.
Future investigations will clarify the extent of this.
\subsubsection*{Difficulties of the PLCool model}
As already briefly mentioned in \S \ref{EvPLCool}, we stress
again here that the most difficult problem of the PLCool
model is related to the best fit energy of the cooling break.
Indeed, both for IR1 and IR2 the break is observed within the NIR
band (see Tab. \ref{TabSEDev}). During IR1 this result is driven
by the combined upper limit on the X-ray emission and by the high
flux and flat photon index in the {\it SINFONI}\ band, inducing a cooling
break suspiciously located at energies just higher than $\sim0.6$~eV,
the upper bound of the {\it SINFONI}\ spectrum. Similar results holds
during IR2. Indeed, at that time, the X-ray flare had already started,
alleviating the problem, however the NIR band showed a slightly flatter
power-law ($\Gamma_{NIR}=1.59\pm0.2$; Tab. \ref{IRtime}), therefore
the cooling break
was observed to again be placed at the upper bound of the {\it SINFONI}\ spectrum
(moreover the steep X-ray slope is not completely reproduced).
We consider a rather unlikely possibility that, by chance, the cooling
break occurred twice within the narrow NIR band.
The biggest pitfall that the PLCool model has to overcome
is the explanation of the early phases of the VB3 flare. Indeed,
at the basis of the PLCool model there is the assumption that
the NIR and X-ray emissions are tied by a broken power law.
Therefore, within this framework one would predict that the NIR
to X-ray emission are strictly related and they follow each other.
The only deviation to this "rule" could be generated by the possible
delay of the NIR radiation associated with the longer NIR synchrotron
cooling time. Therefore, it is expected that the X-ray emission either
rises before or at the same time as the NIR one. One possible way
out (that has been considered in the past to explain the delayed
X-ray emission in the early phases of the very bright flares)
was to assume that the early NIR emission had a very steep slope.
However, we can now rule out that this is happening during VB3.
Indeed, during IR1 bright NIR emission, with a flat slope
($\Gamma_{NIR}=1.48\pm0.2$; Tab. \ref{IRtime}), is observed at
the same time of tight upper limits to the X-ray emission.
On the contrary, no prominent X-ray radiation is observed either
during or before IR1, with upper limits in the 3-10 keV band
of $F_{3-10~keV}<2.3\times10^{-12}$~erg~cm$^{-2}$~s$^{-1}$.
The third difficulty of the PLCool model is to properly reproduce
the IR2 spectrum. A fit with the BPL model shows that the difference
of the X-ray to NIR photon index is significantly higher
$\Delta\Gamma=1.8\pm0.4$ than the one expected by the cooling
break model $\Delta\Gamma=0.5$
(Tab. \ref{TabSEDev}). This resulted in a poor fit of IR2 by the PLCool
model (indeed an F-test suggests that the BPL model provided
a significantly better description of the IR2 spectrum, at
$>99$~\% confidence).
\subsection{PLCoolEv}
\label{SecDiscPLCoolEv}
We then relaxed the requirement that $\gamma_{max}$ has
to be $>10^6$ at all times. We postulated that, $\gamma_{max}$
increases slowly with time (e.g. many times the Alfv\'en speed
crossing time of a source of a size of few Schwarzschild radii),
generating a cut-off that gradually moves to higher energies,
eventually transiting through the X-ray band and producing a
bright X-ray radiation with a delay compared to the start of the
NIR flare. Regardless of the behaviour of $\gamma_{max}$
at the end of the flare, the PLCoolEv (as well as PLCool) model
predicts a delay of a few hundred seconds of the NIR radiation,
compared to the X-ray emission (\S \ref{Evgamma}), in agreement
with a longer duration of the NIR flare.
\
\subsubsection*{Limitations of the model}
For simplicity, we assumed an exponential drop of the high energy
cut off in the synchrotron spectrum, with a shape such that the e-folding
energy is equal to the cut off energy. We note that the high energy
cut off is detected in the observed band only once, during IR2.
During this interval, the X-ray slope is steeper ($\Gamma=3.2\pm0.4$)
than the simultaneous NIR one, however the statistics is not enough
to discriminate its detailed shape. For example, we could not distinguish
either between an exponential or a sub-exponential, or we could not
constrain the broadness of the cut off.
Therefore, should a broadened break be present at that time,
it might potentially affect some of the model parameters. For instance,
broader cutoffs in IR2 and IR4 might allow for a cooling break at higher
energy, implying a weaker magnetic field. Future investigations will clarify
the extent of this.
\subsubsection{Comparison of the PLCoolEv model to the data}
The PLCoolEv model provides an excellent description of the mean
spectrum of VB3 and of its evolution over time.
In fact, it naturally explains the periods during which bright
and flat-spectrum NIR radiation is observed, simultaneous with
no X-ray emission (e.g. IR1). In particular, the passage of the cut-off
within the X-ray band generates: i) shorter flare durations at higher
energies; ii) right at the start of the X-ray flare steeper X-ray
spectra than expected by the cooling break (e.g. IR2) and;
iii) possibly steeper spectra in the flanks of the X-ray flare than
at the peak, such as observed.
We also measured a significant evolution of the cooling break
during the flare. Under the assumption that the escape time
remains constant, this suggests that the strength of the magnetic
field (typically of several tens of Gauss) lowers to values of few Gauss
during the peak of bright flares, to return to high values after that.
For a thermal distribution of electrons with temperature $\theta_E$,
the Synchrotron luminosity is proportional to the square of the magnetic
field ($L_{Synch}\propto N\theta^2_E B^2$), therefore the drop of
the magnetic field strength at the flare peak would appear contradictory.
However, in this scenario, the large Synchrotron luminosity is provided
by the vast increase in the energy of the accelerated particles.
It is likely that the acceleration mechanism is powered by the magnetic
field, that therefore gradually reduces its strength during the flare
(Dodds-Eden et al. 2009; 2010; 2011).
Indeed, a similar process is at work in magnetic reconnection that
is a fundamental process of plasmas in which magnetic energy is
converted into particle acceleration through magnetic field
rearrangement and relaxation (Begelman 1998; Lyubarski 2005;
Zweibel \& Yamada 2009; Sironi et al. 2014; 2016).
The observed drop of the magnetic field, right at the peak of the
flare, is in line with the predictions of magnetic reconnection models.
We conclude that the data are consistent with such an evolution
of the magnetic field.
\subsubsection{Constraints on energy power and source size}
\label{Ssize}
{\it Is the energy stored in the magnetic field enough to power the flare?}
We estimated the total energy emitted during the VB3 flare by considering
that the NIR and X-ray luminosity was at a level of $\sim10$~mJy and
$\sim2\times10^{35}$ erg s$^{-1}$ for about $\sim3.8\times10^3$~s and
$\sim1.8\times10^3$~s, respectively, resulting in a total emitted energy of
$\sim8\times10^{38}$~erg. If we discharge the magnetic energy within
a spherical region with $\sim1.5$~$R_S$ radius, bringing its magnetic
field from $B\sim30-40$~G to $\sim5$~G, then about
$\sim8-15\times10^{38}$~erg are produced. This appears to be
enough energy to power the VB3 flare. Moreover, this suggests that the
source of VB3 had a size $\geq1.5$~$R_S$.
\section{Conclusions}
\label{conclusions}
\begin{itemize}
\item{} The mean X-ray photon index during the very bright
flare VB3 is significantly steeper ($\Gamma_X=2.27\pm0.12$)
than the simultaneous ($\Gamma_{NIR}=1.7\pm0.1$) NIR one,
excluding that the radiative process can be described by a simple
power-law. In particular, the observed steepening
($\Delta\Gamma=0.57\pm0.09$ at 1 $\sigma$)
is consistent with what is expected by Synchrotron emission with
a cooling break ($\Delta\Gamma=0.5$).
\item{} We observe bright $F_{2.2\mu m}=8.9\pm0.1$~mJy and
hard NIR ($\Gamma_{NIR}=1.48\pm0.23$) emission about $\sim10^3$~s
before the start of the X-ray flare. We also observe very steep
X-ray emission ($\Gamma_X=3.2\pm0.4$) at the start of the X-ray flare,
while the contemporaneous NIR photon index was $\Gamma_{NIR}=1.4\pm0.2$.
These results strongly support a scenario where the synchrotron
emitting electron power-law distribution has a cut-off ($\gamma_{max}$)
that is {\it slowly} evolving with time, therefore inducing an evolving
high energy cut-off in the spectrum.
\item{} The data are consistent with an evolution of the magnetic field
strength during the flare (under the assumption of a constant escape time).
Large magnetic field amplitudes ($B=30\pm8$~G)
are observed at the start of the X-ray flare. The magnetic field strength
drops to $B=4.8\pm1.7$~G, at the peak of the X-ray flare, a variation
of a factor of $>6$ in less than $\sim650$~s. It then increases again
in the decreasing flank of the flare ($B=14.3^{+12.3}_{-7.0}$~G).
This is consistent with a scenario where the process that
accelerates the electrons producing the Synchrotron emission
is tapping energy from the magnetic field (such as, e.g. in magnetic
reconnection).
\item{} From the total emitted energy and the variation of the
magnetic field, we estimated that the source size of the VB3 flare
has to be larger than $\geq1.5$~$R_S$, if powered
by magnetic reconnection.
\item{} We observe hints for steeper, by roughly $\Delta\Gamma=0.3$,
X-ray spectra during the rise and the decay of an X-ray flare, compared
to the values at peak. This indicates that, despite the fact that the
photon index is similar between different X-ray flares, there might be
significant spectral evolution during each X-ray flare. This is an
expectation of the PLCoolEv model.
\item{} Bright and very bright {\it XMM-Newton}\ and {\it Chandra}\ flares typically last
significantly ($\sim4.4\sigma$ significance) longer, by $\sim300$~s at
soft X-ray energies, compared to harder ones (2-4 and 6-10~keV, respectively).
This trend appears to join smoothly to the longer duration typically observed
in the NIR band. Again, this is most probably the product of the evolution
of $\gamma_{max}$.
\item{} The three very bright flares, caught by {\it XMM-Newton}\ so far, have very similar
light curves and spectral properties, indicating an analogous physical origin.
This suggests that the results of this study on VB3 could be universal to bright
and very bright flares.
\item{} The best fit column density of neutral absorbing material observed
during the X-ray spectra of the very bright flares of Sgr~A$^\star$\ is constant and
it is consistent with the values observed in nearby sources.
Indeed, the three bright transients within $d_{\rm pro}<1.5^\prime$ from
Sgr~A$^\star$\ (SGR J1745-2900, Swift~J174540.7-290015\ and the foreground component towards AX~J1745.6-2901) show neutral
absorption column densities consistent with the value of Sgr~A$^\star$\
(Coti-Zelati et al. 2015; Ponti et al. 2016a; 2016). This suggests that
the neutral absorption towards Sgr~A$^\star$\ has an ISM origin.
\item{} Synchrotron self Compton models can statistically reproduce the flare
emission and its evolution. On the other hand, they imply unrealistic parameters.
In such a scenario it would
be an unlikely coincidence that the NIR photon index is flatter than the X-ray
one by $\Delta\Gamma=0.5$. Moreover, the evolution of the density,
source radius and magnetic field before, during and after the very bright flare
appears improbable.
\end{itemize}
\section*{Acknowledgments}
The authors wish to thank Jan-Uwe Ness, Ignacio de la Calle, Karl Foster
and the rest of the {\it XMM-Newton}\ and {\it NuSTAR}\ scheduling teams for the enormous
support that made this multi-wavelength campaign possible, as well as the
referee for the careful reading of the paper. GP thanks Lorenzo Sironi,
Hendrik J. van Eerten, Michi Baub\"{o}ck and Francesco Coti-Zelati,
for useful discussion. RT and AG acknowledge support from CNES.
This research has made use both of data obtained with {\it XMM-Newton}, an ESA
science mission with instruments and contributions directly funded by ESA
Member States and NASA, and on data obtained from the Chandra Data Archive.
The GC {\it XMM-Newton}\ monitoring project is supported by the Bundesministerium
f\"{u}r Wirtschaft und Technologie/Deutsches Zentrum f\"{u}r Luft- und Raumfahrt
(BMWI/DLR, FKZ 50 OR 1408 and FKZ 50 OR 1604) and the Max Planck Society.
|
\section{Introduction}
\label{intro}
3D skeleton data records the trajectories of human skeleton joints and is robust to illumination changes and invariant to camera views \cite{han2016space}.
With the prevalence of highly-accurate and affordable devices, action recognition based on
3D skeleton sequence has been attracting increasing attention \cite{xia2012view, vemulapalli2014human, du2015hierarchical, Shahroudy_2016_CVPR, zhu2016co, liu2016spatio, koniusz2016tensor, wang2016action, ke2017skeletonnet}. In this paper, we focus on
skeleton-based 3D action recognition.
To recognize a video action, the temporal information of the sequence
needs to be exploited to understand the dynamics of human postures \cite{niebles2010modeling, gaidon2013temporal, wang2014latent,
fernando2015modeling, ke2016human}. For skeleton data, the spatial structure of the human skeleton is also an important clue for action recognition \cite{zhu2016co}. Each skeleton sequence provides only the trajectory of human skeleton
joints. The time series of the joints can be used in
recurrent neural networks (RNNs) with Long-Short
Term Memory (LSTM) neurons \cite{graves2012neural, graves2013speech} to explore the spatial structure and temporal structure of the skeleton sequence for action recognition \cite{du2015hierarchical, veeriah2015differential, zhu2016co, Shahroudy_2016_CVPR, liu2016spatio}. Although LSTM networks are designed to explore the long-term temporal dependency problem, it is still difficult for LSTM to memorize the information
of the entire sequence with many timesteps \cite{weston2014memory, gu2016recurrent}. In addition, it is also difficult to construct deep LSTM to extract
high-level features \cite{sainath2015convolutional, pascanu2013construct}.
Convolutional neural networks (CNNs) \cite{lecun1995convolutional}
nowadays have achieved great success in
image classification \cite{chatfield2014return, ciregan2012multi, krizhevsky2012imagenet, simonyan2014very, szegedy2015going, xiong2015recognize, ke2014rotation}. However, for video action recognition, it lacks the capacity to
model the long-term temporal dependency of the entire video \cite{wang2016temporal}.
In this paper, instead of directly exploring the long-term temporal information from the skeleton sequences, we first represent the skeleton sequences as clips consisting of only a few frames. With the generated clips, the long-term temporal structure of the skeleton sequence can be effectively learned by using deep CNNs to process the frame images of the generated clips. In addition, the spatial structural information of the human skeleton can be exploited from the entire clips.
More specifically, for each skeleton sequence, we generate three clips corresponding to the three channels of the
cylindrical coordinates of the skeleton sequence. Each clip consists of four frames, which are generated by computing the relative positions
of the joints to four reference joints.
Each frame of the clips describes the temporal information of the entire skeleton sequence, and includes one particular spatial relationship between the joints. The entire clips aggregate multiple frames with different spatial relationships, providing important information of the spatial structure of the skeleton joints.
Since the temporal information of a skeleton sequence is incorporated in the frames of the generated clips,
the long-term temporal structure of the skeleton sequence can be learned by extracting features from the frames
of the generated clips. More specifically, each frame of the generated clips is fed to a deep CNN to extract a CNN feature.
Then the three CNN features of the three clips at the same time-step (See Figure \ref{cnn1}) are concatenated into one feature vector.
Consequently, four feature vectors are extracted from all the time-steps.
Each feature vector represents the temporal information of the skeleton sequence and one particular spatial relationship between the joints.
The feature vectors of different time-steps represent different spatial relationships with intrinsic
relationships among them.
This paper proposes to utilize the intrinsic relationships among different feature vectors for action recognition using a Multi-Task Learning Network (MTLN).
Multi-task learning
aims at improving
the generalization performance by jointly training multiple related tasks and utilizing their intrinsic relationships \cite{caruana1998multitask}.
In the proposed MTLN, the classification of each feature vector is treated as a separate task, and the MTLN jointly learns multiple classifiers each from one feature vector and outputs multiple predictions, each corresponding to one task.
All the feature vectors of the same skeleton sequence have the same label as the skeleton sequence.
During training, the loss value of each task is individually computed using its own class scores. Then the loss values of all tasks are summed up to define the total loss of the network which is then used to learn the network parameters.
During testing, the class scores of all tasks are averaged to form the final prediction of the action class. Multi-task learning simultaneously solves multiple tasks with weight sharing, which can improve the performance
of individual tasks \cite{caruana1998multitask}.
The main contributions of this paper are summarized as follows. \textbf{(1)} We propose to transform each skeleton sequence to a new representation, \ie, three clips, to allow global long-term temporal modelling of the skeleton sequence by using deep CNNs to learn hierarchical features from frame images.
\textbf{(2)} We introduce a MTLN to process all the CNN features of the frames in the generated clips, thus to learn the spatial structure and the temporal information of the skeleton sequence. The MTLN improves the performance by utilizing intrinsic relationships among different frames of the generated clips.
Our experimental results demonstrate that MTLN performs better than concatenating or pooling the features of the frames (See Section \ref{exp}). \textbf{(3)} The proposed method achieves the state-of-the-art performance on three skeleton datasets, including the large scale NTU RGB+D dataset \cite{Shahroudy_2016_CVPR}.
\begin{figure*}
\vspace{-1mm}
\begin{center}
\includegraphics[width=6.5in]{cnn2new1.pdf}
\end{center}
\vspace{-1mm}
\caption{Architecture of the proposed method. Given a skeleton sequence (a), three clips (b) corresponding to the three channels of the cylindrical coordinates are generated.
A deep CNN model (c) and a temporal mean pooling (TMP) layer (d) are used to extract a compact representation from each frame of the clips (see Figure \ref{cnn2} for details).
The output CNN representations of the three clips at the same time-step are concatenated, resulting four feature vectors (e).
Each feature vector represents the temporal information of the skeleton sequence and a particular spatial relationship of the skeleton joints.
The proposed MTLN (f) which includes a fully connected (FC) layer, a rectified linear unit (ReLU), another FC layer and a Softmax layer jointly processes the four feature vectors in parallel and outputs four sets of class scores (g), each corresponding to one task of classification using one feature vector.
During training, the loss values of the four tasks are summed up to define the loss value of the network used to update the network parameters. For testing, the class scores of the four tasks are averaged to generate the final prediction of the action class. }
\label{cnn1}
\vspace{-1mm}
\end{figure*}
\section{Related Works}
In this section, we cover the relevant literature of skeleton-based action recognition methods using hand-crafted features
or using deep learning networks.
\textbf{Hand-crafted Features}~~
In \cite{hussein2013human},
the covariance matrices of the trajectories of the joint positions are computed over hierarchical temporal levels to model the skeleton sequences. In \cite{wang2012mining}, the pairwise relative positions of each joint with other joints are computed to represent each frame of the skeleton sequences, and Fourier Temporal Pyramid (FTP) is used to model the temporal patterns. In \cite{yang2012eigenjoints}, the pairwise relative positions of the joints are also used to characterize posture features, motion features, and
offset features of the skeleton sequences. Principal Component Analysis (PCA) is then applied to the normalized features to compute EigenJoints as representations. In \cite{xia2012view}, histograms of 3D joint locations are computed to represent each frame of the skeleton sequences, and HMMs are used to model the temporal dynamics. In \cite{vemulapalli2014human}, the rotations and translations between various body parts are used as representations, and
a skeleton sequence is modelled as a curve in the Lie group. The temporal dynamics are modelled with FTP.
\textbf{Deep Learning Methods }~~
In \cite{du2015hierarchical}, the skeleton joints are divided
into five sets corresponding to five body parts. They are fed into five LSTMs for feature fusion and classification.
In \cite{zhu2016co}, the skeleton joints are fed to a deep LSTM
at each time slot to learn
the inherent co-occurrence features of skeleton joints.
In \cite{Shahroudy_2016_CVPR}, the long-term context representations of the body parts are learned with a part-aware LSTM.
In \cite{liu2016spatio}, both the spatial and temporal information of skeleton sequences are learned
with a spatial temporal LSTM. A Trust Gate is also proposed to remove noisy joints. This method achieves the state-of-the-art performance on the NTU RGB+D dataset \cite{Shahroudy_2016_CVPR}.
\section{Proposed Method}
An overall architecture of the proposed method is shown in Figure \ref{cnn1}.
The proposed method starts by generating clips of skeleton sequences.
A skeleton sequence of any length is
transformed into three clips each consisting of several gray images. The generated clips are then fed to a deep CNN model to extract CNN features which are used in a MTLN for action recognition.
\subsection{Clip Generation}
\label{imagegene}
\begin{figure} [h!]
\vspace{-3mm}
\begin{center}
\includegraphics[width=3.2in]{visual21.pdf}
\end{center}
\vspace{-3mm}
\caption{Clip Generation of a skeleton sequence. The skeleton joints of each frame are first arranged as a chain by concatenating the joints of each body part (\ie, 1-2-3-...-16). Four reference joints shown in green (\ie, left shoulder 5, right shoulder 8, left hip 11 and right hip 14) are then respectively used to compute relative positions of the other joints to incorporate different spatial relationships between the joints. Consequently, four 2D arrays are obtained by combining the relative positions of all the frames of the skeleton sequence. The relative position of each joint in the 2D arrays is described with cylindrical coordinates. The four 2D arrays corresponding to the same channel of the coordinates are transformed to four gray images and as a clip. Thus three clips are generated from the three channels of the cylindrical coordinates of the four 2D arrays.}
\label{clip1}
\vspace{-1mm}
\end{figure}
Compared to RGB videos which consist of multiple frame images, skeleton sequences only provide the trajectories of the 3D coordinates. This paper proposes to transform the original skeleton sequence to a collection of clips each consisting of several images, thus to allow spatial temporal feature learning using deep neural networks. Intuitively, one could represent the content of each frame of the skeleton sequence as an image to generate a video.
However, if the skeleton sequence has many frames, this method will result in a long video of which the temporal dynamics will be difficult to learn. In addition,
each frame of the generated video will also be very sparse as the number of the skeleton joints is small.
To overcome this problem, we propose to represent the temporal dynamics of the skeleton sequence in a frame image, and then use multiple frames to incorporate different spatial relationships between the joints. An advantage of this method is that for any skeleton sequence of any length, the generated clips contain the same number of frames and the long-term temporal information of the original skeleton sequence can be effectively captured with the powerful CNN representations of the frame images in the generated clips.
As shown in Figure \ref{clip1}, for a skeleton sequence, the skeleton joints of each frame are first arranged as a chain by concatenating the joints of each body part.
Considering that the relative positions between joints provide more useful information than their absolute locations (\eg, the relative location of the hand to the shoulder in ``pushing''), four reference joints,
namely, the left shoulder, the right shoulder, the left hip and the right hip, are respectively
used to compute relative positions of the other joints, thus to incorporate different spatial relationships between joints and provide useful structural information of the skeleton.
These four joints are selected as reference joints due to the fact that they are stable in most actions. They can thus reflect the motions of the other joints. Although the base of the spine is also stable, it is close to the left hip and the right hip. It is therefore discarded to avoid information redundancy. By combing the relative joints of all the frames, four 2D arrays with dimension $(m-1)\times t$ are generated ($m $ is the number of skeleton joints in each frame and $t$ is the number of frames of the skeleton sequence).
The relative positions of joints in the 2D arrays are originally described with 3D Cartesian coordinates. Considering that the cylindrical coordinates are more useful to analyse the motions as each human body utilizes pivotal joint movements to perform an action, the 3D Cartesian coordinates are transformed to cylindrical coordinates in the proposed representation of skeleton sequences. The cylindrical coordinates have been used to extract view-invariant motion features for action recognition in \cite{weinland2006free}.
The four 2D arrays corresponding to the same channel of the 3D cylindrical coordinates are transformed to four gray images by scaling the coordinate values between 0 to 255 using a linear transformation. A clip is then constructed with the four gray images. Consequently, three clips are generated from the three channels of the 3D coordinates of the four 2D arrays.
\subsection{Clip Learning}
Each frame of the generated clips describes the temporal dynamics of all frames of the skeleton sequence and one particular spatial relationship between the skeleton joints in one channel of the cylindrical coordinates.
Different frames of the generated clip describe different spatial relationships and
there exists intrinsic
relationships among them.
A deep CNN is first leveraged to extract a compact representation from each frame of the generated clips to exploit the long-term temporal information of the skeleton sequence.
Then the CNN features of all frames of the generated clips are jointly processed in parallel using multi-task learning, thus to utilize their intrinsic relationships to learn the spatial temporal information for 3D action recognition.
\subsubsection{Temporal Pooling of CNN Feature Maps}
\label{cnnrep}
To learn the features of the generated clips, a deep CNN is firstly employed to extract a compact representation of each frame of the clips.
Since each frame describes the temporal dynamics of the skeleton sequence,
the spatial invariant CNN feature of each frame could thus represent the robust temporal information of the skeleton sequence.
Given the generated clips, the CNN feature of each frame is extracted with the pre-trained VGG19 \cite{simonyan2014very} model.
The pre-trained CNN model is leveraged as a feature extractor due to the fact that the CNN features extracted by the models pre-trained with ImageNet \cite{russakovsky2015imagenet} are very powerful and have been successfully applied in a number of cross-domain applications \cite{donahue2014decaf,girshick2014rich,sharif2014cnn,han2015matchnet}. In addition,
current skeleton datasets are either too small or too noisy to suitably train a deep network.
Although the frames of the generated clips are not natural images, they could still be fed to the CNN model pre-trained with ImageNet \cite{russakovsky2015imagenet} for feature extraction.
The similarity between a natural image and the generated frames is that both of them are matrices with some patterns. The CNN models trained on the large image dataset can be used as a feature extractor to extract representations of the patterns in matrices. The learned representations are generic and can be transferred to novel tasks from the original tasks \cite{yosinski2014transferable, long2015learning}.
\begin{figure}
\vspace{-1mm}
\begin{center}
\includegraphics[width=3.2in]{cnncmb2.pdf}
\end{center}
\vspace{-1mm}
\caption{\small{Temporal mean pooling of the CNN feature maps. (a) An input frame of the generated clips, for which the rows correspond to the different frames of the skeleton sequence and the columns correspond to the different vectors generated from the joints. (b) Output feature maps of the conv5\_1 layer. The size is $14\times 14 \times 512$. Each activation (shown in red) of the feature map is a feature correspond to the local region of the original image (shown with a red square).
(c) Temporal features of all joints of the skeleton sequence, which are obtained by applying mean pooling to each feature map in the row (temporal) dimension.
(d) Output feature, which is achieved by concatenating all the feature maps in (c).
}}
\label{cnn2}
\vspace{-1mm}
\end{figure}
The pre-trained VGG19 \cite{simonyan2014very} model contains 5 sets of convolutional layers conv1, conv2,
..., conv5. Each set includes a stack of 2 or 4 convolutional layers with the same kernel size.
Totally there are 16 convolutional layers and
three fully connected layers in the network.
Although deep neural networks are capable of learning powerful and
generic features which can be used in other novel domains, the features extracted from the different layers have different transferability. Particularly, the features in earlier layers are more generic, while in later layers, the features are more task-specific, which largely rely on the original classes and dataset. The features of the later layers are thus less suitable than those of the earlier layers to transfer to other domains \cite{yosinski2014transferable, long2015learning}.
Therefore,
this paper adopts a compact representation that is derived from the activations of the convolutional layer to exploit the temporal information of a skeleton sequence. The feature maps in the convolutional layer have been successfully applied for action recognition and image retrieval \cite{peng2015encoding, radenovic2016cnn}.
Specifically, the last 3 convolutional layers and fully connected layers of the network are discarded.
Each frame image of the three clips is scaled to $224 \times 224$, and is then duplicated three times to formulate a color image, so that it can be fed to the network.
The output of the convolutional layer conv5\_1 is used as the representation of the input frame,
which is a 3D tensor with size $14\times 14\times 512$,
\ie, 512 feature maps with size $14\times 14$.
The rows of the generated frame correspond to different frames of a skeleton sequence. The dynamics of the row features of the generated image therefore represents the temporal evolution of the skeleton sequence.
Meanwhile, the activations of each feature map in the conv5\_1 layer are the local features corresponding to the local regions in the original input image \cite{peng2015encoding}.
The temporal information of the sequence can thus be extracted from the row features of the feature maps.
More specifically, the feature maps are processed with temporal mean pooling with kernel size $14\times 1$, \ie, the pooling is applied over the temporal, or row dimension, thus to generate a compact fusion representation from all temporal stages of the skeleton sequence.
Let the activation at the $i^{th}$ row and the $j^{th}$ column of the $k^{th}$ feature map be $x_{i,j}^k$. After temporal mean pooling, the output of the $k^{th}$ feature map is given by:
\begin{equation}
\begin{array}{c}
\mathbf{y^k}=\left[y_1^k,\cdots, y_j^k, \cdots,y_{14}^k\right] \\
\\
y_j^k= \dfrac{1}{14}\sum\limits_{i=1}^{14}\max(0,x_{i,j}^k)\\
\end{array}
\end{equation}
The outputs of all feature maps (512) are concatenated to form a 7168D ($14\times512=7168$) feature vector, which represents the temporal dynamics of the skeleton sequence in one channel of the cylindrical coordinates.
\subsubsection{Multi-Task Learning Network (MTLN)}
As shown in Figure \ref{cnn1}(e), the three 7168D features of the three clips at the same time-step are concatenated to form a feature vector, generating four feature vectors in total. Each feature vector represents the temporal dynamics of the skeleton sequence and includes one particular spatial relationship between the joints in one of three cylindrical coordinates.
The four feature vectors have intrinsic relationships between each other.
An MTLN is then proposed to jointly process the four feature vectors to utilize their intrinsic relationships for action recognition.
The classification of each feature vector is treated as a separate task with the same classification label of the skeleton sequence.
The architecture of the network is shown in Figure \ref{cnn1}(f).
It includes two fully connected (FC) layers and a Softmax layer.
Between the two FC layers there is a rectified linear unit (ReLU) \cite{nair2010rectified} to introduce an additional non-linearity.
Given the four features as inputs, the MTLN generates four frame-level predictions, each corresponding to one task. During training, the class scores of each task are used to compute a loss value. Then the loss values of all tasks are summed up to generate the final loss of the network used to learn the network parameters.
During testing, the class scores of all tasks are averaged to form the final prediction of the action class. The loss value of the $k^{th}$ task ($k=1,\cdots,4$) is given by Equation \ref{eql1}.
\begin{equation}
\label{eql1}
\begin{array}{ll}
\ell_k(\mathbf{z_k},\mathbf{y})&=\sum\limits_{i=1}^{m} y_i \left(-log\left(\dfrac{\exp z_{ki}}{\sum\limits_{j=1}^m \exp z_{kj}}\right)\right) \\
&=\sum\limits_{i=1}^{m} y_i \left(log\left(\sum\limits_{j=1}^m \exp z_{kj} \right)-z_{ki}\right)
\end{array}
\end{equation}
where $\mathbf{z_k}$ is the vector fed to the Softmax layer generated from the $k^{th}$ input feature, $m$ is the number of action classes and $y_i$ is the ground-truth label for class $i$.
The final loss value of the network is computed as the sum of the
four individual losses, as shown below in Equation \ref{eql2}:
\begin{equation}
\label{eql2}
\mathcal{L}(Z,\mathbf{y})=\sum\limits_{k=1}^{4} \ell_k (\mathbf{z_k},\mathbf{y})
\end{equation}
where $Z=[\mathbf{z_1},\cdots,\mathbf{z_4}]$.
\section{Experiments and Analysis}
The proposed method is tested on three skeleton action datasets: NTU RGB+D dataset \cite{Shahroudy_2016_CVPR}, SBU kinect interaction dataset \cite{yun2012two} and CMU dataset \cite{cmu2013}.
The main ideas of the proposed method \textbf{Clips + CNN + MTLN } are 1) generating three clips (each clip consists of four frames) from a skeleton sequence, 2) using CNNs to learn global long-term temporal information of the skeleton sequence from each frame of the generated clips, and 3) using MTLN to
jointly train the CNN features of the four frames of the clips to incorporate the spatial structural information for action recognition.
We also conducted the following baselines to demonstrate the advantages of the proposed method:
\textbf{Coordinates + FTP}~~In this baseline, the Fourier Temporal Pyramid (FTP) \cite{wang2012mining} is applied to the
3D coordinates of the skeleton sequences to extract temporal features for action recognition. This baseline
is used to show the benefits of using CNNs for long-term temporal modelling of the skeleton sequences.
\textbf{Frames + CNN}~~In this baseline, the CNN features of single frames instead of the entire generated clips are used for action recognition. In other words, only one feature vector shown in Figure \ref{cnn1}(e) is used to train a neural network for classification.
Thus the loss value of the network is given by Equation \ref{eql1}.
The average accuracy of the four features is provided. This baseline is
used to show the benefits of using the entire generated clips to incorporate the spatial structural information for action recognition.
\textbf{Clips + CNN + Concatenation}~~In this baseline, the CNN features of all frames of the generated clips are concatenated before performing action recognition. In other words, the four feature vectors shown in Figure \ref{cnn1}(e) are concatenated and then fed to a neural network for classification.
This baseline is used to show the benefits of using MTLN to process the features of the entire clips in parallel.
\textbf{Clips + CNN + Pooling}~~In this baseline, max pooling is applied to the CNN features of all frames of the generate clips before performing action recognition.
Same as Clips + CNN + Concatenation, this baseline is also used to show the benefits of using MTLN.
\subsection{Datasets}
\textbf{NTU RGB+D Dataset} \cite{Shahroudy_2016_CVPR}~~
To the best of our knowledge, this dataset is so far the largest skeleton-based human action dataset, with more than 56000 sequences and 4 million frames. There are 60 classes of actions performed by 40 distinct subjects, including both one-person daily actions (e.g., clapping,
reading,
writing) and two-person interactions (\eg, handshaking, hug, pointing).
These actions are captured by three cameras, which are placed at different locations and view points. In total, there are 80 views for this dataset. In this dataset, each skeleton has 25 joints. The
3D coordinates of the joints are provided.
Due to the large view point, intra-class and sequence length variations, the dataset is very challenging.
\textbf{SBU Kinect Interaction Dataset} \cite{yun2012two}~~
This dataset was collected using the Microsoft Kinect sensor. It contains 282 skeleton sequences and
6822 frames. In this dataset, each frame contains two persons performing an interaction. The interactions include
approaching, departing, kicking, punching, pushing, hugging, shaking hands and exchanging. There are 15 joints for each skeleton. This dataset is challenging due to the fact that the joint coordinates exhibit low accuracy \cite{yun2012two}.
\textbf{CMU Dataset} \cite{cmu2013}~~
This dataset contains 2235 sequences and about 1 million frames. For each skeleton, the 3D coordinates of 31 joints are provided.
The dataset has been categorized into 45 classes \cite{zhu2016co}.
All of the actions are performed by only one person. The dataset is very challenging due to the large sequence length variations and intra-class diversity.
\subsection{Implementation Details}
For all datasets, the clips are generated with all frames of the original skeleton sequence without
any pre-processing such as normalization, temporal down-sampling or noise filtering.
The proposed method was implemented using the MatConvNet
toolbox \cite{vedaldi2015matconvnet}.
The number of the hidden unit of the first FC layer is set to 512.
For the second FC layer (\ie, the output layer), the number of the unit is the same as the number of the action classes in each dataset.
The network is trained using the stochastic gradient
descent algorithm. The learning rate is set
to 0.001 and batch size is set to 100. The training is stopped
after 35 epochs. The performance of the proposed method on
each dataset is compared with existing methods using the
same testing protocol.
\subsection{Results}
\label{exp}
\textbf{NTU RGB+D Dataset}~~As in \cite{Shahroudy_2016_CVPR}, the evaluation on this dataset is performed with two standard protocols, \ie, cross-subject evaluation and cross-view evaluation.
In cross-subject evaluation, the sequences of 20 subjects are used for training and the data from 20 other subjects are used for testing.
In cross-view evaluation, the sequences captured by two cameras are used for training and the rest are used for testing.
The results are shown in Table \ref{ntu}. It can be seen that the proposed method performs significantly better than others in both cross-subject and cross-view protocols.
The accuracy of the proposed method is 79.57\% when tested with the cross-subject protocol. Compared to the previous state-of-the-art method (ST-LSTM + Trust Gate \cite{liu2016spatio}), the performance is improved by 10.37\%.
When tested with the cross-view protocol, the accuracy is improved from 77.7\% to 84.83\%.
The improved performance of the proposed method is due to the novel clip representation and feature learning method.
As shown in Table \ref{ntu},
Frames + CNN
achieves an accuracy of about 75.73\% and 79.62\% for the two testing protocols, respectively.
The performances are much better than Coordinates + FTP.
Compared to extracting temporal features of skeleton sequences with FTP and native 3D coordinates,
using CNN to learn the temporal information of skeleton sequences from the generated frames is more robust to noise and temporal variations due to the convolution and pooling operators, resulting in better performances.
From Table \ref{ntu}, it can also be seen that
Frames + CNN also performs better than the previous state-of-the-art method. It clearly shows the effectiveness of the CNN features of the proposed clip representation. The performances are improved by learning entire clips with CNN and MTLN (\ie, Clips + CNN + MTLN). The improvements are about 4\% and 5\% for the two testing protocols, respectively.
It can also be seen that the proposed MTLN (\ie, Clips + CNN + MTLN) performs better than feature concatenation (\ie, Clips + CNN + concatenation) and pooling (\ie, Clips + CNN + pooling).
Frames + CNN, Clips + CNN + concatenation and Clips + CNN + pooling can be viewed as a single-task method, while using MTLN to process multiple frames of the generated clips in parallel utilizes their intrinsic relationships and incorporates the spatial structural information, which improves the performance of the single-task method for action recognition.
\begin{table}[]
\centering
\tabcolsep=0.05cm
\caption {Performance on the NTU RGB+D dataset.}
\begin{tabular}{c|cc}
\hline
\multirow{2}{*}{Methods} & \multicolumn{2}{c}{Accuracy}\\
&Cross Subject&Cross View\\
\hline
Lie Group \cite{vemulapalli2014human}& 50.1\%&52.8\%\\
Skeletal Quads \cite{evangelidis2014skeletal}&38.6\%&41.4\%\\
Dynamic Skeletons \cite{hu2015jointly}&60.2\%&65.2\%\\
Hierarchical RNN \cite{du2015hierarchical}&59.1\%&64.0\%\\
Deep RNN \cite{Shahroudy_2016_CVPR}&59.3\%&64.1\%\\
Deep LSTM \cite{Shahroudy_2016_CVPR}&60.7\%&67.3\%\\
Part-aware LSTM \cite{Shahroudy_2016_CVPR}&62.9\%&70.3\%\\
ST-LSTM \cite{liu2016spatio} & 65.2\%&76.1\%\\
ST-LSTM + Trust Gate \cite{liu2016spatio} & 69.2\%&77.7\%\\
\hline
Coordinates + FTP&61.06\% &74.64\%\\
Frames + CNN & 75.73\%&79.62\%\\
Clips + CNN + Concatenation & 77.05\%&81.11\%\\
Clips + CNN + Pooling & 76.37\%&80.46\%\\
Clips + CNN + MTLN & \textbf{79.57\%}& \textbf{84.83\%}\\
\hline
\end{tabular}
\vspace{-3mm}
\label{ntu}
\end{table}
\textbf{SBU Kinect Interaction Dataset}~~
As in \cite{yun2012two}, the evaluation of this dataset is a 5-fold cross validation, with the provided training/testing splits. Each frame of the skeleton sequences contains two separate human skeletons. In this case,
the two skeletons are considered as two data samples and the clip generation and feature extraction are conducted separately for the two skeletons. For testing, the prediction of actions is obtained by averaging the classification scores of the two samples.
Considering that the number of samples in this dataset is too small, data augmentation is performed to increase the number of samples. More specifically, each frame image of the generated clips are resized to $250\times250$, and then random patches with size of $224\times224$ are cropped from the original image for feature learning using CNN. For this dataset, 20 sub-images are cropped and the total data samples are extended to 11320.
The comparisons of the proposed method with other methods are shown in Table \ref{sbu}. Similar to the NTU RGB+D dataset, CNN features perform better than FTP to learn the temporal information.
It can be seen that when using CNN features of individual frames, the accuracy is 90.88\%, which is similar to the Deep LSTM + Co-occurrence method \cite{zhu2016co}. When incorporating the CNN features of the entire clips using concatenation and pooling methods, the performance is improved by about 2\%.
The performance is improved to 93.57\% when learning the entire clips with MTLN. It clearly shows the benefit
of using MTLN to learn the CNN features entire clips.
Since the joint positions of this dataset are not very accurate \cite{yun2012two}, existing methods including HBRNN \cite{du2015hierarchical} and Co-occurrence LSTM \cite{zhu2016co} remove the joint noise by smoothing the position of each joint using the Svaitzky-Golay filter \cite{savitzky1964smoothing}. In \cite{liu2016spatio}, a Trust Gate is introduced to remove the noisy joints and this improves the accuracy from 88.6\% to 93.3\%.
Our method does not perform any pre-processing to handle the noisy joints, but still performs better than all the others. It clearly shows that the features learned from the generated clips are robust to noise due to the convolution and pooling operators of the deep network.
\begin{table}[]
\centering
\tabcolsep=0.1cm
\caption {Performance on the SBU kinect interaction dataset.}
\begin{tabular}{c|c}
\hline
Methods&Accuracy\\
\hline
\pbox{3cm}
Raw Skeleton \cite{yun2012two} & 49.7\%\\
Joint Feature \cite{ji2014interactive}&86.9\%\\
CHARM \cite{li2015category} &83.9\%\\
Hierarchical RNN \cite{du2015hierarchical}&80.35\%\\
Deep LSTM \cite{zhu2016co}&86.03\%\\
Deep LSTM + Co-occurrence \cite{zhu2016co}&90.41\%\\
ST-LSTM \cite{liu2016spatio} &88.6\%\\
ST-LSTM + Trust Gate \cite{liu2016spatio} & 93.3\%\\
\hline
Coordinates + FTP& 79.75\%\\
Frames + CNN & 90.88\%\\
Clips + CNN + Concatenation &92.86\% \\
Clips + CNN + Pooling &92.26\%\\
Clips + CNN + MTLN & \textbf{93.57\%} \\
\hline
\end{tabular}
\label{sbu}
\end{table}
\begin{table}[t]
\centering
\tabcolsep=0.05cm
\caption {Performance on the CMU dataset.}
\begin{tabular}{c|cc}
\hline
\multirow{2}{*}{Methods} & \multicolumn{2}{c}{Accuracy}\\
&CMU subset&CMU\\
\hline
Hierarchical RNN \cite{du2015hierarchical}&83.13\%&75.02\%\\
Deep LSTM \cite{zhu2016co}&86.00\%&79.53\%\\
Deep LSTM + Co-occurrence \cite{zhu2016co}&88.40\%&81.04\%\\
\hline
Coordinates + FTP &83.44\% &73.61\%\\
Frames + CNN & 91.53\% &85.36\%\\
Clips + CNN + Concatenation &90.97\% &85.76\%\\
Clips + CNN + Pooling &90.66\%& 85.56\%\\
Clips + CNN+ MTLN & \textbf{93.22\%}& \textbf{88.30\%}\\
\hline
\end{tabular}
\label{cmu1}
\end{table}
\textbf{CMU Dataset}~~As in \cite{zhu2016co}, for this dataset, the evaluation is conducted on both the entire dataset with 2235 sequences, and a selected subset of 664 sequences. The subset includes 8 classes of actions, , \ie, basketball, cartwheel, getup, jump, pickup, run,
sit and walk back.
For the entire dataset, the testing protocol is 4-fold cross validation, and for the subset, it is evaluated
with 3-fold cross validation. The training/tesing splits of the different folds are provided by \cite{zhu2016co}.
Similar to the SBU kinect interaction dataset, data augmentation is also conducted on CMU dataset. For the entire dataset, each frame image is used to generate 5 more images and the total data samples are extended to 11175, and for the subset, the total samples are extended to 13280, which is 20 times of the original number.
The results are shown in Table \ref{cmu1}. It can be seen that the performance of the proposed method is much better than previous state-of-the-art methods on both the subset and the entire set. When tested on the subset, the accuracy of the proposed method was about 93.22\%, which is about 5\% better than the previous method \cite{zhu2016co}.
The performance on the entire dataset is improved from 81.04\% to 88.3\%.
\subsection{Discussions}
\textbf{Three gray clips or one color clip?}
As shown in Figure \ref{cnn1}, the frames of the three generated clips are gray images, each corresponding to only one channel of the cylindrical coordinates. Each frame is duplicated three times to formulate a color image for CNN feature learning. The output CNN features of the three channels are concatenated in a feature vector for action recognition. A simple alternative is to generate a color clip with three channels of the cylindrical coordinates, and then extract a single CNN feature from the color frame for action recognition. When this was tested on CMU dataset, the performance is 84.67\%, which is about 4\% worse than the proposed method. This is perhaps due to the fact that the relationship of the three generated channels is different from that of the RGB channels of natural color images. The
RGB channels are arranged in sequence and there is no
matching order between 3D coordinates and RGB channels.
\textbf{The more frames, the better performance?}~~
This paper uses only four reference joints to generate clips, each having four frames. When 6 more joints are selected to generate more frames, \ie, the head, the left hand, the right hand, the left foot, the right foot and the hip, the performance does not improve. When tested on CMU data, the performance is 86.01\%, which is about 2\% worse than the proposed method. This is due to the fact that the other joints are not as stable as the selected four joints, which can introduce noise.
\textbf{Cartesian coordinates or cylindrical coordinates?}~~
As mentioned in Section \ref{imagegene}, the 3D Cartesian coordinates of the vectors between the reference joints and the other joints are transformed to cylindrical coordinates to generate clips. We found that when using the original Cartesian coordinates for clip generation and action recognition, the performance drops. When tested on CMU dataset, the accuracy is 86.21\%, which is about 2\% worse than the proposed method. The cylindrical coordinates are more useful than the Cartesian coordinates to analyse the motions as each human skeleton utilizes pivotal joint movements to perform an action.
\textbf{Features in different layers}~~
As mentioned in Section \ref{cnnrep}, the feature maps in conv5\_1 layer of the pre-trained CNN model is adopted as the representation of each input image. We found that using the features in the earlier layers
decreased the performance. When using the features of the conv4\_1 layer, the accuracy on CMU dataset is 84.59\%, which is about 4\% worse than the proposed method. This is perhaps due to the fact that the features in the earlier layers are not deep enough to capture the salient information of the input image.
We also found that using the features in the later layers
made the performance worse.
When using the features of the fc6 layer, the accuracy on CMU dataset is 83.52\%, which is about 5\% worse than the proposed method.
This is because
the features in the later layers are more task-specific, which largely rely on the original classes and dataset. The features of the later layers are thus less suitable than those of the earlier layers to transfer to other domains \cite{yosinski2014transferable, long2015learning}.
\section{Conclusion}
In this paper, we have proposed to transform a skeleton sequence to three video clips for robust feature learning and action recognition.
We proposed to use a pre-trained CNN model followed by a temporal pooling layer to extract a compact representation of each frame. The CNN features of the three clips at the same time-step are concatenated in a single feature vector,
which describes the temporal information of the entire skeleton sequence and one particular spatial relationship between the joints.
We then propose an MTLN to jointly learn the feature vectors at all the time-steps in parallel, which utilizes their intrinsic relationships and improves the performance for action recognition.
We have tested the proposed method on three datasets, including NTU RGB+D dataset, SBU kinect interaction dataset and CMU dataset. Experimental results have shown the effectiveness of the proposed new representation and feature learning method.
\section{Acknowledgment}
This work was partially supported by Australian Research
Council grants DP150100294, DP150104251, and DE120102960.
This paper used the NTU RGB+D Action Recognition Dataset made available by the ROSE Lab at the Nanyang Technological University, Singapore.
{\small
\bibliographystyle{ieee}
\section{Introduction}
\label{intro}
3D skeleton data records the trajectories of human skeleton joints and is robust to illumination changes and invariant to camera views \cite{han2016space}.
With the prevalence of highly-accurate and affordable devices, action recognition based on
3D skeleton sequence has been attracting increasing attention \cite{xia2012view, vemulapalli2014human, du2015hierarchical, Shahroudy_2016_CVPR, zhu2016co, liu2016spatio, koniusz2016tensor, wang2016action, ke2017skeletonnet}. In this paper, we focus on
skeleton-based 3D action recognition.
To recognize a video action, the temporal information of the sequence
needs to be exploited to understand the dynamics of human postures \cite{niebles2010modeling, gaidon2013temporal, wang2014latent,
fernando2015modeling, ke2016human}. For skeleton data, the spatial structure of the human skeleton is also an important clue for action recognition \cite{zhu2016co}. Each skeleton sequence provides only the trajectory of human skeleton
joints. The time series of the joints can be used in
recurrent neural networks (RNNs) with Long-Short
Term Memory (LSTM) neurons \cite{graves2012neural, graves2013speech} to explore the spatial structure and temporal structure of the skeleton sequence for action recognition \cite{du2015hierarchical, veeriah2015differential, zhu2016co, Shahroudy_2016_CVPR, liu2016spatio}. Although LSTM networks are designed to explore the long-term temporal dependency problem, it is still difficult for LSTM to memorize the information
of the entire sequence with many timesteps \cite{weston2014memory, gu2016recurrent}. In addition, it is also difficult to construct deep LSTM to extract
high-level features \cite{sainath2015convolutional, pascanu2013construct}.
Convolutional neural networks (CNNs) \cite{lecun1995convolutional}
nowadays have achieved great success in
image classification \cite{chatfield2014return, ciregan2012multi, krizhevsky2012imagenet, simonyan2014very, szegedy2015going, xiong2015recognize, ke2014rotation}. However, for video action recognition, it lacks the capacity to
model the long-term temporal dependency of the entire video \cite{wang2016temporal}.
In this paper, instead of directly exploring the long-term temporal information from the skeleton sequences, we first represent the skeleton sequences as clips consisting of only a few frames. With the generated clips, the long-term temporal structure of the skeleton sequence can be effectively learned by using deep CNNs to process the frame images of the generated clips. In addition, the spatial structural information of the human skeleton can be exploited from the entire clips.
More specifically, for each skeleton sequence, we generate three clips corresponding to the three channels of the
cylindrical coordinates of the skeleton sequence. Each clip consists of four frames, which are generated by computing the relative positions
of the joints to four reference joints.
Each frame of the clips describes the temporal information of the entire skeleton sequence, and includes one particular spatial relationship between the joints. The entire clips aggregate multiple frames with different spatial relationships, providing important information of the spatial structure of the skeleton joints.
Since the temporal information of a skeleton sequence is incorporated in the frames of the generated clips,
the long-term temporal structure of the skeleton sequence can be learned by extracting features from the frames
of the generated clips. More specifically, each frame of the generated clips is fed to a deep CNN to extract a CNN feature.
Then the three CNN features of the three clips at the same time-step (See Figure \ref{cnn1}) are concatenated into one feature vector.
Consequently, four feature vectors are extracted from all the time-steps.
Each feature vector represents the temporal information of the skeleton sequence and one particular spatial relationship between the joints.
The feature vectors of different time-steps represent different spatial relationships with intrinsic
relationships among them.
This paper proposes to utilize the intrinsic relationships among different feature vectors for action recognition using a Multi-Task Learning Network (MTLN).
Multi-task learning
aims at improving
the generalization performance by jointly training multiple related tasks and utilizing their intrinsic relationships \cite{caruana1998multitask}.
In the proposed MTLN, the classification of each feature vector is treated as a separate task, and the MTLN jointly learns multiple classifiers each from one feature vector and outputs multiple predictions, each corresponding to one task.
All the feature vectors of the same skeleton sequence have the same label as the skeleton sequence.
During training, the loss value of each task is individually computed using its own class scores. Then the loss values of all tasks are summed up to define the total loss of the network which is then used to learn the network parameters.
During testing, the class scores of all tasks are averaged to form the final prediction of the action class. Multi-task learning simultaneously solves multiple tasks with weight sharing, which can improve the performance
of individual tasks \cite{caruana1998multitask}.
The main contributions of this paper are summarized as follows. \textbf{(1)} We propose to transform each skeleton sequence to a new representation, \ie, three clips, to allow global long-term temporal modelling of the skeleton sequence by using deep CNNs to learn hierarchical features from frame images.
\textbf{(2)} We introduce a MTLN to process all the CNN features of the frames in the generated clips, thus to learn the spatial structure and the temporal information of the skeleton sequence. The MTLN improves the performance by utilizing intrinsic relationships among different frames of the generated clips.
Our experimental results demonstrate that MTLN performs better than concatenating or pooling the features of the frames (See Section \ref{exp}). \textbf{(3)} The proposed method achieves the state-of-the-art performance on three skeleton datasets, including the large scale NTU RGB+D dataset \cite{Shahroudy_2016_CVPR}.
\begin{figure*}
\vspace{-1mm}
\begin{center}
\includegraphics[width=6.5in]{cnn2new1.pdf}
\end{center}
\vspace{-1mm}
\caption{Architecture of the proposed method. Given a skeleton sequence (a), three clips (b) corresponding to the three channels of the cylindrical coordinates are generated.
A deep CNN model (c) and a temporal mean pooling (TMP) layer (d) are used to extract a compact representation from each frame of the clips (see Figure \ref{cnn2} for details).
The output CNN representations of the three clips at the same time-step are concatenated, resulting four feature vectors (e).
Each feature vector represents the temporal information of the skeleton sequence and a particular spatial relationship of the skeleton joints.
The proposed MTLN (f) which includes a fully connected (FC) layer, a rectified linear unit (ReLU), another FC layer and a Softmax layer jointly processes the four feature vectors in parallel and outputs four sets of class scores (g), each corresponding to one task of classification using one feature vector.
During training, the loss values of the four tasks are summed up to define the loss value of the network used to update the network parameters. For testing, the class scores of the four tasks are averaged to generate the final prediction of the action class. }
\label{cnn1}
\vspace{-1mm}
\end{figure*}
\section{Related Works}
In this section, we cover the relevant literature of skeleton-based action recognition methods using hand-crafted features
or using deep learning networks.
\textbf{Hand-crafted Features}~~
In \cite{hussein2013human},
the covariance matrices of the trajectories of the joint positions are computed over hierarchical temporal levels to model the skeleton sequences. In \cite{wang2012mining}, the pairwise relative positions of each joint with other joints are computed to represent each frame of the skeleton sequences, and Fourier Temporal Pyramid (FTP) is used to model the temporal patterns. In \cite{yang2012eigenjoints}, the pairwise relative positions of the joints are also used to characterize posture features, motion features, and
offset features of the skeleton sequences. Principal Component Analysis (PCA) is then applied to the normalized features to compute EigenJoints as representations. In \cite{xia2012view}, histograms of 3D joint locations are computed to represent each frame of the skeleton sequences, and HMMs are used to model the temporal dynamics. In \cite{vemulapalli2014human}, the rotations and translations between various body parts are used as representations, and
a skeleton sequence is modelled as a curve in the Lie group. The temporal dynamics are modelled with FTP.
\textbf{Deep Learning Methods }~~
In \cite{du2015hierarchical}, the skeleton joints are divided
into five sets corresponding to five body parts. They are fed into five LSTMs for feature fusion and classification.
In \cite{zhu2016co}, the skeleton joints are fed to a deep LSTM
at each time slot to learn
the inherent co-occurrence features of skeleton joints.
In \cite{Shahroudy_2016_CVPR}, the long-term context representations of the body parts are learned with a part-aware LSTM.
In \cite{liu2016spatio}, both the spatial and temporal information of skeleton sequences are learned
with a spatial temporal LSTM. A Trust Gate is also proposed to remove noisy joints. This method achieves the state-of-the-art performance on the NTU RGB+D dataset \cite{Shahroudy_2016_CVPR}.
\section{Proposed Method}
An overall architecture of the proposed method is shown in Figure \ref{cnn1}.
The proposed method starts by generating clips of skeleton sequences.
A skeleton sequence of any length is
transformed into three clips each consisting of several gray images. The generated clips are then fed to a deep CNN model to extract CNN features which are used in a MTLN for action recognition.
\subsection{Clip Generation}
\label{imagegene}
\begin{figure} [h!]
\vspace{-3mm}
\begin{center}
\includegraphics[width=3.2in]{visual21.pdf}
\end{center}
\vspace{-3mm}
\caption{Clip Generation of a skeleton sequence. The skeleton joints of each frame are first arranged as a chain by concatenating the joints of each body part (\ie, 1-2-3-...-16). Four reference joints shown in green (\ie, left shoulder 5, right shoulder 8, left hip 11 and right hip 14) are then respectively used to compute relative positions of the other joints to incorporate different spatial relationships between the joints. Consequently, four 2D arrays are obtained by combining the relative positions of all the frames of the skeleton sequence. The relative position of each joint in the 2D arrays is described with cylindrical coordinates. The four 2D arrays corresponding to the same channel of the coordinates are transformed to four gray images and as a clip. Thus three clips are generated from the three channels of the cylindrical coordinates of the four 2D arrays.}
\label{clip1}
\vspace{-1mm}
\end{figure}
Compared to RGB videos which consist of multiple frame images, skeleton sequences only provide the trajectories of the 3D coordinates. This paper proposes to transform the original skeleton sequence to a collection of clips each consisting of several images, thus to allow spatial temporal feature learning using deep neural networks. Intuitively, one could represent the content of each frame of the skeleton sequence as an image to generate a video.
However, if the skeleton sequence has many frames, this method will result in a long video of which the temporal dynamics will be difficult to learn. In addition,
each frame of the generated video will also be very sparse as the number of the skeleton joints is small.
To overcome this problem, we propose to represent the temporal dynamics of the skeleton sequence in a frame image, and then use multiple frames to incorporate different spatial relationships between the joints. An advantage of this method is that for any skeleton sequence of any length, the generated clips contain the same number of frames and the long-term temporal information of the original skeleton sequence can be effectively captured with the powerful CNN representations of the frame images in the generated clips.
As shown in Figure \ref{clip1}, for a skeleton sequence, the skeleton joints of each frame are first arranged as a chain by concatenating the joints of each body part.
Considering that the relative positions between joints provide more useful information than their absolute locations (\eg, the relative location of the hand to the shoulder in ``pushing''), four reference joints,
namely, the left shoulder, the right shoulder, the left hip and the right hip, are respectively
used to compute relative positions of the other joints, thus to incorporate different spatial relationships between joints and provide useful structural information of the skeleton.
These four joints are selected as reference joints due to the fact that they are stable in most actions. They can thus reflect the motions of the other joints. Although the base of the spine is also stable, it is close to the left hip and the right hip. It is therefore discarded to avoid information redundancy. By combing the relative joints of all the frames, four 2D arrays with dimension $(m-1)\times t$ are generated ($m $ is the number of skeleton joints in each frame and $t$ is the number of frames of the skeleton sequence).
The relative positions of joints in the 2D arrays are originally described with 3D Cartesian coordinates. Considering that the cylindrical coordinates are more useful to analyse the motions as each human body utilizes pivotal joint movements to perform an action, the 3D Cartesian coordinates are transformed to cylindrical coordinates in the proposed representation of skeleton sequences. The cylindrical coordinates have been used to extract view-invariant motion features for action recognition in \cite{weinland2006free}.
The four 2D arrays corresponding to the same channel of the 3D cylindrical coordinates are transformed to four gray images by scaling the coordinate values between 0 to 255 using a linear transformation. A clip is then constructed with the four gray images. Consequently, three clips are generated from the three channels of the 3D coordinates of the four 2D arrays.
\subsection{Clip Learning}
Each frame of the generated clips describes the temporal dynamics of all frames of the skeleton sequence and one particular spatial relationship between the skeleton joints in one channel of the cylindrical coordinates.
Different frames of the generated clip describe different spatial relationships and
there exists intrinsic
relationships among them.
A deep CNN is first leveraged to extract a compact representation from each frame of the generated clips to exploit the long-term temporal information of the skeleton sequence.
Then the CNN features of all frames of the generated clips are jointly processed in parallel using multi-task learning, thus to utilize their intrinsic relationships to learn the spatial temporal information for 3D action recognition.
\subsubsection{Temporal Pooling of CNN Feature Maps}
\label{cnnrep}
To learn the features of the generated clips, a deep CNN is firstly employed to extract a compact representation of each frame of the clips.
Since each frame describes the temporal dynamics of the skeleton sequence,
the spatial invariant CNN feature of each frame could thus represent the robust temporal information of the skeleton sequence.
Given the generated clips, the CNN feature of each frame is extracted with the pre-trained VGG19 \cite{simonyan2014very} model.
The pre-trained CNN model is leveraged as a feature extractor due to the fact that the CNN features extracted by the models pre-trained with ImageNet \cite{russakovsky2015imagenet} are very powerful and have been successfully applied in a number of cross-domain applications \cite{donahue2014decaf,girshick2014rich,sharif2014cnn,han2015matchnet}. In addition,
current skeleton datasets are either too small or too noisy to suitably train a deep network.
Although the frames of the generated clips are not natural images, they could still be fed to the CNN model pre-trained with ImageNet \cite{russakovsky2015imagenet} for feature extraction.
The similarity between a natural image and the generated frames is that both of them are matrices with some patterns. The CNN models trained on the large image dataset can be used as a feature extractor to extract representations of the patterns in matrices. The learned representations are generic and can be transferred to novel tasks from the original tasks \cite{yosinski2014transferable, long2015learning}.
\begin{figure}
\vspace{-1mm}
\begin{center}
\includegraphics[width=3.2in]{cnncmb2.pdf}
\end{center}
\vspace{-1mm}
\caption{\small{Temporal mean pooling of the CNN feature maps. (a) An input frame of the generated clips, for which the rows correspond to the different frames of the skeleton sequence and the columns correspond to the different vectors generated from the joints. (b) Output feature maps of the conv5\_1 layer. The size is $14\times 14 \times 512$. Each activation (shown in red) of the feature map is a feature correspond to the local region of the original image (shown with a red square).
(c) Temporal features of all joints of the skeleton sequence, which are obtained by applying mean pooling to each feature map in the row (temporal) dimension.
(d) Output feature, which is achieved by concatenating all the feature maps in (c).
}}
\label{cnn2}
\vspace{-1mm}
\end{figure}
The pre-trained VGG19 \cite{simonyan2014very} model contains 5 sets of convolutional layers conv1, conv2,
..., conv5. Each set includes a stack of 2 or 4 convolutional layers with the same kernel size.
Totally there are 16 convolutional layers and
three fully connected layers in the network.
Although deep neural networks are capable of learning powerful and
generic features which can be used in other novel domains, the features extracted from the different layers have different transferability. Particularly, the features in earlier layers are more generic, while in later layers, the features are more task-specific, which largely rely on the original classes and dataset. The features of the later layers are thus less suitable than those of the earlier layers to transfer to other domains \cite{yosinski2014transferable, long2015learning}.
Therefore,
this paper adopts a compact representation that is derived from the activations of the convolutional layer to exploit the temporal information of a skeleton sequence. The feature maps in the convolutional layer have been successfully applied for action recognition and image retrieval \cite{peng2015encoding, radenovic2016cnn}.
Specifically, the last 3 convolutional layers and fully connected layers of the network are discarded.
Each frame image of the three clips is scaled to $224 \times 224$, and is then duplicated three times to formulate a color image, so that it can be fed to the network.
The output of the convolutional layer conv5\_1 is used as the representation of the input frame,
which is a 3D tensor with size $14\times 14\times 512$,
\ie, 512 feature maps with size $14\times 14$.
The rows of the generated frame correspond to different frames of a skeleton sequence. The dynamics of the row features of the generated image therefore represents the temporal evolution of the skeleton sequence.
Meanwhile, the activations of each feature map in the conv5\_1 layer are the local features corresponding to the local regions in the original input image \cite{peng2015encoding}.
The temporal information of the sequence can thus be extracted from the row features of the feature maps.
More specifically, the feature maps are processed with temporal mean pooling with kernel size $14\times 1$, \ie, the pooling is applied over the temporal, or row dimension, thus to generate a compact fusion representation from all temporal stages of the skeleton sequence.
Let the activation at the $i^{th}$ row and the $j^{th}$ column of the $k^{th}$ feature map be $x_{i,j}^k$. After temporal mean pooling, the output of the $k^{th}$ feature map is given by:
\begin{equation}
\begin{array}{c}
\mathbf{y^k}=\left[y_1^k,\cdots, y_j^k, \cdots,y_{14}^k\right] \\
\\
y_j^k= \dfrac{1}{14}\sum\limits_{i=1}^{14}\max(0,x_{i,j}^k)\\
\end{array}
\end{equation}
The outputs of all feature maps (512) are concatenated to form a 7168D ($14\times512=7168$) feature vector, which represents the temporal dynamics of the skeleton sequence in one channel of the cylindrical coordinates.
\subsubsection{Multi-Task Learning Network (MTLN)}
As shown in Figure \ref{cnn1}(e), the three 7168D features of the three clips at the same time-step are concatenated to form a feature vector, generating four feature vectors in total. Each feature vector represents the temporal dynamics of the skeleton sequence and includes one particular spatial relationship between the joints in one of three cylindrical coordinates.
The four feature vectors have intrinsic relationships between each other.
An MTLN is then proposed to jointly process the four feature vectors to utilize their intrinsic relationships for action recognition.
The classification of each feature vector is treated as a separate task with the same classification label of the skeleton sequence.
The architecture of the network is shown in Figure \ref{cnn1}(f).
It includes two fully connected (FC) layers and a Softmax layer.
Between the two FC layers there is a rectified linear unit (ReLU) \cite{nair2010rectified} to introduce an additional non-linearity.
Given the four features as inputs, the MTLN generates four frame-level predictions, each corresponding to one task. During training, the class scores of each task are used to compute a loss value. Then the loss values of all tasks are summed up to generate the final loss of the network used to learn the network parameters.
During testing, the class scores of all tasks are averaged to form the final prediction of the action class. The loss value of the $k^{th}$ task ($k=1,\cdots,4$) is given by Equation \ref{eql1}.
\begin{equation}
\label{eql1}
\begin{array}{ll}
\ell_k(\mathbf{z_k},\mathbf{y})&=\sum\limits_{i=1}^{m} y_i \left(-log\left(\dfrac{\exp z_{ki}}{\sum\limits_{j=1}^m \exp z_{kj}}\right)\right) \\
&=\sum\limits_{i=1}^{m} y_i \left(log\left(\sum\limits_{j=1}^m \exp z_{kj} \right)-z_{ki}\right)
\end{array}
\end{equation}
where $\mathbf{z_k}$ is the vector fed to the Softmax layer generated from the $k^{th}$ input feature, $m$ is the number of action classes and $y_i$ is the ground-truth label for class $i$.
The final loss value of the network is computed as the sum of the
four individual losses, as shown below in Equation \ref{eql2}:
\begin{equation}
\label{eql2}
\mathcal{L}(Z,\mathbf{y})=\sum\limits_{k=1}^{4} \ell_k (\mathbf{z_k},\mathbf{y})
\end{equation}
where $Z=[\mathbf{z_1},\cdots,\mathbf{z_4}]$.
\section{Experiments and Analysis}
The proposed method is tested on three skeleton action datasets: NTU RGB+D dataset \cite{Shahroudy_2016_CVPR}, SBU kinect interaction dataset \cite{yun2012two} and CMU dataset \cite{cmu2013}.
The main ideas of the proposed method \textbf{Clips + CNN + MTLN } are 1) generating three clips (each clip consists of four frames) from a skeleton sequence, 2) using CNNs to learn global long-term temporal information of the skeleton sequence from each frame of the generated clips, and 3) using MTLN to
jointly train the CNN features of the four frames of the clips to incorporate the spatial structural information for action recognition.
We also conducted the following baselines to demonstrate the advantages of the proposed method:
\textbf{Coordinates + FTP}~~In this baseline, the Fourier Temporal Pyramid (FTP) \cite{wang2012mining} is applied to the
3D coordinates of the skeleton sequences to extract temporal features for action recognition. This baseline
is used to show the benefits of using CNNs for long-term temporal modelling of the skeleton sequences.
\textbf{Frames + CNN}~~In this baseline, the CNN features of single frames instead of the entire generated clips are used for action recognition. In other words, only one feature vector shown in Figure \ref{cnn1}(e) is used to train a neural network for classification.
Thus the loss value of the network is given by Equation \ref{eql1}.
The average accuracy of the four features is provided. This baseline is
used to show the benefits of using the entire generated clips to incorporate the spatial structural information for action recognition.
\textbf{Clips + CNN + Concatenation}~~In this baseline, the CNN features of all frames of the generated clips are concatenated before performing action recognition. In other words, the four feature vectors shown in Figure \ref{cnn1}(e) are concatenated and then fed to a neural network for classification.
This baseline is used to show the benefits of using MTLN to process the features of the entire clips in parallel.
\textbf{Clips + CNN + Pooling}~~In this baseline, max pooling is applied to the CNN features of all frames of the generate clips before performing action recognition.
Same as Clips + CNN + Concatenation, this baseline is also used to show the benefits of using MTLN.
\subsection{Datasets}
\textbf{NTU RGB+D Dataset} \cite{Shahroudy_2016_CVPR}~~
To the best of our knowledge, this dataset is so far the largest skeleton-based human action dataset, with more than 56000 sequences and 4 million frames. There are 60 classes of actions performed by 40 distinct subjects, including both one-person daily actions (e.g., clapping,
reading,
writing) and two-person interactions (\eg, handshaking, hug, pointing).
These actions are captured by three cameras, which are placed at different locations and view points. In total, there are 80 views for this dataset. In this dataset, each skeleton has 25 joints. The
3D coordinates of the joints are provided.
Due to the large view point, intra-class and sequence length variations, the dataset is very challenging.
\textbf{SBU Kinect Interaction Dataset} \cite{yun2012two}~~
This dataset was collected using the Microsoft Kinect sensor. It contains 282 skeleton sequences and
6822 frames. In this dataset, each frame contains two persons performing an interaction. The interactions include
approaching, departing, kicking, punching, pushing, hugging, shaking hands and exchanging. There are 15 joints for each skeleton. This dataset is challenging due to the fact that the joint coordinates exhibit low accuracy \cite{yun2012two}.
\textbf{CMU Dataset} \cite{cmu2013}~~
This dataset contains 2235 sequences and about 1 million frames. For each skeleton, the 3D coordinates of 31 joints are provided.
The dataset has been categorized into 45 classes \cite{zhu2016co}.
All of the actions are performed by only one person. The dataset is very challenging due to the large sequence length variations and intra-class diversity.
\subsection{Implementation Details}
For all datasets, the clips are generated with all frames of the original skeleton sequence without
any pre-processing such as normalization, temporal down-sampling or noise filtering.
The proposed method was implemented using the MatConvNet
toolbox \cite{vedaldi2015matconvnet}.
The number of the hidden unit of the first FC layer is set to 512.
For the second FC layer (\ie, the output layer), the number of the unit is the same as the number of the action classes in each dataset.
The network is trained using the stochastic gradient
descent algorithm. The learning rate is set
to 0.001 and batch size is set to 100. The training is stopped
after 35 epochs. The performance of the proposed method on
each dataset is compared with existing methods using the
same testing protocol.
\subsection{Results}
\label{exp}
\textbf{NTU RGB+D Dataset}~~As in \cite{Shahroudy_2016_CVPR}, the evaluation on this dataset is performed with two standard protocols, \ie, cross-subject evaluation and cross-view evaluation.
In cross-subject evaluation, the sequences of 20 subjects are used for training and the data from 20 other subjects are used for testing.
In cross-view evaluation, the sequences captured by two cameras are used for training and the rest are used for testing.
The results are shown in Table \ref{ntu}. It can be seen that the proposed method performs significantly better than others in both cross-subject and cross-view protocols.
The accuracy of the proposed method is 79.57\% when tested with the cross-subject protocol. Compared to the previous state-of-the-art method (ST-LSTM + Trust Gate \cite{liu2016spatio}), the performance is improved by 10.37\%.
When tested with the cross-view protocol, the accuracy is improved from 77.7\% to 84.83\%.
The improved performance of the proposed method is due to the novel clip representation and feature learning method.
As shown in Table \ref{ntu},
Frames + CNN
achieves an accuracy of about 75.73\% and 79.62\% for the two testing protocols, respectively.
The performances are much better than Coordinates + FTP.
Compared to extracting temporal features of skeleton sequences with FTP and native 3D coordinates,
using CNN to learn the temporal information of skeleton sequences from the generated frames is more robust to noise and temporal variations due to the convolution and pooling operators, resulting in better performances.
From Table \ref{ntu}, it can also be seen that
Frames + CNN also performs better than the previous state-of-the-art method. It clearly shows the effectiveness of the CNN features of the proposed clip representation. The performances are improved by learning entire clips with CNN and MTLN (\ie, Clips + CNN + MTLN). The improvements are about 4\% and 5\% for the two testing protocols, respectively.
It can also be seen that the proposed MTLN (\ie, Clips + CNN + MTLN) performs better than feature concatenation (\ie, Clips + CNN + concatenation) and pooling (\ie, Clips + CNN + pooling).
Frames + CNN, Clips + CNN + concatenation and Clips + CNN + pooling can be viewed as a single-task method, while using MTLN to process multiple frames of the generated clips in parallel utilizes their intrinsic relationships and incorporates the spatial structural information, which improves the performance of the single-task method for action recognition.
\begin{table}[]
\centering
\tabcolsep=0.05cm
\caption {Performance on the NTU RGB+D dataset.}
\begin{tabular}{c|cc}
\hline
\multirow{2}{*}{Methods} & \multicolumn{2}{c}{Accuracy}\\
&Cross Subject&Cross View\\
\hline
Lie Group \cite{vemulapalli2014human}& 50.1\%&52.8\%\\
Skeletal Quads \cite{evangelidis2014skeletal}&38.6\%&41.4\%\\
Dynamic Skeletons \cite{hu2015jointly}&60.2\%&65.2\%\\
Hierarchical RNN \cite{du2015hierarchical}&59.1\%&64.0\%\\
Deep RNN \cite{Shahroudy_2016_CVPR}&59.3\%&64.1\%\\
Deep LSTM \cite{Shahroudy_2016_CVPR}&60.7\%&67.3\%\\
Part-aware LSTM \cite{Shahroudy_2016_CVPR}&62.9\%&70.3\%\\
ST-LSTM \cite{liu2016spatio} & 65.2\%&76.1\%\\
ST-LSTM + Trust Gate \cite{liu2016spatio} & 69.2\%&77.7\%\\
\hline
Coordinates + FTP&61.06\% &74.64\%\\
Frames + CNN & 75.73\%&79.62\%\\
Clips + CNN + Concatenation & 77.05\%&81.11\%\\
Clips + CNN + Pooling & 76.37\%&80.46\%\\
Clips + CNN + MTLN & \textbf{79.57\%}& \textbf{84.83\%}\\
\hline
\end{tabular}
\vspace{-3mm}
\label{ntu}
\end{table}
\textbf{SBU Kinect Interaction Dataset}~~
As in \cite{yun2012two}, the evaluation of this dataset is a 5-fold cross validation, with the provided training/testing splits. Each frame of the skeleton sequences contains two separate human skeletons. In this case,
the two skeletons are considered as two data samples and the clip generation and feature extraction are conducted separately for the two skeletons. For testing, the prediction of actions is obtained by averaging the classification scores of the two samples.
Considering that the number of samples in this dataset is too small, data augmentation is performed to increase the number of samples. More specifically, each frame image of the generated clips are resized to $250\times250$, and then random patches with size of $224\times224$ are cropped from the original image for feature learning using CNN. For this dataset, 20 sub-images are cropped and the total data samples are extended to 11320.
The comparisons of the proposed method with other methods are shown in Table \ref{sbu}. Similar to the NTU RGB+D dataset, CNN features perform better than FTP to learn the temporal information.
It can be seen that when using CNN features of individual frames, the accuracy is 90.88\%, which is similar to the Deep LSTM + Co-occurrence method \cite{zhu2016co}. When incorporating the CNN features of the entire clips using concatenation and pooling methods, the performance is improved by about 2\%.
The performance is improved to 93.57\% when learning the entire clips with MTLN. It clearly shows the benefit
of using MTLN to learn the CNN features entire clips.
Since the joint positions of this dataset are not very accurate \cite{yun2012two}, existing methods including HBRNN \cite{du2015hierarchical} and Co-occurrence LSTM \cite{zhu2016co} remove the joint noise by smoothing the position of each joint using the Svaitzky-Golay filter \cite{savitzky1964smoothing}. In \cite{liu2016spatio}, a Trust Gate is introduced to remove the noisy joints and this improves the accuracy from 88.6\% to 93.3\%.
Our method does not perform any pre-processing to handle the noisy joints, but still performs better than all the others. It clearly shows that the features learned from the generated clips are robust to noise due to the convolution and pooling operators of the deep network.
\begin{table}[]
\centering
\tabcolsep=0.1cm
\caption {Performance on the SBU kinect interaction dataset.}
\begin{tabular}{c|c}
\hline
Methods&Accuracy\\
\hline
\pbox{3cm}
Raw Skeleton \cite{yun2012two} & 49.7\%\\
Joint Feature \cite{ji2014interactive}&86.9\%\\
CHARM \cite{li2015category} &83.9\%\\
Hierarchical RNN \cite{du2015hierarchical}&80.35\%\\
Deep LSTM \cite{zhu2016co}&86.03\%\\
Deep LSTM + Co-occurrence \cite{zhu2016co}&90.41\%\\
ST-LSTM \cite{liu2016spatio} &88.6\%\\
ST-LSTM + Trust Gate \cite{liu2016spatio} & 93.3\%\\
\hline
Coordinates + FTP& 79.75\%\\
Frames + CNN & 90.88\%\\
Clips + CNN + Concatenation &92.86\% \\
Clips + CNN + Pooling &92.26\%\\
Clips + CNN + MTLN & \textbf{93.57\%} \\
\hline
\end{tabular}
\label{sbu}
\end{table}
\begin{table}[t]
\centering
\tabcolsep=0.05cm
\caption {Performance on the CMU dataset.}
\begin{tabular}{c|cc}
\hline
\multirow{2}{*}{Methods} & \multicolumn{2}{c}{Accuracy}\\
&CMU subset&CMU\\
\hline
Hierarchical RNN \cite{du2015hierarchical}&83.13\%&75.02\%\\
Deep LSTM \cite{zhu2016co}&86.00\%&79.53\%\\
Deep LSTM + Co-occurrence \cite{zhu2016co}&88.40\%&81.04\%\\
\hline
Coordinates + FTP &83.44\% &73.61\%\\
Frames + CNN & 91.53\% &85.36\%\\
Clips + CNN + Concatenation &90.97\% &85.76\%\\
Clips + CNN + Pooling &90.66\%& 85.56\%\\
Clips + CNN+ MTLN & \textbf{93.22\%}& \textbf{88.30\%}\\
\hline
\end{tabular}
\label{cmu1}
\end{table}
\textbf{CMU Dataset}~~As in \cite{zhu2016co}, for this dataset, the evaluation is conducted on both the entire dataset with 2235 sequences, and a selected subset of 664 sequences. The subset includes 8 classes of actions, , \ie, basketball, cartwheel, getup, jump, pickup, run,
sit and walk back.
For the entire dataset, the testing protocol is 4-fold cross validation, and for the subset, it is evaluated
with 3-fold cross validation. The training/tesing splits of the different folds are provided by \cite{zhu2016co}.
Similar to the SBU kinect interaction dataset, data augmentation is also conducted on CMU dataset. For the entire dataset, each frame image is used to generate 5 more images and the total data samples are extended to 11175, and for the subset, the total samples are extended to 13280, which is 20 times of the original number.
The results are shown in Table \ref{cmu1}. It can be seen that the performance of the proposed method is much better than previous state-of-the-art methods on both the subset and the entire set. When tested on the subset, the accuracy of the proposed method was about 93.22\%, which is about 5\% better than the previous method \cite{zhu2016co}.
The performance on the entire dataset is improved from 81.04\% to 88.3\%.
\subsection{Discussions}
\textbf{Three gray clips or one color clip?}
As shown in Figure \ref{cnn1}, the frames of the three generated clips are gray images, each corresponding to only one channel of the cylindrical coordinates. Each frame is duplicated three times to formulate a color image for CNN feature learning. The output CNN features of the three channels are concatenated in a feature vector for action recognition. A simple alternative is to generate a color clip with three channels of the cylindrical coordinates, and then extract a single CNN feature from the color frame for action recognition. When this was tested on CMU dataset, the performance is 84.67\%, which is about 4\% worse than the proposed method. This is perhaps due to the fact that the relationship of the three generated channels is different from that of the RGB channels of natural color images. The
RGB channels are arranged in sequence and there is no
matching order between 3D coordinates and RGB channels.
\textbf{The more frames, the better performance?}~~
This paper uses only four reference joints to generate clips, each having four frames. When 6 more joints are selected to generate more frames, \ie, the head, the left hand, the right hand, the left foot, the right foot and the hip, the performance does not improve. When tested on CMU data, the performance is 86.01\%, which is about 2\% worse than the proposed method. This is due to the fact that the other joints are not as stable as the selected four joints, which can introduce noise.
\textbf{Cartesian coordinates or cylindrical coordinates?}~~
As mentioned in Section \ref{imagegene}, the 3D Cartesian coordinates of the vectors between the reference joints and the other joints are transformed to cylindrical coordinates to generate clips. We found that when using the original Cartesian coordinates for clip generation and action recognition, the performance drops. When tested on CMU dataset, the accuracy is 86.21\%, which is about 2\% worse than the proposed method. The cylindrical coordinates are more useful than the Cartesian coordinates to analyse the motions as each human skeleton utilizes pivotal joint movements to perform an action.
\textbf{Features in different layers}~~
As mentioned in Section \ref{cnnrep}, the feature maps in conv5\_1 layer of the pre-trained CNN model is adopted as the representation of each input image. We found that using the features in the earlier layers
decreased the performance. When using the features of the conv4\_1 layer, the accuracy on CMU dataset is 84.59\%, which is about 4\% worse than the proposed method. This is perhaps due to the fact that the features in the earlier layers are not deep enough to capture the salient information of the input image.
We also found that using the features in the later layers
made the performance worse.
When using the features of the fc6 layer, the accuracy on CMU dataset is 83.52\%, which is about 5\% worse than the proposed method.
This is because
the features in the later layers are more task-specific, which largely rely on the original classes and dataset. The features of the later layers are thus less suitable than those of the earlier layers to transfer to other domains \cite{yosinski2014transferable, long2015learning}.
\section{Conclusion}
In this paper, we have proposed to transform a skeleton sequence to three video clips for robust feature learning and action recognition.
We proposed to use a pre-trained CNN model followed by a temporal pooling layer to extract a compact representation of each frame. The CNN features of the three clips at the same time-step are concatenated in a single feature vector,
which describes the temporal information of the entire skeleton sequence and one particular spatial relationship between the joints.
We then propose an MTLN to jointly learn the feature vectors at all the time-steps in parallel, which utilizes their intrinsic relationships and improves the performance for action recognition.
We have tested the proposed method on three datasets, including NTU RGB+D dataset, SBU kinect interaction dataset and CMU dataset. Experimental results have shown the effectiveness of the proposed new representation and feature learning method.
\section{Acknowledgment}
This work was partially supported by Australian Research
Council grants DP150100294, DP150104251, and DE120102960.
This paper used the NTU RGB+D Action Recognition Dataset made available by the ROSE Lab at the Nanyang Technological University, Singapore.
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
\label{sec: Intro}
For more than three decades, topological quantum systems have remained at the centre of attention in condensed matter physics.\cite{top-SC} These diverse systems, from a two-dimensional (2D) electron gas
exhibiting the integer Quantum Hall Effect\cite{TKNN82}, to chiral $p$-wave superconductors and superfluids,\cite{Volovik-book} to topological band insulators,\cite{KM05,TI-review} all share one common feature:
the quantum states in the bulk fall into distinct classes characterized by various topological invariants, which are robust under sufficiently small perturbations. The choice of the bulk invariant is determined by
the symmetry and dimensionality of the system, with a particularly important role played by time reversal symmetry (TRS).
According to Ref. \onlinecite{SRFL08}, for single-particle Hamiltonians the topological invariant either takes an
integer value or is a $\mathbb{Z}_2$ quantity, although this may be significantly modified by interactions.\cite{interactions-Z8}
There is a widely held belief that a nontrivial topology of the bulk manifests itself in the presence of protected gapless quasiparticle states localized near the boundary of the system, which is known as
the bulk-boundary correspondence.\cite{Volovik-book} Archetypal examples include the current-carrying chiral edge states in the quantum Hall\cite{QHE-edge}
or topological band insulators,\cite{KM05} and also the Andreev bound states (ABS) in the chiral $p$-wave\cite{ABS-pwave} and the nodal $d$-wave superconductors.\cite{Hu94}
In some cases, it is possible to obtain an explicit analytical relation between the number of the zero-energy modes and a certain topological invariant in the bulk.\cite{BBC} For instance, the
quantum-Hall edge states are related to the first Chern number of the 2D Brillouin zone, which is also known as the TKNN integer, after Ref. \onlinecite{TKNN82},
while the dispersionless surface ABS in TR invariant unconventional
superconductors are controlled by the phase winding number of the determinant of the off-diagonal Bogoliubov-de Gennes (BdG) Hamiltonian.\cite{STYY11} Note, however, that there are some recent results\cite{MR15}
that call into question the universality of the bulk-boundary correspondence, at least in the $\mathbb{Z}_2$ case.
The extent to which the zero modes are ``protected'', i.e., insensitive to the boundary details, is one of the less-understood aspects of the bulk-boundary correspondence. Given the crucial role played by TRS,
of particular interest here is the fate of the zero modes in the situations when the bulk is TR invariant, while the boundary is not. The goal of this paper is to study the effects on the ABS of (i) the magnetic boundary
scattering and (ii) an intrinsic TRS breaking in the superconducting state, both at the mean-field level and also including fluctuations.
One can capture the essential physics by looking at a half-infinite one-dimensional (1D) superconductor in contact with a ferromagnetic insulator. In contrast to previous works, see Ref. \onlinecite{TRB-1D}, we do not
consider the TRS breaking by an applied magnetic field.
The 1D (or rather, quasi-1D) superconductivity can be realized in a metallic quantum wire on a substrate. The pairing interaction can be either intrinsic
to the wire or extrinsic, i.e., induced by the substrate. Inversion symmetry is absent in this system and a crucial role is played by the Rashba spin-orbit (SO) coupling of the electrons in the wire
with the asymmetric substrate potential, see Refs. \onlinecite{Rashba-model}, \onlinecite{Manchon15}, and the references therein. The Rashba SO coupling lifts the spin degeneracy of the electron states,
producing nondegenerate Bloch bands labeled by ``helicity'', with the wave functions characterized by a nontrivial momentum-space topology.
Its profound consequences for superconductivity have been extensively studied in the last decade, see Refs. \onlinecite{NCSC-book}, \onlinecite{Kneid15}, and \onlinecite{Smid17} for reviews.
The paper is organized as follows. The structure of the single-electron bands and the superconducting pairing in a 1D Rashba wire are discussed in Sec. \ref{sec: electron bands}. In Sec. \ref{sec: ABS}, we present
the semiclassical derivation of the ABS spectrum, which is expressed in terms of the boundary scattering matrix. The latter is calculated in Sec. \ref{sec: FM boundary}. Stability of the ABS zero modes against TRS-breaking
perturbations is studied in Sec. \ref{sec: ABS-stability}.
Throughout the paper we use the units in which $\hbar=k_B=1$, neglecting, in particular, the difference between the quasiparticle momentum and wave vector.
\section{Superconductivity in 1D nondegenerate bands}
\label{sec: electron bands}
We consider a quasi-1D electron gas on a $xy$-plane substrate. Neglecting the lattice periodicity, the three-dimensional (3D) potential $U(x,y,z)$ affecting the electrons is constant in $x$ direction,
but confining in both $y$ and $z$ directions. This system is TR invariant in the normal state but lacks an inversion center, because the substrate breaks the $z\to-z$ mirror reflection symmetry.
The momentum space is one-dimensional, labelled by the wave vector $\bk=k_x\hat{\bm{x}}$, where $-\infty<k_x<\infty$.
The simplest Hamiltonian that captures the essential features of the electronic band structure in a noncentrosymmetric 1D system has the following form:
\begin{equation}
\label{H-Rashba}
\hat H_0=\sum\limits_{k_x}\sum_{s,s'=\uparrow,\downarrow}\left[\epsilon_0(k_x)\delta_{ss'}+\bgam(k_x)\bm{\sigma}_{ss'}\right]\hat b^\dagger_{k_x,s}\hat b_{k_x,s'}.
\end{equation}
This is the 1D version of the well-known Rashba model.\cite{Rashba-model} The first term describes a single spin-degenerate band, for which we use the effective mass approximation:
\begin{equation}
\label{eff-mass}
\epsilon_0(k_x)=\frac{k_x^2}{2m^*}-\epsilon_F,
\end{equation}
where $\epsilon_F=k_F^2/2m^*$ is the Fermi energy (the difference between $\epsilon_F$ and the chemical potential is neglected). The second term in Eq. (\ref{H-Rashba}) is the asymmetric SO coupling, with
$\hat{\bm{\sigma}}$ being the Pauli matrices.
While the momentum space is 1D, the spin space is still 3D, so that the asymmetric SO coupling is described by the 3D pseudovector $\bgam(k_x)$,
which is real and odd in $k_x$ due to the TRS. The simplest expression compatible with both requirements is
\begin{equation}
\label{gamma-Rashba}
\bgam(k_x)=\bm{a}k_x,
\end{equation}
with a real $\bm{a}$. In the absence of additional mirror reflection symmetries of the confining potential,\cite{Sam17} there are no further constraints on the components of $\bm{a}$.
Below we will use the spherical angle parameterization,
\begin{equation}
\label{a-angles}
\bm{a}=|\bm{a}|(\sin\alpha\cos\beta,\sin\alpha\sin\beta,\cos\alpha),
\end{equation}
where $0\leq\alpha\leq\pi$ and $0\leq\beta<2\pi$.
The asymmetric SO coupling lifts the spin degeneracy of the bands almost everywhere in the momentum space. Diagonalizing Eq. (\ref{H-Rashba}), we obtain:
\begin{equation}
\label{Rashba-bands}
\xi_\lambda(k_x)=\frac{k_x^2-k_F^2}{2m^*}+\lambda|\bm{a}||k_x|=\xi_\lambda(-k_x).
\end{equation}
Here the band index $\lambda=\pm$, called the helicity, has the meaning of the spin projection on the direction of motion. Although the two 1D Fermi ``surfaces'', defined by the equations $\xi_\pm(k_x)=0$,
see Fig. \ref{fig: bands}, have different sizes:
\begin{equation}
\label{Fermi-wavevectors}
k_{F,\lambda}=\tilde k_F-\lambda\Lambda,\quad \tilde k_F=\sqrt{k_F^2+\Lambda^2},
\end{equation}
where $\Lambda=m^*|\bm{a}|$, the Fermi velocities are the same in both bands and equal to $\tilde v_F=\tilde k_F/m^*$.
The corresponding eigenstates can be chosen in the following form:
\begin{eqnarray}
\label{Rashba-eigenstates}
|k_x,+\rangle=\frac{1}{\sqrt{2\tilde v_F}}\left(\begin{array}{c}
e^{-i\beta/2}\sqrt{1+\sgn k_x\cos\alpha}\vspace*{5pt} \\
e^{i\beta/2}\sgn k_x\sqrt{1-\sgn k_x\cos\alpha}
\end{array}\right),\nonumber\\ \\
|k_x,-\rangle=\frac{1}{\sqrt{2\tilde v_F}}\left(\begin{array}{c}
-ie^{-i\beta/2}\sqrt{1-\sgn k_x\cos\alpha}\vspace*{5pt} \\
ie^{i\beta/2}\sgn k_x\sqrt{1+\sgn k_x\cos\alpha}
\end{array}\right),\nonumber
\end{eqnarray}
normalized to produce the current of unit magnitude.
The TR invariance of the normal state implies that the states $|k_x,\lambda\rangle$ and $K|k_x,\lambda\rangle$ have the same energy, forming a Kramers doublet
(recall that the TR operator for spin-1/2 particles is $K=i\hat\sigma_2K_0$, where $\hat\sigma_2$ is the Pauli matrix and $K_0$ is complex conjugation). Therefore,
\begin{equation}
\label{t-factor}
K|k_x,\lambda\rangle=t_\lambda(k_x)|-k_x,\lambda\rangle,
\end{equation}
where $t_\lambda$ is a phase factor.\cite{t-factor} Since $K^2=-1$ for fermions, one can show that $t_\lambda(-k_x)=-t_\lambda(k_x)$.
Note that the phase factors are not gauge invariant, and for the eigenstates given by Eq. (\ref{Rashba-eigenstates}) have a particularly simple form:
\begin{equation}
\label{t-sgn-x}
t_\lambda(k_x)=\sgn k_x.
\end{equation}
The helicity bands (\ref{Rashba-bands}) are nondegenerate at all $k_x$, except the TR invariant point $k_x=0$, where the SO coupling (\ref{gamma-Rashba}) vanishes and neither the eigenstates
nor the TR phase factors are well defined. In a lattice model taking into account the momentum space periodicity, there are additional TR invariant points at the 1D Brillouin zone boundaries, where the SO coupling also vanishes.
Next, we use the helicity basis to construct the superconducting Hamiltonian. The only assumption we make is that, whatever the microscopic pairing mechanism, the SO band splitting is large enough
to suppress the pairing of electrons from different bands. In the mean-field approximation we have
\begin{eqnarray}
\label{H-MF}
\hat H &=& \sum_{k_x,\lambda}\xi_\lambda(k_x)\hat c^\dagger_{k_x,\lambda}\hat c_{k_x,\lambda}\nonumber\\
&&+\frac{1}{2}\sum_{k_x,\lambda}\left[\Delta_\lambda(k_x)\hat c^\dagger_{k_x,\lambda}\hat{\tilde c}^\dagger_{k_x,\lambda}+\mathrm{H.c.}\right],
\end{eqnarray}
where the second term represents the intraband Cooper pairing between the states $|k_x,\lambda\rangle$ and $K|k_x,\lambda\rangle$. According to Eq. (\ref{t-factor}), the electron creation operator in
the time-reversed state $K|k_x,\lambda\rangle$ is given by
\begin{equation}
\label{c-tilde c}
\hat{\tilde c}^\dagger_{k_x,\lambda}\equiv K\hat c^\dagger_{k_x,\lambda}K^{-1}=t_\lambda(k_x)\hat c^\dagger_{-k_x,\lambda}.
\end{equation}
Substituting this last expression in $\hat H$ and using the anticommutation of the fermionic creation and annihilation operators, we obtain:
\begin{equation}
\label{Delta-even}
\Delta_\lambda(k_x)=\Delta_\lambda(-k_x),
\end{equation}
therefore the pairing in the helicity representation is necessarily even in momentum. The TR operation acting on the gap functions is equivalent to the complex conjugation, i.e., $\Delta_\lambda(k_x)\to\Delta^*_\lambda(k_x)$,
while under an arbitrary rotation of the band state phases, $|k_x,\lambda\rangle\to e^{i\theta_\lambda(k_x)}|k_x,\lambda\rangle$, the gap functions remain invariant.
In a BCS-type model, the gap functions are nonzero only in the vicinity of the Fermi level, where their $k_x$-dependence can be neglected. In our case, the 1D ``Fermi surface'' is given by the four Fermi wave vectors
$\pm k_{F,\pm}$, see Fig. \ref{fig: bands}, and the superconductivity is described by two complex order parameters $\Delta_+$ and $\Delta_-$.
The stable uniform states are found by minimizing the Ginzburg-Landau free energy.
In addition to the TR invariant states, in which the phases of the gap functions are either $0$ or $\pi$, there is also a phenomenological possibility of TRS-breaking states, both in the multiband and the two-band cases,
\cite{TRSB-states} in which
\begin{equation}
\label{gap-functions}
\Delta_+=|\Delta_+|e^{i\chi},\quad\Delta_-=|\Delta_-|,
\end{equation}
where $0\leq\chi\leq\pi$ is the interband phase difference. The bulk quasiparticle spectrum consists of two electron-hole symmetric branches in each helicity band, given by $\pm\sqrt{\xi_\lambda^2+|\Delta_\lambda|^2}$,
with the energy gap equal to $|\Delta_\lambda|$.
Suppose the gap functions are real. Then, our system belongs to the symmetry class DIII in 1D and, therefore, can be characterized in the bulk by a $\mathbb{Z}_2$
topological invariant.\cite{SRFL08} Due to TR invariance, the boundary zero modes can only come in pairs,\cite{TRI-1D} and the $\mathbb{Z}_2$ invariant is nothing but the parity of the number of such Kramers pairs.
For the Rashba superconductor with just two bands, this invariant has the following form:
\begin{equation}
\label{Z2-invariant}
N_{1D}=\sgn\Delta_+\sgn\Delta_-,
\end{equation}
see Ref. \onlinecite{QHZ10}. The states with $N_{1D}=1$ ($\chi=0$) are topologically trivial, while those with $N_{1D}=-1$ ($\chi=\pi$) are nontrivial and should have two ABS zero modes.
Below, in Secs. \ref{sec: ABS} and \ref{sec: ABS-stability}, we examine the validity of this last statement using an explicit calculation of the ABS energy under general assumptions about the boundary scattering.
\begin{figure}
\includegraphics[width=7cm]{bands.eps}
\caption{The 1D helicity bands and the Fermi points.}
\label{fig: bands}
\end{figure}
\subsection{Spin representation}
\label{sec: Kitaev-chain}
It is instructive to translate the above results about the superconducting pairing in the helicity representation into the spin representation. The fermionic creation and annihilation operators are transformed into the spin basis
using the following relations:
\begin{eqnarray*}
\hat c^\dagger_{k_x,\lambda}=\sum_{s=\uparrow,\downarrow}\langle k_x,s|k_x,\lambda\rangle\hat b_{k_x,s}^\dagger,\\
\hat{\tilde c}^\dagger_{k_x,\lambda}=\sum_{s=\uparrow,\downarrow}\langle k_x,\lambda|k_x,s\rangle\hat{\tilde b}_{k_x,s}^\dagger,
\end{eqnarray*}
where $\hat{\tilde b}_{k_x,s}^\dagger\equiv K\hat b_{k_x,s}^\dagger K^{-1}=(i\hat\sigma_2)_{ss'}\hat b_{-k_x,s'}^\dagger$. Substituting these expressions into
the pairing Hamiltonian (\ref{H-MF}), we obtain:
\begin{eqnarray}
\label{H-MF-spin}
\hat H &=& \sum_{k_x,ss'}\epsilon_{ss'}(k_x)\hat b_{k_x,s}^\dagger\hat b_{k_x,s'}\nonumber\\
&&+\frac{1}{2}\sum_{k_x,ss'}\bigl[\Delta_{ss'}(k_x)\hat b_{k_x,s}^\dagger\hat b_{-k_x,s'}^\dagger+\mathrm{H.c.}\bigr].
\end{eqnarray}
The normal-state energy and the gap function become $2\times 2$ spin matrices:
\begin{eqnarray}
\label{spin-matrices}
&& \hat\epsilon(k_x)=\sum_\lambda\xi_\lambda(k_x)\hat\Pi_\lambda(k_x),\nonumber\\ \\
&& \hat\Delta(k_x)=\sum_\lambda\Delta_\lambda(k_x)\hat\Pi_\lambda(k_x)(i\hat\sigma_2),\nonumber
\end{eqnarray}
where $\hat\Pi_\lambda(k_x)=|k_x,\lambda\rangle\langle k_x,\lambda|$ is the projection operator onto the $\lambda$th band.
From Eq. (\ref{Rashba-eigenstates}) we obtain the following expression for the projection operator:
$$
\hat\Pi_\lambda(k_x)=\frac{1+\lambda\hat{\bgam}(k_x)\hat{\bm{\sigma}}}{2},
$$
where $\hat{\bgam}=\bgam/|\bgam|=\hat{\bm{a}}\sgn k_x$, see Eq. (\ref{gamma-Rashba}). Then, taking into account Eq. (\ref{Rashba-bands}), the normal-state energy matrix in Eq. (\ref{spin-matrices}) reproduces the Rashba
Hamiltonian (\ref{H-Rashba}), while the gap function in the spin representation takes the form
\begin{equation}
\label{gap-spin}
\hat\Delta(k_x)=\Delta_s(i\hat\sigma_2)+\Delta_t\hat{\bgam}(k_x)(i\hat{\bm{\sigma}}\hat\sigma_2).
\end{equation}
Here
$$
\Delta_s=\frac{\Delta_++\Delta_-}{2},\quad \Delta_t=\frac{\Delta_+-\Delta_-}{2},
$$
and we neglected the momentum dependence of $\Delta_\pm$. The expression (\ref{gap-spin}) describes a mixture of spin-singlet ($\Delta_s$) and spin-triplet ($\Delta_t$) pairing. The latter, which is characterized by the
spin vector $\bm{d}(k_x)=\Delta_t\hat{\bgam}(k_x)$, is protected against the pair breaking effect of the Rashba SO band splitting.\cite{NCSC-book}
In a purely triplet superconducting state with $\Delta_+=-\Delta_-=\Delta$, we have $\hat\Delta(k_x)=\Delta\hat{\bgam}(k_x)(i\hat{\bm{\sigma}}\hat\sigma_2)$. Assuming for simplicity that
$\bm{a}\parallel\hat{\bm{y}}$ in Eq. (\ref{gamma-Rashba}), which is required by symmetry for some 1D point groups,\cite{Sam17} we obtain: $\hat\Delta(k_x)=i\Delta\sgn k_x\hat\sigma_0$.
This gap function describes an odd in $k_x$ (``$p$-wave'') pairing state, which is TR invariant (because one can make the gaps real by a gauge transformation) and topologically nontrivial in the bulk, with two zero-energy ABS
at the boundary,\cite{SF09} in agreement with the classification based on the invariant (\ref{Z2-invariant}). This state can also be viewed as two coupled Kitaev chains, see Ref. \onlinecite{Kit01}.
Since each chain contributes one zero-energy ABS, there are two zero modes at the end of the wire, corresponding to two Majorana quasiparticles.\cite{Majoranas} The stability of the
Majorana states in the Kitaev chain and similar models against perturbations, such as interactions or disorder, has been investigated in a number of works, see Refs. \onlinecite{interactions-Z8} and \onlinecite{MF-stability}.
\section{Spectrum of the boundary modes}
\label{sec: ABS}
We consider a half-infinite superconductor at $x\geq 0$. The Bogoliubov quasiparticle wave function in each band is a two-component (electron-hole) spinor.
In the vicinity of the Fermi point $rk_{F,\lambda}$, where $r=\pm$, it
can be represented in the semiclassical approximation as $e^{irk_{F,\lambda}x}\psi_{\lambda,r}(x)$. The ``envelope'' function $\psi_{\lambda,r}$ varies slowly on the scale of the Fermi wavelength
$k_{F,\lambda}^{-1}$ and satisfies the Andreev equation:\cite{And64}
\begin{equation}
\label{And-eq-gen}
\left(\begin{array}{cc}
-iv_{\lambda,r}\dfrac{d}{dx} & \Delta_\lambda \\
\Delta^*_\lambda & iv_{\lambda,r}\dfrac{d}{dx}
\end{array}\right)\psi_{\lambda,r}=E\psi_{\lambda,r}.
\end{equation}
Here $v_{\lambda,r}=r\tilde v_F$ is the group velocity and $\Delta_\lambda\equiv\Delta_\lambda(k_{F,\lambda})=\Delta_\lambda(-k_{F,\lambda})$
is the gap function affecting the quasiparticles near the Fermi point $rk_{F,\lambda}$. To make analytical progress, we neglect self-consistency and assume that the gap functions do not depend on $x$.
We focus on the bound-state solutions of Eq. (\ref{And-eq-gen}) localized near $x=0$. The semiclassical approximation breaks down near the boundary due to a rapid variation of the potential, which leads to the mixing of
the states corresponding to different Fermi wave vectors $\pm k_{F,\pm}$. As a result, the general wave function of the subgap states away from the boundary is given by a superposition
of four semiclassical solutions:
\begin{equation}
\label{Psi-ABS}
\Psi(x)=\sum_{\lambda=\pm}\sum_{r=\pm}\phi(rk_{F,\lambda})e^{-\Omega_\lambda x/\tilde v_F}e^{irk_{F,\lambda}x},
\end{equation}
where $\Omega_\lambda=\sqrt{|\Delta_\lambda|^2-E^2}$ and the Andreev amplitudes have the form
\begin{equation}
\label{Andreev amplitude}
\phi(rk_{F,\lambda})=C(rk_{F,\lambda})\left(\begin{array}{c}
\dfrac{\Delta_\lambda}{E-i\Omega_\lambda\sgn v_{\lambda,r}}\vspace*{5pt} \\ 1
\end{array}\right).
\end{equation}
The ABS energy satisfies $|E|<\min(|\Delta_-|,|\Delta_+|)$.
Depending on the sign of the group velocity, the Fermi wave vectors in the wave function (\ref{Psi-ABS}) are classified as either incident, $k^{\mathrm{in}}_\lambda=-k_{F,\lambda}$,
or reflected, $k^{\mathrm{out}}_\lambda=k_{F,\lambda}$. According to Eq. (\ref{Andreev amplitude}), the corresponding Andreev amplitudes are given by
\begin{equation}
\label{phi-in-out}
\phi\left(k^{\mathrm{in(out)}}_\lambda\right)=C\left(k^{\mathrm{in(out)}}_\lambda\right)\left(\begin{array}{c}
\dfrac{\Delta_\lambda}{E\pm i\Omega_\lambda}\vspace*{5pt} \\ 1
\end{array}\right).
\end{equation}
The boundary condition for the wave function (\ref{Psi-ABS}) cannot be derived using the semiclassical approximation.
As shown in Ref. \onlinecite{Shel-bc}, it can be written as a relation between the Andreev amplitudes for the reflected and incident states:
\begin{equation}
\label{Shelankov-bc}
\phi(k^{\mathrm{out}}_\lambda)=\sum_{\lambda'} S_{\lambda\lambda'}\phi(k^{\mathrm{in}}_{\lambda'}).
\end{equation}
Here the coefficients $S_{\lambda\lambda'}$ form a unitary $2\times 2$ matrix ($S$ matrix), which is an electron-hole scalar, determined by the microscopic details of the boundary scattering
at the Fermi level in the normal state.
Inserting the expressions (\ref{phi-in-out}) into Eq. (\ref{Shelankov-bc}), we obtain the general form of the ABS energy equation, valid for any mechanism of the boundary scattering:
\begin{equation}
\label{ABS-equation}
\frac{E^2-|\Delta_-||\Delta_+|\cos\chi}{\sqrt{(|\Delta_-|^2-E^2)(|\Delta_+|^2-E^2)}} = R,
\end{equation}
where
\begin{equation}
\label{R-def}
R=1-2\frac{S_{--}S_{++}}{S_{-+}S_{+-}}=1+2\frac{|S_{--}|^2}{|S_{-+}|^2}
\end{equation}
(the second equality here follows from the unitarity of the $S$ matrix). As evident from Eq. (\ref{ABS-equation}), the ABS spectrum consists of symmetrical pairs $\pm|E|$;
therefore the zero-energy states, if they exist, are twofold degenerate.
At $\chi=0$, the $\mathbb{Z}_2$ invariant (\ref{Z2-invariant}) is equal to $+1$, placing this bulk state into the topologically trivial class, without any stable zero modes. Indeed, in this case
the left-hand side of Eq. (\ref{ABS-equation}) is negative, while the right-hand side is positive, which means that there are no subgap ABS solutions, zero-energy or not, regardless of the boundary details.
In the other TR invariant superconducting state, corresponding to $\chi=\pi$, we have $N_{1D}=-1$. However, it follows from Eqs. (\ref{ABS-equation}) and (\ref{R-def}) that the zero-energy
solution at $\chi=\pi$ exists only if $R=1$. This last condition is equivalent to the requirement that $S_{--}=S_{++}=0$, i.e., there is no backscattering into the same band.
Thus we see that the zero modes are sensitive to the form of the boundary scattering matrix, even if the bulk state is topologically nontrivial.
\section{The $S$ matrix}
\label{sec: FM boundary}
In this section, we study the structure of the $S$ matrix in the presence of the TRS-breaking boundary scattering. Since the $S$ matrix is an electron-hole scalar,
one can neglect the superconductivity and consider a normal metal in contact with a ferromagnetic insulator, the latter occupying the $x<0$ half-space.
The ferromagnetism is modelled by the exchange splitting $h$ of the potentials affecting the two spin channels:
\begin{equation}
\label{U-FM}
\hat{U}_{FM}=U\hat\sigma_0+h\hat\sigma_3.
\end{equation}
We assume that $U\pm h>\epsilon_F$. The mismatch between the effective masses of quasiparticles on the metallic and insulating sides is neglected for analytical simplicity.
The normal-state Hamiltonian, naively written as $\hat{H}=\epsilon_0(\hat k_x)\hat{\sigma}_{0}+\theta(x)\bgam(\hat k_x)\hat{\bm{\sigma}}+\theta(-x)\hat{U}_{FM}$, where $\hat k_x=-i\nabla_x$,
is not Hermitian due to the SO coupling term. To restore its Hermiticity, we use the following revised form:
$$
\hat{H}=\epsilon_0(\hat k_x) \hat{\sigma}_{0} + \frac{1}{2}\{\bgam(\hat k_x),\theta(x)\}\hat{\bm{\sigma}} + \theta(-x)\hat{U}_{FM},
$$
where $\{...\}$ denotes the anticommutator and $\hat{U}_{FM}$ is given by Eq. (\ref{U-FM}). In the case of the 1D Rashba expression for the SO coupling, see Eq. (\ref{gamma-Rashba}),
the Hamiltonian becomes
\begin{eqnarray}
\label{H-normal}
\hat{H} &=& \epsilon_0(\hat k_x) \hat{\sigma}_{0} + \theta(-x)\hat{U}_{FM} + \theta(x)(\bm{a}\hat{\bm{\sigma}})\hat k_x\nonumber\\
&& +\frac{i}{2}(\bm{a}\hat{\bm{\sigma}})\delta(x),
\end{eqnarray}
The additional delta-function term leads to a modification of the boundary conditions for the spinor wave functions:
\begin{equation}
\label{psi-bc}
\left\{\begin{array}{l}
\psi(+0)=\psi(-0)=\psi(0),\\
\psi'(+0)-\psi'(-0)=-i\Lambda\dfrac{(\bm{a}\hat{\bm{\sigma}})}{|\bm{a}|}\psi(0),
\end{array}\right.
\end{equation}
where $\Lambda=m^*|\bm{a}|$ characterizes the SO coupling strength.
The Fermi-level eigenstates of the Hamiltonian (\ref{H-normal}) are given by the evanescent waves on the insulating side ($x<0$):
\begin{equation}
\label{FM-side}
\psi(x)=C_\uparrow\left(\begin{array}{c}
1 \\ 0
\end{array}\right)e^{\kappa_\uparrow x}
+C_\downarrow\left(\begin{array}{c}
0 \\ 1
\end{array}\right)e^{\kappa_\downarrow x},
\end{equation}
with $\kappa_{\uparrow(\downarrow)}=\sqrt{2m^*(U\pm h-\epsilon_F)}$, and by a superposition of four propagating waves on the metallic side ($x>0$):
\begin{equation}
\label{metal-side}
\psi(x)=\sum_{\lambda=\pm}\left(A_\lambda\chi_\lambda^{\mathrm{in}}e^{-ik_{F,\lambda} x}+B_\lambda\chi_\lambda^{\mathrm{out}}e^{ik_{F,\lambda} x}\right),
\end{equation}
with the Fermi wave vectors defined by Eq. (\ref{Fermi-wavevectors}). The spinor components of the propagating waves are obtained from Eq. (\ref{Rashba-eigenstates}):
\begin{eqnarray*}
\chi_-^{\mathrm{in}}=-i\chi_+^{\mathrm{out}}=-\frac{i}{\sqrt{\tilde v_F}}\left(\begin{array}{c}
e^{-i\beta/2}\cos\dfrac{\alpha}{2}\vspace*{2pt} \\
e^{i\beta/2}\sin\dfrac{\alpha}{2}
\end{array}\right),\\
\chi_-^{\mathrm{out}}=-i\chi_+^{\mathrm{in}}=-\frac{i}{\sqrt{\tilde v_F}}\left(\begin{array}{c}
e^{-i\beta/2}\sin\dfrac{\alpha}{2}\vspace*{2pt} \\
-e^{i\beta/2}\cos\dfrac{\alpha}{2}
\end{array}\right),
\end{eqnarray*}
where we used the spherical angle parameterization of the SO coupling, see Eq. (\ref{a-angles}).
Substituting Eqs. (\ref{FM-side}) and (\ref{metal-side}) into the boundary conditions (\ref{psi-bc}), we can express the amplitudes of the reflected waves
in terms of the amplitudes of the incident waves:
\begin{equation}
\label{S-matrix-def}
B_\lambda=\sum_{\lambda'}S_{\lambda\lambda'}A_{\lambda'}.
\end{equation}
The scattering matrix here is given by
\begin{widetext}
\begin{equation}
\label{S-matrix-result}
\hat S=-\frac{i}{(1+i\rho)^2+\rho_m^2}\left(\begin{array}{cc}
2\rho_m\sin\alpha & -1-\rho^2+\rho_m^2-2i\rho_m\cos\alpha \vspace*{2pt} \\
1+\rho^2-\rho_m^2-2i\rho_m\cos\alpha & 2\rho_m\sin\alpha
\end{array}\right),
\end{equation}
\end{widetext}
where
$$
\rho=\frac{\kappa_\uparrow+\kappa_\downarrow}{2\tilde k_F},\quad \rho_m=\frac{\kappa_\uparrow-\kappa_\downarrow}{2\tilde k_F}
$$
are the dimensionless measures of the TRS-preserving and TRS-breaking boundary scattering, respectively, satisfying the condition $|\rho_m|<\rho$.
If the boundary is nonmagnetic, i.e., $h=0$ and $\rho_m=0$, then Eq. (\ref{S-matrix-result}) takes the form
\begin{equation}
\label{S-nonmagnetic}
\hat S=\frac{1-i\rho_0}{1+i\rho_0}\left(\begin{array}{cc}
0 & i \\
-i & 0
\end{array}\right),
\end{equation}
where $\rho_0=\sqrt{2m^*(U-\epsilon_F)}/\tilde k_F$.
In particular, in the case of an infinitely strong boundary potential, $U\to\infty$, we have $\hat S=\hat\sigma_2$ (note the difference from the dimensional
reduction of the $S$ matrix for the 2D Rashba model,\cite{KS10} which is due to the different phase choice for the helicity eigenstates).
The vanishing of the diagonal matrix elements in Eq. (\ref{S-nonmagnetic}) means that the backscattering into the same helicity band is forbidden by the TRS.
This property does not in fact depend on the particular model of the boundary, as long as it is nonmagnetic, see Appendix \ref{app: S-offdiagonal}.
If the boundary is magnetic, then there are no symmetry reasons for $S_{--}$ and $S_{++}$ to vanish, but it can happen by accident.
For instance, regardless of the value of $\rho_m$, the diagonal elements in Eq. (\ref{S-matrix-result}) are equal to zero if $\sin\alpha=0$,
which corresponds to $\bgam\parallel\hat{\bm{z}}$. However, one can show that such a configuration of the SO coupling is not protected by the point-group symmetry\cite{Sam17}
and is unstable under a small perturbation of the confining potential.
Therefore, for the parameter $R$, which controls the boundary effects on the ABS spectrum, see Eq. (\ref{R-def}), we obtain that $R=1$ if and only if the boundary is nonmagnetic, otherwise $R>1$.
In particular, for the boundary model (\ref{H-normal}), the $S$ matrix is given by Eq. (\ref{S-matrix-result}) and we obtain:
\begin{equation}
\label{R-final}
R=1+\frac{8\rho_m^2\sin^2\alpha}{(1+\rho^2-\rho_m^2)^2+4\rho_m^2\cos^2\alpha}.
\end{equation}
To summarize, the twofold degenerate ABS zero modes exist only if (i) the bulk superconducting state is real and topologically nontrivial, i.e., $\chi=\pi$,
and (ii) the boundary is nonmagnetic, i.e., $S_{--}=S_{++}=0$ . Violation of either of these two conditions results in the gapping of the ABS.
\section{Stability of the zero modes}
\label{sec: ABS-stability}
To illustrate the effects of the TRS-breaking perturbations on the ABS zero modes, we assume equal gap magnitudes in both helicity bands: $|\Delta_-|=|\Delta_+|=\Delta$, and obtain from Eq. (\ref{ABS-equation})
the following expression: $E=\pm\Delta\sqrt{(R+\cos\chi)/(R+1)}$,
where $R$ is given by Eq. (\ref{R-final}). If the bulk is TR invariant, with $\chi=\pi$, but the boundary scattering is not, then the ABS split and move to nonzero energies, as shown in Fig. \ref{fig: E-rho_m}.
Any deviation of the interband phase difference from $\pi$ produces the same effect, see Fig. \ref{fig: E-chi}.
\begin{figure}
\includegraphics[width=7cm]{E-rho_m.eps}
\caption{The ABS energy as a function of the TRS-breaking boundary scattering, for $|\Delta_+|=|\Delta_-|=\Delta$, $\rho=1$, and $\alpha=\pi/2$.}
\label{fig: E-rho_m}
\end{figure}
\begin{figure}
\includegraphics[width=7cm]{E-chi.eps}
\caption{The ABS energy as a function of the interband phase difference, for a nonmagnetic boundary and $|\Delta_+|=|\Delta_-|=\Delta$.}
\label{fig: E-chi}
\end{figure}
We see that the ABS are pushed apart and away from the zero energy if the mean-field state breaks TRS. It is then natural to ask if the zero modes are stable under TRS-breaking fluctuations.
We consider the superconducting state with arbitrary gap magnitudes, in which $h=0$ and $\chi=\pi$ at the mean-field level. However, both the exchange field (or magnetisation)
and the interband phase difference now experience classical thermal fluctuations:
$$
h\to\delta h(x),\quad \chi\to\pi+\delta\chi(x).
$$
Here $\delta h$ and $\delta\chi$ are Gaussian-distributed random fields with zero average in the left and right half-spaces, respectively.
For long-wavelength fluctuations, the ABS is affected only by the local values of the fluctuating fields at the boundary, $\delta h(0)$ and $\delta\chi(0)$. Then, it follows from Eqs. (\ref{R-final}) and (\ref{ABS-equation})
that $R=1+\delta R$ and
\begin{equation}
\label{E-squared-fluct}
E^2=\frac{2|\Delta_-|^2|\Delta_+|^2}{(|\Delta_-|+|\Delta_+|)^2}\left[\delta R+\frac{1}{2}\delta\chi^2(0)\right],
\end{equation}
where
$$
\delta R=2C^2\frac{\delta h^2(0)}{\epsilon^2_F},\quad C=\frac{\sin\alpha}{\rho_0(1+\rho_0^2)}\left(\frac{k_F}{\tilde k_F}\right)^2.
$$
Introducing the renormalized fluctuating fields
$$
m(x)=C\,\frac{\delta h(x)}{\epsilon_F},\qquad \varphi(x)=\frac{\delta\chi(x)}{2},
$$
and also the dimensionless energy
$$
\varepsilon=\frac{E}{\tilde\Delta},\qquad \tilde\Delta=\frac{2|\Delta_-||\Delta_+|}{|\Delta_-|+|\Delta_+|},
$$
we obtain from Eq. (\ref{E-squared-fluct}):
\begin{equation}
\label{conical}
\varepsilon=\pm\sqrt{m^2(0)+\varphi^2(0)}.
\end{equation}
This form of the ABS energy dependence on the TRS-breaking perturbations is rather generic and can be understood as follows.
The Bogoliubov quasiparticle spectrum is found by diagonalizing the BdG Hamiltonian ${\cal H}$, which incorporates the boundary potential and
can be written as ${\cal H}={\cal H}_{\mathrm{TRI}}+\delta{\cal H}$. The first term here is the TR invariant part, corresponding to $h=0$ and $\chi=0$ or $\pi$ in the model considered above. Its
diagonalization in a topologically nontrivial state ($\chi=\pi$) produces two zero-energy ABS at each end of the superconducting wire.
The $\delta{\cal H}$ term contains all TRS-breaking contributions, both from the bulk superconducting state and from the external fields, including a magnetic boundary.
These contributions are assumed to be small and can therefore be treated by
standard means of the degenerate pertubation theory. In the subspace of the zero modes localized near $x=0$,
the TRS-breaking Hamiltonian is represented by a Hermitian $2\times 2$ matrix $\bm{b}\hat{\bm{\tau}}$, where $\bm{b}$ is real and $\hat{\bm{\tau}}$ are the Pauli matrices,
the unit matrix being absent due to the BdG electron-hole symmetry. The ABS energies are given by $\pm|\bm{b}|$, i.e. the zero-mode energy splitting has a nonanalytical, conical, dependence on the TRS-breaking perturbations,
in agreement with the explicit model calculation, see Eq. (\ref{conical}).
\subsection{ABS density of states}
\label{sec: ABS DOS}
The main quantity of interest for us is the fluctuation-averaged density of states (DoS) of the ABS, which is given by the following expression:
$$
N_{\mathrm{ABS}}(\varepsilon)=\left\langle\delta\left[\varepsilon-\sqrt{m^2(0)+\varphi^2(0)}\right]\right\rangle,
$$
at $\varepsilon\geq 0$, while $N_{\mathrm{ABS}}(-\varepsilon)=N_{\mathrm{ABS}}(\varepsilon)$. Introducing the notation $m(0)=m_0$ and $\varphi(0)=\varphi_0$, we obtain:
\begin{eqnarray}
\label{DOS-PP}
N_{\mathrm{ABS}}(\varepsilon) = \iint_{-\infty}^{\infty}dm_0\,d\varphi_0\,P_m(m_0)P_\varphi(\varphi_0)\nonumber\\
\times\delta(\varepsilon-\sqrt{m_0^2+\varphi_0^2}).
\end{eqnarray}
For analytical simplicity one can assume the Gaussian distributions of the local fluctuating fields:
\begin{eqnarray}
\label{P-local}
&& P_m(m_0)=\frac{1}{\sqrt{2\pi\Gamma_m}}e^{-m_0^2/2\Gamma_m},\nonumber\\ \\
&& P_\varphi(\varphi_0)=\frac{1}{\sqrt{2\pi\Gamma_\varphi}}e^{-\varphi_0^2/2\Gamma_\varphi},\nonumber
\end{eqnarray}
where $\Gamma_m=\langle m^2(0)\rangle$ and $\Gamma_\varphi=\langle\varphi^2(0)\rangle$ characterize the magnitudes of the magnetic and the superconducting phase fluctuations, respectively.
In the absence of fluctuations, i.e., at $\Gamma_m=\Gamma_\varphi=0$, we have $P_m=\delta(m_0)$ and $P_\varphi=\delta(\varphi_0)$, and Eq. (\ref{DOS-PP}) yields
\begin{equation}
\label{DOS-MF}
N_{\mathrm{ABS}}(\varepsilon)=\delta(\varepsilon).
\end{equation}
If only one fluctuation channel is present, for instance, $\Gamma_m\neq 0$ but $\Gamma_\varphi=0$, then
\begin{equation}
\label{DOS-1-channel}
N_{\mathrm{ABS}}(\varepsilon)=\frac{1}{\sqrt{2\pi\Gamma_m}}e^{-\varepsilon^2/2\Gamma_m},
\end{equation}
i.e., the delta-function peak in the DoS at zero energy is broadened.
In the general case, with both fluctuation channels taken into account, we obtain from Eq. (\ref{DOS-PP}):
\begin{eqnarray}
\label{DOS-2-channels}
N_{\mathrm{ABS}}(\varepsilon) = \frac{\varepsilon}{\sqrt{\Gamma_m\Gamma_\varphi}}\exp\left[-\frac{\varepsilon^2}{4}\left(\frac{1}{\Gamma_m}+\frac{1}{\Gamma_\varphi}\right)\right]\nonumber\\
\times I_0\left(\frac{\varepsilon^2}{4}\left|\frac{1}{\Gamma_m}-\frac{1}{\Gamma_\varphi}\right|\right),
\end{eqnarray}
where $I_0(z)$ is the modified Bessel function. Instead of the zero-energy peak, see Eqs. (\ref{DOS-MF}) and (\ref{DOS-1-channel}), the DoS now has a dip at low energies, with
$N_{\mathrm{ABS}}(\varepsilon)\simeq\varepsilon/\sqrt{\Gamma_m\Gamma_\varphi}$ at $\varepsilon\to 0$, and two maxima symmetrically located at
$\varepsilon\sim\sqrt{\Gamma_m\Gamma_\varphi/(\Gamma_m+\Gamma_\varphi)}$, see Fig. \ref{fig: DoS-2}.
The most remarkable feature of the expresssion (\ref{DOS-2-channels}) is that
\begin{equation}
\label{DoS-zero}
N_{\mathrm{ABS}}(\varepsilon=0)=0,
\end{equation}
which means that the ABS zero modes are completely destroyed by the TRS-breaking fluctuations.
It is easy to see that the result (\ref{DoS-zero}) holds true regardless of the details of the
fluctuation distribution, as long as the latter is nonsingular. Indeed, at low energy one can replace the distribution functions in Eq. (\ref{DOS-PP}) by their values
at $m_0=0$ and $\varphi_0=0$ and obtain:
$N_{\mathrm{ABS}}(\varepsilon\to 0)=P_m(0)P_\varphi(0)\iint dm_0\,d\varphi_0\,\delta(\varepsilon-\sqrt{m_0^2+\varphi_0^2})\propto\varepsilon$.
\begin{figure}
\includegraphics[width=7cm]{DoS.eps}
\caption{The fluctuation-averaged ABS density of states, for $\Gamma_m=0.02$ and $\Gamma_\varphi=0.01$.}
\label{fig: DoS-2}
\end{figure}
\section{Conclusions}
\label{sec: Conclusion}
We have calculated the spectrum of the Andreev boundary modes in a half-infinite superconducting wire on a substrate. This system is intrinsically noncentrosymmetric, which means that both the electron band structure and
the Cooper pairing are strongly affected by the SO coupling of the Rashba type. We focused on the effects of two types of TR symmetry-breaking perturbations on the ABS spectrum. While the external magnetic field is assumed to
be zero, TRS can still be broken, either by a magnetic boundary scattering or intrinsically in the superconducting bulk, if the interband phase difference is not an integer multiple of $\pi$. We put the wire in contact with a
ferromagnetic insulator and described the boundary scattering of the Andreev wave functions by the semiclassical $S$ matrix.
We have shown that the symmetry and topology of the bulk Hamiltonian alone does not determine the number of the zero-energy ABS.
Even if the $\mathbb{Z}_2$ topological invariant in the bulk points to the presence of a Kramers pair of the zero modes, the magnetic boundary scattering splits them.
Thus the bulk-boundary correspondence, understood as a relation between the number of the zero modes and some topological characteristics of the bulk Hamiltonian, can break down if the symmetry of
the boundary is lower than that of the bulk.
We have found that the zero modes are very sensitive to a ``virtual'' breaking of TRS. Even if the mean-field superconducting state is TR invariant and topologically nontrivial and the boundary is nonmagnetic,
the TRS-breaking fluctuations lead to a complete vanishing of the fluctuation-averaged ABS DoS at zero energy. The overall shape of the DoS is qualitatively changed by the fluctuations, with a ``pseudogap''
minimum developing at low energies.
\acknowledgments
This work was supported by a Discovery Grant (K. S.) and by a USR Award (B. T.), both from the Natural Sciences and Engineering Research Council of Canada.
|
\section*{Acknowledgements}
\textit{Acknowledgements} GT acknowledges support from FAPESP proc. 2014/13120-7 and CNPQ bolsa de
produtividade 301996/2014-8. LT was
supported by the U.S. Department of Energy, Office of
Science, Office of Nuclear Physics under Award No. DE-SC0004286
and Polish National Science Center Grant DEC-2012/06/A/ST2/00390. DM would like to acknowledge CNPQ graduate fellowship n. 147435/2014-5
Parts of this work were done when LT visited Campinas on FAEPEX fellowship number 2020/16, as well as when GT participated in the INT workshop "Exploring the QCD Phase Diagram through Energy Scans
" We thank FAEPEX and the INT organizers for the support provided.
This work is a part of the project INCT-FNA Proc. No. 464898/2014-5.\\
We wish to thank Miklos Gyulassy for enlightening discussions which posed the conceptual challenges that eventually led to this work, and Mike Lisa for showing us experimental literature and useful discussions.
|
\section{Introduction}\label{S:Introduction}
Wormholes are handles or tunnels in
spacetime that link widely separated
regions of our Universe or different universes
altogether. Morris and Thorne \cite{MT88}
proposed the following line element for the
wormhole spacetime:
\begin{equation}\label{E:line1}
ds^{2}=-e^{2\Phi(r)}dt^{2}+\frac{dr^2}{1-b(r)/r}
+r^{2}(d\theta^{2}+\text{sin}^{2}\theta\,
d\phi^{2}),
\end{equation}
using units in which $c=G=1$. In this line
element, $b=b(r)$ is called the \emph{shape
function} and $\Phi=\Phi(r)$ is called the
\emph{redshift function}, which must be
everywhere finite to avoid an event horizon.
For the shape function we must have
$b(r_0)=r_0$, where $r=r_0$ is the radius of
the \emph{throat} of the wormhole. An important
requirement is the \emph{flare-out condition}
at the throat: $b'(r_0)<1$, while $b(r)<r$
near the throat. The flare-out condition can
only be met by violating the null energy
condition
\begin{equation}
T_{\alpha\beta}k^{\alpha}k^{\beta}\ge 0
\end{equation}
for all null vectors $k^{\alpha}$, where
$T_{\alpha\beta}$ is the stress-energy tensor.
Matter that violates this condition is referred to
as ``exotic" in Ref. \cite{MT88}. In particular,
for the radial outgoing null vector $(1,1,0,0)$,
the violation reads
$T_{\alpha\beta}k^{\alpha}k^{\beta}=\rho+
p_r<0$. Here $T^t_{\phantom{tt}t}=-\rho(r)$,
the energy density, and $T^r_{\phantom{rr}r}=
p_r(r)$, the radial pressure.
($T^\theta_{\phantom{\theta\theta}\theta}=
T^\phi_{\phantom{\phi\phi}\phi}=p_t(r)$,
the lateral pressure.)
In Ref. \cite{MT88}, Morris and Thorne also
discussed the tidal forces that a traveler
would be subjected to and subsequently put
the following constraint on the redshift
function: $\Phi'(r)\le g_{\oplus}/
(c^2\sqrt{1-b(r)/r})$. The feasibility of
this assumption will be discussed later.
The purpose of this paper is to revisit
these requirements by starting with the
modified charged wormhole model due to
Kim and Lee \cite{KL01} and assuming a
noncommutative-geometry background.
Similar issues involving traversable
wormholes in Lovelock gravity are discussed
in Refs. \cite{DD09, ZLP15, MZL15a}, with
emphasis on the energy conditions. In
particular, Ref. \cite{MZL15a} considers
higher-dimensional thin-shell wormholes.
Einstein-Gauss-Bonnet wormholes satisfying
the weak energy condition are studied in
Ref. \cite{MZL15b}, while higher-dimensional
evolving wormholes satisfying the null energy
condition are discussed in Ref. \cite{ZLR14}.
Regarding the tidal constraints, it was noted
in Ref. \cite{pK15} that the zero-tidal-force
assumption is incompatible with quantum field
theory in classical general relativity.
\section{Noncommutative geometry}
An important outcome of string theory is
the realization that coordinates may become
noncommutative operators on a $D$-brane
\cite{eW96, SW99}. The result is a
fundamental discretization of spacetime
due to the commutator
$[\textbf{x}^{\mu},\textbf{x}^{\nu}]
=i\theta^{\mu\nu}$, where $\theta^{\mu\nu}$ is
an antisymmetric matrix. According to Refs.
\cite{SS03, NSS06, mR11}, noncommutativity
replaces point-like objects by smeared
objects, thereby eliminating the divergences
that normally appear in general relativity.
Applications of noncommutative geometry are
numerous and varied. One of the first was to
wormholes in semi-classical gravity by
Garattini and Lobo \cite{GL09}. The same
authors analyzed the stability of gravastars
\cite{LG13}. Gravitational lensing of
wormholes in noncommutative geometry was
studied in Ref. \cite{pK16a}, while the
search for higher-dimensional wormholes
was proposed in Ref. \cite{fR12}.
It seems natural to model the smearing by
means of a Gaussian distribution of minimal
length $\sqrt{\alpha}$ instead of the Dirac
delta function \cite{NSS06,mR11, fR12, pK13}.
A simpler but equally effective way is to
assume that the energy density of the static
and spherically symmetric and particle-like
gravitational source has the form
\begin{equation}\label{E:rho}
\rho(r)=\frac{M\sqrt{\alpha}}
{\pi^2(r^2+\alpha)^2};
\end{equation}
(see Refs. \cite{LL12} and \cite{NM08}.)
The point is that the mass $M$ is diffused
throughout the region of linear dimension
$\sqrt{\alpha}$ due to the uncertainty.
It should be noted that noncommutative
geometry is an intrinsic property of spacetime
and does not depend on particular features
such as curvature.
Next, we note that the Einstein field equations
$G_{\mu\nu}=
8\pi T_{\mu\nu}$ result in the
following forms:
\begin{equation}\label{E:Einstein1}
\rho(r)=\frac{b'}{8\pi r^2},
\end{equation}
\begin{equation}\label{E:Einstein2}
p_r(r)=\frac{1}{8\pi}\left[-\frac{b}{r^3}+
2\left(1-\frac{b}{r}\right)\frac{\Phi'}{r}
\right],
\end{equation}
and
\begin{equation}\label{E:Einstein3}
p_t(r)=\frac{1}{8\pi}\left(1-\frac{b}{r}\right)
\left[\Phi''-\frac{b'r-b}{2r(r-b)}\Phi'
+(\Phi')^2+\frac{\Phi'}{r}-
\frac{b'r-b}{2r^2(r-b)}\right].
\end{equation}
Here $p_r(r)$ is the radial pressure and $p_t(r)$
is the lateral pressure.
The conservation law
$T^{\mu\nu}_{\phantom{\mu\nu};\nu}=0$
implies that only two of Eqs.
(\ref{E:Einstein1})-(\ref{E:Einstein3}) are
independent. So Eq. (\ref{E:Einstein3}) can be
obtained from the other two. (See Ref. \cite{sS05}
for further details.)
As a final comment, to make use of Eq.
(\ref{E:rho}), we can keep the standard form
of the Einstein field equations in the sense
that the Einstein tensor retains its original
form, but the stress-energy tensor is modified
\cite{NSS06}. It follows that the length
scale used need not be restricted to the
Planck scale.
\section{The modified Kim-Lee model}
Since black holes can carry an electric charge,
it is natural to assume that wormholes can do
likewise. So it was proposed by Kim and Lee
\cite{KL01} that for a wormhole with constant
electric charge $Q$, the Einstein field
equations take on the form
\begin{equation}\label{E:EFE}
G^{(0)}_{\mu\nu}+G^{(1)}_{\mu\nu}=8\pi
[T^{(0)}_{\mu\nu}+T^{(1)}_{\mu\nu}].
\end{equation}
Given that the usual form is $G^{(0)}_{\mu\nu}=
8\pi T ^{(0)}_{\mu\nu}$, the modified form in
Eq. (\ref{E:EFE}) is obtained by adding the matter
term $T^{(1)}_{\mu\nu}$ to the right side and
the corresponding back reaction term
$G^{(1)}_{\mu\nu}$ to the left side. The proposed
metric is
\begin{equation}\label{E:line2}
ds^2=-\left(1+\frac{Q^2}{r^2}\right)dt^2
+\left(1-\frac{b(r)}{r}+\frac{Q^2}{r^2}\right)^{-1}dr^2\\
+r^2(d\theta^2+\text{sin}^2\theta\,d\phi^2).
\end{equation}
Observe that whenever $b\equiv 0$, the wormhole
becomes a Reissner-Nordstr\"{o}m black hole,
and if $Q=0$, the spacetime reverts to a Morris-Thorne
wormhole, where the shape function $b=b(r)$ meets
the usual requirements. Kim and Lee go on to
show that the metric, Eq. (\ref{E:line2}), is a
self-consistent solution of the Einstein field
equations. The shape function $b=b(r)$ of the
Morris-Thorne wormhole is now replaced by the
effective shape function
\begin{equation}\label{E:eff}
b_{\text{eff}}(r)=b(r)-\frac{Q^2}{r},
\end{equation}
to be discussed later.
To overcome certain difficulties in determining
the tidal constraints, we will use the following
modified metric already introduced in Refs.
\cite{pK11} and \cite{pK16}:
\begin{equation}\label{E:line3}
ds^2=-\left(1+S(r)+\frac{Q^2}{r^2}\right)dt^2\\
+\left(1-\frac{b(r)}{r}
+\frac{Q^2}{r^2}\right)^{-1}dr^2
+r^2(d\theta^2+\text{sin}^2\theta\,d\phi^2).
\end{equation}
here $S(r)$ is a differentiable function such
that $S(r)>0$ and $S'(r)>0$. As noted in Ref.
\cite{pK11}, line element (\ref{E:line3})
remains a valid solution of the Einstein field
equations.
\section{The modified Kim-Lee model in noncommutative
geometry}
In the Kim-Lee model, Eq. (\ref{E:line2}),
and in the modified model, Eq. (\ref{E:line3}),
\begin{equation}
\text{total charge}=\int\int\int_V\rho_q(r)dV,
\end{equation}
where $\rho_q(r)$ is the charge density.
Adapting this to the noncommutative-geometry
background, we assume that the charge
density has the form
\begin{equation}\label{E:charged}
\rho_q(r)=\frac{Q^2\sqrt{\alpha}}
{\pi^2(r^2+\alpha)^2},
\end{equation}
where $Q^2$ refers to the Kim-Lee model. To
obtain the smeared charge $Q^2_{\alpha}(r)$,
we evaluate
\begin{equation}\label{E:Q}
Q^2_{\alpha}(r)=\int^r_{r_0}4\pi (r')^2
\frac{Q^2\sqrt{\alpha}}
{\pi^2(r^2+\alpha)^2}dr'
=\frac{2Q^2\sqrt{\alpha}}{\pi}
\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r}{\sqrt{\alpha}}-
\frac{r}{r^2+\alpha}\right).
\end{equation}
Observe that $Q^2_{\alpha}(r_0)=0$; also,
seen from a large distance, $Q^2_{\alpha}(r)$
becomes $Q^2$.
Similarly, from Eqs. (\ref{E:Einstein1}) and
(\ref{E:eff}), we get for the effective shape
function,
\begin{multline}\label{E:shape1}
b_{\text{eff}}(r)=\frac{4M\sqrt{\alpha}}{\pi}
\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r}{\sqrt{\alpha}}-\frac{r}{r^2+\alpha}
\right)\\
-\frac{1}{r}\frac{2Q^2\sqrt{\alpha}}
{\pi}\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r}{\sqrt{\alpha}}-\frac{r}{r^2+\alpha}
\right)+C.
\end{multline}
Since $Q^2_{\alpha}(r_0)=0$, we obtain from
$b(r_0)=r_0$,
\begin{equation}\label{E:shape2}
b_{\text{eff}}(r_0)=\frac{4M\sqrt{\alpha}}{\pi}
\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r_0}{\sqrt{\alpha}}-\frac{r_0}{r_0^2+\alpha}
\right)-0+C=r_0
\end{equation}
and hence the constant $C$. Thus
\begin{multline}\label{E:shape3}
b_{\text{eff}}(r)=\frac{4M\sqrt{\alpha}}{\pi}
\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r}{\sqrt{\alpha}}-\frac{r}{r^2+\alpha}
\right)\\
-\frac{1}{r}\frac{2Q^2\sqrt{\alpha}}
{\pi}\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r}{\sqrt{\alpha}}-\frac{r}{r^2+\alpha}
\right)
-\frac{4M\sqrt{\alpha}}{\pi}
\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r_0}{\sqrt{\alpha}}-\frac{r_0}{r_0^2+\alpha}
\right)+r_0.
\end{multline}
For a wormhole solution, we require that
$b'_{\text{eff}}(r_0)>0$ and
$b'_{\text{eff}}(r_0)<1$, the flare-out
condition mentioned earlier. To that end,
we determine
\begin{equation}\label{E:derivative}
b'_{\text{eff}}(r_0)=\frac{4M\sqrt{\alpha}}
{\pi}\frac{2r_0^2}{(r_0^2+\alpha)^2}-
\frac{1}{r_0}\frac{2Q^2\sqrt{\alpha}}{\pi}
\frac{2r_0^2}{(r_0^2+\alpha)^2}=
\frac{4r_0^2\sqrt{\alpha}}
{\pi (r_0^2+\alpha)^2}
\left(2M-\frac{1}{r_0}Q^2\right)
\end{equation}
since $Q^2_{\alpha}(r_0)=0$. Looking
ahead to the traversability conditions
which involve measurements in meters, we will
assume that $r_0>1$ m. So to keep
$b'_{\text{eff}}(r_0)$ positive, we must have
\begin{equation}
Q^2\le 2M.
\end{equation}
Since $\sqrt{\alpha}\ll M$, it also follows
that
\begin{equation}\label{E:flare}
b'_{\text{eff}}(r_0)<1,
\end{equation}
as required. Moreover, $b_{\text{eff}}(r)
<r$ near the throat and
\[
\text{lim}_{r\rightarrow\infty}
\frac{b_{\text{eff}}(r)}{r}=0,
\]
so that the wormhole spacetime is
asymptotically flat.
From the exoticity condition in Ref.
\cite{MT88}, $b'_{\text{eff}}(r_0)<1$
is equivalent to the violation of the null
energy condition, requiring the need for
``exotic matter." According to Eqs.
(\ref{E:derivative}) and (\ref{E:flare}),
however, this violation is simply a consequence
of the noncommutative-geometry background.
To study the effect that the tidal forces
may have on the traveler, we return to the
criterion
\begin{equation}\label{E:Phiprime1}
\Phi'(r)\le\frac{g_{\oplus}}
{c^2\sqrt{1-b(r)/r}}
\end{equation}
mentioned in Sec. \ref{S:Introduction}. At
the throat of the wormhole, this condition
is trivially satisfied. It is subsequently
proposed in Ref. \cite{MT88} that the space
stations be far enough away from the throat
so that $b(r)/r$ becomes negligible. The
result is the requirement
\begin{equation}\label{E:Phiprime2}
\Phi'\lesssim (9.2\times 10^{15}\,\text{m})
^{-1}\approx (10^{16}\,\text{m})^{-1}.
\end{equation}
In the modified Kim-Lee model, Eq.
(\ref{E:line3}),
\[
e^{2\Phi}=1+R(r)+\frac{Q^2}{r^2}
\]
yields
\begin{equation}
\Phi(r)=\frac{1}{2}\text{ln}
\left[1+R(r)+\frac{1}{r^2}
\frac{2Q^2\sqrt{\alpha}}{\pi}
\left(\frac{1}{\sqrt{\alpha}}
\text{tan}^{-1}\frac{r}{\sqrt{\alpha}}
-\frac{r}{r^2+\alpha}
\right)\right]
\end{equation}
and
\begin{equation}\label{E:Phiprime3}
\Phi'(r)=\frac{\frac{1}{2}R'(r)
-\frac{1}{r^3}\frac{2Q^2\sqrt{\alpha}}
{\pi}\left(\frac{1}{\sqrt{\alpha}}
\text{tan}^{-1}\frac{r}{\sqrt{\alpha}}
-\frac{r}{r^2+\alpha}\right)
+\frac{2Q^2\sqrt{\alpha}}
{\pi (r^2+\alpha)^2}}
{1+R(r)+\frac{1}{r^2}
\frac{2Q^2\sqrt{\alpha}}{\pi}
\left(\frac{1}{\alpha}\text{tan}^{-1}
\frac{r}{\sqrt{\alpha}}
-\frac{r}{r^2+\alpha}
\right)}.
\end{equation}
As noted earlier, $R(r)>0$ and $R'(r)>0$,
but it quickly becomes apparent that in view
of Inequality (\ref{E:Phiprime2}), $R(r)$
has to be fine-tuned so that
\begin{equation}\label{E:finetune}
\frac{1}{2}R'(r)-\frac{1}{r^3}
\frac{2Q^2\sqrt{\alpha}}{\pi}
\left(\frac{1}{\sqrt{\alpha}}\text{tan}^{-1}
\frac{r}{\sqrt{\alpha}}
-\frac{r}{r^2+\alpha}\right)
\end{equation}
becomes negligibly small near the proposed
location of the station.
\emph{Remark:} While the need for fine-tuning
may not be desirable, it does not present a
serious conceptual problem here: for large $r$,
expression (\ref{E:finetune}) approaches
\[
\frac{1}{2}R'(r)-\frac{Q^2}{r^3}+
\frac{2Q^2\sqrt{\alpha}}
{\pi r^2(r^2+\alpha)}.
\]
So $\frac{1}{2}R'\approx Q^2/r^3$ near the station
where $R(r)$ has the approximate form $-Q^2/r^2
+C$, $C>0$; $C$ can be chosen so that the
denominator in Eq. (\ref{E:Phiprime3})
exceeds unity. As a result
\[
\Phi'(r)\lesssim \frac{2Q^2\sqrt{\alpha}}
{\pi (r^2+\alpha)^2}.
\]
Eqs. (\ref{E:Phiprime2}) and (\ref{E:Phiprime3})
now yield the condition
\begin{equation}\label{E:tidal}
\ \Phi'(r)\lesssim\frac{2Q\sqrt{\alpha}}
{\pi (r^2+\alpha)^2}<(10^{16}\,\text{m})^{-1}.
\end{equation}
This requirement places the following constraint
on the parameter $\alpha$:
\begin{equation}\label{E:constraint}
\sqrt{\alpha}<\frac{(r^2+\alpha)^2\times 10^{-16}}
{(2Q^2)/\pi}\,\text{m}.
\end{equation}
According to Ref. \cite{NSS06}, noncommutativity
is not visible at presently accessible energies
if $\sqrt{\alpha}<10^{-18}$ m. So according to
Inequality (\ref{E:constraint}), the constraint
on $\alpha$ is much less severe and so would in
principle be observable.
\section{Conclusion}
While Morris and Thorne \cite{MT88} had demonstrated
that wormholes could be macroscopic structures
suitable for interstellar travel, they could not
avoid certain problems with traversability. In
particular, the structures would have to be
designed to produce low tidal forces, while
the unavoidable violation of the null energy
condition called for the use of ``exotic matter."
The latter requirement is not only difficult to
meet classically, but may even be unphysical.
The former was met by assuming a constant
redshift function, so that $\Phi'\equiv 0$,
called the ``zero-tidal-force solution."
To overcome these problems, this paper
discusses the modified wormhole model due
to Kim and Lee in conjunction with a
noncommutative-geometry background. The
violation of the null energy condition can
be attributed to the latter, thereby avoiding
the need for exotic matter. The presence of
$Q^2$ in the modified Kim-Lee model allows
the reduction of the tidal forces to
acceptable levels without invoking the
trivial zero-tidal-force assumption, again
made possible by the noncommutative-geometry
background. Finally, the condition on
$\Phi'$ places a constraint on the value
of $\sqrt{\alpha}$, but one that is well
above the limit of observability.
|
\section{Introduction}
The physical world is filled with constraints. You can open a door, but only if it isn't locked. You can douse a fire, but only if a fire is present. You can throw a rock or drop a rock or even, under certain circumstances, converse with a rock, but you cannot traverse it, enumerate it, or impeach it. The term \textit{affordances}~\cite{gibson77} refers to the subset of possible actions which are feasible in a given situation. Human beings detect these affordances automatically, often subconsciously, but it is not uncommon for autonomous learning agents to attempt impossible or even ridiculous actions, thus wasting effort on futile behaviors.
This paper presents a method for affordance extraction based on the copiously available linguistic information in online corpora. Word embeddings trained using Wikipedia articles are treated as a common sense knowledge base that encodes (among other things) object-specific affordances. Because knowledge is represented as vectors, the knowledge base can be queried using linear algebra.
This somewhat counterintuitive notion - the idea that words can be manipulated mathematically - creates a theoretical bridge between the frustrating realities of real-world systems and the immense wealth of common sense knowledge implicitly encoded in online corpora.
We apply our technique to a text-based environment and show that \textit{a priori} knowledge provided by affordance extraction greatly speeds learning. Specifically, we reduce the agent's search space by (a) identifying actions afforded by a given object; and (b) discriminating objects that can be grasped, lifted and manipulated from objects which can merely be observed. Because the agent explores only those actions which `make sense', it is able to discover valuable behaviors more quickly than a comparable agent using a brute force approach. Critically, the affordance agent is demonstrably able to eliminate extraneous actions without (in most cases) discarding beneficial ones.
\section{Related Work}
Our research relies heavily on word2vec \cite{Mikolov2013}, an algorithm that encodes individual words based on the contexts in which they tend to appear. Earlier work has shown that word vectors trained using this method contain intriguing semantic properties, including structured representations of gender and geography~\cite{mikolov2013b,mikolov2013a}. The (by now) archetypal example of such properties is represented by the algebraic expression $vector[$`$king$'$] - vector[$`$man$'$] + vector[$`$woman$'$] = vector[$`$queen$'$]$.
Researchers have leveraged these properties for diverse applications including sentence- and paragraph-level encoding \cite{Kiros2015,Le2014}, image categorization \cite{Frome2013}, bidirectional retrieval \cite{Karpathy2014}, semantic segmentation \cite{Socher2011}, biomedical document retrieval~\cite{Georgios2016}, and the alignment of movie scripts to their corresponding source texts \cite{Zhu2015}. Our work is most similar to \cite{Zhu2014}; however, rather than using a Markov Logic Network to build an explicit knowledge base, we instead rely on the semantic structure implicitly encoded in skip-grams.
Affordance detection, a topic of rising importance in our increasingly technological society, has been attempted and/or accomplished using visual characteristics \cite{song2011,song2015}, haptic data \cite{navarro2012}, visuomotor simulation \cite{schenck2012,schenck2016}, repeated real-world experimentation \cite{Montesano2007,Stoytchev2008}, and knowledge base representations \cite{Zhu2014}.
In 2001 \cite{laird2001} identified text-based adventure games as a step toward general problem solving. The same year at AAAI, Mark DePristo and Robert Zubek unveiled a hybrid system for text-based game play \cite{arkin1998}, which operated on hand-crafted logic trees combined with a secondary sensory system used for goal selection. The handcrafted logic worked well, but goal selection broke down and became cluttered due to the scale of the environment. Perhaps most notably, in 2015 \cite{Narasimhan2015} designed an agent which passed the text output of the game through an LSTM \cite{hochreiter1997} to find a state representation, then used a DQN \cite{mnih2015} to select a Q-valued action. This approach appeared to work well within a small discrete environment with reliable state action pairs, but as the complexity and alphabet of the environment grew, the clarity of Q-values broke down and left them with a negative overall reward. Our work, in contrast, is able to find meaningful state action pairs even in complex environments with many possible actions.
\section{Wikipedia as a Common Sense Knowledge Base}
Google `knowledge base', and you'll get a list of hand-crafted systems, both commercial and academic, with strict constraints on encoding methods. These highly-structured, often node-based solutions are successful at a wide variety of tasks including topic gisting \cite{liu2004}, affordance detection \cite{Zhu2014} and general reasoning \cite{russ2011}. Traditional knowledge bases are human-interpretable, closely tied to high-level human cognitive functions, and able to encode complex relationships compactly and effectively.
It may seem strange, then, to treat Wikipedia as a knowledge base. When compared with curated solutions like ConceptNet~\cite{liu2004}, Cyc~\cite{matuszek2006}, and WordNet~\cite{miller1995}, its contents are largely unstructured, polluted by irrelevant data, and prone to user error. When used as a training corpus for the word2vec algorithm, however, Wikipedia becomes more tractable. The word vectors create a compact representation of the knowledge base and, as observed by~\cite{Bolukbasi2016} and~\cite{Bolukbasi2016a}, can even encode relationships about which a human author is not consciously cognizant. Perhaps most notably, Wikipedia and other online corpora are constantly updated in response to new developments and new human insight; hence, they do not require explicit maintenance.
\begin{figure}\label{genderplot}
\vskip -0.2in
\centering
\includegraphics[width=\linewidth]{images/flattened_plot_gender.png}
\caption{Word vectors projected into the space defined by $vector[$`$king$'$] - vector[$`$queen$'$]$ and $vector[$`$woman$'$] - vector[$`$man$'$]$. In this projection, masculine and feminine terms are linearly separable.}
\end{figure}
However: in order to leverage the semantic structure implicitly encoded within Wikipedia, we must be able to interpret the resulting word vectors. Significant semantic relationships are not readily apparent from the raw word vectors or from their PCA reduction. In order to extract useful information, the database must be queried through a mathematical process. For example, in Figure 1 a dot product is used to project gendered terms onto the space defined by $vector[$`$king$'$] - vector[$`$queen$'$]$ and $vector[$`$woman$'$] - vector[$`$man$'$]$. In such a projection, the mathematical relationship between the words is readily apparent. Masculine and feminine terms become linearly separable, making it easy to distinguish instances of each group.
These relationships can be leveraged to detect affordances, and thus reduce the agent's search space. In its most general interpretation, the adjective \textit{affordant} describes the set of actions which are \textbf{physically possible} under given conditions. In the following subsections, however, we use it in the more restricted sense of actions which \textbf{seem reasonable}. For example, it is physically possible to eat a pencil, but it does not `make sense' to do so.
\subsection{Verb/Noun affordances}
So how do you teach an algorithm what `makes sense'? We address this challenge through an example-based query. First we provide a canonical set of verb/noun pairs which illustrate the relationship we desire to extract from the knowledge base. Then we query the database using the analogy format presented by~\cite{Mikolov2013}. Using their terminology, the analogy sing:song::[?]:[x] encodes the following question: If the affordant verb for `song' is `sing', then what is the affordant verb for [x]?
In theory, a single canonical example is sufficient to perform a query. However, experience has shown that results are better when multiple canonical values are averaged.
More formally, let $W$ be the set of all English-language word vectors in our agent's vocabulary. Further, let $N = \{\vec{n}_1, ... , \vec{n}_j\}, N \subset W$ be the set of all nouns in $W$ and let $V = \{\vec{v}_1, ... , \vec{v}_k\}, V \subset W$ be the set of all verbs in $W$.
Let $C = \{ (\vec{v}_1, \vec{n}_1), ... , (\vec{v}_m, \vec{n}_m) \}$ represent a set of canonical verb/noun pairs used by our algorithm. We use $C$ to define an affordance vector $\vec{a} = 1/m \sum_i (\vec{v}_i - \vec{n}_i)$, which can be thought of as the distance and direction within the embedding space which encodes affordant behavior.
In our experiments we used the following verb/noun pairs as our canonical set:
\begin{quote}
[`sing song', `drink water', `read book', `eat food', `wear coat', `drive car', `ride horse', `give gift', `attack enemy', `say word', `open door', `climb tree', `heal wound', `cure disease', `paint picture']
\end{quote}
We describe a verb/noun pair $(\vec{v},\vec{n})$ as affordant to the extent that $\vec{n} + \vec{a} = \vec{v}$. Therefore, a typical knowledge base query would return the $n$ closest verbs $\{\vec{v}_{c1}, ... , \vec{v}_{cn}\}$ to the point $\vec{n} + \vec{a}$
For example, using the canonical set listed above and a set of pre-trained word vectors, a query using $\vec{n} =$ vector[`sword'] returns the following:
\begin{quote}
[`vanquish', `duel', `unsheathe', `wield', `summon', `behead', `battle', `impale', `overpower', `cloak']
\end{quote}
Intuitively, this query process produces verbs which answer the question, `What should you do with an [x]?'. For example, when word vectors are trained on a Wikipedia corpus with part-of-speech tagging, the five most affordant verbs to the noun `horse' are \{`gallop', `ride', `race', `horse', `outrun'\}, and the top five results for `king' are \{`dethrone', `disobey', `depose', `reign', `abdicate'\}.
The resulting lists are surprisingly logical, especially given the unstructured nature of the Wikipedia corpus from which the vector embeddings were extracted. Subjective examination suggests that affordances extracted using Wikipedia are at least as relevant as those produced by more traditional methods (see Figure 2).
\begin{figure}\label{bentable}
\begin{center}
\begin{tabular}{ | m{3.5em} m{1.0cm} | m{0.88cm} m{0.7cm} | m{1.09cm} m{0.9cm} | }
\hline
\multicolumn{2}{|l|}{\textbf{Our algorithm}}
& \multicolumn{2}{|l|}{\textbf{Co-occurrence}} &
\multicolumn{2}{|l|}{\textbf{Concept~Net}} \\ \hline \hline
vanquish & impale & have & die & kill & harm \\
duel & battle & make & cut & parry & fence \\
unsheath & behead & kill & fight & strike & thrust \\
summon & wield & move & use & slash & injure \\
overpower & cloak & destroy & be & look~cool & cut \\ \hline
\end{tabular}
\caption{Verb associations for the noun `sword' using three different methods: (1) Affordance detection using word vectors extracted from Wikipedia, as described in this section, (2) Strict co-occurrence counts using a Wikipedia corpus and a co-occurrence window of 9 words, (3) Results generated using ConceptNet's CapableOf relationship.}
\end{center}
\end{figure}
It is worth noting that our algorithm is not resilient to polysemy, and behaves unpredictably when multiple interpretations exist for a given word. For example, the verb `eat' is highly affordant with respect to most food items, but the twelve most salient results for `apple' are \{`apple', `package', `program', `release', `sync', `buy', `outsell', `download', `install', `reinstall', `uninstall', `reboot'\}. In this case, `Apple, the software company' is more strongly represented in the corpus than `apple, the fruit'.
\subsection{Identifying graspable objects}
Finding a verb that matches a given noun is useful. But an autonomous agent is often confronted with more than one object at a time. How should it determine which object to manipulate, or whether any of the objects are manipulable? Pencils, pillows, and coffee mugs are easy to grasp and lift, but the same cannot be said of shadows, boulders, or holograms.
To identify affordant nouns - i.e. nouns that can be manipulated in a meaningful way - we again utilize analogies based on canonical examples. In this section, we describe a noun as \textit{affordant} to the extent that it can be pushed, pulled, grasped, transported, or transformed. After all, it would not make much sense to lift a sunset or unlock a cliff.
We begin by defining canonical affordance vectors $\vec{a}_x = \vec{n}_{x1} - \vec{n}_{x2}$ and $\vec{a}_y = \vec{n}_{y1} - \vec{n}_{y2}$ for each axis of the affordant vector space. Then, for each object $\vec{o}_i$ under consideration, a pair of projections $\vec{p}_{o_{ix}} = \vec{o}_i$ dot $\vec{a}_x$ and $\vec{p}_{o_{iy}} = \vec{o}_i$ dot $\vec{a}_y$.
The results of such a projection can be seen in Figure 3. This query is distinct from those described in section 3.1 because, instead of using analogies to test the relationships between nouns and verbs, we are instead locating a noun on the spectrum defined by two other words.
In our experiments, we used a single canonical vector, $vector[$`$\mathit{forest}$'$]$ - $vector[$`$tree$'$]$, to distinguish between nouns of different classes. Potentially affordant nouns were projected onto this line of manipulability, with the word whose projection lay closest to `tree' being selected for further experimentation.
Critical to this approach is the insight that canonical word vectors are most effective when they are thought of as exemplars rather than as descriptors. For example, $vector[$`$\mathit{forest}$'$]$ $-$ $vector[$`$tree$'$]$ and $vector[$`$building$'$]$ $-$ $vector[$`$brick$'$]$ function reasonably well as projections for identifying manipulable items. $vector[$`$big$'$]$ $-$ $vector[$`$small$'$]$, on the other hand, is utterly ineffective.
\begin{figure}\label{manipulate}
\vskip -0.2in
\centering
\includegraphics[width=\linewidth]{images/manipulability.png}
\caption{Word vectors projected into the space defined by $vector[$`$\mathit{forest}$'$] - vector[$`$tree$'$]$ and $vector[$`$mountain$'$] - vector[$`$pebble$'$]$. Small, manipulable objects appear in the lower-left corner of the graph. Large, abstract, or background objects appear in the upper right. An object's manipulability can be roughly estimated by measuring its location along either of the defining axes.}
\end{figure}
\begin{algorithm}
\caption{Noun Selection With Affordance Detection}
\begin{algorithmic}[1]
\scriptsize
\STATE $state$ = game response to last command
\STATE $manipulable\_nouns \gets \{\}$
\FOR{each word $w \in state$}
\IF {$w$ is a noun}
\IF {$w$ is manipulable}
\STATE add $w$ to $manipulable\_nouns$
\ENDIF
\ENDIF
\ENDFOR
\STATE $noun$ = a randomly selected noun from $manipulable\_nouns$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{Verb Selection With Analogy Reduction}
\begin{algorithmic}[1]
\scriptsize
\STATE $navigation\_verbs$ = [`north', `south', `east', `west', `northeast', `southeast', `southwest', `northwest', `up', `down', `enter']
\STATE $manipulation\_verbs$ = a list of 1000 most common verbs
\STATE $essential\_manipulation\_verbs$ = [`get', `drop', `push', `pull', `open', `close']
\STATE $\mathit{affordant}\_verbs$ = verbs returned by Word2vec that match $noun$
\STATE $\mathit{affordant}\_verbs$ = $\mathit{affordant}\_verbs $ $\cap$ \\ $manipulation\_verbs$
\STATE $final\_verbs$ = $navigation\_verbs $ $\cup$ $\mathit{affordant}\_verbs $ $\cup $ \\ $essential\_manipulation\_verbs$
\STATE $verb$ = a randomly selected verb from $final\_verbs$
\end{algorithmic}
\end{algorithm}
\section{Test Environment: A World Made of Words}
In this paper, we test our ideas in the challenging world of text-based adventure gaming.
Text-based adventure games offer an unrestricted, free-form interface: the player is presented with a block of text describing a situation, and must respond with a written phrase. Typical actions include commands such as: `examine wallet', `eat apple', or `light campfire with matches'. The game engine parses this response and produces a new block of text. The resulting interactions, although syntactically simple, provide a fertile research environment for natural language processing and human/computer interaction. Game players must identify objects that are manipulable and apply appropriate actions to those objects in order to make progress.
In these games, the learning agent faces a frustrating dichotomy: its action set must be large enough to accommodate any situation it encounters, and yet each additional action increases the size of its search space. A brute force approach to such scenarios is frequently futile, and yet factorization, function approximation, and other search space reduction techniques bring the risk of data loss. We desire an agent that is able to clearly perceive all its options, and yet applies only that subset which is likely to produce results.
In other words, we want an agent that explores the game world the same way a human does: by trying only those actions that `make sense'. In the following sections, we show that affordance-based action selection provides a meaningful first step towards this goal.
\begin{figure}
\centering
\includegraphics[width=70mm]{images/ZorkText2.png}
\caption{Sample text from the adventure game Zork. Player responses follow a single angle bracket.}
\end{figure}
\subsection{Learning algorithm}
\begin{figure*}
\label{final_results}
\vskip -0.2in
\centering
\includegraphics[width=\linewidth]{images/FINAL_RESULTS_1000_epochs_1000_steps_games_with_points.png}
\caption{Learning trajectories for sixteen Z-machine games. Agents played each game 1000 times, with 1000 game steps during each trial. No agent received any reward on the remaining 32 games. 10 data runs were averaged to create this plot.}
\end{figure*}
Our agent utilizes a variant of Q-learning \cite{watkins1992}, a reinforcement learning algorithm which attempts to maximize expected discounted reward.
Q-values are updated according to the equation
\begin{equation}
\Delta Q(s,a) = \alpha(R(s,a) + \gamma max_{a}Q(s',a) - Q(s,a))
\end{equation}
where $Q(s,a)$ is the expected reward for performing action $a$ in observed state $s$, $\alpha$ is the learning rate, $\gamma$ is the discount factor, and $s'$ is the new state observation after performing action $a$. Because our test environments are typically deterministic with a high percentage of consumable rewards, we modify this algorithm slightly, setting $\alpha=1$ and constraining Q-value updates such that
\begin{equation}
Q'(s,a) = max(~Q(s,a),~Q(s,a) + \Delta Q(s,a)~)
\end{equation}
This adaptation encourages the agent to retain behaviors that have produced a reward at least once, even if the reward fails to manifest on subsequent attempts. The goal is to prevent the agent from `unlearning' behaviors that are no longer effective during the current training epoch, but which will be essential in order to score points during the next round of play.
The agent's state representation is encoded as a hash of the text provided by the game engine. Actions are comprised of verb/object pairs:
\begin{equation}
a = v +~`~~\textit{'}~+ o, v \epsilon V, o \epsilon O
\end{equation}
where $V$ is the set of all English-language verbs and $O$ is the set of all English-language nouns. To enable the agent to distinguish between state transitions and merely informational feedback, the agent executes a `look' command every second iteration and assumes that the resulting game text represents its new state. Some games append a summary of actions taken and points earned in response to each `look' command. To prevent this from obfuscating the state space, we stripped all numerals from the game text prior to hashing.
Given that the English language contains at least 20,000 verbs and 100,000 nouns in active use, a naive application of Q-learning is intractable. Some form of action-space reduction must be used. For our baseline comparison, we use an agent with a vocabulary consisting of the 1000 most common verbs in Wikipedia, an 11-word navigation list and a 6-word essential manipulation list as depicted in Algorithm 2. The navigation list contains words which, by convention, are used to navigate through text-based games. The essential manipulation list contains words which, again by convention, are generally applicable to all in-game objects.
The baseline agent does not use a fixed noun vocabulary. Instead, it extracts nouns from the game text using part-of-speech tags. To facilitate game interactions, the baseline agent augments its noun list using adjectives that precede them. For example, if the game text consisted of `You see a red pill and a blue pill', then the agent's noun list for that state would be [`pill', `red pill', `blue pill']. (And its next action is hopefully `swallow red pill').
In Sections 5.1 and 5.2 the baseline agent is contrasted with an agent using affordance extraction to reduce its manipulation list from 1000 verbs to a mere 30 verbs for each state, and to reduce its object list to a maximum of 15 nouns per state. We compare our approach to other search space reduction techniques and show that the \textit{a priori} knowledge provided by affordance extraction enables the agent to achieve results which cannot be paralleled through brute force methods. All agents used epsilon-greedy exploration with a decaying epsilon.
The purpose of our research was to test the value of affordance-based search space reduction. Therefore, we did not add augmentations to address some of the more challenging aspects of text-based adventure games. Specifically, the agent maintained no representation of items carried in inventory or of the game score achieved thus far. The agent was also not given the ability to construct prepositional commands such as `put book on shelf' or `slay dragon with sword'.
\section{Results}
We tested our agent on a suite of 50 text-based adventure games compatible with Infocom's Z-machine. These games represent a wide variety of situations, ranging from business scenarios like `Detective' to complex fictional worlds like `Zork: The Underground Empire'.
Significantly, the games provide little or no information about the agent's goals, or actions that might provide reward.
During training, the agent interacted with the game engine for 1000 epochs, with 1000 training steps in each epoch. On each game step, the agent received a positive reward corresponding to the change in game score. At the end of each epoch the game was restarted and the game score reset, but the agent retained its learned Q-values.
Our affordance-based search space reduction algorithms enabled the agent to score points on 16/50 games, with a peak performance (expressed as a percentage of maximum game score) of 23.40\% for verb space reduction, 4.33\% for object space reduction, and 31.45\% when both methods were combined. The baseline agent (see Sec.~4.1) scored points on 12/50 games, with a peak performance of 4.45\%. (Peak performance is defined as the maximum score achieved over all epochs, a metric that expresses the agent's ability to comb through the search space and discover areas of high reward.)
Two games experienced termination errors and were excluded from our subsequent analysis; however, our reduction methods outperformed the baseline in both peak performance and average reward in the discarded partial results.
Figures 5 and 7 show the performance of our reduction techniques when compared to the baseline. Affordance-based search space reduction improved overall performance on 12/16 games, and decreased performance on only 1 game.
Examination of the 32 games in which no agent scored points (and which are correspondingly not depicted in Figures 5 and 7) revealed three prevalent failure modes: (1) The game required prepositional commands such as `look at machine' or `give dagger to wizard', (2) The game provided points only after an unusually complex sequence of events, (3) The game required the user to infer the proper term for manipulable objects. (For example, the game might describe `\textbf{something} shiny' at the bottom of a lake, but required the agent to `get shiny \textbf{object}'.) Our test framework was not designed to address these issues, and hence did not score points on those games. A fourth failure mode (4) might be the absence of a game-critical verb within the 1000-word manipulation list. However, this did not occur in our coarse examination of games that failed.
\subsection{Alternate reduction methods}
\begin{figure}
\begin{center}
\begin{tabular}{ | l | l | }
\hline
\textbf{Affordant selection} & \textbf{Random selection} \\ \hline \hline
decorate glass & continue quantity \\
open window & break sack \\
add table & result window \\
generate quantity & stay table \\
ring window & build table \\
weld glass & end house \\
travel passage & remain quantity \\
climb staircase & discuss glass \\
jump table & passage \\
\hline
\end{tabular}
\caption{Sample exploration actions produced by a Q-learner with and without affordance detection. The random agent used nouns extracted from game text and a verb list comprising the 200 most common verbs in Wikipedia.}
\end{center}
\end{figure}
We compared our affordance-based reduction technique with four other approaches that seemed intuitively applicable to the test domain. Results are shown in Figure 7.
\textbf{Intrinsic rewards}: This approach guides the agent's exploration of the search space by allotting a small reward each time a new state is attained. We call these awards intrinsic because they are tied to the agent's assessment of its progress rather than to external events.
\textbf{Random reduction}: When applying search space reductions one must always ask: `Did improvements result from my specific choice of reduced space, or would \textit{any} reduction be equally effective?' We address this question by randomly selecting 30 manipulation verbs to use during each epoch.
\textbf{ConceptNet reduction}: In this approach we used ConceptNet's CapableOf relation to obtain a list of verbs relevant to the current object. We then reduced the agent's manipulation list to include only words that were also in ConceptNet's word list (effectively taking the intersection of the two lists).
\textbf{Co-occurrence reduction}: In this method, we populated a co-occurrence dictionary using the 1000 most common verbs and 30,000 most common nouns in Wikipedia. The dictionary tracked the number of times each verb/noun pair occurred within a 9-word radius. During game play, the agent's manipulation list was reduced to include only words which exceeded a low threshold (co-occurrences \textgreater~3).
Figure 7 shows the performance of these four algorithms, along with a baseline learner using a 1000-word manipulation list. Affordance-based verb selection improved performance in most games, but the other reduction techniques fell prey to a classic danger: they pruned precisely those actions which were essential to obtain reward.
\subsection{Fixed-length vocabularies vs. Free-form learning}
An interesting question arises from our research. What if, rather than beginning with a 1000-word vocabulary, the agent was free to search the entire English-language verb space?
A traditional learning agent could not do this: the space of possible verbs is too large. However, the Wikipedia knowledge base opens new opportunities. Using the action selection mechanism described in Section 4.1, we allowed the agent to construct its own manipulation list for each state (see Section 3.1). The top 15 responses were unioned with the agent's navigation and essential manipulation lists, with actions selected randomly from that set.
A sampling of the agent's behavior is displayed in Figure
6, along with comparable action selections from the baseline
agent described in Section 4.1. The free-form learner is able
to produce actions that seem, not only reasonable, but also
rather inventive when considered in the context of the game
environment. We believe that further research in this direction
may enable the development of one-shot learning for text-based
adventure games.
\section{Conclusion}
The common sense knowledge implicitly encoded within
Wikipedia opens new opportunities
for autonomous agents. In this paper we have shown that
previously intractable search spaces can be efficiently navigated
when word embeddings are used to identify context-dependent
affordances. In the domain of text-based adventure games, this approach is superior
to several other intuitive methods.
Our initial experiments have
been restricted to text-based environments, but the underlying
principles apply to any domain in which mappings can be
formed between words and objects. Steady advances in object recognition and semantic segmentation, combined
with improved precision in robotic systems, suggests
that our methods are applicable to systems including
self-driving cars, domestic robots, and UAVs.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{images/Figure_7_with_legend.png}
\caption{Five verb space reduction techniques compared over 100 exploration epochs. Average of 5 data runs. Results were normalized for each game based on the maximum reward achieved by any agent.}
\end{figure}
\section{Acknowledgements}
Our experiments were run using Autoplay:~a learning environment for interactive fiction (https://github.com/-danielricks/autoplay). We thank Nvidia, the Center for Unmanned Aircraft Systems, and Analog Devices, Inc. for their generous support.
\pagebreak
{\small
\bibliographystyle{named}
|
\section{Introduction}
\bigskip
This paper considers multidimensional reflected stochastic differential equations in a sublinear expectation space. These equations are driven by a new type of Brownian motion associated with a sublinear expectation, which are introduced by Peng \cite{Peng3, Peng5, Peng4} during the past decade. In the seminal works of Peng, he has established a framework of nonlinear It\^o's calculus and related stochastic analysis which does not rely on a single probability measure. This theory provides basic tools to discuss problems in finance with Knigntian uncertainty and in robust statistics.
Moreover, the nonlinear Feynman-Kac formula obtained in \cite{Peng4, HJPS14b} provides an probabilistic representation of fully nonlinear parabolic PDEs via forward-backward systems. \\
In \cite{Peng4}, the so-called $G$-Brownian motion is defined as a continuous process with stationary independent increments. Under the associated $G$-expectation, these increments are subject to $G$-normal distribution with volatility uncertainty between two bounds. According to Denis et al. \cite{DHP}, the $G$-expectation can be regarded as an upper expectation based on a collection of non-dominated martingale measures $\mathcal{P}_G$. Furthermore, a Choquet capacity associated with such a collection can be defined. This leads to a notion of equivalence between two random variables in the sublinear expectation space --- ``quasi-sure'' (q.s.). Instead of ``almost-surely'' in the classical probability theory, we say a property holds quasi-surely if it holds outside a null set for the referred capacity. In the present paper, we shall examine the following equation in the quasi-sure sense,
\begin{equation}
\left\{
\begin{aligned}
& X_{t}=x_{0}+\int_{0}^{t}f\left(
s,X_{s}\right) ds+\int_{0}^{t}h\left( s,X_{s}\right) d\langle B, B\rangle_{s}+\int_{0}^{t}g\left( s,X_{s}\right) dB_{s}+K_{t},~~~0\leq t\leq T;\\
& K_{t} =\int_{0}^{t}\mathbf{n}_s d\left\vert K\right\vert _{s};~~~
\left\vert K\right\vert _{t}=\int_{0}^{t}\mathbf{1}%
_{\left\{ X_{s} \in \partial D\right\} }d\left\vert
K\right\vert _{s},~~~~\text{q.s.}
\end{aligned}
\right.\label{a}
\end{equation}
where $B$ is a $d$-dimensional $G$-Brownian motion; $\langle B, B\rangle$ is the covariance matrix of $B$; $X$ is a process reflecting on the boundaries of domain $\overline{D}$ and
$K$ is a bounded variations process with variation $|K|$ increasing only when $X\in \partial\overline{D}$. \\
In the classical framework, reflected stochastic differential equations driven by Brownian motion have been extensively studied by many authors. Among them, Skorokhod \cite{Sko1961, Sko1962} is the first who introduced diffusion processes with
reflecting boundaries in the 1960s. Later on, reflecting diffusions in a half-space have been investigated by Watanabe \cite{watanabe}, El Karoui \cite{Elk1975}, Yamada \cite{Yamada}, El Karoui and Chleyat-Maurel \cite{Elk1978}, El Karoui et al. \cite{Elk1980}, etc. The study of multi-dimensional stochastic differential equations on a general domain dates back to Stroock and Varadhan \cite{stroock}, in which the existence and uniqueness of weak solutions have been proved when the domain is smooth. Afterwards, solutions of such equations has been built on a convex domain by a direct method in Tanaka \cite{Tanaka}, whereas
Menaldi \cite{menaldi} and Lions et al. \cite{lions} have adopted a penalization method to construct them. Concerning the reflecting problem with a non-convex but ``admissible'' domain, Lions and Sznitman \cite{LionsetSznitman} have first solved the deterministic Skorokhod problem and have applied this result to construct pathwisely an iteration sequence in order to approximate the corresponding reflecting diffusion. The results of \cite{LionsetSznitman} has been later improved in Saisho \cite{Saisho1} and in Saisho and
Tanaka \cite{SaishoetTanaka} by removing the admissibility condition on the domain. \\
In the context of sublinear expectation, we shall discuss the above mentioned equation (\ref{a}), in which a newly defined It\^o stochastic integral with respect to $G$-Brownian motion $\int g dB$ appears in the dynamic. As its counterpart in the classical theory, adapted to Peng's method, this integral is first defined for simple processes and could subsequently be extended to $M^p_G$ due to the $G$-BDG type estimate, where $M^p_G$ is a normed space of processes with slightly additional regularity (see also \cite{hwz}). Thanks to this extension of the It\^o type integral, the notion stochastic differential equations driven by $G$-Brownian motion is brought to this nonlinear stochastic analysis framework. Under the Lipschitz condtions, forward equations are studied by Peng \cite{Peng4} and Gao \cite{Gao}; backward equation are examined in \cite{HJPS14a}. Moreover, scalar $G$-diffusion processes with reflection has been considered by Lin in \cite{Lin13} and the multidimensional problem are solved in Lin \cite{LIN} by the penalization method similar to \cite{menaldi}.\\
The main objective of this paper is to generalize the results of \cite{Lin13, LIN} to the multidimensional case when the reflecting boundary is not necessarily convex. We adopt the same assumptions as \cite{Saisho1, SaishoetTanaka} on the domain and, as the first step, we restrict ourself to the deterministic Skorokhod problems concerning $\alpha$-H\"older continuous paths. Precisely, we revise the estimate for the penalization sequence obtained in \cite{SaishoetTanaka} by introducing the $\alpha$-H\"older coefficient. Since the $G$-Brownian motion is supported on $\mathcal{C}^{0, \alpha}$ ($\alpha<1/2$), this estimates are accordingly applied to prove that the multidimensional $G$-Brownian motion in this non-convex domain can be approximated in the Banach space $M^p_G$ by a sequence of solutions of Lipschitz equations studied in \cite{Gao}. Similar arguments also apply for reflected $G$-diffusion with bounded generators. Finally, we pathwisely construct an iteration sequence by using the deterministic result and conduct a fixed-point argument as \cite{LionsetSznitman} to prove the wellposedness of (\ref{a}) under the bounded and Lipschitz assumption on coefficients. We remark that the equation (\ref{a}) under consideration can also be examined in a weaker ``quasi-sure'' sense, which means it holds $\mathbb{P}$-almost surely for all $\mathbb{P}\in \mathcal{P}_G$. This notion is adopted by Soner et al. for establishing a similar nonlinear stochastic analysis framework -- second order backward stochastic differential equations (cf. \cite{soner}). We could proceed almost the same procedures of the present paper under each $\mathbb{P}\in \mathcal{P}_G$ and obtain a solution $(X^\mathbb{P}, K^\mathbb{P})$ which could be later aggregated to $(X, K)$ by Nutz \cite{nutz}. \\
The paper is organized as follows. In Section 2, we introduce
preliminaries in the framework of $G$-expectation which are necessary for the remainder of this paper. In addition, we revisit the deterministic Skorohod problems in a non-convex domain. In Section 3, we present our main results and Section 4 is devoted to prove the main results.
\bigskip
\section{Preliminaires}
In this section, we shall briefly introduce the $G$-expectation framework. Moreover, we shall discuss the deterministic Skorohod problem in non-convex domains and the sufficient conditions for its solvability.
\subsection{$G$-Brownian motion and $G$-expectation}
Adapting to Peng's framework, we first recall useful notations and results on the $G$-expectation and related $G$-It\^o type stochastic calculus.
The reader interested
in the more detailed description on this topic is referred to
Denis et al. \cite{DHP}, Gao \cite{Gao}, Li and Peng \cite{LIetPeng} and Peng \cite{Peng4}.\\
Let $\Omega$ be the space of all $\mathbb{R}^{d}$-valued
continuous paths with $\omega_0 =0$, noted by $\mathcal{C}_0([0, \infty); \mathbb{R}^d),$
equipped
with the distance
$$
\rho(\omega^1, \omega^2):=\sum^\infty_{N=1} 2^{-N} ((\max_
{t\in[0,N]} | \omega^1_t-\omega^2_t|) \wedge 1),
$$
and $(B_t)_{t\geq 0}$ be the canonical process, i.e., $B_t(\omega):=\omega_t$. For each $t\in \lbrack0,\infty)$, we list the following notations:
\begin{itemize}
\item[$\bullet$] $\Omega_{t}:=\{ \omega_{\cdot \wedge t}:\omega \in \Omega \}$; $\mathcal{F}_{t}:=\mathcal{B}(\Omega_{t})$;
\item[$\bullet$] $L^{0}(\Omega)$: the space of all $\mathcal{B}(\Omega)$-measurable real functions;
\item[$\bullet$] $L_{ip}(\Omega_t):=\{\varphi(B_{t_1}, \cdots, B_{t_n}):n\geq
1,\ 0\leq t_1< \cdots< t_n\leq t,\ \varphi \in C_{b,lip}
(\mathbb{R}^{d\times n})\}$;
\item[$\bullet$] $L_{ip}(\Omega):=\{\varphi(B_{t_1}, \cdots, B_{t_n}):n\geq
1,\ 0\leq t_1< \cdots< t_n\leq \infty,\ \varphi \in C_{b,lip}
(\mathbb{R}^{d\times n})\}$;
\end{itemize}
where $C_{b,lip}
(\mathbb{R}^{d\times n})$ is the collection of all bounded Lipschitz functions on $\mathbb{R}^{d\times n}$.\\
We fix a sublinear continuous and monotone function $G$: $\mathbb{S}^d\rightarrow \mathbb{R}$. For some bounded and closed subset $\Gamma\subset \mathbb{R}^d$, this function can be represented by
$$
G(A_0)=\sup_{Q\in \Gamma}\left\{\frac{1}{2}{\rm tr}\left[A_0Q\ltrans{Q}\right]\right\}, \ \ \ \ {\rm for}\ A_0\in \mathbb{S}^d.
$$
The related $G$-expectation on $(\Omega, L_{ip}(\Omega))$ can be constructed in the following way: for each $\xi\in L_{ip}(\Omega)$ of the form
$$\xi=\varphi(B_{t_1},B_{t_2}-B_{t_1},\cdots,B_{t_n}-B_{t_{n-1}}), \ \ 0\leq t_1<t_2<\cdots<t_n,$$
define
$$\mathbb{E}_G[\xi]:=u_1(0,0),$$
where $u_1(0,0)\in\mathbb{R}$ is obtained by recurrence: for $k=n,\cdots,1$, define $u_k:=u_k(t,x;x_1,\cdots,x_{k-1})$, which is a function in $(t,x)$ parameterized by $(x_1,\cdots,x_{k-1})\in\mathbb{R}^{d\times(k-1)}$, by the solution of the following $G$-heat equation defined on $[t_{k-1},t_k)\times\mathbb{R}^d$:
$$\frac{\partial u_k}{\partial t}-G(D^2 u_k)=0,$$
with the terminal condition
$$u_k(t_k,x;x_1,\cdots,x_{k-1})=u_{{k+1}}(t_{k},x;x_1,\cdots,x_{k-1},x).$$
In particular, $u_n(t_n,x;x_1,\cdots,x_{n-1}):=\varphi(x_1,\cdots,x_{n-1},x)$. We say that the canonical process $(B_t)_{t\geq 0}$ is a $G$-Brownian motion under this sublinear expectation $\mathbb{E}_G[\cdot]$, which is with stationary, independent and $G$-Gaussian distributed increments (see Definition 1.4 and 1.8 in Chap. I of \cite{Peng4} for the definition of $G$-Gaussian distribution and see Definition 3.10 in Chap. I of \cite{Peng4} for the definition of independence under sublinear expectation).\\
We denote by $L_G^p(\Omega)$ (resp. $L_G^p(\Omega_T)$ the completion of $Lip(\Omega)$ (resp. $Lip(\Omega_T)$) with respect to the norm $||\cdot||_p:=\mathbb{E}_G[|\cdot|^p]^{\frac{1}{p}}$, for $p\geq 1$.
We can extend the domain of $G$-expectation $\mathbb{E}_G[\cdot]$ from $L_{ip}(\Omega)$ to $L^{0}(\Omega)$ by the procedure introduced in \cite{DHP},
i.e., constructing an upper expectation $\mathbb{E}[\cdot]$:
\begin{equation*}
\mathbb{E}[X]:=\sup_{\mathbb{P}\in\mathcal{P}_G}{\mathbf E}^\mathbb{P}[X],\ X\in L^0(\Omega),
\end{equation*}
where $\mathcal{P}_G$ is a weakly compact family of martingale measures on $(\Omega,\mathcal{B}(\Omega))$. This upper expectation
coincides with the $G$-expectation $\mathbb{E}_G[\cdot]$ on $L_{ip}(\Omega)$ and thus, on its completion $L^1_G(\Omega)$. Naturally, the Choquet capacity related to the upper expectation can be defined by
$$
\bar{C}(A):=\sup_{\mathbb{P}\in\mathcal{P}_G}\mathbb{P}(A),\ A\in\mathcal{B}(\Omega),
$$
and the notation of ``quasi-surely'' (q.s.) can be introduced as follows:
\begin{definition}[Quasi-sure]
A set $A\in\mathcal{B}(\Omega)$ is called polar if $\bar{C}(A)=0.$ A property
is said to hold quasi-surely if it holds outside a polar
set.
\end{definition}
The following Markov's inequality holds in the context of the upper expectation and the related Choquet capacity (Lemma 13 in \cite{DHP}).
\begin{lemma}[Markov's inequality]\label{MI}
Let $X\in L^0(\Omega)$ satisfying $\mathbb{E}[|X|^p]<\infty$, for $p>0$. Then, for each $a>0$,
$$
\bar{C}(\{|X|>a\})\leq \frac{\mathbb{E}[|X|^p]}{a^p}.
$$
\end{lemma}
We also have a generalized Fatou's lemma (cf. e.g. Lemma 2.11 in Bai and Lin \cite{Bai-Lin}) in the $G$-framework.
\begin{lemma}[Fatou's lemma]\label{fatulem}
Assume that $\{X^n\}_{n\in\mathbb{N}}$ is a sequence in $L^0(\Omega)$ and that for a $Y\in L^0(\Omega)$ satisfying $\mathbb{E}[|Y|]<\infty$ and for all $n\in\mathbb{N}$, $X^n \geq Y$, q.s., then
$$
\mathbb{E}[\liminf_{n\rightarrow\infty}X^n]\leq\liminf_{n\rightarrow\infty}\mathbb{E}[X^n].\\
$$
\end{lemma}
In \cite{Peng4}, Peng introduce the It\^o type stochastic integral with respect to the $G$-Brownian motion by first considering the simple process space:
\begin{align*}
M^0([0,T]; \mathbb{R}) & =\{ \eta:\eta_{t}(\omega)=\sum_{i=0}^{n-1}\xi_{i}%
(\omega)\mathbf{1}_{[t_{i},t_{i+1})}(t),\\
&\ {\rm where}\ n\in \mathbb{N}^*,\ 0=t_{0}<\cdots<t_{n}=T,\ \xi_{i}\in Lip%
(\Omega_{t_{i}}),\ i=0,\cdots,n-1\}.
\end{align*}
\begin{definition}
For $p\geq 1$, we denote by $M_G^p([0,T];\mathbb{R})$ the completion of $M^0([0,T]; \mathbb{R})$ under the following norm:
\begin{equation*}
\|\eta\|_p:=\bigg(\mathbb{E}\bigg[\frac{1}{T}\int_0^T|\eta_t|^p dt\bigg]\bigg)^{1/p}.
\end{equation*}
\end{definition}
Here below is the definition of the $G$-It\^o type sintegral. In the sequel, $B^{\bf a}$ denotes the inner product of ${\bf a}\in \mathbb{R}^d$ and $B$, which is still a $G$-Brownian motion, and $\sigma_{\bf a\ltrans{a}}:=\mathbb{E}[({\bf a}, B_1)^2]$.
\begin{definition}\label{hao}
For each $\eta\in M^0([0, T]; \mathbb{R})$, we define the It\^o type integral
$$
\mathcal{I}_{[0, T]}(\eta)=\int^T_0\eta_tdB^{\bf a}_t:=\sum^{N-1}_{k=0}\xi_k(B^{\bf a}_{t_{k+1}}-B^{\bf a}_{t_k}).
$$
Then, thanks to $G$-It\^o's inequality (cf. Lemma 3.4 in \cite{LIetPeng}), this linear mapping $\mathcal{I}_{[0, T]}$ on $M^0([0, T]; \mathbb{R})$ can be continuously extended to $\mathcal{I}_{[0, T]}: M^2_G([0, T];\mathbb{R})\rightarrow L^2_G(\Omega_T)$ and for each $\eta \in M^2_G([0, T]; \mathbb{R})$, we define $\int^T_0\eta_tdB^{\bf a}_t:=\mathcal{I}_{[0, T]}(\eta)$.
\end{definition}
Moreover, we have the following BDG type inequality (cf. Theorem 2.1 in \cite{Gao}). Define
$$\sigma_{\mathbf{a}\ltrans{\mathbf{a}}}:=\sup_{Q\in\Gamma}{\rm tr}(Q \ltrans{Q}\mathbf{a}\ltrans{\mathbf{a}}).$$
\begin{lemma}\label{th 2.1.1}Let $p\geq2$, $\mathbf{a}\in\mathbb{R}^d$, $\eta\in M_G^p([0,T]; \mathbb{R})$ and $0\leq s\leq t\leq T$. Then,
$$
\bar{\mathbb{E}}\bigg[\sup_{s\leq u\leq
t}\bigg|\int^u_s\eta_rdB^{\mathbf{a}}_r\bigg|^p\bigg]\leq C_p\sigma^{p/2}_{\mathbf{a}\ltrans{\mathbf{a}}} \mathbb{E}\left[\left(\int^t_s |\eta_u|^2du\right)^{p/2}\right],
$$
where $C_p>0$ is a constant independent of $\mathbf{a}, \
\eta$ and $\Gamma.$
\end{lemma}
In the $G$-expectation framework the quadratic variation process $\langle B^{\bf a}\rangle$ is no longer deterministic, which is
formulated by
$$
\langle B^\mathbf{a}\rangle_t:=\lim_{\mu(\pi^N_{[0, T]})\rightarrow 0}\sum^{N-1}_{k=0}(B^\mathbf{a}_{t^N_{k+1}}-B^\mathbf{a}_{t^N_k})^2=(B^\mathbf{a}_t)^2-2\int^t_0B^\mathbf{a}_sdB^\mathbf{a}_s,
$$
where $\pi^N_{[0, T]}$ is a partition of $[0,T]$, i.e., $\pi^N_{[0, T]}=\{t_0, t_1,\ldots,t_N\}$ such that
$0=t_0<t_1<\ldots<t_N=T$, and $\mu(\pi^N_{[0, T]}):=\max_{1\leq i\leq N}|t^N_i-t^N_{i-1}|$.
For two given vectors $\mathbf{a}$, $\bar{\mathbf{a}}\in\mathbb{R}^d$, the mutual variation process of $B^\mathbf{a}$ and
$B^{\bar{\mathbf{a}}}$ is defined by $$\langle
B^\mathbf{a},B^{\bar{\mathbf{a}}} \rangle_t:=\frac{1}{4}(\langle
B^{\mathbf{a}+\bar{\mathbf{a}}}\rangle_t-\langle
B^{\mathbf{a}-\bar{\mathbf{a}}}\rangle_t).$$
By Corollary 5.7 in Chapter III of Peng \cite{Peng4},
for each $0\leq s\leq t\leq T,$
\begin{equation}\label{pquar}
\langle B^\mathbf{a}\rangle_t-\langle B^\mathbf{a}\rangle_s\leq \sigma_{\mathbf{a}\ltrans{\mathbf{a}}}(t-s).
\end{equation}
Let $B^i$ denote the $i$th coordinate of the $G$-Brownian motion $B$ and set $\langle B, B\rangle_t=(\langle B^i, B^j\rangle_t)_{i, j= 1, \ldots, d}$. Thus, the path of $\langle B, B\rangle$ quasi-surely has a bounded density and indeed the stochastic integral for $\eta\in M^1_G([0, T]; \mathbb{R})$ with respect to $\langle B^{\bf a}, B^{\bar{\bf a}}\rangle$ could be defined pathwisely.\\
Finally, we recall that Gao \cite{Gao} proves the $G$-It\^{o} type integral $X_\cdot=\int_0^\cdot\eta_sdB^{\bf a}_s$ has a continuous $\bar{C}$-modification, for any $\eta\in M_G^2([0,T]; \mathbb{R})$.
\subsection{Conditions on the domain}
In order to investigate the reflected $G$-Brownian motion in this paper, we shall first recall the results in \cite{SaishoetTanaka} for the deterministic Skorohod problem in a domain $D\subset\mathbb{R}^d$, $d\in \mathbb{N}^*$. In that paper, the following conditions are assumed:\\
\textbf{CONDITION (A).} For $x\in \partial D$, we denote
$$\mathcal{N}_{x,r}=\{\mathbf{n}\in \mathbb{R}^d: |\mathbf{n}|=1, B(x-r\mathbf{n},r)\cap
D=\emptyset\},\ r>0\quad \mbox{and}\quad \mathcal{N}_{x}=%
\bigcup_{r>0}\mathcal{N}_{x,r},$$
where $B(z, r):=\{y\in \mathbb{R}^d: |y-z|<r|\}$, for $z\in \mathbb{R}^d$.
We assume that there exists a constant $r_{0}>0$ such that $\mathcal{N}_{x}=%
\mathcal{N}_{x,r_{0}}\neq \emptyset$, for all $x\in \partial D$.\\
\textbf{CONDITION (B).} Assume that there exist constants $\delta >0$ and $%
\beta \in \lbrack 1,\infty )$ such that for any $x\in \partial
D$, we can find a unit vector $l_{x}$ such that
\begin{equation*}
\langle l_{x},\mathbf{n}\rangle \geq \frac{1}{\beta },\text{ \ \ for
all \ \ }\mathbf{n\in }\bigcup_{y\in B(x,\delta )\cap \partial D}\mathcal{N}%
_{y},
\end{equation*}%
where $\langle \cdot, \cdot\rangle$ denotes the usual inner product in $\mathbb{R}^d$.\\
Throughout this paper, we consider a domain
$D\subset \mathbb{R}^{d}$ satisfying both Condition (A) and (B). For each
$x\in \mathbb{R}^{d}$ such that $\mbox{dist}(x,\overline{D})<r_{0}$, there exists a unique
$\overline{x}\in \overline{D}$
with
$\left\vert x-\overline{x}\right\vert =\mbox{dist}(x,\overline{D})$. If $x\notin \overline{D}$
we have $\overline{x}\in \partial D$ and $\frac{\overline{x}-x}{\left\vert \overline{x%
}-x\right\vert }\in \mathcal{N}_{\overline{x}}$ (see e.g. Remark 1.3 in \cite{Saisho1}). We keep this notation $\overline{x}$ for the projection of $x$ on $D$ in the remainder of this paper.\\
To consider the solvability of reflected multi-dimensional Skorohod stochastic differential equations on the domain $D$, we assume furthermore \\
\textbf{CONDITION (C).} There exists a bounded function
$\Psi \in \mathcal{C}_{b}^{2}\left( \mathbb{R}%
^{d}\right) $ whose first and second derivatives are also bounded, and there exists a $\delta'>0$, such that
\begin{equation*}
\forall x\in \partial D,~\forall y\in \overline{D},~\forall \mathbf{n\in }%
\mathcal{N}_{x},~\left\langle y-x,\mathbf{n}\right\rangle +\frac{1}{\delta'}%
\left\langle \nabla \Psi \left( x\right) ,\mathbf{n}\right\rangle \left\vert
y-x\right\vert ^{2}\geq 0.
\end{equation*}
We note the bound of $\Psi$ and its derivatives by $L_\Psi$. This condition is critical for proving the a priori estimate (\ref{est10}), which is similar to its analogue in \S 3 of \cite{LionsetSznitman}.
\subsection{Deterministic Skorohod problem}
We recall here the solvability result of deterministic Skorohod problems in the domain $D$ satisfying Conditions $(A)$ and (B), which could be found in \cite{Saisho1}. This result is our starting point of this paper. \\
Assume that $\phi$ is a continuous function taking values in $\mathbb{R}^{d}$ and that $\phi$ is of bounded variation over each finite interval. We denote by
$|\phi |_{t}$ the total variation of $\phi$ over $[0, t]$, i.e.,
\begin{equation*}
\left\vert \phi \right\vert _{t}:=\sup_{0=t_{1}<t_{2}<\cdots
<t_{n}=t, n\in \mathbb{N}} \sum_{k=1}^{n}|\phi _{t_{k}}-\phi
_{t_{k-1}}|.
\end{equation*}%
We also note
$\left\vert\phi \right\vert _{t}^{s}:=|\phi |_{t}-|\phi |_{s},$ $0\leq s\leq t$.\\
For a continuous function $w$ defined on $[0, T]$, $T>0$, taking values in $\mathbb{R}^{d}$
with
$w(0)=0$ and for $x_0\in \overline{D}$,
we consider the Skorohod equation below:
\begin{equation}
\xi _t=x_0+w_t+\phi_t,~~~t\in \left[ 0,T\right] . \label{1.3}
\end{equation}%
\begin{definition}
We call a couple of functions
$\left( \xi ,\phi
\right)$
solution of (\ref{1.3}), if it satisfies (\ref{1.3}) and the following conditions:
\begin{enumerate}[(i)]
\item The function $\xi $ is continuons and takes values in $\overline{D}$;
\item The function $\phi $ is continuous and takes values in $\mathbb{R}^{d}$ with $\phi (0)=0$. Moreover, it is of bounded variation over $[0, T]$ and for all $t\in \left[ 0,T\right]$,
\begin{eqnarray*}
\phi _t &=&\int_{0}^{t}\mathbf{n}_sd\left\vert \phi \right\vert _{s}; \\
|\phi |_{t} &=&\int_{0}^{t}{\bf 1}_{\{\xi _s\in\partial D\}}d\left\vert \phi
\right\vert _{s},
\end{eqnarray*}%
where $\mathbf{n}_s\in \mathcal{N}_{\xi _s}$, if $\xi _s\in \partial D$.
\end{enumerate}
\end{definition}
\begin{theorem}\label{det1}
(Theorem 4.1 in \cite{Saisho1}) Suppose that the domain $D\subset \mathbb{R}^d$ is open and satisfies Conditions (A) and (B). Then there exists a unique solution $\left( \xi ,\phi \right) $ for the deterministic Skorohod problem (\ref{1.3}).
\end{theorem}
\section{Main results}
In this section, we present our main results on the reflected $G$-Brownian motion and on reflected stochastic differential equations driven by $G$-Brownian motion.
\subsection{Reflected $G$-Brownian motion}
We replace the deterministic function $w$ in the Skorohod problem (\ref{1.3}) by the $G$-Brownian motion $B$ and establish the following equation in the ``quasi-sure'' sense:
\begin{equation}
X_{t}=x_{0}+B_{t}+K_{t},~~x_0\in \overline{D},~~0\leq t\leq T \label{1.4}.
\end{equation}%
\begin{definition}\label{defrgb}
We call a couple of processes $\left( X,K\right) $ solves the Skorohod problem for the $G$-Brownian motion %
(\ref{1.4}), if there exists a polar set $A$, such that
\begin{enumerate}[(i)]
\item The processes $X$ et $K$ belong to $M_{G }^{2}\left( \left[ 0,T%
\right]; \mathbb{R}^{d}\right) $, and for all $\omega\in A^c$,
\begin{equation*}
X_{t}(\omega)=x_{0}+B_{t}(\omega)+K_{t}(\omega),~~0\leq t\leq T;
\end{equation*}
\item For all $\omega\in A^c$, $X(\omega)$ is continuous and takes values in $\overline{D}$;
\item For all $\omega\in A^c$, $K(\omega)$ is continuous and takes values in $\mathbb{R}^{d}$ with $K_0(\omega)=0$. Moreover, $K(\omega)$ is of bounded variation over $[0, T]$ and for all $t\in \left[ 0,T\right]$,
\begin{eqnarray*}
K_{t}\left( \omega \right) &=&\int_{0}^{t}\mathbf{n}_s\left(\omega\right) d\left\vert K\right\vert _{s}\left( \omega \right);\\
\left\vert K\right\vert _{t}\left( \omega \right) &=&\int_{0}^{t}\mathbf{1}%
_{\left\{ X_{s}\left( \omega \right) \in \partial D\right\} }d\left\vert
K\right\vert _{s}\left( \omega \right) ,
\end{eqnarray*}
where $\mathbf{n}_s\left(\omega\right)\in \mathcal{N}_{X_{s}\left(
\omega \right)}$, if $X_{s}\left(
\omega \right)\in \partial D$.
\end{enumerate}
In addition, we call $X$ reflected $G$-Brownian motion on the domain $D$.
\end{definition}
We have the following existence and uniqueness theorem for the reflected $G$-Brownian motion. The proof of this theorem is postponed to the next section.
\begin{theorem}\label{trgb}
Suppose that the domain $D\subset \mathbb{R}^d$ is open and satisfies Conditions (A) and (B). Then there exists a couple $\left( X ,K \right)\in \left(M^2_G([0, T]; \mathbb{R}^d)\times M^2_G([0, T]; \mathbb{R}^d)\right)$ which solves the Skorohod problem (\ref{1.4}) whenever $x_0\in \overline{D}$. Moreover, if the problem (\ref{1.4}) admits two solutions $(X, K)$ and $(X', K')$, then the exists a polar set $\widetilde A$, such that for all $\omega\in \widetilde A^c$,
$$
X(\omega)=X'(\omega)~~~~and~~~K(\omega)=K'(\omega),\ \ 0\leq t\leq T.
$$
\end{theorem}
\subsection{Reflected stochastic differential equations driven by $G$-Brownian motion}
In addition to the reflected $G$-Brownian motion, we shall study reflected stochastic differential equations driven by $G$-Brownian motion, which is formulated as
\begin{align}
X_{t}=x_{0}+\int_{0}^{t}f\left(
s,X_{s}\right) ds&+\int_{0}^{t}h^{ij}\left( s,X_{s}\right) d\langle B^i, B^j\rangle_{s}\notag\\&+\int_{0}^{t}g^{j}\left( s,X_{s}\right) dB^j_{s}+K_{t},~~~0\leq t\leq T,~~~~\text{q.s.}. \label{e1}
\end{align}
Here we adopt the Einstein summation convention.
In (\ref{e1}), the process $\langle B, B\rangle$ is the covariation matrix of the $d$-dimensional $G$-Brownian motion $B$. In what follows, we assume that the functions $f$, $h$, $g$ satisfy the following conditions:
\begin{assumption}
The functions $f$, $h^{i j}$, $g^{j}: \Omega\times [0, T]\times \overline{D}\longrightarrow \mathbb{R}^d$, $i$, $j=1, 2, \ldots d$, are functions such that \\
(\textbf{H1}) For all $x\in \overline{D}$, the processes $f(\cdot, x)$, $h^{i j}(\cdot, x)$, $g^{j}(\cdot, x)$ belong to
$M_G^{2}\left( \left[
0,T\right]; \mathbb{R}^d \right)$;\\
(\textbf{H2}) The functions $f$, $h^{i j}$, $g^{j}$ are uniformly bounded by $L_0$ and uniformly $L_0$-Lipschitz, i.e., there exists a constant $L_0>0$ such that for all $(\omega, t)\in \Omega\times [0, T]$,
\begin{equation*}
\|f\left( t,x\right) -f\left( t,y\right)\|+\left\|
h^{ij}\left( t,x\right) -h^{ij}\left( t,y\right) \right\| + \left\|
g^{j}\left( t,x\right) -g^{j}\left( t,y\right) \right\|\leq L_0\left\|
x-y\right\|,~~~~~\forall x,\ y\in \overline{D},
\end{equation*}
where $\|\cdot\|$ denotes the Hilbert-Schmidt norm for matrices.
\end{assumption}
\begin{definition}
We call a couple of processes $\left( X,K\right) $ solves the Skorohod stochastic differential equation (\ref{e1}),
if there exists a polar set $A$, such that
\begin{enumerate}[(i)]
\item The processes $X$ et $K$ belong to $M_G^{2}\left( \left[ 0,T%
\right]; \mathbb{R}^{d}\right) $ and satisfies (\ref{e1});
\item For all $\omega\in A^c$, $X(\omega)$ takes values in $\overline{D}$;
\item For all $\omega\in A^c$, $K(\omega)$ takes values in $\mathbb{R}^{d}$ with $K_0(\omega)=0$. Moreover, $K(\omega)$ is of bounded variation over $[0, T]$ and for all $t\in \left[ 0,T\right]$,
\begin{eqnarray*}
K_{t}\left( \omega \right) &=&\int_{0}^{t}\mathbf{n}_s\left(\omega\right) d\left\vert K\right\vert _{s}\left( \omega \right);\\
\left\vert K\right\vert _{t}\left( \omega \right) &=&\int_{0}^{t}\mathbf{1}%
_{\left\{ X_{s}\left( \omega \right) \in \partial D\right\} }d\left\vert
K\right\vert _{s}\left( \omega \right) ,
\end{eqnarray*}
where $\mathbf{n}_s\left(\omega\right)\in \mathcal{N}_{X_{s}\left(
\omega \right)}$, if $X_{s}\left(
\omega \right)\in \partial D$.
\end{enumerate}
\end{definition}
Using a fixed point type argument, we shall prove in the next section the following existence and uniqueness theorem for the Skorohod stochastic differential equation (\ref{e1}).
\begin{theorem}\label{trgsde}
Suppose that the domain $D\subset \mathbb{R}^d$ is open and satisfies Conditions (A) and (B). Then there exists a unique couple $\left( X ,K \right)\in \left(M^2_G([0, T]; \mathbb{R}^d)\times M^2_G([0, T]; \mathbb{R}^d)\right)$ which solves the Skorohod stochastic differential equation (\ref{e1}) whenever $x_0\in \overline{D}$ and the coefficients $f$, $h$, $g$ satisfy Assumptions (H1) and (H2).
\end{theorem}
\section{Proofs}
In this section, we shall prove Theorem \ref{trgb} and \ref{trgsde}. First, we recall the results for deterministic Skorohod problem in Saisho and Tanaka \cite{SaishoetTanaka} and provide an estimate when the function $w$ is $\alpha$-H\"older continuous, $\alpha\in (0, 1/2)$.
\subsection{Estimates for the deterministic Skorohod problem}
In (\ref{1.3}), we assume in addition that $w$ is an $\alpha$-H\"older continuous function on $[0, T]$, where $\alpha\in (0, 1/2)$, i.e.,
\begin{equation*}
\left\Vert w \right\Vert _{\alpha }=\sup_{0\leq s<t\leq T}\frac{%
\left\vert w _{t}-w _{s}\right\vert }{\left\vert t-s\right\vert
^{\alpha }}<\infty.
\end{equation*}%
Set
\begin{equation*}
\Delta _{s,t}(w ):=\sup \left\{\left\vert w_{t_{2}}-w
_{t_{1}}\right\vert :s\leq t_{1}<t_{2}\leq t\right\},
\end{equation*}%
and $\Vert w\Vert_T:=\sup\{|w|_t: 0\leq t\leq T\}$.\\
We recall the penalization method in \cite{LionsetSznitman, SaishoetTanaka} and define a sequence of equations: for $m\in \mathbb{N}^*$,
\begin{equation}
\xi_{t}^{m}=x_{0}+w_{t}-\frac{m}{2}\int_{0}^{t}\nabla U(\xi_{s}^{m})ds, \label{1.5}
\end{equation}
where $x_0\in \overline{D}$ and $U$ is a function satisfying (a) $U\in \mathcal C^{1}(\mathbb{R}^{d})$ and $U\geq 0$; (b) $U(x)=\left\vert x-\overline{x}\right\vert ^{2}$, if ${\rm dist}(x,\overline{D})\leq r_{0}/2$ ($r_{0}$ is from Condition (A)); (c) $\nabla U$ is bounded and Lipschitz. Indeed, the existence of such function $U$ is ensured by Condition (A) and (B). We denote by $2L$ the Lipschitz constant of $U$. The equation (\ref{1.5}) admits a unique solution $\xi^m$ which is continuous on $[0, T]$.\\
For each $m\in \mathbb{N}^*$, we define
\begin{equation}\label{eps}
\varepsilon
_{m}^{\alpha }\left( w \right) :=\frac{12e^{L}}{m^{\alpha }}\left\Vert
w \right\Vert _{\alpha }.
\end{equation}%
For $\varepsilon >0$, we note
\begin{equation*}
D_{\varepsilon }:=\left\{ x\in \mathbb{R}^{d},~{\rm dist}\left( x,\overline{D}%
\right) <\varepsilon \right\} .
\end{equation*}
Additionally, set
\begin{equation*}
\widetilde{\xi}_{t}^{m} :=\xi_{t/m}^{m},
\end{equation*}
it is easy to verify that
\begin{equation*}
\widetilde{\xi}_{t}^{m} =x_{0}+\widetilde{w}_{t/m} -\frac{1}{2}\int_{0}^{t}\nabla U(\widetilde{\xi}_{s}^{m} )ds,~~~0\leq t\leq mT.
\end{equation*}%
Similar to Lemma 4.2 in \cite{SaishoetTanaka}, we have the lemma below.
\begin{lemma}
\label{l4.2} Assume that the domain $D\subset\mathbb{R}^d$ is open and satisfies Conditions (A) and (B). For each $m\in \mathbb{N}^{\ast }$ such that $0<\varepsilon _{m}^{\alpha }\left( w \right)
<r_{0}/2$ and for all $m'\geq m$, if there exists $u\in (0, m'T)$ such that
$\widetilde{\xi}_{u}^{m'} \in \partial D_{\varepsilon
_{m}^{\alpha }\left( w \right) /2}$, then $\{\widetilde{\xi}_{t}^{m'}\}_{u\leq t\leq T}$ hits $\partial D_{\varepsilon
_{m}^{\alpha }\left( w\right) /3}$ before hitting $\partial
D_{\varepsilon _{m}^{\alpha }\left( w \right) }$.
\end{lemma}
\begin{proof}
This lemma can be proved by slightly modifying the proof of Lemma 4.2 in \cite{SaishoetTanaka}. Precisely, we could verify that for the given $\varepsilon
_{m}^{\alpha}$, the number $m$ itself is large enough to ensure Lemma 4.2 in \cite{SaishoetTanaka} if $\varepsilon_m^\alpha(w)$ is defined as (\ref{eps}). For the convenience of the reader, we briefly prove this lemma.\\
Fix $m'\geq m$ and suppose that $\widetilde{\xi}_{u}^{m'} \in \partial D_{\varepsilon
_{m}^{\alpha }\left( w \right) /2}$. Consider the auxiliary equation:
\begin{equation*}
\eta _t=\widetilde{\xi}_{u}^{m'} -\frac{1}{2}%
\int_{u}^{t}\nabla U(\eta _s)ds,~~~t\geq u.
\end{equation*}%
From Lemma 4.1 in \cite{SaishoetTanaka}, the above equation is solved by
\begin{equation*}
\eta _t=\overline{\widetilde{\xi}_{u}^{m'}}-\left(\overline{%
\widetilde{\xi}_{u}^{m'}}-\widetilde{\xi}_{u}^{m'}\right)\exp \{-(t-u)\},
\end{equation*}%
and the function $\eta$ satisfies
that for all $t\geq u$,
\begin{equation*}
\left\vert \overline{\eta _t}-\eta _t\right\vert =\frac{\varepsilon
_{m}^{\alpha }(w)}{2}\exp \{-(t-u)\}.
\end{equation*}%
Then, we denote
\begin{equation*}
u^{\prime }: =\inf \left\{t>u:\eta _t\in \partial D_{\varepsilon _{m}^{\alpha
}\left(w \right) /4}\right\},
\end{equation*}%
and it is obvious that $u^{\prime }=u+\log 2<u+1$. \\
On the other hand, for $u\leq t\leq m'T$, we have
\begin{eqnarray}
\widetilde{\xi}_{t}^{m'} -\eta _t = w_{t/m'} -w_{u/m'}-\frac{1}{2}%
\int_{u}^{t}\left(\nabla U(\widetilde{\xi}_{s}^{m'} )-\nabla
U(\eta _s)\right) ds,\notag
\end{eqnarray}
which implies
\begin{eqnarray}
\left\vert \widetilde{\xi}_{t}^{m'}-\eta _t\right\vert
\leq \left\vert {w}_{t/m'} -w_{u/m'} \right\vert +L\int_{u}^{t}\left|\widetilde{\xi}_{s}^{m'} -\eta _s\right|ds. \label{46}
\end{eqnarray}
\begin{itemize}
\item If $u^{\prime }\leq m'T$, then for all $u\leq t\leq u^{\prime
}$, we can deduce that
\begin{equation*}
\left\vert {w}_{t/m'} -w_{u/m'} \right\vert \leq \frac{\left\Vert
w \right\Vert _{\alpha }}{{m'}^{\alpha }}\leq \frac{\left\Vert
w \right\Vert _{\alpha }}{{m}^{\alpha }}=\frac{\varepsilon
_{m}^{\alpha }\left( w \right)}{12e^L}.
\end{equation*}%
We apply Gronwall's lemma to (\ref{46}) and obtain
\begin{equation*}
\left\vert \widetilde{\xi}_{t}^{m'} -\eta _t\right\vert \leq \frac{\varepsilon
_{m}^{\alpha }\left( w \right)}{12e^L} e^{L(t-u)}<\frac{\varepsilon
_{m}^{\alpha }\left( w \right) }{12},~~~u\leq t\leq u^{\prime }.
\end{equation*}%
Therefore, for $u\leq t\leq u^{\prime }$,%
\begin{eqnarray*}
\left\vert \overline{\widetilde{\xi}_{t}^{m'}}-\widetilde{\xi}%
_{t}^{m'}\right\vert \leq \left\vert \overline{\eta
_t}-\eta _t\right\vert +\left\vert \widetilde{\xi}_{t}^{m'}-\eta _t \right\vert
<\frac{\varepsilon _{m}^{\alpha }\left( w \right) }{2}+\frac{%
\varepsilon _{m}^{\alpha }\left( w \right) }{12}
<\varepsilon _{m}^{\alpha }\left( w \right),
\end{eqnarray*}%
whereas
\begin{eqnarray}\notag
\left\vert \overline{\widetilde{\xi}_{u^{\prime }}^{m'} }-%
\widetilde{\xi}_{u^{\prime }}^{m'}\right\vert \leq
\left\vert \overline{\eta_{u^{\prime }}}-\eta_{u^{\prime }}\right\vert
+\left\vert \widetilde{\xi}_{u'}^{m'}-\eta_{u'} \right\vert
<\frac{\varepsilon _{m}^{\alpha }\left( w \right) }{4}+\frac{%
\varepsilon _{m}^{\alpha }\left( w \right) }{12}
=\frac{\varepsilon _{m}^{\alpha }\left( w \right) }{3},
\end{eqnarray}%
which implies that $\{\widetilde{\xi}_{t}^{m'}\}_{u\leq t\leq u'}$ hits $\partial D_{\varepsilon
_{m}^{\alpha }\left( w\right) /3}$ before hitting $\partial
D_{\varepsilon _{m}^{\alpha }\left( w \right) }$.
\item If $u^{\prime }>m'T$, then we could repeat the procedure above to prove that for $u\leq t\leq m'T$, $\left\vert \overline{\widetilde{\xi}_{t}^{m'}}-\widetilde{\xi}%
_{t}^{m'}\right\vert <\varepsilon _{m}^{\alpha }\left( w \right)$, which implies that $\{\widetilde{\xi}_{t}^{m'}\}_{u\leq t\leq m'T}$ never hits $\partial
D_{\varepsilon _{m}^{\alpha }\left( w \right) }$.
\end{itemize}
\end{proof}
We now give a proposition which is a straightforward corollary of Lemma \ref{l4.2}. The proof of this proposition is omitted and we refer the reader to Proposition 4.1 in \cite{SaishoetTanaka}.
\begin{proposition}
\label{p4.1}
Assume that the domain $D\subset\mathbb{R}^d$ is open and satisfies Conditions (A) and (B). For each $m\in \mathbb{N}^{\ast }$ such that $0<\varepsilon _{m}^{\alpha }\left( w \right)
<r_{0}/2$ and for all $m'\geq m$,
\begin{equation*}
\xi_{t}^{m'} \in D_{\varepsilon _{m}^{\alpha }\left(
w \right) },\quad\quad 0\leq t\leq T.
\end{equation*}
\end{proposition}
In particular, the results in the remainder of this subsection is based on the fact that if $m\in \mathbb{N}^{\ast }$ is large enough such that $0<\varepsilon _{m}^{\alpha }\left( w \right)
<r_{0}/2$ then
\begin{equation*}
\xi_{t}^{m} \in D_{\varepsilon _{m}^{\alpha }\left(
w \right) },\quad\quad 0\leq t\leq T.
\end{equation*}
For $m$, $n\in \mathbb{N}^*$, set
\begin{eqnarray*}
T_{m,0} &=&\inf \left\{t\geq 0:\overline{\xi_{t}^{m} }\in
\partial D\right\}; \notag \\
t_{m,n} &=&\inf \left\{ t>T_{m,n-1}:\left\vert \overline{\xi_{t}^{m}}-\overline{\xi_{T_{m,n-1}}^{m} }%
\right\vert \geq \delta /2\right\}; \label{0.7} \\
T_{m,n} &=&\inf \left\{t\geq t_{m,n}:\overline{\xi_{t}^{m} }%
\in \partial D\right\},\notag
\end{eqnarray*}
where the constant $\delta$ is from Condition (B).
Moreover, denote by
$$
\phi^m_t:=-\frac{m}{2}\int_{0}^{t}\nabla U(\xi_{s}^{m})ds,~~0\leq
t\leq T.$$
In the remainder of this subsection, we shall provide an estimate of $|\phi^m|^0_T$. First, we prove that for sufficient large $m$, there is a lower bounded for $T_{m, n}-T_{m, n-1}$.\\
For simplicity, note $\gamma:=\frac{2\kappa^2(r_0/2)\beta}{r_0}$ and $\lambda(w):=\exp\left\{\gamma\left(\Vert w\Vert_T+\delta\right)\right\}$, where $\kappa(r_0/2)$ is a constant from Lemma 2.1 in \cite{SaishoetTanaka} such that
for all $|x-\overline{x}|<r_0/2$ and $|y-\overline{y}|<r_0/2$, $|\overline{x}-\overline{y}|\leq \kappa(r_0/2)|x-y|$. Obviously, $\lambda> 1$.
\begin{lemma}
Assume that the domain $D\subset \mathbb{R}^d$ is open and satisfies Conditions (A) and (B). For each $m\in \mathbb{N}^{\ast }$ such that $0<\varepsilon _{m}^{\alpha }\left( w \right)
<\frac{\delta}{180\beta\lambda(w)} \wedge r_{0}/2$ and for each $n\geq 1$ such that $T_{m, n}<\infty$,
$$
\left|T_{m, n}-T_{m, n-1}\right|\geq h:=\left(\frac{\delta}{36\beta\lambda(w)\Vert w\Vert_\alpha}\right)^{1/\alpha},
$$
\end{lemma}
\begin{proof} From Lemma 5.3 in \cite{SaishoetTanaka}, we have
\begin{align}
\Delta_{s, t}(\xi^m)\leq \left(8\beta\lambda(\omega)+1\right) \left(\Delta_{s, t}(w)+\varepsilon^\alpha_{m}(w)\right).\label{est98}
\end{align}
Then,
from Proposition \ref{p4.1} and the definitions of $t_{m, n}$ and $T_{m, n}$, we have
\begin{align*}
\frac{\delta }{2}-2\varepsilon _{m}^{\alpha }\left( w \right)&=\left\vert \overline{\xi_{t_{m,n}}^{m}%
}-\overline{\xi_{T_{m,n-1}}^{m} }
\right\vert -2\varepsilon _{m}^{\alpha }\left( w \right)\\
&\leq \left\vert \xi_{t_{m,n}}^{m} -\xi_{T_{m,n-1}}^{m} \right\vert\\
&\leq 9\beta\lambda(w) \left(\Delta_{T_{m,n}, T_{m, n-1}}(w)+\varepsilon^\alpha_{m}(w)\right),
\end{align*}%
which implies
$$
\Delta_{T_{m,n}, T_{m, n-1}}(w)\geq \frac{\delta-4\varepsilon^\alpha_m}{18\beta\lambda(w)}-\varepsilon^\alpha_m\geq \frac{\delta}{18\beta\lambda(w)}-5\varepsilon^\alpha_m\geq \frac{\delta}{36\beta\lambda(w)}.
$$
Furthermore, we deduce the desired result by the definition of $||w||_\alpha$.
\end{proof}
\begin{proposition}\label{prop44}
Assume that the domain $D\subset \mathbb{R}^d$ is open and satisfies Conditions (A) and (B). For each $m\in \mathbb{N}^{\ast }$ such that $0<\varepsilon _{m}^{\alpha }\left( w \right)
<\frac{\delta}{180\beta\lambda(w)} \wedge r_{0}/2$, we have
\begin{equation}
\left\vert \phi^m \right\vert^0 _{T} \leq C_0\left(\left\Vert w
\right\Vert _{\alpha }^{1+1/\alpha }+\left\Vert w
\right\Vert _{\alpha }\right)\exp \left\{ \gamma \left( 1+ 1/\alpha\right) \left\Vert w\right\Vert
_{T}\right\} , \label{5.1}
\end{equation}%
where $C_0$ depends only on $\alpha$, $\beta$, $L$, $\delta$ and $T$.
\end{proposition}
\begin{proof}
For any $s$, $t$, such that $T_{m, n-1}\leq s\leq t\leq T_{m, n}$, we know from Lemma 5.1 in \cite{SaishoetTanaka} that
$$
\left|\phi^m\right|^s_t\leq \beta\left(\Delta_{s, t}(\xi^m)+\Delta_{s, t}(w)\right),
$$
We combine this inequality with (\ref{est98}) and deduce
$$
\left|\phi^m\right|^s_t\leq 10\beta^2\exp\left\{\gamma\left(\Vert w\Vert_T+\delta\right)\right\} \left(\Delta_{s, t}(w)+\varepsilon^\alpha_{m}(w)\right).
$$
Thus,
\begin{align*}
\left|\phi^m\right|^0_T
&\leq 10\left(\frac{T}{h}+1\right)\beta^2\exp\left\{\gamma\left(\Vert w\Vert_T+\delta\right)\right\} \left(\Delta_{0, T}(w)+\varepsilon^\alpha_{m}(w)\right)\\
&\leq 10\left(\frac{T}{h}+1\right)\beta^2\exp\left\{\gamma\left(\Vert w\Vert_T+\delta\right)\right\} \left(T^\alpha \Vert w\Vert_\alpha+\frac{12e^L}{m^\alpha}\Vert w\Vert_\alpha\right).
\end{align*}
By definition, $\exp\left\{\gamma\Vert w\Vert_T\right\} \geq 1$, then we can complete the proof by recalling the
the definition of $h$.
\end{proof}
\subsection{The existence and uniqueness for the reflected $G$-Brownian motion}
By Theorem \ref{det1}, for each $\omega\in \Omega$, there exists a pair $(X(\omega), K(\omega))$ that solves the deterministic Skorohod problem for $B(\omega)$, i.e.,
\begin{align}\label{eq32}
X(\omega)= x_0+ B(\omega)+K(\omega), \quad\quad x_0\in \overline{D}.
\end{align}
It is easy to see that the pair of processes $(X, K)$ satisfy (i), (ii) and (iii) in Definition \ref{defrgb}. Thus, to prove Theorem \ref{trgb}, it suffices to show that both $X$ and $K$ belong to $M_G^{2}\left( \left[ 0,T%
\right]; \mathbb{R}^{d}\right)$.
\begin{lemma}\label{lem45}
Assume that the domain $D\subset\mathbb{R}^d$ is open and satisfies Conditions (A) and (B). For each $\omega\in \Omega$, we define the pair of processes $(X, K)$ by the unique solution of the deterministic Skorohod problem
$$
X(\omega)= x_0+ B(\omega)+K(\omega), \quad\quad x_0\in \overline{D},
$$
where $B$ is a $G$-Brownian motion. Then, $X$ and $K$ belong to $M_G^{2}\left( \left[ 0,T%
\right]; \mathbb{R}^{d}\right)$.
\end{lemma}
The main idea to prove this lemma is to construct a sequence of $\{(X^m, K^m)\}_{m\in \mathbb{N}^*}$ formed by elements from $(M^p_G([0, T]; \mathbb{R}^d)\times M^p_G([0, T]; \mathbb{R}^d))$ and then to show that the following convergences hold
$$
X^m\longrightarrow X, \quad\quad K^m\longrightarrow K, \quad\quad \mbox{in}\quad M^p_G([0, T]; \mathbb{R}^d),\ \ \ \ {\rm for}\ p\geq 2.
$$
Similarly to the previous subsection, we define for each $m\in \mathbb{N}^*$,
\begin{equation*}
X_{t}^{m}=x_{0}+B_{t}-\frac{m}{2}\int_{0}^{t}\nabla U(X_{s}^{m})ds,~~0\leq
t\leq T,
\end{equation*}%
which is a stochastic differential equations driven by $G$-Brownian motion with bounded Lipschtiz coefficients. For each $m\in \mathbb{N}^*$, the above equation admits a unique solution in $M^p_G([0, T]; \mathbb{R}^d)$, $p\geq 2$ (cf. Theorem 4.2 in \cite{Gao}). Moreover, one can find a version of $X^m$, denoted still by $X^m$, such that there exists a polar set $A^m$, for all $\omega\in (A^m)^c$, $X^m$ is continuous and
$$
X_{t}^{m}(\omega)=x_{0}+B_{t}(\omega)-\frac{m}{2}\int_{0}^{t}\nabla U(X_{s}^{m}(\omega))ds,~~0\leq
t\leq T.
$$
We note $A:=\cup_{m\in \mathbb{N}^*} A^m$, which is still a polar set, and we note
\begin{equation*}
K_{t}^{m}\left( \omega \right) :=-\frac{m}{2}\int_{0}^{t}\nabla
U(X_{s}^{m}\left( \omega \right) )ds.
\end{equation*}
In what follows, we shall find a bound uniform in $m$ for $\{\mathbb{E}\left[\sup_{0\leq t\leq T} |X^m_t|^p\right]\}_{m\in \mathbb{N}^*}$ and $\{\mathbb{E}\left[(|K^m|^0_T)^p\right]\}_{m\in \mathbb{N}^*}$.\\
First, by Kolmogorov's Criterion (cf. Theorem 36 in Denis et al. \cite{DHP}), for any $\alpha\in (0, 1/2)$, there exists a polar set $A'$ such that for all $\omega\in (A')^c$, the path of $G$-Brownian motion $B(\omega)$ is $\alpha$-H\"older continuous. Moreover, for any $p>0$, $\alpha\in (0, 1/2)$,
\begin{equation}\label{pmoment}
\mathbb{E}\left[\Vert B\Vert^p_\alpha\right]=\mathbb{E}\left[ \left( \sup_{0\leq s<t\leq T}\frac{\left\vert
B_{t}-B_{s}\right\vert}{\left\vert t-s\right\vert ^{\alpha }}\right)^p %
\right] <\infty.
\end{equation}
For each $m\in \mathbb{N}^*$, define
$$
\overline{A}^m:=\left(\left\{\omega\in\Omega: \varepsilon^\alpha_m(B_\cdot(\omega)) <\frac{\delta}{180\beta\lambda(B_\cdot(\omega))} \wedge \frac{r_{0}}{2}\right\}\cap A^c\cap (A')^c\right)^c.
$
The following lemma gives an estimate for $c\left(\overline{A}^m\right)^c$.
\begin{lemma}
Fix $\alpha\in (0, 1/2)$. For $m\in \mathbb{N}^*$,
$$
c\left(\overline{A}^m\right)\leq \frac{C_{\alpha, p}}{m^p}, \quad\quad p\geq 1,
$$
where $C_{\alpha, p}$ depends on $T$, $\Gamma$, $\alpha$, $p$, $r_0$, $\delta$, $L$ and $\beta$.
\end{lemma}
\begin{proof} It is clear that
$$
c\left(\overline{A}^m\right)\leq c\left(\left\{\omega\in \Omega: \varepsilon^\alpha_m(B_\cdot(\omega)) \geq \frac{r_{0}}{2}\right\}\right)+ c\left(\left\{\omega\in \Omega: \varepsilon^\alpha_m(B_\cdot(\omega)) \geq \frac{\delta}{180\beta\lambda(B_\cdot(\omega))} \right\}\right).
$$
We calculate by Markov's inequality, for any $p\geq 1$,
\begin{align*}
c\left(\left\{\omega\in \Omega: \varepsilon^\alpha_m(B_\cdot(\omega)) \geq \frac{r_{0}}{2}\right\}\right)&=c\left(\left\{\omega\in \Omega: \Vert B_\cdot(\omega))\Vert_\alpha\geq \frac{r_0e^{-L}}{24}m^\alpha\right\}\right)\leq \frac{C_1\mathbb{E}\left[\left\Vert B\right\Vert^{p/\alpha}_\alpha\right]}{m^p},
\end{align*}
where $C_1>0$ depends on $r_0$ and $L$. On the other hand,
\begin{align*}
&c\left(\left\{\omega\in \Omega: \varepsilon^\alpha_m(B_\cdot(\omega))\geq \frac{\delta}{180\beta\lambda(B_\cdot(\omega))} \right\}\right)\\
&\quad\quad\quad=c\left(\left\{\omega\in\Omega: \Vert B_\cdot (\omega)\Vert_\alpha \exp \left(\gamma \Vert B_\cdot (\omega)\Vert_T\right)\geq \frac{\delta e^{-L}}{2160\beta \exp\{\gamma\delta\}} m^\alpha\right\}\right)\\
&\quad\quad\quad\leq \frac{C_2\mathbb{E}\left[\Vert B\Vert_\alpha^{p/\alpha}\exp \left\{\frac{\gamma p}{\alpha}\Vert B\Vert_T\right\}\right]}{m^p}\\
&\quad\quad\quad\leq \frac{C_2\mathbb{E}\left[\Vert B \Vert_\alpha^{2p/\alpha}\right]^{1/2}\mathbb{E}\left[\exp \left\{\frac{2\gamma p}{\alpha}\Vert B\Vert_T\right\}\right]^{1/2}}{m^p},
\end{align*}
where $C_2>0$ depends on $r_0$, $\delta$, $L$ and $\beta$. By Theorem 3.3 in Luo and Wang \cite{LW}, we have
$$
\mathbb{E}\left[\exp \left\{\frac{2\gamma p}{\alpha}\Vert B\Vert_T\right\}\right]\leq C',
$$
where $C'$ depends on $T$, $\Gamma$, $p$, $\alpha$ and $\gamma$.
We combine this with (\ref{pmoment}) to conclude the desired result.
\end{proof}
\begin{proposition}\label{prop47}
Assume that the domain $D\subset \mathbb{R}^d$ is open and satisfies Conditions (A) and (B). Fix $\alpha\in (0, \frac{1}{2})$, then we have
$$
\mathbb{E}\left[ \sup_{0\leq t\leq T}\left\vert X_{t}^{m}\right\vert ^{p}%
\right] +\mathbb{E}\left[ \left( \left\vert K^{m}\right\vert^0 _{T}\right) ^{p}%
\right] \leq C' _{\alpha, p},
$$
where $C'_{\alpha, p}$ depends on $T$, $\Gamma$, $\alpha$, $p$, $r_0$, $\delta$, $L$ and $\beta$.
\end{proposition}
\begin{proof}
We denote by $\Lambda$ an upper bound of $\left\vert \nabla U\right\vert$, then
\begin{equation}\label{ka}
\mathbb{E}\left[ \left( \left\vert K^{m}\right\vert^0 _{T}\right) ^{p}\mathbf{1%
}_{\overline{A}^{m}}\right] \leq \left( \frac{m}{2}T\Lambda \right) ^{p}\overline{C}%
\left( \overline{A}^{m}\right) =C_{\alpha, p}\left(\frac{T\Lambda}{2}\right)^p.
\end{equation}
From (\ref{5.1}), we have, for each $\omega\in\left (\overline{A}^m\right)^c$,
\begin{equation}
\left\vert K^{m}\right\vert^0_{T}\left( \omega \right) \leq C_0 \left(\left\Vert B_{.}\left( \omega \right)
\right\Vert _{\alpha }^{1+1/\alpha }+\left\Vert B_{.}\left( \omega \right)
\right\Vert _{\alpha }\right)\exp \left\{ \gamma \left( 1+\frac{1}{%
\alpha }\right) \left\Vert B_{.}\left( \omega \right) \right\Vert
_{T}\right\},\label{kac}
\end{equation}%
where $C_0$ is the constant from (\ref{5.1}). Thus,
\begin{align*}
\mathbb{E}\left[ \left( \left\vert K^{m}\right\vert^0 _{T}\right) ^{p}\mathbf{1%
}_{(A^{m})^c}\right] \leq C_0\left(\mathbb{E}\left[\Vert B\Vert_\alpha^{2p(1+\alpha)/\alpha}\right]^{1/2}+\mathbb{E}\left[\Vert B\Vert_\alpha^{2p}\right]^{1/2}\right)\mathbb{E}\left[\exp \left\{\frac{2\gamma p(1+\alpha)}{\alpha}\Vert B\Vert_T\right\}\right]^{1/2}.
\end{align*}
Recall that for some $C_p>0$, which depends only on $p$,
\begin{equation*}
\left\vert X_{t}^{m}\right\vert ^{p}\leq C_{p}\left( \left\vert
x_{0}\right\vert ^{p}+\left\vert B_{t}\right\vert ^{p}+\left\vert
K_{t}^{m}\right\vert ^{p}\right),
\end{equation*}%
we can deduce that
\begin{equation*}
\sup_{0\leq t\leq T}\left\vert X_{t}^{m}\right\vert ^{p}\leq C_{p}\left(
\left\vert x_{0}\right\vert ^{p}+\sup_{0\leq t\leq T}\left\vert
B_{t}\right\vert ^{p}+\left( \left\vert K^{m}\right\vert^0_{T}\right)
^{p}\right).
\end{equation*}%
We take the $G$-expectation on both sides and apply the BDG type inequality, (\ref{ka}) and (\ref{kac}) to conclude the desired result.
\end{proof}
Now we are ready to prove Lemma \ref{lem45}.
\begin{proof}[Proof of Lemma \ref{lem45}]
From (6.4) in \cite{SaishoetTanaka}, we have for $\omega\in (A\cup A')^c$,
$$
X^m_t(\omega)\longrightarrow X_t(\omega), \ \ \ \ \mbox{uniform\ on}\ [0, T],
$$
and
for each $\omega\in (\overline{A}^m)^c$,
\begin{align*}
\sup_{0\leq s\leq t}|X^m_s(\omega)-X_s(\omega)|^2\leq 4\varepsilon^\alpha_m(\omega)&\left(\left|K^m(\omega)\right|^0_t
+\left|K(\omega)\right|^0_t\right)\\&+\frac{\gamma}{2\beta}
\int^t_0 \sup_{0\leq u\leq s}\left(|X^m_u(\omega)-X_u(\omega)|^2\right)d\left(\left|K^m(\omega)\right|^0_s+\left|K(\omega)\right|^0_s\right),
\end{align*}
which implies
\begin{align*}
\sup_{0\leq t\leq T}|X^m_t(\omega)-X_t(\omega)|^2\leq 4\varepsilon^\alpha_m(\omega)\left(\left|K^m(\omega)\right|^0_T
+\left|K(\omega)\right|^0_T\right)\exp\left\{\frac{\gamma}{2\beta}\left(\left|K^m(\omega)\right|^0_T+\left|K(\omega)\right|^0_T\right)\right\}.
\end{align*}
Now we shall prove that
\begin{equation}\label{import4}
\mathbb{E}\left[\sup_{0\leq t\leq T}|X^m_t(\omega)-X_t(\omega)|^2\right]\longrightarrow 0, \ \ \ \ \mbox{as}\ m\longrightarrow \infty.
\end{equation}
For $\epsilon>0$, by Markov's inequality, we could first fix a constant $M_0>0$, such that
\begin{equation}\label{import2}
\left(2C'_{\alpha, 4}\right)^{\frac{1}{2}}\mathbb{E}\left[{\bf 1}_{\Vert B_\cdot\Vert_\alpha\geq M_0}\right]\leq \frac{\left(2C'_{\alpha, 4}\right)^{\frac{1}{2}}\mathbb{E}\left[{\Vert B_\cdot\Vert_\alpha}\right]}{M_0}\leq \frac{\epsilon}{2},
\end{equation}
where $C'_{\alpha, 4}$ is the constant from Proposition \ref{prop47}. Then, we choose $m_0\in\mathbb{N}^*$ sufficiently large such that
\begin{align}\label{import1}
\frac{12e^L}{m_0^\alpha}M_0\leq \frac{r_0}{2}\ \ \mbox{and}\ \ \frac{12e^L}{m_0^\alpha}M_0\leq \frac{\delta}{180\beta\exp\left\{\gamma\left(\delta+ M_0T^\alpha\right)\right\}};
\end{align}
\begin{align}\label{import}
\frac{48e^L}{m_0^\alpha}M_0\left(4C_0M_0^{1+1/\alpha}\exp\left\{\gamma\left(1+\frac{1}{\alpha}\right)M_0T^\alpha\right\}\right)
\exp\left\{ \frac{2\gamma C_0}{\beta}M_0^{1+1/\alpha}\exp\left\{\gamma\left(1+\frac{1}{\alpha}\right)M_0T^\alpha\right\}\right\}\leq \frac{\epsilon}{2},
\end{align}
where $C_0$ is the constant from Proposition \ref{prop44}.
From (\ref{import1}), we know for $\omega\in \{\Vert B_\cdot\Vert_\alpha< M_0\}\cap A^c\cap A'^c$ and $m\geq m_0$,
$$
\varepsilon^\alpha_m(B_\cdot(\omega)) <\frac{\delta}{180\beta\lambda(B_\cdot(\omega))} \wedge \frac{r_{0}}{2}.
$$
It follows that for $m\geq m_0$,
\begin{align*}
\mathbb{E}\left[ \sup_{0\leq t\leq T}\left\vert X_{t}^{m}-X_{t}\right\vert
^{2}\right] \leq \mathbb{E}&\left[ \sup_{0\leq t\leq T}\left\vert
X_{t}^{m}-X_{t}\right\vert ^{2}\mathbf{1}_{\{\Vert B_\cdot\Vert_\alpha< M_0\}\cap A^c\cap A'^c}\right] \\
&+\mathbb{E}\left[ \sup_{0\leq t\leq T}\left\vert X_{t}^{m}-X_{t}\right\vert
^{2}\mathbf{1}_{\{\Vert B_\cdot\Vert_\alpha\geq M_0\}\cup A\cup A'}\right] \\
&\leq \frac{\epsilon}{2}+\mathbb{E}\left[ \sup_{0\leq t\leq T}\left\vert X_{t}^{m}-X_{t}\right\vert
^{4}\right]^{\frac{1}{2}}\mathbb{E}\left[\mathbf{1}_{\{\Vert B_\cdot\Vert_\alpha\geq M_0\}}\right]\leq \epsilon,
\end{align*}
where the last inequality is deduced from (\ref{import2}) and (\ref{import}). Therefore, (\ref{import4}) holds true. From (\ref{import4}), it is obvious that
$$
\mathbb{E}\left[\sup_{0\leq t\leq T}|K^m_t(\omega)-K_t(\omega)|^2\right]\longrightarrow 0, \ \ \ \ \mbox{as}\ m\longrightarrow \infty.
$$
We end the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{trgb}]
We define pathwisely a couple $(X, K)$ by the solution of the deterministic problem (\ref{eq32}). Then, we apply Lemma \ref{lem45} to prove that $X$ and $K$ belong to
$M_G^{2}\left( \left[ 0,T%
\right]; \mathbb{R}^{d}\right)$. Therefore, $(X, K)$ is a couple satisfying Definition \ref{defrgb}. The uniqueness of the solution is inherited from the pathwise uniqueness.
\end{proof}
Instead of the $G$-Brownian motion, if we consider a $G$-It\^o process as
\begin{equation}\label{eq7}
Y_{t}=\int_{0}^{t}\alpha _{s}ds+\int_{0}^{t}\eta _{s}^{ij}d\left\langle B^{i},B^{j}\right\rangle _{s}+\int_{0}^{t}\beta^j_{s}
dB_{s}^{j},~~~~0\leq t\leq T,
\end{equation}%
where $\alpha$, $\eta^{i j}$, $\beta^{j}: \Omega\times [0, T]\longrightarrow \mathbb{R}^d$, $i$, $j=1, 2, \ldots d$, are bounded functions in $M^2_G([0, T]; \mathbb{R}^d)$, then a similar result holds due to the fact that for any $p\geq 2$,
$$
\mathbb{E}\left[\sup_{0\leq t\leq T}\left|Y_t\right|^p\right]\leq C_p,
$$
which can be easily obtained by the BDG type inequality.
\begin{corollary}\label{ritop}
Suppose that the domain $D\subset\mathbb{R}^d$ is open and satisfies Conditions (A) and (B). Then there exists a couple $\left( X ,K \right)\in \left(M^2_G([0, T]; \mathbb{R}^d)\times M^2_G([0, T]; \mathbb{R}^d)\right)$ which solves the Skorohod problem
\begin{align}\label{e3}
X_{t}=x_{0}+Y_{t}+K_{t},~~0\leq t\leq T,
\end{align}
whenever $x_0\in \overline{D}$ and $Y$ is defined by (\ref{eq7}). Moreover, if the above problem admits two solutions $(X, K)$ and $(X', K')$, then the exists a polar set $A$, such that for all $\omega\in A^c$,
$$
X(\omega)=X'(\omega)~~~~and~~~K(\omega)=K'(\omega),\ \ 0\leq t\leq T.
$$
\end{corollary}
\begin{remark}
Indeed, thanks to Proposition \ref{prop47}, we could have a stronger convergence instead of (\ref{import4}), that is, for any $p\geq 2$,
$$
\mathbb{E}\left[\sup_{0\leq t\leq T}|X^m_t(\omega)-X_t(\omega)|^p\right]\longrightarrow 0, \ \ \ \ \mbox{as}\ m\longrightarrow \infty.
$$
Thus, the couple of solution $(X, K)$ in both Theorem \ref{trgb} and Corollary \ref{ritop} could be found in $\left(M^p_G([0, T]; \mathbb{R}^d)\times M^p_G([0, T]; \mathbb{R}^d)\right)$.
\end{remark}
\subsection{The existence and uniqueness for the RGSDE}
Without loss of generality, we consider in this subsection the following equation instead of (\ref{e1}),
\begin{align}
X_{t}=x_{0}+\int_{0}^{t}f\left(
s,X_{s}\right) ds+\int_{0}^{t}g\left( s,X_{s}\right) dB_{s}+K_{t},~~~0\leq t\leq T,~~~~\text{q.s.}. \label{e2}
\end{align}
However, all result here holds for the more general case (\ref{e1}) due to the boundedness of the density of the process $\langle B, B\rangle$ (see \S III-4 in \cite{Peng4}). \\
If the coefficients $f$ and $g$ satisfy Assumptions (H1) and (H2) and (\ref{e2}) admits a solution couple $(X, K)$, then $(X, K)$ can be regarded as the solution couple of the Skorohod problem (\ref{e3}) in the domain $D$ for
$$
Y_t=\int_{0}^{t}f\left(
s,X_{s}\right) ds+\int_{0}^{t}g^{j}\left( s,X_{s}\right) dB^j_{s},~~~0\leq t\leq T.
$$
Then, it is straightforward that for any $p\geq 2$ there exists a constant $C'_p>0$ such that
$$
\mathbb{E}\left[ \sup_{0\leq t\leq T}\left\vert X_{t}\right\vert ^{p}%
\right] +\mathbb{E}\left[ \left( \left\vert K\right\vert^0 _{T}\right) ^{p}%
\right] \leq C' _{p}.
$$
\begin{proposition}\label{propest}
\label{p1} Suppose that the domain $D\subset \mathbb{R}^d$ is open and satisfies Conditions (A), (B) and (C). For $i=1, 2$, the couple $(\widetilde{X}^i, K^i)$ are solutions of the following Skorohod problems
\begin{align*}
\widetilde X_{t}^{i} =x_{0}+\int_{0}^{t}f^{i}\left( s,X_{s}^{i}\right) ds+\int_{0}^{t}g^{i}\left( s,X_{s}^{i}\right)
dB_{s}+K_{t}^{i},~~i=1,~2, \\
\left\vert K^{i}\right\vert _{t} =\int_{0}^{t}\mathbf{1}%
_{\left\{ \widetilde X^i_{s}\in \partial D\right\} }d\left\vert K^{i}\right\vert _{s}~%
{\rm and}\ K_{t}^{i}=\int_{0}^{t}{\bf n}_{s}^{i}d\left\vert K^{i}\right\vert
_{s}~{\rm with}\ {\bf n} _{s}^{i}\in {\mathcal{N}}_{{\widetilde X}_{s}^{i}},
\end{align*}%
where the coefficients $f^i$ and $g^i$ satisfy Assumptions (H1) and (H2).
Then, there exists a constant $C>0$ that depends on $\Gamma$, $d$, $\delta'$, $L_\Psi$ and $L_0$,
\begin{align}
\mathbb{E}\left[ \sup_{0\leq s\leq t}\left\vert
\widetilde X_{s}^{1}-\widetilde X_{s}^{2}\right\vert ^{4}\right] &+\mathbb{E}\left[ \sup_{0\leq
s\leq t}\left\vert K_{s}^{1}-K_{s}^{2}\right\vert ^{4}\right]\\
& \leq C\int^t_0\left(\mathbb{E}\left[ \sup_{0\leq u\leq s}\left\vert
X_{u}^{1}- X_{u}^{2}\right\vert ^{4}\right]+
\mathbb{E}\left[ \sup_{0\leq u\leq s}\left\vert \hat{f}_{u}\right\vert ^{4}%
\right] +\mathbb{E}\left[ \sup_{0\leq u\leq s}\left\vert \hat{g}%
_{u}\right\vert ^{4}\right] \right) ds,\label{est10}
\end{align}%
where
$\hat{f}_{s}:=f^{1}\left( s,X_{s}^{2}\right)
-f^{2}\left( s,X_{s}^{2}\right) $ and $\hat{g}_{s}:=g^{1}\left( s,X_{s}^{2}\right) -g^{2}\left(
s,X_{s}^{2}\right)$.
\end{proposition}
\begin{proof} The proof is similar to the one of Lemma 3.1 in \cite{LionsetSznitman}, so we only display the key steps for the convenience of the readers. First, we have
\begin{eqnarray*}
\left\vert {\widetilde X}_{t}^{1}-{\widetilde X}_{t}^{2}\right\vert ^{2}
&=&2\int_{0}^{t}\left\langle {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2},f^{1}\left(
s,X_{s}^{1}\right) -f^{1}\left( s,X_{s}^{2}\right) +\hat{f}_{s}\right\rangle
ds \\
&&+2\int_{0}^{t}\ltrans{\left({\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right)}\left( g^{1}\left(
s,X_{s}^{1}\right) -g^{1}\left( s,X_{s}^{2}\right) +\hat{g}_{s}\right)
dB_{s} \\
&&+2\int_{0}^{t}\left\langle {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}, {\bf n}
_{s}^{1}\right\rangle d\left\vert K^{1}\right\vert _{s}
-2\int_{0}^{t}\left\langle {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2},{\bf n}
_{s}^{2}\right\rangle d\left\vert K^{2}\right\vert _{s} \\
&&+\int_{0}^{t}\mbox{tr}\left[ \left( g^{1}\left( s,X_{s}^{1}\right)
-g^{1}\left( s,X_{s}^{2}\right) +\hat{g}_{s}\right) d\left\langle
B,B\right\rangle _{s}\ltrans{\left( g^{1}\left( s,X_{s}^{1}\right) -g^{1}\left(
s,X_{s}^{2}\right) +\hat{g}_{s}\right)} \right] ,
\end{eqnarray*}%
and
\begin{align*}
\Psi \left( {\widetilde X}_{t}^{i}\right) =\Psi \left( x_{0}\right)
&+\int_{0}^{t}\left\langle \nabla \Psi
\left( {\widetilde X}_{s}^{i}\right) ,f^{i}\left( s,X_{s}^{i}\right) \right\rangle ds+\int_{0}^{t} \ltrans{\left(\nabla \Psi \left( {\widetilde X}_{s}^{i}\right)\right)} g^{i}\left(
s,X_{s}^{i}\right) dB_{s} \\
&+\int_{0}^{t}\left\langle \nabla \Psi \left( {\widetilde X}_{s}^{i}\right) ,{\bf n}
_{s}^{i}\right\rangle d\left\vert K^{i}\right\vert _{s} +\frac{1}{2}\int_{0}^{t}\mbox{tr}\left[{\bf H}(\Psi({\widetilde X}_s^i))
g^{i}\left( s,X_{s}^{i}\right) d\left\langle
B,B\right\rangle _{s}\ltrans{\left( g^{i}\left( s,X_{s}^{i}\right) \right)}%
\right].
\end{align*}%
Then,
\begin{align*}
&\exp\left\{-\frac{1}{{\delta'} }\left( \Psi \left( {\widetilde X}_{t}^{1}\right) +\Psi
\left( {\widetilde X}_{t}^{2}\right) \right) \right\} \times \left\vert
{\widetilde X}_{t}^{1}-{\widetilde X}_{t}^{2}\right\vert ^{2} \\
&\ \ \ \ =2\int_{0}^{t}\exp \left\{ -\frac{1}{{\delta'} }\left( \Psi \left(
{\widetilde X}_{s}^{1}\right) +\Psi \left( {\widetilde X}_{s}^{2}\right) \right) \right\} \\
&\ \ \ \ \ \ \ \ \ \ \ \ \times \bigg\{
\left\langle {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}, f^{1}\left(
s,X_{s}^{1}\right) -f^{1}\left( s,X_{s}^{2}\right) +\hat{f}_{s}\right\rangle ds\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\ltrans{\left( {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right)} \left( g^{1}\left(
s,X_{s}^{1}\right) -g^{1}\left( s,X_{s}^{2}\right) +\hat{g}_{s}\right) dB_{s}\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\left\langle {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}, {\bf n}
_{s}^{1}\right\rangle d\left\vert K^{1}\right\vert _{s}
-\left\langle {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2},{\bf n}
_{s}^{2}\right\rangle d\left\vert K^{2}\right\vert _{s}\bigg\}
\\
&\ \ \ \ \ \ \ \ +\int_{0}^{t}\exp \left\{ -\frac{1}{{\delta'} }\left( \Psi \left(
{\widetilde X}_{s}^{1}\right) +\Psi \left( {\widetilde X}_{s}^{2}\right) \right) \right\} \\
&\ \ \ \ \ \ \ \ \ \ \ \ \times\mbox{tr}\left[\left( g^{1}\left( s,X_{s}^{1}\right) -g^{1}\left(
s,X_{s}^{2}\right) +\hat{g}_{s}\right) d\left\langle B,B\right\rangle
_{s}\ltrans{\left( g^{1}\left( s,X_{s}^{1}\right) -g^{1}\left( s,X_{s}^{2}\right) +%
\hat{g}_{s}\right)}\right] \\
&\ \ \ \ \ \ \ \ -\frac{1}{{\delta'} }\int_{0}^{t}\exp \left\{ -\frac{1}{{\delta'} }\left( \Psi
\left( {\widetilde X}_{s}^{1}\right) +\Psi \left( {\widetilde X}_{s}^{2}\right) \right) \right\}
\times \left\vert {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right\vert ^{2} \\
&\ \ \ \ \ \ \ \ \ \ \ \ \times\bigg\{
\left\langle \nabla \Psi \left({\widetilde X}_{s}^{1}\right), f^{1}\left(s,X_{s}^{1}\right)\right\rangle
+\left\langle \nabla \Psi \left( {\widetilde X}_{s}^{2}\right), f^{2}\left(s,X_{s}^{2} \right)\right\rangle ds\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\left( \ltrans{\left(\nabla\Psi \left( {\widetilde X}_{s}^{1}\right)\right)} g^{1}\left(
s,X_{s}^{1}\right)+\ltrans{\left(\nabla\Psi \left({\widetilde X}_{s}^{2}\right)\right)} g^{2}\left(
s,X_{s}^{2}\right) \right) dB_{s} \\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\left\langle\nabla \Psi \left( \widetilde{X}_{s}^{1}\right), {\bf n} _{s}^{1}\right\rangle d\left\vert
K^{1}\right\vert _{s}
+\left\langle\nabla \Psi \left( \widetilde{X}_{s}^{2}\right), {\bf n} _{s}^{2}\right\rangle d\left\vert
K^{2}\right\vert _{s}\bigg\} \\
&\ \ \ \ \ \ \ \ -\frac{1}{2{\delta'} }\int_{0}^{t}\exp \left\{ -\frac{1}{{\delta'} }\left( \Psi
\left( {\widetilde X}_{s}^{1}\right) +\Psi \left( {\widetilde X}_{s}^{2}\right) \right) \right\}
\times \left\vert {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right\vert ^{2} \\
&\ \ \ \ \ \ \ \ \ \ \ \ \times\mbox{tr}\bigg[{\bf H}\left(\Psi\left( {\widetilde X}_{s}^{1}\right)\right)
g^{1}\left( s,X_{s}^{1}\right) d\left\langle B,B\right\rangle _{s}\ltrans{\left(
g^{1}\left( s,X_{s}^{1}\right) \right)}\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +{\bf H}\left(\Psi\left( {\widetilde X}_{s}^{2}\right)\right) g^{2}\left(
s,X_{s}^{2}\right) d\left\langle B,B\right\rangle _{s}\ltrans{\left( g^{2}\left(
s,X_{s}^{2}\right) \right)}\bigg] \\
&\ \ \ \ \ \ \ \ +\frac{1}{2}\left( \frac{1}{{\delta'} }\right) ^{2}\int_{0}^{t}\exp \left\{ -%
\frac{1}{{\delta'} }\left( \Psi \left( {\widetilde X}_{s}^{1}\right) +\Psi \left(
{\widetilde X}_{s}^{2}\right) \right) \right\} \left\vert {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right\vert
^{2} \\
&\ \ \ \ \ \ \ \ \ \ \ \ \times \left(\ltrans{\left(\nabla \Psi \left( {\widetilde X}_{s}^{1}\right)\right)} g^{1}\left(
s,X_{s}^{1}\right) +\ltrans{\left(\nabla \Psi \left( {\widetilde X}_{s}^{2}\right)\right)} g^{2}\left(
s,X_{s}^{2}\right)\right)\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \times d\left\langle B,B\right\rangle _{s}\left(\nabla \Psi \left(
{\widetilde X}_{s}^{1}\right) \ltrans{\left(g^{1}\left( s,X_{s}^{1}\right)\right)} +\nabla \Psi \left(
{\widetilde X}_{s}^{2}\right) \ltrans{\left(g^{2}\left( s,X_{s}^{2}\right) \right)}\right)\\
&\ \ \ \ \ \ \ \ -\frac{2}{{\delta'} }\int_{0}^{t}\exp \left\{ -\frac{1}{{\delta'} }\left( \Psi
\left( {\widetilde X}_{s}^{1}\right) +\Psi \left( {\widetilde X}_{s}^{2}\right) \right) \right\} \\
&\ \ \ \ \ \ \ \ \ \ \ \ \times\ltrans{\left( {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right)}\left( g^{1}\left(
s,X_{s}^{1}\right) -g^{1}\left( s,X_{s}^{2}\right) +\hat{g}_{s}\right) \\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \times d\left\langle B,B\right\rangle _{s}\left(\nabla \Psi \left(
{\widetilde X}_{s}^{1}\right) \ltrans{\left(g^{1}\left( s,X_{s}^{1}\right)\right)} +\nabla \Psi \left(
{\widetilde X}_{s}^{2}\right) \ltrans{\left(g^{2}\left( s,X_{s}^{2}\right) \right)}\right).
\end{align*}%
Thanks to Condition (C), we know that the integrals with respect to $d|K|$ are negative. Since the set $\Gamma$, the function $\Psi$ and its derivatives, the functions $f^i$ and $g^i$, $i=1, 2$, are bounded, we have
\begin{align*}
&\exp\left\{-\frac{2M}{{\delta'} } \right\} \times \left\vert
{\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right\vert ^{2} \\
&\ \ \ \ \leq 2\int_{0}^{t}\exp \left\{ -\frac{1}{{\delta'} }\left( \Psi \left(
{\widetilde X}_{s}^{1}\right) +\Psi \left( {\widetilde X}_{s}^{2}\right) \right) \right\}\ltrans{\left( {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right)} \left( g^{1}\left(
s,X_{s}^{1}\right) -g^{1}\left( s,X_{s}^{2}\right) +\hat{g}_{s}\right) dB_{s}\\
&\ \ \ \ \ \ \ \ -\frac{1}{{\delta'} }\int_{0}^{t}\exp \left\{ -\frac{1}{{\delta'} }\left( \Psi
\left( {\widetilde X}_{s}^{1}\right) +\Psi \left( {\widetilde X}_{s}^{2}\right) \right) \right\}
\times \left\vert {\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right\vert ^{2}\\
&\ \ \ \ \ \ \ \ \ \ \ \ \times\left( \ltrans{\left(\nabla\Psi \left( {\widetilde X}_{s}^{1}\right)\right)} g^{1}\left(
s,X_{s}^{1}\right)+\ltrans{\left(\nabla\Psi \left({\widetilde X}_{s}^{2}\right)\right)} g^{2}\left(
s,X_{s}^{2}\right) \right) dB_{s}\\
&\ \ \ \ \ \ \ \ +C\int^t_0 \left(\left\vert
{\widetilde X}_{s}^{1}-{\widetilde X}_{s}^{2}\right\vert ^{2}+ \left\vert
{X}_{s}^{1}-{X}_{s}^{2}\right\vert ^{2}+\left\vert
\hat{f}_{s}\right\vert^{2}+\left\vert
\hat{g}_{s}\right\vert^{2}\right)ds,
\end{align*}%
where $C>0$ is a constant that depends on $\Gamma$, $d$, $\delta'$, $L_\Psi$ and $L_0$, which may vary from line to line in the sequel.
We square both sides and apply the BDG type inequality to obtain
\begin{align*}
\mathbb{E}&\left[ \sup_{0\leq s\leq t}\left\vert
\widetilde X_{s}^{1}-\widetilde X_{s}^{2}\right\vert ^{4}\right]\\
& \leq C\int^t_0\left(\mathbb{E}\left[ \sup_{0\leq u\leq s}\left\vert
\widetilde X_{u}^{1}-\widetilde X_{u}^{2}\right\vert^{4}\right]+\mathbb{E}\left[ \sup_{0\leq u\leq s}\left\vert
X_{u}^{1}- X_{u}^{2}\right\vert ^{4}\right]+
\mathbb{E}\left[ \sup_{0\leq u\leq s}\left\vert \hat{f}_{u}\right\vert ^{4}%
\right] +\mathbb{E}\left[ \sup_{0\leq u\leq s}\left\vert \hat{g}%
_{u}\right\vert ^{4}\right] \right) ds.
\end{align*}%
The desired result follows from the Gronwall inequality.
\end{proof}
\begin{proof}[Proof of Theorem \ref{trgsde}]
The uniqueness of solution is straightforward by Proposition \ref{propest}. We now turn to prove the existence. Indeed, by Corollary \ref{ritop} one can construct a sequence $\{(X^m, K^m)\}_{m\in \mathbb{N}^*}$ by the Picard type iteration starting with $X^0\equiv K^0=0$,
\begin{align*}
Y^{m+1}_t=\int_{0}^{t}f\left( s,X^m_{s}\right) ds+\int_{0}^{t}g\left( s,X^m_{s}\right)
dB_{s},\ \ \ \
X^{m+1}_{t} =x_{0}+Y^{m+1}_t+K^{m+1}_{t}, \\
\left\vert K^{m+1}\right\vert _{t} =\int_{0}^{t}\mathbf{1}%
_{\left\{ X^{m+1}_{s} \in \partial D\right\} }d\left\vert K^{m+1}\right\vert _{s}~%
{\rm and}\ K^{m+1}_{t}=\int_{0}^{t}{\bf n}^{m+1}_{s}d\left\vert K^{i}\right\vert
_{s}~{\rm with}\ {\bf n}^{m+1} _{s}\in {\mathcal{N}}_{{X^{m+1}_{s}}}.
\end{align*}
Thanks to the a priori estimate (\ref{est10}), we can proceed a similar argument as the proof of Theorem 2.1 and 4.1 in \cite{Gao} to find a couple of processes $(X, K)$ such that
\begin{align*}
\mathbb{E}\left[\sup_{0\leq t\leq T}|X^m_t-X_t|^2\right]+\mathbb{E}\left[\sup_{0\leq t\leq T}|K^m_t-K_t|^2\right]&\longrightarrow 0,\\
\mathbb{E}\left[\sup_{0\leq t\leq T}\left|Y^m_t-\int_{0}^{t}f\left( s,X_{s}\right) ds-\int_{0}^{t}g\left( s,X_{s}\right)
dB_{s}\right|^2\right]&\longrightarrow 0,
\ \ \ \ \mbox{as}\ m\longrightarrow \infty,
\end{align*}
and such that there exists a polar set $A$ and subsequence $\{(X^{m_k}, K^{m_k})\}_{k\in \mathbb{N}^*}$, such that for each $\omega\in A^c$, $(X^{m_k}(\omega), K^{m_k}(\omega))$ is the solution couple for the deterministic Skorohod problem with $(Y^{m_k}(\omega), D)$, and
\begin{align*}
\sup_{0\leq t\leq T}|X^{m_k}_t(\omega)-X_t(\omega)|+\sup_{0\leq t\leq T}|K^{m_k}_t(\omega)-K_t(\omega)|&\longrightarrow 0, \\
\sup_{0\leq t\leq T}\left|Y^{m_k}(\omega)-\left(\int_{0}^{t}f\left( s,X_{s}\right) ds-\int_{0}^{t}g\left( s,X_{s}\right)
dB_{s}\right)(\omega)\right |&\longrightarrow 0,
\ \ \ \ \mbox{as}\ k\longrightarrow \infty.
\end{align*}
It is clear that $\left( X ,K \right)\in \left(M^2_G([0, T]; \mathbb{R}^d)\times M^2_G([0, T]; \mathbb{R}^d)\right)$. Besides, for each $\omega\in A^c$, $(X(\omega), K(\omega))$ verified (i) (ii) (iii) of Definiton \ref{defrgb}, which can be proved by the last step of the proof to Theorem 4.1 in \cite{Saisho1}. We complete the proof.
\end{proof}
\bigskip
{\bf Acknowledgment:} Yiqing Lin gratefully acknowledges financial support from the European Research Council (ERC) under grant 321111. Abdoulaye SOUMANA HIMA is grateful for partial financial support from the Lebesgue Center of Mathematics (``Investissements d'avenir'' Program) under grant ANR-11-LABX-0020-01. We thank Ying HU for helpful suggestions.
\bigskip
|
\section{Introduction}
During the last decade the electronic, magnetic, and structural
properties of the iron-based high-temperature superconducting
pnictides and chalcogenides have been the subject of intensive
research\cite{pnictide_discovery,review_superconductors}. These
novel superconducting materials show certain similarities with
the high-$T_c$ cuprate superconductors. Indeed, the iron-based
superconductors (FeSC) adopt a quasi 2D crystal structure where
the iron atoms form a square lattice. The latter are separated
by non-conducting layers containing, for example, alkali, alkaline
earth, or rare earth elements, oxygen and/or fluorine. Moreover,
the superconducting phase of these novel compounds often appears
in the vicinity of a magnetic phase transition and/or structural
instability. In particular, superconductivity in FeSCs often
occurs as a result of the suppression of long-range, single-stripe
antiferromagnetic (AF) order with a wave vector $Q_m = (\pi, \pi)$,
due to electron/hole doping or pressure. This behavior has been
regarded as evidence for the importance of spin fluctuations in
the pairing of electrons in FeSCs.
The newly discovered Fe$_{1+y}$Se is structurally the simplest
among the FeSCs \cite{FeSe_structure}. At ambient pressure it has
been found to become superconducting below $T_{c} \sim$ 8~K close
to its stoichiometric composition \cite{Superconductivity_FeSe}.
FeSe has the same layered structure as the pnictides, containing
layers of edge-sharing FeSe$_4$ tetrahedra, but without separating
(non-conducting) layers \cite{FeSe_structure}. Therefore FeSe is
viewed as the parent compound of Fe-based superconductors which
represents a minimal model material for understanding the mechanism
of superconductivity of FeSCs. Moreover, FeSe itself exhibits
remarkable physical properties. Its critical temperature $T_c~\sim$~8~K
at normal pressure increases to $\sim$~14~K upon isovalent substitution
of Se with Te (corresponding to a negative chemical pressure, i.e.,
lattice expansion \cite{FeSe_Te_doping}), to $\sim$~37~K under
compression \cite{FeSe_hydrostatic}, to $\sim$~40~K by means of
intercalation \cite{FeSe_intercalation}, and all the way up to
$\sim$~65-109~K in the case of a monolayer \cite{FeSe_monolayer}.
In addition, FeSe has been found to exhibit a transition to a nematic
phase below $\sim$~90~K in which the crystal (C$_4$) rotation
symmetry is spontaneously broken \cite{nematicity}. Due to these
intriguing properties FeSe has attracted much recent attention
from both theory and experiment.
Angle-resolved photoemission spectroscopy (ARPES) \cite{FeSe_photoemission_1,FeSe_photoemission_2,FeSe_photoemission_3}
and band structure calculations \cite{DFT_FeSe} reveal that the
electronic structure of FeSe resembles that of the pnictides. It
has a quasi 2D Fermi surface with three concentric hole pockets
at the Brillouin zone $\Gamma$-point and two intersecting elliptical
electron pockets centered at the M-point. The Fermi surface topology
is characterized by an in-plane nesting wave vector $(\pi,\pi)$,
consistent with $s^{\pm}$ pairing symmetry \cite{s_pm_pairing}.
Moreover, experimental studies of the spin excitation spectra of
both pnictides and chalcogenides show an enhancement of short-range
AFM spin fluctuations at vector $(\pi,\pi)$ near the $T_c$ \cite{FeSe_pi_pi_fluctuations}.
These results suggest a common origin of superconductivity in
pnictides and chalcogenides, for example due to spin fluctuations
associated with the suppression of long-range magnetic order.
Unlike the majority of the FeSCs, FeSe is not magnetically ordered
at ambient pressure and composition \cite{FeSe_phase_diag, FeSe_magnetism}.
Its isoelectronic counterpart FeTe, the end member of the Fe(Se,Te)
series, is antiferromagnetic below the N\'eel temperature of 70~K.
However, in contrast to the magnetic phases of the Fe-pnictides,
FeTe exhibits double-stripe AF order with a $(\pi,0)$ propagation
vector \cite{FeSe_T_cT}. Upon compression, FeTe exhibits a transition
to a collapsed-tetragonal phase which is accompanied by a collapse
of magnetic moments \cite{FeSe_T_cT}. All this suggests a reconstruction
of the electronic structure of Fe(Se,Te) upon change of the Se content
or compression.
Photoemission and ARPES measurements of the electronic properties
of Fe(Se,Te) reveal a significant narrowing of the Fe $3d$ bandwidth
as compared to band structure calculations \cite{FeSe_photoemission_1}.
This corresponds to a strong orbital-dependent enhancement of the
quasiparticle mass in the range $\sim$ 3-20 compared with the values
obtained by electronic band structure techniques \cite{FeSe_photoemission_2,FeSe_photoemission_3}.
Moreover, these experiments exhibit a damping of the coherent
quasiparticles in Te-rich Fe(Se,Te), indicating a crossover from a
coherent to incoherent behavior of the electronic structure.
In addition, with increasing Te content, ARPES data for Fe(Se,Te)
show a suppression of the spectral weight intensity associated with
a Fermi surface pocket at the Brillouin zone M-point \cite{FeSe_photoemission_3}.
This behavior is accompanied by an enhancement of spectral weight
at the X-point, implying a possible doping-induced reconstruction
of the electronic structure. Overall these experimental results
point towards the importance of strong orbital-selective correlations.
State-of-the-art methods for the calculation of the electronic
properties of strongly correlated systems, such as the density
functional theory plus dynamical mean-field theory (DFT+DMFT)
approach \cite{dmft,dftdmft} provide a good qualitative and even
quantitative description of the band structure of FeSCs \cite{U_in_superconductors}.
Applications of DFT+DMFT to FeSe yield a band mass enhancement
in the range 2-5 and, in contrast to the pnictides, reveal the
presence of a lower Hubbard band in the spectral function of
FeSe \cite{FeSe_Aichhorn_2010,WB16}. This clearly demonstrates
the importance of correlation effects for the electronic properties
of FeSe. Moreover, our recent DFT+DMFT calculations of the electronic
properties and phase stability of FeSe predict that FeSe undergoes
a phase transformation from a collapsed tetragonal to tetragonal
phase upon expansion of the lattice \cite{FeSe_Leonov_2015}.
The transformation is found to be accompanied by a complete
reconstruction of the topology of the Fermi surface (Lifshitz transition),
a sharp increase of the local moments, and a change of magnetic
correlations due to a transition of the in-plane magnetic wave
vector from $(\pi,\pi)$ to $(\pi,0)$. This behavior was attributed
to a correlation-induced shift of the van Hove singularity
associated with the Fe $xy$ and $xz/yz$ orbitals at the Brillouin
zone M-point across the Fermi level \cite{FeSe_Leonov_2015}.
The present study extends our previous investigation of FeSe
\cite{FeSe_Leonov_2015}. In particular, we now perform fully
charge self-consistent DFT+DMFT calculations to determine the
electronic properties and phase stability of paramagnetic FeSe.
To this end, we take the crystal structure data for the
paramagnetic tetragonal phase of FeSe from experiment \cite{FeSe_structure}
and calculate the total energy as a function of volume. Our
results reveal a substantial change of the total energy upon
inclusion of the effects of charge redistribution caused by
correlation effects. This proves the general importance of
electronic correlations on the charge density and, hence, on
the orbital occupancies. While this influence turns out to
be negligible for the equilibrium volume, it becomes significant
at higher volumes. At the same time the actual results for
the electronic structure and phase stability show no qualitative
difference compared to those calculated without charge
self-consistency \cite{FeSe_Leonov_2015}. Namely, the fully
charge self-consistent calculations still find a structural
phase transition upon expansion of the lattice, which is
associated with a reconstruction of the topology of the Fermi
surface (Lifshitz transition) and is accompanied by a sharp
increase of the local moments. Indeed, our analysis of the
Fermi surface topology and results for the spin susceptibility
$\chi({\bf q})$ support the previously suggested reconstruction
of magnetic correlations from the in-plane magnetic wave vector
$(\pi,\pi)$ to $(\pi,0)$, indicating a competition between
these two magnetic instabilities \cite{NatComm.7.12182}.
Moreover, we find that the individual orbitals contribute
very differently to $\chi({\bf q})$, a fact which may play
a crucial role in explaining the observed nematicity in
Fe(Se,Te) compounds \cite{nematicity}. Our calculations
reveal a pronounced orbital-selective enhancement of
electronic correlation upon expansion of the lattice.
In particular, we observe a crossover from a Fermi-liquid
with a weak self-energy-induced damping at the Fermi energy,
to a non-Fermi-liquid like behavior where the self-energy
almost diverges. The crossover is found to be associated
with a transformation from an itinerant to localized magnetic
moment behavior. Our results clearly demonstrate the crucial
importance of orbital-selective correlations for a realistic
description of the electronic and lattice properties of FeSe.
\section{Method}
In this paper, we employ a state-of-the-art DFT+DMFT computational
scheme \cite{dmft,dftdmft}, which is fully self-consistent in
the charge density, to determine the electronic properties and
phase stability of paramagnetic tetragonal FeSe. It is implemented
\cite{LB08} within the non-spin polarized generalized gradient
approximation (GGA) in DFT using plane-wave pseudopotentials
\cite{pseudopotential}. The approach combines a construction of
the low-energy Hamiltonian for the partially filled Fe $3d$ and
Se $4p$ orbitals in the basis of Wannier functions \cite{WannierH}
with the solution of the DMFT impurity problem using the
continuous-time hybridization-expansion (segment) quantum Monte
Carlo method \cite{ctqmc}. The effects caused by the correlation-induced
charge redistribution are taken into account by solving the DFT+DMFT
equations self-consistently in the charge density.
To investigate the structural stability, we use the atomic positions
and the lattice parameter $c/a$ of paramagnetic tetragonal FeSe
taken from experiment \cite{FeSe_structure}. To this end, we adopt
the crystal structure data (space group $P4/nmm$, the lattice parameter
ratio $c/a$=1.458, and the $z$-position of Se $z$=0.266) and calculate
the total energy as a function of volume. In these calculations,
we consider a uniform expansion or contraction of the lattice
volume, i.e., only the lattice parameter $a$ is varied, while the
$c/a$ ratio is fixed. We use the average Coulomb interaction
$U$~=~3.5~eV and Hund's exchange $J$~=~0.85~eV for the Fe $3d$ shell,
which are typical for the pnictides and chalcogenides according
to different estimations \cite{U_in_superconductors}. The Coulomb
interaction is treated in the density-density approximation.
The spin-orbit coupling is neglected in these calculations.
We employ the fully localized double-counting correction, evaluated
from the self-consistently determined local occupancies, to account
for the electronic interactions already described by DFT. The spectral
functions and angle resolved spectra are evaluated from analytic
continuation of the self-energy using Pad\' e approximants.
We analyze possible magnetic instabilities of FeSe by calculating
the static momentum-dependent susceptibility $\chi({\bf q})$ within
the particle-hole bubble approximation:
\begin{equation}
\chi({\bf q})=-k_{\mathrm{B}}\mathrm{T}\hspace{1mm}\mathrm{Tr}
\sum_{{\bf q},i\omega_{n}}\hat G({\bf k},i\omega_{n})\hat G({\bf k}+{\bf q},i\omega_{n}).
\label{eq_chiofq}
\end{equation}
Here $\mathrm{T}$ is the temperature, $\omega_{n}=(2n+1)\pi k_{\mathrm{B}}\mathrm{T}$
is the Matsubara frequency, $\hat G({\bf k},i\omega_{n})$ is the
interacting lattice Green's function
\begin{equation}
\hat G({\bf k},i\omega_{n})=[(i\omega_{n}+\mu)\hat I - \hat H({\bf k})-\hat \Sigma(i\omega_{n})]^{-1},
\end{equation}
where $\mu$ is the chemical potential, $\hat H({\bf k})$ is the
effective low-energy Hamiltonian in Wannier basis, and $\hat\Sigma(i\omega_{n})$
is the self-energy which includes an energy shift due to the
double-counting correction term.
\section{Results}
\subsection{Phase stability and local magnetic moments}
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth,clip=true]{./Fig_1.eps}
\caption{(Color online)
Total energy (upper panel) and instantaneous local magnetic moments
$\sqrt{\langle m_z^2 \rangle}$ (lower panel) of paramagnetic FeSe as
a function of lattice constant calculated by DFT+DMFT at a temperature
$T=390$ K with (csc) and without (ncsc) charge self-consistency.
The total energy curve obtained with nonmagnetic GGA (nm GGA) is shown
in the upper panel for comparison.
}
\label{Fig_1}
\end{figure}
As a starting point, we compute the electronic structure and phase
stability of paramagnetic FeSe. To this end, we evaluate the total
energy of FeSe as a function of lattice volume by employing a fully
charge self-consistent (csc) DFT+DMFT scheme \cite{V2O3_Leonov_2015,FOR1346_P2_review_2017}
and compare the result with that obtained from non-charge self-consistent
(ncsc) DFT+DMFT calculations \cite{FeSe_Leonov_2015} (Fig. \ref{Fig_1}).
The calculated equilibrium lattice constant $a=7.05$ a.u. at a temperature
$T=390$~K is in good quantitative agreement with the experimental
data \cite{FeSe_structure}, and to a good accuracy coincides with
that obtained within ncsc DFT+DMFT \cite{FeSe_Leonov_2015}. We note
that within the nonmagnetic generalized gradient approximation (GGA)
the equilibrium lattice constant is substantially underestimated \cite{FeSe_Leonov_2015}.
We also observe a substantial change of the total energy when the
correlation-induced charge redistribution is taken into account.
This clearly demonstrates the importance of the feedback of electronic
correlations to the charge density. However, we find that this change
is not very important for the actual value of the equilibrium volume.
It only becomes notable at larger volumes, where it results in a shift
of a lattice anomaly from 7.25 a.u. in the ncsc calculation to 7.45 in
the csc calculations. We also estimate the bulk modulus $K$ for the
equilibrium phase by fitting the obtained energy-volume dependence
using the third-order Birch-Murnaghan equation of state \cite{birch}.
The computed value $K=79$ GPa and its pressure derivative $K' \equiv dK/dP=4.3$
at $T=390$ K are close to those obtained by the ncsc calculations \cite{FeSe_Leonov_2015}.
The computed instantaneous local magnetic moment $\sqrt{\langle m_z^2 \rangle}$
is about 1.9~$\mu_{B}$, corresponding to a fluctuating local magnetic
moment of $\sim$ 0.7~$\mu_B$ \cite{note}. Clearly, it is the inclusion
of the local Coulomb interaction that provides an overall improved
description of the properties of FeSe compared to the DFT results.
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth,clip=true]{./Fig_2.eps}
\caption{(Color online)
Total energy (upper panel) and instantaneous local magnetic moments
$\sqrt{\langle m_z^2 \rangle}$ (lower panel) of paramagnetic FeSe
calculated for different interaction parameters $U$ and $J$ at $T=1160$~K
using the charge self-consistent DFT+DMFT method. The arrows in the
upper panel indicate the position of the energy minima.}
\label{Fig_2}
\end{figure}
Both in the ncsc and csc DFT+DMFT calculations the local magnetic
moment is found to increase upon expansion of the lattice volume
(Fig.~\ref{Fig_1}). We observe that charge self-consistency leads
to a smoother evolution of the local moment and to a reduction of
its absolute value in the whole range of lattice parameters.
Moreover, ncsc and csc calculations both predict an iso-structural
phase transition which is associated with a substantial increase
of the local magnetic moment $\sqrt{\langle m_z^2 \rangle}$ upon
expansion of the lattice. In view of the experimental findings for
the volume and local magnetic moment of FeTe upon compression \cite{FeSe_T_cT},
we interpret this behavior of FeSe as a transition from a
collapsed-tetragonal (equilibrium volume) to tetragonal (expanded
volume) phase which occurs upon expansion of the lattice. The
expansion corresponds to a negative pressure of above $\sim$~-7.6~GPa.
The expanded-volume phase has a significantly smaller bulk modulus
of about 49 GPa. For $a=7.6$ a.u. the calculated local magnetic
moment is $\sim 2.9~\mu_B$ (the fluctuating local moment is $\sim$2.6
$\mu_B$). Our results show that the transition is accompanied by
an increase of the lattice constant from $a=7.35$ a.u. to $a=7.6$~a.u.,
corresponding to an increase of the lattice volume by 11 \%. This
transition persists even if the values of $U$ and $J$ are changed,
as seen in Fig.~\ref{Fig_2}. As expected, a stronger Coulomb
repulsion $U$ between the electrons leads to an increase of the
equilibrium lattice volume. We also observe that for larger $U$
values the phase transition occurs at lower volumes. In any case,
the ncsc and csc calculations both predict a lattice and magnetic
anomaly upon expansion of the unit cell volume. This anomaly is not
found in spin polarized DFT calculations for the $(\pi,0)$ and
$(\pi,\pi)$ antiferromagnetic configurations of iron moments \cite{FeSe_magnetic_DFT},
demonstrating the importance of electronic correlations in FeSe.
\subsection{Spectral properties}
\begin{figure}[t]
\centering
\includegraphics[width=0.35\textwidth,clip=true,angle=-90]{./Fig_3}
\caption{(Color online)
Left panels: Spectral functions of paramagnetic FeSe calculated
within the charge self-consistent DFT+DMFT method (lines) in
comparison with the DFT results (filled areas). Right panels:
imaginary parts of the orbitals contributing to the self-energies
on the Matsubara grid. Top row shows the results obtained for
$a=7.05$ a.u. Bottom row corresponds to $a=7.6$ a.u.
}
\label{Fig_3}
\end{figure}
To explore the mechanism behind this unusual volume dependence we
calculate the spectral properties of FeSe and compare the results
with those obtained from the nsc DFT+DMFT calculations reported
earlier~\cite{FeSe_Leonov_2015}. The spectral functions computed
at the equilibrium volume ($a=7.05$~a.u.) and above the transition
($a=7.6$~a.u.) are shown in Fig.~\ref{Fig_3}. Our results overall
agree with those presented in Ref.~\onlinecite{FeSe_Leonov_2015}.
In particular, we find a substantial renormalization of the Fe $3d$
bands with respect to the DFT results. Indeed, such a behavior is
common for the pnictides and chalcogenides and is in agreement with
previous DFT+DMFT results for FeSe \cite{FeSe_Aichhorn_2010, FeSe_Leonov_2015}.
Upon expansion of the lattice, we observe a strong redistribution
of the spectral weight. In particular, it is seen that the sharp
peak at -0.19 eV below the Fermi energy in the equilibrium volume
phase is absent for larger volumes. This peak originates from the
van Hove singularity of the Fe $xz$/$yz$ and $xy$ bands at the M-point.
Moreover, for both phases the spectrum shows a broad feature at
about -1.2 eV which is associated with the lower Hubbard band \cite{FeSe_Aichhorn_2010,WB16}.
The overall change of the spectral function shape upon expansion
of the lattice agrees well with the evolution of photoemission
spectra of Fe(Se,Te) series obtained upon increase of the Te
content \cite{Yokoya}.
\begin{figure}[t]
\centering
\includegraphics[width=0.47\textwidth,clip=true]{./Fig_4}
\caption{(Color online) {\bf k}-resolved spectral functions
(left column) and partial density of states (right column)
of paramagnetic FeSe calculated by DFT+DMFT for $a=7.05$ a.u.
(red) and $7.6$ a.u. (green). The orbitally-resolved contributions
were evaluated using the maximum entropy method (solid line)
and Pad\'e approximants (broken line).}
\label{Fig_4}
\end{figure}
Next we calculate the {\bf k}-resolved spectral functions of paramagnetic
FeSe along the high-symmetry directions of the Brillouin zone.
In Fig.~\ref{Fig_4} (left panel) we present our results of the
DFT+DMFT calculations for $a=7.05$ a.u. and $a=7.6$ a.u., respectively.
The orbitally-resolved integrated spectral functions are shown
in the right panel of Fig.~\ref{Fig_4}. Our results for the electronic
structure of FeSe are summarized in the left column of Fig.~\ref{Fig_5}.
Upon expansion of the lattice, we observe a remarkable reconstruction
of the electronic structure of FeSe (see Figs.~\ref{Fig_4} and
\ref{Fig_5}) which cannot be described by a simple rescaling or a shift
of the non-correlated DFT band structure. We find that a substantial
part of the spectral weight in the vicinity of $E_{\mathrm F}$ at
the M-point is pushed from below to above the Fermi level, while
the position of the energy bands near the $\Gamma$-point remains
unaffected. This is associated with a correlation-induced shift of
the van Hove singularity at the M-point above the Fermi level and
implies an enhancement of the effect of electron correlations upon
expansion of the lattice of FeSe. We also note that the correlation
effects exhibit a pronounced orbital-dependent character.
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth,clip=true,angle=-90]{./Fig_5}
\caption{(Color online)
Electronic structure (left column) and Fermi surface (right column) of
paramagnetic FeSe in the $\Gamma$-X-M plane of the reciprocal space as
obtained using the charge self-consistent DFT+DMFT method at $T=390$~K.
Top row shows the results for $a=7.05$~a.u. (low volume). Bottom row
corresponds to $a=7.6$ a.u. (high volume)}
\label{Fig_5}
\end{figure}
To analyze this behavior in more detail we evaluate the Fermi surface
of paramagnetic FeSe. In Fig.~\ref{Fig_5} (right column) we display
the contour map of the spectral weight for the plane $k_z=0$ obtained
by integration of the spectral function $A({\bf k},\omega)$ over a
small energy window (5 meV) around the Fermi level. Our results for
the low-volume phase indicate a well-defined (coherent) Fermi surface
(FS) which is similar to that in the pnictides \cite{Nesting_fepn}.
The FS exhibits two elliptic electron-like pockets at the M-point and
two circular concentric hole pockets at the $\Gamma$-point. Similar to
the results obtained from the ncsc calculations \cite{FeSe_Leonov_2015}
the computed FS is characterized by a $(\pi,\pi)$ nesting vector
connecting the electron and hole sheets. A comparison of the calculated
FS of paramagnetic FeSe with experiment shows that the size of the
measured FS pocket is smaller than that obtained within DFT+DMFT.
This is in accordance with previous DFT and DFT+DMFT studies \cite{WB16,WK16},
suggesting, e.g., the importance of non-local correlations effects,
frustration magnetism \cite{GM15}, or spin-orbit interaction effects \cite{BE16}.
Upon expansion of the lattice, we observe an abrupt change of the
topology of the Fermi surface (Lifshitz transition). In particular,
the spectral weight of the electron pockets centered at the M-point
vanishes. The hole pocket encircling the $\Gamma$-point transforms
into a large square-like FS surrounding the M-point with the four
pronounced spots around the $\Gamma$-point. This transition results
in a change of the dominating nesting vector from $(\pi,\pi)$ to
$(\pi,0)$. The observed topological change proceeds similar to the
evolution of the experimental photoemission spectra \cite{FeSe_photoemission_2, FeSe_photoemission_3}
of the doped FeSe$_{1-x}$Te$_{x}$ samples. These data confirm the
emergence of the Fermi surface pocket at the X-point for large
concentrations of Te for $x>0.7$.
\subsection{Orbital-selective renormalization}
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth,clip=true]{./Fig_6}
\caption{(Color online) Orbitally-resolved quasiparticle mass
enhancement $m^{*}/m$ of the Fe $3d$ states in paramagnetic
FeSe as a function of lattice constant calculated by the charge
self-consistent DFT+DMFT approach. The critical region associated
with the electronic and structural transition is indicated by a
red filled rectangle.
}
\label{Fig_6}
\end{figure}
An expansion of the lattice also goes along with a remarkable
orbital-selective renormalization of the Fe $3d$ bands (see Fig~\ref{Fig_6}),
indicating significantly stronger renormalization of the $t_{2}$
bands ($xz/yz$ and $xy$) than of the $e$ bands ($3z^2-r^2$ and $x^2-y^2$).
In Fig.~\ref{Fig_3} (right column) we show the Fe $3d$ imaginary
self-energies for the low- and high-volume phases, respectively.
At the equilibrium volume, the self-energy obeys a Fermi-liquid-like
behavior characterized by a weak damping of quasiparticles. By contrast,
the expanded-volume phase shows a pronounced orbital-selective
behavior, associated with a non-Fermi-liquid behavior of the $t_{2}$
orbitals. Indeed, the self-energies of the $t_{2}$ orbitals decrease
with decreasing Matsubara frequency -- and in the case of the
self-energy of the $xy$-orbital even seems to diverge -- but finally
show an upturn at the lowest Matsubara frequency. At the same time,
the $e$ states remain Fermi-liquid-like, but with a damping which
is about five times stronger than that in the equilibrium phase.
These results agree well with an analysis of the band mass enhancement
${m^*}/{m}=1-\partial \mathrm{ Im }\Sigma(\omega)/\partial\omega|_{\omega=0}$,
which provides a quantitative measure of the correlation strength.
In Fig.~\ref{Fig_6} we display the computed mass enhancement $m^*/m$
as a function of lattice constant. In the vicinity of the equilibrium
lattice constant $m^{*}/m$ lies in the range $1.5$--$2$. Upon
expansion of the lattice it shows a substantial increase followed
by a critical region at $a\sim 7.5$ a.u. (where the electronic and
structural transition occurs), which is characterized by a change
of the sign of its derivative. Furthermore, the effective mass of
the $t_{2}$ electrons exhibit larger renormalizations than in the
$e$ orbitals. Indeed, for the former it reaches $\sim$ 6.5 and 4.5
for the Fe $xy$ and $xz/yz$ states, respectively.
\begin{figure}[t]
\centering
\includegraphics[width=0.47\textwidth,clip=true]{./Fig_7}
\caption{(Color online) Orbitally-resolved local spin correlation
functions $\chi(\tau) = \langle \hat{m}_z(\tau) \hat{m}_z(0)\rangle$
of paramagnetic FeSe calculated using DFT+DMFT for the lattice constant
$a=7.05$ a.u. (left) and $a=7.60$ a.u. (right).
}
\label{Fig_7}
\end{figure}
\subsection{Susceptibility}
The electronic and structural phase transition is accompanied by a
significant growth of the fluctuating local magnetic moment (see
lower panel of Fig.~\ref{Fig_1}). The transition is found to result
in a crossover from an itinerant to localized moment behavior, as it
is seen from the local spin susceptibility
$\chi(\tau ) = \langle \hat{m}_z(\tau) \hat{m}_z(0) \rangle$, where
$\tau$ is the imaginary time. The results for the different orbital
contributions are presented in Fig. \ref{Fig_7}. This behavior is
consistent with the coherence-incoherence transition scenario which
was found experimentally in the Fe(Se,Te) series~\cite{FeSe_photoemission_3}.
Moreover, our calculations reveal a strong orbital-selectivity in
the formation of the local moments upon expansion of the lattice of
FeSe. Here the $xy$ orbital plays a predominant role, while the
contribution of the $xz/yz$ orbitals is substantially weaker. On the
other hand, the $e$ orbitals exhibit an itinerant moment behavior.
\begin{figure}[t]
\centering
\includegraphics[width=0.2\textwidth,clip=true,angle=-90]{./Fig_8}
\caption{(Color online)
Momentum dependence of the local spin susceptibility $\chi({\bf q})$
of paramagnetic FeSe calculated using the charge self-consistent
DFT+DMFT method for $a=7.05$ a.u. (left) and $a=7.6$ a.u. (right).}
\label{Fig_8}
\end{figure}
In addition, we compute the momentum-dependent local spin
susceptibility $\chi({\bf q})$ in the $(q_x,q_y)$ plane for $q_z=0$.
Our results are presented in Fig.~\ref{Fig_8}. The susceptibility
calculated for the equilibrium volume shows a maximum at the
corners of the tetragonal Brilloiun zone at the M-points. This
confirms that the leading magnetic instability of FeSe at ambient
pressure occurs at the wave vector $(\pi,\pi)$, in agreement with
experiment \cite{FeSe_pi_pi_fluctuations}. An expansion of the
lattice volume leads to a dramatic change of $\chi({\bf q})$,
associated with a suppression of the maximum at $(\pi,\pi)$ and
the development of a maximum at $(\pi,0)$. This change of the
magnetic correlations is associated with the change of the Fermi
surface (Lifshitz transition) discussed above. The evolution of
$\chi({\bf q})$ qualitatively agrees with the experimentally
observed transformation of magnetic correlations in the Fe(Se,Te)
series \cite{FeSe_magnetism}. Indeed, our results show a transition
from $(\pi,\pi)$-type antiferromagnetic fluctuations in the
paramagnetic tetragonal phase of FeSe to $(\pi,0)$-type magnetism
upon expansion of the lattice.
Moreover, we calculate the orbital contributions of $\chi({\bf q})$
along the $\Gamma$-X-M-$\Gamma$ path (Fig.~\ref{Fig_9}). For $a=7.05$~a.u.
we observe a strong orbital-selective behavior of magnetic correlations
with a leading contribution originating from the Fe $xy$ orbital. This
orbital leads to a maximum of $\chi({\bf q})$ at the M-point, confirming
that magnetic correlations in FeSe are predominantly of the $(\pi,\pi)$-type.
On the other hand, the behavior of $\chi({\bf q})$ in the high volume
phase is completely different. In particular, for $a=7.6$~a.u. the leading
contribution to $\chi({\bf q})$ is due to the Fe $3z^2-r^2$ orbital which
varies only weakly along the $\Gamma$-X-M-$\Gamma$ path. Our analysis shows
that only the inclusion of all orbital contributions (especially of the
$x^2-y^2$ orbital contribution, which exhibits the most substantial variation
in the reciprocal space and shows a maximum at the X-point) results in
the $(\pi,0)$-type magnetic correlations prevalent in the high-volume
phase of Fe(Se,Te).
\begin{figure}[t]
\centering
\includegraphics[width=0.47\textwidth,clip=true]{./Fig_9}
\caption{(Color online) Orbitally-resolved local spin susceptibility
$\chi({\bf q})$ of paramagnetic FeSe calculated along the $\Gamma$-X-M-$\Gamma$
path using the charge self-consistent DFT+DMFT for $a=7.05$ a.u. (left)
and $a=7.6$ a.u. (right).}
\label{Fig_9}
\end{figure}
Our results for $\chi({\bf q})$ in Fig.~\ref{Fig_9} demonstrate that
for $a=7.05$ a.u. the $xz$ and $yz$ orbitals contribute very differently
to $\chi( {\bf q} )$ along the $\Gamma$-X-M direction. It will be
interesting to check whether this finding, together with the symmetry-induced
splitting between the $xz/yz$ orbitals at the X point, can stabilize
the observed nematicity in FeSe \cite{nematicity}, for example through
the coupling of magnetic fluctuations to phonons near the X point.
\section{Conclusion}
In conclusion, we studied the electronic structure and phase
stability of the tetragonal paramagnetic phase of FeSe using
a fully charge self-consistent implementation of the DFT+DMFT
method. Our results demonstrate the importance of electron
correlation effects which, in particular, trigger the anomalous
behavior of FeSe upon expansion of the lattice volume. We note
that such an expansion can be experimentally realized by the
isovalent substitution of Se with Te. Our results also reveal
a complete change of the electronic structure of paramagnetic
FeSe upon a moderate expansion of the lattice (at -7.6 GPa).
This behavior is associated with a remarkable reconstruction
of the Fermi surface topology (Lifshitz transition) of FeSe and
is accompanied by a change of the in-plane magnetic nesting
vector from $(\pi,\pi)$ to $(\pi,0)$, in agreement with experiment \cite{FeSe_magnetism}.
This behavior is intimately linked with an orbital-selective
transition from itinerant to localized moment behavior, where
the Fe $xy$ orbitals contribute most strongly. The phase transformation
is driven by a correlation-induced shift of the van Hove
singularity of the Fe $t_{2}$ bands at the M-point across the
Fermi level \cite{CE15}. We also observe a strong orbital-selective
renormalization of the Fe $3d$ band structure, with the largest
contribution coming again from the Fe $xy$ orbital, which gives
rise to a non-Fermi-liquid-like behavior above the transition. \cite{WS14}
In view of our results the complex behavior of the chalcogenide
parent system Fe(Se,Te), such as the anomalous increase of the
superconducting temperature upon positive or negative pressure,
appears to be associated with the proximity of the van Hove
singularity of the Fe $t_{2}$ bands at the M-point to the Fermi
level, and with the sensitivity of its position to external
conditions \cite{CE15}. Furthermore, our results for the local
spin susceptibility $\chi({\bf q})$, which exhibits a strong
splitting between the $xz$ and $yz$ orbitals near the X-point,
suggest a spin-fluctuation origin of the nematic phase of
paramagnetic FeSe. This will be the subject of further investigations.
\section{Acknowledgments}
We thank V. Tsurkan, J. Schmalian, and L. H. Tjeng for useful discussions.
I.L. acknowledges support from the Deutsche Forschungsgemeinschaft
through Transregio TRR 80 and the Ministry of Education and Science
of the Russian Federation in the framework of Increase Competitiveness
Program of NUST "MISIS" (K3-2016-027), implemented by a governmental
decree dated 16th of March 2013, N 211. D.V., S.L.S. and V.I.A. are
grateful to the Deutsche Forschungsgemeinschaft for financial support
through the Research Unit FOR 1346.
|
\section{Introduction}
\label{section:introduction}
At the head of the list of discrete Painlev\'e equations described by Sakai sits an elliptic difference equation, which has attracted a great deal of attention in recent times.
This equation \cite{SakaiH2001:MR1882403,MSY2003:MR1958273} has the affine Weyl group symmetry of type $E_8^{(1)}$ and a perennial question is whether it is unique as the only elliptic-difference-type discrete Painlev\'e equation.
While two other candidate equations are now known \cite{ORG2001:MR1877472,RCG2009:MR2525848}, these are related to Sakai's elliptic difference equation.
In this paper, we deduce elliptic-difference equations for the first time, which are not related to Sakai's equation by either a Miura transformation or by projective reduction.
We do this by considering different (non-conjugate) translations on the $E_8^{(1)}$ lattice and describing these translations under the action of Jacobian elliptic functions.
A discrete Painlev\'e equation is an ordinary difference equation, which is iterated by translation in an affine Weyl group, where translation corresponds to vectors in the root- or (equivalently) weight-lattice \cite{book_CSBBLNOPQV2013:Sphere,book_HumphreysJE1992:Reflection}.
Sakai's elliptic difference equation is realized by translation on the $E_8^{(1)}$ lattice.
To describe the construction, we first explain how such translations are characterized.
Fix a point in the $E_8^{(1)}$ lattice.
Then there are 240 nearest neighbors of this point in the lattice, lying at a distance whose squared length is equal to 2.
We refer to the 120 vectors between this fixed point and its possible nearest neighbors as nearest-neighbor-connecting vectors (NVs).
Similarly, there are 2160 next-nearest neighbors, lying at a distance whose squared length is 4.
The 1080 vectors between the fixed point and such next-nearest neighbors will be referred to as next-nearest-neighbor-connecting vectors (NNVs).
In \cite{SakaiH2001:MR1882403} and \cite{ORG2001:MR1877472}, elliptic difference equations were constructed as translations expressed in terms of NVs.
In a previous study \cite{AHJN2016:MR3509963}, we considered another elliptic difference equation (Equation \eqref{eqns:RCGeqn}), which was obtained by reduction from a partial difference equation \cite{RCG2009:MR2525848}.
Curiously, its symmetry group turns out to be $W(F_4^{(1)})$, a sub-group of $W(E_8^{(1)})$.
Moreover, its time iteration turns out not to be given by translation on the $E_8^{(1)}$ lattice.
However, its square (i.e., composition with itself) {\em is} a translation.
We have discovered that this translation is expressed in terms of NNVs, which makes it very different to the elliptic difference equations in earlier papers \cite{SakaiH2001:MR1882403,ORG2001:MR1877472}.
Note that independent study by Carstea {\it et al.} \cite{CDT2017:arXiv170204907} has also led to equations with translations given by NNVs.
Generalizing this surprising insight led us to the discovery of a new elliptic Painlev\'e equation \eqref{eqns:TJ1}, which can be regarded as the generic version of Equation \eqref{eqns:RCGeqn}.
The term generic here refers to the fact that it contains the largest number of parameters possible as an equation with symmetry group $W(E_8^{(1)})$.
It has 8 parameters in addition to the independent variable.
Moreover, we also obtain a projective reduction of Sakai's elliptic difference equation \eqref{eqn:MSY_dP_PR}, which contains 4 parameters in addition to the independent variable.
All of these elliptic difference equations have the same space of initial values, regularized by blowing up 8 points in arbitrary position in $\mathbb{P}^1\times\mathbb{P}^1$.
The space of initial values is labelled as $A_0^{(1)}$.
We emphasize here that this characterization of the space of initial values is not enough to distinguish the four different elliptic difference equations described above.
We provide evidence that different (non-conjugate) translations in $\widetilde{W}\big(E_8^{(1)}\big)$ lead to distinct elliptic difference equations.
All four elliptic difference equations are described in this paper in terms of Jacobian elliptic functions for the sake of uniformity, because this is the form originally given for Equation \eqref{eqns:RCGeqn}.
\subsection{Background}
Discrete Painlev\'e equations are nonlinear ordinary difference equations of second order,
which include discrete analogues of the six Painlev\'e equations: P$_{\rm I}$, $\dots$, P$_{\rm VI}$.
Together with the Painlev\'e equations, the discrete Painlev\'e equations are now regarded as one of the most important classes of equations in the theory of integrable systems (see, e.g., \cite{GR2004:MR2087743}).
Sakai's geometric description of discrete Painlev\'e equations, based on types of space of initial values,
is well known\cite{SakaiH2001:MR1882403}.
This picture relies on compactifying and regularizing space of initial values.
The spaces of initial values are constructed by the blow up of $\mathbb{P}^1\times\mathbb{P}^1$ at eight base points
(i.e. points where the system is ill defined because it approaches $0/0$)
and are classified into 22 types according to the configuration of the base points as follows:
\begin{center}
\begin{tabular}{|l|l|}
\hline
Discrete type&Type of surface\\
\hline
Elliptic&$A_0^{(1)}$\rule[-.5em]{0em}{1.6em}\\
\hline
Multiplicative&$A_0^{(1)\ast}$, $A_1^{(1)}$, $A_2^{(1)}$, $A_3^{(1)}$, \dots, $A_8^{(1)}$, $A_7^{(1)'}$\rule[-.5em]{0em}{1.6em}\\
\hline
Additive&$A_0^{(1)\ast\ast}$, $A_1^{(1)\ast}$, $A_2^{(1)\ast}$, $D_4^{(1)}$, \dots, $D_8^{(1)}$, $E_6^{(1)}$, $E_7^{(1)}$, $E_8^{(1)}$\rule[-.5em]{0em}{1.6em}\\
\hline
\end{tabular}
\end{center}
In each case, the root system characterizing the surface forms a subgroup of the 10-dimensional Picard group.
The symmetry group of each equation, formed by Cremona isometries, arises from the orthogonal complement of this root system inside the Picard group.
Its birational actions give the discrete Painlev\'e equation of interest in each case.
Recently, the following elliptic Painlev\'e equation was obtained from the reduction of Adler's partial difference equation (or, Q4 equation)\cite{RCG2009:MR2525848}:
\begin{subequations}\label{eqns:RCGeqn}
\begin{align}
&\tilde{y}=\frac{(1-k^2{\rm sz}^4){\rm cg}_{\rm e}{\rm dg}_{\rm e}\, xy-({\rm cg}_{\rm e}^2-{\rm cz}^2){\rm cz}\,{\rm dz}-(1-k^2{\rm sg}_{\rm e}^2 {\rm sz}^2){\rm cz}\,{\rm dz}\, x^2}
{k^2({\rm cg}_{\rm e}^2-{\rm cz}^2){\rm cz}\,{\rm dz}\, x^2 y-(1-k^2 {\rm sz}^4){\rm cg}_{\rm e}{\rm dg}_{\rm e}\, x+(1-k^2{\rm sg}_{\rm e}^2{\rm sz}^2){\rm cz}\,{\rm dz}\,y},\\
&\tilde{x}=\frac{(1-k^2{\rm \widehat{sz}}^4){\rm cg}_{\rm o}{\rm dg}_{\rm o}\, \tilde{y} x-({\rm cg}_{\rm o}^2-{\rm \widehat{cz}}^2){\rm \widehat{cz}}\,{\rm \widehat{dz}}-(1-k^2{\rm sg}_{\rm o}^2 {\rm \widehat{sz}}^2){\rm \widehat{cz}}\,{\rm \widehat{dz}}\,\tilde{y}^2}
{k^2({\rm cg}_{\rm o}^2-{\rm \widehat{cz}}^2){\rm \widehat{cz}}\,{\rm \widehat{dz}}\, \tilde{y}^2x-(1-k^2 {\rm \widehat{sz}}^4){\rm cg}_{\rm o}{\rm dg}_{\rm o}\, \tilde{y}+(1-k^2{\rm sg}_{\rm o}^2 {\rm \widehat{sz}}^2){\rm \widehat{cz}}\,{\rm \widehat{dz}}\, x},
\end{align}
\end{subequations}
where $k$ is the modulus of the elliptic sine,
\begin{subequations}
\begin{align}
&{\rm sz}=\sn{z_0},\quad
{\rm \widehat{sz}}=\sn{z_0+\gamma_{\rm e}+\gamma_{\rm o}},
&&{\rm sg}_{\rm e}=\sn{\gamma_{\rm e}},\quad
{\rm sg}_{\rm o}=\sn{\gamma_{\rm o}},\\
&{\rm cz}=\cn{z_0},\quad
{\rm \widehat{cz}}=\cn{z_0+\gamma_{\rm e}+\gamma_{\rm o}},
&&{\rm cg}_{\rm e}=\cn{\gamma_{\rm e}},\quad
{\rm cg}_{\rm o}=\cn{\gamma_{\rm o}},\\
&{\rm dz}=\dn{z_0},\quad
{\rm \widehat{dz}}=\dn{z_0+\gamma_{\rm e}+\gamma_{\rm o}},
&&{\rm dg}_{\rm e}=\dn{\gamma_{\rm e}},\quad
{\rm dg}_{\rm o}=\dn{\gamma_{\rm o}},
\end{align}
\end{subequations}
and
\begin{equation}
\tilde{}\,:(\gamma_{\rm e},\gamma_{\rm o},z_0)\mapsto \big(\gamma_{\rm e},\gamma_{\rm o},z_0+2(\gamma_{\rm e}+\gamma_{\rm o})\big).
\end{equation}
See Appendix \ref{section:Appendix_A} for standard results about Jacobian elliptic functions.
In \cite{AHJN2016:MR3509963}, the geometry of Equation \eqref{eqns:RCGeqn}, i.e., its space of initial values and corresponding Cremona isometries, was investigated.
The space of initial values of Equation \eqref{eqns:RCGeqn} was identified with the elliptic $A_0^{(1)}$-surface and its Cremona isometries collectively form an affine Weyl transformation group of type $E_8^{(1)}$, denoted by $W\big(E_8^{(1)}\big)$.
Moreover, it was shown that Equation \eqref{eqns:RCGeqn} cannot be derived from a translation of $W\big(E_8^{(1)}\big)$ but can be derived by using a projective reduction.
The process of deriving discrete dynamical systems of Painlev\'e type from elements of affine Weyl groups that are of infinite order (but that are not necessarily translations) by taking a projection on an appropriate subspace of parameters is called a projective reduction \cite{KNT2011:MR2773334,KN2015:MR3340349}.
Note that this process is motivated by taking symmetric versions of systems of discrete Painlev\'e equations, which has been known since \cite{RGH1991:MR1125951}.
Although the geometry of Equation \eqref{eqns:RCGeqn} has been clarified,
the realization of Equation \eqref{eqns:RCGeqn} from the action of Cremona isometries was missing
since its base points are parametrized by the Jacobian elliptic function (Jacobi's setting)
and birational actions of Cremona isometries on such setting were not explicitly known.
The present study fills this gap (see Theorem \ref{theorem:birational_tWE8}), that is, we provide the realization of the Equation \eqref{eqns:RCGeqn}
as a half-translation of the extended affine Weyl group $\widetilde{W}\big(E_8^{(1)}\big)$.
In the remainder of the paper, we refer to Sakai's elliptic Painlev\'e equation \cite{SakaiH2001:MR1882403}
as the MSY elliptic Painlev\'e equation \cite{MSY2003:MR1958273} because the former was obtained in $\mathbb{P}^2$
while the latter was provided by resolution in $\mathbb{P}^1\times\mathbb{P}^1$.
We work throughout the paper in $\mathbb{P}^1\times\mathbb{P}^1$.
\subsection{Plan of the paper}
This paper is organized as follows.
In \S \ref{section:cremona}, we construct Cremona isometries for base points \eqref{eqn:basepoints} that generalize those of Equation \eqref{eqns:RCGeqn} and show that these form an affine Weyl group of type $\EE8^{(1)}$, denoted by $W\big(E_8^{(1)}\big)$, under the linear actions on the Picard group.
In \S \ref{section:birational}, using intersection theory, we obtain the birational action of $W\big(E_8^{(1)}\big)$ on the coordinates and parameters of the base points, and we prove that these birational actions also satisfy the fundamental relations of an affine Weyl group of type $\EE8^{(1)}$.
Adding the identity mappings on the Picard group, we obtain the extended affine Weyl group $\widetilde{W}\big(E_8^{(1)}\big)$.
Equation \eqref{eqns:RCGeqn} and three other elliptic Painlev\'e equations are then deduced from the action of this resulting group $\widetilde{W}\big(E_8^{(1)}\big)$.
Finally, we give some concluding remarks in \S \ref{ConcludingRemarks}.
\section{Cremona isometries}
\label{section:cremona}
In this section, we construct Cremona isometries for a generalization of the base points of Equation \eqref{eqns:RCGeqn} and show that these isometries collectively form an affine Weyl group of type $\EE8^{(1)}$.
Recall that Equation \eqref{eqns:RCGeqn} has the following eight base points (see \cite{AHJN2016:MR3509963}):
\begin{subequations}\label{eqn:basepoints_RCG}
\begin{align}
&p_1:(x,y)=\big(\cd{\gamma_{\rm o}+\kappa},\cd{z_0-\gamma_{\rm e}-\gamma_{\rm o}+\kappa}\big),\\
&p_2:(x,y)=\big(\cd{\gamma_{\rm o}+{\rm i} K'},\cd{z_0-\gamma_{\rm e}-\gamma_{\rm o}+{\rm i} K'}\big),\\
&p_3:(x,y)=\big(\cd{\gamma_{\rm o}+2K},\cd{z_0-\gamma_{\rm e}-\gamma_{\rm o}+2K}\big),\\
&p_4:(x,y)=\big(\cd{\gamma_{\rm o}},\cd{z_0-\gamma_{\rm e}-\gamma_{\rm o}}\big),\\
&p_5:(x,y)=\big(\cd{z_0+\kappa},\cd{\gamma_{\rm e}+\kappa}\big),\\
&p_6:(x,y)=\big(\cd{z_0+{\rm i} K'},\cd{\gamma_{\rm e}+{\rm i} K'}\big),\\
&p_7:(x,y)=\big(\cd{z_0+2K},\cd{\gamma_{\rm e}+2K}\big),\\
&p_8:(x,y)=\big(\cd{z_0},\cd{\gamma_{\rm e}}\big),
\end{align}
\end{subequations}
where $K=K(k)$ and $K'=K'(k)$ are complete elliptic integrals and
\begin{equation}\label{eqn:kappa}
\kappa=2K+{\rm i} K',
\end{equation}
which lie on the elliptic curve
\begin{equation}\label{eqn:RCG_curve}
\sn{z_0-\gamma_{\rm e}}^2(1+k^2x^2y^2)+2\cn{z_0-\gamma_{\rm e}}\dn{z_0-\gamma_{\rm e}}xy-(x^2+y^2)=0.
\end{equation}
Note that these base points contain three arbitrary parameters $z_0$, $\gamma_{\rm e}$, $\gamma_{\rm o}$.
We now generalize the number of arbitrary parameters.
Take a new set of base points given by
\begin{equation}\label{eqn:basepoints}
p_i:(x,y)=\big(\cd{c_i+\eta},\cd{\eta-c_i}\big),\quad i=1,\dots,8,
\end{equation}
where $c_i$, $i=1,\dots,8$, and $\eta$ are non-zero complex parameters.
These generalized base points lie on the elliptic curve
\begin{equation}\label{eqn:Jelliptic_curve}
\sn{2\eta}^2(1+k^2x^2y^2)+2\cn{2\eta}\dn{2\eta}xy-(x^2+y^2)=0.
\end{equation}
These base points can be reduced to the ones given in Equation \eqref{eqn:basepoints_RCG} under a specialization of the parameters.
Indeed, the points \eqref{eqn:basepoints} and elliptic curve \eqref{eqn:Jelliptic_curve}
can be respectively reduced to the points \eqref{eqn:basepoints_RCG} and the curve \eqref{eqn:RCG_curve} by assuming
\begin{subequations}\label{eqns:condition_RCG}
\begin{align}
&\cc2=\cc1+2K,\quad
\cc3=\cc1+{\rm i} K',\quad
\cc4=\cc1+\kappa,\\
&\cc6=\cc5+2K,\quad
\cc7=\cc5+{\rm i} K',\quad
\cc8=\cc5+\kappa,
\end{align}
and letting
\begin{equation}
z_0=\eta+\cc5+\kappa,\quad
\gamma_{\rm e}=\cc5-\eta+\kappa,\quad
\gamma_{\rm o}=\eta+\cc1+\kappa.
\end{equation}
\end{subequations}
\begin{remark}
To distinguish between Weierstrass's and Jacobi's setting,
we here denote the surface characterized by the elliptic curve \eqref{eqn:Jelliptic_curve} as $A_0^{(1)J}$-surface.
The relation between $A_0^{(1)}$- and $A_0^{(1)J}$-surface is given in Appendix \ref{section:Appendix_B}.
\end{remark}
Let $\epsilon: X \to \mathbb{P}^1\times\mathbb{P}^1$ denote the blow up of $\mathbb{P}^1\times\mathbb{P}^1$ at the points \eqref{eqn:basepoints}. Moreover, let the linear equivalence classes of the total transform of vertical and horizontal lines in $\mathbb{P}^1\times\mathbb{P}^1$ be denoted respectively by $H_0$ and $H_1$.
The Picard group of $X$, denoted by Pic$(X)$, is given by
\begin{equation}
{\rm Pic}(X)=\mathbb{Z} H_0\bigoplus\mathbb{Z} H_1\bigoplus_{i=1}^8\mathbb{Z} \EE{i},
\end{equation}
where $\EE{i}=\epsilon^{-1}(p_i)$, $i=1,\dots,8$, are exceptional divisors.
The intersection form $(\,|\,)$ is given by a symmetric bilinear form with
\begin{equation}
(H_i|H_j)=1-\delta_{ij},\quad
(H_i|\EE{j})=0,\quad
(\EE{i}|\EE{j})=-\delta_{ij},
\end{equation}
where $\delta_{ij}$ is the Kronecker delta.
The anti-canonical divisor of $X$ is given by
\begin{equation}
-K_X=2H_0+2H_1-\sum_{i=1}^8\EE{i}.
\end{equation}
For later convenience, let
\begin{equation}
\delta=-K_X.
\end{equation}
We define the root lattice $Q(A_0^{(1)\bot})=\bigoplus_{i=0}^8\mathbb{Z}\alpha_i$ by the elements of Pic$(X)$
that are orthogonal to the anti-canonical divisor $\delta$.
The simple roots $\alpha_i$, $i=0,\dots,8$, are given by
\begin{subequations}\label{eqns:simpleroots_alpha}
\begin{align}
&\alpha_1=H_1-H_0,\quad
\alpha_2=H_0-\EE1-\EE2,\quad
\alpha_i=\EE{i-1}-\EE{i},\quad i=3,\dots,7,\\
&\alpha_8=\EE1-\EE2,\quad
\alpha_0=\EE7-\EE8,
\end{align}
\end{subequations}
where
\begin{equation}
\delta=2\alpha_1+4\alpha_2+6\alpha_3+5\alpha_4+4\alpha_5+3\alpha_6+2\alpha_7+3\alpha_8+\alpha_0.
\end{equation}
We can easily verify that
\begin{equation}
(\alpha_i|\alpha_j)
=\begin{cases}
-2,& i=j\\
\ 1, &i=j-1\quad (j=2,\dots,7),\quad \text{or\hspace{0.5em}if}\quad (i,j)=(3,8),(7,0)\\
\ 0, &\text{otherwise}.
\end{cases}
\end{equation}
Representing intersecting $\alpha_i$ and $\alpha_j$ by a line between nodes $i$ and $j$, we obtain the Dynkin diagram of $E_8^{(1)}$ shown in Figure \ref{fig:dynkin_E8}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{affine_E8.eps}
\caption{Dynkin diagram for the root lattice $\bigoplus_{i=0}^8\mathbb{Z}\alpha_i$.}
\label{fig:dynkin_E8}
\end{center}
\end{figure}
\begin{definition}{\rm \cite{DolgachevIV2012:MR2964027,DO1988:MR1007155,LooijengaE1981:MR632841}}\label{def:cremona}
An automorphism of Pic$(X)$ is called a Cremona isometry if it preserves
\begin{description}
\item[(i)]
the intersection form $(\,|\,)$ on Pic$(X)$;
\item[(ii)]
the canonical divisor $K_X$;
\item[(iii)]
effectiveness of each effective divisor of Pic$(X)$.
\end{description}
\end{definition}
It is well-known that the reflections are Cremona isometries.
In this case we define the reflections $s_i$, $i=0,\dots,8$, by the following linear actions:
\begin{equation}\label{eqn:def_si}
s_i.v=v-\cfrac{2(v|\alpha_i)}{(\alpha_i|\alpha_i)}\,\alpha_i,
\end{equation}
for all $v\in {\rm Pic}(X)$.
They collectively form an affine Weyl group of type $\EE8^{(1)}$, denoted by $W\big(E_8^{(1)}\big)$.
Namely, we can easily verify that under the action on the ${\rm Pic}(X)$ the following fundamental relations hold:
\begin{equation}\label{eqns:fundamental_We8}
(s_is_j)^{l_{ij}}=1,
\end{equation}
where
\begin{equation}
l_{ij}=
\begin{cases}
1,& i=j\\
3, &i=j-1\quad (j=2,\dots,7),\quad \text{or\hspace{0.5em}if}\quad (i,j)=(3,8),(7,0)\\
2, &\text{otherwise}.
\end{cases}
\end{equation}
\section{Birational actions of the Cremona isometries for the Jacobi's setting}
\label{section:birational}
In this section, we give the birational actions of the Cremona isometries on the coordinates and parameters of the base points \eqref{eqn:basepoints}.
By using these birational actions, we derive various elliptic Painlev\'e equations.
We focus on a particular example first to explain how to deduce such birational actions.
Recall $H_0$ and $H_1$ are given by the linear equivalence classes of vertical lines $x=\text{constant}$ and horizontal lines $y=\text{constant}$, respectively.
Applying the reflection operator $s_2$ given by \eqref{eqn:def_si} to $H_1$, we find
\begin{equation}
s_2.H_1=H_0+H_1-E_1-E_2,
\end{equation}
which means that $s_2(y)$ can be described by the curve of bi-degree $(1,1)$ passing through base points $p_1$ and $p_2$ with multiplicity $1$.
(See \cite{KNY2015:arXiv150908186K} for for more detail.)
This result leads us to the birational action given below in Equation \eqref{eqn:action_E8_J_y}.
Similarly, from the linear actions of $s_i$, $i=0,\dots,8$, we obtain their birational actions on the coordinates and parameters of the base points \eqref{eqn:basepoints} as follows.
The actions of the generators of $W\big(E_8^{(1)}\big)$ on the coordinates $(x,y)$ are given by
\begin{subequations}\label{eqns:action_E8_J_para_xy}
\begin{align}
&s_1(x)=y,\quad
s_1(y)=x,\\
&\bfrac{s_2(y)-\cd{2\eta-\frac{\cc1-\cc2}{2}}}{s_2(y)-\cd{2\eta+\frac{\cc1-\cc2}{2}}}
\bfrac{x-\cd{\eta+\cc1}}{x-\cd{\eta+\cc2}}
\bfrac{y-\cd{\eta-\cc2}}{y-\cd{\eta-\cc1}}\notag\\
&\qquad =\bfrac{1-\frac{\cd{\eta-\cc2}}{\cd{\eta}}}{1-\frac{\cd{\eta-\cc1}}{\cd{\eta}}}
\bfrac{1-\frac{\cd{\eta+\cc1}}{\cd{\eta}}}{1-\frac{\cd{\eta+\cc2}}{\cd{\eta}}}
\bfrac{1-\frac{\cd{2\eta-\frac{\cc1-\cc2}{2}}}{\cd{\frac{\cc1+\cc2}{2}}}}{1-\frac{\cd{2\eta+\frac{\cc1-\cc2}{2}}}{\cd{\frac{\cc1+\cc2}{2}}}},
\label{eqn:action_E8_J_y}
\end{align}
while those on the parameters $\cc{i}$, $i=1,\dots,8$, and $\eta$ are given by
\begin{align}
&s_0(\cc7)=\cc8,\quad
s_0(\cc8)=\cc7,\quad
s_1(\eta)=-\eta,\\
&s_2(\eta)=\eta-\frac{2\eta+\cc1+\cc2}{4},\quad
s_2(\cc{i})=
\begin{cases}
\cc{i}-\dfrac{3(2\eta+\cc1+\cc2)}{4},& i=1,2\\
\cc{i}+\dfrac{2\eta+\cc1+\cc2}{4},& i\neq 1,2,
\end{cases}\\
&s_k(\cc{k-1})=\cc{k},\quad
s_k(\cc{k})=\cc{k-1},\quad
k=3,\dots,7,\\
&s_8(\cc1)=\cc2,\quad
s_8(\cc2)=\cc1.
\end{align}
\end{subequations}
Note that
\begin{equation}
\lambda=\sum_{i=1}^8\cc{i}
\end{equation}
is invariant under the action of $W\big(E_8^{(1)}\big)$.
For Jacobi's elliptic function $\cd{u}$ it is well known that shifts by half periods give the following relations:
\begin{equation}
\cd{u+2K}=-\cd{u},\quad
\cd{u+{\rm i} K'}=\dfrac{1}{k\, \cd{u}}.
\end{equation}
These identities motivate our search for the transformations that are identity mappings on the ${\rm Pic}(X)$.
Indeed, we define such transformations $\iota_i$, $i=1,\dots,4$, by the following actions:
\begin{subequations}\label{eqns:action_iota}
\begin{align}
&\iota_1:(\cc{1},\dots,\cc{8},\eta,x,y)
\mapsto\left(\cc{1}-\frac{{\rm i} K'}{2},\dots,\cc{8}-\frac{{\rm i} K'}{2},\eta-\frac{{\rm i} K'}{2},\frac{1}{kx},y\right),\\
&\iota_2:(\cc{1},\dots,\cc{8},\eta,x,y)
\mapsto\left(\cc{1}-\frac{{\rm i} K'}{2},\dots,\cc{8}-\frac{{\rm i} K'}{2},\eta+\frac{{\rm i} K'}{2},x,\frac{1}{ky}\right),\\
&\iota_3:(\cc{1},\dots,\cc{8},\eta,x,y)
\mapsto\left(\cc{1}-K,\dots,\cc{8}-K,\eta-K,-x,y\right),\\
&\iota_4:(\cc{1},\dots,\cc{8},\eta,x,y)
\mapsto\left(\cc{1}-K,\dots,\cc{8}-K,\eta+K,x,-y\right).
\end{align}
\end{subequations}
Adding the transformations $\iota_i$, we extend $W\big(E_8^{(1)}\big)$ to
\begin{equation}
\widetilde{W}\big(E_8^{(1)}\big)=\lrangle{\iota_1,\iota_2,\iota_3,\iota_4}\rtimesW\big(E_8^{(1)}\big).
\end{equation}
In general, for a function $F=F(c_i,\eta,x,y)$, we let an element
$w\in\widetilde{W}\big(E_8^{(1)}\big)$ act as $w.F=F(w.c_i,w.\eta,w.x,w.y)$, that is,
$w$ acts on the arguments from the left.
\begin{theorem}\label{theorem:birational_tWE8}
Under the birational actions \eqref{eqns:action_E8_J_para_xy} and \eqref{eqns:action_iota},
the generators of $\widetilde{W}\big(E_8^{(1)}\big)=\lrangle{\iota_1,\iota_2,\iota_3,\iota_4}\rtimes\lrangle{s_0,\dots,s_8}$ satisfy the following fundamental relations:
\begin{subequations}\label{eqns:fundamental_tWe8}
\begin{align}
&(s_is_j)^{l_{ij}}=(\iota_i\iota_j)^{m_{ij}}=1,\\
&\iota_is_j=s_j\iota_i,\quad i=1,2,3,4,~ j\neq 1,2,\quad
\iota_{\{1,2,3,4\}}s_1=s_1\iota_{\{2,1,4,3\}},\\
&\iota_1s_2=s_2\iota_1\iota_2,\quad
\iota_2s_2=s_2\iota_2,\quad
\iota_3s_2=s_2\iota_3\iota_4,\quad
\iota_4s_2=s_2\iota_4,
\end{align}
\end{subequations}
where
\begin{align}
&l_{ij}=
\begin{cases}
1,& i=j\\
3, &i=j-1\quad (j=2,\dots,7),\quad \text{or\hspace{0.5em}if}\quad (i,j)=(3,8),(7,0)\\
2, &\text{otherwise},
\end{cases}\\
&m_{ij}=
\begin{cases}
1,& i=j\\
2, &\text{otherwise}.
\end{cases}
\end{align}
\end{theorem}
\begin{proof}
Let us define the function $g(z)$ by
\begin{equation}
g(z)
=\bfrac{1-\frac{\cd{\eta-\cc2}}{\cd{\eta-z}}}{1-\frac{\cd{\eta-\cc1}}{\cd{\eta-z}}}
\bfrac{1-\frac{\cd{\eta+\cc1}}{\cd{\eta+z}}}{1-\frac{\cd{\eta+\cc2}}{\cd{\eta+z}}}
\bfrac{1-\frac{\cd{2\eta-\frac{\cc1-\cc2}{2}}}{\cd{z+\frac{\cc1+\cc2}{2}}}}{1-\frac{\cd{2\eta+\frac{\cc1-\cc2}{2}}}{\cd{z+\frac{\cc1+\cc2}{2}}}}.
\end{equation}
The action \eqref{eqn:action_E8_J_y} can be expressed as
\begin{equation}
\bfrac{s_2(y)-\cd{2\eta-\frac{\cc1-\cc2}{2}}}{s_2(y)-\cd{2\eta+\frac{\cc1-\cc2}{2}}}
\bfrac{x-\cd{\eta+\cc1}}{x-\cd{\eta+\cc2}}
\bfrac{y-\cd{\eta-\cc2}}{y-\cd{\eta-\cc1}}=g(0).
\end{equation}
By replacing the Jacobian elliptic function to the Jacobi theta-functions using \eqref{eqn:Jelliptic_JTheta}
and then applying the addition formula \eqref{eqn:add_JTH}, the function $g(z)$ can be rewritten as
\begin{equation}
g(z)=\frac{\JTK{\eta-\cc1}\JTK{\eta+\cc2}\JTK{2\eta+\frac{\cc1-\cc2}{2}}}{\JTK{\eta-\cc2}\JTK{\eta+\cc1}\JTK{2\eta-\frac{\cc1-\cc2}{2}}},
\end{equation}
which gives
\begin{equation}
g(z)=g(0).
\end{equation}
Therefore, the action \eqref{eqn:action_E8_J_y} can be also expressed as
\begin{equation}
\bfrac{s_2(y)-\cd{2\eta-\frac{\cc1-\cc2}{2}}}{s_2(y)-\cd{2\eta+\frac{\cc1-\cc2}{2}}}
\bfrac{x-\cd{\eta+\cc1}}{x-\cd{\eta+\cc2}}
\bfrac{y-\cd{\eta-\cc2}}{y-\cd{\eta-\cc1}}=g(c_3),
\end{equation}
which leads the proof of the relation $(s_2s_3)^3=1$.
The other relations can be directly verified by using the actions \eqref{eqns:action_E8_J_para_xy} and \eqref{eqns:action_iota}.
Therefore, we have completed the proof.
\end{proof}
Now we are in a position to derive Equation \eqref{eqns:RCGeqn} from the Cremona transformations
associated with $A_0^{(1)J}$-surface.
Note that for convenience we use the following notations
for the composition of the reflections $s_i$ and for the summation of the parameters $\cc{i}$:
\begin{align}
&s_{i_1\cdots i_m}=s_{i_1}\dots s_{i_m},\quad i_1\cdots i_m\in\{0,\dots,8\},\\
&\cc{j_1\cdots j_n}=\cc{j_1}+\cdots+\cc{j_n},\quad j_1\cdots j_n\in\{1,\dots,8\},
\end{align}
respectively.
Let
\begin{equation}
\RJ1=s_{56453483706756452348321 56453483706756452348321706734830468}\iota_4\iota_3\iota_2\iota_1.
\end{equation}
The actions of $\RJ1$ on the root lattice $Q(A_0^{(1)\bot})$ and parameter space are not translational as the following shows:
\begin{align}
&\RJ1:
\begin{pmatrix}
\alpha_0\\\alpha_1\\\alpha_2\\\alpha_3\\\alpha_4\\\alpha_5\\\alpha_6\\\alpha_7\\\alpha_8
\end{pmatrix}
\mapsto
\left(\begin{array}{ccccccccc}
-1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
-1 & -1 & -4 & -6 & -5 & -4 & -3 & -2 & -3 \\
0 & 0 & 1 & 2 & 1 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & -1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & -1 & 0 & 0 & 0 & 0 \\
1 & 1 & 2 & 4 & 4 & 3 & 3 & 2 & 2 \\
0 & 0 & 0 & 0 & 0 & 0 & -1 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & -1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -1 \\
\end{array}\right)
\begin{pmatrix}
\alpha_0\\\alpha_1\\\alpha_2\\\alpha_3\\\alpha_4\\\alpha_5\\\alpha_6\\\alpha_7\\\alpha_8
\end{pmatrix},\\
&\RJ1(\cc{i})=-\cc{i}+\frac{\cc{1234}-\cc{5678}}{4}-\kappa,\quad i=1,\dots,4,\\
&\RJ1(\cc{j})=-\cc{j}+\frac{\cc{1234}+3\cc{5678}}{4}-\kappa,\quad j=5,\dots,8,\quad
\RJ1(\eta)=\eta+\frac{\lambda}{2},
\end{align}
where $\kappa$ is defined by \eqref{eqn:kappa}.
However, when the parameters take special values \eqref{eqns:condition_RCG},
the action of $\RJ1$ becomes the translational motion in the parameter subspace:
\begin{equation}\label{eqn:action_RJ1_special_para}
\RJ1:(\gamma_{\rm e},\gamma_{\rm o},z_0)\mapsto(\gamma_{\rm e},\gamma_{\rm o},z_0+2(\gamma_{\rm e}+\gamma_{\rm o})-2\kappa),
\end{equation}
and then the action on the coordinates:
\begin{equation}
\RJ1:(x,y)\mapsto(\tilde{x},\tilde{y}),
\end{equation}
gives Equation \eqref{eqns:RCGeqn}.
Here, we also consider the translation
\begin{equation}\label{eqn:def_TJ1}
\TJ1={\RJ1}^2
\end{equation}
whose actions on the root lattice $Q(A_0^{(1)\bot})$ and parameter space are given by
\begin{align}
&\TJ1(\alpha_1)=\alpha_1-2\delta,\quad
\TJ1(\alpha_5)=\alpha_5+\delta,\\
&\TJ1:(\cc{i},\cc{i+4},\eta)\mapsto (\cc{i}-\lambda,\cc{i+4}+\lambda+4\kappa,\eta+\lambda-2\kappa),\quad i=1,\dots,4.
\label{eqn:action_TJ1_para}
\end{align}
The action on the coordinates:
\begin{equation}
\TJ1:(x,y)\mapsto(\overline{x},\overline{y}),
\end{equation}
gives the following elliptic Painlev\'e equation:
\begin{subequations}\label{eqns:TJ1}
\begin{align}
&\bfrac{k\,\cd{\eta-\cc8+\kappa}\overline{y}+1}{k\,\cd{\eta-\cc7+\kappa}\overline{y}+1}
\bfrac{\tilde{x}-\cd{\eta-\cc7+\frac{\cc{5678}}{2}+\lambda+\kappa}}{\tilde{x}-\cd{\eta-\cc8+\frac{\cc{5678}}{2}+\lambda+\kappa}}\notag\\
&\qquad =\gG{\frac{\cc{5678}-2\cc5+\lambda}{2},\frac{\cc{5678}-2\cc6+\lambda}{2},\frac{\cc{5678}-2\cc7+\lambda}{2},\frac{\cc{5678}-2\cc8+\lambda}{2},\eta+\frac{\lambda}{2}+\kappa}\notag\\
&\qquad\qquad \times\frac{\pP{\frac{\cc{5678}-2\cc5+\lambda}{2},\frac{\cc{5678}-2\cc6+\lambda}{2},\frac{\cc{5678}-2\cc7+\lambda}{2},\eta+\frac{\lambda}{2}+\kappa}{\tilde{x},\tilde{y}}}
{\pP{\frac{\cc{5678}-2\cc5+\lambda}{2},\frac{\cc{5678}-2\cc6+\lambda}{2},\frac{\cc{5678}-2\cc8+\lambda}{2},\eta+\frac{\lambda}{2}+\kappa}{\tilde{x},\tilde{y}}},\\
&\bfrac{k\,\cd{\eta+\cc4+\kappa}\overline{x}+1}{k\,\cd{\eta+\cc3+\kappa}\overline{x}+1}
\bfrac{k\,\cd{\eta-\cc3+2\lambda+\kappa}\overline{y}+1}{k\,\cd{\eta-\cc4+2\lambda+\kappa}\overline{y}+1}\notag\\
&\qquad=\gG{\eta-\cc1+\frac{\cc{1234}}{4}+\lambda,\eta-\cc2+\frac{\cc{1234}}{4}+\lambda,\eta-\cc3+\frac{\cc{1234}}{4}+\lambda,\eta-\cc4+\frac{\cc{1234}}{4}+\lambda,\frac{\cc{5678}+2\lambda}{4}+\kappa}\notag\\
&\qquad\qquad \times\frac{\pP{\eta-\cc1+\frac{\cc{1234}}{4}+\lambda,\eta-\cc2+\frac{\cc{1234}}{4}+\lambda,\eta-\cc3+\frac{\cc{1234}}{4}+\lambda,\frac{\cc{5678}+2\lambda}{4}+\kappa}{\frac{-1}{k\overline{y}},\tilde{x}}}
{\pP{\eta-\cc1+\frac{\cc{1234}}{4}+\lambda,\eta-\cc2+\frac{\cc{1234}}{4}+\lambda,\eta-\cc4+\frac{\cc{1234}}{4}+\lambda,\frac{\cc{5678}+2\lambda}{4}+\kappa}{\frac{-1}{k\overline{y}},\tilde{x}}},
\end{align}
where $\tilde{x}=\RJ1(x)$ and $\tilde{y}=\RJ1(y)$ are given by
\begin{align}
&\bfrac{k\,\cd{\eta+\cc8-\frac{\cc{5678}}{2}}\tilde{y}+1}{k\,\cd{\eta+\cc7-\frac{\cc{5678}}{2}}\tilde{y}+1}
\bfrac{x-\cd{\eta+\cc7}}{x-\cd{\eta+\cc8}}
=\gG{\cc5,\cc6,\cc7,\cc8,\eta}
\frac{\pP{\cc5,\cc6,\cc7,\eta}{x,y}}{\pP{\cc5,\cc6,\cc8,\eta}{x,y}},\\
&\bfrac{k\,\cd{\eta-\cc4+\frac{\cc{1234}}{2}}\tilde{x}+1}{k\,\cd{\eta-\cc3+\frac{\cc{1234}}{2}}\tilde{x}+1}
\bfrac{k\,\cd{\eta+\cc3+\frac{\cc{5678}}{2}}\tilde{y}+1}{k\,\cd{\eta+\cc4+\frac{\cc{5678}}{2}}\tilde{y}+1}\notag\\
&\qquad=\gG{\eta+\cc1+\frac{\cc{5678}}{4},\eta+\cc2+\frac{\cc{5678}}{4},\eta+\cc3+\frac{\cc{5678}}{4},\eta+\cc4+\frac{\cc{5678}}{4},\frac{\cc{5678}}{4}}\notag\\
&\qquad\qquad\times\frac{\pP{\eta+\cc1+\frac{\cc{5678}}{4},\eta+\cc2+\frac{\cc{5678}}{4},\eta+\cc3+\frac{\cc{5678}}{4},\frac{\cc{5678}}{4}}{\frac{-1}{k\tilde{y}},x}}
{\pP{\eta+\cc1+\frac{\cc{5678}}{4},\eta+\cc2+\frac{\cc{5678}}{4},\eta+\cc4+\frac{\cc{5678}}{4},\frac{\cc{5678}}{4}}{\frac{-1}{k\tilde{y}},x}}.
\end{align}
\end{subequations}
Here, for conciseness, we provide the terms $\gG{\aaa1,\aaa2,\aaa3,\aaa4,b}$,
$\pQ{\aaa1,\aaa2,\aaa3,\aaa4,\aaa5,b}{X}$ and
$\pP{\aaa1,\aaa2,\aaa3,b}{X,Y}$ in Appendix \ref{section:Appendix_A}.
\begin{remark}
The periodicity of Jacobian elliptic functions in Equations \eqref{eqns:RCGeqn} and \eqref{eqns:TJ1} allows us to take the actions \eqref{eqn:action_RJ1_special_para} and \eqref{eqn:action_TJ1_para} to be
\begin{align}
&\RJ1:(\gamma_{\rm e},\gamma_{\rm o},z_0)\mapsto(\gamma_{\rm e},\gamma_{\rm o},z_0+2(\gamma_{\rm e}+\gamma_{\rm o})),\\
&\TJ1:(\cc{i},\cc{i+4},\eta)\mapsto (\cc{i}-\lambda,\cc{i+4}+\lambda,\eta+\lambda),\quad i=1,\dots,4,
\end{align}
respectively, without loss of generality.
\end{remark}
\begin{remark}\label{remark:projective_reduction}
The relation \eqref{eqn:def_TJ1} is the key to the reduction of Equation \eqref{eqns:TJ1} to Equation \eqref{eqns:RCGeqn}.
Under the conditions \eqref{eqns:condition_RCG}, it gives the projective reduction of Equation \eqref{eqns:TJ1} to Equation \eqref{eqns:RCGeqn}.
(See \cite{KNT2011:MR2773334,KN2015:MR3340349} for more detail.)
We therefore refer to Equation \eqref{eqns:RCGeqn} as the projectively-reduced equation of Equation \eqref{eqns:TJ1}.
Conversely, we refer to Equation \eqref{eqns:TJ1} as the generic version of Equation \eqref{eqns:RCGeqn}.
Note that the word ``generic'' here has a specific meaning.
An equation is called generic only when it has the same number of parameters as the corresponding surface-type discrete Painlev\'e equation in Sakai's list \cite{SakaiH2001:MR1882403}.
\end{remark}
\begin{remark}
Often a discrete Painlev\'e equation can be extended up to a generic version by singularity confinement \cite{GRP1991:MR1125950,RGH1991:MR1125951}.
However, Equation \eqref{eqns:RCGeqn} has a reduced number of parameters after applying the singularity confinement criterion (see \cite{RCG2009:MR2525848}).
The reason is some of the coefficients in Equation \eqref{eqns:TJ1} become zero under the condition \eqref{eqns:condition_RCG}.
In such a case, singularity confinement cannot be used to extend the equation up to the generic version.
We note that a similar observation has also been reported for another example in \cite{HHNS2015:MR3317164}.
\end{remark}
Next, we provide a new expression for the MSY elliptic Painlev\'e equation investigated in \cite{SakaiH2001:MR1882403,MurataM2004:MR2108677,MSY2003:MR1958273,ORG2002:MR1946932,ORG2001:MR1877472}.
Moreover, we also construct its projectively-reduced equation.
Define a translation of $\widetilde{W}\big(E_8^{(1)}\big)$ by
\begin{equation}
\TJ2={\RJ2}^2,
\end{equation}
where
\begin{equation}
\RJ2=s_{23483256457067348356452348321},
\end{equation}
whose action on the root lattice $Q(A_0^{(1)\bot})$ is given by
\begin{equation}
\TJ2(\alpha_1)=\alpha_1-2\delta,\quad
\TJ2(\alpha_2)=\alpha_2+\delta.
\end{equation}
The translation $\TJ2$ actually corresponds to the translation ${\rm dP}(A_0^{(1)})$ in \cite{MSY2003:MR1958273}.
However, it was expressed in terms of Weierstrass elliptic functions.
We now convert this translation to Jacobian elliptic functions by using the birational actions \eqref{eqns:action_E8_J_para_xy}.
Consider the following action of $\TJ2$:
{\allowdisplaybreaks
\begin{subequations}\label{eqns:MSY_dP}
\begin{equation}
\TJ2:
\left(\begin{matrix}
\cc1,\cc2,\cc3,\cc4\\
\cc5,\cc6,\cc7,\cc8
\end{matrix},\eta,x,y\right)
\mapsto
\left(\begin{matrix}
\cc1,\cc2,\cc3,\cc4\\
\cc5,\cc6,\cc7,\cc8
\end{matrix},\eta+\lambda,\overline{x},\overline{y}\right).
\end{equation}
Here, $\overline{x}$ and $\overline{y}$ are given by
\begin{align}
&\bfrac{\overline{y}-\cd{\eta-\cc8+\lambda}}{\overline{y}-\cd{\eta-\cc7+\lambda}}
\bfrac{x-\cd{\eta+\cc7}}{x-\cd{\eta+\cc8}}\notag\\
&\quad=\bfrac{1-\frac{\cd{\eta-\cc8+\lambda}}{\cd{3\eta+\cc6+\lambda}}}{1-\frac{\cd{\eta-\cc7+\lambda}}{\cd{3\eta+\cc6+\lambda}}}
\bfrac{1-\frac{\cd{2\eta+\frac{\cc7-\cc8}{2}+\lambda}}{\cd{\cc6+\frac{\cc7+\cc8}{2}-\lambda}}}
{1-\frac{\cd{2\eta-\frac{\cc7-\cc8}{2}+\lambda}}{\cd{\cc6+\frac{\cc7+\cc8}{2}-\lambda}}}
\bfrac{1-\frac{\cd{2\eta-\frac{\cc7-\cc8}{2}+\lambda}}{\cd{2\eta+\cc6-\frac{\cc7+\cc8}{2}+\lambda}}}
{1-\frac{\cd{2\eta+\frac{\cc7-\cc8}{2}+\lambda}}{\cd{2\eta+\cc6-\frac{\cc7+\cc8}{2}+\lambda}}}
\bfrac{1-\frac{\cd{\eta+\cc8+\frac{\cc{1234}}{2}}}{\cd{\eta+\cc5+\frac{\cc{1234}}{2}}}}
{1-\frac{\cd{\eta+\cc7+\frac{\cc{1234}}{2}}}{\cd{\eta+\cc5+\frac{\cc{1234}}{2}}}}\notag\\
&\qquad\times\bfrac{\frac{\pQ{\cc5,\cc6,\cc3,\cc7,\cc{1234},\eta}{x}}{\pP{\cc1,\cc2,\cc4,\eta}{x,y}}-\frac{x-\cd{\eta+\cc3}}{x-\cd{\eta+\cc4}}\,\frac{\pQ{\cc5,\cc6,\cc4,\cc7,\cc{1234},\eta}{x}}{\gG{\cc1,\cc2,\cc3,\cc4,\eta}\pP{\cc1,\cc2,\cc3,\eta}{x,y}}}
{\frac{\pQ{\cc5,\cc6,\cc3,\cc8,\cc{1234},\eta}{x}}{\pP{\cc1,\cc2,\cc4,\eta}{x,y}}-\frac{x-\cd{\eta+\cc3}}{x-\cd{\eta+\cc4}}\,\frac{\pQ{\cc5,\cc6,\cc4,\cc8,\cc{1234},\eta}{x}}{\gG{\cc1,\cc2,\cc3,\cc4,\eta}\pP{\cc1,\cc2,\cc3,\eta}{x,y}}},\\
&\bfrac{\overline{x}-\cd{\eta+\cc1+\lambda}}{\overline{x}-\cd{\eta+\cc2+\lambda}}
\bfrac{\overline{y}-\cd{\eta-\cc2+\lambda}}{\overline{y}-\cd{\eta-\cc1+\lambda}}\notag\\
&\quad=\bfrac{1-\frac{\cd{\eta+\cc1+\lambda}}{\cd{3\eta-\cc3+3\lambda}}}{1-\frac{\cd{\eta+\cc2+\lambda}}{\cd{3\eta-\cc3+3\lambda}}}
\bfrac{1-\frac{\cd{2\eta+\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{\cc3+\frac{\cc1+\cc2}{2}}}}
{1-\frac{\cd{2\eta-\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{\cc3+\frac{\cc1+\cc2}{2}}}}
\bfrac{1-\frac{\cd{2\eta-\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{2\eta-\cc3+\frac{\cc1+\cc2}{2}+2\lambda}}}
{1-\frac{\cd{2\eta+\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{2\eta-\cc3+\frac{\cc1+\cc2}{2}+2\lambda}}}
\bfrac{1-\frac{\cd{\eta-\cc1+\frac{\cc{1234}}{2}+\lambda}}{\cd{\eta-\cc4+\frac{\cc{1234}}{2}+\lambda}}}
{1-\frac{\cd{\eta-\cc2+\frac{\cc{1234}}{2}+\lambda}}{\cd{\eta-\cc4+\frac{\cc{1234}}{2}+\lambda}}}\notag\\
&\qquad\times\bfrac{\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\frac{\lambda}{2}-\cc6,\frac{\lambda}{2}-\cc2,\cc{1234},\eta+\frac{\lambda}{2}}{\overline{y}}}{\pP{\frac{\lambda}{2}-\cc8,\frac{\lambda}{2}-\cc7,\frac{\lambda}{2}-\cc5,\eta+\frac{\lambda}{2}}{\overline{y},x}}-\frac{\overline{y}-\cd{\eta-\cc6+\lambda}}{\overline{y}-\cd{\eta-\cc5+\lambda}}\,\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\frac{\lambda}{2}-\cc5,\frac{\lambda}{2}-\cc2,\cc{1234},\eta+\frac{\lambda}{2}}{\overline{y}}}{\gG{\frac{\lambda}{2}-\cc8,\frac{\lambda}{2}-\cc7,\frac{\lambda}{2}-\cc6,\frac{\lambda}{2}-\cc5,\eta+\frac{\lambda}{2}}\pP{\frac{\lambda}{2}-\cc8,\frac{\lambda}{2}-\cc7,\frac{\lambda}{2}-\cc6,\eta+\frac{\lambda}{2}}{\overline{y},x}}}
{\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\frac{\lambda}{2}-\cc6,\frac{\lambda}{2}-\cc1,\cc{1234},\eta+\frac{\lambda}{2}}{\overline{y}}}{\pP{\frac{\lambda}{2}-\cc8,\frac{\lambda}{2}-\cc7,\frac{\lambda}{2}-\cc5,\eta+\frac{\lambda}{2}}{\overline{y},x}}-\frac{\overline{y}-\cd{\eta-\cc6+\lambda}}{\overline{y}-\cd{\eta-\cc5+\lambda}}\,\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\frac{\lambda}{2}-\cc5,\frac{\lambda}{2}-\cc1,\cc{1234},\eta+\frac{\lambda}{2}}{\overline{y}}}{\gG{\frac{\lambda}{2}-\cc8,\frac{\lambda}{2}-\cc7,\frac{\lambda}{2}-\cc6,\frac{\lambda}{2}-\cc5,\eta+\frac{\lambda}{2}}\pP{\frac{\lambda}{2}-\cc8,\frac{\lambda}{2}-\cc7,\frac{\lambda}{2}-\cc6,\eta+\frac{\lambda}{2}}{\overline{y},x}}}.
\end{align}
\end{subequations}
}
Now consider the special choices of parameters:
\begin{equation}
\cc5+\cc4=\cc6+\cc3=\cc7+\cc2=\cc8+\cc1=\frac{\lambda}{2}.
\end{equation}
Then, the transformation $\RJ2$ acts on this parameter subspace and variables $x$, $y$ as
\begin{equation}
\RJ2:
\left(\cc1,\cc2,\cc3,\cc4,\eta,x,y\right)
\mapsto
\left(\cc1,\cc2,\cc3,\cc4,\eta+\frac{\lambda}{2},\tilde{x},x\right).
\end{equation}
This leads to a scalar second-order mapping for $x$ alone, given by
\begin{align}\label{eqn:MSY_dP_PR}
&\bfrac{\tilde{\tx}-\cd{\eta+\cc1+\lambda}}{\tilde{\tx}-\cd{\eta+\cc2+\lambda}}
\bfrac{\tilde{x}-\cd{\eta-\cc2+\lambda}}{\tilde{x}-\cd{\eta-\cc1+\lambda}}\notag\\
&\quad=\bfrac{1-\frac{\cd{\eta+\cc1+\lambda}}{\cd{3\eta-\cc3+3\lambda}}}{1-\frac{\cd{\eta+\cc2+\lambda}}{\cd{3\eta-\cc3+3\lambda}}}
\bfrac{1-\frac{\cd{2\eta+\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{\cc3+\frac{\cc1+\cc2}{2}}}}
{1-\frac{\cd{2\eta-\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{\cc3+\frac{\cc1+\cc2}{2}}}}
\bfrac{1-\frac{\cd{2\eta-\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{2\eta-\cc3+\frac{\cc1+\cc2}{2}+2\lambda}}}
{1-\frac{\cd{2\eta+\frac{\cc1-\cc2}{2}+2\lambda}}{\cd{2\eta-\cc3+\frac{\cc1+\cc2}{2}+2\lambda}}}
\bfrac{1-\frac{\cd{\eta-\cc1+\frac{\cc{1234}}{2}+\lambda}}{\cd{\eta-\cc4+\frac{\cc{1234}}{2}+\lambda}}}
{1-\frac{\cd{\eta-\cc2+\frac{\cc{1234}}{2}+\lambda}}{\cd{\eta-\cc4+\frac{\cc{1234}}{2}+\lambda}}}\notag\\
&\qquad\times\bfrac{\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\cc3,\frac{\lambda}{2}-\cc2,\cc{1234},\eta+\frac{\lambda}{2}}{\tilde{x}}}{\pP{\cc1,\cc2,\cc4,\eta+\frac{\lambda}{2}}{\tilde{x},x}}-\frac{\tilde{x}-\cd{\eta+\cc3+\frac{\lambda}{2}}}{\tilde{x}-\cd{\eta+\cc4+\frac{\lambda}{2}}}\,\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\cc4,\frac{\lambda}{2}-\cc2,\cc{1234},\eta+\frac{\lambda}{2}}{\tilde{x}}}{\gG{\cc1,\cc2,\cc3,\cc4,\eta+\frac{\lambda}{2}}\pP{\cc1,\cc2,\cc3,\eta+\frac{\lambda}{2}}{\tilde{x},x}}}
{\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\cc3,\frac{\lambda}{2}-\cc1,\cc{1234},\eta+\frac{\lambda}{2}}{\tilde{x}}}{\pP{\cc1,\cc2,\cc4,\eta+\frac{\lambda}{2}}{\tilde{x},x}}-\frac{\tilde{x}-\cd{\eta+\cc3+\frac{\lambda}{2}}}{\tilde{x}-\cd{\eta+\cc4+\frac{\lambda}{2}}}\,\frac{\pQ{\frac{\lambda}{2}-\cc4,\frac{\lambda}{2}-\cc3,\cc4,\frac{\lambda}{2}-\cc1,\cc{1234},\eta+\frac{\lambda}{2}}{\tilde{x}}}{\gG{\cc1,\cc2,\cc3,\cc4,\eta+\frac{\lambda}{2}}\pP{\cc1,\cc2,\cc3,\eta+\frac{\lambda}{2}}{\tilde{x},x}}},
\end{align}
which is the projectively-reduced equation of the MSY elliptic Painlev\'e equation \eqref{eqns:MSY_dP}.
\begin{remark}\label{remark:differenceTj1Tj2}
Recall that a Kac translation $T_\alpha:{\rm Pic}(X)\to {\rm Pic}(X)$ is defined by
\begin{equation}
T_\alpha(\lambda)=\lambda+(\delta|\lambda)\alpha-\left(\frac{(\alpha|\alpha)(\delta|\lambda)}{2}+(\alpha|\lambda)\right)\delta,\quad
\lambda\in {\rm Pic}(X),
\end{equation}
and, therefore, its squared length is given by $-(\alpha|\alpha)$.
In \cite{ORG2001:MR1877472}, Ohta {\it et al.} discovered that NVs are equivalent to each other under the operations of the Weyl group $W\big(E_8^{(1)}\big)=\langle s_0,\dots,s_8\rangle$
and the same holds for NNVs.
$\TJ1$ and $\TJ2$ are expressed by Kac transformations as the following:
\begin{equation}
\TJ1=T_{2H_0-\sum_{i=5}^8E_i},\quad
\TJ2=T_{H_0-H_1}.
\end{equation}
Since
$$-\left(2H_0-\sum_{i=5}^8E_i\left|2H_0-\sum_{i=5}^8E_i\right)\right.=4,\quad -(H_0-H_1|H_0-H_1)=2,$$
$\TJ1$ and $\TJ2$ are a NNV and a NV, respectively.
In this sense, the translation $\TJ1$ differs from $\TJ2$.
\end{remark}
Remark \ref{remark:differenceTj1Tj2} shows that Equation \eqref{eqns:TJ1} forms a new system of elliptic difference equations, which differs from the MSY elliptic difference equation given originally by Sakai.
Since the generic versions of equations are different, Equation \eqref{eqn:MSY_dP_PR} differs from Equation \eqref{eqns:RCGeqn}.
In this sense, Equation \eqref{eqn:MSY_dP_PR} is also a new elliptic difference equation.
\section{Concluding remarks}
\label{ConcludingRemarks}
In this paper, we have constructed the birational actions of the Cremona isometries $\widetilde{W}\big(E_8^{(1)}\big)$ for the $A_0^{(1)J}$-surface.
Equation \eqref{eqns:RCGeqn} is realized in terms of birational actions of Cremona isometries for the first time.
Using birational actions, we further derived the elliptic Painlev\'e equations \eqref{eqns:TJ1}, \eqref{eqns:MSY_dP} and \eqref{eqn:MSY_dP_PR}.
Equation \eqref{eqns:MSY_dP} is a new expression of the important elliptic Painlev\'e equation studied in \cite{SakaiH2001:MR1882403,MurataM2004:MR2108677,MSY2003:MR1958273,ORG2002:MR1946932,ORG2001:MR1877472}.
For a long time, this equation remained the unique elliptic Painlev\'e equation known with 8 parameters.
In this paper, we provided a new elliptic Painlev\'e equation which also has 8 parameters.
We showed that it must be different because it is realized in terms of translations in $\widetilde{W}\big(E_8^{(1)}\big)$, which are not conjugate to those that give rise to the previously known elliptic difference equation.
Furthermore, we provided a projectively-reduced version of the MSY elliptic Painlev\'e equation, namely
Equation \eqref{eqn:MSY_dP_PR}.
We also showed that Equation \eqref{eqns:TJ1} is a generic version of Equation \eqref{eqns:RCGeqn}.
These results lead to more questions.
The search for Lax pairs of these new equations remains open.
Moreover, special function solutions of the projectively-reduced equations \eqref{eqn:MSY_dP_PR} and \eqref{eqns:RCGeqn} remain to be found.
\subsection*{Acknowledgment}
The authors would like to express their sincere thanks to Profs M. Noumi, Y. Ohta, T. Takenawa and Y. Yamada for inspiring and fruitful discussions.
This research was supported by an Australian Laureate Fellowship \# FL120100094 and grant \# DP160101728 from the Australian Research Council and JSPS KAKENHI Grant Number JP17J00092.
Nakazono gratefully acknowledges support from Erwin Schr\"{o}dinger International Institute for Mathematics and Physics (ESI) that enabled his attendance at the conference on Elliptic Hypergeometric Functions in Combinatorics, Integrable Systems and Physics where this work was presented in March 20-24, 2017.
|
\section{Introduction}\label{intro}
The preferential attachment mechanism, in which edges and nodes are
added to the network based on probabilistic rules, provides an
appealing description for the evolution of a network. The rule for how
edges connect nodes depends on node degree; {large degree nodes}
attract more edges. The idea is applicable to both directed and
undirected graphs and is often the basis for studying social networks,
collaborator and citation networks, and recommender
networks. Elementary descriptions of the preferential attachment model
can be found in \cite{easley:kleinberg:2010} while more
mathematical treatments are available in \cite{durrett:2010b,vanderHofstad:2017,bhamidi:2007}.
Also see \cite{kolaczyk:csardi:2014} for a statistical survey of
methods for network data, \cite{rinaldoEtAl2013} for
consideration of statistics of an undirected network and
\cite{yanEtal:2016} for asymptotics of a directed exponential
random graph models. Limit theory for estimates of an undirected
preferential attachment model was considered in \cite{GAO2017}.
{For many networks, empirical evidence} supports the hypothesis
that in- and out-degree distributions follow a power law. This
property has been shown to hold in linear
preferential attachment models, which makes {preferential
attachment} an attractive choice
for network modeling \cite{durrett:2010b,vanderHofstad:2017,
krapivsky:2001,krapivsky:redner:2001,bollobas:borgs:chayes:riordan:2003}.
While the marginal degree power laws in a simple linear preferential
attachment model were established in
\cite{krapivsky:2001,krapivsky:redner:2001,bollobas:borgs:chayes:riordan:2003},
the joint regular variation (see \cite{resnickbook:2008,
resnickbook:2007})
which is akin to a {\it joint power law}, was only recently
established
\cite{resnick:samorodnitsky:towsley:davis:willis:wan:2016, resnick:samorodnitsky:2015}.
In addition, it was shown in \cite{wang:resnick:2016} that the joint probability mass function of the in- and out-degrees is multivariate regularly varying. This is a key result as the degrees of a network are integer-valued.
In this paper, we {discuss methods of} fitting a simple linear
preferential attachment model, which {is parametrized by}
$\boldsymbol{\theta} =(\alpha,\beta,\gamma,\delta_{\text{in}},\delta_{\text{out}})$. The first
three parameters, $\alpha,\beta,\gamma$, correspond to probabilities of the 3
scenarios for adding an edge and hence sum to 1, i.e.,
$\alpha+\beta+\gamma=1$. The other two, $\delta_{\text{in}}$ and $\delta_{\text{out}}$, are tuning
parameters related to growth rates. The tail indices of the marginal
power laws for the in- and out-degrees can be expressed as explicit
functions of $\boldsymbol{\theta}$ (see \eqref{c1} and \eqref{c2} below). The graph
$G(n)=(V(n),E(n))$, where $V(n)$ is the set of nodes and $E(n)$ is the
set of edges at the $n$th iteration, evolves {based on postulates} that describe how new edges and nodes are formed. This construction of the network is Markov in the sense that the probabilistic rules for obtaining $G(n+1)$ once $G(n)$ is known do not require prior knowledge of earlier stages of the construction.
The Markov structure of the model allows us to construct a likelihood
function based on {observing}
$G(n_0),G(n_0+1),\ldots,G(n_0+n)$. After deriving the likelihood
function, we show that there exists a unique maximum at
$\hat{\boldsymbol{\theta}} =
(\hat\alpha,\hat\beta,\hat\gamma,\hatdin,\hat{\delta}_{\text{out}})$ and that the
resulting maximum likelihood estimator is strongly consistent and
asymptotically normal. The normality is proved using a martingale
central limit theorem applied to the score function. The limiting
distribution also reveals that $(\hat\alpha,\hat\beta,\hat\gamma)$,
$\hatdin$, and $\hat{\delta}_{\text{out}}$ are asymptotically independent.
From these results, asymptotic properties of the MLE for the power law indices can be derived.
For some network data, only a snapshot of the nodes and edges is
available at a single point in time, that is, only $G(n)$ is available
for some $n$. In such cases, we propose an estimation procedure for
the parameters of the network using an approximation to the
likelihood and method of moments. This also produces strongly consistent
estimators. {These estimators perform} reasonably well compared to the MLE where the entire evolution of the network is known {but predictably there is some loss of efficiency.}
We illustrate the estimation procedure for both scenarios {using}
simulated data. {S}imulation plays an important role in the process
of modeling networks {since} it provides a way to assess the
performance of model fitting procedures in the idealized setting of
knowing the true model. {Also}, after fitting a model to real
data, simulation provides a check on the quality of fit. Departures
from model assumptions can often be detected via simulation of
multiple realizations from the fitted network. Hence it is important
to have efficient simulation algorithms for producing realizations of
the preferential attachment network for a given set of parameter
values.
We adopt a simulation method, learned from Joyjit Roy, that was {inspired} by \cite{atwood:2015} and is similar to that of \cite{tonelli:concas:locci:2010}.
{Our fitting methods are implemented} in a real data setting using
the Dutch Wiki talk network {\cite{kunegis:2013}.} While one should not
expect the simple
5-parameter (later extended to 7-parameter) linear preferential
attachment model to fully explain a network with millions of edges, it
does provide a reasonable fit to the tail behavior of the degree
distributions. We are also able to detect important structural
features in the network through
{fitting the model over} separate time intervals.
{Often it is difficult to believe in the existence of a true model, especially one whose parameters remain constant over time. Allowing, as we do, a preferential attachment model with only a few parameters and no possibility for node removal may seem simplistic and unrealistic for social network data. Of course, preferential attachment is only one mechanism for network formation and evidence for its use in fields outside data networks is mixed \cite{jones:handcock:2003, jones2003social} and we restrict attention to linear preferential attachment. Even imperfect models have the potential to capture salient properties in the data, such as heavy-tailedness of the in-degree and out-degree distributions, and to identify departures from model assumptions. While maximum likelihood estimation is essentially the gold standard for cases when the underlying model is a good representation of the data, it may perform poorly in case the model is far from being appropriate. In forthcoming work, we consider a semi-parametric estimation approach for network models that exhibit heavy-tailed degree distributions. This alternative estimation methodology borrows ideas from extreme value theory. }
The rest of the paper is structured as follows. In Section \ref{sec:sim}, we formulate the linear
preferential attachment network model and present an efficient simulation method for the network. Section \ref{sec:estMLE} gives parameter estimators when either the full history is known or when only a single snapshot {in time} is available. We test these estimators against simulated data in Section \ref{sec:estSim} and then explore the Wiki talk network in Section \ref{sec:estReal}.
\section{Model specification and simulation}\label{sec:sim}
In this section, we present the linear preferential attachment model in detail and provide a fast simulation algorithm for the network.
\subsection{The linear preferential attachment model}\label{subsec:linpref}
The directed edge preferential attachment model
\cite{bollobas:borgs:chayes:riordan:2003,krapivsky:redner:2001}
constructs a growing directed random graph $G(n)=(V(n),E(n))$ whose
dynamics depend on five non-negative real numbers
$\alpha,
\beta, \gamma$, $\delta_{\text in}$ and $\delta_{\text out}$, where $\alpha+\beta+\gamma=1$ and $\delta_{\text{in}},\delta_{\text{out}} >0$. To avoid degenerate situations,
assume that each of the numbers $\alpha,
\beta, \gamma$ is strictly smaller than 1.
We obtain a new graph $G({n})$
by adding one edge to the existing graph $G({n-1})$ and
index the constructed graphs by the number $n$ of edges in $E(n)$.
We start with an arbitrary initial finite directed
graph $G({n_0})$ with at least one node and $n_0$ edges.
For $n >n_0$,
$G(n)=(V(n),E(n))$ is a graph with $|E(n)|=n$ edges and a random number $|V(n)|=N(n)$ of
nodes. If $u\in V(n)$, $D_{\rm in}^{(n)}(u)$ and $D_{\rm out}^{(n)}(u)$
denote the in- and out-degree of $u$ respectively in $G(n)$.
There are three
scenarios that we call the $\alpha$, $\beta$ and
$\gamma$-schemes, which are activated by flipping a
3-sided coin whose outcomes are $1,2,3$ with probabilities
$\alpha,\beta,\gamma$. More formally, we have an iid sequence
of multinomial random
variables $\{J_n, n>n_0\}$ with cells labelled $1,2,3$ and cell
probabilities $\alpha,\beta,\gamma$.
Then the graph $G(n)$ is
obtained from $G(n-1)$ as follows.
\tikzset{
>=stealth',
punkt/.style={
rectangle,
rounded corners,
draw=black, very thick,
text width=6.5em,
minimum height=2em,
text centered},
pil/.style={
->,
thick,
shorten <=2pt,
shorten >=2pt,}
}
\newsavebox{\mytikzpic}
\begin{lrbox}{\mytikzpic}
\begin{tikzpicture}
\begin{scope}[xshift=0cm,yshift=1cm]
\node[draw,circle,fill=white] (s1) at (2,0) {$v$};
\node[draw,circle,fill=gray!30!white] (s2) at (.5,-1.5) {$w$};
\draw[->] (s1.south west)--(s2.north east){};
\draw[dashed] (0,-2.2) circle [x radius=2cm, y radius=15mm];
\end{scope}
\begin{scope}[xshift=5cm,yshift=1cm]
\node[draw,circle,fill=gray!30!white] (s1) at (.5,-1.5) {$v$};
\node[draw,circle,fill=gray!30!white] (s2) at (-.5,-2.5) {$w$};
\draw[->] (s1.south west)--(s2.north east){};
\draw[dashed] (0,-2.2) circle [x radius=2cm, y radius=15mm];
\end{scope}
\begin{scope}[xshift=10cm,yshift=1cm]
\node[draw,circle,fill=white] (s1) at (2,0) {$w$};
\node[draw,circle,fill=gray!30!white] (s2) at (.5,-1.5) {$v$};
\draw[->] (s2.north east)--(s1.south west){};
\draw[dashed] (0,-2.2) circle [x radius=2cm, y radius=15mm];
\end{scope}
\node at (0,-3.5) {$\alpha$-scheme};
\node at (5,-3.5) {$\beta$-scheme};
\node at (10,-3.5) {$\gamma$-scheme};
\end{tikzpicture}
\end{lrbox}
\begin{figure}[h]
\centering
\usebox{\mytikzpic}
\end{figure}
\begin{itemize}
\item
If $J_n=1$ (with probability
$\alpha$), append to $G(n-1)$ a new node $v\in V(n)\setminus V(n-1)$ and an edge
$(v,w)$ leading
from $v$ to an existing node $w \in V(n-1)$.
Choose the existing node $w\in V(n-1)$ with probability depending
on its in-degree in $G(n-1)$:
\begin{equation} \label{eq:probIn}
\textbf{P}[\text{choose $w\in V(n-1)$}] = \frac{D_{\rm in}^{(n-1)}(w)+\delta_{\text
in}}{n-1+\delta_{\text in}N(n-1)} \,.
\end{equation}
\item If $J_n=2$ (with probability $\beta$), add a directed edge
$(v,w) $ to $E({n-1})$ with $v\in V(n-1)=V(n) $ and $w\in V(n-1)=V(n) $ and
the existing nodes $v,w$ are chosen independently from the nodes of $G(n-1)$ with
probabilities
\begin{equation} \label{eq:probBoth}
\textbf{P}[\text{choose $(v,w)$}] = \Bigl(\frac{D_{\rm out}^{(n-1)}(v)+\delta_{\text
out}}{n-1+\delta_{\text out}N(n-1)}\Bigr)\Bigl(
\frac{D_{\rm in}^{(n-1)}(w)+\delta_{\text
in}}{n-1+\delta_{\text in}N(n-1)}\Bigr).
\end{equation}
\item If $J_n=3$ (with probability
$\gamma$), append to $G(n-1)$ a new node $w\in V(n)\setminus V(n-1)$ and an edge $(v,w)$ leading
from the existing node $v\in V(n-1)$ to the new node $w$. Choose
the existing node $v\in V(n-1)$ with probability
\begin{equation} \label{eq:probOut}
\textbf{P}[\text{choose $v \in V(n-1)$}] = \frac{D_{\rm out}^{(n-1)}(v)+\delta_{\text
out}}{n-1+\delta_{\text out}N(n-1)}\,.
\end{equation}
\end{itemize}
{
Note that this construction allows the possibility of having self loops in the case where $J_n=2$,
but the proportion of edges that are self loops goes to 0 as $n\to\infty$. Also, multiple edges are allowed between two nodes.}
{
\subsection{Power law of degree distributions}
Given an observed network with $n$ edges, let $N_{ij}(n)$ denote the number of nodes with in-degree $i$ and out-degree $j$. If the network is generated from the linear preferential attachment model described above, then from \cite{bollobas:borgs:chayes:riordan:2003},
there exists a proper probability distribution $\{f_{ij}\}$ such
that almost surely
\begin{equation}\label{pij}
\frac{N_{ij}(n) }{N(n)} \to f_{ij}=:\frac{p_{ij}}{1-\beta},\quad n\to\infty.
\end{equation}
Consider the limiting marginal in-degree distribution $p^{\text{in}}_i:=\sum_j p_{ij}$.
It is calculated from \cite[Equation (3.10)]{bollobas:borgs:chayes:riordan:2003} that
\begin{align*}
p^{\text{in}}_0 &= \frac{\alpha}{1+a_1(\delta_{\text{in}})\delta_{\text{in}}},\\
p^{\text{in}}_i &= \frac{\Gamma(i+\delta_{\text{in}})\Gamma(1+\delta_{\text{in}}+a_1(\delta_{\text{in}})^{-1})}{\Gamma(i+1+\delta_{\text{in}}+a_1(\delta_{\text{in}})^{-1})\Gamma(1+\delta_{\text{in}})}\left(\frac{\alpha\delta_{\text{in}}}{1+a_1(\delta_{\text{in}})\delta_{\text{in}}}+\frac{\gamma}{a_1(\delta_{\text{in}})}\right),\quad i\ge 1,
\end{align*}
where
$$a_1(\lambda) := \frac{\alpha + \beta}{1+\lambda(1-\beta)}, \quad \lambda>0.$$
Moreover, $p^{\text{in}}_i$ satisfies
\begin{align}
p^{\text{in}}_i &:= \sum_{j=0}^\infty p_{ij} \sim C_{\text{in}} i^{-\iota_\text{in}}\mbox{ as }i\to\infty, \quad\text{as long as }\alpha\delta_{\text{in}}+\gamma>0,\label{asyI}
\end{align}
for some finite positive constant $C_{\text{in}}$,
where the power index
\begin{equation}\label{c1}
\iota_\text{in} = 1+\frac{1+\delta_{\text{in}}(\alpha+\gamma)}{\alpha+\beta}
\end{equation}
Similarly, the limiting marginal out-degree distribution has the same property:
\begin{align*}
p^{\text{out}}_j &:= \sum_{i=0}^\infty p_{ij} \sim C_{\text{out}} i^{-\iota_\text{out}}\mbox{ as }j\to\infty, \quad\text{as long as }\gamma\delta_{\text{out}}+\alpha>0
\end{align*}
for $C_{\text{out}}$ positive and
\begin{equation}\label{c2}
\iota_\text{out} = 1+\frac{1+\delta_{\text{out}}(\alpha+\gamma)}{\beta+\gamma}.
\end{equation}
}
\subsection{Simulation algorithm}\label{subsec:sim}
\begin{algorithm}[t]
\label{algo:1}
\SetAlgoLined\DontPrintSemicolon
\SetKwProg{myalg}{Algorithm}{}{}
\myalg{}{
\vspace{.05in}
\KwIn{$\alpha,\beta,\delta_{\text in},\delta_{\text out}$, the parameter values; $G(n_0) = (V(n_0),E(n_0))$, the initialization graph; $n$, the targeted number edges}
\KwOut{$G(n) = (V(n),E(n))$, the resulted graph}
\vspace{.05in}
$t\gets n_0$\;
\While{$t < n$}{
$N(t) \gets |V(t)|$\;
Generate $U\sim Uniform(0,1)$\;
\uIf{$U<\alpha$}{
$v^{(1)} \gets N(t)+1$\;
$v^{(2)} \gets $\textsf{ Node\_Sample}$(E(t),2,\delta_{\text in})$\;
$V(t) \gets {\textsf{Append}}(V(t),N(t)+1)$\;
}
\uElseIf{$\alpha\le U<\alpha+\beta$}{
$v^{(1)} \gets $\textsf{ Node\_Sample}$(E(t),1,\delta_{\text out})$\;
$v^{(2)} \gets $\textsf{ Node\_Sample}$(E(t),2,\delta_{\text in})$\;
}
\uElseIf{$U\ge\alpha+\beta$}{
$v^{(1)} \gets $\textsf{ Node\_Sample}$(E(t),1,\delta_{\text out})$\;
$v^{(2)} \gets N(t)+1$\;
$V(t) \gets {\textsf{Append}}(V(t),N(t)+1)$\;
}
$E(t+1) \gets $ {\textsf{Append}}$(E(t), (v^{(1)},v^{(2)}))$\;
$t \gets t+1$\;
}
\Return{$G(n)=(V(n),E(n))$}\;
}
\algrule[.8pt]
\setcounter{AlgoLine}{1}
\SetKwProg{myproc}{Function}{}{}
\myproc{\textsf{ Node\_Sample}}{
\vspace{.05in}
\DontPrintSemicolon
\KwIn{$E(t)$, the edge list up to time $t$; $j=1,2$, the node to be sample, representing outgoing and incoming nodes, respectively; $\delta \in \{\delta_{\text in},\delta_{\text out}\}$, the offset parameter}
\KwOut{the sampled node, $v$}
Generate $W\sim Uniform(0,t+N(t)\delta)$\;
\uIf{$W\le t$}{
$v \gets v_{\lceil W\rceil}^{(j)}$\;
}
\uElseIf{$W > t$}{
$v \gets \left\lceil \frac{W-t}{\delta}\right\rceil$\;
}
\Return{$v$}\;
}
\caption{Simulating a directed edge preferential attachment network}
\end{algorithm}
We describe an efficient simulation procedure for the preferential attachment network given the parameter values $(\alpha,\beta,\gamma,\delta_{\text in},\delta_{\text out})$, where $\alpha+\beta+\gamma=1$. The simulation cost of the algorithm is linear in time.
{This algorithm, which was provided by Joyjit Roy during his graduate work at Cornell University, is presented below for completeness.} {Note that this simulation algorithm is specifically designed for the case where the preferential attachment probabilities \eqref{eq:probIn}--\eqref{eq:probOut} are linear in the degrees. A similar idea for the simulation of the Yule-Simon process appeared in \cite{tonelli:concas:locci:2010}. Efficient simulation methods for the case where the preferential attachment probabilities are non-linear are studied in \cite{atwood:2015}, where their algorithm trades some efficiency for the flexibility to model non-linear preferential attachment.}
Using the notation from the introduction, at time $t=0$, we initiate with an arbitrary graph $G(n_0) = (V(n_0),E(n_0))$ of $n_0$ edges, where the elements of $E(n_0)$ are represented in form of $(v_i^{(1)},v_i^{(2)}) \in V(n_0)\times V(n_0)$, $i=1,\ldots,n_0$, with $v_i^{(1)},v_i^{(2)}$ denoting the outgoing and incoming vertices of the edge, respectively. To grow the network, we update the network at each stage from $G(n-1)$ to $G(n)$ by adding a new edge $(v_{n}^{(1)},v_{n}^{(2)})$. Assume that the nodes are labeled using positive integers starting from 1 according to the time order in which they are created, and let the random number $N(n) = |V(n)|$ denote the total number of nodes in $G(n)$.
Let us consider the situation where an existing node is to be chosen from $V(n)$ as the vertex of the new edge. Naively sampling from the multinomial distribution requires $O(N(n))$ evaluations, where $N(n)$ increases linearly with $n$. Therefore the total cost to simulate a network of $n$ edges is $O(n^2)$. This is significantly burdensome when $n$ is large, which is usually the case for observed networks. Algorithm~\ref{algo:1} describes a simulation algorithm which uses the alias method \cite{kronmal:peterson:1979} for node sampling. Here sampling an existing node from $V(n)$ requires only constant execution time, regardless of $n$. Hence the cost to simulate $G(n)$ is only $O(n)$. This method allows generation of a graph with $10^7$ nodes on a personal laptop in less than 5 seconds.
To see that the algorithm indeed produces the intended network, it suffices to consider the case of sampling an existing node from $V(n-1)$ as the incoming vertex of the new edge. In the function \textsf{Node\_Sample} in Algorithm~\ref{algo:1}, we generate $W\sim \text{Uniform}(0,n-1+N(n-1)\delta_{\text in})$ and set
$$
v \gets
v_{\lceil W\rceil}^{(j)}\,\mathbf1_{\{W\le n-1\}} + \left\lceil \frac{W-(n-1)}{\delta_{\text in}}\right\rceil \,\mathbf1_{\{W>n-1\}}.
$$
Then
\begin{eqnarray*}
\textbf{P}\left( v = w \right) &=& \textbf{P}\left( v_{\lceil W\rceil}^{(j)} = w \right) \textbf{P}\left( W \le n-1 \right) + \textbf{P}\left( \left\lceil \frac{W-(n-1)}{\delta_{\text in}}\right\rceil = w \right) \textbf{P}\left( W > n-1 \right) \\
&=& \frac{D_{\text {in}}^{(n-1)}(w)}{n-1} \, \frac{n-1}{n-1 + N(n-1)\delta_{\text in}} + \frac{1}{N(n-1)} \, \frac{N(n-1) \delta_{\text in}}{n-1 + N(n-1) \delta_{\text in}} \\
&=& \frac{D_{\text {in}}^{(n-1)}(w) +\delta_{\text in}}{n-1 + N(n-1) \delta_{\text in}},
\end{eqnarray*}
which corresponds to the desired selection probability \eqref{eq:probIn}.
\section{Parameter estimation: MLE based on the full network history}\label{sec:estMLE}
In this section, we estimate the preferential attachment
parameter vector $(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}})$ under two
{assumptions about what data is available.}
In the first scenario, the full evolution of the network is
observed, from which the likelihood function can be computed.
The resulting MLE is {strongly} consistent and asymptotically {normal}. For the
second {scenario}, the data only consist of one snapshot of the network
with $n$ edges, without the knowledge of the network history that
produced these edges. {For this scenario we give}
an estimation approach {through approximating the score function and moment matching}, which produces parameter estimators that are
also {strongly} consistent but less efficient than those based on the full
evolution of the network.
In both cases, {the estimators are uniquely determined.}
\subsection{Likelihood calculation}\label{MLEdef}
Assume the network begins with the graph $G(n_0)$ (consisting of
$n_0$ edges) and then evolves {according to the description in
Section \ref{subsec:linpref}}
with parameters
$(\alpha, \beta, \delta_{\text{in}}, \delta_{\text{out}})$, where $\delta_{\text{in}},\delta_{\text{out}}>0$ and
$\alpha,\beta$ are non-negative probabilities. The $\gamma$ is
implicitly defined by $\gamma=1-\alpha-\beta$.
{To avoid trivial cases, we will also assume $\alpha,\beta,\gamma<1$ for the rest of the paper.}
For MLE estimation we restrict the parameter space for $\delta_{\text{in}},\delta_{\text{out}}$ to be $[\epsilon,K]$, for some sufficiently small $\epsilon>0$ and large $K$. In particular, the true value of $\delta_{\text{in}},\delta_{\text{out}}$ is assumed to be contained in $(\epsilon,K)$.
Let $e_t=(v^{(1)}_t, v^{(2)}_t)$ be the newly created edge when the random graph evolves from
$G(t-1)$ to $G(t)$. {W}e sometimes refer to $t$ as the
time rather than the number of edges.
{Assume} we {observe the initial graph $G(n_0)$ and} the edges $\{e_t\}_{t=n_0+1}^n$ in the order of their formation. For $t=n_0+1,\ldots,n$, the values of the following variables are known:
\begin{itemize}
\item
$N(t)$, the number of nodes in graph $G(t)$;
\item
$D_{\text{in}}^{(t-1)}(v)$, $D_{\text{out}}^{(t-1)}(v)$, the in- and out-degree of node $v$ in $G(t-1)$, for {all} $v\in V(t-1)$;
\item
$J_t$, the scenario under which $e_t$ is created.
\end{itemize}
Then the likelihood function is
\begin{align}
L & (\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}|\ G(n_0), (e_t)_{t=n_0+1}^n) \nonumber\\
=& \prod_{t=n_0+1}^n \left(\alpha\frac{D_{\text{in}}^{(t-1)}(v^{(2)}_t)+\delta_{\text{in}}}{t-1+\delta_{\text{in}} N(t-1)}\right)^{\textbf{1}_{\{J_t=1\}}} \nonumber\\
&\times\prod_{t=n_0+1}^n\left(\beta
\Bigl(\frac{D_{\text{in}}^{(t-1)}(v^{(2)}_t)+\delta_{\text{in}}}{t-1+\delta_{\text{in}} N(t-1)} \Bigr)\Bigl(\frac{D_{\text{out}}^{(t-1)}(v^{(1)}_t)+\delta_{\text{out}}}{t-1+\delta_{\text{out}} N(t-1)}\Bigr)\right)^{\textbf{1}_{\{J_t=2\}}}\nonumber\\
&\times\prod_{t=n_0+1}^n \left((1-\alpha-\beta)\frac{D_{\text{out}}^{(t-1)}(v^{(1)}_t)+\delta_{\text{out}}}{t-1+\delta_{\text{out}} N(t-1)}\right)^{\textbf{1}_{\{J_t=3\}}}\label{jointL}
\end{align}
and the log likelihood function is
\begin{align}
\log& L(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}|\ G(n_0), (e_t)_{t=n_0+1}^n) \label{logL}\\%\nonumber\\
=& \log\alpha \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=1\}}
+ \log\beta \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=2\}} + \log(1-\alpha-\beta) \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}}\nonumber\\
&+ \sum_{t=n_0+1}^n \log\left(D_{\text{in}}^{(t-1)}(v^{(2)}_t)+\delta_{\text{in}}\right)\textbf{1}_{\{J_t\in\lbrace 1,2\rbrace\}}
+ \sum_{t=n_0+1}^n \log\left(D_{\text{out}}^{(t-1)}(v^{(1)}_t)+\delta_{\text{out}}\right)\textbf{1}_{\{J_t\in\lbrace 2,3\rbrace\}} \nonumber\\
&- \sum_{t=n_0+1}^n \log(t-1+\delta_{\text{in}} N(t-1))\textbf{1}_{\{J_t\in\lbrace 1,2\rbrace\}}
- \sum_{t=n_0+1}^n \log(t-1+\delta_{\text{out}} N(t-1))\textbf{1}_{\{J_t\in\lbrace 2,3\rbrace\}}.\nonumber
\end{align}
The score functions for $\alpha, \beta, \delta_{\text{in}}, \delta_{\text{out}}$ are calculated as follows:
\begin{align}
&\frac{\partial}{\partial\alpha} \log L(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}|\ G(n_0), (e_t)_{t=n_0+1}^n)
= \frac{1}{\alpha} \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=1\}} - \frac{1}{1-\alpha-\beta} \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}},
\label{score_alpha}\\
&\frac{\partial}{\partial\beta} \log L(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}| \ G(n_0), (e_t)_{t=n_0+1}^n)
= \frac{1}{\beta} \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=2\}} - \frac{1}{1-\alpha-\beta} \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}},
\label{score_beta}\\
&\frac{\partial}{\partial\delta_{\text{in}}} \log L(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}|\
G(n_0), (e_t)_{t=n_0+1}^n) \label{score_din}\\
&\quad = \sum_{t=n_0+1}^n \frac{1}{D_{\text{in}}^{(t-1)}(v^{(2)}_t)+\delta_{\text{in}}}
\textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\}}
- \sum_{t=n_0+1}^n \frac{N(t-1)}{t-1+\delta_{\text{in}} N(t-1)}\textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\}} ,
\nonumber \\
&\frac{\partial}{\partial\delta_{\text{out}}} \log L(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}|\
G(n_0), (e_t)_{t=n_0+1}^n)
\nonumber\\
&\quad = \sum_{t=n_0+1}^n \frac{1}{D_{\text{out}}^{(t-1)}(v^{(1)}_t)+\delta_{\text{out}}}
\textbf{1}_{\{J_t\in\lbrace 2,3\rbrace\}}
- \sum_{t=n_0+1}^n \frac{N(t-1)}{t-1+\delta_{\text{out}} N(t-1)}\textbf{1}_{\{J_t\in\lbrace 2,3\rbrace\}}.
\nonumber
\end{align}
Note that the {score functions \eqref{score_alpha}, \eqref{score_beta} for} $\alpha$ and $\beta$ do not depend on $\delta_{\text{in}}$ and $\delta_{\text{out}}$.
One can show that the Hessian matrix of the log-likelihood for $(\alpha,\beta)$ is positive definite. Setting \eqref{score_alpha} and \eqref{score_beta} to zero gives the unique MLE estimates for $\alpha$ and $\beta$,
\begin{align}
\hat{\alpha}^{MLE} &= \frac{1}{n-n_0} \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=1\}}\label{alpha_MLE},\\
\hat{\beta}^{MLE} &= \frac{1}{n-n_0} \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=2\}}\label{beta_MLE}.
\end{align}
These estimates are strongly consistent by applying the strong law of large numbers for the $\{J_t\}$ sequence.
{Next, consider the first term of the score function for $\delta_{\text{in}}$ in \eqref{score_din}, and we have}
\begin{align*}
\sum_{t=n_0+1}^n \frac{1}{D_{\text{in}}^{(t-1)}(v^{(2)}_t)+\delta_{\text{in}}}\textbf{1}_{\left\{J_t\in\lbrace 1, 2\rbrace\right\}}
&= \sum_{i=0}^\infty \frac{1}{i+\delta_{\text{in}}}\sum_{t=n_0+1}^n \textbf{1}_{\left\{D_{\text{in}}^{(t-1)}(v^{(2)}_t)=i, J_t\in\lbrace 1, 2\rbrace\right\}}.
\end{align*}
Observe that $\left\{D_{\text{in}}^{(t-1)}(v^{(2)}_t)=i, J_t\in\lbrace 1,
2\rbrace\right\}$ describes the event that the in-degree of node
$v^{(2)}_t {\in V(t-1)}$ is $i$ at time $t-1$ and is augmented to $i+1$ at time
$t$. For {each $i\ge1$, such {an} event happens at some stage $t\in\{n_0+1, n_0+2,\ldots, n\}$ only for those nodes
with
in-degree $\le i$ at time $n_0$ and in-degree $> i$ at time $n$. }
Let $N_{ij}(n)$ denote the number of nodes with in-degree $i$ and
out-degree $j$ at time $n$, and $N^{\text{in}}_i(n)$ and $N^{\text{in}}_{>i}(n)$ to be
the number of nodes with in-degree equal to $i$ and greater than $i$,
respectively, i.e.,
\begin{align*}
N^{\text{in}}_i(n) = \sum_{j=0}^\infty N_{ij}(n), & \quad N^{\text{in}}_{>i}(n) = \sum_{k>i}N^{\text{in}}_{k} (n).
\end{align*}
Then
$$\sum_{t=n_0+1}^n \textbf{1}_{\left\{D_{\text{in}}^{(t-1)}(v^{(2)}_t)=i,J_t\in\lbrace 1, 2\rbrace\right\}}=N^{\text{in}}_{>i}(n) - N^{\text{in}}_{>i}(n_0),\quad i\ge 1.$$
On the other hand, when $i=0$, $\left\{D_{\text{in}}^{(t-1)}(v^{(2)}_t)=0,J_t\in\lbrace 1, 2\rbrace\right\}$ occurs for some $t$ if and only if all of the following three events happen:
\begin{itemize}
\item[(i)] $v^{(2)}_t$ has in-degree $>0$ at time $n$;
\item[(ii)] $v^{(2)}_t$ does not have in-degree $>0$ at time $n_0$;
\item[(iii)] $v^{(2)}_t$ was not created under the $\gamma$-scheme (otherwise it would have been born with in-degree 1).
\end{itemize}
This implies:
\[
\sum_{t=n_0+1}^n \textbf{1}_{\left\{D_{\text{in}}^{(t-1)}(v^{(2)}_t)=0,J_t\in\lbrace 1, 2\rbrace\right\}}=N^{\text{in}}_{>0}(n) - N^{\text{in}}_{>0}(n_0) - \sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}},
\]
since there are, in total, $\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}}$ nodes
created under the $\gamma$-scheme. {Therefore,}
\begin{align}\label{approx-p1}
\sum_{t=n_0+1}^n \frac{1}{D_{\text{in}}^{(t-1)}(v^{(2)}_t)+\delta_{\text{in}}}\textbf{1}_{\left\{J_t\in\lbrace 1, 2\rbrace\right\}}
&= \sum_{i=0}^\infty \frac{1}{i+\delta_{\text{in}}}\sum_{t=n_0+1}^n \textbf{1}_{\left\{D_{\text{in}}^{(t-1)}(v^{(2)}_t)=i, J_t\in\lbrace 1, 2\rbrace\right\}}\nonumber\\
&= \sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n) - N^{\text{in}}_{>i}(n_0)}{i+\delta_{\text{in}}}-\frac{\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}}}{\delta_{\text{in}}}.
\end{align}
Setting {the score function \eqref{score_din} for $\delta_{\text{in}}$} to 0 and dividing both sides by $n-n_0$ leads to
\begin{align} \label{MLEdin}
\frac{1}{n-n_0}\sum_{i=0}^\infty &\frac{N^{\text{in}}_{>i}(n) - N^{\text{in}}_{>i}(n_0)}{i+\delta_{\text{in}}}\nonumber\\
&-\frac{1}{\delta_{\text{in}}(n-n_0)}\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}}
-\frac{1}{n-n_0}\sum_{t=n_0+1}^n \frac{N(t-1)}{t-1+\delta_{\text{in}} N(t-1)}\textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\}}=0,
\end{align}
{where the only unknown parameter is $\delta_{\text{in}}$.}
In Section~\ref{subsec:consistency}, we show that the solution to \eqref{MLEdin} actually maximizes the likelihood function in $\delta_{\text{in}}$.
Similarly, the MLE for $\delta_{\text{out}}$ can be solved from
\begin{align*}
\frac{1}{n-n_0}\sum_{j=0}^\infty &\frac{N^{\text{out}}_{>j}(n) - N^{\text{out}}_{>j}(n_0)}{j+\delta_{\text{out}}}\nonumber\\
&-\frac{\frac{1}{n-n_0}\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=1\}}}{\delta_{\text{out}}}
-\frac{1}{n-n_0}\sum_{t=n_0+1}^n \frac{N(t-1)}{t-1+\delta_{\text{out}} N(t-1)}\textbf{1}_{\{J_t\in\lbrace 2, 3\rbrace\}}=0,
\end{align*}
where $N^{\text{out}}_{>j}(n)$ is defined in the same fashion as $N^{\text{in}}_{>i}(n)$.
\begin{Remark}\label{sufficiency}
The arguments leading to \eqref{approx-p1} allow us to rewrite the likelihood function \eqref{jointL}:
\begin{align*}
L&(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}| \ G(n_0), (e_t)_{t=n_0+1}^n) \\
=&\ \alpha^{\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=1\}}}
\ \beta ^{\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=2\}}}
\ (1-\alpha-\beta) ^{\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}}}\nonumber\\
& \times\prod_{t=n_0+1}^n (t-1+\delta_{\text{in}} N(t-1))^{-\textbf{1}_{\left\{J_t\in\lbrace 1, 2\rbrace\right\}} }
\ (t-1+\delta_{\text{out}} N(t-1)) ^{-\textbf{1}_{\left\{J_t\in\lbrace 2,3\rbrace\right\}}}\\
&{\times \prod_{t=n_0+1}^n \left[\prod_{i=0}^\infty \ (i+\delta_{\text{in}})^{\textbf{1}_{\left\{D_{\text{in}}^{(t-1)}(v^{(2)}_t)=i,J_t\in\lbrace 1, 2\rbrace\right\}}}
\prod_{j=0}^\infty (j+\delta_{\text{out}})^{\textbf{1}_{\left\{D_{\text{out}}^{(t-1)}(v^{(1)}_t)=j,J_t\in\lbrace 2, 3\rbrace\right\}}}\right]}\\
= & \alpha^{\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=1\}}}
\ \beta ^{\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=2\}}}
\ (1-\alpha-\beta) ^{\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}}}\nonumber\\
& \times\prod_{t=n_0+1}^n (t-1+\delta_{\text{in}} N(t-1))^{-\textbf{1}_{\left\{J_t\in\lbrace 1, 2\rbrace\right\}} }
\ (t-1+\delta_{\text{out}} N(t-1)) ^{-\textbf{1}_{\left\{J_t\in\lbrace 2,3\rbrace\right\}}}
\ \delta_{\text{in}}^{-\textbf{1}_{\{J_t=3\}}}\ \delta_{\text{out}}^{-\textbf{1}_{\{J_t=1\}}}\\
&\times \prod_{i=0}^\infty \ (i+\delta_{\text{in}})^{N^{\text{in}}_{>i}(n)-N^{\text{in}}_{>i}(n_0)}
\ {\prod_{j=0}^\infty} (j+\delta_{\text{out}})^{N^{\text{out}}_{>j}(n)-N^{\text{out}}_{>j}(n_0)}.
\end{align*}
Hence by the factorization theorem, {$N(n_0)$,} $(J_t)_{t=n_0+1}^n$, $(N^{\text{in}}_{>i}(n)-N^{\text{in}}_{>i}(n_0))_{i\ge 0}$, $(N^{\text{out}}_{>j}(n)-N^{\text{out}}_{>j}(n_0))_{j\ge 0}$ are sufficient statistics for $(\alpha,\beta,\delta_{\text{in}}, \delta_{\text{out}})$.
\end{Remark}
\subsection{Consistency of MLE}\label{subsec:consistency}
{We remarked after} \eqref{alpha_MLE} and \eqref{beta_MLE} that
$\hat\alpha^{MLE}$ and $\hat\beta^{MLE}$ converge almost surely to
$\alpha$ and $\beta$. We now prove that the MLE of $(\delta_{\text{in}},\delta_{\text{out}})$
is also {strongly} consistent.
Note that if we initiate the network with $G(n_0)$ (for both $n_0$ and $N(n_0)$ finite), then almost surely for all $i,j\ge0$,
\[
\frac{N^{\text{in}}_{>i}(n_0)}{n}\le\frac{N(n_0)}{n}\to 0, \quad
\frac{N^{\text{out}}_{>j}(n_0)}{n}\le\frac{N(n_0)}{n}\to 0,
\quad \text{as }n\to\infty,
\]
and $(n-n_0)/n\to 1$.
In other words, $n_0$, $N^{\text{in}}_{>i}(n_0)$, $N^{\text{out}}_{>j}(n_0)$ are all
{$o(n)$.}
{So for simplicity,}
we assume that the graph is initiated with finitely many nodes and no
edge{s}, that is, $n_0=0$ and $N(0)\ge1$.
In particular, these assumptions imply the sum of the in-degrees at time $n$ is equal to $n$.
Let $\Psi_n(\cdot),\Phi_n(\cdot)$ be the functional forms of the terms in the log-likelihood function \eqref{logL} involving $\delta_{\text{in}}$ and $\delta_{\text{out}}$ respectively, normalized by $1/n$, i.e.,
\begin{align*}
\Psi_n(\lambda) &:= \sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n)}{n}\log(i+\lambda) - \frac{\log\lambda}{n}\sum_{t=1}^n \textbf{1}_{\lbrace J_t=3\rbrace}
- \frac{1}{n}\sum_{t=1}^n \log\left(t-1+\lambda N(t-1)\right)\textbf{1}_{\lbrace J_t\in\{1, 2\}\rbrace},\\
\Phi_n(\mu) &:= \sum_{j=0}^\infty \frac{N^{\text{out}}_{>j}(n)}{n}\log(j+\mu) - \frac{\log\mu}{n}\sum_{t=1}^n \textbf{1}_{\lbrace J_t=1\rbrace}
- \frac{1}{n}\sum_{t=1}^n \log\left(t-1+\mu N(t-1)\right)\textbf{1}_{\lbrace J_t\in\{2, 3\}\rbrace}.
\end{align*}
The following theorem gives the consistency of the MLE of $\delta_{\text{in}}$ and $\delta_{\text{out}}$.
\begin{Theorem} \label{thm:consistency}
Suppose $\delta_{\text{in}}, \delta_{\text{out}} \in (\epsilon,K) \subset (0,\infty)$.
Define
\[
\hatdin^{MLE}=\hatdin^{MLE}(n) := \operatornamewithlimits{argmax}_{\epsilon\le\lambda\le K} \Psi_n(\lambda),\qquad
\hat{\delta}_{\text{out}}^{MLE}=\hat{\delta}_{\text{out}}^{MLE}(n) := \operatornamewithlimits{argmax}_{\epsilon\le\mu\le K} \Phi_n(\mu).
\]
{Then these are the MLE estimators of $\delta_{\text{in}}, \delta_{\text{out}}$ and they} are
strongly consistent; that is,
$
\hatdin^{MLE}\stackrel{\text{a.s.}}{\longrightarrow} \delta_{\text{in}},\qquad \hat{\delta}_{\text{out}}^{MLE}\stackrel{\text{a.s.}}{\longrightarrow} \delta_{\text{out}}, \qquad n \to \infty.
$$
\end{Theorem}
\begin{proof}[Proof of Theorem \ref{thm:consistency}]
{
We only verify the consistency of $\hatdin^{MLE}$ since similar
arguments apply to $\hat{\delta}_{\text{out}}^{MLE}$.
Define
\begin{equation}
\psi_n(\lambda) := \Psi'_n(\lambda) = \sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n)/n}{i+\lambda}-\frac{\frac{1}{n}\sum_{t=1}^n \textbf{1}_{\{J_t=3\}}}{\lambda}
-\frac{1}{n}\sum_{t=1}^n \frac{N(t-1)}{t-1+\lambda N(t-1)}\textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\}}.\label{psin}
\end{equation}
Let us consider a limit version of $\psi_n$:
\begin{equation}
\psi(\lambda) := \sum_{i=0}^\infty\frac{p^{\text{in}}_{>i}(\delta_{\text{in}})}{i+\lambda} -\frac{\gamma}{\lambda}
-(1-\beta)a_1(\lambda), \label{defpsi}
\end{equation}
where $p^{\text{in}}_{>i}(\delta_{\text{in}}):= \sum_{k>i} p^{\text{in}}_k(\delta_{\text{in}})$ with $p^{\text{in}}_k(\delta_{\text{in}}):=p^{\text{in}}_k$ as defined in \eqref{asyI}, and
\[
a_1(\lambda) := \frac{\alpha+\beta}{1+\lambda(1-\beta)},\qquad \lambda>0.
\]
Here we write $p^{\text{in}}_{i}(\delta_{\text{in}})$ to emphasize the dependence on $\delta_{\text{in}}$.
In Lemmas~\ref{phi} and \ref{unifconv}, provided in the appendix, it is shown that}
{$\psi(\cdot)$ has a unique zero at $\delta_{\text{in}}$,}
{where $\psi(\lambda)>0$ when $\lambda<\delta_{\text{in}}$ and $\psi(\lambda)<0$ when $\lambda>\delta_{\text{in}}$,
and
\begin{equation}\label{eq:unifconv}
\sup_{\lambda{\ge\epsilon}}|\psi_n(\lambda)-\psi(\lambda)|{\to} 0.
\end{equation}
Since $\psi$ is continuous, for any
$\kappa>0$ arbitrarily small, there exists $\varepsilon_\kappa>0$ such
that $\psi(\lambda)>\varepsilon_\kappa$ for $\lambda \in
[\epsilon,\delta_{\text{in}}-\kappa ]$ and $\psi(\lambda)<-\varepsilon_\kappa$ for $\lambda \in [\delta_{\text{in}}+\kappa,K]$. From \eqref{eq:unifconv},
\begin{equation} \label{kpbound}
\textbf{P}\left(\exists N_\kappa\ s.t. \sup_{n>N_\kappa}\sup_{\lambda\in[\epsilon,K]}|\psi_n(\lambda) - \psi(\lambda)|< \varepsilon_\kappa/2\right) = 1.
\end{equation}
Note
$\sup_{\lambda\in[\epsilon,K]}|\psi_n(\lambda) - \psi(\lambda)|< \varepsilon_\kappa/2$ implies
$$
\psi_n(\lambda) \ge \psi{(\lambda)} - \sup_{\lambda\in[\epsilon,K]}|\psi_n(\lambda) - \psi(\lambda)| \ge \varepsilon_\kappa - \varepsilon_\kappa /2 > 0,\quad \lambda \in [\epsilon,\delta_{\text{in}}-\kappa),
$$
and
$$
\psi_n(\lambda) \le \psi{(\lambda)} + \sup_{\lambda\in[\epsilon,K]}|\psi_n(\lambda) - \psi(\lambda)| \le -\varepsilon_\kappa + \varepsilon_\kappa /2 < 0,\quad \lambda \in (\delta_{\text{in}}+\kappa,K].
$$
These jointly indicate that $\delta_{\text{in}}-\kappa \le \hatdin^{MLE}\le \delta_{\text{in}}+\kappa$.
Hence \eqref{kpbound} implies
$$
\textbf{P}\left(\lim_{n\to\infty} |\hatdin^{MLE}-\delta_{\text{in}}|\le \kappa\right) = 1,
$$
for arbitrary $\kappa>0$. That is, $\hatdin^{MLE} \stackrel{\text{a.s.}}{\longrightarrow} \delta_{\text{in}}$.
}
\end{proof}
\subsection{Asymptotic normality of MLE}\label{subsec:asynorm}
In the following theorem, we establish the asymptotic normality for the MLE estimator
$$ \label{mleest}
\hat{\boldsymbol{\theta}}_n^{MLE} = (\hat{\alpha}^{MLE},\, \hat{\beta}^{MLE},\, \hatdin^{MLE},\, \hat{\delta}_{\text{out}}^{MLE}).
$$
\begin{Theorem}\label{asymp_normality}
Let $\hat{\boldsymbol{\theta}}_n^{MLE} $ be the MLE estimator for $\boldsymbol{\theta}$, the parameter vector of the preferential attachment model. Then
\begin{equation} \label{MLElimit}
\sqrt{n}(\hat{\boldsymbol{\theta}}^\text{MLE}_n - {\boldsymbol{\theta}})\stackrel{d}{\rightarrow} N\left(\mathbf0,\Sigma(\boldsymbol{\theta})\right),
\end{equation}
where
\begin{equation} \label{fisher}
\Sigma^{-1}(\boldsymbol{\theta}) = I(\boldsymbol{\theta})
:= \begin{bmatrix}
\frac{1-\beta}{\alpha(1-\alpha-\beta)} & \frac{1}{1-\alpha-\beta} & 0 & 0 \\
\frac{1}{1-\alpha-\beta} & \frac{1-\alpha}{\beta(1-\alpha-\beta)} & 0 & 0 \\
0 & 0 & I_\text{in} & 0\\
0 & 0 & 0 & I_\text{out}
\end{bmatrix},
\end{equation}
with
\begin{align}
I_\text{in} &:= \sum_{i=0}^\infty \frac{p^{\text{in}}_{>i}}{(i+\delta_{\text{in}})^2} - \frac{\gamma}{\delta_{\text{in}}^2} - \frac{(\alpha+\beta)(1-\beta)^2}{\left(1+\delta_{\text{in}}(1-\beta)\right)^2}, \label{I_in}\\
I_\text{out} &:= \sum_{j=0}^\infty \frac{p^{\text{out}}_{>j}}{(j+\delta_{\text{out}})^2} - \frac{\alpha}{\delta_{\text{out}}^2} - \frac{(\gamma+\beta)(1-\beta)^2}{\left(1+\delta_{\text{out}}(1-\beta)\right)^2}. \nonumber
\end{align}
In particular, $I(\boldsymbol{\theta})$ is the asymptotic Fisher information matrix for the parameters, and hence the MLE estimator is efficient.
\end{Theorem}
\begin{Remark}
From Theorem~\ref{asymp_normality}, the estimators
$(\hat{\alpha}^{MLE},\, \hat{\beta}^{MLE})$, $\hatdin^{MLE}$, and
$\hat{\delta}_{\text{out}}^{MLE}$ are asymptotically independent.
\end{Remark}
{
\begin{proof} [Proof of Theorem~\ref{asymp_normality}]
We first show the limiting distributions for $(\hat{\alpha}^{MLE},\, \hat{\beta}^{MLE})$, $\hatdin^{MLE}$, and $\hat{\delta}_{\text{out}}^{MLE}$, respectively.
From \eqref{alpha_MLE} and \eqref{beta_MLE},
$$
(\hat{\alpha}^{MLE},\, \hat{\beta}^{MLE}) = \frac{1}{n} \sum_{t=1}^n \left(\textbf{1}_{\{J_t=1\}},\textbf{1}_{\{J_t=2\}}\right),
$$
where $\{J_t\}$ is a sequence of iid random variables. Hence the limiting distribution of the pair $\left(\hat{\alpha}^{MLE},\hat{\beta}^{MLE}\right)$ follows directly from standard central limit theorem {for sums of independent random variables}.
Next we show the asymptotic normality for $\hatdin^{MLE}$; the argument for $\hat{\delta}_{\text{out}}^{MLE}$ is similar.
Recall from \eqref{score_din} that the score function for $\delta_{\text{in}}$ can be written as
$$
\left.\frac{\partial}{\partial\delta_{\text{in}}} \log L(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}) \right|_\delta =: \sum_{t=1}^n u_t(\delta),
$$
where $u_t$ is defined by
\begin{equation} \label{ut_def}
u_t(\delta):=\frac{1}{D_{\text{in}}^{(t-1)}(v^{(2)}_t)+\delta} \textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\}}
- \frac{N(t-1)}{t-1+\delta N(t-1)}\textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\}}.
\end{equation}
The MLE estimator $\hatdin^{MLE}$ can be obtained by solving $\sum_{t=1}^n u_t(\delta)=0$. By a Taylor expansion of $\sum_{t=1}^n u_t(\delta)$,
\begin{align}
0 & = \sum_{t=1}^n u_t(\hatdin^{MLE})= \sum_{t=1}^n u_t(\delta_{\text{in}}) + (\hatdin^{MLE} - \delta_{\text{in}}) \sum_{t=1}^n \dot{u}_t(\hatdin^*), \label{score_taylor}
\end{align}
where $\dot{u}_t$ denotes the derivative of $u_t$ and $\hatdin^*=\delta_{\text{in}} + \xi(\hatdin^{MLE} - \delta_{\text{in}})$ for some $\xi \in [0,1]$.
An elementary transformation of \eqref{score_taylor} gives
$$
n^{1/2} (\hatdin^{MLE} - \delta_{\text{in}}) =
\left(- \frac{1}{n^{-1}\sum_{t=1}^n \dot{u}_t(\hatdin^*)} \right)
\left(n^{-1/2} \sum_{t=1}^n u_t(\delta_{\text{in}}) \right).
$$
To establish
$$
n^{1/2} (\hatdin^{MLE} - \delta_{\text{in}}) \stackrel{d}{\rightarrow} N(0,I_\text{in}^{-1}),
$$
where $I_\text{in}$ is as defined in \eqref{fisher},
it suffices to show the following two results:
\begin{enumerate}
\item[(i)]
$n^{-1/2} \sum_{t=1}^n u_t(\delta_{\text{in}}) \stackrel{d}{\rightarrow} N(0,I_\text{in})$,
\item[(ii)]
$n^{-1}\sum_{t=1}^n \dot{u}_t(\hatdin^*) \stackrel{p}{\rightarrow} -I_\text{in}$.
\end{enumerate}
These are proved in Lemmas~\ref{normality_lemma2} and \ref{normality_lemma3} in the appendix, respectively.
To establish the joint asymptotic normality of the MLE estimator $\hat{\boldsymbol{\theta}}_n^{MLE}$, denote the joint score function vector for $\boldsymbol{\theta}$ by
$$
\frac{\partial}{\partial\boldsymbol{\theta}} \log L(\boldsymbol{\theta}) =: \mathbf S_n(\boldsymbol{\theta}) = \left(S_n(\alpha),S_n(\beta),S_n(\delta_{\text{in}}),S_n(\delta_{\text{out}})\right)^T,
$$
where $S_n(\alpha),S_n(\beta),S_n(\delta_{\text{in}}),S_n(\delta_{\text{out}})$ are the score functions for $\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}}$, respectively. A multivariate Taylor expansion gives
\begin{equation} \label{joint_score_taylor}
\mathbf0 = \mathbf S_n\left(\hat{\boldsymbol{\theta}}_n^{MLE}\right) = \mathbf S_n(\boldsymbol{\theta}) + \dot{\mathbf S}_n\left(\hat{\boldsymbol{\theta}}_n^*\right) \left(\hat{\boldsymbol{\theta}}_n^{MLE} -\boldsymbol{\theta} \right),
\end{equation}
where $ \dot{\mathbf S}_n$ denotes the Hessian matrix of the log-likelihood function $\log L(\boldsymbol{\theta})$, and {$\hat{\boldsymbol{\theta}}_n^* = \boldsymbol\theta + \boldsymbol\xi \circ\left(\hat{\boldsymbol{\theta}}_n^{MLE} -\boldsymbol{\theta} \right)$ for some vector $\boldsymbol\xi \in [0,1]^4$, where ``$\circ$" denotes the Hadamard product.} From Remark~\ref{sufficiency}, the likelihood function $L(\boldsymbol{\theta})$ can be factored into
$$
L(\boldsymbol{\theta}) = f_1(\alpha,\beta)f_2(\delta_{\text{in}})f_3(\delta_{\text{out}}).
$$
Hence
\begin{equation} \label{score_hes_conv}
\frac{1}{n}\dot{\mathbf S}_n(\hat{\boldsymbol{\theta}}_n^*) =
\begin{bmatrix}
\frac{\partial^2\log L_n(\hat{\boldsymbol{\theta}}_n^*)}{\partial\alpha^2} & \frac{\partial^2\log L_n(\hat{\boldsymbol{\theta}}_n^*)}{\partial\alpha\partial\beta} & 0 & 0 \\
\frac{\partial^2\log L_n(\hat{\boldsymbol{\theta}}_n^*)}{\partial\beta\partial\alpha} & \frac{\partial^2\log L_n(\hat{\boldsymbol{\theta}}_n^*)}{\partial\beta^2} & 0 & 0 \\
0 & 0 &\frac{\partial^2\log L_n(\hat{\boldsymbol{\theta}}_n^*)}{\partial\delta_{\text{in}}^2} & 0\\
0 & 0 & 0 & \frac{\partial^2\log L_n(\hat{\boldsymbol{\theta}}_n^*)}{\partial\delta_{\text{out}}^2}
\end{bmatrix}
\stackrel{p}{\rightarrow} I(\boldsymbol{\theta})
\end{equation}
as implied in the previous part of the proof, where $I(\boldsymbol{\theta})$ is as defined in \eqref{fisher} and is positive semi-definite.
Note that $(S_n(\alpha),S_n(\beta)),S_n(\delta_{\text{in}}),S_n(\delta_{\text{out}})$ are pairwise uncorrelated. As an example, observe that
\begin{align*}
\textbf{E}[S_n(\alpha)S_n(\delta_{\text{in}})] =&\ \int \frac{\partial\log L(\boldsymbol{\theta})}{\partial\alpha}\frac{\partial\log L(\boldsymbol{\theta})}{\partial\delta_{\text{in}}} L(\boldsymbol{\theta})d\mathbf{x}\\
=&\ \int\frac{\partial\log f_1(\alpha,\beta)}{\partial\alpha}\frac{\partial\log f_2(\delta_{\text{in}})}{\partial\delta_{\text{in}}} f_1(\alpha,\beta)f_2(\delta_{\text{in}})f_3(\delta_{\text{out}}) d\mathbf{x} \\
=&\ \int\frac{\partial f_1(\alpha,\beta)}{\partial\alpha}\frac{\partial f_2(\delta_{\text{in}})}{\partial\delta_{\text{in}}} f_3(\delta_{\text{out}}) d\mathbf{x} \\
=&\ \frac{\partial^2}{\partial\alpha\partial\delta_{\text{in}}} \int L(\boldsymbol{\theta}) d\mathbf{x} \\
= &\ 0
= \textbf{E}[S_n(\alpha)]\textbf{E}[S_n(\delta_{\text{in}})].
\end{align*}
Using the Cram\'er-Wold device, the joint convergence of $\mathbf S_n(\boldsymbol{\theta})$ follows easily, i.e.,
$$
n^{-1/2} \mathbf S_n(\boldsymbol{\theta}) \stackrel{d}{\rightarrow} N(\mathbf0, I(\boldsymbol{\theta})).
$$
From here, the result of the theorem follows from \eqref{joint_score_taylor} and \eqref{score_hes_conv}.\end{proof}
}
\section{Parameter estimation based on one snapshot}\label{OneSnapshot}
Based only {on} the single snapshot $G(n)$, we propose
a parameter estimation procedure.
We assume that the choice of the snapshot does not depend on any
endogenous information related to the network. The snapshot merely
represents a point in time where the data is available.
Since no information on
the initial graph $G(n_0)$ is {available, we merely assume $n_0$ and $N(n_0)$ are fixed and $n\to\infty$}.
Among the sufficient statistics for $(\alpha,\beta,\delta_{\text{in}},\delta_{\text{out}})$ derived in Remark \ref{sufficiency}, $\left(N^{\text{in}}_{>i}(n)\right)_{i\ge 0}$, $\left(N^{\text{out}}_{>j}(n)\right)_{j\ge 0}$ are computable from $G(n)$, but the $(J_t)_{t=1}^n$ are not. However, when $n$ is large, we can use the following approximations according to the proof of {Lemma}~\ref{unifconv}:
$$
\frac{1}{n}\sum_{t=n_0+1}^n \textbf{1}_{\{J_t=3\}} \approx 1-\alpha-\beta,
$$
and
\[
\frac{1}{n}\sum_{t=n_0+1}^n \frac{N(t)}{t+\delta_{\text{in}} N(t)}\textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\} }
\approx (\alpha + \beta) \frac{1-\beta}{1+\delta_{\text{in}} (1-\beta)}.
\]
Substituting in \eqref{MLEdin}, we {estimate $\delta_{\text{in}}$ in terms of $\alpha$ and $\beta$ by solving}
\begin{equation} \label{approx_din}
\sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n)/n}{i+\delta_{\text{in}}}-\frac{1-\alpha-\beta}{\delta_{\text{in}}}
-\frac{(\alpha+\beta)(1-\beta)}{1+(1-\beta)\delta_{\text{in}}} = 0.
\end{equation}
Note that a strongly consistent estimator of $\beta$ can be obtained directly from $G(n)$:
$$\label{tildebeta}
\tilde\beta= 1-\frac{N(n)}{n}
\stackrel{\text{a.s.}}{\longrightarrow} \beta.
$$
To obtain an estimate for $\alpha$, we make use of the recursive formula for $\{p^{\text{in}}_i\}$ in \eqref{rec_pin1}:
\begin{equation}\label{alpha-p0}
\left(1+\frac{(\alpha+\beta)\delta_{\text{in}}}{1+(1-\beta)\delta_{\text{in}}}\right)p^{\text{in}}_0 = \alpha,
\end{equation}
and replace $p^{\text{in}}_0$ by $N^{\text{in}}_0(n)/n$ for large $n$,
\begin{equation}\label{alpha1}
\left(1+\frac{(\alpha+\beta)\delta_{\text{in}}}{1+(1-\beta)\delta_{\text{in}}}\right)\frac{N^{\text{in}}_0(n)}{n} = \alpha.
\end{equation}
Plug the strongly consistent estimator $\tilde{\beta}$ into \eqref{approx_din} and \eqref{alpha1}, and
we claim that solving the system of equations:
\begin{subequations}
\label{onesnap}
\begin{align}
& \sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n)/n}{i+\delta_{\text{in}}}-\frac{1-\alpha-\tilde{\beta}}{\delta_{\text{in}}}
-\frac{(\alpha+\tilde{\beta})(1-\tilde{\beta})}{1+(1-\tilde{\beta})\delta_{\text{in}}} = 0,\label{onesnap1}\\
& \left(1+\frac{(\alpha+\tilde{\beta})\delta_{\text{in}}}{1+(1-\tilde{\beta})\delta_{\text{in}}}\right)\frac{N^{\text{in}}_0(n)}{n} = \alpha, \label{onesnap2}
\end{align}
\end{subequations}
gives the unique solution $(\tilde{\alpha},\tilde{\delta}_{\text{in}})$ which is strongly consistent for $(\alpha,\delta_{\text{in}})$.
\begin{Theorem} \label{ss_consist}
The solution $(\tilde{\alpha},\tilde{\delta}_{\text{in}})$ to {the system of equations in \eqref{onesnap}} is unique and strongly consistent
for $(\alpha,\delta_{\text{in}})$, i.e.
$$ \tilde{\alpha}\stackrel{\text{a.s.}}{\longrightarrow}\alpha,\quad \tilde{\delta}_{\text{in}}\stackrel{\text{a.s.}}{\longrightarrow}\delta_{\text{in}}.$$
\end{Theorem}
The proof of Theorem~\ref{ss_consist} is given in Section~\ref{subsec:proof3}.
\vspace{.1in}
The parameters $\tilde{\delta}_{\text{out}}$ and $\tilde\gamma$ can be estimated by a mirror argument. We summarize the estimation procedure for $(\alpha,\beta,\gamma,\delta_{\text{in}},\delta_{\text{out}})$ from the snapshot $G(n)$ as follows:
\begin{enumerate}
\item[1.] {Estimate $\beta$} by $\tilde{\beta}=1-N(n)/n$.
\item[2.] Obtain $\tilde{\delta}_{\text{in}}^0$ by solving {(i.e., matching \eqref{onesnap1} and \eqref{onesnap2})}
$$
\sum_{i=1}^\infty \frac{N^{\text{in}}_{>i}(n)}{n}\frac{i}{i+\delta_{\text{in}}}(1+\delta_{\text{in}}(1-\tilde\beta)) =\frac{\frac{N^{\text{in}}_0(n)}{n} + \tilde\beta }{1-\frac{N^{\text{in}}_0(n)}{n} \frac{\delta_{\text{in}}}{1+(1-\tilde\beta)\delta_{\text{in}}}}.
$$
\item[3.] {Estimate $\alpha$} by
$$
\tilde\alpha^0 = \frac{\frac{N^{\text{in}}_0(n)}{n} + \tilde\beta}{1-\frac{N^{\text{in}}_0(n)}{n} \frac{\tilde{\delta}_{\text{in}}^0}{1+(1-\tilde\beta)\tilde{\delta}_{\text{in}}^0}} - \tilde\beta.
$$
\item[4.] Obtain $\tilde{\delta}_{\text{out}}^0$ by solving
$$
\sum_{j=1}^\infty \frac{N^{\text{out}}_{>j}(n)}{n}\frac{j}{j+\delta_{\text{out}}}(1+\delta_{\text{out}}(1-\tilde\beta)) = \frac{\frac{N^{\text{out}}_0(n)}{n} + \tilde\beta }{1-\frac{N^{\text{out}}_0(n)}{n} \frac{\delta_{\text{out}}}{1+(1-\tilde\beta)\delta_{\text{out}}}}.
$$
\item[5.] {Estimate $\gamma$} by
$$
\tilde\gamma^0 = \frac{\frac{N^{\text{out}}_0(n)}{n} + \tilde\beta }{1-\frac{N^{\text{out}}_0(n)}{n} \frac{\tilde{\delta}_{\text{out}}^0}{1+(1-\tilde\beta)\tilde{\delta}_{\text{out}}^0}} - \tilde\beta.
$$
\end{enumerate}
\smallskip
{
Note that even though all three estimators $\tilde\alpha^0,\tilde\beta,\tilde\gamma^0$ are strongly consistent and hence $\tilde\alpha^0+\tilde\beta+\tilde\gamma^0\convas1$, Step 1--5 do not necessarily imply the strict equality
\begin{equation} \label{eq:probeq}
\tilde\alpha^0+\tilde\beta+\tilde\gamma^0=1.
\end{equation}
We recommend adding the following two steps for a re-normalization to overcome this defect.
}
\begin{enumerate}
\item[6.] Re-normalize the probabilities
$$ \label{eq:renormalize}
(\tilde\alpha,\tilde\beta,\tilde\gamma) \leftarrow \left(\frac{\tilde\alpha^0(1-\tilde\beta)}{\tilde\alpha^0+\tilde\gamma^0},\tilde\beta,\frac{\tilde\gamma^0(1-\tilde\beta)}{\tilde\alpha^0+\tilde\gamma^0}\right).
$$
\item[7.] Plug $\tilde\alpha$ into \eqref{onesnap1} to update the estimate of $\delta_{\text{in}}$, i.e., solve for $\tilde{\delta}_{\text{in}}$ from
$$
\sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n)/n}{i+\tilde{\delta}_{\text{in}}}-\frac{1-\tilde\alpha-\tilde{\beta}}{\tilde{\delta}_{\text{in}}}
-\frac{(\tilde\alpha+\tilde{\beta})(1-\tilde{\beta})}{1+(1-\tilde{\beta})\tilde{\delta}_{\text{in}}} = 0.
$$
Similarly, solve for $\tilde{\delta}_{\text{out}}$ from
$$
\sum_{j=0}^\infty \frac{N^{\text{out}}_{>j}(n)/n}{j+\tilde{\delta}_{\text{out}}}-\frac{1-\tilde\gamma-\tilde{\beta}}{\tilde{\delta}_{\text{out}}}
-\frac{(\tilde\gamma+\tilde{\beta})(1-\tilde{\beta})}{1+(1-\tilde{\beta})\tilde{\delta}_{\text{out}}} = 0.
$$
\end{enumerate}
\section{Simulation study}\label{sec:estSim}
We now apply the estimation procedures described in Sections~\ref{sec:estMLE} and \ref{OneSnapshot} to simulated data,
which allows us to compare the estimation results using the full history of the network with {that using} just one snapshot.
Algorithm~\ref{algo:1} is used to simulate realizations of the preferential attachment network.
\subsection{MLE}\label{estSim:MLE}
\begin{figure}[t]\center
\includegraphics[width=12cm, height=10cm]{QQ1.pdf
\caption{Normal QQ-plots {in black} for normalized estimates in \eqref{normalize} under $5000$ replications of a preferential attachment network with $10^5$ edges and $\boldsymbol{\theta}=(0.3,0.5,2,1)$. The fitted lines in {blue} are the traditional qq-lines {(given by R)} used to check normality of the estimates.
The red dashed line represents the $y=x$ line in all plots.}\label{qq-MLE}
\end{figure}
For the scenario of observing the full history of the network, we simulated 5000 independent replications of the preferential attachment network with $10^5$ edges under the true parameter values
\begin{equation}\label{params}
\boldsymbol{\theta} = \left(\alpha,\beta,\delta_{\text{in}}, \delta_{\text{out}}\right) = (0.3,\, 0.5,\, 2,\, 1).
\end{equation}
For each realization, the MLE estimate of the parameters was computed and standardized as
{
\begin{equation}\label{normalize}
\frac{\sqrt{n}\left((\hat{\boldsymbol{\theta}}_n^{MLE})_i-(\boldsymbol{\theta})_i\right)}{\hat{\sigma}_{ii}},
\end{equation}
where} $(\hat{\boldsymbol{\theta}}_n)_i$ and $(\boldsymbol{\theta})_i$ denote the $i$-th components of $\hat{\boldsymbol{\theta}}_n^{MLE}$ and $\boldsymbol{\theta}$ respectively, and
$\hat{\sigma}_{ii}^2$ is the $i$-th diagonal component of the matrix $\hat{\Sigma}:= \Sigma(\hat{\boldsymbol{\theta}}_n^{MLE})$.
{The explicit formula for the entries of $\hat{\Sigma}$ is}
$
\hat{\Sigma} =
\begin{bmatrix}
\hat{\alpha}^{MLE}\left(1-\hat{\alpha}^{MLE}\right) & -\hat{\alpha}^{MLE}\hat{\beta}^{MLE}&0 & 0\\
-\hat{\alpha}^{MLE}\hat{\beta}^{MLE} & \hat{\beta}^{MLE}\left(1-\hat{\beta}^{MLE}\right) & 0 & 0 \\
0 & 0 & \hat{I}^{-1}_\text{in} & 0 \\
0 & 0 & 0 & \hat{I}^{-1}_\text{out}
\end{bmatrix},
$$
where, {see \eqref{fisher} and \eqref{I_in},}
\begin{align*}
\hat{I}_\text{in} &= \sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n)/n}{\left(i+\hatdin^{MLE}\right)^2} - \frac{1-\hat{\alpha}^{MLE}-\hat{\beta}^{MLE}}{\left(\hatdin^{MLE}\right)^2} - \frac{\left(\hat{\alpha}^{MLE}+\hat{\beta}^{MLE}\right)\left(1-\hat{\beta}^{MLE}\right)^2}{\left(1+\hatdin^{MLE}\left(1-\hat{\beta}^{MLE}\right)\right)^2},\\
\hat{I}_\text{out} &= \sum_{j=0}^\infty \frac{N^{\text{out}}_{>j}(n)/n}{\left(j+\hat{\delta}_{\text{out}}^{MLE}\right)^2} - \frac{\hat{\alpha}^{MLE}}{\left(\hat{\delta}_{\text{out}}^{MLE}\right)^2} - \frac{\left(1-\hat{\alpha}^{MLE}\right)\left(1-\hat{\beta}^{MLE}\right)^2}{\left(1+\hat{\delta}_{\text{out}}^{MLE}(1-\hat{\beta}^{MLE})\right)^2}.
\end{align*}
{By the consistency of the MLEs combined with the convergence of $\{N^{\text{in}}_{>i}(n)/n\}$ and $\{N^{\text{out}}_{>j}(n)/n\}$, see \eqref{eq:Nin_over}, we have that $\hat{\Sigma}_n \overset{a.s.}\to \Sigma$.
The QQ-plots of the normalized MLEs are shown in Figure~\ref{qq-MLE}, all of which line up quite well with the $y=x$ line (the red dashed line).
This is consistent with the asymptotic theory described in Theorem~\ref{asymp_normality}.
Confidence intervals for $\boldsymbol{\theta}$ can be obtained using this theorem.}
Given a single realization, {an approximate $(1-\varepsilon)$-confidence interval} for $(\boldsymbol{\theta})_i$ is
$$
(\hat{\boldsymbol{\theta}}_n^{MLE})_i \pm z_{\varepsilon/2}\sqrt{\frac{\hat{\sigma}_{ii}^2}{n}}\quad\mbox{for } i=1,\ldots,4,
$$
where $z_{\varepsilon/2}$ is the upper $\varepsilon/2$ quantile of $N(0, 1)$
\subsection{One snapshot}\label{subsec:oness}
\begin{figure}[t]\center
\includegraphics[width=12cm, height =10cm]{QQ2.pdf
\caption{Normal QQ-plots for the normalized estimates in \eqref{normalize-oness} under $5000$ replications of a preferential attachment network with $10^5$ edges and $\boldsymbol{\theta}=(0.3,0.5,2,1)$. The fitted lines in {blue} are the traditional qq-lines used to check normality of the estimates.
The red dashed line represents the $y=x$ line in all plots.}\label{qq-onesnapshot}
\end{figure}
We used the same simulated data as in Section~\ref{estSim:MLE} to obtain parameter estimates {$\tilde{\boldsymbol{\theta}}_n := (\tilde{\alpha},\, \tilde{\beta},\, \tilde{\delta}_{\text{in}},\, \tilde{\delta}_{\text{out}})$} through only the final snapshot, {i.e.{,} the set of directed edges without timestamps,} following the procedure described at the end of Section~\ref{OneSnapshot}.
{For the purpose of comparison with MLE, Figure~\ref{qq-onesnapshot} gives the QQ-plots for the normalized estimates from the snapshots using the same standardizations for the MLEs, i.e.,
\begin{equation}\label{normalize-oness}
\frac{\sqrt{n}\left((\tilde{\boldsymbol{\theta}}_n)_i-(\boldsymbol{\theta})_i\right)}{\hat{\sigma}_{ii}},\quad i=1,\ldots,4,
\end{equation}
where $(\tilde{\boldsymbol{\theta}}_n)_i$ denotes the $i$-th components of $\tilde{\boldsymbol{\theta}}_n$.}
Again, the fitted lines in {blue} are the traditional QQ-lines and the red dashed lines are the $y=x$ line.
The QQ-plot for $\tilde{\beta}$ exhibits {the} same shape as for $\hat{\beta}^{MLE}$, {since the two estimates are identical.}
{
From Figure~\ref{qq-onesnapshot}, we see that the snapshot estimates of all four parameters are consistent and approximately normal, i.e., the QQ-plots are linear. However, the slope{s} of the QQ-lines for $\tilde{\alpha},\tilde{\delta}_{\text{in}}, \tilde{\delta}_{\text{out}}$ are much steeper than the diagonal line, indicating a loss of efficiency for $\tilde{\boldsymbol{\theta}}_n$ compared with $\hat{\boldsymbol{\theta}}_n$. Indeed the estimator variance is inflated for all parameters except for $\beta$, where $\tilde\beta$ coincides with the true MLE. This is as expected since knowing only the final snapshot provides far less information than the whole network history.
}
Recall that for a consistent estimator $T_n$ of a one-dimensional parameter $\theta$ constructed from a random sample of size $n$, the asymptotic relative efficiencies (ARE) of $T_n$ is defined by
\[
ARE(T_n) := \lim_{n\to\infty}\frac{\text{Var}(\sqrt{n}T_n^*)}{\text{Var}(\sqrt{n}T_n)},
\]
where $T_n^*$ denotes the asymptotically efficient estimator.
{We may compute the ARE's for the snapshot parameter estimates }
\begin{subequations}
\label{eff}
\begin{align}
ARE(\tilde{\alpha}) & =\lim_{n\to\infty}\frac{n\text{Var}(\hat{\alpha}^{MLE})}{n\text{Var}(\tilde{\alpha})}
\approx \frac{\widehat{\text{Var}}(\hat{\alpha}^{MLE})}{\widehat{\text{Var}}(\tilde{\alpha})}
\approx 0.398, \\
ARE(\tilde{\delta}_{\text{in}})&=\lim_{n\to\infty}\frac{n\text{Var}(\hatdin^{MLE})}{n\text{Var}(\tilde{\delta}_{\text{in}})}
\approx \frac{\widehat{\text{Var}}(\hatdin^{MLE})}{\widehat{\text{Var}}(\tilde{\delta}_{\text{in}})}
\approx 0.392,\\
ARE(\tilde{\delta}_{\text{out}}) & = \lim_{n\to\infty}\frac{n\text{Var}(\hat{\delta}_{\text{out}}^{MLE})}{n\text{Var}(\tilde{\delta}_{\text{out}})}
\approx \frac{\widehat{\text{Var}}(\hat{\delta}_{\text{out}}^{MLE})}{\widehat{\text{Var}}(\tilde{\delta}_{\text{out}})}
\approx 0.226,
\end{align}
\end{subequations}
{where $\widehat\text{Var}$ denotes the {sample} variance of the parameter estimate based on the 5000 replications. Note that $ARE(\tilde\beta)=1$ since $\tilde\beta = \hat\beta^{MLE}$.}
Given a single realization, the variances of the snapshot estimates can be estimated through resampling as follows.
{
Using the estimated parameter $\tilde{\boldsymbol{\theta}}_n$,
simulate $10^4$ independent bootstrap replicates of the network with $n=10^5$ edges.
{For each simulated network, the snapshot estimate,
$\tilde{\boldsymbol{\theta}}^*_n := \left(\tilde{\alpha}^*,\,
\tilde{\beta}^*,\, \tilde{\delta}_{\text{in}}^*,\, \tilde{\delta}_{\text{out}}^*\right)$, is
computed. The sample variance of these $10^4$ snapshot estimates
can then be used as an approximation for the variance of
$\tilde{\boldsymbol{\theta}}_n$ so that assuming asymptotic
normality, a $(1-\varepsilon)$-confidence interval for $\boldsymbol \theta$ can
be approximated by
}
$$
{(\tilde{\boldsymbol{\theta}}_n)_i\pm z_{\varepsilon/2}\sqrt{\widehat{\text{Var}}\left((\tilde{\boldsymbol{\theta}}^*_n)_i\right)}\quad\mbox{for } i=1,\ldots,4,}
$$
where $z_{\varepsilon/2}$ is the upper $\varepsilon/2$ quantile of $N(0, 1)$.}
\subsection{Sensitivity test}
Now {we {investigate} the sensitivity {of} our estimates while values of
the parameters $(n,\alpha, \beta,\delta_{\text{in}}, \delta_{\text{out}})$ are allowed to vary.}
First consider the impact of $n$, the number of edges in the network. {To do so we held the parameters {fixed with values given by} \eqref{params}: $\left(\alpha,\beta,\delta_{\text{in}}, \delta_{\text{out}}\right) = (0.3,\, 0.5,\, 2,\, 1)$ and varied the value of $n$.} The QQ-plots {(not presented)} for standardized estimates using both full MLE and one-snapshot methods {were} produced to check the asymptotic normality. {When $n = 500,1000$, diagnostics revealed departures from normality for {both} the MLE {and} the snapshot estimates. However, after increasing $n$ to $10000$, estimates obtained from both approaches appeared normally distributed as expected.}
{For each value of $n$ in Table~\ref{varyn}, 5000 replicates of the network with $n$ edges and parameters $\boldsymbol{\theta}=(0.3, 0.5, 2,1)$ were generated.
For each realization, the MLE's $\hat{\boldsymbol{\theta}}_n^{MLE}$ were computed using the full history of the network and the one-snapshot estimates $\tilde{\boldsymbol{\theta}}_n$ were obtained using the 7-step snapshot method proposed in Section~\ref{OneSnapshot}, pretending that only the last snapshot $G(n)$ was available. The mean for these two estimators were recorded in Table~\ref{varyn}.}
{There is little bias for both estimates of $\alpha$ and $\beta$, even for small values of $n$. On the other hand, there is some bias for estimated $\delta_{\text{in}}$ and $\delta_{\text{out}}$ for $n\le 5000$. The magnitude of the biases for both types of estimates {decrease} as $n$ increases.}
Also the ARE's of the snapshot estimator {stay within a narrow band} as $n$ increases.
\begin{table}[t]
\centering
\caption{\smallskip{Mean of $\hat{\boldsymbol{\theta}}_n^{MLE}$ and $\tilde{\boldsymbol{\theta}}_n$ with ARE's of $\tilde{\boldsymbol{\theta}}_n$ relative to $\hat{\boldsymbol{\theta}}_n^{MLE}$ for $\boldsymbol{\theta}=(0.3, 0.5, 2,1)$ under different choices of $n$.}}\label{varyn}
\begin{tabular}{ccccc}
\hline
$n$& $Mean(\hat{\boldsymbol{\theta}}_n^{MLE})$ &
$Mean(\tilde{\boldsymbol{\theta}}_n)$ & $ARE(\tilde{\boldsymbol{\theta}}_n)$\\
\hline
$1000$ & (0.300, 0.500, {2.076}, {1.054}) & ({0.301}, 0.500, {2.128}, {1.066}) & (0.408, 1.000, 0.397, 0.228) \\
$5000$ & (0.300, 0.500, {2.022}, {1.013}) & ({0.301}, 0.500, {2.036}, {1.010}) & (0.414, 1.000, 0.386, 0.236) \\
$10000$ & (0.300, 0.500, {2.011}, {1.006}) & (0.301, 0.500, {2.019}, {1.006}) & (0.408, 1.000, 0.388, 0.232) \\
$50000$ & (0.300, 0.500, {2.003}, {1.002}) & (0.300, 0.500, {2.005}, {1.002}) & (0.399, 1.000, 0.393, 0.230) \\
$100000$ & (0.300, 0.500, 2.001, 1.001) & (0.300, 0.500, 2.003, 1.000) & (0.392, 1.000, 0.382, 0.223)\\
\hline
\end{tabular}
\end{table}
{
Next we held $(n, \delta_{\text{in}}, \delta_{\text{out}}) = (10^5, 2, 1)$ fixed and experimented with various values of $(\alpha, \beta)$ in Table~\ref{varyab}.
For each choice of $(\alpha, \beta)$, 5000 independent realizations of the network were generated and the means of the MLE $\hat{\boldsymbol{\theta}}_n^{MLE}$ and the one-snapshot estimates $\tilde{\boldsymbol{\theta}}_n$ were recorded.}
Overall, the biases for $\hat{\boldsymbol{\theta}}_n^{MLE}$ are remarkably small for virtually all combinations of parameter values, except for those parameter choices where one of $(\alpha, \beta)$ is extremely small. The biases for the snapshot estimates $\tilde{\boldsymbol{\theta}}_n$ exhibit a similar property, but the magnitudes of the biases are consistently larger than those in the MLE case.
{
In general, the snapshot estimators are able to achieve $20\%$--$50\%$ efficiency over the range of parameters considered. The loss of efficiency might be less than one would expect given the substantial reduction in the data available to produce the snapshot estimates. It is worth noting that in the case where $(\alpha,\beta)=(0.7,0.2)$, the efficiencies of the snapshot estimators for $\alpha$ and $\delta_{\text{in}}$ are much larger (.73 and .79, respectively). A heuristic explanation for this increase is that the parameter $\gamma =1-\alpha-\beta=0.1$ is relatively small. By the implicit constraints used for the snapshot estimates, we have
$$
\tilde\alpha+\tilde\gamma=1-\tilde \beta=1-\hat\beta^{MLE}=\hat\alpha^{MLE}+\hat\gamma^{MLE},
$$
that is, the snapshot estimate of the sum $\alpha+\gamma$ is the same as the MLE for the sum. Now if $\gamma$ is small, one would expect the resulting estimates to also be small so that $\tilde\alpha$ would be nearly the same as $\hat \alpha^{MLE}$. Hence the ARE would be close to 1. On the other hand, in the case of a larger $\gamma$, see the bottom row of Table~\ref{varyab} in which $\gamma=0.6$, the ARE for $\alpha$ is not as large (.42), but the ARE for $\tilde{\delta}_{\text{out}}$ is (.63).}
\begin{table}[t]
\caption{\smallskip{Mean of $\hat{\boldsymbol{\theta}}_n^{MLE}$ and $\tilde{\boldsymbol{\theta}}_n$ with ARE's of $\tilde{\boldsymbol{\theta}}_n$ relative to $\hat{\boldsymbol{\theta}}_n^{MLE}$ for $(n,\delta_{\text{in}},\delta_{\text{out}})=(10^5,2,1)$ under different choices of $(\alpha,\beta)$.}}
\label{varyab}
\centering
\begin{tabular}{ccccc}
\hline
$(\alpha, \beta)$ & $Mean(\hat{\boldsymbol{\theta}}_n^{MLE})$ &
$Mean(\tilde{\boldsymbol{\theta}}_n)$ & $ARE(\tilde{\boldsymbol{\theta}}_n)$\\
\hline
(0.001, 0.99) & (0.001, 0.990, {2.034}, {1.016}) & (0.001, 0.990, {2.071}, {1.049}) & (0.291, 1.000, 0.147, 0.316) \\
(0.01, 0.9) & (0.010, 0.900, {2.004}, 1.001) & (0.010, 0.900, {2.008}, {1.004}) & (0.331, 1.000, 0.207, 0.381) \\
(0.1, 0.8) & (0.100, 0.800, 2.003, 1.001) & (0.100, 0.800, {2.004}, {1.002}) & (0.353, 1.000. 0.264, 0.216)\\
(0.2, 0.6) & (0.200, 0.600, 2.002, 1.001) & (0.200, 0.600, 2.003, 1.001) & (0.364, 1.000, 0.309, 0.236) \\
(0.5, 0.3) & (0.500, 0.300, 2.001, 1.001) & (0.500, 0.300, {2.002}, 1.000) & (0.472, 1.000, 0.529, 0.202) \\
(0.7, 0.2) & (0.700, 0.200, 2.002, 1.000) & ({0.700}, 0.200, {2.002}, 1.000) & (0.726, 1.000, 0.793, 0.217) \\
(0.1, 0.3) & (0.100, 0.300, 2.001, 1.001) & ({0.100}, 0.300, {2.002}, 1.000) & (0.420, 1.000, 0.313, 0.629) \\
\hline
\end{tabular}
\end{table}
\section{Real network example}\label{sec:estReal}
In this section, we explore fitting a preferential attachment model to
a social network. As illustration, we chose the Dutch Wiki talk
network dataset, available on KONECT \cite{kunegis:2013}
(\url{http://konect.uni-koblenz.de/networks/wiki_talk_nl}). The nodes
represent users of Dutch Wikipedia, and an edge from node A to node B
refers to user A writing a message on the talk page of user B at a
certain time point. The network consists of 225,749 nodes (users) and
1,554,699 edges (messages). All edges are recorded with timestamps.
In order to accommodate all the edge formulation scenarios appeared in the dataset, we extend our model by appending the following two interaction schemes ($J_n=4,5$) in addition to the existing three ($J_n=1,2,3$) described in Section \ref{subsec:linpref}.
\begin{itemize}
\item If $J_n=4$ (with probability $\xi$), append to $G(n-1)$ two new nodes $v,w\in V(n)\setminus V(n-1)$ and an edge connecting them $(v,w)$.
\item If $J_n=5$ (with probability $\rho$), append to $G(n-1)$ a new node $v\in V(n)\setminus V(n-1)$ {with} self loop $(v,v)$
\end{itemize}
These scenarios have been observed in other social network data, such as the Facebook wall post network (\url{http://konect.uni-koblenz. de/networks/facebook-wosn-wall}), etc. They occur in small proportions and can be easily accommodated by a slight modification in the model fitting procedure. The new model has parameters $(\alpha,\beta,\gamma,\xi,\delta_{\text{in}},\delta_{\text{out}})$, and $\rho$ is implicitly defined through $\rho = 1- (\alpha+\beta+\gamma+\xi)$. Similar to the derivations in Section \ref{sec:estMLE}, the MLE estimators for $\alpha,\beta,\gamma,\xi$ are
\begin{align*}
\hat{\alpha}^{MLE} = \frac{1}{n} \sum_{t=1}^n \textbf{1}_{\{J_t=1\}}, & \quad \hat{\beta}^{MLE} = \frac{1}{n} \sum_{t=1}^n \textbf{1}_{\{J_t=2\}},\\
\hat{\gamma}^{MLE} = \frac{1}{n} \sum_{t=1}^n \textbf{1}_{\{J_t=3\}}, & \quad \hat{\xi}^{MLE} = \frac{1}{n} \sum_{t=1}^n \textbf{1}_{\{J_t=4\}},
\end{align*}
and $\delta_{\text{in}},\delta_{\text{out}}$ can be obtained through solving
\begin{align*}
\sum_{i=0}^\infty \frac{N^{\text{in}}_{>i}(n)/n}{i+\delta_{\text{in}}}-\frac{\frac{1}{n}\sum_{t=1}^n \textbf{1}_{\{J_t\in\{3,4,5\}\}}}{\delta_{\text{in}}}
-\frac{1}{n}\sum_{t=1}^n \frac{N(t)}{t+\delta_{\text{in}} N(t)}\textbf{1}_{\{J_t\in\lbrace 1, 2\rbrace\}}&=0, \\
\sum_{j=0}^\infty \frac{N^{\text{out}}_{>j}(n)/n}{j+\delta_{\text{out}}}-\frac{\frac{1}{n}\sum_{t=1}^n \textbf{1}_{\{J_t\in\{1,4,5\}\}}}{\delta_{\text{out}}}
-\frac{1}{n}\sum_{t=1}^n \frac{N(t)}{t+\delta_{\text{out}} N(t)}\textbf{1}_{\{J_t\in\lbrace 2,3 \rbrace\}}&=0.
\end{align*}
We first naively fit the linear preferential attachment model to the full network using MLE. The MLE estimators are
\begin{align} \label{eq:wikiparBefore}
(\hat\alpha,\hat\beta,\hat\gamma,&\hat\xi,\hat\rho,\hatdin,\hat{\delta}_{\text{out}})= \\
& (3.08\times10^{-3}, 8.55\times10^{-1}, 1.39\times10^{-1}, 4.76\times10^{-5},3.06\times10^{-3}, 0.547, 0.134 ).\nonumber
\end{align}
To evaluate the goodness-of-fit, 20 network realizations were simulated from the fitted model. We overlaid the {empirical} in- and out-degree frequencies of the original network with that of the simulations. If the model fits the data well, the {degree frequencies} of the data should lie within the range formed by that of the simulations, {which gives an informal confidence region for the degree distributions}. From Figure \ref{fig:wikiDegConst0}, we see that while the data roughly agrees with the simulations in the out-degree frequencies, the deviation in the in-degree frequencies is noticeable.
\begin{figure}[t]
\includegraphics[width=5in]{wikiDegDist0.pdf}
\caption{{Empirical in- and out-degree frequencies of the full Wiki talk network (red) and that from 20 realizations of the linear preferential attachment network with fitted parameter values \eqref{eq:wikiparBefore} from MLE (blue). The scatter plots for the degree frequencies from the 20 simulations are overlaid together to form {an informal confidence region} for the degree distribution of the fitted model}}
\label{fig:wikiDegConst0}
\end{figure}
To better understand the discrepancy in the in-degree frequencies,
{we examined the link data and their timestamps} and discovered
bursts of messages originating from certain nodes over small
time intervals. {According to Wikipedia policy \cite{wikipedia:2016}, certain administrating accounts are allowed to send group messages to multiple users simultaneously.} These bursts presumably represent broadcast announcements generated from these accounts.
These administrative broadcasts can {also be detected} if we
apply the linear preferential attachment model to the network in local
time intervals. We {divided the total time frame down to sub-intervals of varying length
each containing the formation of $10^4$ edges. The number $10^4$ is chosen to ensure good asymptotics as shown in Table~\ref{varyn}.
This process generated 155 networks,
$$
G(n_{k-1}),\dots, G(n_k-1), \quad k=1,\dots,155.
$$
For each of the 155 datasets, we fit a
preferential attachment model using MLE.}
The resulting estimates $(\hatdin,\hat{\delta}_{\text{out}})$ are plotted against the
corresponding timeline on the upper left panel of Figure
\ref{fig:wikiParEst}. Notice that $\hatdin$ exhibits large spikes at
various times. Recall from \eqref{eq:probIn}, a large value of $\delta_{\text{in}}$
indicates that the probability of an existing node $v$ receiving a new
message becomes less dependent on its in-degree, i.e., previous
popularity. These spikes appear to be directly related to the
occurrences of group messages. This plot is truncated after the day
2016/3/16, on which a massive group message of size 48,957 was sent
and the model can no longer be fit.
{We identified 37 users who have sent, at least once, 40 or
more consecutive messages in the message history. This is evidence
that group messages were
sent by this user. We presume these nodes are administrative
accounts}; they are responsible for about $30\%$ of
the total messages sent. Since their behavior cannot be regarded as
normal social interaction, we excluded messages from these accounts
from the dataset in our analysis. We then also removed
nodes with zero in- and out-degrees.
The re-estimated parameters after the data cleaning are displayed
in the other three panels of Figure
\ref{fig:wikiParEst}. Here all parameter estimates are quite stable
through time.
\begin{figure}[t]
\includegraphics[width=5in]{wikiPar.pdf}
\caption{Local parameter estimates of the linear preferential attachment model for the full and reduced Wiki talk network. Upper left: $(\hatdin,\hat{\delta}_{\text{out}})$ for the full network. Upper right, lower left, lower right: $(\hatdin,\hat{\delta}_{\text{out}})$, $(\hat\beta,\hat\gamma)$, $(\hat\alpha,\hat\xi,\hat\rho)$ for the reduced network, respectively.}
\label{fig:wikiParEst}
\end{figure}
The reduced network now contains 112,919 nodes and 1,086,982 edges, to
which we fit the linear preferential attachment model. The fitted
parameters based on MLE for our reduced dataset are
\begin{align}\label{eq:wikiparAfter}
(\hat\alpha,\hat\beta,\hat\gamma,&\hat\xi,\hat\rho,\hatdin,\hat{\delta}_{\text{out}})=\\
&(6.95\times10^{-3}, 8.96\times10^{-1}, 9.10\times10^{-2}, 1.44\times10^{-4},5.61\times10^{-3}, 0.174, 0.257 ).\nonumber
\end{align}
Again the degree distributions of the data and 20 simulations from the
fitted model are displayed in Figure \ref{fig:wikiDegConst}. The
out-degree distribution of the data agrees reasonably well with the
simulations. For the in-degree distribution, the fit is better than that for the entire dataset (Figure \ref{fig:wikiDegConst0}). However, for smaller in-degrees, the fitted model over-estimates the in-degree frequencies. We speculate that in many social networks, the out-degree is {in line} with that predicted by the preferential attachment model. An individual node would be more likely to reach out to others if having done so many times previously. For in-degrees, the situation is complicated and may depend on a multitude of factors. For instance, the choice of recipient may
depend on the community that the sender is in, the topic being
discussed in the message, etc. As an example a group leader might send messages to his/her team on a regular basis. Such examples violate the base assumptions of the preferential attachment model and could result in the deviation between the data and the simulations.
\begin{figure}[t]
\includegraphics[width=5in]{wikiDegDist.pdf}
\caption{{Empirical in- and out-degree frequencies of the reduced Wiki talk network (red) and that from 20 realizations of the linear preferential attachment network with fitted parameter values \eqref{eq:wikiparAfter} from MLE (blue). }}
\label{fig:wikiDegConst}
\end{figure}
{
Next we consider the estimation method of Section~\ref{OneSnapshot} applied to a single snapshot of the data. In order to implement this procedure, we donned blinders and assumed that our dataset consists only of the information of the wiki data at the last timestamp. That is, information about administrative broadcasts, and other aspects of the data learned by looking at the previous history of the data are unavailable. In particular, we would have no knowledge of the existence of the two additional scenarios corresponding to $J_n=4, 5$. With this in mind, we fit the three scenario model using the methods in Section~\ref{OneSnapshot}. The fitted parameters are
\begin{equation}\label{eq:wikiparSnap}
(\tilde\alpha,\tilde\beta,\tilde\gamma,\tilde{\delta}_{\text{in}},\tilde{\delta}_{\text{out}})=
(5.80\times10^{-4}, 8.55\times10^{-1}, 1.45\times10^{-1},0.199, 0.165 ).
\end{equation}
The comparison of the degree distributions between the data and simulations from the fitted model is displayed in Figure~\ref{fig:wikiSnap} and is not too dissimilar to the plots in Figure~\ref{fig:wikiDegConst0} that are based on maximum likelihood estimation using the full network data. In particular, the out-degree distribution is matched reasonably well, but the fitted model does a poor job of {capturing} the in-degree distribution.
}
\begin{figure}[t]
\includegraphics[width=5in]{wikiSnap.pdf}
\caption{{Empirical in- and out-degree frequencies of the full Wiki talk network (red) and that from 20 realizations of the linear preferential attachment network with fitted parameter values \eqref{eq:wikiparSnap} from the snapshot estimator (blue). }}
\label{fig:wikiSnap}
\end{figure}
We see from this example that while the linear preferential attachment model {is} perhaps too simplistic for
the Wiki talk network dataset, it has the ability to
{illuminate} some gross features, such as the out-degrees, as well as to
capture important structural changes such as the group message
behavior. {Consequently, despite its limitation, this model} may be used as a building block for more flexible models. Modification to the existing model formulation and
more careful analysis of change points in parameters is
a direction for future research.
\section{Acknowledgement}
Research of the four authors was partially supported by Army MURI grant W911NF-12-1-0385. Don Towsley from University of Massachusetts introduced us to the model
and within his group, James Atwood graciously supplied us with a
simulation algorithm designed for a class of growth models broader
than the one specified in Section \ref{subsec:linpref}{; this later became \cite{atwood:2015}}. Joyjit Roy,
formerly of Cornell, created an efficient algorithm designed to
capitalize on the linear growth structure.
{Finally, we appreciate the many helpful and sensible comments
of the referees and editors.}
|
\section{Introduction}
This work offers some observations on a classic topic: the relationship between mobile-immobile solute equilibrium, kinetics, and the retarded advection-dispersion equation (ADE).
Our primary motivation for revisiting this subject lies in the fact that two contradictory approaches to modeling mobile-immobile mass transfer have co-existed in the literature for decades, both underpinned by seemingly exact mathematical arguments. The first approach---usage of a retardation factor in the ADE---has been ostensibly derived under fast mass transfer, or "local equilibrium" conditions. In parallel, a second group of authors have shown that mass transfer is always dispersive and, for first-order kinetic mass transfer, analytically quantified its effect. We concur with this second group of authors that mass-transfer is always dispersive. However, the error underpinning the ostensibly exact retarded ADE derivation does not appear to have been pinpointed in the literature. Furthermore, we document below numerous places in the literature where the retarded ADE is treated as exact, in which this error is not harmless.
Our secondary motivation is to correct a potential misconception regarding the relationship between degree of local equilibrium and degree of validity of the retarded ADE. While usage of the latter is sometimes referred to as the "local equilibrium assumption" (LEA), we present an example transport simulation that respects local equilibrium, but in which the retarded ADE is a very poor proxy for true behavior. It is actually the case that the solute remobilization rate is the control on the validity of the retarded ADE.
The remediation of contaminated groundwater sites is a topic of persistent interest in industrialized societies. Remediation is generally quite expensive---the U.S. National Research Council recently estimated that the cost to clean up existing sites in the United States at over \$100 billion over the next 30 years \cite{us2013alternatives}. This high cost necessitates the development of accurate yet tractable groundwater models. Unfortunately, the trade-off between accuracy and tractability is made difficult by the presence of pore-scale mass transfer processes that are too complex to model at their natural support scale, but which also have macroscopically observable effects. One of the most important such processes is adsorption, the reversible chemical interaction between dissolved contaminants and solid-phase components of the pore structure.
In hydrogeologic and engineering models, it is common to model mobile-immobile (e.g., sorbing) solute transport with the retarded ADE. This is the equation:
\begin{equation}
R\frac{\partial c}{\partial t}(x,t)=-v\frac{\partial c}{\partial x}(x,t)+D\frac{\partial^{2}c}{\partial x^{2}}(x,t),\label{eq:ADE retarded}
\end{equation}
where $R$ $\left[\mathrm{dimensionless}\right]$ is a constant retardation factor, $c$ $\left[\mathrm{ML^{-3}}\right]$ is aqueous solute concentration, $t$ $\left[\mathrm{T}\right]$ is time, $x$ $\left[\mathrm{L}\right]$ is the spatial coordinate, $v$ $\left[\mathrm{LT^{-1}}\right]$ is the advection velocity, and $D$ $\left[\mathrm{L^{2}T^{-1}}\right]$ is a Fickian dispersion coefficient. (In all unit expressions, $[\mathrm{M}]$ represents mass, $[\mathrm{L}]$ represents length, and $[\mathrm{T}]$ represents time.) This equation applies as well to mobile-immobile mass transfer (MIMT) processes other than sorption.
The $R$ on the LHS accumulation term of \eqref{eq:ADE retarded} can be viewed in two different ways: as a rescaling factor for time, or as a rescaling factor for solute accumulation. Based on this, there are two possible understandings of the retardation factor in homogeneous media. It may be conceived as the ratio of groundwater velocity to mean solute velocity \citep[e.g.,][]{rajaram_time_1997}, or as the ratio of total (mobile and immobile) solute concentration to mobile solute concentration at equilibrium. This second conception motivates the idea that ``local equilibrium'' mass transfer provides support for usage of the retarded ADE.
However, there can be no exact equilibrium under transient conditions (only fast kinetics). While this may seem innocuous, its impact may be significant. To understand the degree of approximation that is occurring relative to fast kinetic behavior, we will consider the explicit transport equations for advection and dispersion in the presence of first-order single-rate mass transfer. The relevant equations may be written \citep[][p. 133]{fetter_contaminant_1993}:
\begin{equation}
\begin{array}{rl} \frac{\partial c}{\partial t}(x,t)+\frac{\partial s}{\partial t}(x,t) & =-v\frac{\partial c}{\partial x}(x,t)+D\frac{\partial^2 c}{\partial x^2}(x,t)\\ \\ \frac{\partial s}{\partial t}(x,t) & =\lambda c(x,t)-\mu s(x,t) \end{array},\label{eq:Explicit treatment of sorption}
\end{equation} where $s$ $\left[\mathrm{ML^{-3}}\right]$ is the immobile concentration, $\lambda$ $\left[\mathrm{T^{-1}}\right]$ is the probability per unit time of immobilization of mobile solute, and $\mu$ $\left[\mathrm{T^{-1}}\right]$ is the probability per unit time for mobilization of immobile solute. We show in Appendix \ref{sec: eulerian derivation} how \eqref{eq:ADE retarded} is a special case of \eqref{eq:Explicit treatment of sorption}, in the $\mu \rightarrow \infty$ limit. So while the first-order kinetic model is itself an idealization, it is no more so than the retarded ADE and additionally captures the true behavior of solute being continuously mobile or immobile for finite intervals. Equations of form \eqref{eq:Explicit treatment of sorption} are widely used in the literature to capture general MIMT processes (see conceptual discussion in \citet{Fernandez-garcia2015}, \citet{valocchi_validity_1985}, and \citet{bahr_direct_1987}). They are applicable over a range of advection velocities \citep{Zhang2008a,Zhang2009} and spatial support scales \citep{Raoof2010}. Thus, the analysis of the system they describe is relevant to a large variety of hydrogeologic problems. This single-rate paradigm, while not encompassing all forms of MIMT---for example, non-linear sorption and mobile-immobile phenomena with heavy-tailed immobile-state waiting times \citep{Margolin2003,Schumer2003} are not covered---remains of sufficient generality to reveal the nature of the retardation factor approximation.
That first-order kinetic MIMT has a dispersive effect (i.e., that capture and release of particles independent of one another drives spatial spreading of the distribution of $c$) has long been recognized. In \citet{giddings_molecular_1955}, equations for the spreading of breakthrough curves at the output of a chromatograph, using essentially the assumptions of chemical non-equilibrium, were derived. \citet{valocchi_validity_1985} and \cite{Goltz1987} performed thorough parametric studies of moments for a variety of MIMT processes and their contributions to the spreading of plumes in the subsurface. Many other authors have considered aspects of this topic, as well. Regardless of the rapidity of the MIMT, the retarded ADE does not capture dispersion due to mass transfer: the time-scaling retardation factor does not change the qualitative shape of the solution by adjusting the relative weights of advection and dispersion. In fact, that \eqref{eq:ADE retarded} fails to capture dispersion encoded by \eqref{eq:Explicit treatment of sorption} was made explicitly in a numerical study by \citet{elfeki_modeling_2007}.
At the same time, however, the substitution of retardation factors from equilibrium batch experiments---i.e., the use of \eqref{eq:ADE retarded}---to modify transport equations in the presence of kinetic sorption is frequently presented in expository works as though it is exact. A derivation of the retarded ADE by such means is presented as mathematically exact in the canonical \emph{Hydraulics of Groundwater} text \cite[p. 242]{bear_hydraulics_1979}, given ``equilibrium'' sorption, and in the authors' experience is believed by many hydrogeologists to be exact. \citet{bouwer_simple_1991} also developed a relationship between a soil distribution coefficient and retardation factor by assuming that all solute released at the same instant has been, at any moment, immobile for the exact same amount of time. In a recent textbook \citep[p. 208]{hiscock_kevin_hydrogeology:_2014}, the Bouwer result is also reported without any explicit indication that transport with mass transfer is a dispersive process in which different particles may be immobile for different fractions of any given time interval (although a caveat is given that the Bouwer result assumes instantaneous sorption and equilibrium---i.e., instantaneous desorption---which implies no effect of sorption at all.) Other textbook treatments similarly provide ostensibly exact paths to \eqref{eq:ADE retarded} without indication that dispersion is being suppressed. \citet[p. 66]{zheng_applied_1995} provide an extensive derivation leading to an apparently exact \eqref{eq:ADE retarded}, but silently introduce an approximation analogous to \eqref{eq:Desorption approximation}, below. \citet[p. 117]{fetter_contaminant_1993} similarly discusses linear isotherms in a transport-free context, and then introduces a retardation constant into the ADE, apparently exactly.
In practice, \eqref{eq:ADE retarded} has also been used for the interpretation of push-pull tracer tests aimed at quantifying $D$ and $R$ \citep{schroth_situ_2000}; ignoring the dispersive effect of sorption. The retarded ADE has also commonly been incorporated in numerical codes that handle more complicated geometries. As the user guide for the popular MT3DMS transport modeling software states, ``{[}i{]}t is generally assumed that equilibrium conditions exist between the aqueous-phase and solid-phase concentrations and that the sorption reaction is fast enough relative to groundwater velocity so that it can be treated as instantaneous....Equilibrium-controlled sorption isotherms are generally incorporated into the transport model through the use of the retardation factor'' \citep[p. 12]{zheng_mt3dms:_1999}. So while it is well established that kinetic mass transfer is a cause of dispersion, the use of retardation factors that ignore it under ``local equilibrium'' conditions is common in practical subsurface hydrology, as well as in the literature. In particular, we note that this is the practice in remediation studies performed on EPA Superfund sites \citep{Chen1999,Zheng1991}, as well as U.S. DOE sites \citep{Rogers1992}. In light of the above, new conceptual arguments pinpointing the approximation being made in the apparently exact derivation of \eqref{eq:ADE retarded} appear timely.
Regarding the relationship between degree of local equilibrium and usage of the the retarded ADE, there is more to be said. \citet{wallach_small_1998} and \citet{valocchi_validity_1985} acknowledge dispersion due to mass transfer and identify validity of the local equilibrium assumption (LEA) with applicability of the retarded ADE in light of large hydrodynamic dispersion relative to MIMT-driven dispersion (see Appendix \ref{sec: safety}). However, they do not directly investigate the degree of local disequilibrium. By contrast, \citet{bahr_direct_1987} qualify the extent to which fast kinetic mass transfer leads to pointwise local equilibrium (i.e., reduces the difference between $s$ and $(R-1)c$), without directly addressing dispersion due to mass transfer. However, a direct discussion of the degree of support that a given maximum amount of local disequilibrium provides for a given maximum amount of dispersion due to mass transfer (including the potentially surprising answer, \emph{zero}) does not seem to exist in the literature.
In Section 2, we examine mathematically the the implications of the two conceptions of the retardation factor and show how the derivation of the ADE makes a hidden assumption--akin to the gambler's falacy--that hides its inexactitude. In Section 3 we present a numerical study of plume evolution on a heterogeneous 2D conductivity field, as modeled with rapid first-order MIMT and with a retarded ADE. We show a substantial difference in plume evolution despite the fact that local equilibrium is maintained by the mobile and immobile plumes, highlighting the incorrectness of using the term "local equilibrium assumption" to refer to assumed ADE validity. In Section 4, we sum up what we have demonstrated and draw lessons from it. In Appendix \ref{sec: eulerian derivation}, we show how the retarded ADE may be derived as a special case of first-order MIMT in an Eulerian context, and that the remobilization rate is the parameter that controls the divergence between the formulations. In Appendix \ref{sec: safety}, we explicitly discuss past results concerning when it is proper to employ the retarded ADE, highlighting the centrality of the remobilization rate.
\section{Hidden assumptions in the retarded ADE}
In this section, we establish that interpreting the retarded ADE as exact is to essentially ask for ergodicity to equalize the \emph{absolute} \emph{amount} of time that each particle is immobile in some long time interval, rather than the \emph{fraction} of time immobile. This conflation of absolute and relative frequencies is tantamount to the gambler's fallacy. This fallacy \citep[e.g.,][]{Ayton2004,Sundali2006} represents the erroneous belief that the law of large numbers requires negative auto-correlation in sequences of independent events in order to obtain ``balance'' (informally, that if one has just flipped an unbiased coin for a long string of tails, then heads is now more probable than tails in future flips). In our context, instead of the two states of a coin, we imagine solute particles periodically making a Markovian selection between mobile and immobile states.
It is immediately apparent from viewing $R$ as a scaling factor for time in \eqref{eq:ADE retarded}, that values of $R$ different from unity do not cause any extra dispersion: they simply map the concentration profile at $t$ for any given initial distribution to that at $t/R$ in the case when $R=1$, for the same initial distribution. This is to say: it generates the distribution that would occur if \textit{every} particle spent $t/R$ of the time immobile. If different particles spend different amounts of time immobile during the interval $[0,t]$, then this will represent an additional source of dispersion (which becomes clear when the case $D=0$, $v>0$ in \eqref{eq:ADE retarded} is considered).
\subsection{Relations between mobile and immobile concentrations}
The hypothesis of local equilibrium is \emph{local} in both space
and time: it constrains the fraction of the solute particles at a
given location (i.e., small representative pore volume), at any given
time that are mobile (or equivalently, the instantaneous probability
that a given individual particle is mobile). The retarded ADE it ostensibly
justifies depends on a constraint on the \emph{exact amount of time}
in a given time interval that \emph{each} of the particles is mobile.
In other words, the retardation factor approach attempts to
equate an aggregate \emph{spatial} relationship with a deterministic
temporal quantity. It is important to understand the actual conceptual
relationship between this spatial constraint and temporal particle
behavior. To do so we consider the simplest possible mobile-immobile ``transport''
system---a batch experiment with first-order MIMT---freeing
us from the need to consider extraneous processes.
Specifically, we consider a steady-state batch system consisting of $N_{m}$ mobile particles and $N_{i}$ immobile particles, where these numbers are both large. For our analysis, we employ the conceptual model implied by system \eqref{eq:Explicit treatment of sorption} which, as we have already mentioned, is a generalization of the retarded ADE, and allows for explicit treatment of individual mobile and immobile intervals. Our analysis proceeds in a similar spirit to that of \cite{benson_simple_2009}, considering the aggregate behavior that results from independent particles, each of which has the same defined probability distributions for lengths of its mobile and immobile intervals. In this system, the mobile particles have probability $\lambda$ of immobilization per unit time, and the immobile particles have a probability $\mu$ of remobilization per unit time. It follows the expected duration of a single immobilization event is $\mu^{-1}$. We assume all particles are mutually independent and define $K$ as the rate of immobilization, in particles per second: $K=\lambda N_{m}$. Based on the equilibrium conception of retardation and the principle of conservation of mass, the retardation factor satisfies
\begin{equation}
R=\frac{\mathrm{E}\left[N_{i}+N_{m}\right]}{\mathrm{E}\left[N_{m}\right]},\label{eq:Retarded}
\end{equation}
where $\mathrm{E}[\cdot]$ represents mathematical expectation. Little's law is the intuitive statement that the expected number of
particles in a state is equal to their rate of arrival multiplied
by their expected wait in that state \citep[p. 37]{bhat_simple_2008}.
If the state of interest is the immobile state, this implies $\mathrm{E}[N_{i}]=\mathrm{E}\left[K\right]/\mu$.
Then we can conclude that
\begin{equation}
\begin{array}{rl}
R & =\frac{\mathrm{E}[N_{m}]\left(\lambda/\mu+1\right)}{\mathrm{E}[N_{m}]}\\
\\
& =1+\lambda/\mu.
\end{array}
\label{eq: retardation}
\end{equation}
It is possible to take this aggregate (multi-particle) spatial behavior
and draw conclusions about the temporal behavior of any single particle.
However, we shall see that the validity of these assumptions only
constrain the \emph{expected} behavior of any particular particle.
This is to say that if we define $F_{t}$ to be the a random variable
representing the amount of time a particular particle is mobile in the
interval $[0,t]$, the assumptions underlying the retardation approach
will correctly establish that $\mathrm{E}[F_{t}]=t/R$. They will
not, however, establish that $F_{t}=t/R$, which is what would be
required for the retarded ADE to be exact. The former condition is
naturally weaker---constraining only the average of a whole population
of solute particles---whereas the latter states that \emph{each} solute
particle in a population is immobile for the same amount of time. It
is useful to consider these claims precisely.
\subsection{The expected time a single particle is immobile is fixed by $R$}
By symmetry of particle behavior (i.e., all have the same tendencies
to immobilize and remobilize), the retardation approach implies that each particle
is expected (in the mathematical sense) to spend $1/R$ of the \emph{time}
mobile. To see this, imagine a steady-state, batch system in which $N$
particles are immobilizing and remobilizing independently of each other. Define,
for particle $n$, the indicator function
\begin{equation}
I_{n}(t)\equiv\begin{cases}
0 & \textrm{if immobile at \ensuremath{t}}\\
1 & \textrm{if mobile at }t
\end{cases},\label{eq:}
\end{equation}
which is only non-zero in such cases as the particle is mobile at time $t$. Define
\begin{equation}
\Omega_{t}(N)\equiv\intop_{0}^{t}{\displaystyle \frac{1}{N}\sum_{n=1}^{N}I_{n}(\tau)}d\tau.\label{eq:omega def}
\end{equation}
Then
\begin{equation}
\underset{N\rightarrow\infty}{lim}\Omega_{t}(N)=\intop_{0}^{t}\frac{1}{R}d\tau=\frac{t}{R}\label{eq:integration 1st}
\end{equation}
which follows, because in the limit $N\rightarrow\infty$, a sample
mean converges to the expectation (by the law of large numbers), and
the expected value of an indicator function is the probability of
being mobile, and $1/R$ of the $N$ particles are mobile at every instant.
Because of linearity, it is possible to rearrange the order of summation
and integration, so that
\begin{equation}
\begin{array}{rl}
\underset{N\rightarrow\infty}{lim}\Omega_{t}(N) & =\underset{N\rightarrow\infty}{lim}\frac{1}{N}{\textstyle {\displaystyle \sum_{n=1}^{N}\intop_{0}^{t}{\displaystyle I_{n}(\tau)}d\tau}}\\
\\
& =\mathrm{E}\left[\intop_{0}^{t}{\displaystyle I_{n}(\tau)}d\tau\right]\\
\\
& =\mathrm{E}[F_{t}]
\end{array}\label{eq:series 1st}
\end{equation}
Since $I_{n}(t)$ is just the indicator function that is unity when
the particle is mobile, the integral represents the amount of time in
the interval $[0,t]$ in which particle $n$ is mobile. Combining \eqref{eq:integration 1st}
and \eqref{eq:series 1st}, we see $\mathrm{E}[F_{t}]=t/R$.
\subsection{The absolute time a single particle is immobile is not fixed by $R$}
The conclusion that $\mathrm{E}[F_{t}]=t/R$ is the strongest that
can be made. The stronger statement, that $F_{t}=t/R$, for any given
particle, is false. To see this, consider a system over some interval
in which all the particles with even index are always immobile and all
the particles with odd index are never immobile during the interval
$t$ (not because they are qualitatively different, just that the
particles are independent and this is one possible, though not likely,
configuration). Then this system satisfies \eqref{eq:omega def} and
\eqref{eq:series 1st} for $R=2$, though it is not true for any particle
that $\frac{t}{2}=\intop_{0}^{t}{\displaystyle I_{n}(t)}dt$.
It is true that, applying the law of large numbers for large $t$,
it follows that after a long time (i.e., a large number of immobilization and remobilization events), the actual fraction of time every particle
spends mobile, $F_{t}/t$, approaches $1/R$. Naturally, if each particle
were to spend \emph{exactly} $1/R$ of the time immobile, for all $t$,
then we could compute $F_{t}\equiv t/R$ (valid for every particle)
and the retardation factor approach would be exact. However,
there is no reason to expect $F_{t}$ to converge to $t/R$ as $t\rightarrow\infty$.
While possibly unintuitive, such situations are common: consider that
as $t\rightarrow\infty$, $\left(t+1\right)/t\rightarrow1$ but $\left(t+1\right)\nrightarrow t$.)
The law of large numbers concerns itself exclusively with \emph{relative}
frequencies, not absolute frequencies. This is a subtle distinction,
but an important one: this distinction is what the gambler's
fallacy (discussed earlier) turns on.
\section{Local equilibrium and retarded ADE validity}
In this section, we directly consider the degree of support which ``local equilibrium'' (this is to say, fast kinetics) provides to the usage of the retarded ADE. We perform two particle tracking simulations: one employing first-order MIMT, governed by \eqref{eq:Explicit treatment of sorption}, and one employing the retarded ADE \eqref{eq:ADE retarded} with the corresponding $R$ \eqref{eq: retardation}. In so doing, we are able to monitor the degree of local equilibrium between mobile and immobile plumes in the first-order MIMT model, and its coherence with the retarded ADE model that purports to capture it.
Our study begins by generating a 40 by 80 m random log-hydraulic conductivity field with a multi-Gaussian correlation structure described by an exponential semivariogram with correlation length 5 m, geometric mean conductivity 1e-4 m/s, and $\sigma^2_{\ln K} = 2$ (moderate heterogeneity), discretized into blocks 1 m on a side. The resulting conductivity field is shown in Figure \ref{fig: flow field}. This log-conductivity field is used with the finite-volume numerical flow and transport solver PFLOTRAN \citep{lichtner2015pflotran} to determine the steady-state cell-center velocities. For this computation no-flow boundary conditions are imposed at $x=0$ and $x=40$, a constant pressure of 111135 Pa is imposed at $y=80$, and constant pressure of 101325 Pa is imposed at $y = 0$. The resulting velocity field is also illustrated in Figure \ref{fig: flow field}. Both particle tracking simulations are performed by randomly introducing 500 particles into a circle of radius 2 m, centered at $x=25$ m, $y = 75$ m. When particles are mobile, their positions are tracked by making successive steps of constant duration 0.1 h, during which they passively follow the flow lines. At the end of each step, a small random translation is added to model local-scale dispersion, described by longitudinal dispersivity 0.01 m, and transverse dispersivity 0.001 m.
For the MIMT simulation, the times of successive immobilization and remobilization events for each particle are generated by draws from exponential random number generators with rate parameters $\lambda=10$ and $\mu=\frac{1}{3}$, respectively. The retarded ADE simulation was performed by disabling particle immobilization altogether and using an alternative velocity field, with identical directions to those used in the MIMT simulation, but all of whose magnitudes were divided by $R$, where $R=1+\frac{\lambda}{\mu} = 31$. Plume concentrations from both particle tracking simulations are determined at $t = 1$ y and $t = 5$ y by performing kernel density estimation using the locations of all particles at the relevant time. These plumes are shown in Figure \ref{fig:plumes}. From examination of the figure, the strong divergence of the two models is apparent.
Approximate local equilibrium for the MIMT model was established by comparing mobile and immobile plumes at fixed times, and by tabulating each plume's spatial moments over time and verifying their coherence. Graphs of the first two spatial moments are presented in Figure \ref{fig: moments} to illustrate how closely the mobile and immobile plumes cohere. Note that the ratio of immobile to mobile particles is always approximately 30, resulting in smoother immobile particle graphs. We have thus demonstrated an example of a realistic system in which \textit{local equilibrium is satisfied, but performance of the retarded ADE is very poor}. So the use of the term ``local equilibrium assumption'' to refer to the assumption of retarded ADE validity is misleading. Indeed, as we note in Appendix \ref{sec: safety}, classic ``local equilibrium'' metrics actually quantify the relative strengths of MIMT-driven dispersion and hydrodynamic dispersion. They are legitimate metrics for the validity of of the retarded ADE, but do not concern local equilibrium, per se.
\begin{figure}
\centerline{
\includegraphics[trim={7.5cm 0cm 2.5cm 1cm}, clip, scale=0.75]{FIG_kfield.pdf}
\includegraphics[trim={5.5cm 0cm 5.5cm 1cm}, clip, scale=0.75]{FIG_flowfield.pdf}
}
\caption{Left: Heat map of $\log_{10} K$, where $K$ is the local hydraulic conductivity of the field in which the particle tracking simulations were performed. Right: quiver plot of heterogeneous velocity field computed by PFLOTRAN using the $K$-field. Each cell-center velocity is indicated by an arrow whose length represents its relative speed and orientation indicates its direction.}
\label{fig: flow field}
\end{figure}
\begin{figure}
\centering
{\fontfamily{phv}\selectfont
\begin{tabular}{c c c}
& \parbox{7cm}{\centering\textbf{\textit{t} = 1 y}} & \parbox{7cm}{\centering\textbf{\textit{t} = 5 y}}\\
\parbox{0.25cm}{\rotatebox{90}{\textbf{Exact first-order MIMT}}}&
\parbox{7cm}{\includegraphics[trim={5.5cm 0cm 5.5cm 0cm}, clip, scale=0.7]{FIG_plume_sorp_8760h_500.pdf}}&
\parbox{7cm}{\includegraphics[trim={5.5cm 0cm 5.5cm 0cm}, clip, scale=0.7]{FIG_plume_sorp_43800h_500.pdf}}\\
\parbox{0.25cm}{\rotatebox{90}{\textbf{Retarded ADE approximation}}}&
\parbox{7cm}{\includegraphics[trim={5.5cm 0cm 5.5cm 1cm}, clip, scale=0.7]{FIG_plume_tard_8760h_500.pdf}}&
\parbox{7cm}{\includegraphics[trim={5.5cm 0cm 5.5cm 1cm}, clip, scale=0.7]{FIG_plume_tard_43800h_500.pdf}}\\
\end{tabular}
}
\caption{Heat maps of plume concentration at two times under exact first-order MIMT and the retarded ADE approximation. All plumes used the same velocity field and release location. Hue closer to the red end of the spectrum indicates higher concentration, but scales differ between heat maps.}
\label{fig:plumes}
\end{figure}
\begin{figure}
\centering
\includegraphics[trim={2cm 2cm 2cm 2cm}, clip, scale=0.5]{FIG_mean.pdf}\\
\includegraphics[trim={2cm 2cm 2cm 2cm}, clip, scale=0.5]{FIG_variance.pdf}
\caption{Spatial moments of mobile (solid blue lines) and immobile (dashed black lines) plumes. Top: centroid $x$-coordinate. Bottom: plume $x$-coordinate variance.}
\label{fig: moments}
\end{figure}
\section{Summary and concluding discussion}
The two major contributions of this work are the following:
\begin{enumerate}
\item We note that the retarded ADE for \textit{arbitrary} $R$ is derived in canonical sources in a seemingly exact fashion, and is often treated as exact in the literature. We also note that the retarded ADE approach is a special case of first-order kinetic MIMT equations in the limit of instantaneously fast remobilization (i.e. when $R=1$), and otherwise neglects the dispersion that has long been known to be a feature of kinetic mass transfer. We resolve the contradiction by showing how the ostensibly exact derivation of the retarded ADE silently introduces an approximation that is equivalent to the gambler's fallacy.
\item Through a numerical study of transport in a heterogeneous aquifer, we demonstrate that use of the term ``local equilibrium assumption'' to describe the assumption of ADE validity is misleading. In our example, we demonstrate an MIMT-generated plume that is largely disjoint from the plume predicted by use of the corresponding retarded ADE, despite the fact that local equilibrium between mobile and immobile plumes holds everywhere. We note that previously published criteria for the validity of the ``local equilibrium assumption'' are actually criteria for conditions in which the dispersive effect of MIMT is overwhelmed by that of local-scale hydrodynamic fluctuations. We concur that this is the correct condition for usage of the retarded ADE.
\end{enumerate}
Despite common assumptions to the contrary, the dispersive effect of MIMT, even under ``local equilibrium'' conditions, can not be discarded a priori. Since this extra dispersion may cause un-modeled early- or late-time breakthrough, how to treat it presents a practical question to working hydrogeologists and environmental engineers. The critical role of the remobilization rate, $\mu$, in driving dispersion at late time is clear (see Appendices \ref{sec: eulerian derivation} and \ref{sec: safety}), with only truly \emph{instantaneous} remobilization (i.e., no mass transfer) recovering (\ref{eq:ADE retarded}), and the dispersive effect of sorption increasing as $\mu$ shrinks. Furthermore, MIMT generates anomalous (asymmetric) plumes that are not well described by an ADE at early time. \cite{hansen_effective_2015} presented the guideline, for small $\mu$, in the $D \rightarrow 0$ limit, that an ADE model becomes adequate after time $t \mu > 70$. Slow mobile-immobile kinetics, and thus and small values of $\mu$, are in reality widespread \citep{pignatello_mechanisms_1996}, so these limitations are practically important.
We hope that by revisiting this classic topic, we are able to clear up some misconceptions that---as we established in the introduction---continue to persist in the literature, and which have the potential to adversely impact remedial actions.
{
|
\section{Introduction}\label{sec:intro}
In recent years, the systems and control community has turned to study of networked systems and multi-agent systems in the context of social sciences. Of particular interest are social networks, where groups of people interact with acquaintances through \emph{interpersonal relationships}.
One problem of ``opinion dynamics'' has been of particular interest; how do the opinions of individuals for a given issue evolve as they discuss this issue in a social network? A recent survey on opinion dynamics is presented in \cite{friedkin2015_socialsurvey}. An important aspect of opinion dynamics is \emph{social power}, which in one sense can be considered as the weight/power/influence an individual has on the opinion discussion, relative to the weight/power/influence of the other individuals in the social network. This relativity arises due to interpersonal relationships and their strengths (which may be unidirectional). This concept is studied in the seminal works \cite{french1956_socialpower,degroot1974OpinionDynamics}. The evolution of social power is studied in \cite{friedkin2011_powerevolution}. The paper \cite{parsegov2017_multiissue} studies the case where multiple, interdependent issues are simultaneously discussed. Selecting the most influential individual in social diffusion models is studied in \cite{kempe2003_maxspread}. A social network with stubborn individuals who remain attached to their initial opinions is studied in \cite{friedkin1990_FJsocialmodel,stubborn}. The centralised DeGroot-Friedkin model for the evolution of social power is proposed and analysed in \cite{jia2015opinion_SIAM}. Distributed discrete- and continuous-time DeGroot-Friedkin models are studied in \cite{xu2015_modified_DF} and \cite{chen2017_DFdistributed} respectively.
According to French Jr. and Snyder in \cite{french1959_leadership}, ``\emph{leadership} is the potential social influence of one part of the group over another.'' From the perspective of opinion dynamics, a leader can therefore be seen as an individual or a group of individuals that has a disproportionate amount of control over the opinion discussion process. In the context of social power, one can therefore refer to a leader/leader group as the \emph{socially dominant} individual/group of individuals. The fact that social power tends to accumulate with one individual or a subgroup of individuals in a social network is reported empirically in \cite{friedkin2011_powerevolution} and theoretically in \cite{jia2015opinion_SIAM}. This individual or subgroup is defined explicitly by the interpersonal relationships in the social network. Motivated by this concept of social dominance/leadership, and using the DeGroot-Friedkin model to describe the social network, we begin with network topologies which have a single socially dominant individual, and seek to study \emph{control strategies, including introduction of new individuals into the network and/or establishment of new interpersonal relationships, that will cause the social dominance to shift to another individual}. We now introduce the DeGroot-Friedkin model to better motivate the formal problem statement which follows in the sequel. In order to allow readers to quickly grasp the concepts of the new model and understand the motivations, in the following subsection, where possible we leave out definitions and exact mathematical results; these will be included in Section~\ref{section:background_problem}. The terms ``self-weight'', ``individual social power'' and ``social power" will be used interchangeably.
\subsection{The DeGroot-Friedkin Model}\label{ssec:df_model}
The discrete-time DeGroot-Friedkin model comprises a consensus model for describing the opinion dynamics (details are given below) and a mechanism for updating self-weights (the weight an individual applies to its own opinion value in the consensus process). We define $\mathcal{S} = \{0, 1, 2, \ldots\}$ to be the set of indices of sequential issues which are being discussed by the social network. For a given issue $s$, the social network discusses the issue using the discrete-time DeGroot consensus model (with constant weights throughout the discussion of the issue). At the end of the discussion (i.e. when the DeGroot model has effectively reached steady state), each individual reflects upon, and judges its impact on the discussion. This mechanism is termed reflected self-appraisal, with ``reflection'' referring to the fact that adjustments to weights are made after discussion on an issue. The individual then updates its own self-weight and discussion begins on the next issue $s+1$ (using the same consensus model but now with adjusted weights). We now explain the mathematical modelling of the mechanism for updating opinions within an issue, and the updating of self-weights from one issue to the next.
\subsubsection{DeGroot Consensus of Opinions}
For each issue $s \in \mathcal{S}$, each individual updates its opinion $y_i(s,\cdot) \in \mathbb{R}$ at time $t+1$ as
\begin{equation}
y_i(s, t+1) = w_{ii}(s) y_i(s, t) + \sum_{j\neq i}^n w_{ij}(s) y_j(s, t)
\end{equation}
where $w_{ii}(s)$ is the self-weight individual $i$ places on its own opinion and $w_{ij}$ is the weight given by agent $i$ to the opinion of its neighbour individual $j$. As will be made apparent in the sequel, $\sum_{j = 1}^n w_{ij} = 1$, which implies that individual $i$'s new opinion value $y_i(s, t+1)$ is a convex combination of its own opinion, and the opinions of its neighbours at the current time instant. The opinion dynamics for the entire social network may be expressed as
\begin{equation}\label{eq:opinion_network}
\vect y(s, t+1) = \mat W(s) \vect y(s, t)
\end{equation}
where $\vect y(s, t) = [y_1(s, t) \; \cdots \; y_n(s, t)]^\top$ is the vector of opinions of the $n+1$ agents in the network at time instant $t$. This model was studied in \cite{degroot1974OpinionDynamics} with $\mathcal{S} = \{0\}$ (i.e. only one issue was discussed). The dynamics of \eqref{eq:opinion_network}, and the graphical conditions required for opinions to converge, have been well studied. Next, we describe the model for the updating of $\mat W(s)$ (specifically $w_{ii}(s)$ via a reflected self-appraisal mechanism that occurs at the end of discussion of an issue $s$). For simplicity, we assume that each individual's opinion, $y_i(s,t)$, is a scalar. Kronecker products may be used if each individual's opinion state is a vector $\vect y_i \in \mathbb{R}^p, p \geq 2$.
\subsubsection{Friedkin's Self-Appraisal Model for Determining Self-Weight}
The Friedkin component of the model proposes a method for updating the self-weight (individual social power, self-confidence or self-esteem) of individual $i$, which is denoted by $x_i(s) = w_{ii}(s) \in [0,1]$ (the $i^{th}$ diagonal term of $\mat W(s)$) \cite{jia2015opinion_SIAM}. Define the vector $\vect x(s) = [x_1(s) \; \cdots \; x_n (s)]^\top$ as the vector of self-weights for the individuals of the social network, with starting self-weight $0 \leq x_i(0) \leq 1$ satisfying $\sum_i x_i(0) = 1$. The self-weight vector $\vect x(s)$ is updated at the end of issue $s$ as
\begin{equation}\label{eq:x_update}
\vect x(s+1) = \vect{\zeta}(s)
\end{equation}
where $\vect \zeta(s)^\top$ is the unique nonnegative left eigenvector of $\mat W(s)$ associated with the eigenvalue $1$, normalised such that $\vect 1_n^\top \vect \zeta(s) = 1$, see \cite{jia2015opinion_SIAM}. When individual $i$ adjusts its value of $w_{ii}(s) = x_i(s)$, it necessarily must adjust the weights $w_{ij}(s), j\neq i$ to maintain $\sum_{j=1}^n w_{ij} = 1$. The precise structure of $\mat{W}(s)$, and its properties giving rise to the existence of $\vect{\zeta}(s)^\top$, will be provided in the sequel. See Remark~\ref{rem:selfweight_update} in Section \ref{xxx} for comments on the motivation for this update mechanism. Convergence properties will be presented in the sequel, but under mild assumptions on the social network topology, it is shown that $\lim_{s\to \infty} \vect{\zeta}(s)^\top = \vect{x}^*$ where $\vect{x}^*$ is the constant vector of \emph{social power at equilibrium}.
\subsection{Contributions}
In order to simplify the problem of achieving change of a socially dominant leader, we will only consider \emph{social power at equilibrium $\vect{x}^*$} in this paper. It was shown in \cite{jia2015opinion_SIAM} that if the social network has a specific topology termed the star topology, all social power at equilibrium accumulates with a single individual $k$ as issues are sequentially discussed, in an ``autocratic configuration''. In this paper, we show that by \emph{strategic} introduction of new individuals and/or new interpersonal relationships into the social network, not only is the autocratic configuration broken but \emph{if the new relationship is sufficiently strong, other identifiable individual(s) will have social power at equilibrium \emph{greater than individual $k$}.} Specifically, we derive necessary and sufficient conditions \emph{based on local information} for the relationship strength. This is in contrast to many control strategies on networked systems which rely on global information \cite{max,consent}. In fact, a number of different strategies are considered. We also propose a strategy whereby two socially dominant individuals in separate networks can combine their networks and together remain socially dominant.
While the results are initially presented mathematically as inequalities, we provide detailed analysis and interpretation. In doing so, we show that the strategies are remarkably intuitive and precisely what one would expect when considered from a sociological context. The fact that the strategies affect the \emph{social power of individuals} in a social network which is sequentially discussing issues implies that we have developed control strategies for \emph{affecting/influencing the opinion dynamics process}.
\subsection{Paper Structure}
In Section~\ref{section:background_problem}, we provide notations, an introduction to graph theory, and convergence results for the DeGroot-Friedkin model. At the same time, a formal problem statement is given. The main results are presented in Section~\ref{sec:main_result}. Simulations are presented in Section~\ref{sec:simulations} and conclusions are drawn in Section~\ref{sec:conclusions}.
\section{Background and Formal Problem Statement}\label{section:background_problem}
We begin by introducing some mathematical notations used in the paper. Let $\vect 1_n$ and $\vect 0_n$ denote, respectively, the $n\times 1$ column vectors of all ones and all zeros. For a vector $\vect x\in\mathbb{R}^n$, $0\preceq\vect x$ and $0 \prec \vect x$ indicate component-wise inequalities, i.e., for all $i\in\{1,2,\ldots,n\}$, $0\leq x_i$ and $0<x_i$, respectively. Let $\Delta_n$ denote the $n$-simplex, the set which satisfies $\{\vect x\in \mathbb{R}^n : 0 \preceq \vect x, \vect 1_n^\top \vect x = 1 \}$. The canonical basis of $\mathbb{R}^n$ is given by $\mathbf{e}_1, \ldots, \mathbf{e}_n$. Define $\wt{\Delta}_n = \Delta_n \backslash \{ \mathbf{e}_1, \ldots, \mathbf{e}_n \}$ and $\text{int}(\Delta_n) = \{\vect x\in \mathbb{R}^n : 0 \prec \vect x, \vect 1_n^\top \vect x = 1 \}$. For the rest of the paper, we shall use the terms ``node'', ``agent'', and ``individual'' interchangeably.
\subsection{Graph Theory}
The interaction between individuals in a social network is modelled using a weighted directed graph, denoted as $\mathcal{G} = (\mathcal{V}, \mathcal{E})$. Each individual agent is a node in the finite, nonempty set of nodes $V = \{v_1, \ldots, v_n\}$. The set of ordered edges is $\mathcal{E} \subseteq \mathcal{V}\times \mathcal{V}$. We denote an ordered edge as $e_{ij} = (v_i, v_j) \in \mathcal{E}$, and because the graph is directed, in general $e_{ij}$ and $e_{ji}$ may not both exist. An edge $e_{ij}$ is said to be outgoing with respect to $v_i$ and incoming with respect to $v_j$. The presence of an edge $e_{ij}$ connotes that individual $j$ learns of, and takes into account, the opinion value of individual $i$ when updating its own opinion. The incoming and outgoing neighbour set of $v_i$ are respectively defined as $\mathcal{N}_i^+ = \{v_j \in \mathcal{V} : e_{ji} \in \mathcal{E}\}$ and $\mathcal{N}_i^- = \{v_j \in \mathcal{V} : e_{ij} \in \mathcal{E}\}$. The relative interaction matrix $\mat C\in\mathbb{R}^{n\times n}$ associated with $\mathcal{G}$ has nonnegative entries $c_{ij}$, termed ``relative interpersonal weights'' in \cite{jia2015opinion_SIAM}. The entries of $\mat C$ have properties such that $0 < c_{ij} \leq 1 \Leftrightarrow e_{ji} \in \mathcal{E}$ and $c_{ij} = 0$ otherwise. It is assumed that $c_{ii} = 0$ (i.e. with no self-loops), and we impose the restriction that $\sum_{j\in\mathcal{N}_i^+} c_{ij} = 1$ (i.e. that $\mat C$ is a row-stochastic matrix).
A directed path is a sequence of edges of the form $(v_{p_1}, v_{p_2}), (v_{p_2}, v_{p_3}), \ldots$ where $v_{p_i} \in \mathcal{V}, e_{ij} \in \mathcal{E}$. Node $i$ is reachable from node $j$ if there exists a directed path from $v_j$ to $v_i$. A graph is said to be strongly connected if every node is reachable from every other node. The relative interaction matrix $\mat C$ is irreducible if and only if the associated graph $\mathcal{G}$ is strongly connected ($\mat C$ is known in some literature as the weighted adjacency matrix). If $\mat C$ is irreducible then it has a unique (up to a scaling) left eigenvector $\vect{\gamma}^\top$, with all entries strictly positive, associated with the eigenvalue 1 (Perron-Frobenius Theorem, see \cite{godsil2001algebraic}). Henceforth, we shall call this left eigenvector $\vect{\gamma}^\top$ the \emph{dominant left eigenvector of $\mat C$} and assume that $\vect{\gamma}^\top$ has been normalised such that it has the property $\vect{\gamma}^\top\vect 1_n = 1$.
\subsection{Convergence Results for the DeGroot-Friedkin Model}\label{xxx}
We now provide additional, specific details on the model. For a given issue, the influence matrix $\mat W(s)$ is defined as follows
\begin{equation}\label{eq:W_matrix}
\mat W(s) = \mat X(s) + (\mat I_n - \mat X(s))\mat C
\end{equation}
where $\mat C$ is the relative interaction matrix associated with the graph $\mathcal{G}$, and the matrix $\mat X(s) \doteq diag[\vect x(s)]$. From the fact that $\mat C$ is row-stochastic with zero diagonal entries, \eqref{eq:W_matrix} implies that $\mat W(s)$ is a row-stochastic matrix. Note that $\vect{\zeta}(s)^\top \vect{1}_n = 1$ implies that $\vect x(s) \in \Delta_n$ for all $s$. From \eqref{eq:W_matrix} and the fact that $\mat{C}$ is constant, it is apparent that by adjusting $w_{ii}(s+1) = \zeta_i(s)$, individual $i$ also scales $w_{ij}(s+1), j \neq i$ by $(1-w_{ii}(s+1))$ in order to maintain the row-stochastic property of $\mat{W}(s)$.
\begin{remark}[Social Power]\label{rem:selfweight_update}
The precise motivation behind using \eqref{eq:x_update} as the updating model for $\vect{x}(s)$ is detailed in \cite{jia2015opinion_SIAM}, but we provide a brief overview here in the interest of making this paper self-contained. The definition of $\mat{W}$ in \eqref{eq:W_matrix} ensures that, for any given $s$, there holds $\lim_{t \to \infty} y(s,t) = (\vect{\zeta}(s)^\top \vect y(s,0))\vect{1}_n$. In other words, for any given issue $s$, the opinions of every individual in the social network reach a consensus value $\vect{\zeta}(s)^\top \vect y(s,0)$ equal to a convex combination of their initial opinion values $\vect y(s,0)$. The elements of $\vect{w}(s)^\top$ are the convex combination coefficients, i.e. $\zeta_{i}(s)$ represents precisely the amount of weight/power that individual $i$ had on the opinion discussion for issue $s$. For a given issue $s$, $\zeta_i(s)$ is a manifestation of individual $i$'s social power in the social network, as it is in effect the ability of individual $i$ to control the outcome of a discussion \cite{cartwright1959social_book}. The reflected self-appraisal mechanism therefore describes an individual $1)$ observing how much power it had on the discussion of issue $s$ (the nonnegative quantity $\zeta_{i}(s)$) and, $2)$ for the following issue $s+1$, adjusting its self-weight to be equal to this power, i.e. $x_i(s+1) = w_{ii}(s+1) = \zeta_i(s)$. As can be observed from \eqref{eq:W_matrix} and because $\mat{C}$ is constant, adjusting $w_{ii}(s+1)$ also adjusts the interpersonal weights $w_{ij}(s+1)$.
\end{remark}
It is shown in [Lemma 2.2, \cite{jia2015opinion_SIAM}] that the system \eqref{eq:x_update} describing the update of self-weights, is equivalent to
\begin{equation}\label{eq:DF_system}
\vect x(s+1) = \vect F(\vect x(s))
\end{equation}
where the nonlinear vector-valued function $\vect F(\vect x(s))$ is defined as
\begin{align}\label{eq:map_F_DF}
\vect F( \vect x(s) ) = \begin{cases}
\mathbf e_i & \hspace*{-6pt} \text{if } x_i(s) = \mathbf e_i, \text{for any } i \\ \\
\alpha (\vect x(s)) \begin{bmatrix} \frac{c_1}{1-x_1(s)} \\ \vdots \\ \frac{c_n}{1-x_n(s)} \end{bmatrix} & \text{otherwise }
\end{cases}
\end{align}
with $\alpha(\vect x(s)) = 1/\sum_{i=1}^n \frac{c_i}{1- x_i(s)}$. Much of this paper will deal with scenarios where the underlying graph has a star topology or its variants, the definition and relevance of which are now given.
\begin{definition}[Star topology]\label{def:star}
A strongly connected graph\footnote{While it is indeed possible to have a star graph that is not strongly connected, this paper similarly to \cite{jia2015opinion_SIAM} deals only with graphs which are strongly connected.} $\mathcal{G}$ is said to have star topology if there exists a node $i$, which is called the centre node, such that every edge of $\mathcal{G}$ is either to or from node $i$
\end{definition}
Note that the irreducibility of $\mat C$ implies that the star topology must include edges in both directions between the centre node $v_i$ and every other node $v_j, j \neq i$. We now provide a lemma and a theorem regarding the convergence of $\vect{F}(\vect{x}(s))$ as $s\to \infty$.
\begin{lemma}[Lemma 3.2, \cite{jia2015opinion_SIAM}]\label{lem:star}
Suppose that $n \geq 3$, and suppose further that $\mathcal{G}$ has star topology, which without loss of generality has centre node $v_1$. Let $\mat{C}$ be the row-stochastic and irreducible adjacency matrix, with zero diagonal entries, associated with $\mathcal{G}$. Then for all initial conditions $\vect{x}(0) \in \wt{\Delta}_n$, the self-weights $\vect x(s)$ converge to the fixed point $\vect{x}^* = \mathbf{e}_1$ as $s \to \infty$.
\end{lemma}
\begin{theorem}[Theorem 4.1, \cite{jia2015opinion_SIAM}]\label{thm:DFmain}
For $n\geq 3$, consider the DeGroot-Friedkin dynamical system \eqref{eq:DF_system} with a relative interaction matrix $\mat C$ that is row-stochastic, irreducible, and has zero diagonal entries. Assume that the digraph $\mathcal{G}$ associated with $\mat C$ does not have star topology and define $\vect\gamma^\top$ as the dominant left eigenvector of $\mat C$. Then,
\begin{enumerate}[label=(\roman*)]
\item \label{prty:thm_DFmain01} For all initial conditions $\vect x(1) \in \wt{\Delta}_n $, the self-weights $\vect x(s)$ converge to $\vect x^*$ as $s\to\infty$. Here, $\vect x^* \in \wt{\Delta}_n $ is the unique fixed point satisfying $\vect x^* = \vect F(\vect x^*)$.
\item \label{prty:thm_DFmain02} There holds $x^*_i < x^*_j$ if and only if $\gamma_i < \gamma_j$, for any $i,j$, where $\gamma_i$ is the $i^{th}$ entry of the dominant left eigenvector $\vect\gamma$. There holds $x^*_i = x^*_j$ if and only if $\gamma_i = \gamma_j$.
\item \label{prty:thm_DFmain03} The unique fixed point $\vect x^*$ is determined only by $\vect\gamma^\top$, and is independent of the initial conditions.
\end{enumerate}
\end{theorem}
An interpretation of Lemma~\ref{lem:star} and Theorem~\ref{thm:DFmain} is given in below in Remark~\ref{rem:autocratic}.
\subsection{Formal Problem Statement}\label{ssec:problem_def}
In this paper, we investigate \emph{how additional nodes and/or edges strategically connected to a star topology can change the social power at equilibrium, $\vect{x}^*$}. To that end, we begin first by providing definitions which will aid in describing our problem and discussing results obtained. Moreover, we are interested in comparing the social power of individuals within the network at equilibrium, i.e. when $s\to \infty$. We will therefore refer to the equilibrium value $x_i^*$ as the social power of individual $i$ when there is no ambiguity (as opposed to the evolving $x_i(s)$ when $s < \infty$).
To simplify the problem, we do not study the evolution of the opinions $\vect y(s,t)$. Under the assumption that $\mat C$ is irreducible, it is shown in \cite{jia2015opinion_SIAM} that, for any issue $s$, the opinions always converge as $\lim_{t\to \infty} \vect{y}(s,t) = (\zeta(s)^\top \vect{y}(s,0))\vect{1}_n$. As discussed above, we are interested in individual social power of the network.
\begin{definition}[Autocratic Network]\label{def:autocrat}
A social network is said to be an autocratic configuration, with node $v_i$ being the autocrat, if $\vect x(s) = \mathbf{e}_i$.
\end{definition}
\begin{definition}[Social dominance/leadership]\label{def:leader}
Node $v_i$ is said to be the socially dominant/leader node in the network if $x_i^* > x_j^*$ for all $j \neq i$. In other words, at equilibrium, the social power of individual $i$ is greater than the social power of any other individual in the social network.
\end{definition}
\begin{remark}[Autocratic tendency]\label{rem:autocratic}
Lemma~\ref{lem:star} has an important social connotation. One can consider $x_i(0)$ as individual $i$'s \emph{estimate} of its social power when the social network is first formed, before any issue discussion. For any initial estimate $\vect{x}(0) \in \wt{\Delta}_n$ (that is, no individual $i$ believes $x_i(0) = 1$), the star topology network tends to an autocratic configuration at equilibrium, $\vect x^* = \mathbf{e}_1$. This implies that, for the first few issues, opinion discussion will occur with everyone contributing to the final consensus value. However, as more issues are discussed, the centre individual increasingly guides the outcome of discussions until, for $s = \infty$, only the centre individual's opinion value matters.
\end{remark}
\begin{remark}\label{rem:trust}
In \cite{jia2015opinion_SIAM}, the constant entries $c_{ij}$ of $\mat{C}$ are termed ``relative interpersonal weights'', and we will keep with this terminology. However, one can also consider $c_{ij}$ as the amount of ``trust'' individual $i$ has for individual $j$ or the strength of ``influence'' individual $j$ has on individual $i$. In other words, $c_{ij}$ captures the strength of a unidirectional relationship (unidirectional since $c_{ij} \neq c_{ji}$ in general).
\end{remark}
For a given graph $\mathcal{G}$ with star topology, with centre node $v_1$, let us call the other nodes \emph{subject nodes} in the sense that they are subjects to the autocrat centre node. In Fig.~\ref{fig:star_top}, these are nodes $v_i, i = 2, ..., 7$. We are going to study how the autocracy can be disrupted by introduction of a perturbation to the star graph. This leads us to define a new type of node. An \emph{attacker node} is a node $v_j$ which forms edges $e_{ji}, e_{ij}$ with some node $v_i$, $i \neq 1$, i.e. a subject node. In doing so, we modify the graph $\mathcal{G}$ to become $\bar{\mathcal{G}}$ which is no longer a star. In Fig.~\ref{fig:single_attack}, node $v_8$ is the attacker node, forming edges with node $v_7$. We call node $v_j$ an attacker node because, as will become apparent in the sequel, the weights $c_{ji}$ and $c_{ij}$ determine the social power $x_1^*$ of the autocrat node $v_1$. In other words, $v_j$ attacks the social dominance of $v_1$. Note that two edges, $e_{ji}, e_{ij}$ must be formed to ensure that $\bar{\mathcal{G}}$ remains strongly connected. Actually, there are a number of interesting ways to attack the social dominance of $v_1$, and we list some of the most important/fundamental methods. For each topology variation we list below, we provide an example in the Figures~\ref{fig:single_attack}-\ref{fig:leader_group}.
\begin{topology}[Single Attack]\label{top:single}
Suppose that $n\geq 4$. Suppose further that $\mathcal{G}$ has star topology, with $v_1$ being the centre node, and with $n-2$ subject nodes, $v_i, i = 2, ..., n-1$. A single attacker node $v_n$ attaches to subject node $v_{n-1}$ by forming edges $e_{n-1,n}$ and $e_{n,n-1}$. This forms the modified graph $\bar{\mathcal{G}}$
\end{topology}
\begin{topology}[Coordinated Double Attack]\label{top:two_attack}
Suppose that $n\geq 5$. Suppose further that $\mathcal{G}$ has star topology, with $v_1$ being the centre node, and with $n-3$ subject nodes, $v_i, i = 2, ..., n-2$. Two attacker nodes $v_{n-1}$ and $v_n$ attach to subject node $v_{n-2}$ by forming the set of edges $\{e_{n-2,n-1}, e_{n-1,n-2}, e_{n-2,n}, e_{n,n-2}\}$. This forms $\bar{\mathcal{G}}$.
\end{topology}
\begin{topology}[Uncoordinated Double Attack]\label{top:two_attack_two_sub}
Suppose that $n \geq 5$. Suppose further that $\mathcal{G}$ has star topology, with $v_1$ being the centre node, and with $n-3$ subject nodes, $v_i, i = 2, ..., n-2$. One attacker node $v_{n-1}$ attaches to subject node $v_{n-3}$ with edges $e_{n-3,n-1}, e_{n-1,n-3}$. A second attacker node $v_{n}$ attaches to subject node $v_{n-2}$ with edges $e_{n-2,n}, e_{n,n-2}$. This forms $\bar{\mathcal{G}}$.
\end{topology}
\begin{topology}[Two Dissenting Subjects]\label{top:no_attack}
Suppose that $n \geq 4$. Suppose further that $\mathcal{G}$ has star topology, with $v_1$ being the centre node, and with $n-1$ subject nodes, $v_i, i = 2, ..., n$. There are no attacker nodes. Subject nodes $v_{n-1}$ and $v_n$ form edges $e_{n,n-1}, e_{n-1,n}$, forming $\bar{\mathcal{G}}$.
\end{topology}
The following topology variation is motivated by the concept of a \emph{leadership group} where two leaders exist, and seek to maintain their collective social dominance.
\begin{topology}[Leadership group]\label{top:leader_group}
Suppose that $\mathcal{G}_1$ and $\mathcal{G}_2$ respectively have $n \geq 3$ and $m\geq 3$ nodes, with node set $\mathcal{V}_1 = \{1, ..., n\}$ and $\mathcal{V}_2 = \{ n+1, ..., n+m\}$ respectively. Both $\mathcal{G}_1$ and $\mathcal{G}_2$ have star topology; the centre nodes for $\mathcal{G}_1$ and $\mathcal{G}_2$ are $v_1$ and $v_{n+1}$ respectively. Let $\bar{\mathcal{G}}$ be the graph formed by merging $\mathcal{G}_1$ and $\mathcal{G}_2$ by insertion of the edges $e_{1,n+1}$ and $e_{n+1,1}$. Nodes $v_1, v_{n+1}$ form a \emph{leadership group} with subjects $v_2, ..., v_n, v_{n+2}, ..., v_{n+m}$.
\end{topology}
In the next section, we investigate the above topological variations of the star graph. Note that Topology Variations~\ref{top:single}-\ref{top:leader_group} have modified graphs $\bar{\mathcal{G}}$ which \emph{do not have star topology}. From the properties of $\vect{F}(\vect{x}(s))$ established in \cite{jia2015opinion_SIAM} and detailed in Lemma~\ref{lem:star} and Theorem~\ref{thm:DFmain}, it immediately follows that $x_1^* < 1$ for all Topology Variations. In other words, $v_1$ is no longer the autocrat but if the perturbation from star topology (caused by the new edges) is small, one expects that $v_1$ remains \emph{socially dominant}. What will we show is that if the interpersonal weights associated with these new edges exceed a given threshold, the socially dominant node changes from $v_1$ to some other node. It is worth emphasising at this stage that, in Variations~\ref{top:single}-\ref{top:two_attack_two_sub}, it is useless for an attacker node $v_n$ to attach to the centre node $v_1$ instead of a subject node; the topology remains a strongly connected star, and there is no change in the autocratic nature of $v_1$'s social dominance.
Note that when new edges are introduced, we assume each individual $i$ adjusts its weights $c_{ij}$ to ensure that the new $\mat{C}$ is row-stochastic. Take Topology Variation~\ref{top:leader_group} as an example. Separately, the relative interaction matrix $\mat{C}_1$ (respectively $\mat{C}_2$) associated with $\mathcal{G}_1$ (respectively $\mathcal{G}_2$) is assumed to be row-stochastic. The relative interaction matrix $\bar{\mat C}$ associated with $\bar{\mathcal{G}}$ is also implicitly assumed to be row-stochastic with zero diagonal. That is, we assume that after the addition of edges $e_{1,n+1}$ and $e_{n+1,1}$, adjustments are made to the original weights $c_{1,j}$ and $c_{n+1,k}$ to ensure $\bar{\mat{C}}$ is row-stochastic.
\begin{remark}[Ordering of Social Power]
Although Theorem~\ref{thm:DFmain} states that $\vect{x}^*$ is uniquely determined by $\vect{\gamma}^\top$, there are no results available which allow one to analytically compute the \emph{value} of $\vect{x}^*$ given $\vect{\gamma}^\top$. What is available is Statement~\ref{prty:thm_DFmain02} of Theorem~\ref{thm:DFmain}, which states that the ordering $\vect{x}_i^*$ is consistent with the ordering of $\gamma_i$. In this paper, we are therefore interested in the ordering of individual social power, as opposed to the precise values of social power. This is reflected in Definition~\ref{def:leader}.
\end{remark}
\begin{figure*}
\begin{minipage}{0.32\linewidth}
\begin{center}
\resizebox{0.75\columnwidth}{!}{\input{star_top.pdf_tex}}
\caption{Star Topology with red Centre Node $v_1$, and blue subject nodes, $n = 7$.}
\label{fig:star_top}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\begin{center}
\resizebox{0.9\columnwidth}{!}{\input{topvar1.pdf_tex}}
\caption{Topology Variation~\ref{top:single} (Single Attacker) with $n = 8$, attacker nodes are green. }
\label{fig:single_attack}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\begin{center}
\resizebox{1.1\columnwidth}{!}{\input{topvar2.pdf_tex}}
\caption{Topology Variation~\ref{top:two_attack} (Coordinated Double Attacker) with $n = 9$, attacker nodes are green.}
\label{fig:two_attack}
\end{center}
\end{minipage}
\end{figure*}
\begin{figure*}
\begin{minipage}{0.32\linewidth}
\begin{center}
\resizebox{1\columnwidth}{!}{\input{topvar3.pdf_tex}}
\caption{Topology Variation~\ref{top:two_attack_two_sub} (Uncoordinated Double Attacker) with $n = 9$.}
\label{fig:two_attack_two_sub}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\begin{center}
\resizebox{1\columnwidth}{!}{\input{topvar4.pdf_tex}}
\caption{Topology Variation~\ref{top:no_attack} (Two Dissenting Subjects) with $n = 7$.}
\label{fig:no_attack}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.32\linewidth}
\begin{center}
\resizebox{1\columnwidth}{!}{\input{topvar5.pdf_tex}}
\caption{Topology Variation~\ref{top:leader_group} (Leadership Group) with $n = m = 4$.}
\label{fig:leader_group}
\end{center}
\end{minipage}
\end{figure*}
\section{Main Results}\label{sec:main_result}
In order to maintain the flow of this paper, and to place focus on discussion of the \emph{social connotations} of each result, we place all proofs in the appendix. We firstly present theorems and corollaries for each topological variation, and then discuss their social implications.
\subsection{Topology Variation~\ref{top:single}: Single Attacker}
Now, firstly consider Topology Variation~\ref{top:single}. The relative interaction matrix $\mat C(\beta)$ associated with $\bar{\mathcal{G}}$ can be expressed as
\begin{equation}\label{eq:C_singleattack}
\mat C(\beta) =
\begin{bmatrix}
0 & c_{12} & c_{13} & \hdots & c_{1,n-1} & 0 \\
1 & 0 & 0 & \hdots & 0 & 0 \\
1 & 0 & 0 & \hdots & 0 & 0 \\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
1-\beta & 0 & 0 & \hdots & 0 & \beta \\
0 & 0 & 0 & \hdots & 1 & 0
\end{bmatrix}
\end{equation}
where $\beta = c_{n-1,n} \in (0,1)$ is the influence exerted by the attacker node $v_n$ on subject node $v_{n-1}$. The following theorem details how the social power of each individual changes as $\beta$ changes.
\begin{theorem}[Single Attack]\label{thm:single}
For a social network with Topology Variation~\ref{top:single}, with initial conditions $\vect{x}(0) \in \wt{\Delta}_n$, and described by the DeGroot-Friedkin model, the following statements are true:
\begin{enumerate}[label=(\roman*)]
\item \label{prty:thm_single01} For all values of $\beta \in (0,1)$, there holds $x_i^* < x_1^*$, for all $i \neq 1, n-1, n$ and $x_n^* < x_{n-1}^*$
\item \label{prty:thm_single02} There holds 1) $x_1^* > x_i^*, \forall\, i \neq 1$ if and only if $\beta < 1- c_{1,n-1} = \sum_{i = 2}^{n-2} c_{1,i}$,
or 2) $x_{n-1}^* > x_i^*, \forall\, i \neq n-1$
if and only if $\beta > 1- c_{1,n-1}$. There holds $x_1^* = x_{n-1}^* > x_i^*, \forall\, i \neq 1,n-1$ if and only if $\beta_1 = 1- c_{1,n-1}$
\item \label{prty:thm_single03} There holds $x_n^* > x_1^*$ if and only if $\beta > 1/(1+c_{1,n-1})$.
\end{enumerate}
\end{theorem}
\begin{corollary}[Generalised Placement of Single Attacking Node]\label{cor:single}
Suppose that instead of attaching to subject node $v_{n-1}$, attacker node $v_n$ can attach to any subject node $v_i, i \in \{2, ..., n-1\}$ by forming edges $e_{n,i}, e_{i,n}$. The lower bound on $\beta = c_{n,i}$ required to have $x_{n-1}^* > x_1^*$ is minimised if $v_n$ attaches to $v_k$ where $k = \text{argmax}_{j\in \{2, ..., n-1\}} \; c_{1,j}$.
\end{corollary}
The above mathematical results can be interpreted in the following social context. From Statement~\ref{prty:thm_single01}, we conclude that individuals $2$ to $n-2$, i.e. subject nodes $v_i, i\in \{2, .., n-2\}$ will never have greater social power at equilibrium, $x_i^*$ than the centre individual $v_1$ with $x_1^*$, regardless of how $\beta$ changes. In addition, the attacker node will never have greater social power than the subject node $v_{n-1}$ which it is attached to.
Recall from Remark~\ref{rem:trust} that $c_{ij}$ can be considered the trust level accorded to individual $j$ by individual $i$. Then according to Statement~\ref{prty:thm_single02}, centre individual $v_1$ \emph{remains the socially dominant individual in the social network} only if subject $v_{n-1}$ trusts attacker $v_n$ less than the total sum of trust accorded to subjects $v_i, i\in \{2, ..., n-2\}$ by centre node $v_1$. \emph{In order to become socially dominant, and to undermine the authority of the centre node $v_1$, individual $v_{n-1}$ must trust the attacker $v_n$}.
Lastly, Statement~\ref{prty:thm_single03} reveals that the attacker can also obtain social power greater the centre individual $v_1$ if $\beta$, i.e. the trust accorded to the attacker $v_n$ by subject node $v_{n-1}$, is sufficiently large. We therefore conclude that the leadership/social dominance within a social network with Topology Variation~\ref{top:single} can be shifted from the original leader (centre node) to a subject via introduction of a single attacker and strengthening of the newly formed interpersonal relationships.
Corollary~\ref{cor:single} delivers an intuitive and powerful, socially relevant result. It states that the single attacker node $v_n$ should seek to form an interpersonal relationship with \emph{the subject node $v_k$ that centre node $v_1$ trusts the most}. This will minimise the required amount of trust subject $v_k$ accords attacker $v_n$ before centre node $v_1$ loses social dominance.
\subsection{Topology Variation~\ref{top:two_attack}: Coordinated Double Attack}
Consider now Topology Variation~\ref{top:two_attack}. Firstly, define $\beta_{1} = c_{n-2,n-1} \in (0,1)$ and $\beta_{2} = c_{n-2,n} = (0,1)$ as the two adjustable interpersonal weights. Note that because $\mat{C}$ is assumed to be row-stochastic, it is implied that $\beta_1 + \beta_2 + c_{n-2,1} = 1$ which in turn implies $\beta_1 + \beta_2 < 1$ because $c_{n-2,1} > 0$. We omit displaying the exact form of $\mat{C}(\beta_1, \beta_2)$ due to spatial limitations.
\begin{theorem}[Coordinated Double Attack]\label{thm:two_attack}
For a social network with Topology Variation~\ref{top:two_attack}, with initial conditions $\vect{x}(0) \in \wt{\Delta}_n$, and described by the DeGroot-Friedkin model, the following statements are true:
\begin{enumerate}[label=(\roman*)]
\item \label{prty:thm_two01} For all $\beta_1, \beta_2 \in (0,1)$, there holds $x_i^* < x_1^*$ for all $i \neq 1, n-2, n-1, n$, and $x_n^*, x_{n-1}^* < x_{n-2}^*$.
\item \label{prty:thm_two02} There holds 1) $x_1^* > x_i^*, \forall, i \neq 1$ if and only if $\beta_1+\beta_2 < 1- c_{1,n-2} = \sum_{i = 2}^{n-3} c_{1,i}$, or 2) $x_{n-2}^* > x_i^*, \forall, i \neq n-2$
if and only if $\beta_1+\beta_2 > 1- c_{1,n-2}$. There holds $x_1^* = x_{n-2}^*$ if and only if $\beta_1+\beta_2 = 1- c_{1,n-2}$.
\item \label{prty:thm_two03} There holds $x_n^* > x_1^*$ (respectively $x_{n-1}^* > x_1^*$) if and only if $\beta_2 > (1-\beta_1)/(1+c_{1,n-2})$ (respectively $\beta_1 > (1-\beta_2)/(1+c_{1,n-2})$ ).
\item \label{prty:thm_two04} There holds $x_{n-1}^* < x_{n}^*$ or $x_{n-1}^* > x_{n}^*$ if and only if $\beta_1 < \beta_2$ or $\beta_1 > \beta_2$ respectively. If $\beta_1 = \beta_2$, then $x_{n-1}^* = x_{n}^*$.
\end{enumerate}
\end{theorem}
\begin{corollary}[Generalised Placement of Coordinated Double Attack]\label{cor:two_attack}
Suppose that instead of attaching to subject node $v_{n-2}$, attacker nodes $v_{n-1}, v_{n}$ can attach to any subject node $v_i, i \in \{2, ..., n-2\}$ by forming the set of edges $\{ e_{n-1,i}, e_{i,n-1}, e_{n,i}, e_{i,n} \}$. The lower bound on $\beta_1+ \beta_2 = c_{n-1,i} + c_{n,i}$ required to have $x_{n-2}^* > x_1^*$ is minimised if $v_{n-1}$ and $v_n$ attach to $v_k$ where $k = \text{argmax}_{j\in \{2, ..., n-2\}} \; c_{1,j}$.
\end{corollary}
Due to spatial limitations, we discuss social implications of Theorem~\ref{thm:two_attack} only if the conclusions differ significantly from the discussion in the previous subsection.
The key result is Statement~\ref{prty:thm_two02}, which indicates that the \emph{combined trust} given to attackers $v_{n-1}$ and $v_n$ by subject node $v_{n-2}$ must exceed the \emph{combined trust} given to subjects $v_2, ..., v_{n-3}$ by centre node $v_1$, in order for centre node $v_1$ to lose social dominance (and thus subject $v_{n-2}$ becomes the socially dominant individual). It is most interesting to note that \emph{it is only the sum} of the trust/influence $\beta_1 + \beta_2$ that is relevant, and \emph{there is no requirement on the individual magnitudes of $\beta_1, \beta_2$}.
Regarding Statement~\ref{prty:thm_two03}, we observe that the inequality, which if satisfied ensures that attacker $v_n$ has social power greater than centre $v_1$, is a function of $\beta_1, \beta_2$ and $c_{1,n-2}$. As detailed in the proof in Appendix~\ref{app:pf_twoattack}, there always exists a $\beta_1, \beta_2$ satisfying $\beta_1 + \beta_2 < 1$ which ensures both attacker nodes $v_{n-1}, v_{n}$ have social power greater than the centre $v_1$.
\subsection{Topology Variation~\ref{top:two_attack_two_sub}: Uncoordinated Double Attack}
Define $\beta_1 = c_{n-3,n-1} \in (0,1)$ and $\beta_2 = c_{n-2,n} \in (0,1)$.
\begin{theorem}\label{thm:twotwo}
For a social network with Topology Variation~\ref{top:two_attack_two_sub}, with initial conditions $\vect{x}(0) \in \wt{\Delta}_n$, and described by the DeGroot-Friedkin model, the following statements are true:
\begin{enumerate}[label=(\roman*)]
\item \label{prty:thm_twotwo01} For all values of $\beta_1,\beta_2 \in (0,1)$, there holds $x_i^* < x_1^*$ for all $i \in\{2,\ldots,n-4\}$, and $x_{n-1}^* < x_{n-3}^*$ and $x_{n}^* < x_{n-2}^*$.
\item \label{prty:thm_twotwo02} There holds $x_1^* > x_i^*$ for all $i \neq 1$ if and only if $\beta_1 < 1- c_{1,n-3}$
and $\beta_2 < 1-c_{1,n-2}$. If $\beta_1 > 1- c_{1,n-3}$ (respectively $\beta_2 > 1 - c_{1,n-2}$), then $x_{n-3}^* > x_1^*$ (respectively $x_{n-2}^* > x_1^*$).
\item \label{prty:thm_twotwo03} For $i\in\{1,2\}$, there holds $x_{n-2+i}^* > x_1^*$ if and only if $\beta_i > 1/(1+c_{1,n-4+i})$.
\item \label{prty:thm_twotwo04} There holds $x_{n-3}^* > x_{n-2}^*$ if and only if $\frac{1-\beta_2}{1-\beta_1}>\frac{c_{1,n-2} }{c_{1,n-3}}$. Equivalently, $x_{n-3}^* > x_{n-2}^*$ if and only if $\frac{c_{1,n-3}}{c_{n-3,1}}>\frac{c_{1,n-2} }{c_{n-2,1}}$
\end{enumerate}
\end{theorem}
The most interesting conclusion drawn from Theorem~\ref{thm:twotwo} is when we compare to Theorem~\ref{thm:two_attack} which concerns Topology Variation~\ref{top:two_attack}. With Topology Variation~\ref{top:two_attack}, for the centre individual $v_1$ to lose its social dominance we require \emph{the sum of the trust values $\beta_1 + \beta_2$} to exceed a lower bound, and there are no separate lower bounding inequalities for $\beta_1$ or $\beta_2$. With Topology Variation~\ref{top:two_attack_two_sub}, centre individual $v_1$ loses social dominance if and only if either $\beta_1$ or $\beta_2$ exceed their respective lower bounding inequalities. Importantly, these two lower bounding inequalities \emph{are independent of each other}. \emph{This clearly points to the fact that a \textbf{coordinated attack} on the social dominance of the centre node is more desirable, an idea which is socially intuitive}.
From Statement \ref{prty:thm_twotwo03}, both attacker nodes have larger social power than the centre node if and only if
$\beta_1 > 1/(1+c_{1,n-3})$ and $\beta_2 > 1/(1+c_{1,n-2})$,
which implies that $\beta_1+\beta_2 > 1/(1+c_{1,n-3})+1/(1+c_{1,n-2})$. From Statement \ref{prty:thm_two03} Theorem \ref{thm:two_attack}, with Topology Variation \ref{top:two_attack}, both attacker nodes have larger social power than the centre node if and only if $\beta_2 > (1-\beta_1)/(1+c_{1,n-2})$ and $\beta_1 > (1-\beta_2)/(1+c_{1,n-2})$, which implies that $\beta_1+\beta_2 > 2/(2+c_{1,n-2})$. Since both $1+c_{1,n-2}$ and $1+c_{1,n-3}$ are smaller than $2+c_{1,n-2}$, it follows that
$1/(1+c_{1,n-3})+1/(1+c_{1,n-2})> 2/(2+c_{1,n-2})$, which implies that \emph{a \textbf{coordinated attack} on the social dominance of the centre node is also more efficient for the attackers}.
\subsection{Topology Variation~\ref{top:no_attack}: Two Dissenting Subjects}
Topology Variation~\ref{top:no_attack} is different to the ones studied above in the sense that there are no attacker nodes. Instead, one can consider this variation as one where two subjects form a relationship in \emph{dissent} from the leader. Firstly, let $\beta_1 = c_{n-1,n} \in (0,1)$ and $\beta_2 = c_{n,n-1} \in (0,1)$. Analysis yields the following result.
\begin{theorem}[Two Dissenting Subjects]\label{thm:no_attack}
For a social network with Topology Variation~\ref{top:no_attack}, with initial conditions $\vect{x}(s) \in \wt{\Delta}_n$, and described by the DeGroot-Friedkin model, the following statements are true:
\begin{enumerate}[label=(\roman*)]
\item \label{prty:thm_no01} For all $\beta_1, \beta_2 \in (0,1)$, there holds $x_i^* < x_1^*$ for all $i \neq 1, n-1, n$.
\item \label{prty:thm_no02} There holds $x_n^* > x_1^*$ if and only if $\beta_1 > (1-c_{1,n})/(c_{1,n-1}+\beta_2)$ with $\beta_1 \in (0,1)$. There exists such a $\beta_1 \in (0,1)$ only if $\beta_2 > \sum_{i = 2}^{n-2} c_{1,i}$.
\item \label{prty:thm_no04} There holds $x_{n-1}^* > x_1^*$ if and only if $\beta_2 > (1-c_{1,n-1})/(c_{1,n}+\beta_1)$ with $\beta_2 \in (0,1)$. There exists such a $\beta_2 \in (0,1)$ only if $\beta_1 > \sum_{i = 2}^{n-2} c_{1,i}$.
\item \label{prty:thm_no06} There holds $x_n^* < x_{n-1}^*$ if and only if $\beta_2 > \beta_1 c_{1,n} + c_{1,n-1} (c_{1,n} - 1)$ or equivalently $\beta_1 < (\beta_2 + c_{n-1} (1 - c_{1,n}))$
\end{enumerate}
\end{theorem}
Note that the inequality in statement~\ref{prty:thm_no02} can be rewritten as $\beta_2 > (1-c_{1,n} - \beta_1 c_{1,n-1} )/\beta_1$ with $\beta_2 \in (0,1)$ which is satisfiable only if $\beta_1 > (1-c_{1,n})/(1+c_{1,n-1})$. Similarly, the inequality in statement~\ref{prty:thm_no04} is equivalent to $\beta_1 > (1- c_{1,n-1} - \beta_2 c_{1,n} )/\beta_2$ with $\beta_1 \in (0,1)$ which is satisfiable only if $\beta_2 > (1-c{1,n-1})/(1+c_{1,n})$.
We now interpret Statement~\ref{prty:thm_no02}, which we believe is the key result of the theorem. A similar conclusion can be drawn for Statement~\ref{prty:thm_no04} but we omit this due to spatial limitations. In order to make centre node $v_1$ lose social dominance, the dissent subject nodes $v_{n-1}$ and $v_{n}$ \emph{must adopt a cooperative strategy}. From their definitions, we can interpret $\beta_1$ as the trust given by $v_{n-1}$ to $v_n$ while $\beta_2$ is the trust given by $v_{n}$ to $v_{n-1}$. \emph{A necessary condition} for individual $v_n$ to have social power greater than centre node $v_1$ is that $\beta_2 > \sum_{i = 2}^{n-2} c_{1,i}$. This means that not only must $v_{n-1}$ trust $v_n$ sufficiently (as given by the inequality $\beta_1 > (1-c_{1,n})/(c_{1,n-1}+\beta_2)$), but individual $v_n$ must \emph{reciprocate} by ensuring that it trusts $v_{n-1}$ sufficiently ($\beta_1 > \sum_{i = 2}^{n-2} c_{1,i}$). Unless the two dissenting nodes build a cooperative and sufficiently strong bilateral relationship, centre node $v_1$ will remain socially dominant.
\subsection{Topology Variation~\ref{top:leader_group}: Leadership Group}
With $\beta_1 = c_{1,n+1} \in (0,1)$ and $\beta_2 = c_{n+1,1} \in (0,1)$, the following result is obtained
\begin{theorem}\label{thm:leader_group}
For a social network with Topology Variation~\ref{top:leader_group}, with initial conditions $\vect{x}(s) \in \wt{\Delta}_{n+m}$, and described by the DeGroot-Friedkin model, the following statements are true:
\begin{enumerate}[label=(\roman*)]
\item \label{prty:thm_leader01} For all $\beta_1 \in (0,1)$ and for all $\beta_2 \in (0,1)$ there holds $x_i^* < x_1^*$ and $x_k^* < x_{n+1}^*$ for $i \in \{2, ..., n\}$ and $k \in \{n+2, ..., n+m\}$.
\item \label{prty:thm_leader02} There holds $x_1^* < x_{n+1}^*$ or $x_1^* > x_{n+1}^*$ if and only if $\beta_2 < \beta_1$ or $\beta_2 > \beta_1$ respectively. If $\beta_1 = \beta_2$, then $x_1^* = x_{n+1}^*$.
\item \label{prty:thm_leader03} For $k \in \{n+2, ..., n+m\}$ there holds $x_k^* > x_1^*$ if and only if $c_{n+1,k}(\beta_1/\beta_2) > 1$. For $i \in \{ 2, ..., n\}$, holds $x_i^* > x_{n+1}^*$ if and only if $c_{1,i}(\beta_2/\beta_1) > 1$.
\end{enumerate}
\end{theorem}
Statement~\ref{prty:thm_leader02} shows that the ratio of $\beta_1/\beta_2$ determines whether centre node $v_1$ or centre node $v_{n+1}$ is socially dominant. Statement~\ref{prty:thm_leader03} delivers a surprising and interesting result on how leaders can \emph{cooperatively} protect themselves and maintain collective social dominance. Let $i \in \{2, ..., n\}$ and $k \in \{n+1, ..., n+m\}$. Consider from centre individual $v_1$'s point of view. While $x_i^* < x_1^*$ is guaranteed, in order to ensure that $v_1$ has greater social power than subject $v_k$ (i.e. subjects of centre individual $v_{n+1}$), individual $v_1$ must ensure that $c_{n+1,k}(\beta_1/\beta_2) < 1$. This inequality always holds, regardless of the value of $c_{n+1,k} < 1$, if $\beta_1 = \beta_2$. I.e. if the trust level $v_1$ accords to $v_{n+1}$ is equal to the trust level $v_{n+1}$ accords to $v_1$, regardless of the magnitude of $\beta_1 = \beta_2$, $v_1$ has greater social power than \emph{all} subject nodes \emph{including the subjects of $v_{n+1}$}. It can appear to be surprising because this holds even if $c_{n+1,k} >> \beta_1, \beta_2$. Yet such a result is intuitive if we consider $\beta_1/\beta_2$ as the ratio of the trust $v_1$ places on $v_{n+1}$ (and indirectly the trust $v_1$ places on subject $v_{k}$) versus the trust $v_{n+1}$ places on $v_1$ (and indirectly the trust $v_k$ places on $v_1$).
\section{Simulations}\label{sec:simulations}
In this section, we provide 2 short simulations to highlight some of our most interesting results. We do not provide comprehensive simulations for each Topology Variation due to spatial limitations.
Firstly, we simulate Topology Variation~\ref{top:single} as it is the fundamental strategy, with $n = 8$. The top row of the matrix $\mat{C}$ is given by $[0,\, 0.15,\, 0.15,\, 0.2,\, 0.05,\, 0.15,\, 0.3,\, 0]$. Figure~\ref{fig:SimTopVar1} shows the social power at equilibrium $x_i^*$, for selected individuals, as a function of $\beta = c_{78}$. Centre $v_1$ loses social dominance when $\beta > 0.7$ as stated in Theorem~\ref{thm:single}.
Next, we simulate Topology Variation~\ref{top:no_attack} to show the need for cooperation between two dissenting individuals in order to displace the centre node. The top row of $\mat{C}$ is $[0,\, 0.1,\, 0.1,\, 0.2,\, 0.05,\, 0.05,\, 0.2,\, 0.3]$. Figure~\ref{fig:SimTopVar4_fail} shows the social power at equilibrium $x_i^*$, for selected individuals as a function of $\beta_1 = c_{78}$ when $\beta_2 = 0.49$ (i.e. when $\beta_2 < \sum_{i = 2}^{n-2} c_{1,i}$). In accordance with Theorem~\ref{thm:no_attack}, Statement~\ref{prty:thm_no02}, dissent subject $v_8$ never achieves social power greater than centre $v_1$ because there does not exist a $\beta_1 \in (0,1)$ satisfying the required inequality. Figure~\ref{fig:SimTopVar4_win} shows the same simulation scenario but now with $\beta_2 = 0.55 > \sum_{i = 2}^{n-2} c_{1,i}$. In accordance with Statement~\ref{prty:thm_no02} of Theorem~\ref{thm:no_attack}, $x_8^* > x_1^*$ when $\beta_1 > 0.93$.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{Figures/SimTopVar1.pdf}
\caption{Simulation of Topology Variation~\ref{top:single} with $n = 8$.}
\label{fig:SimTopVar1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{Figures/SimTopVar4_fail.pdf}
\caption{Simulation of Topology Variation~\ref{top:no_attack} with $n = 8$, when $\beta_2 < \sum_{i = 2}^{n-2} c_{1,i}$.}
\label{fig:SimTopVar4_fail}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{Figures/SimTopVar4_win.pdf}
\caption{Simulation of Topology Variation~\ref{top:no_attack} with $n = 8$, when $\beta_2 > \sum_{i = 2}^{n-2} c_{1,i}$.}
\label{fig:SimTopVar4_win}
\end{center}
\end{figure}
\section{Conclusions}\label{sec:conclusions}
Social networks with a star topology converge to an autocratic configuration, with the centre individual holding all the social power, as the number of issues discussed tend towards infinity. This paper proposed a number of different strategies, involving introduction of new individuals and/or new interpersonal relationships into the social network, in order to move social dominance from the centre individual to a subject individual. Necessary and sufficient conditions are developed, and interpretation of these conditions showed the strategies are sociologically intuitive. Numerous future directions exist. Firstly, we wish to generalise the results on uncoordinated attack and coordinated attack to arbitrary numbers of attacker nodes. Different leadership groups, and dissent topologies will also be explored. We also wish to investigate whether such straightforward strategies exist for more general topologies, and lastly we shall study strategies concerning social power for a subgroup of individuals.
|
\section{Coupling to the multimode radiation field}
The atomic system is coupled to the quantized multimode radiation field. Within the dipole approximation, the Hamiltonian that describes the dynamics of the atomic ensemble and the radiation field reads $H_\mathrm{af}=\sum_{j=1}^N\hbar\omega_\mathrm{a}\mathbf{b}_j^\dagger\cdot\mathbf{b}_j +\sum_{\mathbf{q}\lambda}\hbar\omega_{\mathbf{q}}a_{\mathbf{q}\lambda}^\dagger a_{\mathbf{q}\lambda}+i\hbar\sum_{j=1}^N\sum_{\mathbf{q}\lambda}
\mathbf{g}_{\mathbf{q}\lambda}\cdot\left(a_{\mathbf{q}\lambda}^\dagger
\mathbf{s}_j e^{-i\mathbf{q}\cdot\mathbf{r}_j}-
\mathbf{s}_ja_{\mathbf{q}\lambda}e^{i\mathbf{q}\cdot\mathbf{r}_j}\right)$. The first term represents the atomic energy, with $\hbar\omega_\mathrm{a}=2\pi\hbar c/\lambda$ being the energy difference between the ground ($\left|g\right>$) and the three degenerate $J=1$ manifold states ($\left|-\right>$, $\left|0\right>$ and $\left|+\right>$ for $m_J=-1,0$ and $+1$, respectively) and the vector transition operator for the $j$-th atom (located at $\mathbf{r}_j$) being defined as $\mathbf{b}_j=\left(\left|g\right>_j\!\left<+\right|,\left|g\right>_j\!\left<-\right|,\left|g\right>_j\!\left<0\right|\right)$. The second term of the Hamiltonian represents the radiation field, and here $\hbar\omega_{\mathbf{q}}$ is the energy of a photon with momentum $\mathbf{q}$ and polarization $\lambda$ and $a_{\mathbf{q}\lambda}$ is the annihilation operator of such a photon ($[a_{\mathbf{q}\lambda},a_{\mathbf{q}'\lambda'}^\dag] =\delta_{\mathbf{q}\mathbf{q}'}\delta_{\lambda\lambda'}$). The last term represents the coupling between the two systems, with $\mathbf{s}_j=\mathbf{b}_j^\dag+\mathbf{b}_j$ and the coupling coefficient being given by $\mathbf{g}_{\mathbf{q}\lambda}=d\sqrt{\frac{\omega_{\mathbf{q}}}{2\epsilon_0\hbar
V}} \hat{e}_{\mathbf{q}\lambda}$, where $V$ is the quantization volume, $d$ the modulus of the transition dipole moment between the ground and the three degenerate excited states, and $\hat{e}_{\mathbf{q}\lambda}$ the unit polarization vector of the photon ($\mathbf{q}\cdot\hat{e}_{\mathbf{q}\lambda}=0$).
Within the dipole and Born-Markov approximations, and tracing out the environment, one obtains the quantum master equation for the atomic density matrix $\rho$, $\dot{\rho}=-\frac{i}{\hbar}\left[H,\rho\right]+{\cal D}(\rho)$, which describes the dynamics of the atomic degrees of freedom \cite{Lehmberg70,Agarwal70,James93}. The many-body Hamiltonian that describes the coherent evolution of the system reads $H=\hbar\sum_{j\neq l}\mathbf{b}^\dag_{j}\cdot\overline{W}_{jl}\cdot\mathbf{b}_{l}$, characterized by the coefficient matrix
\begin{equation}\label{eq:V}
\overline{W}_{jl}=\left(\begin{array}{ccc}
A_{jl} & B_{jl}e^{-2i\phi_{jl}} & 0\\
B_{jl}e^{2i\phi_{jl}} & A_{jl} & 0\\
0 & 0 & C_{jl}
\end{array}\right)
\end{equation}
with $A_{jl}$ and $B_{jl}$ being defined in the main text [Eqs. (5) and (6), respectively] and
\begin{eqnarray}
C_{jl}&=&\frac{3\Gamma}{2}\left[-\frac{\sin{\kappa_{jl}}}{\kappa_{jl}^2} -\frac{\cos{\kappa_{jl}}}{\kappa_{jl}^3}\right]\label{eq:CC}
\end{eqnarray}
for $j\neq l$. The dissipator ${\cal D}(\rho)$ reads ${\cal D}(\rho) = \sum_{jl}\mathbf{b}_j\cdot\overline{\gamma}_{jl} \cdot\rho\mathbf{b}_l^\dag-\frac{1}{2}\left\{\mathbf{b}_j^\dag\cdot\overline{\gamma}_{jl} \cdot\mathbf{b}_l,\rho\right\}$, where
\begin{equation}\label{eq:diss1}
\overline{\gamma}_{jl}=\left(\begin{array}{ccc}
A'_{jl} & B'_{jl}e^{-2i\phi_{jl}} & 0\\
B'_{jl}e^{2i\phi_{jl}} & A'_{jl} & 0\\
0 & 0 & C'_{jl}
\end{array}\right),
\end{equation}
with $A'_{jl}$ and $B'_{jl}$ being defined in the main text [Eqs. (7) and (8), respectively] and
\begin{eqnarray}
C'_{jl}&=&3\Gamma\left[-\frac{\cos{\kappa_{jl}}}{\kappa_{jl}^2} +\frac{\sin{\kappa_{jl}}}{\kappa_{jl}^3}\right]\label{eq:CCp}
\end{eqnarray}
It is clear from the form of Eqs. (\ref{eq:V}) and (\ref{eq:diss1}) that the dynamics of the internal level $m_J=0$ is disconnected from the other two. We will consider thus only the dynamics on the subspace formed by the states $m_J=\pm1$.
\section{Ewald summation}
\subsection{General notions}
We briefly review the Ewald summation technique \cite{Bonsall77,Kantorovich04} introduced originally to cope with the evaluation of the electrostatic potential energy of a gas of electrons interacting through Coulomb force. Here we wish to evaluate the discrete sum (\ref{eq:Wk}). Considering two-dimensional lattices with unit cell defined by two lattice vectors $(\mathbf{a}_1, \mathbf{a}_2)$, a vector ${\mathbf r}$ connecting two atoms $\alpha$ and $\beta$ can be written as
\begin{equation}
{\mathbf r} = {\pmb \tau}_{\alpha \beta} + {\bf T},
\label{eq:r}
\end{equation}
where ${\bf T} = (m \mathbf{a}_1, n \mathbf{a}_2),\; m,n \in \mathbb{Z}$ is the lattice translation vector, ${\pmb \tau}_{\alpha \beta} = {\pmb \tau}_{\beta} - {\pmb \tau}_{\alpha}$ and ${\pmb \tau}_\alpha$ denotes the position of the $\alpha$-th atom in the unit cell. We also define the reciprocal lattice vectors in the usual way as ${\mathbf G} = (m \mathbf{b}_1, n \mathbf{b}_2),\; m,n \in \mathbb{Z}$, where $\mathbf{b}_j,\;j=1,2$ are the primitive reciprocal lattice vectors.
Let us consider an algebraically decaying function of the atomic separation $v({\mathbf r})$ and say we want to evaluate the sum
\begin{equation}
V({\bf k}) = \sum_{{\mathbf r}\neq 0} {\rm e}^{i {\bf k r}} v({\bf r})
\label{eq:Vk direct}
\end{equation}
The Ewald's trick consists of separating the contributions of $v$ to the sum to the short (S) and long (L) ranged part by introducing a regulating function $\xi$ through the identity
\begin{equation}
1 = {}^S\xi(r) + {}^L\xi(r),
\end{equation}
so that
\begin{equation}
V({\bf k}) = {}^S V({\bf k}) + {}^L V({\bf k}) - {}^{\rm self} V({\bf k}),
\label{eq:Vk pot}
\end{equation}
where
\begin{subequations}
\begin{align}
{}^S V({\bf k}) &= \sum_{\mathbf r\neq 0} {\rm e}^{i {\bf k r}} {}^S\xi(r) v({\bf r}) \label{eq:FS}\\
{}^L V({\bf k}) &= \sum_{\mathbf r} {\rm e}^{i {\bf k r}} {}^L\xi(r) v({\bf r}) \label{eq:FL}\\
{}^{\rm self} V({\bf k}) &= \lim_{r \to 0} {\rm e}^{i {\bf k r}} {}^L\xi(r) v({\bf r}). \label{eq:Fself}
\end{align}
\end{subequations}
The function $\xi$ is chosen such that ${}^S\xi(r) v({\bf r})$ decays exponentially fast as $r \rightarrow \infty$ and (\ref{eq:Fself}) is finite. Due to the fast decay of ${}^S\xi$, the short-range part can be readily obtained by evaluating the sum in (\ref{eq:FS}) in real space. We are thus left with the evaluation of (\ref{eq:FL}).
We proceed as follows. Using (\ref{eq:r}), (\ref{eq:FL}) becomes
\begin{widetext}
\begin{eqnarray}
{}^L V({\bf k}) &=& \sum_{\alpha,\beta \in {\rm unit\;cell}} \sum_{\bf T} {\rm e}^{i {\bf k}({\pmb \tau}_{\alpha \beta} + {\bf T})} {}^L \xi(|{\pmb \tau}_{\alpha \beta} + {\bf T}|) v({\pmb \tau}_{\alpha \beta} + {\bf T}) \nonumber \\
&=& \sum_{\alpha,\beta \in {\rm unit\;cell}} {\rm e}^{i {\bf k}{\pmb \tau}_{\alpha \beta}} \int {\rm d}^2{\mathfrak r} \sum_{\bf T} \delta^{2}({\boldsymbol{\mathfrak r}}-{\bf T}) {\rm e}^{i {\bf k}{\boldsymbol{\mathfrak r}} } {}^L \xi(|{\pmb \tau}_{\alpha \beta} + {\boldsymbol{\mathfrak r}}|) v({\pmb \tau}_{\alpha \beta} + {\boldsymbol{\mathfrak r}}) \nonumber \\
&=& \sum_{\bf G} \sum_{\alpha,\beta \in {\rm unit\;cell}} {\rm e}^{i {\bf k}{\pmb \tau}_{\alpha \beta}} \int {\rm d}^2 {\mathfrak r} {\rm e}^{i {\bf k + G}{\boldsymbol{\mathfrak r}} } {}^L \xi(|{\pmb \tau}_{\alpha \beta} + {\boldsymbol{\mathfrak r}}|) v({\pmb \tau}_{\alpha \beta} + {\boldsymbol{\mathfrak r}}) \nonumber \\
&=& \sum_{\bf G} \sum_{\alpha,\beta \in {\rm unit\;cell}} {\rm e}^{-i {\bf G}{\pmb \tau}_{\alpha \beta}} \int {\rm d}^2 r {\rm e}^{i {\bf \tilde{G}}{\bf r}} {}^L \xi(r) v({\bf r}),
\label{eq:FL transf}
\end{eqnarray}
\end{widetext}
where we in the third line we have used the identity (normalizing the volume of the unit cell to 1)
\begin{equation}
\sum_{\bf T} \delta^{2}({\boldsymbol{\mathfrak r}}-{\bf T}) = \sum_{\bf G} {\rm e}^{i {\bf G}{\bf r}}
\end{equation}
and in the last line we made a substitution of variables ${\bf r} = {\pmb \tau}_{\alpha \beta} + {\boldsymbol{\mathfrak r}}$ and defined $\widetilde{\bf G} = {\bf k + G}$.
Next, due to the form of (\ref{eq:W}),(\ref{eq:A}),(\ref{eq:B}), we consider the following functions $v$
\begin{subequations}
\begin{align}
{}_{1,n}v({\bf r}) &= \frac{\cos \kappa r}{\kappa r} {\rm e}^{i n \phi({\bf r})} \label{eq:1nv}\\
{}_{2,n}v({\bf r}) &= \frac{\sin \kappa r}{(\kappa r)^2} {\rm e}^{i n \phi({\bf r})} \label{eq:2nv}\\
{}_{3,n}v({\bf r}) &= \frac{\cos \kappa r}{(\kappa r)^3} {\rm e}^{i n \phi({\bf r})}, \label{eq:3nv}
\end{align}
\end{subequations}
where $n=0,2$ and $\kappa = 2\pi/\lambda$.
\subsection{Analytical approach to the $1/r$ term: the divergence}
We start with the evaluation of the integral in (\ref{eq:FL transf}) with the function ${}_{1,n}v({\bf r})$
\begin{eqnarray}
{}_{1,n}\mathfrak{V} &\equiv & \int {\rm d}^2 r {\rm e}^{i {\bf \widetilde{G}}{\bf r}} {}^L \xi(r) {}_{1,n}v({\bf r}) \nonumber \\
&=& \kappa^{-1} \int_0^\infty {\rm d}r \left(1-{\rm e}^{-\gamma r} \right) \cos(\kappa r)\!\!\! \int_0^{2\pi}\!\!\! {\rm d} \varphi {\rm e}^{i ({\widetilde{G}}r \cos{(\varphi-\eta)} + n \phi({\bf r})) } \nonumber \\
&=& \kappa^{-1} \int_0^\infty {\rm d}r \left(1-{\rm e}^{-\gamma r} \right) \cos(\kappa r) J_n(\widetilde{G}r),
\end{eqnarray}
where we have chosen ${}^L\xi(r) = 1-{\rm e}^{-\gamma r}$, $\gamma >0$ is an arbitrary control parameter, $\eta$ is the polar angle of $\widetilde{\bf G}$ and $J_n$ is the $n$-th Bessel function of the first kind. Next, we split the integration as
\begin{equation}
{}_{1,n}\mathfrak{V} ={}_{1,n}\mathfrak{V}_0^c+{}_{1,n}\mathfrak{V}_c^\infty,
\end{equation}
where
\begin{equation}
{}_{1,n}\mathfrak{V}_a^b=\kappa^{-1} \int_a^b {\rm d}r \left(1-{\rm e}^{-\gamma r} \right) \cos(\kappa r) J_n(\widetilde{G}r).
\end{equation}
Finally, we substitute the asymptotic expression for the Bessel function (for large $\widetilde{G}r$)
\begin{equation}
J_n(\widetilde{G}r) \approx \sqrt{\frac{2}{\pi \widetilde{G}}} \frac{1}{\sqrt{r}} \cos \left( \widetilde{G} r -\phi_n\right),
\label{eq:Bessel approx}
\end{equation}
where $\phi_n = \frac{\pi}{4}(2 n + 1)$, to ${}_{1,n}\mathfrak{V}_c^\infty$ so that
\begin{equation}
{}_{1,n}\mathfrak{V}_c^\infty \approx I_0^\infty - I_0^c,
\end{equation}
where
\begin{equation}
I_a^b = \frac{1}{\kappa}\sqrt{\frac{2}{\pi \widetilde{G}}} \int_c^d\!\! {\rm d}r \frac{1-{\rm e}^{-\gamma r}}{\sqrt{r}} \cos(\kappa r) \cos(\widetilde{G}r-\phi_n).
\label{eq:Iab}
\end{equation}
Importantly, we note, that only the term $I_0^\infty$ can give rise to the divergence in (\ref{eq:Vk pot}), which we now analyze in detail. Substituting the identity
\begin{equation}
\frac{1}{\sqrt{r}} = \frac{1}{\sqrt{\pi}} \int_0^\infty {\rm d}t t^{-\frac{1}{2}} {\rm e}^{-t r}
\end{equation}
and using
\begin{equation}
\cos{a}\cos{b} = \frac{1}{2} \left[ \cos(a+b) + \cos(a-b) \right]
\end{equation}
we get
\begin{widetext}
\begin{equation}
\kappa \pi \sqrt{\frac{\widetilde{G}}{2}} I_0^\infty = \int_0^\infty {\rm d}t t^{-\frac{1}{2}} \int_0^\infty {\rm d}r \left( 1-{\rm e}^{-\gamma r} \right) {\rm e}^{-t r} \left[ \cos((\widetilde{G} + \kappa)r - \phi_n) + \cos((\widetilde{G} - \kappa)r - \phi_n) \right].
\label{eq:I 0 infty}
\end{equation}
\end{widetext}
This amounts to evaluation of the integral
\begin{eqnarray}
II &=& \int_0^\infty {\rm d}t t^{-\frac{1}{2}} \int_0^\infty {\rm d}r \left( 1-{\rm e}^{-\gamma r} \right) {\rm e}^{-t r} \cos(\mathcal{G} r - \phi_n) \nonumber \\
&=& \frac{{\rm e}^{-i \phi_n}}{2} \int_0^\infty {\rm d}t \frac{1}{\sqrt{t}} \int_0^\infty {\rm d}r \left[ {\rm e}^{-t r} - {\rm e}^{-(t+\gamma) r} \right] {\rm e}^{i \mathcal{G} r} + {\rm c.c.} \nonumber \\
&=& \frac{{\rm e}^{-i \phi_n}}{2} \int_0^\infty {\rm d}t \frac{1}{\sqrt{t}} \left( \frac{1}{t-i \mathcal{G}} - \frac{1}{t+\gamma-i \mathcal{G}}\right) + {\rm c.c.},
\label{eq:II}
\end{eqnarray}
where $\mathcal{G} = \widetilde{G} \pm \kappa$. The final integration in (\ref{eq:II}) can be done by methods of contour integration along the contour depicted in Fig. \ref{fig:contour}, i.e.
\begin{equation}
\oint_C = \int_{C_1} + \int_R + \int_{C_2} + \int_\epsilon.
\label{eq:cont int}
\end{equation}
Using the property of the branch cut, the integration along $C_2$ becomes
\begin{eqnarray}
\int_{C_2} &=& \int_R^\epsilon {\rm d}z \frac{1}{\sqrt{z}} \left( \frac{1}{z-z_0} - \frac{1}{z-z_1} \right) \nonumber \\
&=& \int_R^\epsilon {\rm d}z \frac{1}{{\rm e}^{\frac{1}{2}(\log |z| + i 2 \pi)}} \left( \frac{1}{z-z_0} - \frac{1}{z-z_1} \right) \nonumber \\
&=& \int_R^\epsilon {\rm d}z \frac{1}{-\sqrt{z}} \left( \frac{1}{z-z_0} - \frac{1}{z-z_1} \right) \nonumber \\
&=& \int_\epsilon^R {\rm d}z \frac{1}{\sqrt{z}} \left( \frac{1}{z-z_0} - \frac{1}{z-z_1} \right).
\end{eqnarray}
\begin{center}
\begin{figure}[h!]
\includegraphics[width=8cm]{integration_contour_v2.png}
\caption{integration contour}
\label{fig:contour}
\end{figure}
\end{center}
Omitting the c.c. term, we express the leftmost part of (\ref{eq:II}) as
\begin{eqnarray}
II_- &=& \frac{{\rm e}^{-i \phi_n}}{2} \int_0^\infty\!\!\! {\rm d}z \frac{1}{\sqrt{z}} \left( \frac{1}{z-i \mathcal{G}} - \frac{1}{z+\gamma-i \mathcal{G}}\right)\\
&=& \int_0^\infty {\rm d}z g(z) = \int_0^\infty {\rm d}z \frac{f(z)}{z-z_0} - \frac{f(z)}{z-z_1}.
\end{eqnarray}
It can be immediately seen, that $\int_R \rightarrow 0$ since $g(z) \rightarrow R^{-\frac{3}{2}}$ in the limit of large $|z|$. Next, since $\gamma>0$, assuming $\mathcal{G} \neq 0$ and substituting $z=\epsilon {\rm e}^{i \varphi}$, we get from the estimation lemma
\begin{equation}
\left| \int_\epsilon \right| \leq |g(z)| 2\pi \epsilon \propto 2 \pi \frac{\epsilon}{\sqrt{\epsilon}} \xrightarrow[]{\epsilon \rightarrow 0^+} 0.
\end{equation}
We are thus left with
\begin{eqnarray}
II_- &=& \frac{1}{2} \oint_C = i \pi \left( f(z_0) - f(z_1) \right)\label{eq:II minus}\\\nonumber
&=& i \pi \frac{{\rm e}^{-i \phi_n}}{2} \left[ \frac{1}{\sqrt{i \mathcal{G}}} - \frac{1}{\sqrt{-\gamma + i \mathcal{G}}} \right],
\end{eqnarray}
where we have defined the function $f$ and the poles
\begin{eqnarray}
f(z) &=& \frac{{\rm e}^{-i \phi_n}}{2}\frac{1}{\sqrt{z}} \nonumber \\
z_0 &=& i \mathcal{G} \nonumber \\
z_1 &=& -\gamma + i \mathcal{G}.
\end{eqnarray}
\begin{widetext}
Plugging (\ref{eq:II minus}) back to (\ref{eq:II}) and evaluating the complex conjugated term we get
\begin{eqnarray}
\frac{2}{\pi} II &=& i {\rm e}^{-i \phi_n} \left[ \frac{1}{\sqrt{i \mathcal{G}}} - \frac{1}{\sqrt{-\gamma + i \mathcal{G}}} \right] + {\rm c.c} \nonumber \\
&=& i \cos(\phi_n) \left[ \frac{1}{\sqrt{i \mathcal{G}}} - \frac{1}{\sqrt{-i \mathcal{G}}} + \frac{1}{\sqrt{-\gamma-i \mathcal{G}}} - \frac{1}{\sqrt{-\gamma + i \mathcal{G}}} \right]
+ \sin(\phi_n) \left[ \frac{1}{\sqrt{i \mathcal{G}}} + \frac{1}{\sqrt{-i \mathcal{G}}} - \frac{1}{\sqrt{-\gamma-i \mathcal{G}}} + \frac{1}{\sqrt{-\gamma + i \mathcal{G}}} \right] \nonumber \\
&=& 2 \left\{ \frac{\cos(\phi_n) \sin(\frac{1}{2} {\rm arg}(i \mathcal{G}))}{|\mathcal{G}|^\frac{1}{2}} - \frac{\cos(\phi_n) \sin(\frac{1}{2} {\rm arg}(\gamma + i \mathcal{G}))}{|\gamma + i \mathcal{G}|^\frac{1}{2}} + \frac{\sin(\phi_n) \cos(\frac{1}{2} {\rm arg}(i \mathcal{G}))}{|\mathcal{G}|^\frac{1}{2}} + \frac{\sin(\phi_n) \cos(\frac{1}{2} {\rm arg}(\gamma + i \mathcal{G}))}{|\gamma + i \mathcal{G}|^\frac{1}{2}} \right\}.
\label{eq:II expansion}
\end{eqnarray}
Next, we note that this result can be further simplified in specific cases: for $\mathcal{G} \in \mathbb{R}$ and $n=0,2$ respectively, we get the following options
\begin{eqnarray}
&n=0:&\; \phi_0 = \frac{\pi}{4} \rightarrow \sin(\frac{\pi}{4}) = \frac{1}{\sqrt{2}}; \;\;\; \cos(\frac{\pi}{4}) = \frac{1}{\sqrt{2}} \nonumber \\
&n=2:&\; \phi_2 = \frac{5 \pi}{4} \rightarrow \sin(\frac{5 \pi}{4}) = \frac{1}{\sqrt{2}}; \;\;\; \cos(\frac{5 \pi}{4}) = -\frac{1}{\sqrt{2}} \nonumber \\
&\mathcal{G}>0:&\; {\rm arg}(i \mathcal{G}) = \frac{\pi}{2} \rightarrow \sin(\frac{1}{2} {\rm arg}(i \mathcal{G})) = \frac{1}{\sqrt{2}}; \;\;\; \cos(\frac{1}{2} {\rm arg}(i \mathcal{G})) = \frac{1}{\sqrt{2}} \nonumber \\
&\mathcal{G}<0:&\; {\rm arg}(i \mathcal{G}) = -\frac{\pi}{2} \rightarrow \sin(\frac{1}{2} {\rm arg}(i \mathcal{G})) = -\frac{1}{\sqrt{2}}; \;\;\; \cos(\frac{1}{2} {\rm arg}(i \mathcal{G})) = \frac{1}{\sqrt{2}}.
\label{eq:na}
\end{eqnarray}
Lets first consider the $n=0$ case. Using (\ref{eq:na}), the relation (\ref{eq:II expansion}) simplifies to
\begin{eqnarray}
&\mathcal{G}>0:& \; \frac{1}{\pi} II_{n=0}(\mathcal{G}) = \frac{1}{|\mathcal{G}|^\frac{1}{2}} - \frac{1}{\sqrt{2}|\gamma + i \mathcal{G}|^\frac{1}{2}} \left( \sin(\frac{1}{2}{\rm arg}(\gamma + i \mathcal{G})) - \cos(\frac{1}{2}{\rm arg}(\gamma + i \mathcal{G})) \right) \nonumber \\
&\mathcal{G}<0:& \; \frac{1}{\pi} II_{n=0}(\mathcal{G}) = - \frac{1}{\sqrt{2}|\gamma + i \mathcal{G}|^\frac{1}{2}} \left( \sin(\frac{1}{2}{\rm arg}(\gamma + i \mathcal{G})) - \cos(\frac{1}{2}{\rm arg}(\gamma + i \mathcal{G})) \right).
\end{eqnarray}
and similar expression is obtained for $n=2$. Those results can be combined to a single expression
\begin{equation}
\frac{1}{\pi} II_n(\mathcal{G}) = \frac{(-1)^{n}}{|\mathcal{G}|^\frac{1}{2}} \theta(\mathcal{G}) - \frac{1}{\sqrt{2}|\gamma + i \mathcal{G}|^\frac{1}{2}} \left( \sin(\frac{1}{2}{\rm arg}(\gamma + i \mathcal{G})) - \cos(\frac{1}{2}{\rm arg}(\gamma + i \mathcal{G})) \right),
\label{eq:II n=0}
\end{equation}
where $\theta$ is the Heaviside step function. Remarkably, the function (\ref{eq:II n=0}) is finite in the limit $\mathcal{G} \rightarrow 0^-$, but diverges as $\mathcal{G} \rightarrow 0^+$, which constitutes the main analytic result of this section.
\end{widetext}
\emph{Remark:} The discontinuity and one-sided divergence could be seen already in (\ref{eq:II}) by setting $\mathcal{G}=0$. This would result in the diverging integral in the vicinity of the origin, namely the part $\int_\epsilon \propto \int {\rm d}t \frac{1}{t\sqrt{t}}$ in (\ref{eq:cont int}), which, according to the estimation lemma, scales as $\propto \frac{1}{\sqrt{\epsilon}}$.
\subsection{Numerical implementation}
In the previous section we have found that (\ref{eq:FS}),(\ref{eq:Fself}) can be evaluated numerically, while (\ref{eq:FL}) can be written as
\begin{equation}
{}^{\phantom{1,}L}_{1,n}V({\bf k}) = \sum_{\bf G} \sum_{\alpha,\beta \in {\rm unit\;cell}} {\rm e}^{-i {\bf G}{\pmb \tau}_{\alpha \beta}} \left[ {}_{1,n}\mathfrak{V}_0^c - I_0^c + I_0^\infty \right].
\label{eq:1nV}
\end{equation}
Here, the expressions ${}_{1,n}\mathfrak{V}_0^c$ and $I_0^c$ are finite and can by obtained numerically. In principle similar decomposition can be obtained when considering the functions (\ref{eq:2nv}),(\ref{eq:3nv}), however we found it more convenient to evaluate the corresponding element (\ref{eq:Vk direct}) by direct summation in real space.
The final subtlety is related to the fact, that for ${}^L \xi = 1-{\rm e}^{-\gamma r}$ the self-interaction term (\ref{eq:Fself}) is not defined for ${}_{1,n=2}v$ as it depends on the direction in which the origin is approached (which is due to the dependence of ${}_{1,n=2}v$ on the polar angle of $\bf r$). This problem can be circumvented by considering ${}^L \xi = 1-{\rm e}^{-\gamma r^2}$ instead in which case the self-interaction term vanishes. Here, it is still possible to evaluate the integral in (\ref{eq:Iab}) symbolically, which we then use in (\ref{eq:1nV}).
In (\ref{eq:1nV}) the tuning parameters $\gamma, c$ are chosen such that the approximations used (eq. (\ref{eq:Bessel approx})) are well satisfied.
|
\section{Introduction}\label{intro}
With the increasing ability of controlling quantum systems, quantum
metrology has become a major and lively field. One important task in
quantum metrology is the parameter estimation of quantum channels,
also known as quantum channel estimation.
It is important for at least two
reasons. Firstly, because to carry out experiments, it is crucial to
know exactly the processes applied to the system. Channel
estimation is then used as quantum process tomography.
Secondly, quantum channel estimation is useful for
investigating the utility of a system for measuring
some relevant
physical quantity, and
for increasing the sensitivity of
sensors.
The field has attracted a lot of attention when it became clear that
for a given number of probes
using entanglement between the probes
can
increase the sensitivity compared to the classical case \cite{vittorio_giovannetti_quantum_enhanced_2004,giovannetti_quantum_2006}.
A second possibility of using entanglement is to introduce an ancillary
system and to entangle it with the original system, while
still applying the
quantum channel only to the original system.
It was shown that such channel
extensions can
sometimes enhance the
sensitivity
\cite{fujiwara_quantum_2001,dariano_using_2001}, in the sense of
increasing the Quantum Fisher Information (QFI).
Among quantum channels,
unitary channels play a particular role. In these, the time evolution
is obtained through propagation with a unitary operator, obtained through the
Hamiltonian that depends on the parameter to be estimated,
and quantum channel estimation for unitary channels is therefore also
known as Hamiltonian parameter estimation.
Phase-shift
Hamiltonians correspond to the special case where the parameter to be estimated
multiplies a Hermitian generator.
The corresponding
parameter estimation problem is particularly relevant due to the
importance of phase measurements in physics. A typical example of such
a situation is the estimation of a phase in an
interferometer. For unitary channels, we can go beyond
channel extension by adding an arbitrary
parameter-independent operator
to the Hamiltonian, which may include
even interactions with
ancillary systems that can be initially entangled with the original
system. Adding a parameter-independent Hamiltonian to the original
Hamiltonian is known as ``Hamiltonian extension''
\cite{fraisse_hamiltonian_2016}.
For
phase shifts, Hamiltonian extension does not improve the best possible
sensitivity.
This was shown in
\cite{boixo_generalized_2007,fraisse_hamiltonian_2016} by calculating
an upper bound for the
QFI for unitary channels and showing that the bound is saturated for
phase shift Hamiltonians.
Here we investigate the
problem of saturating the upper bound calculated in
\cite{boixo_generalized_2007,fraisse_hamiltonian_2016} for general
Hamiltonians, where it is typically not saturated yet,
by extending the
Hamiltonian.
Indeed, from a pure metrological point of view it is
interesting to check
whether or not one can go beyond the optimization over initial state
and POVM measurement
by engineering also the Hamiltonian to increase the
sensitivity, without introducing additional parameter
dependence.
In particular, one would like to know whether ancilla-assisted
schemes can provide an
advantage for general Hamiltonian parameter estimation. We find
that
for general parameter dependent Hamiltonians, the
sensitivity can be increased by adding a part to the Hamiltonian
given by its local derivative with respect to the parameter,
multiplied with a large, parameter-independent prefactor.
This can be
understood intuitively as a ``signal flooding'', \emph{i.e.}~the
relative weight of
the useful part of the Hamiltonian is enhanced. A second way to
saturate the upper bound, is to subtract the Hamiltonian taken at a
fixed value of the parameter from the original Hamiltonian. At this
specific value of the parameter the new Hamiltonian saturates the
upper bound. Neither scheme needs any ancillas.
A third opportunity for Hamiltonian extension exists if the
eigenvalues of the Hamiltonian are independent of the parameter. This
leads to a periodic time-dependence of the channel QFI. An
addition to the Hamiltonian that breaks the parameter-independence
of the eigenvalues can create a quadratic time
scaling in the channel QFI and hence for large enough times
strongly increase the sensitivity.
While typically this method does not
saturate the upper bound, it is still useful for metrology as it
allows one to use
time as a resource. {The engineering of the time scaling to
obtain the $t^2$ scaling was also shown in the context of metrology
with feedback controls
\cite{yuan_optimal_2015,yuan_sequential_2016}. }
We illustrate our results with NV-center
magnetometry, where we find that optimal sensitivity can be reached by
adding a strong external magnetic field in the direction of the field
to be measured, {\em i.e.}~flooding the signal
with a strong
known signal of the same kind. No ancillas are required to reach
optimal sensitivity. We also study the qualitatively different estimation of a direction of
a magnetic field with a single spin-1 as a probe
which allows us to compare
the three different methods.
\section{Quantum metrology}
\subsection{Channel QFI}
In parameter estimation theory,
the goal is to infer the value of a parameter $\theta$ given
the realization of an $m$-sample of a random variable whose
distribution depends on $\theta$. This is done by an estimator
$\est{\theta}$. A common property required for estimators is to be
unbiased, meaning that in an infinitesimal interval about the true
value of the parameter $\theta$, on average the estimator should give
that true value of the parameter, $\moy{\est{\theta}}=\theta$. A second
desirable property is to have a variance as small as possible, such
that the estimate fluctuates as little as possible about the true value of the
parameter.
Parameter estimation theory applies naturally to the metrology of
quantum systems.
The most general measurements correspond to
POVMs (Positive Operator Valued Measure), \emph{i.e.} sets
$\ens{E_\xi}$ of positive semi-definite operators fulfilling the
closure relation $\sum_\xi E_\xi =\ensuremath{ \mathcal{I} }$. Given a state $\rho_\theta$,
a POVM generates a probability distribution $\mu_\theta(\xi) =
\tr{E_\xi \rho_\theta}$, mapping the problem of estimating a parameter
of a quantum state to the problem of parameter estimation. This
approach leads to a fundamental theorem, known as the quantum /
theorem \cite{helstrom_quantum_1969,braunstein_statistical_1994}, which sets a
bound on the variance of any unbiased estimator
given a state $\rho_\theta$,
\begin{equation}
\Vcl{\est{\theta}} \geq \frac{1}{\qfi{\rho_\theta}{\theta}}\;,
\end{equation}
with $\qfi{\rho_\theta}{\theta}$ the Quantum Fisher Information (QFI),
defined as $\qfi{\rho_\theta}{\theta} = \tr{\rho_\theta
L_\theta^2}$. The so-called symmetric logarithmic derivative
$L_\theta$ is defined implicitly by $\frp{\rho_\theta}{\theta}
=(\rho_\theta L_\theta+L_\theta \rho_\theta)/2$. The
bound results from
a double optimization: a first optimization
over all unbiased estimators, and then a second optimization over all
possible POVMs. When repeating the measurement $m$ times independently
--- which amounts to generating an $m$-sample --- the quantum /
theorem reads $\Vcl{\est{\theta}} \geq
(m\qfi{\rho_\theta}{\theta})^{-1}$. The bound can always be saturated
in the limit of large numbers $m$ of measurements by using the maximum
likelihood estimator and building the POVM based on the eigenvectors
of $L_\theta$.
Often the parameter to be estimated characterizes
a physical process, \emph{i.e.}~a quantum
channel:
One starts with an initial state $\rho_0$
of the probe that is independent of the parameter.
Then we let the quantum channel $\mc{E}_\theta$ act on the probe,
giving as a result the state $\rho_\theta = \mc{E}_\theta (\rho_0)$.
A POVM measurement is then performed, whose outcome is fed into
$\hat{\theta}_\text{est}$ and provides us with an estimate of $\theta$.
We see that in this case we have a new degree of freedom in the QFI,
the choice of the initial state. One thus introduces a new quantity,
the channel QFI $C$, which corresponds to the largest QFI
reachable for a given channel,
\begin{equation}
\cqfi{\mc{E}_\theta}{\theta}=\max_{\rho_0} \qfi{\mc{E}_\theta(\rho_0)}{\theta}\;.
\end{equation}
Importantly, due to the convexity of the QFI, it is enough to maximize
over pure states, $\rho_0=\ddens{\psi_0}$.
\subsection{Channel extensions and Hamiltonian extensions}
We now turn our attention to extensions for metrology.
It has been shown that by adding an ancilla
to the probe but still acting only with the channel on the probe,
\emph{i.e.}~applying $\mc{E}\otimes \ensuremath{ \mathrm{Id} }$ to the whole system, one
can, for certain channels, improve the channel QFI
\cite{dariano_using_2001,fujiwara_quantum_2001}. We call such
extensions ``channel extensions'', in contrast to ``Hamiltonian
extensions'', defined below. For
unitary channels, channel extensions
do not increase the
channel QFI: $\cqfi{\mc{U}_{{H(\theta)}} \otimes
\ensuremath{ \mathrm{Id} }}{\theta}=\cqfi{\mc{U}_{{H(\theta)}}}{\theta}$ (see
section \ref{sec.cqfi} for a short proof).
But when extending the Hamiltonian $H(\theta)$
by adding to it another parameter-independent
Hamiltonian $H_1$,
\begin{equation}\label{eq:def_ext_ph_shift}
H_{\mathrm{ext}}(\theta)=
H(\theta)\otimes \ensuremath{ \mathcal{I} } + H_{1}\;,
\end{equation}
we are not anymore in the situation covered by channel
extension.
Eq.\eqref{eq:def_ext_ph_shift} is our formal definition of a ``Hamiltonian
extension''. $H_\text{1}$ can be a coupling between the original
system and an ancilla, but can also refer to a Hamiltonian that acts
non-trivially only on the Hilbert space of the original system or on the
Hilbert space of the ancilla.
Can such a Hamiltonian
extension lead to an improvement in the channel QFI? It was shown
in \cite{fraisse_hamiltonian_2016} that for the specific case of
phase shift Hamiltonians, \emph{i.e.} Hamiltonians of the form
$H(\theta) = \theta G$, such Hamiltonian extensions cannot improve the
channel QFI, a
result that was
already known in the context of
many-body interaction metrology
\cite{boixo_generalized_2007}.
Here we investigate the question more generally for arbitrary Hamiltonians
$H(\theta)$.
\section{Metrology with unitary channels}
\subsection{Channel QFI {and semi-norm} }\label{sec.cqfi}
From now and for the rest of the article we will focus on unitary
channels. These channels correspond to the unitary evolution of the
state of a closed
system
described by the Schr\"odinger equation. Given the Hamiltonian
$H(\theta)$, the effect of the channel $\mc{U}_{H(\theta)}$ on the
initial state $\rho_0$ is given by
$\mc{U}_{H(\theta)}(\rho_0)=U_{H(\theta)}
\rho_0{U_{H(\theta)}}^\dagger $ with $U_{H(\theta)}=\e{-\ensuremath{ \mathrm{i\,} } t
H(\theta)}$ the evolution operator (we take $\hbar=1$ throughout
the paper apart from the section \ref{sec:NV}). Since we only deal
with unitary channels we adopt the notation
$\cqfi{\mc{U}_{H(\theta)}}{\theta} \equiv \cqfi{H(\theta)}{\theta} $
for the channel QFI.
In general, calculating the QFI is a difficult task as it requires one
to diagonalize the density matrix. For pure states the expression is
still simple and reads,
$\qfi{\ddens{\psi_\theta}}{\theta}=\ppscal{\dot{\psi}_\theta}-\vert
\pscal{\psi_\theta}{\dot{\psi}_\theta} \vert^2$, where here and
throughout the article the dot stands for the derivative with respect
to the parameter to be estimated, {\em
i.e.}~$\ket{\dot{\psi}_\theta}=\frp{\ket{\psi_\theta}}{\theta}$. Introducing
the local generator
\begin{equation}
\msc{H} = \ensuremath{ \mathrm{i\,} } {U_{H(\theta)}}^\dagger \dot{U}_{H(\theta)}\;,
\end{equation}
we can write the QFI as
\begin{equation}\label{eq:qfi_gen_ham_3}
\qfi{\ddens{\psi_\theta}}{\theta} =4\Vstate{\msc{H}}{\ket{\psi_0}}\;,
\end{equation}
with $ \Vstate{\msc{H}}{\ket{\psi_0}} \equiv
\moyvec{\msc{H}^2}{\psi_0} - \moyvec{\msc{H}}{\psi_0}^2$,
and consequently the channel QFI as
\begin{equation}
\cqfi{H(\theta)}{\theta}= 4 \max_{\ket{\psi_0}\in\mc{H}} \Vstate{\mathscr{H}}{\ket{\psi_0}}\;.
\end{equation}
The maximization {\cite{giovannetti_quantum_2006}}
can be done as follows: Using Popoviciu's inequality
\cite{Popoviciu35}, which states that for a random variable $X$, with
minimal value $a$ and maximal value $b$, the variance of $X$ is upper
bounded by $(b-a)^2/4$, and then noticing that in
eq.\eqref{eq:qfi_gen_ham_3} the variance saturates its upper bound
for states of the form $(\ket{h_{\mathrm{M}}}+\e{\ensuremath{ \mathrm{i\,} } \ensuremath{\varphi}}
\ket{h_{\mathrm{m}}})/\sqrt{2}$
(where $\ket{h_{\mathrm{M}}}$ and $\ket{h_{\mathrm{m}}}$ correspond,
respectively, to
eigenvectors of $\msc{H}$ with maximal eigenvalue ($h_{\mathrm{M}}$) and minimal eigenvalue ($h_{\mathrm{m}}$)),
we have
\begin{equation}
\cqfi{H(\theta)}{\theta}= (h_{\mathrm{M}}-h_{\mathrm{m}})^2\;.
\end{equation}
In order to simplify the calculation of the channel QFI, and following the method in \cite{boixo_generalized_2007}, we introduce the semi-norm
\begin{equation}
\pnorm{A}{sn} = a_{\mathrm{M}}-a_{\mathrm{m}}
\end{equation}
with $ a_{\mathrm{M}}$ and $ a_{\mathrm{m}}$ the maximal and minimal
eigenvalues of $A$ { (we call such eigenvalues \emph{extremal}
eigenvalues, and their associated eigenvectors, \emph{extremal}
eigenvectors ; we also call maximal (resp.~minimal) eigenvector an
eigenvector corresponding to a maximal (resp.~minimal)
eigenvalue). We then have the simple expression for the channel
QFI
\begin{equation}\label{eq:cqfi_sn}
\cqfi{H(\theta)}{\theta}= \pnorm{\msc{H}}{sn}^2\;.
\end{equation}
As it will be important in the following, let us show the triangle
inequality for this semi-norm. Let $A=B+C$. Then
$\pnorm{A}{sn}=\moyvec{A}{a_{\mathrm{M}}} -
\moyvec{A}{a_{\mathrm{m}}}=\moyvec{B}{a_{\mathrm{M}}}+\moyvec{C}{a_{\mathrm{M}}}
- \moyvec{B}{a_{\mathrm{m}}} - \moyvec{C}{a_{\mathrm{m}}}$, with $\ket{a_{\mathrm{M}}}$ (resp.~$\ket{a_{\mathrm{m}}}$) a maximal (resp.~minimal) eigenvector {of $A$}.
By definition $\moyvec{B}{a_{\mathrm{M}}} \leq b_{\mathrm{M}}$ and
$-\moyvec{B}{a_{\mathrm{m}}} \leq -b_{\mathrm{m}}$ with $
b_{\mathrm{M}}$ and $ b_{\mathrm{m}}$ the maximal and minimal
eigenvalues of $B$. In the same way we have
$\moyvec{C}{a_{\mathrm{M}}} \leq c_{\mathrm{M}}$ and
$-\moyvec{C}{a_{\mathrm{m}}} \leq -c_{\mathrm{m}}$ with $
c_{\mathrm{M}}$ and $ c_{\mathrm{m}}$ the maximal and minimal
eigenvalues of $C$. We thus get $\pnorm{A}{sn}=\pnorm{B+C}{sn} \leq
b_{\mathrm{M}}+c_{\mathrm{M}}-b_{\mathrm{m}}-c_{\mathrm{m}} =
\pnorm{B}{sn}+\pnorm{C}{sn}$ which is exactly the triangle
inequality. {When $B$ and $C$ have no degenerate extremal eigenvalues } the equality is reached for $\ket{b_{\mathrm{M}}} \propto
\ket{c_{\mathrm{M}}}$ and $\ket{b_{\mathrm{m}}} \propto
\ket{c_{\mathrm{m}}}$, meaning that both operators have to share the
same extremal eigenvectors. For the degenerate case, the triangle
inequality is saturated
if and only if the intersection of the invariant
subspaces of the maximal (resp.~minimal) eigenvalue of $B$ and of $C$
is not empty \footnote{Say $A$ is an operator acting on $E$. Then a
subspace $M\subset E$ is an invariant subspace of $A$ if and only if
$AM\subset M$. We also say that $M$ is stable by $A$}. Stated
otherwise, $B$ and $C$ should share (in the sense of proportionality)
at least one {maximal and one minimal eigenvector}.
Since the eigenvalues
are preserved by
similarity transformations, we also have $\pnorm{U A
U^{-1}}{sn}=\pnorm{A}{sn}$ for any unitary $U$, regardless of whether
$U$ depends on $\theta$ or not.
We can use eq.\eqref{eq:cqfi_sn} to show that channel extension does
not increase the channel QFI of unitary channels: For a unitary
channel $\mc{U}_{H(\theta)}$ the extended channel
$\mc{U}_{H(\theta)}\otimes \ensuremath{ \mathrm{Id} }$ can be written as
$\mc{U}_{H(\theta)\otimes \ensuremath{ \mathcal{I} }}$. We furthermore have
${U_{H(\theta)\otimes \ensuremath{ \mathcal{I} }}}^\dagger \dot{U}_{H(\theta)\otimes
\ensuremath{ \mathcal{I} }}={U_{H(\theta)}}^\dagger \dot{U}_{H(\theta)} \otimes \ensuremath{ \mathcal{I} }$
showing that $\cqfi{\mc{U}_{H(\theta)}\otimes \ensuremath{ \mathrm{Id} }}{\theta}=
\pnorm{\msc{H}\otimes
\ensuremath{ \mathcal{I} }}{sn}=\pnorm{\msc{H}}{sn}=\cqfi{\mc{U}_{H(\theta)}}{\theta}$,
using eq.\eqref{eq:cqfi_sn} plus the fact that $\msc{H}\otimes \ensuremath{ \mathcal{I} }$
has the same eigenvalues as $\msc{H}$.
\subsection{Upper bound for the channel QFI} \label{sec:upper_bound}
\begin{lemma}[Upper bound for channel QFI \cite{boixo_generalized_2007}]\label{Boixo}
For general Hamiltonians $H(\theta)$, with associated evolution operator $U_{H(\theta)}=\e{-\ensuremath{ \mathrm{i\,} } t H(\theta) }$, the channel QFI is upper bounded {as}
\begin{equation}\label{eq:main_ineq}
\cqfi{ H(\theta) }{ \theta } \leq t^2 \pnorm{\dot{H}(\theta)}{sn}^2\;.
\end{equation}
\end{lemma}
\begin{lemma}[Saturation of the bound]\label{sat}
In the case where $\dot{H}(\theta)$ has no degenerate extremal
eigenvalues,
equality in \eqref{eq:main_ineq} is reached if and
only if the extremal eigenvectors of $\dot{H}(\theta)$ are also
eigenvectors of $H(\theta)$.
In the degenerate case,
equality in \eqref{eq:main_ineq} is reached if and
only if there exist $\ket{\psi} \in \mc{P}$ and $\ket{\phi} \in \mc{D}$ such that
$V(\alpha) \ket{\psi} \in \mc{P}$ and $V(\alpha) \ket{\phi} \in
\mc{D}$ for all $\alpha \in [0,1]$ with
$\mc{P}=\spn{\ket{1},\cdots,\ket{a}}$
(resp.~$\mc{D}=\spn{\ket{b},\cdots,\ket{d}}$)
the invariant subspace of $\dot{H}(\theta)$ associated with its
maximal
(resp.~minimal) eigenvalue, and where $V(\alpha)=\e{-\ensuremath{ \mathrm{i\,} } \alpha t
H(\theta)}$. A sufficient condition is that there exists an
eigenvector of $H(\theta)$
in $\mc{P}$ and another one in $\mc{D}$.
The inequality is saturated by
$\ket{\psi_\mathrm{opt}} =(\ket{M}+\ket{m})/\sqrt{2}$ with $\ket{M}
\in \mc{P}$ and $\ket{m} \in \mc{D}$, {\em i.e.}~a balanced
superposition of a maximal eigenvector and a minimal eigenvector of
$\dot{H}(\theta)$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{Boixo}\cite{boixo_generalized_2007}]
In general \cite{wilcox_exponential_1967,snider_perturbation_1964}, for a matrix $M(x)$ depending on the parameter $x$, the derivative of its exponential with respect to $x$ is given by
\begin{equation}\label{eq:der_exp}
\frd{\e{M(x)}}{x} =\int_{0}^1 \e{\alpha M(x)}\frd{M(x)}{x}\e{(1-\alpha) M(x)}\diff \alpha\;.
\end{equation}
From this expression we can re-express the local generator of the translation,
\begin{equation}
\msc{H} = t \int_{-1}^0 W(\alpha,\theta) \diff \alpha\;,
\end{equation}
with $W(\alpha,\theta) =V(\alpha)\dot{H}(\theta) V(\alpha)^\dagger$,
where $V(\alpha)=\e{-\ensuremath{ \mathrm{i\,} } \alpha t H(\theta)}$.
Applying the triangle inequality to the semi-norm of $\msc{H}$, and noticing that $\pnorm{W(\alpha,\theta)}{sn}=\pnorm{\dot{H}(\theta)}{sn}$ we obtain
\begin{equation} \label{eq:first_ineq}
\pnorm{\msc{H}}{sn} \leq t \int_{-1}^0 \pnorm{ W(\alpha,\theta) }{sn}\diff \alpha = t\pnorm{\dot{H}(\theta)}{sn}\;.
\end{equation}
\end{proof}
\begin{proof}[Proof of Lemma \ref{sat}]
Since $H(\theta)$ is Hermitian, so is
$\dot{H}(\theta)$ and we can then write it in its orthonormal eigenbasis
as $\dot{H}(\theta) =\sum_{i=1}^d e_i \ddens{i}$, with $e_1=\cdots=e_a > e_i > e_{b}=
\cdots =e_d
$ for all $a<i<b$, where $a$ is the dimension of the
invariant subspace of maximal eigenvalue and $d-b+1$ is the dimension
of the invariant subspace with minimal eigenvalue.
Since $W(\alpha,\theta)$ is related to
$\dot{H}(\theta)$ by a similarity transformation, we can write it as
$W(\alpha,\theta) = \sum_{i=1}^d e_i
\ddens{\alpha_i}$ with $\ket{\alpha_i} = V(\alpha) \ket{i}$. The
condition for equality in eq.\eqref{eq:first_ineq} and hence in
eq.\eqref{eq:main_ineq} is that there exists a vector $\ket{\psi}$
which is an eigenvector of $W(\alpha,\theta)$ with eigenvalue $e_1$
simultaneously for all values of $\alpha \in [-1,0]$, and a vector
$\ket{\phi}$ which is an eigenvector of $W(\alpha,\theta)$ with
eigenvalue $e_d$ simultaneously for all values of $\alpha \in
[-1,0]$.
Consider first the case where
$\dot{H}(\theta)$ has no degenerate extremal eigenvalues,
\emph{i.e.}~$a=1$ and $b=d$.
Then the condition for equality
is equivalent to
$\ket{1}$
and $\ket{d}$ being
eigenvectors of $V(\alpha)$ for all $ \alpha \in [-1,0]$.
Let us see how we can re-express this condition so that it
only involves {eigenvectors} of
$\dot{H}(\theta)$ and $H(\theta)$. By expressing the
Hamiltonian in its eigenbasis, $H(\theta) = \sum_i h_i \ddens{h_i}$,
we can write $V(\alpha)=\e{-\ensuremath{ \mathrm{i\,} } \alpha t h_i}\ddens{h_i}$.
The eigenvectors of $H(\theta)$ are also eigenvectors of $V(\alpha)$ but
not the other way round: Indeed, $\e{-\ensuremath{ \mathrm{i\,} } \alpha t h_i}$ may be equal to $\e{-\ensuremath{ \mathrm{i\,} } \alpha t h_j}$
while $h_i \neq h_j$. In such cases we can construct eigenvectors of
$V(\alpha)$ which are not eigenvectors of $H(\theta)$ by linearly
combining $\ket{h_i}$ and $\ket{h_j}$. Nevertheless this
can happen only for a countable number of $\alpha$-values
given $t$, whereas in all other
cases the eigenvectors of $V(\alpha)$ must also be eigenvectors of
$H(\theta)$. Hence, the condition $\e{-\ensuremath{ \mathrm{i\,} } \alpha t h_i}=\e{-\ensuremath{ \mathrm{i\,} } \alpha t h_j}$ cannot be satisfied for all
$ \alpha \in [-1,0]$ if $h_i \neq h_j$,
and therefore the condition for equality in \eqref{eq:first_ineq} can
be stated as: $\ket{1}$ and $\ket{d}$ should be eigenvectors of
$H(\theta)$.
In the degenerate case the
eigenspace of $W(\alpha,\theta)$ with
eigenvalue $e_1$ is spanned by $\lbrace
V(\alpha)\ket{1},\cdots,V(\alpha)\ket{a} \rbrace$. The existence of a
common eigenvector of $W(\alpha,\theta)$ for $\alpha \in [-1,0]$ with
eigenvalue $e_1$ means that there should exist a vector
$\ket{\psi}$ independent of $\alpha$ that can be written as
$\sum_{i=1}^a \psi_i(\alpha) V(\alpha) \ket{i} =V(\alpha)
\sum_{i=1}^a \psi_i(\alpha) \ket{i} =V(\alpha)
\ket{\varphi(\alpha)}$ where the last equality defines
$\ket{\varphi(\alpha)}$. Since $\ket{\varphi(\alpha)} \in
\mc{P}$ this is equivalent to say that there should exist a vector
$\ket{\psi}$ such that $ V(-\alpha)\ket{\psi}$ belongs to $\mc{P}$ for
$\alpha \in[-1,0]$. This is true especially for $\alpha=0$ showing
that also $\ket{\psi}$ belongs
to $\mc{P}$. A similar treatment
for the lowest eigenvalue shows that the necessary and sufficient
condition for equality in eq.\eqref{eq:main_ineq}
is equivalent to
the existence of a vector $\ket{\psi} \in
\mc{P}$ such that $V(\alpha)\ket{\psi} \in \mc{P}$ for $\alpha
\in[0,1]$, and of a vector $\ket{\phi}\in \mc{D}$ such that
$V(\alpha)\ket{\phi} \in \mc{D}$ for $\alpha \in[0,1]$.
The sufficient condition is found by observing that the above condition is fulfilled if $H(\theta)$ has an eigenvector in $\mc{P}$ and an eigenvector in $\mc{D}$.
\end{proof}
In the case of phase shifts, the condition for equality in
eq.\eqref{eq:main_ineq}
is fulfilled as $\dot{H}(\theta)$ and $H(\theta)$ are simultaneously diagonalizable, showing that for phase
shifts $\cqfi{\theta G}{\theta}=\pnorm{G}{sn}$. The Hamiltonian
extension $G_{\mathrm{ext}}(\theta)$ of a phase shift
is not a phase shift anymore. We thus have
$\cqfi{G_{\mathrm{ext}}(\theta)}{\theta} \leq
\pnorm{\dot{G}_{\mathrm{ext}}(\theta)}{sn}=\pnorm{G \otimes
\ensuremath{ \mathcal{I} }}{sn}=\pnorm{G }{sn} = \cqfi{\theta G}{\theta}$, showing that
quantum metrology with phase shift Hamiltonians cannot profit from
Hamiltonian extension, in addition to not being able to profit from channel extensions.
\section{Saturating the bound} \label{sec:saturating_bound}
We have seen that Hamiltonian extension fails to provide an advantage
in terms of channel QFI for phase shift Hamiltonians. Nevertheless
the question of the order between $\pnorm{\mathscr{H}}{sn}$ and
$\pnorm{\mathscr{H}_{\mathrm{ext}}}{sn}$ is still open
for arbitrary Hamiltonians $H(\theta)$, {where
$\mathscr{H}_{\mathrm{ext}}=\ensuremath{ \mathrm{i\,} } U_{H_{\mathrm{ext}}}{}^\dagger \dot{U}_{H_{\mathrm{ext}}} $ is the local generator of the extended Hamiltonian with $U_{H_{\mathrm{ext}}}=\e{-\ensuremath{ \mathrm{i\,} } t H_{\mathrm{ext}}}$.}
We only have the two inequalities
\begin{align}
\pnorm{\mathscr{H}}{sn}^2 &\leq t^2 \pnorm{\dot{H}(\theta)}{sn}^2 \label{upper_bound_1}\\
\pnorm{\mathscr{H}_{\mathrm{ext}}}{sn}^2 &\leq t^2 \pnorm{\dot{H}(\theta)}{sn}^2 \label{upper_bound_2} \,,
\end{align}
where the second one follows from the fact that $H_\text{1}$ is
independent of $\theta$.
The interesting question here is thus whether for
a given $H(\theta)$ that does not saturate \eqref{upper_bound_1}
we can saturate the bound \eqref{upper_bound_2} by tuning the interaction
Hamiltonian in $H_{\mathrm{ext}}$, and therefore increase the
sensitivity, \emph{i.e.}~have
$\pnorm{\mathscr{H}_{\mathrm{ext}}}{sn}^2>\pnorm{\mathscr{H}}{sn}^2$.
To answer this question, we first look at the specific case of {what
we call} ``broken phase-shift''
before treating the general case.
\subsection{Restoring a broken phase shift}\label{sec:restoring}
Let us consider the Hamiltonian $K(\theta) = \theta G + F$,
along with the corresponding unitary operator $U_{K(\theta) }=\e{-\ensuremath{ \mathrm{i\,} } K(\theta)t}$
and the corresponding local generator $\mathscr{K}=\ensuremath{ \mathrm{i\,} }
U_{K(\theta)}{}^\dagger\dot{U}_{K(\theta)}$. For $K(\theta)$ we have
from eq.\eqref{eq:main_ineq}
$\pnorm{\mathscr{K}}{sn}^2 \le t^2 \pnorm{G}{sn}^2$.
We assume that the conditions for equality in eq.\eqref{eq:main_ineq}
are not fulfilled and we therefore have $\pnorm{\mathscr{K}}{sn}^2 <
t^2\pnorm{G}{sn}^2$. Our goal is to design a Hamiltonian extension $K_{\mathrm{ext}}(\theta)$ for $K(\theta)$ such that the channel QFI for $K_{\mathrm{ext}}(\theta)$
saturates inequality \eqref{eq:main_ineq},
i.e.
\begin{equation}
\cqfi{K_{\mathrm{ext}}(\theta)}{\theta} =t^2 \pnorm{G}{sn}^2
\;.\label{16}
\end{equation}
In order to saturate the bound two solutions appear directly.
Loosely
speaking, we can either cancel the part that spoils the Hamiltonian,
$ F$, or increase the useful part of the Hamiltonian, $\theta
G$. Both of them correspond to a Hamiltonian extension but without
ancillas, just adding an extra part to the Hamiltonian. In the first
case the extra Hamiltonian is $ - F$, and the {corresponding extended
Hamiltonian} becomes
the phase shift $\theta G${, which saturates the bound
\eqref{16}}.
While this method may appear artificial,
it clearly demonstrates
the possibility of enhancing the sensitivity in the case where
the parameter is not coded in a simple phase shift. In fact, it is
not even necessary to add the full $-F$. Assuming for simplicity
non-degenerate extremal eigenvalues of {$\dot{K}(\theta)$},
we know from
Lemma \ref{sat} that only the
extremal eigenvectors of {
$\dot{K}(\theta)$} have to be eigenvectors of
$K(\theta)$. With this one shows easily that a corrected
$\tilde{K}(\theta)\equiv K(\theta)+R$ {saturates the upper bound}
if and only if
$\bra{m} R \ket{n}=-\bra{m} F \ket{n}$ for $n=1$ with $m \in\{2,\cdots
,d\}$ and for
$n=d$ with $m\in\{1,\cdots, d-1\}$}. {\em I.e.}~loosely speaking, $F$ needs
to be substracted only in
the subspace of the extremal eigenvectors.
The second strategy is to
add a Hamiltonian $\beta G$. The extended Hamiltonian in this case
reads $K_{\mathrm{ext}}(\theta,\beta) =( \theta +\beta)G + F$,
\emph{i.e.} we have
$K_{\mathrm{ext}}(\theta,\beta)=K(\theta+\beta)$.
Hence, for large $\beta$, the eigenvectors of $K_\text{ext}$ become
those of $G$, {\em i.e.}~$K_\text{ext}$ and $\dot{K}_\text{ext}$ have then
the same extremal eigenvectors and the bound {\eqref{16}}
can be saturated.
Importantly, this transformation does not
correspond to just
a re-parametrization: In a re-parametrization we keep
the same probability distribution, and we only change what we consider
to be the parameter of the probability distribution. In the present
case
we change the probability distribution, but
we keep the original parameter, since it corresponds to the physical quantity
in which we are interested.
Let us formalize this in terms of Fisher Information (FI).
We consider the distribution $\mu_\theta(x)$. The FI of this distribution for the parameter $\theta$ at the point $\theta_0$ is $\cfi{\mu_{\theta}(x)}{\theta}\vert_{\theta=\theta_0}$. If we now consider a new parameter $g \equiv g(\theta)$, then the FI for the same probability distribution at $g_0\equiv g(\theta_0)$ is given by
\begin{equation}
\cfi{\mu_{\theta}(x)}{g}\vert_{g=g_0}=\frac{\cfi{\mu_{\theta}(x)}{\theta}\vert_{\theta=\theta_0}}{\left( \frp{g(\theta)}{\theta} \vert_{\theta=\theta_0}\right)^2}\;.
\end{equation}
In the case of a shift of the parameter, say $g(\theta)=\theta + y$ we
get $\frp{g(\theta)}{\theta}=1$ and the FI
is not changed. This
corresponds to the intuitive picture that by just changing what we
consider to be the parameter in the probability distribution we do not
really gain more information.
Fundamentally different is the change of the probability
distribution. Consider the new probability distribution
$\tilde{\mu}_\theta(x) = \mu_{f(\theta)}(x)$. Then the FI of the new
distribution, still for the same parameter $\theta$,
is given by
\begin{equation}
\cfi{\mu_{f(\theta)}(x)}{\theta}\vert_{\theta=\theta_0} = \left(
\frp{f(\theta)}{\theta} \bigg\vert_{\theta=\theta_0}\right)^2
\cfi{\mu_{\theta}(x)}{\theta}\vert_{\theta=f(\theta_0)}\;.
\end{equation}
In the specific case of a shift of the parameter in the probability distribution, $f(\theta,\beta)=\theta+\beta$ we obtain \begin{equation}
\cfi{\mu_{\theta+\beta}(x)}{\theta}\vert_{\theta=\theta_0} = \cfi{\mu_{\theta}(x)}{\theta}\vert_{\theta=\theta_0 +\beta}\;.
\end{equation}
This means that when we use the new distribution $\tilde{\mu}$, the
resulting FI
at $\theta_0$ equals the FI
of the original
distribution $\mu$ at the point $\theta_0+\beta$ (see Figure
\ref{fig0}).
\begin{figure}
\centering\includegraphics[scale=0.75]{shift_FI_1}
\caption{Change of probability distribution which amounts to
a shift of the parameter in the FI. While working at the same value of
the parameter, the second distribution (bottom plot) offers an
increased QFI. }\label{fig0}
\end{figure}
If we now go back to physics, and if we consider $\beta$
to be a free parameter, then we can tune it to work at the most
favorable position of the Hamiltonian in terms of
$\theta$.
In the case of the broken phase shift, the most favorable values of
$\theta$ are the large values of $\theta$, where the effect of the
term $ F$ becomes negligible. Then, whatever is the original value of
$\theta_0$ we can, by adding the Hamiltonian $\beta G$ and taking
$\beta$ arbitrarily large, approach arbitrarily the upper bound of the QFI.
\subsection{Maximum sensitivity by "Signal flooding"}
We have seen that in order to saturate the upper bound of the channel
QFI for the "broken phase shift", $K(\theta) = \theta G + F$,
we can add a
Hamiltonian proportional to $G $, the generator of the
original phase shift. This generator corresponds to the first
derivative of the total Hamiltonian $G=\dot{K}(\theta) $. We show in
this section that this
is actually a general result: By adding a term proportional to the
first derivative of the Hamiltonian, we can bring the channel QFI
arbitrarily close to its upper bound. We denote the extended
Hamiltonian obtained in this way as
\begin{equation}
H_{\mathrm{fl}}(\theta,\beta)=H(\theta) +\beta \dot{H}(\theta_0) \;.
\end{equation}
Instead of working directly with the channel QFI we
first consider
the QFI for an arbitrary pure state.
Starting with an initial state $\ket{\psi_0}$ the QFI is given by
\begin{equation}
\qfi{U_{H_{\mathrm{fl}}} \ket{\psi_0}}{\theta} =\Vstate{\msc{H}_{\mathrm{fl}}}{\ket{\psi_0}}\;,
\end{equation}
with the local generator $\msc{H}_{\mathrm{fl}}=\ensuremath{ \mathrm{i\,} }
U_{\mathrm{fl}}{}^\dagger {\dot{U}_{\mathrm{fl}}}$
and the evolution operator $U_{\mathrm{fl}}=\e{-\ensuremath{ \mathrm{i\,} } H_{\mathrm{fl}}(\theta,\beta)}$.
The important part in this expression is the derivative of the
evolution operator. Using eq.\eqref{eq:der_exp} we can write it as
\begin{align*}
\dot{U}_{\mathrm{fl}}\vert_{\theta=\theta_0}&=\frp{U_{\mathrm{fl}}}{\theta}\bigg\vert_{\theta=\theta_0} \nonumber \\
&= - \ensuremath{ \mathrm{i\,} } t \int_{0}^1 \e{- \ensuremath{ \mathrm{i\,} } \alpha t H_{\mathrm{fl}}(\theta_0,\beta)}\dot{H}(\theta_0) \e{ -\ensuremath{ \mathrm{i\,} } (1-\alpha) t H_{\mathrm{fl}}(\theta_0,\beta) }\diff \alpha \;.
\end{align*}
The derivative of the evolution operator with respect to $\beta$ is independent of the value of $\beta$ and equals
\begin{multline}
\frp{U_{\mathrm{fl}}\vert_{\theta=\theta_0}}{\beta}= - \ensuremath{ \mathrm{i\,} } t \int_{0}^1 \e{- \ensuremath{ \mathrm{i\,} } \alpha t H_{\mathrm{fl}}(\theta_0,\beta)} \\ \times \frp{H_{\mathrm{fl}}(\theta_0,\beta)}{\beta} \e{ -\ensuremath{ \mathrm{i\,} } (1-\alpha) t H_{\mathrm{fl}}(\theta_0,\beta) }\diff \alpha \;,
\end{multline}
and since $\frp{H_{\mathrm{fl}}(\theta_0,\beta)}{\beta} =\dot{H}(\theta_0)$, we have
\begin{equation}
\frp{U_{\mathrm{fl}}}{\theta}\bigg\vert_{\theta=\theta_0}=\frp{U_{\mathrm{fl}}\vert_{\theta=\theta_0}}{\beta}\;.
\end{equation}
We thus have
\begin{equation}\label{eq:equality_theta_beta}
\qfi{U_{\mathrm{fl}} \ket{\psi_0}}{\theta} \vert_{\theta=\theta_0}= \qfi{U_{\mathrm{fl}}\ket{\psi_0}}{\beta}\vert_{\theta=\theta_0}\;,
\end{equation}
\emph{i.e.}~the QFI for $\theta$ at $\theta_0$ is equal to the QFI for $\beta$ at $\theta=\theta_0$.
In the limit of large $\beta$ we have {$H_{\mathrm{fl}}(\theta,\beta)\vert_{\beta \gg 1} \simeq
\beta \dot{H}(\theta_0)$}.
This Hamiltonian is, with respect to $\beta$, a phase shift
Hamiltonian which implies that
$\cqfi{H_{\mathrm{fl}}(\theta,\beta)\vert_{\beta \gg 1}}{\beta} \simeq t^2
\pnorm{\dot{H}(\theta_0)}{sn}^2$. Since eq.\eqref{eq:equality_theta_beta}
is true for
all states $\ket{\psi_0}$, it is also true for the channel QFI, {\em
i.e.}~{$\cqfi{H_{\mathrm{fl}}(\theta,\beta)|_{\beta\gg
1}}{\theta}=\cqfi{H_{\mathrm{fl}}(\theta,\beta)|_{\beta\gg 1}}{\beta}\simeq t^2
\pnorm{\dot{H}(\theta_0)}{sn}^2$},
showing
\emph{in fine} that ``signal flooding'' allows one
to saturate the upper bound for the channel
QFI,
\begin{equation}\label{eq:saturating_bound}
\cqfi{H_{\mathrm{fl}}(\theta,\beta)\vert_{\beta \gg 1}}{\theta}\vert_{\theta=\theta_0} \simeq t^2 \pnorm{\dot{H}(\theta_0)}{sn}^2\;.
\end{equation}
The interpretation of this result
follows from Lemma \ref{sat}: by adding a large term proportional to $
\dot{H}(\theta_0)$ to $H(\theta)$,
we bring the eigenvectors of
$H_{\mathrm{fl}}(\theta)$ close to
those
of $\dot{H}(\theta_0)$. When the added
part
dominates completely the Hamiltonian the conditions for equality given
in the Lemma \ref{sat} are fulfilled, and the bound is saturated.
Of course, for applying this method, in general we
need to know the value of
$\theta$ already in order to be able to add the Hamiltonian $\beta
\dot{H}(\theta_0)$. But the situation is not worse than what one
encounters when optimizing the POVM, which through $L_\theta$ usually
also depends on $\theta_0$:
The framework of the QFI and its operational
meaning are local, but
the QFI is still a useful quantity. One typically assumes that we
know already "roughly" the value of $\theta$ and that knowledge can be
used to find a near-optimal POVM. In the present context, we would
also use this prior knowledge to determine the Hamiltonian $\beta
\dot{H}(\theta_0)$ to be added. { Moreover, for the physically important
case of a broken phase shift, $\dot{H}(\theta)$ is independent of $\theta_0$,
and hence no knowledge at all of $\theta_0$ is required for flooding
the signal}.
We emphasize that the Hamiltonian we add does not
depend on $\theta$, but only on $\theta_0$. Adding a Hamiltonian that
depends on $\theta$ may for sure increase the channel QFI to values
actually larger than the upper bound, but requires not only prior
information, but also the need to design a way to add an extra
dependence on the
parameter. This would be comparable to
adding an extra dependence on the parameter through a
$\theta$-dependent POVM \cite{seveso_quantum_2017}.
\subsection{Subtracting the Hamiltonian}
The first method discussed in section \ref{sec:restoring} in the context of a
broken phase shift, namely subtracting the disturbing part
$F=K(\theta_0)- \theta_0 \dot{K}(\theta_0)$ from the
Hamiltonian $K(\theta)$
can be generalized
further: We can
subtract the entire Hamiltonian at $\theta_0$ from the Hamiltonian,
leading to a new Hamiltonian
\begin{equation}
H_{\mathrm{sub}}(\theta) = H(\theta) - H(\theta_0)\;.
\end{equation}
This is a valid Hamiltonian extension in the sense that we added a
$\theta$-independent operator to the Hamiltonian while keeping its
parametric derivative,
$\dot{H}_{\mathrm{sub}}(\theta)=\dot{H}(\theta)$.
Moreover, at $\theta=\theta_0$ this Hamiltonian vanishes,
$H_{\mathrm{sub}}(\theta_0) =0$,
and
therefore commutes with any operator, in particular with its own
derivative, $\com{H_{\mathrm{sub}}(\theta_0)
}{\dot{H}_{\mathrm{sub}}(\theta_0) }=0$. This implies
that we thus saturate the bound:
\begin{equation}
\cqfi{H_{\mathrm{sub}}(\theta) }{\theta} \vert_{\theta=\theta_0}= t^2 \pnorm{\dot{H}(\theta_0)}{sn}^2\;.
\end{equation}
In full generality we can subtract $H(\theta_0)- Q$, with
$\com{Q}{\dot{H}(\theta_0)}=0$ and still saturate the bound. Below
we show for the example of the measurement of the direction of the
magnetic field how a locally vanishing Hamiltonian can be realized.
We now check how stable the method is if one does not subtract exactly
$\dot{H}(\theta_0)$ but rather $\dot{H}(\theta_0+\ensuremath{\varepsilon})$. We define the
extended Hamiltonian $H_{\mathrm{sub},\ensuremath{\varepsilon}}(\theta)\equiv H(\theta) -
H(\theta_0+\ensuremath{\varepsilon})$. To second order in $\ensuremath{\varepsilon}$ we have
\begin{equation}
H_{\mathrm{sub},\ensuremath{\varepsilon}}(\theta) = H(\theta)- H(\theta_0) -\ensuremath{\varepsilon} \dot{H}(\theta_0)-\ensuremath{\varepsilon}^2 \ddot{H}(\theta_0)/2 +\mc{O}(\ensuremath{\varepsilon}^3)\;.
\end{equation}
To obtain the channel QFI we need the eigenvalues of $\msc{H}_{\mathrm{sub},\ensuremath{\varepsilon}}$,
the
local generator corresponding to $H_{\mathrm{sub},\ensuremath{\varepsilon}}(\theta)$.
Using $ e^{X}Ye^{-X}=Y+\left[X,Y\right]+{\frac {1}{2!}}[X,[X,Y]]+\cdots $
in eq.\eqref{eq:der_exp}
and the fact that $H_{\mathrm{sub},\ensuremath{\varepsilon}}(\theta_0)
= -\ensuremath{\varepsilon} \dot{H}(\theta_0)-\ensuremath{\varepsilon}^2 \ddot{H}(\theta_0)/2 +\mc{O}(\ensuremath{\varepsilon}^3)$, we obtain, up to second order,
\begin{equation}
\msc{H}_{\mathrm{sub},\ensuremath{\varepsilon}} = t( \dot{H}(\theta_0) - \ensuremath{ \mathrm{i\,} } \frac{\ensuremath{\varepsilon}^2 t}{2} \Gamma +\mc{O}(\ensuremath{\varepsilon}^3) ) \;,
\end{equation}
where $\Gamma=[\ddot{H}(\theta_0,\dot{H}(\theta_0)]/2$.
We can now use
perturbation theory to see how the eigenvalues of $ \msc{H}_{\mathrm{sub},\ensuremath{\varepsilon}}$ are
affected by $\ensuremath{\varepsilon}$. In its eigenbasis, $\dot{H}(\theta_0)$ is written
$\dot{H}(\theta_0)=\sum_{i=1}^d e_i \ddens{i}$. We denote the
eigenvalues of $\msc{H}_{\mathrm{sub},\ensuremath{\varepsilon}}$ by $e_i^{(\ensuremath{\varepsilon})}$. Assuming
non-degenerate $e_i^{(\ensuremath{\varepsilon})}$, we have to first order perturbation
theory in $\ensuremath{\varepsilon}^2$
\begin{equation}
e_i^{(\ensuremath{\varepsilon})}= t e_i-\ensuremath{ \mathrm{i\,} }\frac{ \ensuremath{\varepsilon}^2t^2}{2}\moyvec{\Gamma}{i} +\mc{O}(\ensuremath{\varepsilon}^3)\;.
\end{equation}
Provided that $\ensuremath{\varepsilon}$ is small enough, no new degeneracies will
appear. If $e_1$ and $e_d$ are respectively the maximal and minimal
eigenvalue of $\dot{H}(\theta_0)$, then the channel QFI of the
extended channel is given up to second order in $\ensuremath{\varepsilon}$ by
\begin{multline}
\cqfi{H_{\mathrm{sub},\ensuremath{\varepsilon}}(\theta)}{\theta}=t^2 \pnorm{\dot{H}(\theta_0)}{sn}^2 \\-\ensuremath{ \mathrm{i\,} } \ensuremath{\varepsilon}^2 t^3(e_1-e_d)(\moyvec{\Gamma}{1}-\moyvec{\Gamma}{d})\;.
\end{multline}
This shows that errors of the order $\epsilon$ in the value of
$\theta_0$ lead to a channel QFI reaching the upper bound up to
a correction of order $\ensuremath{\varepsilon}^2$.
{ It was shown in \cite{yuan_optimal_2015} that one can also saturate the upper bound with a different method: by breaking the
evolution operator $U_{H(\theta)}$ in $m$ evolution operators with
evolution time $\tau=t/m$, and interspersing controls
$U_{\mathrm{c}}$ one can increase the channel QFI. An optimal
choice of controls leads to the saturation of the upper bound. }
Neither of the three methods
signal flooding, using additional interspersed controls, and
Hamiltonian subtraction
makes any use of
ancillas. This shows that in all generality, ancillas are not
\emph{necessary} to achieve the maximal
sensitivity when estimating a
Hamiltonian parameter. While this result was already known for phase
shift Hamiltonians
\cite{boixo_generalized_2007,fraisse_hamiltonian_2016},
these
methods show that it is the case for
{\em any} Hamiltonian.
\subsection{Engineering the time dependence of the channel QFI}\label{sec:scaling}
Hamiltonian extension can also be used to modify the behavior of the
channel QFI with time or other relevant resources.
It was shown in \cite{pang_quantum_2014,pang_erratum:_2016} that
in case the eigenvalues of the Hamiltonian do not depend on the
parameter to be estimated,
the QFI behaves periodically in $t$
(this discussion on $t$ applies actually to any
phase shift parameter). In general for $H(\theta)=\sum_{i=1}^d \lambda_i(\theta) \ddens{\psi_i(\theta)}$ (we assume for simplicity the Hamiltonian non-degenerate) the local generator is given \cite{pang_quantum_2014} by
\begin{multline}
\msc{H}= t \sum_{i=1}^d \frp{ \lambda_i(\theta)}{\theta} \ddens{\psi_i(\theta)}+ 2\sum_{k \neq l} \e{-\ensuremath{ \mathrm{i\,} } t \frac{ \lambda_k(\theta)- \lambda_l(\theta)}{2}}\\ \times \sin( \frac{ \lambda_k(\theta)- \lambda_l(\theta)}{2}) \pscal{\psi_l(\theta)}{\partial_\theta \psi_k(\theta)} \dens{\psi_l(\theta)}{\psi_k(\theta)}\;,
\end{multline}
clearly showing that if $ \frp{ \lambda_i(\theta)}{\theta}=0$,
\emph{i.e} if the eigenvalues of $H(\theta)$ are
$\theta$-independent, only the periodic term survives. The fact that
the channel QFI (and more generally the QFI) behaves only
periodically with the time prohibits the use of time as a
resource: We cannot increase the sensitivity to arbitrarily large
values by increasing the evolution time. This is particularly
harmful since quantum metrology typically provides a quadratic scaling with time
(for time-independent Hamiltonians), to be compared to the linear
scaling obtained by classical averaging.
To show how Hamiltonian extension can help to engineer the time
dependence we assume that the eigenvalues of the original Hamiltonian
$H(\theta)$ do not depend on $\theta$: $\lambda_i(\theta) \rightarrow
\lambda_i$. We then consider the Hamiltonian extension
\begin{equation}
H_V(\theta)=H(\theta)+ \ensuremath{\varepsilon} V\;.
\end{equation}
where $V$ can be any operator. For small $\ensuremath{\varepsilon}$ we can use the
perturbation theory to first order to find the perturbed eigenvalues
$\lambda_i^{\ensuremath{\varepsilon} V}$ of $H_V(\theta)$. Assuming non-degenerate
$\lambda_i$, we have
\begin{equation}
\lambda_i^{\ensuremath{\varepsilon} V}=\lambda_i +\ensuremath{\varepsilon} \moyvec{V}{\psi_i(\theta)}\;.
\end{equation}
We see that the introduction of $\epsilon V$ in the Hamiltonian makes
the eigenvalue depend on $\theta$, as long as
$\moyvec{V}{\psi_i(\theta)}$ is not constant as function of $\theta$.
Under this
condition this Hamiltonian extension
introduces a quadratic
scaling with time in the QFI. Despite the fact that for a general $V$
this Hamiltonian extension does typically not allow one to saturate the upper bound
it offers the advantage that one does not need to know the exact
value of the parameter to implement it.
{ The method of engineering the time dependence shows that in general,
not only
$\dot{H}(\theta_0)$ can be used to increase the channel QFI. One can
check case by case if adding another operator helps to increase the
best sensitivity.
For example it was shown for a broken phase shift
of the form $K=\theta G + \eta F$ that the channel
QFI is not always a monotone function of $\eta$ \cite{de_pasquale_quantum_2013}. Thus, for certain values of $\eta$ the channel QFI can be increased
by increasing $\eta$, an effect
that the authors call ``dithering''.}
\section{Example of applications}\label{sec:NV}
\subsection{NV center magnetometry}
The nitrogen-vacancy defects in diamonds, also known as NV centers,
correspond to defects in a diamond crystal lattice, where a
substitutional nitrogen atom comes
with a vacancy in one of the
neighbouring sites. Such NV centers exist in three forms, a neutral
one, a positively charged one, and a negatively charged one. The
latter provides a promising system for magnetometry, since it has a
spin triplet which can be monitored efficiently through optical
processes, and offers a coherence time that can be as high as a few
ms (see
\cite{schirhagl_nitrogen-vacancy_2014,rondin_magnetometry_2014} for
recent reviews).
Neglecting the interactions with the $^{14}$N nuclear spin as
well as
the bath of the $^{13}$C nuclear spins, the Hamiltonian $H_{\mathrm{NV}}$ for the triplet
state of the NV center can be written \cite{rondin_magnetometry_2014}
\begin{equation}\label{eq:ham_NV}
H_{\mathrm{NV}}= g \mu_\mathrm{B} ( B_x S_x +B_y S_y + B_z
S_z)/\hbar+ D S_z^2/\hbar + E (S_x^2-S_y^2)/\hbar\;,
\end{equation}
with $g$ the Land\'e factor, $\mu_\mathrm{B}$
the Bohr magneton, $D$ and $E$ the zero field splitting parameters and
$S_x, S_y$ and $S_z$ the dimensionless spin-1 matrices, fulfilling
$\com{S_i}{S_j}=\ensuremath{ \mathrm{i\,} } \hbar \ensuremath{\varepsilon}_{ijk}
\,S_k \;\forall \;i,j,k \in \lbrace
x,y,z \rbrace$. The zero field splitting
has two components, the axial one, with parameter $D$ (taken as
$D=2\pi\times 2.87\,$GHz), and the off-axis one, with parameter $E$ (taken as
$E=2\pi\times 5\,$MHz). The parameter that one seeks to estimate is $B_z$, the
magnetic field in the $z$ direction which is the direction of the
axis from N to V.
This Hamiltonian has the form of a "broken phase shift"
$H_{\mathrm{NV}}=B_z G +F$ with $G=g \mu_\mathrm{B} S_z/\hbar$.
The channel QFI is bounded by $(t/\hbar)^2\pnorm{G}{sn}^2 =(t g
\mu_\mathrm{B} /\hbar )^2 \pnorm{S_z/\hbar}{sn}^2
= 4(t g \mu_\mathrm{B} /\hbar)^2$.
Due to
the part which does not commute with $S_z$, \emph{i.e.}~the magnetic
transverse field and the off-axis zero field splitting, the channel
QFI decreases for small value of $B_z$, while for high values of
$B_z$ the channel QFI reaches the upper bound (see Figure
\ref{fig1}).
As we have seen in section \ref{sec:saturating_bound}, by adding the
derivative $\dot{H}_{\mathrm{NV}}$ of the Hamiltonian to the original
Hamiltonian we can saturate the upper bound. The shifted Hamiltonian is
\begin{equation}\label{eq:Ham_NV_ext}
\tilde{H}_{\mathrm{NV}}=H_{\mathrm{NV}} +\beta
\dot{H}_{\mathrm{NV}}=H_{\mathrm{NV}} + \beta g \mu_B S_Z/\hbar\;.
\end{equation}
We represent in Figure~\ref{fig1} the effect of $\beta$ on the
channel QFI. We see that by increasing the value of $\beta$ we {can get
arbitrarily close}
to the upper bound.
As pointed out when
discussing the "broken phase shift", the effect of $\beta$ amounts
here to evaluating the QFI at shifted value of the parameter
(at $\theta+\beta$ instead of at $\theta$).
For magnetometry with NV-centers {it is actually already known that
adding an additional magnetic field
can help to measure weak magnetic
fields}. In \cite{rondin_magnetometry_2014} this was discussed in the
the context of reaching a linear Zeeman effect. In addition,
adding a bias field makes already sense from the perspective of
shifting the precession signal up to higher frequency, where it can be
distinguished from noise more easily. Here we see that independently
of such specific considerations, ``flooding the signal''
by adding the known parameter-derivative of the Hamiltonian with a
large factor is a very general method that allows one to overcome
pernicious effects of other parts of the
Hamiltonian and reach maximal possible sensitivity.
\begin{figure}
\includegraphics[scale=0.25]{NV_signal_flood}
\caption{Channel QFI for the NV center
($t=10^{-3}$s, $B_x=10^{-1}$T and $B_y=0$T), as a function of the
parameter to be estimated, $B_z$. The dotted line represents the
upper bound, the dashed line the channel QFI of the original Hamiltonian \eqref{eq:ham_NV} and the
continuous line the channel QFI of the extended channel
\eqref{eq:Ham_NV_ext} for different values of $\beta$: From bottom to top we have $\beta=10^{-6}$, $\beta=10^{-3}$ and
$\beta=10^{-1}$. }\label{fig1}
\end{figure}
The behaviour of the channel QFI for very low values of $B_z$ deserves
some more comments. As it can be seen in Figure \ref{fig1} when $B_z$
becomes very small the channel QFI reaches a plateau. This is a quite
general feature for Hamiltonians of the form of a "broken phase shift"
$K(\theta)=\theta G + F$. The channel QFI $\cqfi{K(\theta)}
{\theta}\vert_{\theta=0}$
at $\theta=0$ can be obtained by calculating
$\msc{K}\vert_{\theta=0}$ which reads
\begin{equation}\label{eq:local_generator_theta_0}
\msc{K}\vert_{\theta=0}=t\sum_{i}g_{ii}\ddens{i}+\ensuremath{ \mathrm{i\,} } \sum_{i\neq j} g_{ij} \frac{1-\e{\ensuremath{ \mathrm{i\,} } t ( f_i-f_j)}}{f_i-f_j} \dens{i}{j}\;,
\end{equation}
with $ F = \sum_i f_i \ddens{i}$, where the $f_i$s and the
$\ket{i}$s are respectively the eigenvalues (assumed non-degenerate
here)
and eigenvectors of $F$,
and $G=\sum_{i,j} g_{ij} \dens{i}{j}$, with $g_{ij}^*=g_{ji}$. To get
more insight into this expression
notice that the channel QFI vanishes only when maximal and minimal
eigenvalues of $\msc{K}$ coincide,
{\em i.e.}~when $\msc{K}$
is proportional to the
identity operator. In general this will be the case
if $t(f_i-f_j) $ is an
integer and $g_{11}=g_{dd}$.
This condition
is not necessary
since $G$
can be
sparse and therefore some $g_{ij}$ may already be equal to zero. Still,
this simple analysis shows that particular cases excepted,
it is a
quite general feature that the channel QFI for a broken phase shift
does not vanish for
small values of the parameter.
\subsection{Estimation of a direction of a magnetic field}
The estimation of a component
of a magnetic field using a NV center studied in
the previous section
leads to a broken phase shift. We now consider
the estimation of one of the spherical angles characterizing the direction of a
magnetic field with a free spin-1 as a probe, a situation
that does not correspond to a broken
phase shift. The Hamiltonian
is given by
\begin{equation}\label{eq:Ham_direction}
H(B,\theta,\ensuremath{\varphi}) =g \mu_{\mathrm{B}} \mathbf{B}\cdot
\mathbf{S}/\hbar\;,
\end{equation}
with $ \mathbf{B}= B(\sin(\theta) \cos(\ensuremath{\varphi}),\sin(\theta) \sin(\ensuremath{\varphi}),
\cos(\theta) )$
and $\mathbf{S}=(S_x,S_y,S_z)$.
We want to estimate the parameter $\theta$ in a scalar parameter
setting (we consider $B$ and $\ensuremath{\varphi}$ as known). The channel QFI for
the corresponding channel is bounded as
\begin{equation}
\cqfi{H(B,\theta,\ensuremath{\varphi})}{\theta}\leq \left(t / \hbar\right)^2
\pnorm{\frp{H(B,\theta,\ensuremath{\varphi})}{\theta}}{sn}^2\;. \label{39}
\end{equation}
The eigenvalues of $H(B,\theta,\ensuremath{\varphi}) $ are $0$ and $\pm g
\mu_{\mathrm{B}} B$, from
which we get $\pnorm{\frp{H(B,\theta,\ensuremath{\varphi})}{\theta}}{sn}^2 = 4 (g
\mu_{\mathrm{B}} B)^2$. The local generator $\msc{H}_\theta$ of the
translation in $\theta$ can be computed exactly and its eigenvalues
are $0$ and $\pm 2 \sin(g \mu_{\mathrm{B}}Bt/(2\hbar))$. This gives a
channel QFI equal to
\begin{equation}
\cqfi{H(B,\theta,\ensuremath{\varphi})}{\theta}=16 \sin^2( g \mu_{\mathrm{B}}Bt/(2\hbar)) \;.
\end{equation}
We see that the channel QFI has a periodic time dependence since
the
eigenvalues of $H(B,\theta,\ensuremath{\varphi}) $ are $\theta$-independent (see
dashed line in Figure \ref{fig:direction_plot}). However, the upper
bound still scales quadratically with time. As a result, for large
time, the discrepancy between the actual channel QFI and its upper
bound increases. Notice that for this Hamiltonian the role of
time $t$ in the channel QFI is the same as the role of the strength $B$ of
the magnetic field.
\subsubsection{Signal flooding}
We now examine
how "signal flooding" helps to increase the channel QFI. We consider the extended Hamiltonian
\begin{equation}\label{eq:Ham_sig_flod_direction}
H_{\mathrm{fl}}(B,\theta,\ensuremath{\varphi}) = H(B,\theta,\ensuremath{\varphi}) +\beta \frp{H(B,\theta,\ensuremath{\varphi})}{\theta}\bigg\vert_{\theta=\theta_0}\;.
\end{equation}
We represent in the Figure \ref{fig:direction_plot} the effect of
signal flooding by plotting the channel QFI of
$H_{\mathrm{fl}}(B,\theta,\ensuremath{\varphi})$ for different values of $\beta$. In the same
way as for the NV center Hamiltonian we see that increasing the value
of $\beta$ allows one to get arbitrarily close to the upper bound.
In contrast to
the estimation of a broken phase shift (\emph{e.g.}~estimation of $B_z$ for the NV center), here signal flooding does not
correspond to a
shift of the value of $\theta$.
\begin{figure}
\centering\includegraphics[scale=0.25]{direction_plot}
\caption{Channel QFI for a direction of the magnetic field ($B=10^{-9}$T, $\ensuremath{\varphi}=\pi /4$ and $\theta=\pi /3$), as a function of the time. The dotted line represents the
upper bound, the dashed line the channel QFI of the original Hamiltonian \eqref{eq:Ham_direction}. The
continuous
line represents the channel QFI for the "signal flooding" Hamiltonian \eqref{eq:Ham_sig_flod_direction} for different values of $\beta$: From bottom to top we have $\beta=0.2$, $\beta=0.75$ and
$\beta=5$. The dotted-dashed line represents the channel QFI for the extended Hamiltonian \eqref{eq:Ham_quad}, for different values of $\kappa$: From bottom to top we have $\kappa=1$, $\kappa=10$ and $\kappa=10^9$. }\label{fig:direction_plot}
\end{figure}
\subsubsection{Time engineering}
Having eigenvalues independent of $\theta$, $H(B,\theta,\ensuremath{\varphi})$ leads to a channel QFI with a periodic time scaling. In line with section \ref{sec:scaling} we now show how we can restore the quadratic time scaling using a Hamiltonian extension. We consider the extended Hamiltonian
\begin{equation}\label{eq:Ham_quad}
H_{S_z} (B,\theta,\ensuremath{\varphi})=H(B,\theta,\ensuremath{\varphi}) +\kappa B g \mu_{\mathrm{B}}
S_z/\hbar\;,
\end{equation}
which correspond to the original Hamiltonian with an additional
magnetic field in the $z$ direction with a strength $\kappa B$.
The eigenvalues of $H_{S_z} (B,\theta,\ensuremath{\varphi})$ are $0$ and $\pm g
\mu_{\mathrm{B}} \sqrt{B^2+\kappa^2 +2B \kappa \cos(\theta)}$, showing
that the additional magnetic field makes the eigenvalues depend
on $\theta$, and therefore allow one
to create a quadratic time scaling. We have represented in Figure
\ref{fig:direction_plot} the channel QFI of the extended Hamiltonian
for different values of $\kappa$. This clearly shows that the
additional magnetic field introduces a quadratic scaling, but also
that the larger $\kappa$, \emph{i.e.}~the strength of the
additional field, the larger is the channel QFI. In contrast
to signal flooding, for very large values of $\kappa$ we do not
reach the upper bound (compare the case $\beta = 5$ and $\kappa=
10^9$): Indeed $H_{S_z} (B,\theta,\ensuremath{\varphi})$ do not fulfil the condition
of Lemma \ref{sat} and when it dominates completely the Hamiltonian we
do still not reach the upper bound.
\subsubsection{Hamiltonian subtraction}
\begin{figure}
\centering\includegraphics[scale=0.25]{ham_sub}
\caption{Effect of a perturbation in "Hamiltonian subtraction" for the
estimation of a direction of a magnetic field. ($B=10^{-9}$T,
$\ensuremath{\varphi}=\pi /4$, $\theta_0=\pi /3$ and $t=10^{-2}$s). The dotted line
represent the upper bound and the dashed line the channel QFI of the
original Hamiltonian \eqref{eq:Ham_direction}. The plain line
represent the channel QFI of the Hamiltonian
\eqref{eq:Ham_sub_magnetic}.}.\label{fig:sub_ham}
\end{figure}
Finally, we show how we saturate the upper bound using Hamiltonian
subtraction. Hamiltonian subtraction here amounts to adding a magnetic
field of the same strength but
in opposite direction to the original magnetic field. To
see the effect of a
slight deviation from the correct direction of the added field,
we study
the perturbed Hamiltonian
\begin{equation}\label{eq:Ham_sub_magnetic}
H_{\mathrm{sub},\ensuremath{\varepsilon}}(B,\theta,\phi)=H(B,\theta,\phi)-H(B,\theta_0+\ensuremath{\varepsilon},\phi)\;.
\end{equation}
We can
calculate the channel QFI exactly,
\begin{multline}\label{eq:cQFI_pert}
\cqfi{H_{\mathrm{sub},\ensuremath{\varepsilon}}(B,\theta,\phi)}{\theta}\vert_{\theta=\theta_0}=4(g \mu_\mathrm{B}
B \frac{t}{\hbar} )^2 \cos^2(\frac{\ensuremath{\varepsilon}}{2} )\\ + 4 \sin^2(g
\mu_\mathrm{B} B \frac{t}{\hbar} \sin(\frac{\ensuremath{\varepsilon}}{2}))\;.
\end{multline}
One verifies
that for values of $\ensuremath{\varepsilon}=0$ this
saturates the upper bound \eqref{39}.
It
is interesting to notice that the correction of second order in $\ensuremath{\varepsilon}$
in the channel QFI vanishes exactly, and the leading order corrections
are of order $\ensuremath{\varepsilon}^4$, demonstrating the stability of the method to
perturbation in the direction of the subtracted magnetic field, as
represented in Figure \ref{fig:sub_ham}.
\section{Conclusion} \label{ccl}
We investigated the problem of single Hamiltonian parameter
estimation, and the effect of Hamiltonian extension on the precision
with which one can estimate the parameter. It was already known that
in the case of a phase shift, Hamiltonian extension does not lead to
an increase of the channel QFI (QFI optimized over all input
states) \cite{fraisse_hamiltonian_2016,boixo_generalized_2007}.
But for more general Hamiltonians, Hamiltonian extension
may increase the channel QFI. In particular, we found two ways of
engineering the Hamiltonian to saturate the upper bound of the
channel QFI: ($i$)"Signal flooding" which consists of adding to the
original Hamiltonian a large term proportional to its
derivative; and $(ii)$ "Hamiltonian subtraction" which consists in
subtracting the Hamiltonian at a fixed value of the parameter from the
original Hamiltonian.
Neither method makes use of any
ancillas, showing that adding subsystems is in general not necessary
for unitary parameter estimation to achieve the best precision.
We applied "signal flooding" to the Hamiltonian of an NV
center. Such systems are used to measure small magnetic fields. We
showed how adding a
magnetic field in the $z$-direction
helps to increase the maximal possible sensitivity of the
measurement of
the magnetic field's component in the same
direction.
We also illustrated both methods with the estimation of a direction
of a magnetic field.
Finally we showed that in cases where the eigenvalues of the original
Hamiltonian are parameter-independent, {adding almost any constant
Hamiltonian} can lead to a quadratic increase of the channel QFI with
time, whereas for the original Hamiltonian it is bounded and periodic
in time. Such is the
situation for the measurement of the polar angle of the magnetic
field, probed with a free spin. This will typically not enable
saturation of the bound of the channel QFI, but can nevertheless be
very advantageous for large measurement times, and relatively straight-forward to
implement.
The scenario considered in this work is an ideal one.
Our figure of merit, the channel QFI, constitutes a valid
and achievable benchmark for quantum parameter estimation
that is difficult to reach
in practice.
Even if the theory gives us the optimal POVM and
the optimal state, they may be hard to implement.
Therefore, it may be interesting to see to what extend
Hamiltonian extensions
also work in
situations where the optimal POVM or the
optimal state cannot be implemented.
\FloatBarrier
{\bf Acknowledgment} We thank Fabienne Schneiter for useful discussions.
|
\section*{Introduction}
The field of DNA-mediated particle assembly has undergone remarkable progress over recent years\cite{jones2015programmable}, owing, at least in part, to its potential as a powerful platform for rational, bottom-up design and engineering of complex materials, and motivated by recent successful translations into applications as diverse as sensing\cite{Barnaby2015}, photonics\cite{Ross2015}, and catalysis. The growing number of synthetic pathways and design strategies to fabricate DNA-functionalized particles (DFPs) has led to the development of a diverse palette of tailorable building blocks from which to choose, comprised of particles of a wide range of inorganic to organic compositions, a near continuum of particle sizes spanning nanometers to micrometers, precise DNA sequence control and thus tailorable hybridization, diverse chemistries for DNA grafting/association, and fine tunability of the grafting density.\cite{Kim2006,zhang2013general,Oh2015}
Accompanying this expanding diversity of building blocks has been a parallel development of specific to generalized design principles that have begun to link molecular-scale DFP function with mechanisms of assembly and the resulting uni- or multi-modal crystalline structures. To this end, the growing combination of theory, simulations, and experiments, has helped to overcome some of the challenges in the field. For example, re-entrant melting strategies\cite{Angioletti2012, Rogers2015} have been successfully developed to alleviate the very narrow temperature ranges for efficient crystallization of DFPs.
The most common route to induce attraction between DFPs, and thus program their assembly, leverages the direct or indirect (i.e., with additional DNA linker strand) hybridization of complementary DNA strands tethered separately to two types of particles. Under suitable conditions in such systems, particles with complementary DNA functionality (i.e., `unlike' particles) form attractive contacts among multiple strands of hybridizable DNA, whereas particles bearing the same DNA functionality (i.e., `like' particles) typically interact via repulsive, non-hybridizable DNA-mediated steric interactions. By tailoring DFP properties such as particle size, DNA sequence, DNA strand length, and strand grafting density, three-dimensional (3D) assembly of a diversity of stoichiometric and symmetry structures including CsCl, AlB$_2$, Cr$_3$Si, Cs$_6$C$_{60}$, NaTl, and others has been demonstrated for nanoparticle systems~\cite{Nykypanchuk2008, Macfarlane2011, Li2012,Auyeung2014,Cigler2010}.
In these cases, assembly is ultimately tailored by controlling properties of particles that are each functionalized with only a single type of ssDNA. As an alternative, mixtures of two complementary DNA strands in a desired ratio can be grafted to particles as a means for tuning pair-interactions. This, so-called multiflavoring scheme was proposed to selectively program 3D crystallization of DFP into body-centered cubic (BCC) or closed-packed structures (CP)~\cite{Scarlett2011,Casey2012}. Furthermore, Zhang and coauthors showed that such muliflavored DFPs can also be utilized as reconfigurable systems that are capable of dynamically transforming into BCC and CP lattices with the addition of suitable DNA sequences~\cite{Zhang2015}. This suggests that control over interparticle interactions may serve as a new handle for dramatically expanding the structural diversity of crystalline assemblies and for realizing novel smart and adaptable materials.
\begin{figure}[h!]
\center
\includegraphics[width = 6.5 in]{Figure1_revised.pdf}
\caption{Multiflavoring scheme. Programming the particle interactions through functionalizing with a blend of conjugate DNA strands, $\alpha$DNA (depicted as red strands) and $\beta$DNA (depicted as blue strands) at desired ratios. The three resulting pair interactions ($E_{AA}, E_{BB}, E_{AB}=E_{BA}$) can be independently controlled as shown in the quadrant of panels.
}
\label{fig1}
\end{figure}
Although nanoparticle-based DNA-mediated assembly enables realization of a diversity of crystalline structures, the assembly, especially in two-dimensions, of micron-sized DFPs has been slow to come. The appeal of micron-sized particle assemblies, and the motivation of efforts to overcome this barrier, is driven from an applications perspective by their desirable optical properties~\cite{Park2014}, and, from a fundamental science perspective, by the ability to employ simpler optical monitoring for direct mechanistic insight into DFP crystallization~\cite{Gasser2001}. Micron-sized DFPs possess relatively short-range interactions compared to their sizes, and extremely narrow melting transition ranges (1-2$^{o}$C), which leads to tedious annealing protocols with many days of incubation for successful crystallization ~\cite{Biancaniello2005,Kim2006,Dreyfus2010,DiMichele2013,Wang2015,Wang2015i}. Recently, it was shown that DNA strand displacement can be used to widen the melting transition range and to shorten the effective crystallization time~\cite{Rogers2015}. Previous studies on the 2D assembly of DNA-tethered colloids have sought to better understand their association/dissociation transitions~\cite{Dreyfus2009,Dreyfus2010,song2016effect}, and have relied on surface-mediated templating of DFP monolayers~\cite{Hartmann2002,Zou2005,Puchner2008,Shyr2008},
but the direct formation of DNA-mediated binary 2D superlattices remains a serious challenge.
The current paradigm for binary superlattice formation primarily leverages entropic packing effects to generate structures with tailored lattice symmetries. This represents a natural extension of extensively characterized binary hard sphere-like mixtures~\cite{kung2014template} as well as ionic crystals~\cite{leunissen2005ionic}.
Here, we present an enthalpic design strategy, which takes advantage of tuning the strength of interparticle attractions between equally-sized, but distinctly labeled particles in a binary (A, B) mixture. Interactions between like (AA and BB) and unlike (AB) particle pairs are independently controlled to guide their assembly into compositionally ordered lattices. Extensive computer simulations enable comprehensive investigation of this binary assembly, leading to the prediction of compositionally ordered two-dimensional (2D) structures such as square, alternating string, honeycomb, Kagome and square Kagome that we have verified experimentally by leveraging a previously proposed multiflavoring approach~\cite{Casey2012, Scarlett2011} to tune interparticle interactions among micron-sized DFPs by judicious blending of complementary ssDNA functionality as shown in Figure~\ref{fig1}.
First, we have performed molecular dynamics (MD) simulations to investigate the two-dimensional (2D) assembly of binary micron-sized DFPs over a wide range of possible combinations in their interaction strengths (see Figure 1 for details).
For simplicity in presenting and interpreting results, we keep unlike pair interaction strength, $E_{\mathrm {AB}}$, fixed and vary like pair interaction strengths, $E_{\mathrm {AA}}$ and $E_{\mathrm {BB}}$, independently. All of the results are presented as a function of normalized pair energies, $E_{\mathrm {AA}}$/$E_{\mathrm {AB}}$ and $E_{\mathrm {BB}}$/$E_{\mathrm {AB}}$.
To model the short-range interparticle interactions mediated by DNA interactions between micron-sized colloidal particles, we use a suitable functional form (see Supplementary Information (SI) and Fig. S1 for details) that captures important details of the underlying repulsive (particle-particle and osmotic repulsion due to overlap between DNA chains) as well as attractive (due to DNA hybridization) parts of the potential. Similar pair potentials between DFPs and other complex systems have been used in the past to successfully capture their self-assembly properties~\cite{rabideau2007computational,tkachenko2002morphological,martinez2011design, Auyeung2014,mahynski2015grafted}.
To identify suitable conditions for superlattice formation, we conduct simulations over a large range of temperatures and identify putative melting transition temperatures based on the changes in potential energy as a function of temperature (Fig. S2). At temperatures close to the melting transition, sufficiently large crystalline assemblies are observed that can be used for further analysis. We note that most of the simulation results reported in this paper are based on systems at relatively low number density ($\rho$ = 0.10) to identify design parameters suitable for enthalpically-driven assembly of binary superlattices. To identify different lattice symmetries and underlying compositional order in the binary superlattices that are formed, we use three complementary methods for which details are provided in the SI: (i) nearest neighbor analysis (NNA) (SI Figs. S3), (ii) common neighbor analysis (CNA)~\cite{Stukowski2012} (SI Fig. S4), and (iii) visual inspection using visual molecular dynamics (VMD)~\cite{Humphrey1996}.
Extensive results from MD simulations are summarized in Figure~\ref{fig2} as order diagrams depicting 2D binary crystal symmetries resulting from tuned interparticle interactions as a function of $E_{\mathrm {AA}}$/$E_{\mathrm {AB}}$, $E_{\mathrm {BB}}$/$E_{\mathrm {AB}}$.
NNA (Table 1) enabled interpretation of structural symmetries from computational results, which was confirmed by CNA and visual inspection. These data show that various lattice symmetries (square, hexagonal, rhombic) can be obtained by simply changing $E_{\mathrm {AA}}$/$E_{\mathrm {AB}}$ and/or $E_{\mathrm {BB}}$/$E_{\mathrm {AB}}$, as depicted in the snapshots shown in Figure~\ref{fig2}(A-E), of representative binary crystal structures. For low values of like pair energies, one can expect to form non-close-packed square lattice structures as quantified by NNA and depicted as the hatched area in Fig.~\ref{fig2}(a-b). Such structures are otherwise difficult to obtain without introducing additional factors such as surface templating~\cite{ferraro2014graphoepitaxy}.
Previous investigations of 3D assembly of DFPs, where like pair interactions are commonly purely repulsive ($E_{\mathrm {AA}}$/$E_{\mathrm {AB}}$ = $E_{\mathrm {BB}}$/$E_{\mathrm {AB}}$ = 0), have also found that non-closed-packed BCC structures are stabilized~\cite{Li2012}.
As shown in Fig.~\ref{fig2}a, the predicted two-dimensional square lattices display a high degree of compositional order in terms of the coordinated arrangement of A and B particles. Further increasing like-pair energies leads to the formation of close-packed hexagonal lattices, with an interesting rhombic structure formed for the case specified as E in Fig.~\ref{fig2}, and identified only visually. The rhombic structure is of the A(AB$_2$) type in this case.
\begin{figure}[h!]
\center
\includegraphics[width = 6.5 in]{Fig2_JM.pdf}
\caption{Order diagram showing a variety of 2D binary crystals controlled as a function of interparticle binding strengths. The color coded lines drawn on the crystal images (A-E) are guides to the eye for illustrating the corresponding lattice symmetries. Color maps show the identification of the crystals using \textbf{(a)} $N_{BB}-N_{AA}$ and \textbf{(b)} $N_{AB}$ as the order parameter.}
\label{fig2}
\end{figure}
\begin{table}[h!]
\begin{center}
\begin{tabular}{ c c c c }
\hline
{\bf Crystal} & {\bf N$_{\mathrm {AB}}$} & {\bf N$_{\mathrm {AA}}$} & {\bf N$_{\mathrm {BB}}$} \\
\hline
Square & 4 & 0 & 0 \\
Alternating string & 4 & 1 & 1 \\
Honeycomb & 4 & 0 & 2 \\
Kagome & 3 & 0 & 3 \\
Square Kagome & 3 & 0 & 3 \\
\hline
\end{tabular}
\caption{Reference nearest neighbor counts for perfect two-dimensional binary superlattices, where N$_{\mathrm {AB}}$ is the number of unlike contacts and N$_{\mathrm {AA}}$ or N$_{\mathrm {BB}}$ is the number of like contacts.}
\label{table1}
\end{center}
\end{table}
For hexagonally packed structures, we identify conditions for many different binary superlattices, namely alternating strings, honeycomb, Kagome, and square Kagome (Fig.~\ref{fig2}(B-D)).
In a small region of moderate like pair energies, hexagonal structures are compositionally ordered as alternating strings (AS) (Fig.~\ref{fig2}B). Detailed analysis of MD trajectories (not shown here) suggest the formation of AS lattices due to the transformation of nuclei, initially grown as square into hexagonal lattices (see SI movie). Such diffusionless transformations have previously been observed in the context of DFPs in 3D systems~\cite{casey2012driving}, but the underlying mechanism is still not fully resolved~\cite{jenkins2014hydrodynamics}. We are currently investigating this issue in more detail, which will be discussed in a future publication.
With increasing like pair energies along the diagonal axis ($E_{\mathrm {AA}}$/$E_{\mathrm {AB}}$ = $E_{\mathrm {BB}}$/$E_{\mathrm {AB}}$), hexagonal structures become compositionally disordered, which is expected as the energetic difference in interparticle interactions becomes smaller. Interestingly, when $E_{\mathrm {AA}}$/$E_{\mathrm {AB}} > E_{\mathrm {BB}}$/$E_{\mathrm {AB}}$, hexagonal structures can be found in many compositionally ordered arrangements such as honeycomb, Kagome and square Kagome lattices. As evident in Fig.~\ref{fig2}a, honeycomb arrangement is quite prevalent in a large region of the order diagram.
Kagome and square Kagome arrangements are observed when $E_{\mathrm {AA}}$/$E_{\mathrm {AB}} \ll E_{\mathrm {BB}}$/$E_{\mathrm {AB}}$ $\approx$ 1. In these cases, contacts between unlike pairs (AB) are only slightly more favored energetically than the like pair (BB), which gives rise to the formation of binary superlattices with non-equimolar A:B stoichiometries (e.g., 1:2 for honeycomb and 1:3 for Kagome).
These simulation results ultimately demonstrate that simple tuning of interparticle attractions can lead to the formation of a diverse array of binary superlattice structures in 2D systems
without requiring blending of particles of different sizes, application of external fields, or introduction of surface structuring or other external factors. In order to test the validity of these results as possible guidelines for rational design of binary superlattices, we have carried out complementary experimental studies of 2D particle assembly wherein blending of complementary ssDNA particle functionality was used as a means for controlling interparticle interactions and mediating assembly of distinct fluorescently labeled particle populations.
Specifically, like-sized (1.5 $\mu$m) red (A) and green (B) fluorescent silica particles were separately functionalized with pure or blended complementary single-stranded $\alpha$DNA and $\beta$DNA using sequential silanization and cyanuric chloride chemistries as reported previously\cite{song2016effect}. Their assembly was followed via optical and fluorescence microscopy. While the blending ratio, $\gamma_i=\alpha$DNA/($\alpha$DNA+$\beta$DNA), was tuned independently between pure $\alpha$DNA and pure $\beta$DNA for B-type particles ($\gamma_B=0-1$), A-type particles were functionalized with pure $\alpha$DNA ($\gamma_A=1$) as a way to experimentally traverse the most relevant parameter space mapped out by our computations ($E_{AA}/E_{AB}=0, E_{BB}/E_{AB}=[0,1]$). This allowed for programming of attractive interaction strengths between like ($E_{\mathrm {AA}}$, $E_{\mathrm {BB}}$) and unlike ($E_{\mathrm {AB}}$) particles (Fig. \ref{fig1}).
Our analysis suggests that increases in the relative loading of $\alpha$DNA on the B-type particles should lead to a monotonic decrease in the strength of attractive interactions between unlike particles ($E_{AB}=E_{BA}$) while the strength of attraction between B-type particles ($E_{BB}$) should monotonically increase. Such tunability of interparticle interaction strengths should ultimately result in a scenario where interactions between one set of like particles (B) become equivalent to or even exceed interactions between unlike particles.
As an initial confirmation of the tunability of the interparticle attraction, we measured the melting transition of both unary (only B-type) and binary (both A- and B-type) DFPs at different $\gamma_B$ while holding $\gamma_A=1$. With increasing $\gamma_B$, the melting temperature of the unary mixture (Supplementary Fig. S5) shifts to lower temperatures, whereas that of the binary mixture shifts to higher temperatures. These data clearly demonstrate that with increasing $\gamma_B$ the strength of interactions between B-type particles systematically decreases while the strength of interactions between unlike particles systematically increases.
\begin{figure}[h!]
\center
\includegraphics[width = 3.25 in]{Fig3_MAS.pdf}
\caption{Identification of structural order within 2D DFP lattices derived from solutions of different particle number ratios, $n$, and DFPs with different DNA blending ratios, $\gamma_B$ ($\gamma_A=1$). The phase diagram shows the characteristic structures observed under the specified conditions, including square (grey squares), pentagonal (white pentagons), and and hexagonal (black hexagons) lattices for which the lattice symmetry group or mixtures of groups are depicted with concomitantly shaped symbols. Particle assembly images show representative snapshots of square, pentagonal, and hexagonal lattices.
}
\label{fig3}
\end{figure}
Beyond determination of the melting temperature, experimental characterization focused on identification of crystal structures emerging upon self-assembly of the DFPs with systematically controlled ssDNA blending ratios $\gamma_B$ ($\gamma_A=1$) and various particle number ratios, $n=N_B/N_A$, where $N_{\mathrm A}$ and $N_{\mathrm B}$ are the number of A and B particles in solution, respectively. The tuning of of $n$ was motivated by previous studies wherein it was shown that solution stoichiometry can have a significant impact on the formation of three-dimensional binary superlattices formed by DNA-mediated interactions~\cite{vo2015stoichiometric}.
Changing the solution stoichiometry away from 1:1 (A:B) tends to favor structures with symmetries that closely resemble the stoichiometry of the solution itself. For example AB$_2$ and AB$_3$ solid stoichiometric structures may be favored over AB for 1:2 and 1:3 binary mixtures, respectively.
Here, we have assembled 2D structures from solutions of A and B DFPs of specified $n$, varied from 0.25 to 4. Under these conditions, the ssDNA blending ratio for B particles, $\gamma_B$, was systematically varied while A particles were functionalized only with $\alpha$DNA ($\gamma_A=1$). After the completion of thermal annealing, images of the assembled DFPs were collected at room temperature using an inverted optical microscope. Local structural analysis and calculation of pair correlations for broader fields of particles was carried out to identify the formation of specific 2D crystal structures, to determine their symmetry, and, by way of combined analysis of bright field and fluorescence images, to evaluate the compositional (i.e., A-B) order of the structures.
The structural order of 2D assemblies of A and B DFPs as a function of $\gamma_{\mathrm B}$
and $n$
are summarized in Fig. \ref{fig3}. The symbol shapes reflect the observed crystalline lattices ranging among square, pentagonal, and hexagonal symmetries as well as mixtures thereof. As $\gamma_{\mathrm B}$
is increased with $n$ held constant (e.g., $n=1$), we observe the formation of square lattice configurations at the lowest blending ratios considered ($E_{\mathrm {BB}}$ much less than $E_{\mathrm {AB}}$), the onset of minority hexagonal structures at higher blending ratios, eventual transition to majority hexagonal structures, and final dominance by hexagonal lattices at the highest blending ratios ($E_{\mathrm {BB}}$ comparable to $E_{\mathrm {AB}}$). These results are quite consistent with the computationally-derived order diagram presented in Fig.~\ref{fig2} with the exception of the fact that we also observe the formation of a minority phase of pentagonal lattices for intermediate blending ratios, which was not expected from the simulation results. The exact reasoning behind the appearance of such lattices in our experimental systems is somewhat unclear, but this may derive from the difference in local concentration of A and B particles relative to the prescribed bulk solution stoichiometry itself.
As interparticle interactions between A particles are purely repulsive ($\gamma_{\mathrm A}=1$),
increasing the amount of B particles in the solution may favor the formation of lattices with a higher number of BB contacts. We generally observe such a change with increasing $n=N_{B}/N_{A}$. For example, as $n$ is increased while holding $\gamma=0.4$, the dominance of square lattice structures at the lowest $n$ studied gives way to structures that are largely hexagonal. Here, again, we have identified persistent minority petagonal structures for $n$ greater than or equal to 1. In addition, we find that underpopulation of B particles in the solution ($n=0.25$) can lead to the formation of pure pentagonal structures. This is likely due to the higher local availability of B particles with respect to the solution stoichiometry, but inadequacy at the same time to stabilize hexagonal lattice structures. Entropic packing effects undoubtedly contribute to this observation as well, but these are difficult to decouple from the role of stoichiometry in the lattice selection.
In short, our experimental data validate the computational findings, demonstrating that the formation of specific lattice structures can be tailored simply by the combined design of the strength of interparticle interactions through control of the ssDNA blending ratio, $\gamma_B$, and the particle number ratio, $n$, as opposed to surface patterning or other external factors. In addition to structural order, however, we are interested in assessing how well compositional order can also be controlled. Specifically, we employ multi-channel red and green fluorescence imaging to differentiate A and B particles, respectively, within DFP assemblies to determine the achievable compositional coordination and symmetry (Fig. \ref{fig4}).
We find generally that compositional order can, indeed, be tuned by controlling the blending ratio $\gamma_i$, and thus interparticle interaction strength as wel as the particle stoichiometry, $n$. Moreover, unlike the simulations, simultaneous control over stoichiometry, $n$, and interparticle attraction strength may be critical for realizing higher symmetry lattices such as Kagome and square Kagome, and for controlling polymorphism.
Fig.~\ref{fig4} depicts the diversity of achievable compositional order and symmetry accompanying the previously discussed structural tunabilty.
For low $\gamma_B$ ($\gamma_A=1$), particles preferentially organize into square lattices. In those structures, A and B particles are compositionally well-ordered as shown in (Fig. \ref{fig4}a) for $n=1$. This is fully consistent with the particle arrangement observed in the simulations, as very few compositional defects are seen for square lattices.
\begin{figure*}[h!]
\center
\includegraphics[width = 6.5 in]{Fig4_MAS.pdf}
\caption{\footnotesize Identification of compositional order in representative fluorescence images of 2D DFP crystals comprised of A (dark) and B (bright) particles. Insets show B-specific fluorescence images (i.e., green channel) associated with specified regions (dashed boxes) of the crystal structure comprised of
\textbf{(a)} square lattices ($\gamma_B=0.2$, $n=1$),
\textbf{(b)} alternating strings ($\gamma_B=0.2$, $n=1$),
\textbf{(c)} honeycomb structures ($\gamma_B=0.3$, $n=2$),
\textbf{(d)} Kagome and honeycomb structures ($\gamma_B=0.3$, $n=2$),
\textbf{(e)} square Kagome and honeycomb structures ($\gamma_B=0.3$, $n=2$),
\textbf{(f)} demixed A and B particles ($\gamma_B=0.4$, $n=2$)
\textbf{(g)} pentagonal structures ($\gamma_B=0.2$, $n=0.25$), and
\textbf{(h)} honeycomb structures ($\gamma_B=0.2$, $n=4$). Scale bars in each image represent 10$\mu$m.}
\label{fig4}
\end{figure*}
As shown in Fig. 3, one can observe hexagonal lattices by either increasing the like-pair energies or $n$ (e.g., Fig. \ref{fig4}h). If we hold $n$ constant at 2, and change the blending ratio (i.e., to increase like-pair attraction between B particles), we observe compositional ordering into alternating string structures (Fig. \ref{fig4}b), honeycomb structures (Fig. \ref{fig4}c), Kagome structures (Fig. \ref{fig4}d), and square Kagome structures (Fig. \ref{fig4}e) as well as segregated B particles (Fig. \ref{fig4}f). In full consistency with our simulation results (Fig. \ref{fig2}), we have observed the coexistence of honeycomb and Kagome structures for the same blending ratio ($\gamma_B=0.3$).
As previously discussed, we also observe the emergence of pure pentagonal lattices at the lowest $n$ studied here ($n=0.25$). Fluorescence images provide insight into the compositional order of these lattices, which is shown in Fig. \ref{fig4}g. As expected, each A particle is surrounded by 5 B particles, thereby maximizing A-B and B-B coordination.
Taken together with the computational analysis presented before, the validating insight provided by laboratory assembly of DFPs reveals that DNA-mediated particle assembly can enable the formation of 2D binary crystalline structures with a diversity of both structural and compositional order and symmetry through enthalpically rather than entropically controlled processes.
Specifically, without any template or tuning of the size of multi-modal DFP mixtures, square, hexagonal lattice and pentagonal aggregates with alternating strand, honeycomb, Kagome and square Kagome binary compositional symmetries emerge as a result of the controlled interplay between attractive interparticle interactions and tailored solution stoichiometry. This enthalpic handle for controlling structural and compositional diversity enhances the control afforded by entropic effects for realizing material complexity and offers potentially new routes to the design and synthesis of material function, hierarchically intercalated structures, or the sacrificial templating of hierarchically porous materials.
\section*{Acknowledgments}
We thank Professor Ben Rogers (Brandeis University) for critical input with the experimental setup and for helpful discussions. This work was supported by the U.S. Department of Energy, Office of Basic Energy Science, Division of Material Sciences and Engineering under Award (DE-SC0013979). This research used resources of the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported under Contract No. DE-AC02-05CH11231.
Use of the high-performance computing capabilities of the Extreme Science and Engineering Discovery Environment (XSEDE), which is
supported by the National Science Foundation, project no. TG-MCB120014, is also gratefully acknowledged.
\section*{Methods}
\subsection*{Particle functionalization}
Green (B) and red (A) fluorescent 1.5-$\mu$m-diameter silica particles (micromod Partikeltechnologie GmbH, Germany) were separately functionalized with blends of complementary single-stranded 5$^\prime$-primary amine-modified DNA (ssDNA, Integrated DNA Technologies, Inc), using silanization and cyanuric chloride chemistries as reported previously.\cite{Steinberg2004} The specific sequences of the complementary ssDNA, which we refer to as $\alpha$DNA and $\beta$DNA, were 5$^\prime$-NH$_2$-(CH$_{2}$)$_{6}$-(T)$_{50}$-TAATGCCTGTCTACC-3$^\prime$ and 5$^\prime$-NH$_2$-(CH$_2$)$_{6}$-(T)$_{50}$-TGAGTTGCGGTAGAC-3$^\prime$, respectively.
Briefly, silica particles were first functionalized with (3-aminopropyl) triethoxysilane (APTES, Acros Organics) in ethanol. Following solvent exchange, cyanuric chloride (CCl, Sigma-Aldrich) was reacted with the primary amine group of APTES on the particles in a solution of acetonitrile and N,N-diisopropylethylamine (Sigma-Aldrich). The covalent attachment of amine-modified oligonucleotides to the surface of CCl-functionalized silica was carried out in salty borate buffer (pH 8.5), with the $\alpha$DNA:$\beta$DNA molar composition of the functionalization solution tuned to match the desired blending ratios, $\gamma_i=\alpha$DNA/($\alpha$DNA+$\beta$DNA), of $\alpha$DNA and $\beta$DNA on the red ($\gamma_A$) and green ($\gamma_B$) particle surface. After completion of the reaction, DNA-functionalized particles were washed and redispersed in TE buffer (pH 8.0), consisting of 10 mM Tris-HCl and 1 mM EDTA, with 100 mM NaCl. The surface density of DNA was estimated by following the same functionalization procedure, but with non-fluorescent silica particles and fluorescently labeled ssDNA (5$^\prime$-NH$_{2}$-(CH$_{2}$)$_{6}$-TTTTTTATGTATCAAGGT-Cy5-3$^\prime$). Fluorimeter measurements estimate the total ssDNA density at ca. 38,000 strands/$\mu$m$^2$. Since the fluorescent DNA bears a different sequence and lower number of base pairs (18 bps) compared to the nominal DNA employed in this study (65 bps), we anticipate this estimated density may serve as an upper bound for the actual density of the 65-bp ssDNA on the silica surface.
\subsection*{Sample preparation} Suspensions containing mixtures of A and B ssDNA-functionalized particles (DFPs) at desired particle number ratios, $n=N_A/N_B$, where $N_A$ and $N_B$ are the number density of A and B particles, respectively, were prepared from stock suspensions of each particle type by washing and redispersion in 100 mM NaCl TE buffer. 0.1 wt$\%$ Pluronic F88 (BASF) was included to protect against non-specific binding between the particles themselves as well as between the particles and the glass surface of the coverslip chamber employed for imaging of particle assembly. The coverslip microchamber was comprised of two plasma-treated coverslips bonded together on three sides by a melted and solidified parafilm seal ($\sim$ 250 $\mu$m thickness). Sample loading was achieved by injecting DFP solution through the remaining open side of the chamber, followed by sealing of the chamber with UV-curable optical adhesive (Norland 63).
\subsection*{Melting curve}
Samples were prepared by loading unary (B particles) or 1:1 binary (A and B particles) mixtures of particles in the microchamber at a concentration leading to a surface density of approximately 0.01 particles/$\mu m^{2}$ following the gravity sedimentation of the DFPs, owing to the silica density of ca. 2 g/cm$^3$. A Peltier thermoelectric module (TE Technology, Inc) driven by a dipolar thermoelectric temperature controller (TE Technology, Inc) was employed to control the sample temperature.
Melting curves were measured by sequentially cooling the DFP samples through the melting transition, with sample equilibration for 15 min at points above and below the transition temperature and for at least 30 min within the transition region at each temperature point. Following equilibration, at least five independent images were collected using an inverted optical microscope (Nikon Eclipse TE2000U, 40x air-immersion objective, 1.5x amplifier). The singlet fraction, defined as the number of individual particles that are not incorporated in a particle aggregate/crystal, was determined by quantitative comparison of the area of individual particles with the area of particles contained in aggregates by common microscopy image analysis\cite{Dreyfus2009, Crocker1996}.
\subsection*{Colloidal crystallization}
Sample chambers containing suspensions of binary DFP mixtures with tailored particle number ratio, $n$, and ssDNA blending ratios, $\gamma_i$, for $i$=A and B particles were attached to the block of a PCR machine (DNA Thermal Cycler 480, Perkin-Elmer) using silicone grease and incubated with a prescribed temperature trajectory through the melting transition to form crystal structures. Specifically,
the temperature annealing was initiated above the melting temperature ($\sim$45 $^{o}$C), where most of the particles exist in their singlet state. The sample temperature was reduced at 1 $^{o}$C sequential increments, with each step maintained for $\sim$ 4 hours. The cooling was terminated at room temperature. Resulting two-dimensional DFP structures were observed by fluorescence microscopy (Nikon Eclipse TE2000U) using a 60x oil-immersion objective.
\subsection*{Image processing and pair correlation function}
All image analysis was based on the image processing routines developed by Crocker and Grier\cite{Crocker1996} and code developed by Eric Weeks and implemented in the software package IDL (Exelis Visual Information Solutions). Pair correlation functions with respect to all particles (A and B particles) were calculated from independent brightfield images whereas pair correlation functions with respect, specifically, to green (A) or red (B) particles were calculated from multi-channel fluorescence images.
\subsection*{Molecular dynamics simulation details}
We perform molecular dynamics simulations using the software package LAMMPS~\cite{Plimpton1995} in a canonical ensemble at low packing fraction of 8\%. All the quantities below are reported in LJ units.
The temperature was maintained by a Langevin thermostat with a damping coefficient $\tau$ = 2
in a 2D simulation box with periodic boundary conditions applied in all (x and y) directions.
The pair interactions between the particles are modeled using SI eq. 1 to capture interparticle interactions between DFPs. Each simulation is conducted for at least $10^8$ time steps with a step size of $\Delta$t = 0.001.
The system contains a total of 400 or 402 particles, depending on the particle number ratio, $n$, of A and B particles.
\subsection*{Identification of the binary crystals formed in MD simulations}
In addition to the nearest neighbor analysis (see SI Fig. S3) and visual inspection, we have implemented the common neighbor analyses (CNA) for the identification of 2D crystals. The python scripting interface of OVITO\cite{Stukowski2009}, originally available for the 3D crystals, was used to calculate the CNA indices - a set of integer triplets (see
SI Fig. S4 for details) of each particle within the formed crystals. We compare these values with those calculated for reference perfect crystals to distinguish the following binary 2D superlattices: square, hexagonal, alternating string, honeycomb, Kagome and square Kagome. In the determination of CNA indices for the identification of the overall arrangement of DFPs of diameter D$_{p}$, i.e. hexagonal and square lattices, a cut-off radius of R$_{cut}$ = 1.6 D$_{p}$ was used to define the neighborhood, whereas, in the identification of the compositionally ordered crystals, R$_{cut}$ was taken as 1.87 D$_{p}$. The structural identifications were further confirmed by the visual inspection of the grown crystals. On the other hand, the rhombic crystals reported in this work, were only visually identified.
\section{Supporting information text}
\subsection{Effective pair potential model for DNA-functionalized particles}
The effective pairwise interactions between DNA-functionalized particles (DFPs) are derived from previously published simulation data using a sequence-specific coarse-grained model~\cite{Ding2014}. Advanced sampling techniques (replica exchange molecular dynamics and umbrella sampling) were used to obtain free energy or potential of mean force (PMF) as a function of interparticle distance. We use the following expression to fit these data to represent various contributions to the interparticle interactions between DFPs, (i) repulsive interactions between particle cores, (ii) repulsive interactions due to DNA chain overlap, and (iii) attractive interactions due to DNA hybridization.
\begin{equation}
U(r)=\epsilon_{o}\left(\frac{\sigma_{o}}{r-r_{shift}}\right)^{n}+\frac{A_{0}}{1+\text{exp}({A_{1}(r-A_{2}))}}-\frac{B_{0}}{1+\text{exp}({B_{1}(r-B_{2}))}}
\end{equation}
Figure S\ref{pmfs} shows an example comparison between the simulation data and eq. 1. By varying $B_{0}$, we can model potentials with different attractive well depths if particle size and overall grafting density is held constant. Even though the PMFs represented by eq. 1 are reflective of nanosized particles, the qualitative shape of the potential is quite similar to the one measured experimentally for micron-sized particles~\cite{Rogers2011}. We simply rescale the potential to reflect features of micron-sized particles with an appropriate DNA length used in the experimental part of this work. Specifically, we have used the following values for various parameters in eq. 1: $\epsilon_{0}$ = 10.0$\epsilon$, $\sigma_{0}$ = 0.2$\sigma$, r$_{shift}$ = 0.8$\sigma$, n = 36, A$_{0}$ = 11.03$\epsilon$, A$_{1}$ = 404.4$\sigma^{-1}$, A$_{2}$ = 1.0174$\sigma$, B$_{1}$ = 1044.5$\sigma^{-1}$ and B$_{2}$ = 1.031$\sigma$. B$_{0}$ was changed from 0 to 1.32 to vary the attractive well depth between 0 and 1.0$\epsilon$, respectively.
\subsection{Nearest neighbor analysis (NNA)}
To distinguish the formation of various binary superlattices, we calculate the average number of like (AA and BB) and unlike contacts (AB), defined as N$_{K}$ = $\sum n_{K}^{i} / n$ with K = \{AB, AA, BB\}.
Where, n is the total number of particles analyzed in a given lattice and $n_{K}^{i}$ is the number of contacts (AB, AA or BB) between particle $i$ and its nearest neighbors.
|
\section{Introduction}
A variety is one of significant notions in universal algebraic geometry. Let us explain its importance. Let $\LL$ be a language. An $\LL$-equation is an atomic formula $t(X)=s(X)$, where $t,s$ are $\LL$-terms. Let $A$ be an algebraic structure of $\LL$ ($\LL$-algebra). Suppose we are dealing with $\LL$-equations over $A$, and $A$ belongs to a certain variety $\Vbf$ of $\LL$-algebras. Then we can easily simplify $\LL$-terms (parts of equations) using the identities of $\Vbf$. For example, any term $t(X)$ over a semigroup $S$ admits an elimination of brackets, since $S$ satisfies the associativity law (identity) $x(yz)=(xy)z$. Similarly, any ring with zero multiplication admits a reduction of any term to a multiplication-free expression.
Another important notion in universal algebraic geometry is the property of being equationally Noetherian. Recall that an algebraic structure $A$ is $\LL$-equationally Noetherian if any system of $\LL$-equations is equivalent over $A$ to its finite subsystem. The problem about the connections between varieties and equationally Noetherian algebras was posed by B.~Plotkin in~\cite{Plotkin}.
\bigskip
\noindent {\bf Problem 1.} Is there a variety $\Vbf$ of $\LL$-algebras such that every $A\in\Vbf$ is $\LL$-equationally Noetherian algebra?
\bigskip
This problem has positive solutions for many varieties. For example, in~\cite{ModSH} it was proved that all elements of a group variety $\mathbf{V}$ are $\LL_g$-equationally Noetherian ($\LL_g=\{\cdot,{ }^{-1},1\}$ is the group language), if and only if the free group $F_{\mathbf{V}}(X)\in\Vbf$ has the $\mathrm{max-n}$ property for every finite set $X$. It follows that the variety of all metabelian groups satisfies Problem~1.
Let us formulate the central problem of the current paper (it slightly differs from the original Plotkin`s problem).
\bigskip
\noindent {\bf Problem 2.} Is there a variety $\Vbf$ of $\LL$-algebras such that each $A\in\Vbf$ is $\LL(A)$-equationally Noetherian algebras?
\bigskip
The language $\LL(A)=\LL\cup \{a\mid a\in A\}$ above is the extension of $\LL$ by new constant symbols $a\in A$ corresponding to all elements of $A$. Recall that $\LL(A)$ defines the wider class of equations than $\LL$, because it allows to use constants in equations. Thus, an $\LL$-equationally Noetherian algebra $A$ is not necessarily $\LL(A)$-equationally Noetherian.
We prove (Theorems~\ref{th:group_var_from_N},\ref{th:rings_var_from_N}) that a variety of groups (rings) satisfies Problem~2, if and only if it is abelian (respectively, has zero multiplication). The similar result holds for monoids (Theorem~\ref{th:monoid_var_from_N}).
The obtained results are based on equational properties of direct powers of groups (Theorem~\ref{th:group_restricted_power_is_not_q}) rings (Theorem~\ref{th:ring_restricted_power_is_not_q}) and monoids (Theorem~\ref{th:monoid_restricted_power_is_not_q}). Moreover, in Theorem~\ref{th:magma_restricted_power_is_not_q} we consider direct powers of arbitrary algebras with binary operation (magmas). Actually, in such theorems we prove that an infinite direct power of a certain group (ring, monoid, magma) is not $\qq$-compact.
Let us mention an application of our results. In~\cite{BMRom} it was proved that any wreath product (restricted and unrestricted) $W=A\mathrm{wr}B$ of a non-abelian group $A$ and infinite group $B$ is not $\LL(W)$-equationally Noetherian. Theorem~\ref{th:group_restricted_power_is_not_q} allows us to obtain a simple proof of this fact.
In Section~\ref{sec:var_of_groups_in_L_g} we deal with the language $\LL_g$ (the language of groups), and we study axiomatic ranks and finite basis problem for group varieties $\Vbf$ such that each $G\in\Vbf$ is $\LL_g$-equationally Noetherian.
\section{Preliminaries}
Let us give some basic notions of universal algebraic geometry following~\cite{DMR1}--~\cite{DMR3}. Let $\LL$ be a functional language. In the current paper we consider as special cases, languages of the following types: $\LL_m=\{\cdot,1\}$ (monoid language), $\LL_g=\{\cdot,{}^{-1},1\}$ (group language), $\LL_r=\{+,-,\cdot,0\}$ (ring language).
Let $A$ be an algebraic structure of the language $\LL$ ($\LL$-algebra). An {\it equation over $\LL$ ($\LL$-equation)} is an equality of two terms of $\LL$:
$$
p(X)=q(X).
$$
The examples of equations in various languages are: $[x,y]=1$ (here $[x,y]=x^{-1}y^{-1}xy$), $x^{-1}y^3x=y^2$ (language $\LL_g$); $xy=yx$, $x^2=1$ (language $\LL_m$); $x^2+y^2=z^2$, $xy+xy+yz=0$ (language $\LL_r$).
A system of $\LL$-equations ($\LL$-system for shortness) is an arbitrary set $\Ss$ of $\LL$-equations. {\it Notice that we will consider only systems which depend on a finite set of variables $X=\{x_1,x_2,\ldots,x_n\}$}. The set of all solutions of $\Ss$ in $A$ is denoted by $\V_A(\Ss)\subseteq A^n$.
An $\LL$-algebra $A$ is {\it $\LL$-equationally Noetherian} if any infinite $\LL$-system $\Ss$ is equivalent over $A$ to some finite subsystem $\Ss^\pr\subseteq \Ss$. An $\LL$-algebra $A$ is {\it $\qq$-compact} in the language $\LL$ if for any infinite $\LL$-system $\Ss$ and an $\LL$-equation $p(X)=q(X)$ such that
$$
\V_A(\Ss)\subseteq \V_A(p(X)=q(X))
$$
there exists a finite subsystem $\Ss^\pr\subseteq\Ss$ with
$$
\V_A(\Ss^\pr)\subseteq \V_A(p(X)=q(X)).
$$
According to the definitions, any $\LL$-equationally Noetherian algebra is $\qq$-compact in $\LL$.
Let $A$ be an $\LL$-algebra. By $\LL(A)$ we denote the language $\LL\cup\{a\mid a\in A\}$ extended by new constants symbols which correspond to elements of $A$. The language extension allows us to use constants in equations. The examples of equations in extended languages are the following: $[x,g]=1$ (language $\LL_g(G)$, $g\in G$); $xm=mx$ (language $\LL_m(M)$, $m\in M$); $r\cdot x^2+r^\pr\cdot y^2=0$ (language $\LL_r(R)$, $r,r^\pr\in R$). Obviously, the class of $\LL(S)$-equations is wider than the class of $\LL$-equations, so an $\LL$-equationally Noetherian algebra may lose this property in the language $\LL(A)$. For simplicity, we will say that $A$ is {\it equationally Noetherian} if $A$ is $\LL(A)$-equationally Noetherian and $A$ is {\it $\qq$-compact} is it is $\qq$-compact in the language $\LL(A)$.
The following proposition contains simple examples of equationally Noetherian algebraic structures.
\begin{proposition}
\begin{enumerate}
\item Any abelian group $G$ is equationally Noetherian.
\item Any ring $R$ with zero multiplication is equationally Noetherian.
\end{enumerate}
\label{pr:examples_of_noeth}
\end{proposition}
\begin{proof}
The first statement is well-known in universal algebraic geometry (see~\cite{DMR1,DMR2}).
Let $R$ be a ring with zero multiplication, i.e. $ab=0$ for all $a,b\in R$. Therefore, any $\LL_r(R)$-equation is equivalent to one of the following simple expressions $x_i=x_j$, $x_i=a$, $a=b$, where $a,b\in R$. So any system $\Ss$ does not contain an infinite number of pairwise non-equivalent $\LL_r(R)$-equations. Thus, $\Ss$ should be equivalent to its finite subsystem.
\end{proof}
Let $A$ be an $\LL$-algebra. The infinite \textit{direct power} $A^{\infty}=\prod_{i=1}^\infty A$ is an $\LL$-algebra of all sequences
\[
(a_1,a_2,\ldots,a_n,\ldots),\; a_i\in A,
\]
and the definition of each function $f\in\LL$ over $A^{\infty}$ is element-wise.
Let $G^\infty$ be the direct power of a group (monoid) $G$, and $\g=(g_1,g_2,\ldots)\in G^\infty$. Then the {\it support} $\supp(\g)$ is the set of indexes with $g_i\neq 1$. The support $\supp(\r)$ of an element $\r=(r_1,r_2,\ldots)$ in the direct power $R^\infty$ of a ring $R$ is defined as $\{i\mid r_i\neq 0\}$.
The {\it restricted direct power} $G^{(\infty)}$ of a group $G$ is a subgroup in $G^\infty$ of all elements $\g$ with $|\supp(\g)|<\infty$. Similarly, one can define the restricted direct power for monoids and rings.
\section{Direct products}
\begin{theorem}
Let $G$ be a non-abelian group. Then the restricted direct power $G^{(\infty)}$ is not $\qq$-compact.
\label{th:group_restricted_power_is_not_q}
\end{theorem}
\begin{proof}
Let $a\in G$, and $\Ss_a(x)$ be an infinite system of $G^{(\infty)}$-equations defined as follows:
$$
\Ss_a(x)=
\begin{cases}
[x,(a,1,1,1\ldots)]=\one,\\
[x,(a,a,1,1\ldots)]=\one,\\
[x,(a,a,a,1\ldots)]=\one,\\
\ldots\\
\end{cases}
$$
Denote $\Ss(x)=\bigcup_{a\in G}\Ss_a$. Obviously,
$$
\V_{G^{(\infty)}}(\Ss(x))=\{\g=(g_1,g_2,g_3\ldots)\mid g_i\in Z(G), \supp(\g)<\infty \}.
$$
and therefore $\V_{G^{(\infty)}}(\Ss(x))=Z(G^{(\infty)})$ (here $Z(G)$ is the center of $G$).
Let $\Ss(x,y)=\Ss(x)\cup\Ss(y)$ ($\Ss(y)$ is a clone of $\Ss(x)$, where all occurrences of $x$ are replaced to $y$). Then
$$
\V_{G^{(\infty)}}(\Ss(x,y))=\{(\g,\h)\mid \g,\h\in Z(G^{(\infty)})\}.
$$
and the following inclusion
\begin{equation}
\V_{G^{(\infty)}}(\Ss(x,y))\subseteq\V_{G^{(\infty)}}([x,y]=\one)
\end{equation}
holds.
Let us consider a finite subsystem $\Ss^\pr(x,y)$ of $\Ss(x,y)$. Without loss of generality one can assume that
$$
\Ss^\pr(x,y)=\bigcup_{a\in T} \left(\Ss_a^\pr(x)\cup\Ss_a^\pr(y)\right),
$$
where $T$ is a finite subset of $G$ and $\Ss_a^\pr(x),\Ss_a^\pr(y)$ are the first $n$ equations of the systems $\Ss_a(x),\Ss_a(y)$ respectively.
Since $G$ is not abelian, there exists a pair $\al,\beta\in G$ with $[\al,\beta]\neq 1$. Let us consider the following elements
$$
\g=(\underbrace{1,1,\ldots,1}_{\mbox{$n$ times}},\al,1,1,\ldots)\in G^{(\infty)},
$$
$$
\h=(\underbrace{1,1,\ldots,1}_{\mbox{$n$ times}},\beta,1,1,\ldots)\in G^{(\infty)}.
$$
One can directly check that $(\g,\h)\in\V_{G^{(\infty)}}(\Ss^\pr(x,y))$, but $[\g,\h]\neq 1$. Therefore, the inclusion
\begin{equation}
\V_{G^{(\infty)}}(\Ss^\pr(x,y))\subseteq\V_{G^{(\infty)}}([x,y]=\one)
\end{equation}
fails. Thus, $G^{(\infty)}$ is not $\qq$-compact.
\end{proof}
Since $G^{(\infty)}$ is embedded into $G^\infty$, we immediately obtain the following result.
\begin{corollary}
The direct power $G^{\infty}$ is not $\qq$-compact for any non-abelian group $G$.
\label{cor:th:group_power_is_not_N}
\end{corollary}
\bigskip
In \cite{BMRom}, it is proved that if $G$ is non-abelian and $A$ is infinite group, then the wreath product $W=G\mathrm{wr}A$ (restricted or unrestricted) is not equationally Noetherian. Now, we see that this is an instant corollary of Theorem~\ref{th:group_restricted_power_is_not_q}, since $G^{(\infty)}$ is embedded into $W$.
\begin{corollary}
Let $G\mathrm{wr}A$ be the wreath product of a non-abelian group $G$ and infinite group $A$. Then this group is not equationally Noetherian.
\end{corollary}
\bigskip
Let us study ring equations.
\begin{theorem}
Let $R$ be a ring with a nonzero multiplication. Then the restricted direct power $R^{(\infty)}$ is not $\qq$-compact.
\label{th:ring_restricted_power_is_not_q}
\end{theorem}
\begin{proof}
Let us consider an infinite system
$\Ss_a(x)$ of $R^{(\infty)}$-equations defined as follows:
$$
\Ss_a(x)=
\begin{cases}
(a,0,0,0,\ldots)\cdot x=\zero,\\
(a,a,0,0,\ldots)\cdot x=\zero,\\
(a,a,a,0,\ldots)\cdot x=\zero,\\
\ldots
\end{cases}
$$
Denote $\Ss(x)=\bigcup_{a\in R}\Ss_a$. Obviously,
$$
\V_{R^{(\infty)}}(\Ss(x))=\{\r=(r_1,r_2,r_3\ldots)\mid r_i\in \Ann_r(R), \supp(\r)<\infty \},
$$
where
$$
\Ann_r(R)=\{b\in R\mid ab=0\mbox{ for all $a\in R$}\}
$$
is the right annihilator of $R$. Therefore $\V_{R^{(\infty)}}(\Ss(x))=\Ann_r(R^{(\infty)})$.
Let $\Ss(x,y)=\Ss(x)\cup\Ss(y)$ ($\Ss(y)$ is a clone of $\Ss(x)$, where all occurrences of $x$ are replaced to $y$). Then
\[
\V_{R^{(\infty)}}(\Ss(x,y))=\{(\r,\r^\pr)\mid \r,\r^\pr\in \Ann_r(R^{(\infty)})\}.
\]
and the following inclusion
\begin{equation}
\V_{G^{(\infty)}}(\Ss(x,y))\subseteq\V_{G^{(\infty)}}(xy=\zero)
\end{equation}
holds.
Let us consider a finite subsystem $\Ss^\pr(x,y)$ of $\Ss(x,y)$. Without loss of generality one can assume that
\[
\Ss^\pr(x,y)=\bigcup_{a\in T} \left(\Ss_a^\pr(x)\cup\Ss_a^\pr(y)\right),
\]
where $T$ is a finite subset of $R$ and $\Ss_a^\pr(x),\Ss_a^\pr(y)$ are the first $n$ equations of the systems $\Ss_a(x),\Ss_a(y)$ respectively.
By the condition, there exists a pair $\al,\beta\in R$ with $\al\beta\neq 0$. Let us consider the following elements
\[
\r=(\underbrace{0,0,\ldots,0}_{\mbox{$n$ times}},\al,0,0,\ldots)\in R^{(\infty)},
\]
\[
\r^\pr=(\underbrace{0,0,\ldots,0}_{\mbox{$n$ times}},\beta,0,0,\ldots)\in R^{(\infty)}.
\]
One can directly check that $(\r,\r^\pr)\in\V_{R^{(\infty)}}(\Ss^\pr(x,y))$, but $\r\r^\pr\neq 0$. Therefore, the inclusion
\begin{equation}
\V_{R^{(\infty)}}(\Ss^\pr(x,y))\subseteq\V_{R^{(\infty)}}(xy=\zero)
\end{equation}
fails. Thus, $R^{(\infty)}$ is not $\qq$-compact.
\end{proof}
We have the analogue of Corollary~\ref{cor:th:group_power_is_not_N} for rings.
\begin{corollary}
Let $R$ be a ring with non-zero multiplication. Then the direct power $R^{\infty}$ is not $\qq$-compact.
\label{cor:th:ring_power_is_not_N}
\end{corollary}
\bigskip
Also Theorem~\ref{th:group_restricted_power_is_not_q} allows to prove the following statement.
\begin{theorem}
Let $M$ be a non-commutative monoid. Then the restricted direct power $M^{(\infty)}$ is not $\qq$-compact.
\label{th:monoid_restricted_power_is_not_q}
\end{theorem}
\begin{proof}
One should rewrite a system $\Ss_a$ from Theorem~\ref{th:group_restricted_power_is_not_q} as a system of $\LL_m$-equations.
\[
\Ss_a(x)=
\begin{cases}
x(a,1,1,1\ldots)=(a,1,1,1\ldots)x,\\
x(a,a,1,1\ldots)=(a,a,1,1\ldots)x,\\
x(a,a,a,1\ldots)=(a,a,a,1\ldots)x,\\
\ldots\\
\end{cases}
\]
and the rest proof is similar to the argument of Theorem~\ref{th:group_restricted_power_is_not_q}.
\end{proof}
One can generalize the above arguments even for wider class of algebras. A {\it magma} is an arbitrary algebraic structure of the language $\LL_s=\{\cdot\}$. The natural example of a magma is the following.
\begin{example}
Let $\LL$ be an arbitrary language. If $A$ is an $\LL$-algebra and $p(x,y)$ is an $\LL$-term, then the the set $A$ equipped with the binary operation $p(x,y)$ is a magma. Let us denote the obtained magma by $(A,p)$
\label{ex:magma}
\end{example}
\bigskip
One can naturally give the definitions of the commutativity and center for magmas, and the proof of the following result is similar to Theorem~\ref{th:monoid_restricted_power_is_not_q}.
\begin{theorem}
Let $M$ be a non-commutative magma with non-empty center. Then the direct power $M^{\infty}$ is not $\qq$-compact.
\label{th:magma_restricted_power_is_not_q}
\end{theorem}
\bigskip
Theorem~\ref{th:magma_restricted_power_is_not_q} and Example~\ref{ex:magma} give a result about varieties of $\LL$-algebras.
\begin{corollary}
If every element $A$ of a variety $\mathbf{V}$ of $\LL$-algebras is equationally Noetherian, then for any $\LL$-term $p(x,y)$, the magma $(A, p)$ is commutative or has empty center.
\end{corollary}
\bigskip
The results of Theorems~\ref{th:group_restricted_power_is_not_q},~\ref{th:ring_restricted_power_is_not_q},
~\ref{th:monoid_restricted_power_is_not_q} shows that an infinite direct product is not necessarily equationally Noetherian. However, the finite direct products of equationally Noetherian algebraic structures are equationally Noetherian.
\begin{proposition}
Suppose that $A$ and $B$ are two $\LL$-algebras. Let $A$ be $\LL(A)$-equationally Noetherian and $B$ be $\LL(B)$-equationally Noetherian. Then $C=A\times B$ is $\LL(C)$-equationally Noetherian.
\end{proposition}
\begin{proof}
First, let $p(X)=p(x_1, \ldots, x_n, (a_1, b_1), \ldots, (a_m, b_m))$ be a term in the language $\LL(C)$. Note that, here all $(a_i, b_i)$ are coefficients from $C$. There are two new terms corresponding to $p(X)$; the first one is a term in the language $\LL(A)$ which we denote it by $p_A$ and the second one is a term in the language $\LL(B)$ which will be denoted by $p_B$. The term $p_A$ is obtained from $p(X)$ by replacing every variable $x_i$ by a new variable $y_i$ and every coefficient $(a_i, b_i)$ by $a_i$. The same is true also for $p_B$, it is obtained from $p(X)$ by replacing every variable $x_i$ by a new variable $z_i$ and every coefficient $(a_i, b_i)$ by $b_i$. Let $p(X)=q(X)$ be an equation in the language $\LL(C)$. Let $((\alpha_1, \beta_1), \ldots, (\alpha_n, \beta_n))\in C^n$ be a solution of this equation. Then we have
$$
(p_A(\alpha_1, \ldots, \alpha_n, \overline{a}), p_B(\beta_1, \ldots, \beta_n, \overline{b}))=(q_A(\alpha_1, \ldots, \alpha_n, \overline{a}), q_B(\beta_1, \ldots, \beta_n, \overline{b})),
$$
where $\overline{a}=(a_1,\ldots,a_m)$, $\overline{b}=(b_1,\ldots,b_m)$.
So $(\alpha_1, \ldots, \alpha_n)$ is a solution of $p_A(Y)=q_A(Y)$ and $(\beta_1, \ldots, \beta_n)$ is a solution of $p_B(Z)=q_B(Z)$. The converse is also true. In general, let $\Ss$ be a system of equations in the language $\LL(C)$. Let
$$
\Ss^A=\{ p_A(Y)=q_A(Y):\ (p(X)=q(X))\in \Ss\}
$$
and similarly define $\Ss^B$. An easy argument shows that there is a bijection between the algebraic set $\V_C(\Ss)$ and the Cartesian product $\V_A(\Ss^A)\times \V_B(\Ss^B)$. Now, let
$$
\V_C(\Ss_1)\supseteq \V_C(\Ss_2)\supseteq \V_C(\Ss_3)\supseteq \cdots
$$
be a descending chain of algebraic sets in the space $C^n$. Correspondingly, we have a chain
$$
\V_A(\Ss_1^A)\times \V_B(\Ss_1^B)\supseteq \V_A(\Ss_2^A)\times \V_B(\Ss_2^B)\supseteq \V_A(\Ss_3^A)\times \V_B(\Ss_3^B)\supseteq\cdots,
$$
and hence, two partial chains
$$
\V_A(\Ss_1^A)\supseteq \V_A(\Ss_2^A)\supseteq \V_A(\Ss_3^A)\supseteq \cdots,
$$
$$
\V_B(\Ss_1^B)\supseteq \V_B(\Ss_2^B)\supseteq \V_B(\Ss_3^B)\supseteq \cdots.
$$
Since $A$ and $B$ are equationally Noetherian in their own languages, there exists an integer $k$ such that
$$
\V_A(\Ss_k^A)=\V_A(\Ss_{k+1}^A)=\cdots, \ \ \V_B(\Ss_k^B)=\V_B(\Ss_{k+1}^B)=\cdots,
$$
and hence, we have
$$
\V_C(\Ss_k)=\V_C(\Ss_{k+1})=\cdots.
$$
This shows that $C$ is $\LL(C)$-equationally Noetherian.
\end{proof}
\section{Equationally Noetherian Varieties}
In this section, we solve Problem~2 for varieties of groups, rings and monoids. Actually, we show more generally that if a class of groups (rings) is closed under the restricted direct power, the all elements of this class are either abelian (respectively, with zero multiplication) or at least one element of the class is not equationally Noetherian.
\begin{theorem}
Let $\Vbf$ be a variety of groups. Each $G\in \Vbf$ is equationally Noetherian, if and only if $\Vbf$ is abelian.
\label{th:group_var_from_N}
\end{theorem}
\begin{proof}
The ``if'' part of the statement follows from Proposition~\ref{pr:examples_of_noeth}. Let us prove the ``only if'' part.
Assume there exists a non-abelian group $G\in \Vbf$. Since any variety is closed under direct products, we have $G^\infty\in\Vbf$. By Theorem~\ref{th:group_restricted_power_is_not_q}, $G^\infty$ is not equationally Noetherian, and we obtain the contradiction.
\end{proof}
The theorem above can be used for the case of linear groups. Recall that a group $G$ is linear, if there exists a Noetherian ring $R$ and a natural number $n$, such that $G$ embeds in $GL_n(R)$. It is well-known that every linear group is equationally Noetherian. So, we have
\begin{corollary}
If a class of linear groups is closed under restricted or unrestricted direct power, then all elements of that class are abelian.
\end{corollary}
\begin{theorem}
Let $\Vbf$ be a variety of rings. Each $R\in \Vbf$ is equationally Noetherian, if and only if all elements of $\Vbf$ have zero multiplication.
\label{th:rings_var_from_N}
\end{theorem}
\begin{proof}
The ``only if'' part of the statement directly follows from Theorem~\ref{th:ring_restricted_power_is_not_q}. The ``if'' part follows from Proposition~\ref{pr:examples_of_noeth}.
\end{proof}
Now we describe monoid varieties $\Vbf$, where each $M\in\Vbf$ is equationally Noetherian.
\begin{lemma}
Let $L_2=\{0,1\}$ be a two-element semilattice ($0\cdot 0=0\cdot 1=1\cdot 0=0$, $1\cdot 1=1$). Then the direct power $L_2^\infty$ is not equationally Noetherian.
\end{lemma}
\begin{proof}
Let us consider an infinite system of $L_2^{\infty}$-equations
\begin{equation}
\label{eq:system_S_for_L_2}
\Ss=
\begin{cases}
x\cdot(1,1,1,\ldots)=x,\\
x\cdot(0,1,1,\ldots)=x,\\
x\cdot(0,0,1,\ldots)=x,\\
\ldots
\end{cases}
\end{equation}
We have $\V_{L_2^\infty}(\Ss)=\{(0,0,0,\ldots)\}$ . However, the solution set of the first $n$ equations of $\Ss$ is
\[
\{(\underbrace{0,0,\ldots,0}_{\mbox{$n$ times}},a_{n+1},a_{n+2},\ldots)\mid a_i\in L_2\}.
\]
Thus, $\Ss$ is not equivalent to its finite subsystems, so $L_2^{\infty}$ is not equationally Noetherian.
\end{proof}
\begin{corollary}
Let $\Vbf$ be a variety of monoids such that each $M\in \Vbf$ is equationally Noetherian. Then $L_2\notin \Vbf$.
\label{cor:L_2_notin_V}
\end{corollary}
\begin{theorem}
Let $\Vbf$ be a variety of monoids. Each $M\in \Vbf$ is equationally Noetherian, if and only if $\Vbf$ is a variety of abelian groups defined by the identity $x^n=1$.
\label{th:monoid_var_from_N}
\end{theorem}
\begin{proof}
The ``if'' part of the statement follows from Proposition~\ref{pr:examples_of_noeth}. Let us prove the ``only if'' part.
Theorem~\ref{th:monoid_restricted_power_is_not_q} immediately gives that $\Vbf$ is an abelian variety.
By Corollary~\ref{cor:L_2_notin_V}, there exists an identity $p(X)=q(X)$ such that $p(X)=q(X)$ is true in any $M\in\Vbf$ and $p(X)=q(X)$ does not hold in $L_2$. According to the properties of $L_2$, there exists a variable $x$ occurring in $p(X)$ not in $q(X)$. Let us substitute all variables (except $x$) in $p(X)=q(X)$ to $1$, and we obtain an identity $x^n=1$ which holds in $\Vbf$. Thus, any $M\in\Vbf$ is a group (the inverse $a^{-1}$ of any $a\in M$ is $a^{n-1}$).
\end{proof}
\section{$\LL_g$-equations and varieties of groups}
\label{sec:var_of_groups_in_L_g}
Although the variety of abelian groups is a unique example where every element $G$ is equationally Noetherian, there are many examples of varieties every group of which is $\LL_g$-equationally Noetherian (1-equationally Noetherian for shortness). We can use results of~\cite{ModSH} to find such examples. Let us formulate the result of~\cite{ModSH} for groups of language $\LL_g$ (below $\mathrm{max-n}$ is the following property of a group $G$: every ascending chain of normal subgroups $H\vartriangleleft G$ becomes stationary).
\begin{theorem}\textup{\cite{ModSH}}
All groups of a variety $\mathbf{V}$ are 1-equationally Noetherian, if and only if the free group $F_{\mathbf{V}}(X)\in\Vbf$ has the $\mathrm{max-n}$ property for every finite set $X$.
\label{th:noeth_iff_max-n}
\end{theorem}
\bigskip
Since any finitely generated metabelian group has the $\mathrm{max-n}$ property, all metabelian groups are 1-equationally Noetherian. It is also easy to see that every finitely generated nilpotent group has the maximal condition on its subgroups so applying Theorem~\ref{th:noeth_iff_max-n}, we see that every element of a nilpotent variety is 1-equationally Noetherian. Below, we show that there is a close connection between such property of varieties and the property of being finitely based.
Suppose $\mathbf{V}$ is a variety of groups and $\mathbf{W}$ is a subvariety. We say that $\mathbf{W}$ is relatively finitely based, if it can be defined by a finite number of identities inside $\mathbf{V}$. The relative axiomatic rank of $\mathbf{W}$ is finite, if it can be defined using finite number of variables inside $\mathbf{V}$. We apply Theorem~\ref{th:noeth_iff_max-n}, to prove the following result.
\begin{theorem}
Let $\mathbf{V}$ be a variety of groups such that its all elements are 1-equationally Noetherian. Let $\mathbf{W}$ be a subvariety with the finite relative axiomatic rank. Then $\mathbf{W}$ is relatively finitely based.
\end{theorem}
\begin{proof}
Let $R=\mathrm{id}^{\mathbf{V}}_X(\mathbf{W})$ be the set of all $\mathbf{V}$-identities defining $\mathbf{W}$ inside $\mathbf{V}$. We can assume that $X$ is finite. Clearly we have $R\unlhd F_{\mathbf{V}}(X)$. Since every element of $\mathbf{V}$ is 1-equationally Noetherian, so by Theorem 1.1, the group $F_{\mathbf{V}}(X)$ has $\mathrm{max-n}$ and hence $R$ is finitely generated as a normal subgroup. So, there are elements $u_1, \ldots, u_m\in F_{\mathbf{V}}(X)$, such that $R$ is the normal closure of $\{u_1, \ldots, u_m\}$. Suppose
$$
\mathbf{W}^{\prime}=\{ G\in \mathbf{V}: G\models (u_1=1), \ldots, (u_m=1)\}.
$$
We show that $\mathbf{W}^{\prime}=\mathbf{W}$. Clearly $\mathbf{W}\subseteq \mathbf{W}^{\prime}$. Let $G\in \mathbf{W}^{\prime}$ and $v\in R$. Then $v=\prod w_iu_{t_i}^{\pm1}w_i^{-1}$ for some elements $w_1, w_2, \ldots$ and $u_{t_1}, u_{t_2}, \ldots$. Therefore $G\models (v=1)$ and hence $G\in \mathbf{W}$. This proves that $\mathbf{W}$ is relatively finitely based.
\end{proof}
\bigskip
|
\section{Details for the mean field self consistent calculations}
Consider superconductor (SC) order parameters $\Delta_{2\mathbf{q}}=\frac{U}{N}\sum_{\mathbf{k}}\left\langle c_{\mathbf{\mathbf{q}+\mathbf{k}\uparrow}}c_{\mathbf{Q}-\mathbf{k}\downarrow}\right\rangle $,
with $\mathbf{q}=0$ for BCS and $\mathbf{q}=\mathbf{Q}_{\pm}$ for
PDW orders. With the PDW orders, the original Brillouin zone (BZ)
will be folded up into sub-BZ. In the present study, we choose $\mathbf{Q_{+}=-Q_{-}}=(0,2\pi/3)$,
so the folded BZ is $1/3$ of the original BZ, and the mean field
Hamiltonian can be written as
\[
H_{\text{MF}}=\frac{1}{2}\sum_{\mathbf{k}}\Psi_{\mathbf{k}}^{\dagger}\mathcal{H}_{\text{MF}}(\mathbf{k})\Psi_{\mathbf{k}},
\]
\begin{equation}
\mathcal{H}_{\text{MF}}(\mathbf{k})=\left[\begin{array}{cc}
\hat{\mathcal{H}}_{0}(\mathbf{k}) & \hat{\Delta}(\mathbf{k})\\
\hat{\Delta}^{\dagger}(\mathbf{k}) & -\hat{\mathcal{H}}_{0}^{\text{T}}(-\mathbf{k})
\end{array}\right],\label{eq:mf}
\end{equation}
where the basis for the folded BZ is denoted as
\[
\Psi_{\mathbf{k}}=(c_{\mathbf{Q}_{+}+\mathbf{k}\uparrow},c_{\mathbf{k}\uparrow},c_{\mathbf{Q}_{-}+\mathbf{k}\uparrow},(\uparrow\to\downarrow);c_{\mathbf{Q}_{+}-\mathbf{k}\uparrow}^{\dagger},c_{-\mathbf{k}\uparrow}^{\dagger},c_{\mathbf{Q}_{-}-\mathbf{k}\uparrow}^{\dagger},(\uparrow\to\downarrow))^{\text{T}},
\]
with $k_{x}\in[-\pi,\pi)$ and $k_{y}\in[-\pi/3,\pi/3)$. The explicit
form of $\hat{\mathcal{H}}_{0}(\mathbf{k})$ is obtained by restricting
the momentum of the 2D topological Dirac metal Hamiltonian $\mathcal{H}_{0}(\mathbf{k})=(m_{z}-2t_{x}\cos k_{x}-2t_{y}\cos k_{y})\sigma_{z}+2t_{\text{so}}\sin k_{x}\sigma_{x}-\mu$
within a sub-BZ. The order parameter $\hat{\Delta}$ in the matrix
form reads
\begin{equation}
\hat{\Delta}=\left[\begin{array}{cc}
& \Delta_{[\mathbf{\mathbf{Q}}]}\\
-\Delta_{[\mathbf{\mathbf{Q}}]}
\end{array}\right]\label{eq:delta}
\end{equation}
with
\[
\Delta_{[\mathbf{Q}]}=\left[\begin{array}{ccc}
\Delta_{2\mathbf{Q}_{+}} & \Delta_{2\mathbf{Q}_{-}} & \Delta_{0}\\
\Delta_{2\mathbf{Q}_{-}} & \Delta_{0} & \Delta_{2\mathbf{Q}_{+}}\\
\Delta_{0} & \Delta_{2\mathbf{Q}_{+}} & \Delta_{2\mathbf{Q}_{-}}
\end{array}\right].
\]
Utilizing Eq. \eqref{eq:mf} and Eq. \eqref{eq:delta}, one can interatively
solve the Hamiltonian and compute $\Delta_{2\mathbf{q}}$'s until
convergence. The mean field phase diagram versus attractive interaction
strength $U$ and chemical potential $\mu$ are shown in Fig. 2 in
the main text.
\section{Derivations for the generic reduced formula of the Chern-Simons invariant}
Now we are going to prove that, for a 2D superconductor with vortices,
the Chern-Simons invariant $\nu_{3}$~\cite{Pontryagin,WenXiao-Gang2008,JeffreyTeo2010}
defined in the base space $(k_{x},k_{y},\phi)\in T^{2}\times S^{1}$,
with $\phi$ denoting the emergent dimension for vorticity,
\begin{align*}
\nu_{3} & =-\frac{1}{4\pi^{2}}\int_{T^{2}\times S^{1}}\mathcal{Q}_{3}\;\mathrm{mod}\;2,\\
\mathcal{Q}_{3} & =\mathrm{Tr}\left[\mathcal{A}d\mathcal{A}-\frac{2\text{i}}{3}\mathcal{A}^{3}\right],
\end{align*}
where $\mathcal{A}_{\lambda\lambda'}(\mathbf{k},\phi)=\text{i}\langle \psi_{\lambda}|d\psi_{\lambda'}\rangle$ is the
one-form Berry connection ($|\psi_{\lambda}\rangle$ is the corresponding eigenvector
of the Hamiltonian, and the trace is performed on the filled bands), takes the following simple form
\[
\nu_{3}=\frac{1}{2\pi}\sum_{i}n_{i}\oint_{\partial\vec{S}_{i}}\nabla_{\mathbf{k}}\arg\Delta_{\mathbf{Q}_{i}}(\mathbf{k})\cdot d\mathbf{k}\;\mathrm{mod}\;2,
\]
where the orientable area $\vec{S}_{i}$ denotes the region enclosed
by the $i$-th Fermi surface ($\partial S_{i}=\text{FS}_{i}$ in the
main text), $\Delta_{\mathbf{Q}_{i}}$ is the superconductor
order parameter projected onto the $i$-th Fermi surface, and $n_{i}$ denotes the integer vortex winding number attached to $\Delta_{\mathbf{Q}_{i}}$.
For a general BdG Hamiltonian
\[
H=\sum_{\mathbf{k}\in T^{2}}C_{\mathbf{k}}^{\dagger}\hat{H}_{0}(\mathbf{k})C_{\mathbf{k}}+\sum_{i,\mathbf{k}}c_{\mathbf{Q}_{i}+\mathbf{k},\alpha}^{\dagger}\hat{\Delta}^{\alpha\beta}_{\mathbf{Q}_{i}}c_{\mathbf{Q}_{i}-\mathbf{k},\beta}^{\dagger}+\mathrm{h.c.},
\]
where $C_{\mathbf{k}}=(c_{\alpha,\mathbf{k}},c_{\beta,\mathbf{k}},\cdots,c_{\gamma,\mathbf{k}},\cdots)^{\text{T}}$,
with $\alpha$ incorporating the band and spin indices. Suppose that the normal band Hamiltonian
$\hat{H}_{0}(\mathbf{k})$ has multiple bands $\epsilon_{n}(\mathbf{k})$,
and only one of such bands, with (normalized) eigenvector $u_{\mathbf{k}}$,
cuts the chemical potential, and we call this band middle band. The
middle band gives rise to $N$ Fermi surfaces with possible
Berry phases, and $\hat{\Delta}_{\mathbf{Q}_{i}}$ are pairing terms that
can fully gap the whole system. The following results can be easily
generalized to the system with multiple middle bands.
We assume that each Fermi surface is circular and centered at some
momentum $\mathbf{Q}_{i}/2$, otherwise one can always continuously
deform the original Hamiltonian to the current form without gap closing.
One can imagine that each Fermi surface is equipped with a PDW order
parameter $\hat{\Delta}_{\mathbf{Q}_{i}}$ and we further assume that
\emph{the system is fully gapped }\textit{only if}\emph{ all the $\hat{\Delta}_{\mathbf{Q}_{i}}$'s
are non-vanishing}. Here we consider some of $\hat{\Delta}_{\mathbf{Q}_{i}}$'s
acquire a winding $\hat{\Delta}_{\mathbf{Q}_{i}}\rightarrow\hat{\Delta}_{\mathbf{Q}_{i}}e^{-\text{i}n_{i}\phi}$
with $\phi\in[0,2\pi)$ and $n_{i}\in\mathbb{Z}$. Together with the 2D physical space, the bulk BdG Hamiltonian can then be written down in a synthetic 3D torus $T^{3}=T^{2}\times S^{1}$ spanned by $(\mathbf{k},\phi)$.
\begin{figure}[!ht]
\subfloat[]{\includegraphics[scale=0.7]{s01a}}
\subfloat[]{\includegraphics[scale=0.5]{s01b}}
\caption{\label{fig:Schematic} (a) Schematic diagram of the real smooth function
$\Omega_{\mathbf{Q}_{i}}(\mathbf{k})$. (b) Schematic diagram of $\vec{S}_{i}$ and $\mathcal{\vec{F}}_{i}$, the black lines denote the Fermi surfaces, the region enclosed by black line denotes $\vec{S}$, and the green area denotes the patch $\mathcal{\vec{F}}$.}
\end{figure}
Consider a continuous deformation $H\rightarrow H'$, with $H'\equiv H[\hat{\Delta}_{\mathbf{Q}_{i}}\rightarrow\hat{\Delta}_{\mathbf{Q}_{i}}\Omega_{\mathbf{Q}_{i}}(\mathbf{k})]$,
where $\Omega_{\mathbf{Q}_{i}}(\mathbf{k})$ is a positive real smooth
truncation function with $\Omega_{\mathbf{Q}_{i}}(\mathbf{k}\in\vec{S}_{i})=1$,
and decays to zero at a short distance from the Fermi surface, as
shown in Fig.\ \ref{fig:Schematic}(a). We denote the region with
$\Omega_{\mathbf{Q}_{i}}(\mathbf{k})\neq0$ as $\vec{\mathcal{F}}=\bigsqcup_{i}\mathcal{\vec{F}}_{i}$
(disjoint union of $\mathcal{N}$ orientable areas $\mathcal{\vec{F}}_{i}$)
and $\vec{\bar{\mathcal{F}}}=T^{2}-\mathcal{\vec{F}}$, note that
$\vec{S}_{i}\subset\mathcal{\vec{F}}_{i}$, as shown in Fig.\ \ref{fig:Schematic}(b).
Since the whole system remains fully gapped for a continuous deformation
$H\rightarrow H'$, we have (``mod 2'' temporarily omitted)
\[
\nu_{3}=-\frac{1}{4\pi^{2}}\int_{T^{2}\times S^{1}}\mathcal{Q}_{3}[H]=-\frac{1}{4\pi^{2}}\int_{T^{2}\times S^{1}}\mathcal{Q}_{3}[H'].
\]
One can easily see that
\[
\int_{\vec{\bar{\mathcal{F}}}\times S^{1}}\mathcal{Q}_{3}[H']=0,
\]
since in the region $\vec{\bar{\mathcal{F}}}$, $\mathcal{A}_{\phi}=0$
and $\partial_{\phi}\mathcal{A}_{k_{x},k_{y}}=0$, then $\mathcal{A}d\mathcal{A}=\text{Tr}[\mathcal{A}_{k_{x}}(\partial_{k_{y}}\mathcal{A}_{\phi}-\partial_{\phi}\mathcal{A}_{k_{y}})+\mathcal{A}_{k_{y}}(\partial_{\phi}\mathcal{A}_{k_{x}}-\partial_{k_{x}}\mathcal{A}_{\phi})+\mathcal{A}_{\phi}(\partial_{k_{x}}\mathcal{A}_{k_{y}}-\partial_{k_{y}}\mathcal{A}_{k_{x}})]dk_{x}\wedge dk_{y}\wedge d\phi=0$,
also $\mathcal{A}^{3}\sim\mathrm{Tr[}\mathcal{A}_{k_{x}}\mathcal{A}_{k_{y}}\mathcal{A}_{\phi}-\mathcal{A}_{k_{y}}\mathcal{A}_{k_{x}}\mathcal{A}_{\phi}]dk_{x}\wedge dk_{y}\wedge d\phi=0$,
hence,
\[
\nu_{3}=-\frac{1}{4\pi^{2}}\int_{\mathcal{\vec{F}}\times S^{1}}\mathcal{Q}_{3}[H']=-\frac{1}{4\pi^{2}}\sum_{i}\int_{\mathcal{\vec{F}}_{i}\times S^{1}}\mathcal{Q}_{3}[H'].
\]
While in general the Hamiltonian $\hat{H}_{0}$ incorporates multiple normal bands, we can consider the weak SC pairing regime, in which case only the states around each Fermi surface will be effectively paired up. Ignoring the pairing between a state around the Fermi surface and that from other bands does not affect the topology of the system. In this way, the BdG $H'$ further reduces to an effective one band Hamiltonian projected to the middle
band. For $\mathcal{\vec{F}}_{i}$, the effective BdG Hamiltonian takes the form
\[
h_{\mathbf{Q}_{i}}(\mathbf{k},\phi)=\left[\begin{array}{cc}
\epsilon_{\mathbf{Q}_{i}+\mathbf{k}} & \Delta_{\mathbf{Q}_{i}}(\mathbf{k})\Omega_{\mathbf{Q}_{i}}(\mathbf{k})e^{-\text{i}n_{i}\phi}\\
\Delta_{\mathbf{Q}_{i}}^{*}(\mathbf{k})\Omega_{\mathbf{Q}_{i}}(\mathbf{k})e^{\text{i}n_{i}\phi} & -\epsilon_{\mathbf{Q}_{i}-\mathbf{k}}
\end{array}\right],
\]
where $\Delta_{\mathbf{Q}_{i}}\equiv\langle u_{\mathbf{k}}|\hat{\Delta}_{\mathbf{Q}_{i}}|u^{*}_{-\mathbf{k}}\rangle$ is the superconductor order parameter
projected onto the $i$-th Fermi surface. Furthermore, in each $\mathcal{\vec{F}}_{i}$,
only one $\Delta_{\mathbf{Q}_{i}}$ is non-vanishing and captures
the Berry curvature in the corresponding $\mathcal{\vec{F}}_{i}$
only, so the Chern-Simons invariant $\nu_{3}$ can be written as a
sum of the corresponding ``Chern-Simons invariant'' in different
$\mathcal{\vec{F}}_{i}$. Since $\mathcal{A}_{k_{x},k_{y},\phi}$
are local functions of $(k_{x},k_{y},\phi)$, and for effective only
one band, $\mathcal{A}_{k_{x},k_{y},\phi}$ are just numbers, hence
$\mathcal{A}^{3}=0$, thus we can decompose $\nu_{3}$ as
\[
\nu_{3}=-\frac{1}{4\pi^{2}}\sum_{i}\int_{\mathcal{\vec{F}}_{i}\times S^{1}}\mathcal{Q}_{3}[h_{\mathbf{Q}_{i}}(\mathbf{k},\phi)]=-\frac{1}{4\pi^{2}}\sum_{i}\int_{\mathcal{\vec{F}}_{i}\times S^{1}}\mathcal{A}d\mathcal{A}[h_{\mathbf{Q}_{i}}].
\]
In this case, the Chern-Simons invariant reduces to the Hopf invariant
that captures the linking number of the inverse images of two points
in the target space $S^{2}$ of $h_{\mathbf{Q}_{i}}(\mathbf{k},\phi)$.
For brevity, we consider below a particular $\mathcal{\vec{F}}=\mathcal{\vec{F}}_{i}$
for some $i$ and ignore the subscript $i$. To proceed, one needs the eigenvectors of $h_{\mathbf{Q}}(\mathbf{k},\phi)$.
In the original basis, the eigenvectors for the middle band are of
the form $|\psi_{\mathbf{k}\pm}\rangle=(\alpha_{\mathbf{k}\pm}u_{\mathbf{k}},\beta_{\mathbf{k}\pm}u_{-\mathbf{k}}^{*})^{\text{T}}$,
where $(\alpha_{\mathbf{k}\pm},\beta_{\mathbf{k}\pm})^{\text{T}}$
are the eigenvectors of $h_{\mathbf{Q}}(\mathbf{k},\phi)$ written in the
eigen-basis. For $h_{\mathbf{Q}}(\mathbf{k},\phi)$, there are nonetheless
two choices of eigenvectors (without normalization),
\begin{equation}
\left(\begin{array}{c}
\alpha_{+,\mathbf{k}\pm}\\
\beta_{+,\mathbf{k}\pm}
\end{array}\right)\propto\left(\begin{array}{c}
\xi_{\mathbf{k}}\pm\sqrt{\xi_{\mathbf{k}}^{2}+|\Delta_{\mathbf{Q}}(\mathbf{k})\Omega_{\mathbf{Q}}(\mathbf{k})|^{2}}\\
\Delta_{\mathbf{Q}}^{*}(\mathbf{k})\Omega_{\mathbf{Q}}(\mathbf{k})e^{\text{i}n\phi}
\end{array}\right)\label{eq:hole}
\end{equation}
or
\begin{equation}
\left(\begin{array}{c}
\alpha_{-,\mathbf{k}\pm}\\
\beta_{-,\mathbf{k}\pm}
\end{array}\right)\propto\left(\begin{array}{c}
\Delta_{\mathbf{Q}}(\mathbf{k})\Omega_{\mathbf{Q}}(\mathbf{k})e^{-\text{i}n\phi}\\
-\xi_{\mathbf{k}}\pm\sqrt{\xi_{\mathbf{k}}^{2}+|\Delta_{\mathbf{Q}}(\mathbf{k})\Omega_{\mathbf{Q}}(\mathbf{k})|^{2}}
\end{array}\right),\label{eq:electron}
\end{equation}
where $\xi_{\mathbf{k}}\equiv\frac{\epsilon_{\mathbf{Q}+\mathbf{k}}+\epsilon_{\mathbf{Q}-\mathbf{k}}}{2}$.
One can check that both of them are the eigenvectors of $h_{\mathbf{Q}}(\mathbf{k},\phi)$.
Actually, for Fermi surfaces, there are two cases with $\xi_{\mathbf{k}}>0$
or $\xi_{\mathbf{k}}<0$ inside the region $\vec{S}$. For these two
different cases, we choose different eigenvector, i.e., Eq. \eqref{eq:hole}
for $\xi_{\mathbf{k}}>0$ and Eq. \eqref{eq:electron} for $\xi_{\mathbf{k}}<0$.
The Berry connections are $\mathcal{A}_{\eta}=\text{i}\langle\psi_{\mathbf{k}-}|\partial_{\eta}|\psi_{\mathbf{k}-}\rangle$,
with $\eta=k_{x},k_{y},\phi$, or explicitly
\begin{align*}
\mathcal{A}_{\pm,\phi} & =\text{i}(\alpha_{\pm,\mathbf{k}-}^{*}u_{\mathbf{k}}^{\dagger},\beta_{\pm,\mathbf{k}-}^{*}u_{-\mathbf{k}}^{\text{T}})\partial_{\phi}\left(\begin{array}{c}
\alpha_{\pm,\mathbf{k}-}u_{\mathbf{k}}\\
\beta_{\pm,\mathbf{k}-}u_{-\mathbf{k}}^{*}
\end{array}\right)\\
& =\begin{cases}
\text{i}\beta_{+,\mathbf{k}-}^{*}\partial_{\phi}\beta_{+,\mathbf{k}-}=-n|\beta_{+,\mathbf{k}-}|^{2}\text{\ensuremath{\xrightarrow{\Delta_{\mathbf{Q}}\rightarrow0}}}-n\Theta_{\vec{S}}, & \text{for }\xi_{\mathbf{k}}>0\text{ in }\vec{S}\\
\text{i}\alpha_{-,\mathbf{k}-}^{*}\partial_{\phi}\alpha_{-,\mathbf{k}-}=+n|\alpha_{-,\mathbf{k}-}|^{2}\text{\ensuremath{\xrightarrow{\Delta_{\mathbf{Q}}\rightarrow0}}}+n\Theta_{\vec{S}}, & \text{for }\xi_{\mathbf{k}}<0\text{ in }\vec{S}
\end{cases},
\end{align*}
\begin{align*}
\mathcal{A}_{\pm,\mathbf{k}} & \equiv(\mathcal{A}_{\pm,k_{x}},\mathcal{A}_{\pm,k_{y}})=\text{i}(\alpha_{\pm,\mathbf{k}-}^{*}u_{\mathbf{k}}^{\dagger},\beta_{\pm,\mathbf{k}-}^{*}u_{-\mathbf{k}}^{\text{T}})\nabla_{\mathbf{k}}\left(\begin{array}{c}
\alpha_{\pm,\mathbf{k}-}u_{\mathbf{k}}\\
\beta_{\pm,\mathbf{k}-}u_{-\mathbf{k}}^{*}
\end{array}\right)\\
& =\text{i}(|\alpha_{\pm,\mathbf{k}-}|^{2}-|\beta_{\pm,\mathbf{k}-}|^{2})u_{\mathbf{k}}^{\dagger}\nabla_{\mathbf{k}}u_{\mathbf{k}}+\text{i}(\alpha_{\pm,\mathbf{k}-}^{*}\nabla_{\mathbf{k}}\alpha_{\pm,\mathbf{k}-}+\beta_{\pm,\mathbf{k}-}^{*}\nabla\beta_{\pm,\mathbf{k}-})\\
& =\pm(1-2\Theta_{\vec{S}})\mathcal{A}_{0,\mathbf{k}}+\mathcal{A}_{\pm,1,\mathbf{k}},
\end{align*}
here we have used the trick $\Delta_{\mathbf{Q}}\to0$ without closing
the bulk gap, the upper (lower) sign means $\xi_{\mathbf{k}}>0$ ($\xi_{\mathbf{k}}<0$)
inside the region $\vec{S}$, $\Theta_{\vec{S}}\equiv\Theta(\xi_{\mathbf{k}}>0)$
($\Theta(\xi_{\mathbf{k}}<0)$) for the case $\xi_{\mathbf{k}}>0$
($\xi_{\mathbf{k}}<0$), and denotes the step function is 1 inside
the region $\vec{S}$ and 0 else, and $\nabla_{\mathbf{k}}\equiv(\partial_{k_{x}},\partial_{k_{y}})$.
In the last line, we denote $\mathcal{A}_{0,\mathbf{k}}\equiv\text{i}u_{\mathbf{k}}^{\dagger}\nabla_{\mathbf{k}}u_{\mathbf{k}}$,
and
\begin{align*}
\mathcal{A}_{\pm,1,\mathbf{k}} & \equiv\text{i}(\alpha_{\pm,\mathbf{k}-}^{*}\nabla_{\mathbf{k}}\alpha_{\pm,\mathbf{k}-}+\beta_{\pm,\mathbf{k}-}^{*}\nabla_{\mathbf{k}}\beta_{\pm,\mathbf{k}-})\\
& =\pm\text{i}\frac{\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}}\nabla_{\mathbf{k}}(\Delta_{\mathbf{Q}}^{*}\Omega_{\mathbf{Q}})-\Delta_{\mathbf{Q}}^{*}\Omega_{\mathbf{Q}}\nabla_{\mathbf{k}}(\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}})}{4\sqrt{\xi_{\mathbf{k}}^{2}+\Omega_{\mathbf{Q}}^{2}|\Delta_{\mathbf{Q}}|^{2}}\left(\sqrt{\xi_{\mathbf{k}}^{2}+\Omega_{\mathbf{Q}}^{2}|\Delta_{\mathbf{Q}}|^{2}}\mp\xi_{\mathbf{k}}\right)}\\
& \text{\ensuremath{\xrightarrow{\Delta_{\mathbf{Q}}\rightarrow0}\pm}}\text{i}\Theta_{\vec{S}}\frac{\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}}\nabla_{\mathbf{k}}(\Delta_{\mathbf{Q}}^{*}\Omega_{\mathbf{Q}})-\Delta_{\mathbf{Q}}^{*}\Omega_{\mathbf{Q}}\nabla_{\mathbf{k}}(\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}})}{2|\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}}|^{2}}.
\end{align*}
Let's denote $\vec{v}=\text{i}\frac{\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}}\nabla_{\mathbf{k}}(\Delta_{\mathbf{Q}}^{*}\Omega_{\mathbf{Q}})-\Delta_{\mathbf{Q}}^{*}\Omega_{\mathbf{Q}}\nabla_{\mathbf{k}}(\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}})}{2|\Delta_{\mathbf{Q}}\Omega_{\mathbf{Q}}|^{2}}$,
direct substitution of $\Delta_{\mathbf{Q}}=\Delta_{\mathbf{k}}e^{\text{i}\theta_{\mathbf{k}}}$
seems to give $\vec{v}=\nabla_{\mathbf{k}}\theta_{\mathbf{k}}$. But
note that $\theta_{\mathbf{k}}$ is a multivalued field, for a general
SC order parameter with phase winding $m$, there will be a branch
cut from $2\pi m\to0$. Consider the region close to the branch cut,
$\theta_{\mathbf{k}}$ behaves like $2\pi m\Theta$, where $\Theta$
is a step function, so the correct result should be \cite{Kleinert2008}
\[
\vec{v}=\nabla_{\mathbf{k}}\theta_{\mathbf{k}}+\mathbf{A}_{d},
\]
where the vector field $\mathbf{A}_{d}$ is the \emph{defect gauge
field} compensating for the discontinuity in $\theta_{\mathbf{k}}$
and the properties are
\begin{align}
\nabla_{\mathbf{k}}\times\nabla_{\mathbf{\mathbf{k}}}\theta_{\mathbf{k}} & =0,\label{eq:defect1}\\
\oint_{\partial\vec{S}}\nabla_{\mathbf{k}}\theta_{\mathbf{k}}\cdot d\mathbf{k} & =2\pi m,\label{eq:defect2}\\
\nabla_{\mathbf{k}}\times\mathbf{A}_{d} & =2\pi m\delta(\mathbf{k})\hat{z}=\hat{z}\delta(\mathbf{k})\oint_{\partial\vec{S}}\nabla_{\mathbf{\mathbf{k}}}\theta_{\mathbf{k}}\cdot d\mathbf{k},\label{eq:defect3}\\
\oint_{\partial\vec{S}}\mathbf{A}_{d}\cdot d\mathbf{k} & =\begin{cases}
0 & \text{if }\partial\text{ avoids the branch cut}\\
2\pi m & \text{otherwise}
\end{cases},\label{eq:defect4}
\end{align}
where $\partial\vec{S}$ is the boundary of the region containing
the origin. Eq. \eqref{eq:defect3} is corresponded to the boundary
of the branch cut. In order to make these results consistent, the
chain rule of differentiation must be modified to
\[
\nabla_{\mathbf{k}}e^{\text{i}\theta_{\mathbf{k}}}=\text{i}(\nabla_{\mathbf{k}}\theta_{\mathbf{k}}+\mathbf{A}_{d})e^{\text{i}\theta_{\mathbf{k}}}.
\]
Note that $\vec{v}$ is invariant under the gauge transformations
\begin{align*}
\theta_{\mathbf{k}} & \to\theta_{\mathbf{k}}+\lambda_{\mathbf{k}},\\
\mathbf{A}_{d} & \to\mathbf{A}_{d}-\nabla_{\mathbf{k}}\lambda_{\mathbf{k}}.
\end{align*}
An example is that, for a $p+\text{i}p$ SC, the SC order parameter
can be expressed as $\Delta=\Delta_{\mathbf{k}}e^{\text{i}\phi_{\mathbf{k}}}$,
with the azimuthal angle $\phi_{\mathbf{k}}\in[0,2\pi)$. In this
case, the field $\theta_{\mathbf{k}}=\phi_{\mathbf{k}}$ is discontinuous
over the positive $k_{x}$-axis, thus $\nabla_{\mathbf{k}}\theta_{\mathbf{k}}=\hat{\phi}/k-2\pi\Theta(k_{x})\delta(k_{y})\hat{k}_{y}$.
Note that we expect $\vec{v}=\hat{\phi}/k$, hence $\vec{v}=\nabla_{\mathbf{k}}\theta_{\mathbf{k}}+\mathbf{A}_{d}\text{, and }\mathbf{A}_{d}=2\pi\Theta(k_{x})\delta(k_{y})\hat{k}_{y}$.
Now, return to our proof for the Chern-Simons invariant. Consider
the defect gauge field, we have
\[
\mathcal{A}_{\pm,1,\mathbf{k}}=\pm\Theta_{\vec{S}}(\nabla_{\mathbf{k}}\text{arg}\Delta_{\mathbf{Q}}+\mathbf{A}_{d}).
\]
In our effective one band case, since $\mathcal{A}d\mathcal{A}=[\mathcal{A}_{k_{x}}(\partial_{k_{y}}\mathcal{A}_{\phi}-\partial_{\phi}\mathcal{A}_{k_{y}})+\mathcal{A}_{k_{y}}(\partial_{\phi}\mathcal{A}_{k_{x}}-\partial_{k_{x}}\mathcal{A}_{\phi})+\mathcal{A}_{\phi}(\partial_{k_{x}}\mathcal{A}_{k_{y}}-\partial_{k_{y}}\mathcal{A}_{k_{x}})]dk_{x}\wedge dk_{y}\wedge d\phi$,
and $\partial_{\phi}\mathcal{A}_{k_{x},k_{y}}=0$, one can readily
show that
\begin{equation*}
\nu_{3}=\sum_{i}\nu^{(i)}_{3}\;\mathrm{mod}\;2,
\end{equation*}
and
\begin{align*}
\nu^{(i)}_{3} & =-\frac{1}{2\pi}\left(\frac{1}{2\pi}\int_{0}^{2\pi}d\phi\right)\int_{\mathcal{\vec{F}}_{i}}\left[\mathcal{A}_{i,\pm,\phi}\nabla_{\mathbf{k}}\times\mathcal{A}_{i,\pm,\mathbf{k}}+\mathcal{A}_{i,\pm,\mathbf{k}}\times\nabla_{\mathbf{k}}\mathcal{A}_{i,\pm,\phi}\right]\cdot d^{2}\mathbf{k}\\
& =\pm n_{i}\frac{1}{2\pi}\int_{\mathcal{\vec{F}}_{i}}\left[\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times\mathcal{A}_{i,\pm,\mathbf{k}}+\mathcal{A}_{i,\pm,\mathbf{k}}\times\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}\right]\cdot d^{2}\mathbf{k}.
\end{align*}
For the part involving $\mathcal{A}_{i,0,\mathbf{k}}$, we have
\begin{align*}
\nu_{3}^{(i)[\mathcal{A}_{i,0,\mathbf{k}}]} & =\frac{n_{i}}{2\pi}\int_{\mathcal{\vec{F}}_{i}}\left[\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times[(1-2\Theta_{\vec{S}_{i}})\mathcal{A}_{i,0,\mathbf{k}}]+(1-2\Theta_{\vec{S}_{i}})\mathcal{A}_{i,0,\mathbf{k}}\times\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}\right]\cdot d^{2}\mathbf{k}\\
& =-\frac{n_{i}}{2\pi}\int_{\mathcal{\vec{F}}_{i}}[\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times\mathcal{A}_{i,0,\mathbf{k}}-\mathcal{A}_{i,0,\mathbf{k}}\times\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}+2\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}\times\mathcal{A}_{i,0,\mathbf{k}}\\
&\hspace{14pt}+2\Theta_{\vec{S}_{i}}\mathcal{A}_{i,0,\mathbf{k}}\times\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}]\cdot d^{2}\mathbf{k}\\
& =-\frac{n_{i}}{2\pi}\left(\int_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times\mathcal{A}_{i,0,\mathbf{k}}\cdot d^{2}\mathbf{k}-\oint_{\partial\vec{S}_{i}}\mathcal{A}_{i,0,\mathbf{k}}\cdot d\mathbf{k}\right).
\end{align*}
Since the Chern-Simons invariant is gauge-independent, we can consider
a smooth gauge, using the Stokes' theorem, $\nu_{3}^{(i)[\mathcal{A}_{i,0,\mathbf{k}}]}$
just vanishes.
Next we need to consider for $\pm\Theta_{\vec{S}_{i}}(\nabla_{\mathbf{k}}\text{arg}\Delta_{\mathbf{Q}_{i}}+\mathbf{A}_{i,d})$,
which can be divided into two parts $\pm\Theta_{\vec{S}_{i}}\mathbf{A}'_{i}$,
with $\mathbf{A}'_{i}=\nabla_{\mathbf{k}}\text{arg}\Delta_{\mathbf{Q}_{i}}$
or $\mathbf{A}_{i,d}$,
\begin{align*}
\nu_{3}^{(i)[\mathbf{A}'_{i}]} & =\frac{n_{i}}{2\pi}\int_{\mathcal{\vec{F}}_{i}}\left[\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times[\Theta_{\vec{S}_{i}}\mathbf{A}'_{i}]+\Theta_{\vec{S}_{i}}\mathbf{A}'_{i}\times\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}\right]\cdot d^{2}\mathbf{k}\\
& =\frac{n_{i}}{2\pi}\int_{\mathcal{\vec{F}}_{i}}\left[\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times\mathbf{A}'_{i}+\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}\times\mathbf{A}'_{i}+\Theta_{\vec{S}_{i}}\mathbf{A}'_{i}\times\nabla_{\mathbf{k}}\Theta_{\vec{S}_{i}}\right]\cdot d^{2}\mathbf{k}\\
& =\frac{n_{i}}{2\pi}\int_{\mathcal{\vec{F}}_{i}}\Theta_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times\mathbf{A}'_{i}\cdot d^{2}\mathbf{k}=\frac{n_{i}}{2\pi}\int_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times\mathbf{A}'_{i}\cdot d^{2}\mathbf{k}.
\end{align*}
Since $\nabla_{\mathbf{k}}\times\nabla_{\mathbf{k}}\text{arg}\Delta_{\mathbf{Q}_{i}}=0$,
$\nu_{3}^{(i)[\nabla_{\mathbf{k}}\text{arg}\Delta_{\mathbf{Q}_{i}}]}$ vanishes.
Next, by the property \eqref{eq:defect3} of $\mathbf{A}_{i,d}$,
\begin{equation*}
\nu_{3}^{(i)[\mathbf{A}_{i,d}]}=\frac{n_{i}}{2\pi}\int_{\vec{S}_{i}}\nabla_{\mathbf{k}}\times\mathbf{A}_{i,d}\cdot d^{2}\mathbf{k}=\frac{n_{i}}{2\pi}\oint_{\partial\vec{S}_{i}}\nabla_{\mathbf{k}}\text{arg}\Delta_{\mathbf{Q}_{i}}\cdot d\mathbf{k}\int_{\vec{S}_{i}}\delta(\mathbf{k})\hat{z}\cdot d^{2}\mathbf{k}=\frac{n_{i}}{2\pi}\oint_{\partial\vec{S}_{i}}\nabla_{\mathbf{k}}\text{arg}\Delta_{\mathbf{Q}_{i}}\cdot d\mathbf{k}.
\end{equation*}
Note that this is \emph{not} Stokes' theorem. Gathering
all the results together, we have
\[
\nu_{3}=\frac{1}{2\pi}\sum_{i}n_{i}\oint_{\partial\vec{S}_{i}}\nabla_{\mathbf{k}}\arg\Delta_{\mathbf{Q}_{i}}(\mathbf{k})\cdot d\mathbf{k}\;\mathrm{mod}\;2.
\]
Note that the pairing $\Delta_{\mathbf{Q}_{i}}(\mathbf{k})$ has captured
the band topology. For the case with SC pairings between two different
Fermi surfaces, using the similar arguments as in \cite{LiuX-J2017},
the formula is still applicable. This completes the proof.
\end{document}
|
\section*{Acknowledgment}
We greatly thank Y. Tanaka, P. Burset, and R. S. Deacon for fruitful discussions.
This work was partially supported by Grant-in-Aid for Young Scientific Research (A) (No. JP15H05407), Grant-in-Aid for Scientific Research (A) (No. JP16H02204), Grant-in-Aid for Scientific Research (S) (No. JP26220710), JSPS Research Fellowship for Young Scientists (No. JP14J10600), JSPS Program for Leading Graduate Schools (MERIT) from JSPS, Grant-in-Aid for Scientific Research on Innovative Area, "Nano Spin Conversion Science" (No.JP15H01012), Grant-in-Aid for Scientific Research on Innovative Area, "Topological Materials Science" (Grant No. JP16H00984) from MEXT, CREST, and the Murata Science Foundation.
\clearpage
\section*{Supplemental Material}
\begin{figure}[t]
\centering \includegraphics[width=0.45\linewidth]{FigSstack.pdf} \caption{Schematics of the InAs heterostructure material stack. The2DEG exists in the 4 nm-thick InAs QW.}
\label{stack}
\end{figure}
\begin{figure}[t]
\centering \includegraphics[width=0.75\linewidth]{nbti_S.pdf} \caption{Measurements of the critical field of a superconducting NbTi 150nm thin film. $dV/dI$ as functions of magnetic field and bias current measured at 2.0 K. The NbTi holds superconductivity when $dV/dI$ is equal
to $0$. The critical field is 7.0 T.}
\label{nbti}
\end{figure}
\begin{figure}[t]
\centering \includegraphics[width=0.95\linewidth]{Figs2_S.pdf} \caption{Conductance as functions of magnetic field and top gate voltage.}
\label{figs2}
\end{figure}
\begin{figure}[t]
\centering \includegraphics[width=0.95\linewidth]{Figs3.pdf} \caption{Transconductance as functions of bias voltage and top gate voltage obtained at 2.4 T.}
\label{figs3}
\end{figure}
\begin{figure}[t]
\centering \includegraphics[width=0.95\linewidth]{Figs4.pdf}
\caption{Typical results of the $dI/dV$ vs $V_{\rm sd}$ in the plateau regime obtained at 4 T are shown. The result on the $\nu = 1$ and 2 plateau is shown in (a) and (b), respectively.}
\label{figs4}
\end{figure}
\begin{figure}[t]
\centering \includegraphics[width=0.95\linewidth]{Figs1.pdf} \caption{The calculated spin-triplet superconducting proximity gap energy is
shown. The values are $\sim$0.1 meV.}
\label{figs1}
\end{figure}
\begin{figure}[t]
\centering \includegraphics[width=0.75\linewidth]{Figs_Rxx.eps} \caption{The sheet resistance as a function of magnetic field is shown. The lines are incrementally shifted for clarity. We
measured the resistance in the various carrier density by tuning the
gate voltage. We subtract the peak height of the resistance around
4 T.}
\label{figsR}
\end{figure}
\clearpage
\subsection{Device fabrication}
We used an InAs QW grown by molecular beam epitaxy with the density $3\times10^{11}~{\rm cm^{-2}}$
and the mobility $3\times10^{5}~{\rm cm^{2}V^{-1}s^{-1}}$.
The 2DEG is formed in the 4 nm-thick
InAs layer. The material stack of the InAs heterostructure is schematically shown in Fig.~\ref{stack}.
A mesa was first defined in the substrate by wet etching
with an etchant based on ${\rm H_{3}PO_{4}}$. Then, NbTi was
sputtered to form the superconducting electrodes on the mesa edges,
following a procedure of wet etching to make the clean edge
exposed, sulfur passivation to avoid oxidization of the edge, and
in-situ Ar plasma cleaning. Finally a gate electrode metal
of Titanium and Gold was deposited on top to address the low QH filling
regime even under a low magnetic field. The top gate
is placed on an insulating layer made from cross-linked PMMA.
This fabrication procedure creates no superconducting
material on the top surface of the mesa.
This is specially devised to control the carrier density not only
of the mesa but also near the junction using the top gate voltage
(discussed later).
The two junctions are separated by 20 ${\rm \mu}$m, so this device is assumed to have two independent contact regions.
\subsection{NbTi superconductivity}
To characterize the superconducting properties of
the NbTi, we performed a current-bias measurement of the
differential resistance $dV/dI$ at
various out-of-plane magnetic fields $B$ for a 150 nm-thick NbTi thin film device
at 2 K which is lower than the NbTi critical temperature of 6.5
K. The measured data shows a supercurrent branch as $dV/dI$=0 k${\rm \Omega}$ in dark
purple near the zero current bias in Fig.~\ref{nbti}.
From this data, we evaluate the
critical field of $B=$7.0 T. Herein, the coexistence of superconducting
state and QH states can be realized if the 2DEG is in
the spin-resolved QH regime for $B<7.0$ T.
\subsection{Superconducting proximity at 0 T}
In Fig. 2(a) and (c) in the main text, there are dip structures around $V_{\rm sd}=\pm 0.7$ mV. These dip structures cannot be expected from the normal BTK model. These dip structures have also been reported in experimental studies of junctions of three dimensional topological insulators and superconductor~\cite{snelderarxiv2015}. Some theoretical works predict existence of spin-triplet superconducting proximity on such junctions which generate the dip structures~\cite{bursetprb2014, bursetprb2015}. In our case, strong spin-orbit interaction on the interface can affect the superconducting proximity even at 0 T and invokes finite spin-flip process. Therefore, we suspect that the dip structure may be related to the spin-triplet superconducting proximity even at 0 T.
\subsection{Transconductance in the QH regime}
We measured the conductance as functions of magnetic field and top
gate voltage in order to estimate quality of our InAs quantum Hall
effect and Zeeman effect. Figure \ref{figs2} shows the conductance
as functions of $B$ and $V_{{\rm tg}}$. As seen in this figure,
clear quantized conductance plateaus appear at $B>0.6$~T and the
Zeeman splitting is found at $B>\sim1$~T. 2.4 and 4 T is large enough
to study the coexistence between the spin-resolved QH state and superconductor.
Furthermore, we measured the transconductance defined by the deviation
of $dI/dV$ at 2.4 T. The results are plotted in Fig.~\ref{figs3}.
The diamond-shaped structure can be found. We focus on the bright
areas, namely the plateau-transition regime in the main text.
\subsection{Sub-gap feature in QH plateau regime}
In the main text, we focus on the sub-gap feature in QH plateau-transition
regime. In this section, we show the measured $dI/dV$ vs $V_{\rm sd}$
in the plateau regime. The typical results obtained at 4 T are shown in Fig.~\ref{figs4}.
Panel (a) and (b) indicates the result on the $\nu = 1$ and 2 plateau, respectively.
Even in the plateau regime, there is a sub-gap feature with a small dip or gap-like structure. These results
are similar to the sub-gap features observed for $\Delta g < 0.4$ of Fig. 3(b) and (c) in the main text.
The estimated gap from these sub-gap features are nearly equal to the 0.35 mV calculated from the superconducting gap at 0 T.
This supports our assumption in the main text that the spin-singlet superconducting proximity gap at 4 T is the same as the superconducting gap obtained at 0 T.
We conclude that the QH edge states don't contribute to the Andreev
reflection in the device.
\subsection{Numerical calculation}
We executed the numerical calculation using the model in which we
assume the two channels, $\alpha$ and $\beta$. The fitting parameters are the superconducting gaps $\Delta_{\alpha}$
and $\Delta_{\beta}$, barrier strength $Z_{\alpha}$ ($Z_{\beta}=1$),
normal state interface conductance $G_{n}$, effective temperature
$\omega$, the relative contribution of channel $\alpha$, $P$ and offset
conductance $G_{{\rm offset}}$. We define $G_{{\rm int}}^{\alpha\beta}(V_{{\rm sd}},Z_{\alpha},T,\Delta_{\alpha},\Delta_{\beta})$
as eqn. (2) in the main text. Then the fitting function can be written
as
\[
\Bigl(\frac{1}{G_{n}\cdot G_{{\rm int}}^{\alpha\beta}(V_{{\rm sd}},Z_{\alpha},T,2\Delta_{\alpha},2\Delta_{\beta})}+\frac{1}{G_{{\rm offset}}}\Bigr)^{-1}.
\]
We take care to confirm that our device has two independent superconducting-QH bulk
state junctions. Therefore, the sub-gap features and the position in $V_{\rm sd}$
of the side peaks are consistent with $2\Delta_{\alpha}$ and $2\Delta_{\beta}$.
To account for the effective temperature, we approximated the deviation of
the Fermi-Dirac distribution function in eqn.~(1) in the main text
as the Gaussian function, $\exp{(-((E-V_{{\rm sd}})/2\omega)^{2})}$,
where $\omega$ is ideally equal to $T$ but now $\omega$ includes
the broadening due to inelastic scattering, inhomogeneity of the gap
and the local heating~\cite{woodsprb2004,degravenl2011}. To execute
the fitting, we constrict the fitting ranges for all the parameters,
and especially we tightly constrict the $\Delta_{\alpha}$ and $\Delta_{\beta}$
from the curve shapes. In order to reproduce the curve shape around
the zero bias voltage, we changed the fitting range for each of the
curves because the differential conductance of the 2DEG appears as
background and the conductance has a large dependence on the bias
voltage near the plateau regime. Due to this background dependence,
we could not reproduce the curve shapes in the two lower curves of the
left panel and middle panel, and the lowest curve of the right panel
in Fig.4(b) of the main text. In these cases, we evaluated only $\Delta_{\alpha}$
and $\Delta_{\beta}$ from the sub-gap peak features ($2\Delta_{\alpha}$ and
$2\Delta_{\beta}$ are indicated on the panels in Fig.4(b) as open
and closed hexagons).
Our fitting scheme includes many free parameters and results are sensitive to the constriction of the variable range. However, the estimated gap energies and $P$ produce relatively constant results with different fitting ranges, so we think it is valuable to discuss these parameters.
All fits are executed with a genetic algorithm (GenCurvefit package for Igor Pro).
\subsection{Calculation of the proximity gap energy}
As written in the main text, we analyzed our experimental data with
the model to evaluate $\Delta_{\alpha}$ and $\Delta_{\beta}$, the superconducting gap energies. However, these values
are enlarged from the true bulk and proximity superconducting gap energies due to
dissipation induced from the bulk state of the mesa. In the plateau
regime, the transport is non-dissipative in the mesa, while the transport
is dissipative in the plateau-transition regime due to the QH bulk
state. Herein, in the plateau-transition regime, applied $V_{{\rm sd}}$
between two superconductors is divided into the voltage on the junctions
and on the mesa, then the deduced $\Delta_{\beta}$ gives a larger
gap energy as the contribution of the QH bulk state in the transport
becomes larger. The equivalent circuit is represented in Fig. 2(d) in the main text.
Consequently, $\Delta_{\beta}$ produces a peak in
the middle of the plateau-transition regime where the bulk contribution
becomes maximum. The true bulk superconducting gap energies (corresponding to the gap
for channel $\beta$), $0.35$ meV, can be evaluated from $\Delta_{\beta}$
near the plateau regime. From this gap energy, we calculated the true
superconducting proximity gap energy, $\sim0.1$ meV as $0.35\times\Delta_{\alpha}/\Delta_{\beta}$.
The calculated gap, $\Delta_{{\rm triplet}}$ as a function
of $g$, $dI/dV$ at $V_{{\rm sd}}=3.5$~mV in unit of $e^{2}/h$
is shown in Fig.\ref{figs1}.
\subsection{The maximum position of the bulk contribution}
Our results for $P$ and $Z_{\alpha}$ have a maximum and minimum,
respectively, at $\Delta g\sim0.7$. In this section, we estimate
how large $\rho_{xx}$ should be to make the bulk contribution have
a maximum at $\Delta g\sim0.7$.
As the shape of our device is square, the two-terminal conductance is
written by $G_{2t}=\sqrt{\sigma_{xx}+\sigma_{xy}}$. $\sigma_{xx}$
and $\sigma_{xy}$ are the longitudinal conductivity and Hall conductivity~\cite{lippmannzna1958, jensenjpc1972, abanin2008},
respectively. $\sigma_{xx}$ has a maximum when the bulk contribution
is maximum and the situation is given by $\sigma_{xy}=0.5e^{2}/h+ne^{2}/h$
($n=0,1,2,3...$), at which change in the filling factor $\Delta\nu$
is equal to 0.5. Herein, if the bulk contribution is maximum at $\Delta g\sim0.7$
(namely $G_{2t}=0.7e^{2}/h+ne^{2}/h$), $\sigma_{xx}(\sigma_{xy})$
should be 0.49$e^{2}/h$ (0.5$e^{2}/h$), 0.80$e^{2}/h$ (1.5$e^{2}/h$),
and 1.02$e^{2}/h$ (2.5$e^{2}/h$) in the $n=0,1,{\rm and}~2$ cases,
respectively. From these conductivity, we can calculate the longitudinal
resistance,$\rho_{xx}=\sigma_{xx}/(\sigma_{xx}^{2}+\sigma_{xy}^{2})$,
resulting in 1.0$h/e^{2}$ ($\sim26{\rm k\Omega}$), 0.28$h/e^{2}$($\sim7.0{\rm k\Omega}$),
and 0.14$h/e^{2}$($\sim3.6{\rm k\Omega}$) in the $n=0,1,{\rm and}~2$
cases, respectively. The calculated longitudinal resistance should
be obtained when the bulk contribution is maximum, meaning $\rho_{xx}$
is maximum with the calculated resistance in the region corresponding
to the plateau-transition regime. We measured a Hall bar device at
2 K, fabricated from the same InAs quantum well wafer as we used for the superconducting devices and the results are shown in Fig.~\ref{figsR}. $\rho_{xx}$
has some peaks consistent with the finite bulk contribution. The maximum
of $\rho_{xx}$ at 4 T is at 4 k${\rm \Omega}$ and 1.5 k${\rm \Omega}$
corresponding to the $n=1{\rm and}~2$ cases, respectively. The measured
resistances are comparable to the estimated resistance based on the
assumption that the bulk contribution is maximum at $\Delta g\sim0.7$
($G_{2t}=0.7e^{2}/h+ne^{2}/h$).
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
Large-scale magnetic fields are an essential part of magnetohydrodynamic (MHD) turbulence. Even when a large-scale magnetic field is not imposed on the system externally (as may be done in laboratory devices, for example), it is generated by turbulence due to magnetic dynamo action. Such a magnetic field plays a crucial role in magnetic turbulence at small scales. Indeed, unlike a uniform large-scale velocity field, the large-scale magnetic field cannot be removed by a Galilean transformation, and it mediates the energy cascade at all scales. Turbulent plasmas with high Reynolds numbers are therefore anisotropic at small scales with the eddies stretched along a local strong background field. This allows various reduced models, such as reduced MHD models, gyrokinetics models, or models with reduced spatial dimensionality \citep[e.g.,][]{dmitruk_etal_2005, perez2008, tobias2011, mason2012, schekochihin_etal_2009, camporeale2011,wu2013,karimabadi2013,franci2015}, to accurately describe the local dynamics. Recently, there has been a widespread application of these reduced models to understand the dissipation of the turbulent cascade in the solar wind \citep{camporeale2011,wu2013,karimabadi2013,franci2015, howes_etal_2008, howes_etal_2008b, boldyrev2011, howes_etal_2011, tenbarge_howes_2013, tenbarge_etal_2013, told_etal_2015} and in the solar corona \citep{einaudi1999, oughton_etal_2001, dmitruk_etal_2002, rappazzo_etal_2007, rappazzo_etal_2008, wan_etal_2014}.
On the other hand, a large fraction of energy is now known to be dissipated in a small fraction of the volume that is often characterized by strong variations in the large-scale magnetic field \citep[e.g.,][ and references therein]{zhdankin_etal_2016b}. In fact, one may think of MHD turbulence without an imposed large-scale field as consisting of subdomains where the large-scale magnetic field is strong, separated by thin boundaries where the direction of large-scale field changes abruptly. This structure is consistent, for example, with the Borovsky picture of solar wind turbulence as an ensemble of tightly packed flux tubes \cite[][]{borovsky2008}, and with statistical significance of strong rotational magnetic discontinuities observed in other studies~\cite[e.g.,][]{bruno_etal_2001, li2008, zhdankin_etal_2012, greco_etal_2009}. The regions of strong magnetic field variations are very intermittent, i.e., they occupy a small volume and contain only a small fraction of the total energy. However, they may contain a significant fraction of the magnetic energy dissipation. If so, the reduced model may not be used to properly describe turbulent energy dissipation. The question of the extent to which energy dissipation is skewed toward the regions of strong variation of the magnetic field is therefore of principal importance for phenomenological and numerical modeling of MHD turbulence. This work presents the quantitative statistical analysis of this issue.
In this work, we utilize numerical simulations of MHD turbulence to investigate the local relationship between the energy dissipation rate and the relative variation of the magnetic field. We find that a significant fraction of the energy dissipation occurs in regions where this variation is large, although this fraction slowly decreases with increasing size of the inertial range. We therefore argue that caution is required when applying reduced models to systems where the inertial interval is not sufficiently long. For instance, we estimate that in systems where the inertial interval spans less than three orders of magnitude, more than 15\% of the energy dissipation occurs in the regions with strong variations of the magnetic field, where the reduced models are not applicable. These regions occupy a very small fraction of the volume, however, the dissipation inside them is very intense. Models that fail to properly account for such regions may lead to an incomplete description of energy dissipation and particle heating in MHD plasma turbulence.
\section{Methods}
For our analysis we numerically solve the incompressible MHD equations for the plasma velocity $\boldsymbol{v}(\boldsymbol{x},t)$ and the magnetic field $\boldsymbol{B}(\boldsymbol{x},t)=\boldsymbol{B}_0+\boldsymbol{b}(\boldsymbol{x},t)$ (where $\boldsymbol{B}_0 = B_0 \hat{\boldsymbol{z}}$ is the uniform background field):
\begin{eqnarray}
& \partial_t \boldsymbol{v} + (\boldsymbol{v} \cdot \nabla) \boldsymbol{v} = - \nabla p + (\nabla \times \boldsymbol{B}) \times \boldsymbol{B} + \nu \nabla^2 \boldsymbol{v} + \boldsymbol{f}_1, \nonumber \\
& \partial_t \boldsymbol{B} = \nabla \times (\boldsymbol{v} \times \boldsymbol{B}) + \eta \nabla^2 \boldsymbol{B} +\boldsymbol{f}_2, \label{eq:mhd}
\end{eqnarray}
where $p$ is the plasma pressure, along with $\nabla \cdot \boldsymbol{v} = \nabla \cdot \boldsymbol{B} = 0$. For simplicity, we take the resistivity to be equal to the viscosity, $\eta = \nu$, in the simulations. The turbulence is driven by random large-scale external forces $\boldsymbol{f_1}$ and $\boldsymbol{f_2}$ that are applied in Fourier space and have amplitudes chosen so that $v_{rms}\approx 1$. The forces have no component in the direction of the background field and are solenoidal in the $xy$-plane. The Fourier coefficients are non-zero only for wavenumbers $k_{x,y,z}=\pm 1$ or $ \pm 2$, and in such cases the coefficients are chosen from a Gaussian distribution and are refreshed on average every $0.1L/(2\pi v_{rms} )$ time units, where $L$ is the size of the domain (that is, the force is updated approximately 10 times per large-scale turnover time). While the particular choice of the force's statistical properties and correlation time does not affect the spectra of turbulence at smaller scales \cite[][]{mason2006,mason_cb08}, this setup allows us to supply energy in large-scale Alfv\'enic fluctuations in a controlled fashion. The equations are solved on a triply periodic domain using standard pseudospectral methods. Time advancement of the diffusive terms is carried out exactly using the integrating factor method, while the remaining terms are treated using a third-order Runge-Kutta scheme. Further details of the numerical approach can be found in \cite{cattaneo_etal_2003}. A background magnetic field with relatively small amplitude $B_0 = 0.5 b_{\rm rms}$ is imposed. For the present analysis, we focus on a simulation with a $1024^3$ lattice and a Reynolds number $\operatorname{Re}= v_{\rm rms} (L/2\pi) / \nu \sim 5500$; simulations with smaller $\operatorname{Re}$ give similar results. We carry out our analysis on 5 snapshots, each separated by 2 eddy turnover times.
We aim to understand how the relative amplitude of the magnetic fluctuations is correlated with the local energy dissipation rate. For that we have designed the following statistical approach. We subdivide the simulation domain into cubes of size $\Delta x$ and consider the statistical properties of fluctuations in the cubes. Consider a cube of size $\Delta x$, which is centered at the point $\boldsymbol{x}$. The local mean magnetic field in this cube is then given by
\begin{align}
\bar{\boldsymbol{B}}_{\Delta x}(\boldsymbol{x}) = \frac{1}{(\Delta x)^3} \int\limits_{\Delta x} d^3x' \boldsymbol{B}(\boldsymbol{x} + \boldsymbol{x}') \, ,
\end{align}
while the root-mean-square (rms) magnetic field fluctuation is given by
\begin{align}
B_{\text{rms},\Delta x}(\boldsymbol{x}) = \left[ \frac{1}{(\Delta x)^3} \int\limits_{\Delta x} d^3x' \left|\boldsymbol{B}(\boldsymbol{x} + \boldsymbol{x}') - \bar{\boldsymbol{B}}_{\Delta x}(\boldsymbol{x})\right|^2 \right]^{1/2} \,
\end{align}
where $\int_{\Delta x}$ denotes an integral across the volume of the cube. The local strength of the fluctuations is then characterized by the ratio $R_{\Delta x} = B_{\text{rms},\Delta x}/\bar{B}_{\Delta x}$. The presence of a strong local magnetic field is implied by $R_{\Delta x} \ll 1$. The local energy dissipation rate in a cube centered at point $\boldsymbol{x}$ is given by
\begin{align}
{\mathcal E}_{\Delta x}(\boldsymbol{x}) &= \int\limits_{\Delta x} d^3x' \left[ \eta \left|\boldsymbol{j}(\boldsymbol{x} + \boldsymbol{x}')\right|^2 + 2 \nu {\bf \sigma}(\boldsymbol{x} + \boldsymbol{x}') : {\bf \sigma}(\boldsymbol{x} + \boldsymbol{x}') \right] \, ,
\end{align}
where $\boldsymbol{j} = \nabla \times \boldsymbol{B}$ is the current density and ${\bf \sigma} = [\nabla \boldsymbol{v} + (\nabla \boldsymbol{v})^T]/2$ is the rate-of-strain tensor. We note that ${\mathcal E}_{\Delta x}$ includes contributions from both resistive and viscous dissipation, but our results are broadly similar when either dissipation mechanism is considered individually. We measure the above quantities for cubes of varying size $\Delta x$ in order to understand the scale dependence of the field fluctuations and energy dissipation. Our major object of study is the correlation between the local intensity of fluctuations $R_{\Delta x}$ and the local energy dissipation rate ${\mathcal E}_{\Delta x}$.
For reference, in Fig.~\ref{fig:profile} we show contours of $R_{\Delta x} = 1/3$ overlaid on an image of the local energy dissipation rate ${\mathcal E}_{\Delta x}$ in an arbitrarily chosen 2D plane of the simulation, for $\Delta x/L = 1/256$. There is evidently a strong degree of correlation between the two quantities, with both often being concentrated in thin, sheet-like coherent structures.
\begin{figure}
\includegraphics[width=\columnwidth]{fig1.eps}
\centering
\caption{\label{fig:profile} Contour plot of fluctuation-to-mean ratio $R_{\Delta x}$ (red) overlaid on an image of the local energy dissipation rate ${\mathcal E}_{\Delta x}$ (blue) for $\Delta x/L = 1/256$, in an $xy$ slice of the simulation. The contours are taken at $R_{\Delta x} = 1/3$, and the colorbar indicates $30 \times {\mathcal E}_{\Delta x} / \langle {\mathcal E}_{\Delta x} \rangle$. A strong correlation between the magnetic field variations and the energy dissipation is observed.}
\end{figure}
\section{Results}
\subsection{The mean field and the fluctuations}
We begin by analysing the statistical properties of the {\em mean} local quantities $\langle \bar{B}_{\Delta x} \rangle$, $\langle B_{\rm{rms}, \Delta x} \rangle$ and $\langle R_{\Delta x} \rangle$, where the angular brackets denote averaging over all the cubes of size $\Delta x$ in the simulation domain. Fig.~\ref{fig:unconditioned} shows the scaling of these quantities versus $\Delta x$. We see that the local mean field measured at progressively smaller scales, $\langle \bar{B}_{\Delta x} \rangle\vert_{\Delta x\to 0}$, approaches the value of the large-scale fluctuations, $\langle B_{{\rm rms},\Delta x} \rangle\vert_{\Delta x\to L}$, confirming that the large-scale magnetic field fluctuations act as a local background field for the small-scale fluctuations. This important fact, quantified in Fig.~\ref{fig:unconditioned}, is behind the applicability of models of MHD turbulence that assume a strong imposed uniform magnetic field.
In addition, we see that, to a very good approximation, $\langle R_{\Delta x} \rangle \propto (\Delta x)^{1/2}$ in the inertial range, while the fluctuations $\langle B_{{\rm rms},\Delta x} \rangle$ and mean $\langle \bar{B}_{\Delta x} \rangle$ are not as well fit by power laws. One may, however, roughly approximate $\langle B_{{\rm rms},\Delta x} \rangle \propto (\Delta x)^{1/3}$, which is broadly consistent with the scaling ($1/3$) of magnetic field increments in the Goldreich-Sridhar phenomenology \cite[][]{goldreich1995}, and not far from the scaling ($1/4$) predicted in the model of scale-dependent dynamic alignment \cite[e.g.,][]{boldyrev2006,perez_mason2012}. We however note that these phenomenologies assume the presence of a {\em strong} and {\em constant} large-scale magnetic field in the whole domain, whereas in our measurement we instead average over the cubes with all possible values of the mean field.
Finally, we note that the observed scaling of the fluctuations-to-mean ratio, $\langle R_{\Delta x} \rangle \propto (\Delta x)^{1/2}$, implies that anisotropy grows significantly with decreasing scale. For example, $\langle R_{\Delta x} \rangle$ decreases from $1$ to roughly~$1/10$ after $\Delta x$ decreases by only two decades. This implies that the reduced models of MHD turbulence should formally be valid for a description of energy distribution in the bulk of the small-scale fluctuations in most space and astrophysical systems.
\begin{figure}
\includegraphics[width=\columnwidth]{fig2.eps}
\centering
\caption{\label{fig:unconditioned} The mean ratio $\langle R_{\Delta x} \rangle = \langle B_{{\rm rms},\Delta x}/ \bar{B}_{\Delta x} \rangle$ (blue), mean fluctuations $\langle B_{{\rm rms},\Delta x} \rangle$ (red), and mean field $\langle \bar{B}_{\Delta x} \rangle$ (green) versus scale $\Delta x$. The power-law scalings $\Delta{x}^{1/2}$ (black, solid) and $\Delta{x}^{1/3}$ (black, dashed) are shown for reference.}
\end{figure}
\subsection{The energy dissipation}
The picture changes significantly when we consider the {\em energy dissipation}, which is known to be very intermittent, that is, not space filling \cite[e.g.,][]{osman2012,zhdankin_etal_2016b}. We now analyze the correlation of the local energy dissipation rate ${\mathcal E}_{\Delta x}$ with $R_{\Delta x}$.
In Fig.~\ref{fig:correlation}, we show 2D joint probability density functions of ${\mathcal E}_{\Delta x}$ and each of $\bar{B}_{\Delta x}$, $B_{{\rm rms},\Delta x}$, and $R_{\Delta x}$ separately. We find a strong correlation between the dissipation and the fluctuations, such that the most intense dissipation indeed takes place in regions of large relative fluctuations in the magnetic field. In particular, we find that the results can be fit rather well by a quadratic scaling, ${\mathcal E}_{\Delta x} \propto R_{\Delta x}^2$. The scaling of dissipation with absolute fluctuations can be approximated by ${\mathcal E}_{\Delta x} \propto B_{{\rm rms},\Delta x}^3$ for inertial-range fluctuations and ${\mathcal E}_{\Delta x} \propto B_{{\rm rms},\Delta x}^2$ for weaker fluctuations. On the other hand, there is very little correlation between the dissipation and local mean field $\bar{B}_{\Delta x}$, consistent with the mean field being set by the background, large-scale eddies.
While the scaling of ${\mathcal E}_{\Delta x}$ with $R_{\Delta x}$ is non-trivial to explain, we note that a cubic scaling of the dissipation with respect to $B_{{\rm rms},\Delta x}$ can be expected on general grounds. This is because the local dissipation should be comparable to the local magnetic energy divided by the cascade time, which can be estimated by the local eddy turnover time, $\tau_{\Delta x} \sim \Delta x / B_{{\rm rms},\Delta x}$ (in simulation units). We then arrive at
\begin{align}
{\mathcal E}_{\Delta x} \sim (\Delta{x})^3 \frac{B_{{\rm rms},\Delta x}^2}{\tau_{\Delta x}} \sim (\Delta{x})^2 B_{{\rm rms},\Delta x}^3 \, . \label{eq:corr}
\end{align}
\begin{figure}
\includegraphics[width=\columnwidth]{fig3a.eps}
\includegraphics[width=\columnwidth]{fig3b_update.eps}
\includegraphics[width=\columnwidth]{fig3c.eps}
\centering
\caption{\label{fig:correlation} The 2D joint probability density function of the local energy dissipation rate ${\mathcal E}_{\Delta x}$ and mean magnetic field $\bar{B}_{\Delta x}$ (top panel), rms fluctuations $B_{\text{rms},\Delta x}$ (center panel), and rms-to-mean ratio $R_{\Delta x} = B_{\text{rms},\Delta x}/\bar{B}_{\Delta x}$ (bottom panel), for $\Delta x/L = 1/128$ (the plots for other $\Delta x$ are similar). A quadratic scaling (solid black line) and cubic scaling (dashed black line) are shown for reference.}
\end{figure}
To assess more quantitatively the extent to which dissipation occurs in regions with the large fluctuations-to-mean ratio, $R_{\Delta x} \gtrsim 1$, we consider the cumulative distribution of energy dissipation conditioned on $R_{\Delta x}$, which we denote by ${\mathcal E}_{{\rm cum}, R_{\Delta x}} (R_{\rm thr})$. In particular, we set a threshold $R_{\rm thr}$ and measure the fraction of the total energy dissipation that occurs in cubes with $R_{\Delta x} > R_{\rm thr}$. The results are shown in Fig.~\ref{fig:dist_cum_1}, along with the volume occupied by cubes exceeding the threshold. The cumulative distributions extend to large values of $R_{\Delta x}$, implying that, indeed, a significant fraction of energy dissipation may occur in regions with $R_{\Delta x} \gtrsim 1$. However, the tail of the distribution function shifts downwards for decreasing $\Delta x$. Hence, at sufficiently small scales, the majority of energy dissipation should occur in regions where $R_{\Delta x}$ is small. This means that asymptotically in the limit of very large Reynolds number, both the small-scale fluctuations and the energy dissipation are adequately captured by the reduced models.
As we now demonstrate, however, this convergence is rather slow. Figure~\ref{fig:dist_cum} shows the dependence of the cumulative energy dissipation ${\mathcal E}_{{\rm cum}, R_{\Delta x}}(R_{\rm thr})$ on the scale $\Delta x$, for several values of $R_{\rm thr}$. From the plot corresponding to $R_{\rm thr}=1$ we estimate a scaling ${\mathcal E}_{{\rm cum}, R_{\Delta x}} (1) \propto (\Delta x)^{0.8}$. The curves with $R_{\rm thr} < 1$ are similar; they seem to have the same scaling but shifted upward with respect to the curve with $R_{\rm thr} = 1$. To understand the implications of this slow convergence we analyze the following example. Consider the curve corresponding to the threshold $R_{\rm thr} = 1/4$. Such a threshold approximately separates the cubes where the reduced MHD model provides a good description for the spectrum of MHD turbulence ($R_{\Delta x}<1/4$) from the cubes where it does not ($R_{\Delta x}>1/4$) \cite[e.g.,][]{mason2006,mason_cb08,mason2012}. Assuming that we may extrapolate the observed scaling to very small $\Delta x$, we can estimate for this curve: ${\mathcal E}_{{\rm cum}, R_{\Delta x}} (1/4) \sim 40 \times (\Delta x/L)^{0.8}$. The fraction of the dissipation occurring inside the cubes with $R_{\Delta x}<1/4$ will thus exceed $85\%$ if $\Delta x/L < 10^{-3}$, and $97\%$ if $\Delta x/L < 10^{-4}$. This means, for example, that the reduced MHD model will correctly capture more than $97\%$ of the energy dissipation only if the inertial interval of the turbulence extends to scales smaller than $\Delta x/L < 10^{-4}$. In this case, less than $3\%$ of the magnetic energy dissipation will happen inside the cubes where magnetic fluctuation-to-mean ratio exceeds $1/4$, that is, where the reduced models are not applicable.
\begin{figure}
\includegraphics[width=\columnwidth]{fig4.eps}
\caption{\label{fig:dist_cum_1} The fraction of total energy dissipation occurring in cubes of size $\Delta x$ with the ratio of local fluctuations-to-mean exceeding a threshold, $R_{\Delta x} > R_{\rm thr}$, for $\Delta x/L \in \{ 1/256, 1/128, 1/64, 1/32, 1/16, 1/8, 1/4 \}$ (blue to red, solid lines). The corresponding fraction of volume occupied by the cubes is also shown (dotted lines).}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{fig5.eps}
\caption{\label{fig:dist_cum} The fraction of total energy dissipated (solid colored lines) and volume occupied (dashed colored lines) in cubes of size $\Delta x$ with the ratio of local fluctuations-to-mean exceeding a threshold, $R_{\Delta x} > R_{\rm thr}$, for $R_{\rm thr} = 2$ (magenta), $1$ (blue), $1/2$ (red), and $1/4$ (green). Power-law scalings $\Delta x^{0.8}$ (solid black line) and $\Delta x^{1.2}$ (dashed black line) are shown for reference (black).}
\end{figure}
The regions with $R_{\rm thr} > 1/4$, corresponding to $3\%$ of all the energy dissipation in the considered example, are, however, extremely intense. From Fig.~\ref{fig:dist_cum} we estimate the volume occupied by the structures with $R_{\rm thr} > 1$ as $V_{R_{\Delta x}}(1)\propto (\Delta x)^{1.2}$. Assuming that the same scaling holds for smaller values of $R_{\rm thr}$, we can estimate from Fig.~\ref{fig:dist_cum} that $V_{R_{\Delta x}}(1/4)\sim 150\times (\Delta x/L)^{1.2}$. The cubes of the size $\Delta x/L=10^{-4}$, which correspond to $R_{\rm thr} > 1/4$, considered in the previous example, will therefore occupy only about 0.2\% of the total volume. They include significant variations in the magnetic field direction. In cases where the energy dissipation or particle acceleration effects are strongly skewed toward the environments with strong variations of the magnetic field direction \cite[e.g.,][]{chen2015,chasapis2015,tessein2016}, these effects will not be adequately captured by the reduced models.
\section{Conclusions}
Magnetic plasma turbulence is intrinsically anisotropic, meaning that small scale fluctuations experience a large-scale magnetic field that mediates nonlinear interactions. This happens even if the strong magnetic field is not imposed externally, since such a field is self-consistently generated by turbulence itself. It is therefore customary in studies of MHD turbulence to assume the presence of a strong uniform background magnetic field. This assumption, introduced mostly phenomenologically in previous studies, is put on a firmer, quantitative ground in our work. In particular, we argued that the so-called reduced models, that is, models assuming a strong background field and correspondingly anisotropic fluctuations (reduced MHD, gyrokinetics, models with reduced dimensionality, etc.), should describe the turbulent energy distribution correctly, when their inertial interval is sufficiently long. We however also established that such models may not work as well for describing the energy dissipation in systems where the inertial interval for magnetic fluctuations is not long enough. The reason is that the dissipation is extremely spatially intermittent. It is skewed toward the regions where the magnetic field fluctuations are relatively large compared to the mean field, as happens at the boundaries between nearly uniformly magnetized domains. For example, we estimated that more than 3\% of the energy dissipation is not captured by the reduced models if the MHD inertial interval extends over less than four orders of magnitude. Such constraints may be relevant for some natural systems (e.g., solar wind turbulence \cite[][]{kiyani_etal_2015}), and they may also be essential for laboratory experiments, say liquid metal experiments, where the magnetic Reynolds numbers are not large enough \cite[e.g.,][]{lathrop2005}. Moreover, the regions of strong energy dissipation occupy very small volumes, and, therefore, they may be extremely intense. No matter how large the Reynolds number is, the reduced models always miss a certain fraction of intense dissipation events generated by turbulence, which may become important in phenomena involving higher-order moments of field variations, say, transport phenomena.
\section*{Acknowledgements}
The authors thank the referee, Alexander Schekochihin, for helpful comments. VZ acknowledges support from NSF grant AST-1411879. SB is partly supported by the National Science Foundation under the grant NSF AGS-1261659 and by the Vilas Associates Award from the University of Wisconsin - Madison. JM acknowledges the support of the EPSRC, through grant EP/M004546/1. We acknowledge PRACE for awarding us access to resource FERMI based in Italy at CINECA, and the STFC DiRAC HPC Facility for access to the COSMA Data Centric system at Durham University and MINERVA at the University of Warwick.
\bibliographystyle{mnras}
|
\subsection{Proof of \theoremref{apxapxefnash}}\label{sec:approxefnash}
We will first prove that if there exists an $\epsilon_{ef}$-envy-proof Nash equilibrium, then there exists a $k$-uniform $(2\epsilon+\epsilon_{ef})$-envy-proof $\epsilon$-Nash equilibrium. Then by an exhaustive search, we can check all $\binom{n+k-1}{k}^2$ possible $k$-uniform $\epsilon$-Nash equilibrium in time $O(n^{3k})$. Suppose after the exhaustive search, we cannot find any $k$-uniform $\epsilon$-Nash equilibrium that is $(2\epsilon+\epsilon_{ef})$-envy-proof, then by contradiction, there does not exist any $\epsilon_{ef}$-envy-proof Nash equilibrium in the game.
For simplicity we assume the utilities have values between 0 and 1, and both players have $n$ pure strategies. We note that the result can be generalized without these restrictions.
\begin{lemma}
If a game $G=(\{u_0,u_1\},\{a_0^1,...,a_0^n\},\{a_1^1,...,a_1^n\})$ admits an $\epsilon_{ef}$-envy-proof Nash equilibrium $\mathbf{x^*}\equiv(\mathbf{x_0^*},\mathbf{x_1^*})$, then for every $\epsilon<1$, for every $k\geq \frac{3\ln n}{\epsilon^2}$, there exists a $k$-uniform $(2\epsilon+\epsilon_{ef})$-envy-proof $\epsilon$-Nash equilibrium.
\end{lemma}
\begin{proof}
First we construct a profile $\hat{x}\equiv(\mathbf{\hat{x}_0},\mathbf{\hat{x}_1})$ by sampling from $(\mathbf{x_0^*},\mathbf{x_1^*})$: For $b=\in\{0,1\}$, form multiset $S_b$ by sampling $k$ pure strategies with replacement according to $\mathbf{x_b^*}$ independently. Let $\mathbf{\hat{x}_b}$ be a mixed strategy that assigns $1/k$ probability to each member in $S_b$.
Second let $GOAL$ be the event when $\mathbf{\hat{x}}$ is a $(2\epsilon+\epsilon_{ef})$-envy-proof $\epsilon$-Nash equilibrium, and define events\footnote{Recall that when $\mathbf{x_{\bar{b}}^*}$ is a mixed strategy, our notation $u_b(a_b^i:\mathbf{x_{\bar{b}}^*})$ denotes the the expected utility over the randomness of $\mathbf{x_{\bar{b}}^*}$, and $(a_b^i:\mathbf{x_{\bar{b}}^*})$ denotes an unordered strategy profile where $\P_b$ plays $a_b^i$, and $\P_{\bar{b}}$ plays $\mathbf{x_{\bar{b}}^*}$.}
\begin{align*}
\pi_{b,i} & = \{u_b(a_b^i:\mathbf{x_{\bar{b}}^*})\leq u_b(\mathbf{x^*})+\epsilon\}, \forall b\in\{0,1\},i\in [n], and \\
\theta_{b,i} & = \{u_b(a_b^i:\mathbf{x_{\bar{b}}^*})- u_b(\mathbf{x^*}) \leq u_{\bar{b}}(a_b^i:\mathbf{x_{\bar{b}}^*})- u_{\bar{b}}(\mathbf{x^*})+2\epsilon+\epsilon_{ef}\}, \forall b\in\{0,1\},i\in [n] \\
& \text{so, } GOAL = \bigcap_{b,i} (\pi_{b,i}\cap \theta_{b,i}).
\end{align*}
Third, in order to apply probability arguments, we need to somehow decompose $\pi_{b,i}$ and $\theta_{b,i}$. Define the following events:
\begin{align*}
\phi_b &=\{|u_b(\mathbf{\hat{x}})-u_b(\mathbf{x^*})|\leq \epsilon/2 \}, \forall b \\
\psi_{b,i} &=\{|u_b(a_b^i:\mathbf{\hat{x}_{\bar{b}}})-u_b(a_b^i:\mathbf{x^*_{\bar{b}}})|\leq \epsilon/2 \}, \forall b
\end{align*}
It can be verified that $\phi_b\cap \psi_{b,i} \subseteq \pi_{b,i}, \forall b,i$, and $\bigcap_{b'} (\phi_{b'}\cap \psi_{b',i}) \subseteq \theta_{b,i}, \forall b,i$
To analyze the probability of event $\phi_b$, we further define
\begin{align*}
\phi_{b,1} &=\{|u_b(\mathbf{\hat{x}_b}:\mathbf{x^*_{\bar{b}}})-u_b(\mathbf{x^*})|\leq \epsilon/4 \}, \forall b \\
\phi_{b,2} &=\{|u_b(\mathbf{\hat{x}})-u_b(\mathbf{\hat{x}_b}:\mathbf{x^*_{\bar{b}}})|\leq \epsilon/4 \}, \forall b
\end{align*}
Clearly, $\phi_{b,1}\cap \phi_{b,2}\subseteq \phi_{b}$. Put them together: $\bigcap_{b}\phi_{b,1}\bigcap_b\phi_{b,2}\bigcap_{b,i}\psi_{b,i}\subseteq GOAL$
Finally we bound the probability of the events.
Due to the construction of $\mathbf{\hat{x}_b}$, $u_b(\mathbf{\hat{x}_b}:\mathbf{x^*_{\bar{b}}})$ (resp. $u_b(\mathbf{\hat{x})}$) is like the average of $k$ independent random variables of the same expected values $u_b(\mathbf{x^*})$ (resp. $u_b(\mathbf{\hat{x}_b}:\mathbf{x^*_{\bar{b}}})$). Hence by Hoeffding bound \cite{Hoeffding63},
\begin{equation*}
\Pr[\phi_{b,j}^c]\leq 2e^{-k\epsilon^2/8}, \forall b\in\{0,1\},j\in\{1,2\}
\end{equation*}
Similarly,
\begin{equation*}
\Pr[\psi_{b,i}^c]\leq 2e^{-k\epsilon^2/2}, \forall b\in\{0,1\},i\in[n]
\end{equation*}
Then by union bound, we have
\begin{equation*}
\Pr[GOAL^c]\leq \sum_{b,j} \Pr[\phi_{b,j}^c]+\sum_{b,i} \Pr[\psi_{b,i}^c]\leq 8e^{-k\epsilon^2/8}+4ne^{-k\epsilon^2/2}<1
\end{equation*}
Since $GOAL$ happens with a nonzero probability, there exists a $k$-uniform $(2\epsilon+\epsilon_{ef})$-envy-proof $\epsilon$-Nash equilibrium.
\end{proof}
\subsection{Proof of \theoremref{envyfreeNashHard}}\label{sec:envyfreeNashHard}
Let $\phi$ be a given Boolean formula in conjunctive normal form. Let $V=\{x^1,...,x^n\}$ be the set of variables, $L=\{+x^1,-x^1,...,+x^n,-x^n\}$ the set of corresponding literals, and $C$ the set of clauses in $\phi$. Define a function ${\ensuremath{{\text{LitSet}}}}\xspace: V\cup C \rightarrow 2^{L}$ that outputs the set of literals corresponding to a variable or a clause. That is, ${\ensuremath{{\text{LitSet}}}}\xspace(x)=\{+x,-x\}$ if $x\in V$ and ${\ensuremath{{\text{LitSet}}}}\xspace(x)=\{\ell| x \text{ contains } \ell\}$ if $x\in C$. We construct a symmetric 2-player game $G(\phi)$ in normal form as follows. Let $f$ be a default action, $\Sigma=V\cup L\cup C\cup \{f\}$, $\beta=-n$ (or any sufficiently small value), and the utility function be
\begin{itemize}
\item $u_0(\ell_0,\ell_1)=u_1(\ell_0,\ell_1)= \left\{\begin{matrix}
n-1, & \ell_0\neq -\ell_1 \\
\beta, & \ell_0= -\ell_1
\end{matrix}\right. , \forall \ell_0,\ell_1\in L$
\item $u_0(x_0,x_1)=u_1(x_0,x_1)=\beta, \forall x_0,x_1\in V\cup C$.
\item $u_0(x,\ell)=u_1(\ell,x)= \left\{\begin{matrix}
n, & \ell\not\in {\ensuremath{{\text{LitSet}}}}\xspace(x) \\
0, & \ell \in {\ensuremath{{\text{LitSet}}}}\xspace(x)
\end{matrix}\right. ,\forall x\in V\cup C$
\item $u_0(\ell,x)=u_1(x,\ell)= \left\{\begin{matrix}
n-2, & \ell\not\in {\ensuremath{{\text{LitSet}}}}\xspace(x) \\
2(n-1), & \ell \in {\ensuremath{{\text{LitSet}}}}\xspace(x)
\end{matrix}\right. ,\forall x\in V\cup C$
\item $u_0(f,x)=u_1(x,f)=
\left\{\begin{matrix}
n+2, & x=f\\
n, & x\in V\cup C\\
n-1, & x\in L
\end{matrix}\right.$
\item $u_0(x,f)=u_1(f,x)=
\left\{\begin{matrix}
n+1, & x\in V\cup C\\
n-1, & x\in L\\
\end{matrix}\right.$
\end{itemize}
Take $\phi=(x^1\vee -x^2)\wedge (-x^1\vee x^2)$ for example. The game $G(\phi)$ is showed in \tableref{sat}. Note that $\phi$ has exactly two solutions, either assigning both variables to true or both to false
, and the game has exactly 3 equilibria: 1) both player play $\{+x^1,+x^2\}$ uniformly randomly ; 2) both player play $\{-x^1,-x^2\}$ uniformly randomly; 3) both players play $f$. The first two are envy-proof while the last one is not.
\begin{table}[]
\centering
\caption{The table shows the game $G(\phi)$, where $\phi=(x^1\vee -x^2)\wedge (-x^1\vee x^2)$. By $\times$ we denote $(\beta,\beta)$ since $\beta$ is just a sufficiently small value for both the players to avoid running into the state.}
\label{tab:sat}
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|}
\hline
& $x^1$ & $x^2$ & $+x^1$ & $-x^1$ & $+x^2$ & $-x^2$ & $(x^1\vee -x^2)$ & $(-x^1\vee x^2)$ & $f$ \\ \hline
$x^1$ & $\times$ & $\times$ & 0,2 & 0,2 & 2,0 & 2,0 & $\times$ & $\times$ & 3,2 \\ \hline
$x^2$ & $\times$ & $\times$ & 2,0 & 2,0 & 0,2 & 0,2 & $\times$ & $\times$ & 3,2 \\ \hline
$+x^1$ & 2,0 & 0,2 & 1,1 & $\times$ & 1,1 & 1,1 & 2,0 & 0,2 & 1,1 \\ \hline
$-x^1$ & 2,0 & 0,2 & $\times$ & 1,1 & 1,1 & 1,1 & 0,2 & 2,0 & 1,1 \\ \hline
$+x^2$ & 0,2 & 2,0 & 1,1 & 1,1 & 1,1 & $\times$ & 0,2 & 2,0 & 1,1 \\ \hline
$-x^2$ & 0,2 & 2,0 & 1,1 & 1,1 & $\times$ & 1,1 & 2,0 & 0,2 & 1,1 \\ \hline
$(x^1\vee -x^2)$ & $\times$ & $\times$ & 0,2 & 2,0 & 2,0 & 0,2 & $\times$ & $\times$ & 3,2 \\ \hline
$(-x^1\vee x^2)$ & $\times$ & $\times$ & 2,0 & 0,2 & 0,2 & 2,0 & $\times$ & $\times$ & 3,2 \\ \hline
$f$ & 2,3 & 2,3 & 1,1 & 1,1 & 1,1 & 1,1 & 2,3 & 2,3 & 4,4 \\ \hline
\end{tabular}
\end{table}
We will proof the following claim, which implies \theoremref{envyfreeNashHard}.
\begin{claim}
If $(\ell^1,...,\ell^n)$ satisfies $\phi$, then there is an envy-proof Nash equilibrium of $G(\phi)$, where both players play $\{\ell^1,...,\ell^n\}$ uniformly randomly. The only other Nash equilibrium is for both players to play $f$ and is not envy-proof.
\end{claim}
\begin{proof}
Let $\mathbf{\mu}$ be the uniform distribution over $\{\ell^1,...,\ell^n\}$.
First we show that the profile $(\mathbf{\mu},\mathbf{\mu})$ is indeed an envy-proof Nash equilibrium. Because $G(\phi)$ is a symmetric game, w.l.o.g., we fix player $\P_0$'s strategy to be $\mathbf{\mu}$ and consider changes in player $\P_1$'s strategy. Since $u_1(\ell^i,\ell^j)=n-1$ for all $i,j\in[n]$, playing any distribution $\mathbf{\nu}$ over $\{\ell^1,...,\ell^n\}$ gives the same expected utility $u_1(\mathbf{\mu},\mathbf{\nu})=n-1$\footnote{Recall that when $\mathbf{\mu},\mathbf{\nu}$ is a random variable over the action space, the notation $u_1(\mathbf{\mu},\mathbf{\nu})$ denotes the expected utility $E_{\mathbf{\mu},\mathbf{\nu}}[u_1(\mathbf{\mu},\mathbf{\nu})]$}. Playing the negation of some $\ell^i$ gives $u_1(\mathbf{\mu},-\ell^i)=\frac{1}{n}\cdot \beta+\frac{n-1}{n}\cdot n<n-1$, playing a variable or a clause $x$ in $V\cup C$ gives
$u_1(\mathbf{\mu},x)=\frac{1}{n}\cdot 0+\frac{(n-1)}{n}\cdot n=n-1$, and playing $f$ gives $u_1(\mathbf{\mu},f)=n-1$. Hence $u_1(\mathbf{\mu},\mathbf{a})\leq u_1(\mathbf{\mu},\mathbf{\mu})$ for any $\mathbf{a}$; $(\mathbf{\mu},\mathbf{\mu})$ is a Nash equilibrium. On the other hand, note $u_0(\mathbf{\mu},\mathbf{\nu})=u_1(\mathbf{\mu},\mathbf{\nu})$, $u_0(\mathbf{\mu},-\ell^i)=u_1(\mathbf{\mu},-\ell^i)$, $u_0(\mathbf{\mu},x)=\frac{1}{n}\cdot 2(n-1)+\frac{(n-1)}{n}\cdot (n-2)=n-1=u_1(\mathbf{\mu},x)$, and $u_0(\mathbf{\mu},f)=n-1=u_1(\mathbf{\mu},f)$, which together imply $\triangle_1 u_1(\mathbf{\mu},\mathbf{\mu})=\triangle_1 u_0(\mathbf{\mu},\mathbf{\mu})$. Hence $(\mathbf{\mu},\mathbf{\mu})$ is envy-proof.
As for the other Nash equilibrium $(f,f)$, since $u_b(f,f)$ is strictly larger than any other utility value, clearly it is a Nash equilibrium. However it is not envy-proof because $-1=u_1(f,x)-u_1(f,f)>u_0(f,x)-u_0(f,f)=-2$ for all $x\in V\cup C$.
Next we argue that there is no other Nash equilibria. If $\P_1$ plays $f$, clearly the unique best response for $\P_0$ is to also play $f$ since $u_b(f,f)$ gives the unique maximum utility. On the other hand, if $\P_1$ plays a mixed strategy $\mathbf{x_1}$ over $V\cup C$, and $\P_0$ plays a mixed strategy $\mathbf{x_2}$ over $\Sigma-\{f\}$, the social utility $\bar{u}(\mathbf{x_1},\mathbf{x_2})$ is at most $2(n-1)$. If $\P_0$'s utility is smaller than $n$, it is better for $\P_0$ to play $f$ instead (since $u_0(f,x)=n$ for all $x\in V\cup C$). Otherwise, it implies $\P_1$'s utility is at most $n-2$, and it is better for $P_1$ to play $f$ instead (since $u_1(x,f)\geq n-1$ for all $x$). Hence the profile cannot be a Nash equilibrium. The same analysis holds if $\P_1$ plays a mixed strategy over $\Sigma$ with nonzero probability to play actions in $V\cup C$ and $\P_0$ plays any mixed strategy over $\Sigma$ for the following two reasons: 1) Because $u_0(f,\ell)=u_0(\ell,\ell)$ and $u_1(\ell,f)=u_1(\ell,\ell)$, playing $f$ instead would not decrease this part's portion of utility. 2) Because $(f,f)$ gives the maximum utility, even the mixed strategy of $\P_1$ involves nonzero probability of $f$, playing $f$ instead would not decrease this part's portion of utility either, and vice versa. Hence the support of a Nash equilibrium cannot involves actions in $V\cup C$.
So, we only need to check strategy profiles over $L\cup \{f\}$. However, if $\P_1$ puts nonzero probability on $f$, then again, playing $f$ is a strictly better strategy for $\P_0$ since $u_0(f,f)$ is the unique maximum utility and $u_0(f,\ell)=u_0(\ell,\ell)$. Hence except $(f,f)$ the support of any other Nash equilibrium can only involve actions in $L$.
The remaining argument to exclude any other Nash equilibrium over $L$ is similar to that of \cite{ConitzeSa08}. Note that in this case, the expected utility for each player is at most $n-1$. If for some $x\in V$, $\P_0$ puts a probability on playing either $+x$ or $-x$ less than $\frac{1}{n}$, then the expected utility for $\P_1$ to play $x$ is strictly greater than $\frac{1}{n}\cdot 0+\frac{n-1}{n}\cdot n= n-1$.
Furthermore, if for some $\ell\in L$, $\P_0$ puts positive probability on $\ell$ while $\P_1$ on $-\ell$, their utility would be less than $n-1$, and switching to $f$ would be a strictly better strategy.
Hence, these cannot be Nash equilibrium, and we can assume that for each $x\in V$, exactly one of $\{+x,-x\}$ is played with probability $\frac{1}{n}$, and $\P_0$ and $\P_1$ choose the same literal set. Finally, if they play a set of literals whose corresponding assignment does not satisfy a clause $c$, then playing $c$ instead would give utility $n$ and would be a better strategy for both players.
\end{proof}
\section{Warm-up for Immunity and Envy-proofness}\label{sec:connection}
We first discuss the connection between the three notions: Nash Equilibrium, immunity, and envy-proofness in two-player games. Interestingly, their connection is very clean.
Let's formalize and compare the definitions of these notions in general two-player games:
\begin{definition}[Immunity for 2-Player Games\footnote{We note that if we adapt the formula of $(\epsilon,t)$-immunity by Gradwohl and Reingold \cite{GradwohlRe14} to two-player games without the relaxation $\epsilon$, it would be $u_{\bar{b}}(x_b':\mathbf{x_{\bar{b}}})\leq u_{\bar{b}}(\mathbf{x})$ instead. However, we insist on our form that captures the idea that a Byzantine player has no incentive to deviate from an immune profile if he cannot decrease his rival player's payoff by doing so.}]
In a two-player game $G=(\{u_0,u_1\},\{A_0,A_1\})$, we say a strategy profile $\mathbf{x}$ is immune if and only if:
\begin{equation*}\label{eq:twoImmune}
\forall b, \forall x_b'\in A_b, u_{\bar{b}}(x_b':\mathbf{x_{\bar{b}}})\geq u_{\bar{b}}(\mathbf{x}).
\end{equation*}
\end{definition}
\begin{definition}[Envy-proofness for 2-Player Games]
In a two-player game $G=(\{u_0,u_1\},\{A_0,A_1\})$, we say a strategy profile $\mathbf{x}$ is envy-proof if and only if
\begin{align}\label{eq:envyfree}
&\notag \forall b, \forall x_b'\in A_b,
u_b(\mathbf{x_b'}:\mathbf{x_{\bar{b}}})-u_b(\mathbf{x})\leq u_{\bar{b}}(\mathbf{x_b'}:\mathbf{x_{\bar{b}}})-u_{\bar{b}}(\mathbf{x}), \\
&\text{abbreviated as \ } \forall b, \triangle_{b}u_b(\mathbf{x})\leq \triangle_{b}u_{\bar{b}}(\mathbf{x}).
\end{align}
\end{definition}
\paragraph{Compare envy-proofness with immunity.} Immunity is a stability notion against a complete irrational adversary who can be arbitrarily malicious. She would take any action if it can cause a loss of the other player. On the contrary, envy-proofness is against a partial rational player who is envious about the other player's payoff and compare their payoffs all the time. We strengthen that the goal of envy-proofness is not to divide a cake equally or to make everyone happy (which is a too strong requirement in many real cases). Instead, envy-proofness captures the notion that each player has no incentive to change her mind even if she is an envious player.
\paragraph{Compare envy-proofness with Nash equilibria.} While a Nash equilibrium says that a rational player $\P_b$ has no incentive to deviate if $u_b(x_b':\mathbf{x_{\bar{b}}})\leq u_b(\mathbf{x})$, an envy-proof profile suggests that an envious player has no incentive to deviate if $u_b(x_b':\mathbf{x_{\bar{b}}})-u_{\bar{b}}(x_b':\mathbf{x_{\bar{b}}})\leq u_b(\mathbf{x})-u_{\bar{b}}(\mathbf{x})$, which replaces the utility function $u_b(\cdot)$ in the Nash with the difference of the utilities, $(u_b-u_b')(\cdot)$.
\begin{theorem}\label{thm:NashEnvyImmune}
For a 2-player game $G=(\{u_0,u_1\},\{A_0,A_1\})$, an immune Nash equilibrium is envy-proof.
\end{theorem}
\begin{proof}
Let $\mathbf{x}$ be an immune Nash equilibrium. $\forall b, \forall x_b'$, a Nash equilibrium states
\begin{equation*}
u_b(x_b':\mathbf{x_{\bar{b}}})- u_b(\mathbf{x})\leq 0,
\end{equation*}
and an immune profile suggests
\begin{equation*}
u_{\bar{b}}(x_b':\mathbf{x_{\bar{b}}})- u_{\bar{b}}(\mathbf{x})\geq 0,
\end{equation*}
which together imply
\begin{equation*}
\triangle_{b}u_b(\mathbf{x})\leq 0 \leq \triangle_{b}u_{\bar{b}}(\mathbf{x}).
\end{equation*}
\end{proof}
\section{On (Non)-Existence}\label{sec:existence}
We know that every finite game admits at least one Nash equilibrium \cite{Nash50}, but how about the immune profile and the envy-proof profile? Also, does every game admit an immune Nash equilibrium or at least an envy-proof Nash equilibrium? Here we answer these questions for two-player games. Particularly, we show that while every two-player game admits at least one Nash equilibrium, one immune profile, and one envy-proof profile, they are not necessarily to cover each other. That is, some games do not admit any envy-proof Nash equilibrium.
\begin{lemma}\label{lem:immune}
Every finite two-player game admits an immune profile.
\end{lemma}
\begin{proof}
Let $G=(\{u_0,u_1\},\{A_0,A_1\})$ be a 2-player game. An immune profile $\mathbf{x}$ requires
\begin{equation*}
\forall b, \forall x_b', u_{\bar{b}}(x_b':\mathbf{x_{\bar{b}}})\geq u_{\bar{b}}(\mathbf{x}),
\end{equation*}
Define a game $G'=(\{-u_1,-u_0\},\{A_0,A_1\})$. A Nash equilibrium $\mathbf{y}$ of $G'$ satisfies
\begin{equation*}
\forall b, \forall y_b', -u_{\bar{b}}(y_b':\mathbf{y_{\bar{b}}})\leq -u_{\bar{b}}(\mathbf{y}).
\end{equation*}
Hence $\mathbf{y}$ must be an immune profile of $G$.
Since there is always a Nash equilibrium in $G'$, there is also an immune profile in $G$.
\end{proof}
\begin{lemma}\label{lem:envyfree}
Every finite two-player game admits an envy-proof profile.
\end{lemma}
\begin{proof}
Let $G=(\{u_0,u_1\},\{A_0,A_1\})$ be a 2-player game. From \eqref{eq:envyfree}, an envy-proof profile $\mathbf{x}$ says
\begin{equation*}
\forall b, \triangle_{b}u_b(\mathbf{x})\leq \triangle_{b}u_{\bar{b}}(\mathbf{x})
\end{equation*}
Define $G'=(\{u_0-u_1,u_1-u_0\},\{A_0,A_1\})$. A Nash equilibrium $\mathbf{y}$ of $G'$ says
\begin{equation*}
\forall b, \forall y_b', u_b(y_b':\mathbf{y_{\bar{b}}})-u_{\bar{b}}(y_b':\mathbf{y_{\bar{b}}})\leq u_b(\mathbf{y})-u_{\bar{b}}(\mathbf{y}).
\end{equation*}
Hence $\mathbf{y}$ must be an envy-proof profile of $G$.
Since there is always a Nash equilibrium in $G'$, there is also an envy-proof profile in $G$.
\end{proof}
\begin{figure*}[!tp] \large
\centering
\begin{tabular}{|l|l|l|}
\hline
& movie & shopping \\ \hline
movie & 4,4 & 1,3 \\ \hline
shopping & 3,1 & 0,0 \\ \hline
\end{tabular}
\caption{A coordination game. (movie, movie) is the only Nash equilibrium while (shopping, shopping) is the only envy-proof strategy profile.}
\label{fig:nonexistenvy}
\end{figure*}
\begin{lemma}\label{lem:NoEnvyFreeNash}
There are two-player games which do not admit any envy-proof Nash equilibrium.
\end{lemma}
\begin{proof}
Consider a two-player coordination game as described in \figureref{nonexistenvy}. Note that the game admits only one Nash equilibrium (movie, movie) and only one envy-proof strategy profile (shopping, shopping). Hence there is no envy-proof Nash equilibrium in the game.
\end{proof}
\begin{corollary}\label{cor:NoImmuneNash}
There are two-player games which do not admit any immune Nash equilibrium.
\end{corollary}
\begin{proof}
This is from \theoremref{NashEnvyImmune} and \lemmaref{NoEnvyFreeNash}.
\end{proof}
\section{On Efficient Solvability}\label{sec:solvability}
After the discussion of (non)-existence, we are curious about whether these notions are efficiently solvable. Particularly for finding a Nash equilibrium, we already know the problem is PPAD-complete, but how about finding an immune profile and find an envy-proof profile?
\begin{lemma}\label{lem:immuneHard}
Finding an immune profile in a finite two-player game is PPAD-complete.
\end{lemma}
\begin{proofsketch}
This is similar to the proof of \lemmaref{immune} but argued in an inverse way.
Consider any 2-player game $G=(\{u_0,u_1\},\{A_0,A_1\})$. Define a game $G'=(\{-u_1,-u_0\},\{A_0,A_1\})$. Then the problem of finding a Nash equilibrium of $G$ is reduced to finding an immune profile of $G'$. Since finding Nash is PPAD-complete, it is PPAD-complete, too.
\end{proofsketch}
\begin{lemma}\label{lem:envyfreeEasy}
Finding an envy-proof profile in a finite two-player game is polynomial-time solvable.
\end{lemma}
\begin{proofsketch}
In light of the proof of \lemmaref{envyfree}, finding an envy-proof profile can reduce to finding a Nash equilibrium of a zero-sum game. Since zero-sum two-player game is polynomial-time solvable, so is finding an envy-proof profile.
\end{proofsketch}
Recall that from \theoremref{NashEnvyImmune}, envy-proofness is a potentially weaker notion than immunity through the reduction sense. Here by \lemmaref{immuneHard} and \lemmaref{envyfreeEasy}, it is clear to see this point.
\section{Envy-proof Nash and Immune Nash}\label{sec:EnvyImmune}
Now we learned that finding a Nash profile and finding an immune profile are PPAD-hard; meanwhile find an envy-proof profile is efficiently solvable. However, how about immune Nash equilibria and envy-proof Nash equilibria? We have also illustrated that neither an immune Nash equilibrium nor an envy-proof Nash equilibrium does always exist in a game. Hence a follow-up question is whether we can determine their existence in a game efficiently, and whether we can find a solution efficiently if there does exist one.
Gilboa and Zemel \cite{GilboaZe89} and Conitze and Sandholm \cite{ConitzeSa08} show that
it can be NP-complete to determine the existence of a Nash equilibrium with certain simple constraints.
Their results however cannot imply the problem here, nor does their construction in the proof work out. Nevertheless, following the line of reduction from SATISFIABILITY, we show that determining the existence of an envy-proof Nash equilibrium is NP-complete.
\begin{theorem}\label{thm:envyfreeNashHard}
Even in symmetric 2-player games, it is NP-complete to determine whether there is an envy-proof Nash equilibrium.
\end{theorem}
We defer the proof to \sectionref{envyfreeNashHard}. The idea is to construct a game such that every legal assignment of a Boolean formula corresponds to an envy-proof profile in the game, and every Nash equilibrium, except a default Nash equilibrium, corresponds to a satisfiable assignment of the formula. Meanwhile, the default, always-existing Nash equilibrium serves as an absorbing state in the game but is not envy-proof.
On the other hand, somehow surprisingly, while finding an immune profile is harder than finding an envy-proof profile (by \lemmaref{immuneHard} and \lemmaref{envyfreeEasy}), the following theorem suggests that determining the existence of an immune Nash equilibrium is easier than determining the existence of an envy-proof Nash equilibrium, and immune Nash equilibria can be efficiently located given they exist in a game.
\begin{theorem}\label{thm:immuneNashEasy}
Let $G=(\{u_0,u_1\},\{A_0,A_1\})$ be a normal form 2-player game. It is polynomial-time solvable to determine whether there exists an immune Nash equilibrium in $G$ and to output a solution if there exists.
\end{theorem}
\begin{proof}
By definition, a strategy profile $\mathbf{x}$ is an immune Nash equilibrium if and only if
\begin{equation}\label{eq:ImmuneNash}
\forall b, \forall x_b'\in A_b,x'_{\bar{b}}\in A_{\bar{b}},
u_b(\mathbf{x_b}:x_{\bar{b}}')\geq u_b(\mathbf{x})\geq u_b(x_b',\mathbf{x_{\bar{b}}}).
\end{equation}
Now define two games $G_0, G_1$: For all $b, G_b=(\{u_b,-u_b\},A_b)$. Note $\mathbf{x}$ is a Nash equilibrium of $G_0$ and $G_1$ if and only if $\mathbf{x}$ satisfies \eqref{eq:ImmuneNash}. Hence $\mathbf{x}$ is an immune Nash equilibrium of $G$ if and only if $\mathbf{x}$ is a Nash equilibrium of $G_0$ and $G_1$ at the same time. Because both $G_0$ and $G_1$ are zero-sum games,
finding their Nash equilibria equals solving linear programming $LP_0$ and $LP_1$: For $b=0$ and $1$, let $M_b$ be the matrix form of $u_b$. Then,
\begin{align}
LP_b: & \max z_b\\
\notag s.t. \ \ & \mathbf{x}^T M_b\geq z_b \mathbf{1}^T \\
&\notag \mathbf{x}^T\mathbf{1} = 1 \\
&\notag \mathbf{x} \geq \mathbf{0}
\end{align}
Clearly $G$ admits an immune Nash equilibrium if and only if $LP_0$ and $LP_1$ share a common optimum solution. To check whether they do, we solve $LP_0$ and $LP_1$ first. Let $v_b=\max z_b, \forall b$. We claim that the common optimum solution space of $LP_0$ and $LP_1$ is nonempty if and only if the following inequations are feasible:
\begin{equation}\label{eq:immuneNashLP}
\forall b, \mathbf{x}^T M_b\geq v_b \mathbf{1}^T
\end{equation}
Indeed, if $\mathbf{x}$ is a common optimum solution of $LP_0$ and $LP_1$, then certainly it fulfills \eqref{eq:immuneNashLP}. On the other hand, if $\mathbf{x}$ fulfills \eqref{eq:immuneNashLP}, it implies a feasible solution to $LP_b$ with value at least $v_b$ for each $b$. As $v_b$ is the optimum value of $LP_b$, $\mathbf{x}$ is then an optimum solution to $LP_b$ for each $b$.
Since linear programming is polynomial-time solvable, so is the problem.
\end{proof}
\input{app-satproof}
\section{On Approximation}\label{sec:approx}
Here we study the additive approximation of the three stability notions. For example, while an immune profile says that each player's unilateral deviation cannot decrease the rival player's utility, an $\epsilon$-immune profile says that each player's unilateral deviation cannot decrease the rival player's utility more than $\epsilon$ (in absolute value). Intuitively, an approximate immune profile is stable against a lazy Byzantine player who attacks only when he can make sufficiently huge loss of his opponent.
\begin{definition} \label{def:approx}
Let $G=(\{u_0,u_1\},\{A_0,A_1\})$ be a 2-player game and $\mathbf{x}$ be a strategy profile.
\begin{itemize}
\item $x$ is an $\epsilon$-Nash equilibrium if $\forall b, \forall x_b'$, $u_b(x_b':\mathbf{x_{\bar{b}}})\leq u_b(\mathbf{x})+\epsilon$.
\item $x$ is $\epsilon$-Immune if $\forall b, \forall x_b'$, $u_{\bar{b}}(x_b':\mathbf{x_{\bar{b}}})\geq u_{\bar{b}}(\mathbf{x})-\epsilon$.
\item $x$ is $\epsilon$-Envy-proof if $\forall b$, $\triangle_{b}u_b(\mathbf{x})\leq \triangle_{b}u_{\bar{b}}(\mathbf{x})+\epsilon$.
\end{itemize}
\end{definition}
From \lemmaref{immuneHard}, we learned that finding an immune profile of a two-player game $G$ is PPAD-complete. In fact, the proof of \lemmaref{immuneHard} indicates that finding an immune profile of $G$ is the same as finding a Nash equilibrium of another game $G'$ which switches the negative utilities of the two players. Thus, applying a sub-exponential algorithm for finding an $\epsilon$-Nash equilibrium by Lipton, Markakis and Mehta \cite{LiptonMaMe03}, we immediately get the following lemma.
\begin{lemma}\label{lem:approximmune}
For a two-player game $G=(\{u_0,u_1\},A)$, where $|A|=n$, there is a $n^{O(\frac{\ln n}{\epsilon})}$ algorithm to find an $\epsilon$-Immune profile.
\end{lemma}
The algorithm from Lipton et al. \cite{LiptonMaMe03} is based on exhaustively searching all possible $k$-\textit{uniform} $\epsilon$-Nash equilibria, where "$k$-uniform" means that each strategy is a uniform distribution over a multiset of size $k$. They show, through the probability method, that there always exists a $k$-uniform $\epsilon$-Nash equilibria, where the players' utility is close to an exact Nash equilibria so the algorithm works.
Now let's look at a strategy profile with two stability characters again.
Note that by Definition \ref{def:approx}, we say a strategy profile is $\epsilon$-envy-proof Nash equilibrium if it is a Nash equilibrium that is $\epsilon$-envy-proof. Similarly, a strategy profile is an $\epsilon_1$-envy-proof $\epsilon_2$-Nash equilibrium if it is an $\epsilon_2$ Nash equilibrium that is $\epsilon_1$-envy-proof.
From \theoremref{envyfreeNashHard}, we learned that finding an envy-proof Nash equilibrium of a two-player game is NP-complete. So similarly, we resort for potential approximate solutions.
However the theorem from \cite{LiptonMaMe03} does not apply here.
The reason is that even if we can find an $\epsilon$-Nash equilibrium and even if the players' utility can also be close to some exact envy-proof Nash equilibrium, it is unnecessary that the $\epsilon$-Nash equilibrium remains envy-proof.
Nevertheless by adapting the probability method to the problem of approximate envy-proofness here, it can be proved that there is a sub-exponential algorithm for either outputting $\epsilon$-envy-proof $\epsilon$-Nash equilibrium or deciding non-existence of any envy-proof Nash equilibrium in a game.
In the following theorem, we show a more generalized result that has a $2\epsilon$-additive approximation gap between deciding the existence and deciding the non-existence of an approximate envy-proof Nash equilibrium in a game.
\begin{theorem}\label{thm:apxapxefnash}
Let $G=(\{u_0,u_1\},A)$ be a two-player game with $|A|=n$, $\epsilon>0$, and $k=\frac{3\ln n}{\epsilon^2}$.
\begin{enumerate}
\item There is a $O(n^{3k})$-time algorithm for computing a $\epsilon'$-envy-proof $\epsilon$-Nash equilibrium, where $\epsilon'$ is the minimum approximation factor of envy-proofness over all $k$-uniform $\epsilon$-Nash equilibria.
\item If $\epsilon'>2\epsilon$, then there is no $(\epsilon'-2\epsilon)$-envy-proof Nash equilibrium in $G$.
\end{enumerate}
\end{theorem}
\input{app-apxproof}
\section{Multiplayer Games}\label{sec:multiplayer}
We leave several extended questions about multiplayer games for the future, but before the end of this article, we would like to give some more points of view about fault tolerance against envious players in multi-player games. We will define coalitional envy-proofness as a natural extension and relaxation of envy-proofness in multi-player games and define a $\gamma$-varied game that captures the smooth change of the derivative utilities between players.
\subsection{Preliminaries for Multiplayer Games}
First let us reminisce two fault tolerance notions in multiplayer games given in \cite{GradwohlRe14} that generalize approximate Nash equilibria and approximate immune profiles respectively.
\begin{definition}[$(\epsilon,t)$-immunity\cite{GradwohlRe14}]\label{def:etimmune}
In a $m$-player game $G=(\{u_i\}_{i\in[m]},A)$, a strategy profile $\mathbf{x}=\mathbf{x_1}\times\cdots\times\mathbf{x_m}$ is $(\epsilon,t)$-immune if for every set $S\subset N$ of size at most $t$, every $x_S'\in A^{|S|}$, and every $j\not\in S$,
\begin{equation*}
u_j(\mathbf{x}_{-S}:x_S')\geq u_j(\mathbf{x})-\epsilon
\end{equation*}
\end{definition}
\begin{definition}[$(\epsilon,t)$-coalitional Nash equilibrium\cite{GradwohlRe14}] \label{def:etNash}
In a $m$-player game $G=(\{u_i\}_{i\in[m]},A)$, a strategy profile $\mathbf{x}=\mathbf{x_1}\times\cdots\times\mathbf{x_m}$ is $(\epsilon,t)$-coalitional Nash equilibrium if for every set $S\subset N$ of size at most $t$, every $x_S'\in A^{|S|}$, and every $i\in S$,
\begin{equation*}
u_i(\mathbf{x}_{-S}:x_S')\leq u_i(\mathbf{x})+\epsilon
\end{equation*}
\end{definition}
According to \cite{GradwohlRe14}, a profile is said to be $(\epsilon,t)$-immune "if players’ expected utilities do not decrease by more than $\epsilon$ when
any $t$ other players deviate arbitrarily." Nevertheless the formula in their paper is written as $u_i(\mathbf{x})\geq u_i(\mathbf{x}_{-S}:x_S')-\epsilon$. We checked that their main results would hold even if the inequality were revised to $u_i(\mathbf{x}_{-S}:x_S')\geq u_i(\mathbf{x})-\epsilon$. Since the later one is more natural and matches the described notion, we adopt it in \definitionref{etimmune}.
On the other hand, the notion of coalition has a long history in game theory, and similar notions include strong Nash equilibria \cite{Aumann59}, coalition-proof Nash equilibria \cite{BernheimBeWh87} and coalition-proof correlated strategies \cite{MorenoWo96}, etc. Compared to these notions, however, as in \cite{GradwohlRe14}, here we focus on the notion of fault tolerance with a bound on the number of corrupted players.
In \cite{GradwohlRe08,GradwohlRe14}, one main result is about the robustness of Nash equilibria in a $\lambda$-continuous game. We note that the game is equivalent to an alternatively stated $\gamma$-sensitive game in \cite{KearnsPaRoUl13}) for $\gamma=\lambda/m$, where $m$ is the number of players. Essentially each player's utility is quite insensitive to another player's change in the game, which is a relaxation of the insensitive game.
\begin{definition}[$\gamma$-sensitive game\cite{KearnsPaRoUl13,GradwohlRe08,GradwohlRe14}]
A game $G=(\{u_i\}_{i\in[m]},A)$ is said to be $\gamma$-sensitive if
\begin{equation*}
\forall i, \forall x\in A^{m}, \forall x'_i\in A, \forall j\neq i, |u_j(x_{-i}:x_i)-u_j(x)|\leq \gamma.
\end{equation*}
\end{definition}
From \cite{GradwohlRe14}, it was observed that $\gamma$-sensitive games are strongly fault-tolerant with respective to immunity and coalition as the following theorem.
\begin{theorem}[\cite{GradwohlRe14}]\label{thm:sensitiveImmune}
Let $G$ be an $m$-player $\gamma$-sensitive game. Then every Nash equilibirum of $G$ is $(\gamma t,t)$-immune and $(3\gamma t,t+1)$-coalitional.
\end{theorem}
\subsection{Coalitional Envy-proofness}
Now we extend the definition of envy-proofness for the corresponding fault-tolerant notion in multi-player games, which we call $(\epsilon,t)$-coalitional envy-proofness as follows.
\begin{definition}[$(\epsilon,t)$-coalitional envy-proofness] \label{def:etenvyfree}
In a $m$-player game $G=(\{u_i\}_{i\in[m]},A)$, a strategy profile $\mathbf{x}=\mathbf{x_1}\times\cdots\times\mathbf{x_m}$ is $(\epsilon,t)$-coalitional envy-proof if for every set $S\subset N$ of size at most $t$, every $x_S'\in A^{|S|}$, every $i\in S$, and every $j\not\in S$,
\begin{equation*}
u_i(\mathbf{x}_{-S}:x_S')- u_i(\mathbf{x}) \leq u_j(\mathbf{x}_{-S}:x_S')- u_j(\mathbf{x}) +\epsilon
\end{equation*}
\end{definition}
In addition, we define a $\gamma$-varied game that is a natural relaxation of a (generalized) team gameas follows.
\begin{definition}[$\gamma$-varied game]
A game $G=(\{u_i\}_{i\in[m]},A)$ is said to be $\gamma$-varied if
\begin{equation*}
\forall i,j,k, x, |\triangle_i u_j(x)-\triangle_i u_k(x)|\leq \gamma.
\end{equation*}
\end{definition}
Then we have an analog observation for fault-tolerance with respective to envy-proofness in $\gamma$-varied games as follows.
\begin{theorem}\label{thm:variedEnvyfree}
Let $G$ be an $m$-player $\gamma$-varied game. Then every Nash equilibrium of $G$ is $(\gamma t,t)$-coalitional envy-proof.
\end{theorem}
\begin{proofsketch}
Let $\mathbf{x}$ be a Nash equilibrium of $G$. For every $S\subset N, |S|\leq t$, let $S\equiv\{k_1,...,k_t\}$ and $S_{\ell}\equiv\{k_1,...,k_\ell\}, \forall \ell\in [t]$. Then
for every $x_S'\in A^{|S|}$, and every $i\in S, j\not\in S$
\begin{align*}
& \left( u_i(\mathbf{x}_S:x_S')- u_i(\mathbf{x}) \right) - \left( u_j(\mathbf{x}_S:x_S')- u_j(\mathbf{x})\right)\\
= & \left( u_i(\mathbf{x}_{-k_1}:x_{k_1}')-u_i(\mathbf{x}) + \sum_{\ell=2}^t u_i(\mathbf{x}_{-S_{\ell}}:x'_{S_{\ell}})-u_i(\mathbf{x}_{-S_{\ell-1}}:x'_{S_{\ell-1}}) \right)\\
- & \left( u_j(\mathbf{x}_{-k_1}:x_{k_1}')-u_j(\mathbf{x}) + \sum_{\ell=2}^t u_j(\mathbf{x}_{-S_{\ell}}:x'_{S_{\ell}})-u_j(\mathbf{x}_{-S_{\ell-1}}:x'_{S_{\ell-1}}) \right)\\
\leq & \gamma t
\end{align*}
\end{proofsketch}
Next, we generalize \theoremref{NashEnvyImmune} for the corresponding fault-tolerant notions in multi-player games.
\begin{theorem}\label{thm:eNashEnvyImmune}
Let $\mathbf{x}$ be a Nash equilibrium in a game $G$. If $\mathbf{x}$ is $(\epsilon_1,t)$-immune and $(\epsilon_2,t)$-coalitional, then $\mathbf{x}$ is $(\epsilon_1+\epsilon_2,t)$-coalitional envy-proof.
\end{theorem}
\begin{proofsketch} For every $S\subset N$ of size $|S|\leq t$, every $x_S'\in A^{|S|}$, and every $i\in S, j\not\in S$,
by $(\epsilon_2,t)$-coalition,
\begin{equation*}
u_i(\mathbf{x}_{-S}:x'_S)\leq u_i(x)+\epsilon_2,
\end{equation*}
\noindent and by $(\epsilon_1,t)$-immunity,
\begin{equation*}
u_j(\mathbf{x}_{-S}:x'_S)\geq u_j(x)+\epsilon_1.
\end{equation*}
Thus,
\begin{equation*}
u_i(\mathbf{x}_{-S}:x'_S)- u_i(x) \leq
u_j(\mathbf{x}_{-S}:x'_S)- u_j(x)+\epsilon_1+\epsilon_2.
\end{equation*}
\end{proofsketch}
Then by \theoremref{eNashEnvyImmune} and \theoremref{sensitiveImmune}, we have the following corollary.
\begin{corollary}
Let $G$ be a $\gamma$-sensitive game. Then every Nash equilibirum of $G$ is $(4\gamma t,t)$-coalitional envy-proof.
\end{corollary}
Another corollary is about the existence of coalitional envy-proof Nash equilibria in anonymous games, in which the utility of each player is a function of his own action and the empirical distribution of the other players’ actions. Formally, a game $G=(\{u_i\}_{i\in[m]},A)$ is anonymous if for all $i$, for all permutation $\sigma:[m]\backslash \{i\}\rightarrow [m]\backslash \{i\}$, for all $x\in A^m$, $u_i(x)=u_i(x_{\sigma(1)},...,x_{\sigma(i-1)},x_i,x_{\sigma(i+1)},...,x_{\sigma(m)})$.
In \cite{GradwohlRe14}, Gradwohl and Reingold showed that if an $m$-player anonymous game has a constant action space for each player, then the game has an $\epsilon$-Nash equilibrium that is $(\epsilon,t)$-immune and $(4\epsilon,t+1)$-coalitional for $t=O(\sqrt{m})$. Together with \theoremref{eNashEnvyImmune}, we have the following observation.
\begin{corollary}
For every constant $d\in\ensuremath{\mathbb N}\xspace,\epsilon\in (0,1)$, there exists a constant $c$ such that for every anonymous game $G=(\{u_i\}_{i\in[m]},A)$ with $|A|=d$, for $t=c\sqrt{m}$, $G$ admits a $\epsilon$-Nash equilibrium which is $(5\epsilon,t)$-coalitional envy-proof.
\end{corollary}
\section{Introduction}
We study an inherent yet unpredictable factor of players in games that is related to the notorious envy\footnote{In English, malicious envy is more like jealousy, but here we adopt a general term \textit{envy} from the philosophy.} of humanity.
It is observed that in a company, the (un)happiness of employees can depend on, instead of their own salary,
the comparison between their salary and their colleagues'. In some scenarios, the situation can be worse. For example, a (somehow) irrational bidder in an auction can set a bid higher than the true value of an item just because he is unhappy to see his rival bidder win the item.
In general, the outcome of games can deviate from the predicted equilibria, and the players' behavior can disobey their presumed utility function. Even though one remedy is to redefine the utility function and make it as accurate as possible, sometimes the utility function of a real player is hard to predict or to accurately model, and due to the complex humanity, sometimes faulty players can just become spontaneous or unwilling to follow a reasonable utility under a model. Hence it is attractive if an equilibrium has more robust properties against these inherent factors and can prevent potential irrational behavior of players.
To capture the inherent irrational factor due to envy, we define a property called \emph{envy-proofness}.
An envious player would not be satisfied with her current state if she discovers that she can gain more utility "than others", or she would lose less utility "than others" by changing her strategy.
An envy-proof strategy profile is stable with respect to this agitation. That is, A player cannot gain a competitive edge with respect to the utility change over the other players. Note that an envy-proof state may not be \emph{envy-free}: A player in an envy-proof state can envy the other's utility compared to her own, but she has no incentive to deviate.
Envy-proofness is independent of Nash equilibria in general; a Nash equilibrium can be envy-proof or not.
Let's look at its definition for two-player games first.
\begin{definition}[Envy-proof Profile]
In a two-player game $G=(\{u_0,u_1\},\{A_0,A_1\})$, we say a strategy profile $\mathbf{x}= (\mathbf{x_0}, \mathbf{x_1})$, a random variable over $A_0\times A_1$, is envy-proof if and only if the following holds:
\begin{align}\label{eq:envyfree0}
&\notag \forall b, \forall x_b'\in A_b,
u_b(\mathbf{x_b'}:\mathbf{x_{\bar{b}}})-u_b(\mathbf{x})\leq u_{\bar{b}}(\mathbf{x_b'}:\mathbf{x_{\bar{b}}})-u_{\bar{b}}(\mathbf{x}), \\
&\notag (\text{where the notation } u_b(\mathbf{x})=E_{x\sim \mathbf{x}}[u_b(x)]), \\
&\notag \text{abbreviated as} \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\forall b, \triangle_{b}u_b(\mathbf{x})\leq \triangle_{b}u_{\bar{b}}(\mathbf{x})
\end{align}
\end{definition}
\begin{figure}[ht] \large
\centering
\begin{tabular}{|l|l|l|}
\hline
& movie & shopping \\ \hline
movie & 4,4 & 1,3 \\ \hline
shopping & 3,1 & 3,3 \\ \hline
\end{tabular}
\caption{A coordination game. (movie, movie) is a Nash equilibrium but not envy-proof, while (shopping, shopping) is an envy-proof Nash equilibrium.}
\label{fig:envyfree}
\end{figure}
As an example, consider the coordination game described in \figureref{envyfree}. The two players like shopping no mater whether they do it alone or together. On the other hand, they only like to go to the movies together. While (movie, movie) and (shopping, shopping) both sound like a good coordinated strategy profile, one day, the two players quarrel with each other, behave irrationally, and even try to make each other unhappy regardless self-happiness. Then (movie, movie) is no longer a good strategy since it is unstable. It is unstable because each player has an incentive to change the strategy, and assuming the other's strategy is fixed, this deviation can cause the other player to lose more happiness (losing 3 points vs losing 1 point). On the other hand, (shopping, shopping) remains a good strategy because any unilateral deviation of a player can only cause his own loss.
In addition to envy-proofness, we study another
stability notion called immunity.
In \cite{GradwohlRe08,GradwohlRe14}, Gradwohl and Reingold proposed and studied $(\epsilon,t)$-immunity for special games in a fault-tolerance genre. Essentially a strategy profile is $(\epsilon,t)$-immune if
any $t$ players cannot deviate and make some other player's expected utility decrease more than $\epsilon$ (in absolute value).
They showed that in some special games (which we will review briefly later), every Nash equilibrium is $(\epsilon,t)$-immune.
While we will adapt the envy-proof notion to fault-tolerance issues in multi-layer games later in \sectionref{multiplayer},
we focus more on the fundamental properties such as the existence and the computational complexity of exact as well as approximate solutions in general games. Adapting the notion of immunity to two player games, we say a strategy profile is immune if each player's unilateral deviation cannot decrease the other player's utility.
\begin{definition}[Immunity Profile]
In a two-player game $G=(\{u_0,u_1\},\{A_0,A_1\})$, we say a strategy profile $\mathbf{x}$ is immune if and only if:
\begin{equation*}\label{eq:twoImmune0}
\forall b, \forall x_b'\in A_b, u_{\bar{b}}(x_b':\mathbf{x_{\bar{b}}})\geq u_{\bar{b}}(\mathbf{x}),
\end{equation*}
\end{definition}
This paper is then centered around the three stability notions: Nash equilibria, envy-proofness, and immunity and developed from multiple dimensions. First we start by inspecting general two-player games and illustrate a clean connection between the three stability notions in \sectionref{connection}. And in \sectionref{existence}, we study the (non)existence of these notions. We show that while every two-player game has at least an immune strategy profile and an envy-proof strategy profile, an immune Nash equilibrium and an envy-proof Nash equilibrium do not always exist.
Then following these observations regarding existence, in \sectionref{solvability} we are concerned about whether they can
be efficiently found in a game. We show that while finding an immune profile is \emph{PPAD-complete}, finding an envy-proof strategy profile is polynomial-time solvable.
Based on the above discoveries, we check advanced properties of envy-proof Nash equilibria and immune Nash equilibria. In \sectionref{EnvyImmune}, somehow surprisingly, we show while determining the existence of envy-proof Nash equilibria is \emph{NP-complete}, determining the existence of immune Nash equilibria is polynomial-time solvable. Next, because of the PPAD-hardness of finding an immune strategy profile and the NP-hardness of determining the existence of an envy-proof Nash equilibrium, in \sectionref{approx}, we relax the definition of immunity and envy-proofness to resort to approximate solutions. Finally in \sectionref{multiplayer}, we discuss the multi-player fault-tolerance version of these notions. Particularly we show possitive results of $(\epsilon,t)$-coalitional envy-proofness for $\gamma$-varied games,
$\gamma$-sensitive games, and anonymous games.
\section{Related Works and Discussion}\label{sec:relatedwork}
\subsection{Terminology Disambiguation for Envy-Related Notions}
Envy-related notions appeared a few times in several areas. In 1958, Gamow and Stern \cite{GamowSt58} introduce an envy-free concept for fair cake-cutting and chore division.
Essentially the envy-free concept means that each partner believes that their share is at least as great as any other share. Note that the envy-freeness in the problem is quite different from our envy-proofness for general games. Particularly instead of comparing the utilities of the players, we want to compare the change in utilities of the players. That is, the utility between players can be substantially different but a strategy profile would be envy-proof if by any unilateral deviation of a player, the increase of her utility is less or the decrease of her utility is more than the others'.
The term envy-freeness and local envy-freeness also appeared in the problem of generalized second-price (GSP) auction, given by Varian \cite{Varian07} and Edelman et al. \cite{EdelmanOsSc07} independently.
The concept there is to capture the utility optimization of rational players (while excludes some plausible equilibria that cannot provide a good utility). In contrast, the envy-proofness in this paper is motivated by irrational players. Particularly an envy-free bid profile in GSP is necessarily an equilibrium, while an envy-proof profile in a game is not necessary a Nash equilibrium. For example, an irrational bidder in an auction can deviate to decrease his payoff if this can decrease his rival bidder' payoff more.
\subsection{Irrational Players in Games}
While the rationality of the players is a primal assumption in game theory, recently variant points of view are proposed to study or to explain irrational player behaviors in the computer science literature in recent years.
\paragraph{Fault Tolerance.} One of the closest directions to our work is about fault tolerance of Nash equilibria against certain kinds of malicious attacks in large games. Following the line, Bernheim et al. \cite{BernheimBeWh87}, Kalai \cite{Kalai04}, Gradwohl and Reingold \cite{GradwohlRe10,GradwohlRe14} showed that Nash equilibria are robust against certain kind of attacks in general games satisfying specific properties. Particularly, Gradwohl and Reingold \cite{GradwohlRe14} showed that an honest player's payoff in a Nash equilibrium remains almost the same when the number of Byzantine faulty players is less than a threshold. The kind of tolerance is called immunity.
An instant question is that even though approximate immune Nash equilibria exist in games with a special form, it is unclear whether exact/approximate immune profiles or immune Nash equilibria exist in general games and whether it is computationally hard to find a solution. We answered these questions in this work, and so in some sense, we provide a fundamental inspections for the complement points following the line. Also, we study a new notion called envy-proofness that enriches the player's behaviors between fully rational and fully malicious. Interestingly, we will show Nash equilibria, immunity, and envy-proofness are different but relative notions in these questions.
\paragraph{Spitefulness.} Another related direction is about the spiteful/altruistic setting. The setting says that a player's "true" utility is a combination of his own and the others.
(See Chen\cite{PAChen11thesis} and Chen and Micali{ChenMicali16} for detailed reviews and references on this line.)
One major difference between their direction and ours is that the spiteful/altruistic setting assumes certain knowledge of the external factors (called externalities) that decides the true utility function.
This is a necessary step in their approaches since their goals are about how the profit of an auction problem or the social welfare of a game would compromise due to these externalities and how an (approximate) solution of an optimization problem according to the redefined utilities can be obtained.
In contrast, we consider scenarios whether the external factors are unpredictable and focus on issues about Nash equilibria tolerant to irrational players. That is, we assume the metric of the given utility functions are accepted by all the parties and then study equilibria robust against (arbitrarily) malicious or envious behaviors of the players. Nevertheless, we believe that the robustness of Nash equilibria studied in this paper can potentially extend to related problems in their setting. Also, while we show general impossibility and computational hardness results of some of our stability notions, it would be interesting to inspect on more specific problems, as those in this line, for the future studies.
Also, the \emph{price of malice}, defined as the ratio between the social welfare when there are malicious players and the social welfare when there is none, was studies by Moscibroda et al. \cite{MoscibrodaScWa10} in a virus inoculation game and by Babaioff et al. \cite{MalicePapadimitriou07} and Roth \cite{MaliceRoth08} in congestion games. However their works are incomparable to ours since we leave out questions about social welfare for now and focus on questions about equilibria themselves.
\paragraph{Rational Cryptography.} In contrast to a standard assumption where each player is either honest in following the protocol or arbitrarily malicious, rational cryptography concerns scenarios involving both rational and Byzantine players. See \cite{RationalVadhan09,RationalMicali14,RationalMicali13,RationalMicali12,RationalKatz12}, to name a few, and surveys by Nielsen \cite{SurveyNielsen07} and Katz \cite{SurveyKatz09}. On one hand, this setting helps the implementation of protocols go beyond the impossibility regarding fairness in the standard setting. On the other hand, it brings new problems in protocol designs that characterize the complex features of player behaviors.
The tolerance threshold of Byzantine players remains an important issue in secure multiparty computation with rational/partial-rational players, and the design of such protocols usually involve the equilibrium properties (instead of social welfare). Therefore the fundamental studies of the existence, complexity, and approximation of the strategy profiles with respect to the stability notions in the paper has a potential for future development in secure protocols under a rich composition of players.
\section*{Acknowledgement} The author would like to thank Ruta Metha for valuable comments and discussion.
\input{main.bbl}
\fi
\end{document}
\section{Related Work}
\section{Definition}
\begin{definition}[Game]
A normal-form game is described by $G=(\{u_i\}_{i\in[m]},\{A_i\}_{i\in[m]})$ as follows:
\begin{itemize}
\item $m$ indicates that there are $m$ players $\P_1,...,\P_m$.
\item $u_i$ is the utility function of $\P_i$.
\item $A_i$ is a finite set of $\P_i$'s actions.
\end{itemize}
\end{definition}
When $A_i$'s are all the same, we also denote the game by $G=(\{u_i\}_{i\in[m]},A)$. We use a bold font $\mathbf{x}$ to denote a random variable and w.l.o.g. a distribution over a given space.
We will slightly abuse the notation for the expected value of a utility function $u$ by writing $u(\mathbf{x})=E_{x\sim \mathbf{x}}[u(x)]$
\begin{definition}[Strategy Profile] For a game $G=(\{u_i\}_{i\in[m]},\{A_i\}_{i\in[m]})$, a pure strategy profile is a vector $x=(x_1,...,x_m)$ in $\bigotimes_{j=1}^m A_j$, and a mixed strategy profile (or profile for short) is a random variable $\mathbf{x}=(\mathbf{x_1},...,\mathbf{x_m})$ over $\bigotimes_{j=1}^m A_j$.
\end{definition}
We will also denote a strategy profile by $(\mathbf{x}_{-i}:\mathbf{x_i'})$, where $\mathbf{x}_{-i}$ is a product distribution over $\bigotimes_{j\neq i}^m A_j$, and $(\mathbf{x}_{-i}:\mathbf{x_i'})
=(\mathbf{x_i'}:\mathbf{x}_{-i})=(\mathbf{x}_1,...,\mathbf{x_{i-1}},\mathbf{x_{i}'},\mathbf{x_{i+1}}...,\mathbf{x_m})$
\begin{definition}[Nash Equilibrium]
Let $G=(\{u_i\}_{i\in[m]},\{A_i\}_{i\in[m]})$ be a game. We say a strategy profile $\mathbf{x}$ is a (mixed) Nash equilibrium if and only if:
\begin{equation*}\label{eq:twoNash}
\forall i, \forall x_i'\in A_i, u_i(x_i':\mathbf{x_{-i}})\leq u_i(\mathbf{x}),
\end{equation*}
\end{definition}
For two-payer games, we will denote the players by $\P_0$, $\P_1$ and the game by $G=(\{u_0,u_1\},\{A_0,A_1\})$. Similarly, a profile $(\mathbf{x_0},\mathbf{x_1})$ indicates that $\P_0$ plays $\mathbf{x_0}$, and $\P_1$ plays $\mathbf{x_1}$, while a profile
$(\mathbf{x_b}:\mathbf{x_{\bar{b}}})$, the same as $(\mathbf{x_{\bar{b}}}:\mathbf{x_b})$, means $\P_b$ plays $\mathbf{x_b}$ and $\P_{\bar{b}}$ plays $\mathbf{x_{\bar{b}}}$ for some $b\in\{0,1\}$.
|
\section*{Acknowledgments}
We acknowledge financial support from the Deutsche
Forschungsgemeinschaft (DFG) through SFB 787 (MS,MR, AK). This project has also received
funding from the European Unions Horizon 2020 research
and innovation programme under grant agreement No
696656 - Graphene Flagship (EM) and the Swedish Research Council (GB, EM). Finally,
MS gratefully acknowledges inspiring discussions with Samuel Brem (Chalmers University).
\section{Appendix}
To access the exciton dynamics in TMDs, we define a many-particle Hamilton operator including: the dispersion of electrons,\cite{Kormanyos2015} phonons,\cite{Li2013,Jin2014} and photons, the interaction of electrons with a classical electromagnetic field,\cite{Berghauser2014,Selig2016,Thranhardt2000} the carrier-carrier interaction treated within the Hartree Fock approximation,\cite{Berghauser2014} the carrier-phonon interaction treated in an effective deformation potential approximation,\cite{Li2013,Jin2014} and the quantum mechanical interaction of electrons and photons.\cite{Thranhardt2000}
Exploiting the Heisenberg equation of motion and the fundamental commutation relations, we derive time- and momentum-resolved equations for the microscopic polarization $P_\mathbf{Q}$ and the
incoherent exciton occupation
$ N_{\mathbf{Q}}$.
We focus on the low excitation regime, where all terms in third order to the exciting
field can be neglected. In this regime, exciton-exciton and exciton-electron interactions are weak and can be neglected.\cite{Thranhardt2000}
In this limit, the Bloch equation for the microscopic polarization reads:
\begin{align}
&\partial_t P_\mathbf{Q}(t)=\frac{1}{i \hbar}\left( \frac{\hbar^2 \mathbf{Q}^2}{2M}+E - i \gamma_{\text{rad}}\,\delta_\mathbf{Q,0}\right) P_\mathbf{Q}(t) \nonumber \\
&+ \sum_\mathbf{q} \varphi^{*}_\mathbf{q} \,\mathbf{M}^{cv}_\mathbf{q} \cdot \mathbf{A}(t) \,\delta_{\mathbf{Q},0} \nonumber\\
&-\frac{\pi}{\hbar}\sum_{\mathbf{K},\alpha,\pm} |g^{\alpha}_\mathbf{K}|^2 P_\mathbf{Q}(t) (\frac{1}{2}\pm\frac{1}{2}+n_\mathbf{K}^{\alpha}) \delta(\Delta E_\mathbf{Q+K,Q}^{\alpha \pm}).\label{CoExc}
\end{align}
The first line describes the oscillation of the excitonic polarization, where $E$ is the energy of the investigated $A_{1s}$ exciton and $\gamma_{\text{rad}}$ the radiative decay rate of the polarization. The latter is calculated by self-consistent solution of the Maxwell equations and the Bloch equation.\cite{Knorr1996,Selig2016} The second line describes the excitation of the polarization with a classical external field that is determined by the scalar product of the optical matrix element $\mathbf{M}^{cv}_\mathbf{q}=\frac{e_0}{m_0} \langle \mathbf{q} c | \nabla | \mathbf{q} v \rangle $ and the vector potential $\mathbf{A}(t)$ of the exciting field.
Finally, the third line describes the decay of the microscopic polarization due to exciton-phonon scattering leading to the formation of incoherent exciton densities. Here, $g^{\alpha}_\mathbf{q}$ denotes the exciton-phonon coupling element, $n_\mathbf{q}^{\alpha}$ the phonon occupation with the momentum $\mathbf{q}$ and the mode $\alpha$, and finally $\Delta E_\mathbf{K_1,K_2}^{\alpha \pm}=E_\mathbf{K_1}-E_\mathbf{K_2}\pm \hbar \Omega^{\alpha}_\mathbf{K_1-K_2} $ expresses the condition for the energy conservation, where $\hbar \Omega^{\alpha}_\mathbf{q}$ is the energy of the involved phonon, cf. Ref. \onlinecite{Selig2016} for more details on the implementation of exciton-phonon scattering.
In the same limit of low excitation, we derive the equation of motion for the incoherent exciton densities yielding:
\begin{align}
&\partial_t N_{\mathbf{Q}}(t)=\nonumber \\
&\frac{2\pi}{\hbar}\sum_{\mathbf{K},\alpha,\pm}|g^{\alpha}_\mathbf{K}|^2 |P_{\mathbf{Q+K}}(t)|^2 \left(\frac{1}{2}\pm\frac{1}{2}+n_\mathbf{K}^{\alpha}\right) \delta\left(\Delta E_\mathbf{Q,Q+K}^{\alpha \pm}\right)\nonumber \\
&-\frac{2 \pi}{\hbar}\sum_{\mathbf{K},\alpha,\pm}|g^{\alpha}_\mathbf{K}|^2 N_\mathbf{Q}(t)\, \left(\frac{1}{2}\pm\frac{1}{2}+n_\mathbf{K}^{\alpha}\right) \delta\left( \Delta E_\mathbf{Q+K,Q}^{\alpha \pm}\right)\nonumber \\
&+\frac{2 \pi}{\hbar}\sum_{\mathbf{K},\alpha,\pm}|g^{\alpha}_\mathbf{K}|^2 N_\mathbf{Q+K}(t)\, \left(\frac{1}{2}\pm\frac{1}{2}+n_\mathbf{K}^{\alpha}\right) \delta \left( \Delta E_\mathbf{Q,Q+K}^{\alpha \pm}\right)\nonumber \\
&-\frac{2\pi}{\hbar} \sum_{\mathbf{k},\sigma} |M^{\sigma}_\mathbf{Q,k}|^2\, \delta_{\mathbf{Q,k_{\parallel}}} \, N_\mathbf{Q}(t)\, \delta(E_\mathbf{k_\parallel}-\hbar \omega_\mathbf{k}^{\sigma}).\label{InExc}
\end{align}
The first line describes the formation of incoherent exciton densities due to the non radiative decay of coherent excitons. The second and the third line contain exciton phonon scattering terms where the second contributes to out scattering and the third to in scattering processes. In all three lines the $+$ terms denote phonon emission processes, whereas the $-$ terms denote phonon absorption processes. The last line describes the radiative decay of incoherent excitons due to spontaneous emission of photons.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{WSe2Condyn300.png}
\end{center}
\caption{\textbf{Exciton dynamics in WSe$_2$ at \unit[300]{K}}}
\label{supp_Con300}
\end{figure}
Now, we discuss the exciton formation and thermalization in WSe$_2$ at \unit[300]{K}, cf. Fig. \ref{supp_Con300}. Phonon-mediated exciton formation takes place at the same energetic spectral positions as at \unit[77]{K} (discussed in the main text), since the excitonic bandstructure and the phonon dispersion do not depend on temperature. Higher temperatures result in larger phonon occupations and thus a more efficient exciton-phonon scattering can take place resulting in a faster exciton formation and thermalization. Here, we observe already \unit[1]{ps} after the optical excitation a Bose-Einstein distribution of the exciton occupations.
In MoSe$_2$, the exciton formation occurs mainly at the excitonic $\Gamma$ valley resultig in $K-K$ excitons. This is due to the efficient scattering with intravalley acoustic phonons, cf. Figs \ref{supp_Con77Mo} and \ref{supp_Con300Mo}. Increased phonon-assisted exciton formation occurs around \unit[36]{meV}, which coincides with the optical phonon energy in MoSe$_2$ \cite{Jin2014}. Furthermore, we find exciton formation in K-K' states at \unit[10]{meV} above the minimal excitonic energy through the absorption of acoustic $K$ phonons. The energy of the latter is \unit[16]{meV} and the separation between the K-K and the K-K' states is about \unit[7]{meV}. The small amount of created K-$\Lambda$ excitons can be traced back to off-resonant absorption processes. Scattering with phonons leads to thermalization of the exciton occupation, which reaches a steady state after \unit[2]{ps}. Here, K-K states show the highest occupation, since they are the energetically lowest lying states in MoSe$_2$, cf. Fig. \ref{supp_val300}.
At \unit[300]{K}, the formation of excitons is already finished after some tens of fs due to a more efficient exciton-phonon scattering resulting from a larger phonon occupations at higher temperatures. The thermalization of the exciton occupations is also faster resulting in thermalized steady state exciton densities already after approximately \unit[500]{fs}, cf. Fig. \ref{supp_val300}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{MoSe2Condyn77.png}
\end{center}
\caption{\textbf{Exciton dynamics in MoSe$_2$} at \unit[77]{K}.}
\label{supp_Con77Mo}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{MoSe2Condyn300.png}
\end{center}
\caption{\textbf{Exciton dynamics in MoSe$_2$} at \unit[300]{K}.}
\label{supp_Con300Mo}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{valleys300ps.eps}
\end{center}
\caption{\textbf{Exciton densities in different valleys in WSe$_2$ and MoSe$_2$ at \unit[300]{K}}}
\label{supp_val300}
\end{figure}
Note that we treat the exciton-phonon interaction in second-order Born-Markov approximation \cite{Kochbuch} that has been already demonstrated to be an excellent approximation to reproduce the experimentally measured homogeneous linewidths of excitonic resonances.\cite{Selig2016} At higher temperatures, multiple exciton-phonon processes might become important, which would further accellerate the formation and thermalization of excitons.
\bibliographystyle{unsrt}
|
\section{Introduction}
The physics of systems of confined particles has attracted the
interest of researchers from many different areas working on both, theoretical and experimental aspects \cite{wigner_1934,wineland_1987}. One feature of such quantum systems that has recently gained
impulse is the study of entropic measures of entanglement \cite{osenda_2015, garagiola_2016, koscik_2015, lopezrosa_2015,simonovic_2015}. Among the many kind of systems that could be addressed using the physics of the confined systems, in the present work we focus on Wigner molecules, which are the finite-size analogue of Wigner crystals, named after the seminal work of E. Wigner \cite{wigner_1934}. Since the late 80's, when the
first confined linear chains of ions where reported by D. Wineland \cite{wineland_1987}, there has been an increasing capacity to confine, control and manipulate such entities, and has turned the Wigner molecules into a new platform to test the known, and look for new, traits of quantum
mechanics \cite{drewsen_2015}.
Trapped ions are not the only physical systems that allow the
formation of entities like the Wigner molecules (also named Coulomb crystals \cite{drewsen_2015}). The experimental observation of strongly correlated states in quantum dots has attracted considerable interest \cite{cavaliere_2009}. Wigner molecules have also been observed in two-dimensional semiconductor heteroestructures \cite{andrei_1998,piot_2008}, semiconductor quantum dots \cite{kallikaos_2008},
one-dimensional quantum wires \cite{ellenberger_2006,singha_2010,kristinsdottir_2011,meyer_2009}, carbon
nanotubes \cite{deshpande_2008,pecker_2013}, and in crystalline states
for dusty plasma \cite{melzer_2003}. Several theoretical
studies \cite{jauregui_1993,guclu_2008,mendl_2014,gambeta_2014,cavaliere_2014,
cavaliere_2015,cavaliere_2015_2,kylanpaa_2016} have demonstrated that
the physics of these systems with reduced dimensionality usually does not
depend on the shape of the confinement but on its symmetries and
strength (see Ref. \cite{pecak_2016} for an example where shape does have an
influence). In the present work we use a harmonic confinement as a model potential, and represent different physical situations using different interaction potentials. Wigner molecules arise when the interparticle interaction strength is much larger than the kinetic energy. The latter can be related to the temperature of the system \cite{drewsen_2015} and also to the density, or confining energy, of the particles \cite{filinov_2001}. Hence Wigner localization is expected for low density systems or for large
interaction strengths.
The Calogero and Moshinsky models are the most salient examples of
analytically solvable models of confined particles, including the exact
computation of the entanglement entropies
\cite{osenda_2015,garagiola_2016,manzano_2010,dehesa_2012,benavides_2014,
koscik_2010,koscik_2013, koscik_saha_2015}, in this sense it can
also be mentioned the spherium model \cite{toranzo_2015}, and the quasi-solvable
Hook model \cite{taut_1993}. In particular, the Calogero model has been widely
studied in condensed matter physics and has experienced several
revivals \cite{Polychronakos_1993,Polychronakos_2006}, such as the discovery of
an explicit relation of the Calogero model with the fractionary quantum hall
effect \cite{Azuma_1994} and fractional statistics \cite{Murthy_1994}.
{In a previous paper~\cite{garagiola_2016}, we have studied the
behavior of the von Neumann and R\'enyi entropies of the one- and
two-dimensional Calogero model for two particles. By considering anisotropic
confinement in the two-dimensional case we showed that the one-dimensional
regime is reached when the anisotropy of the trap increases, and we also
demonstrated that the R\'enyi entropies present a non analytical behavior in
the neighborhood of those values of the interaction strength parameter for which
the reduced density matrix has finite support.}
Motivated by this, we consider anisotropic harmonic confinement and compute the
exact expression for the occupation numbers or occupancies of the
two-dimensional ground state wave function in the large interaction strength
limit for two particles which interact via different potentials depending on the
distance between the particles. The exact natural orbitals are obtained from the
Schmidt decomposition of the ground state wave function in the same limit and
the occupancies are used to evaluate several quantum information measurements
such as von Neumann and R\'enyi entropies in closed form. {The
method presented here is a generalization of the strategy
developed in Refs.~\cite{koscik_2015,koscik_2010,koscik_2015_2,glasser_2013}.
The two particle one-dimensional systems with Coulomb and inverse powers
interactions are addressed in~\cite{koscik_2015,koscik_2015_2,glasser_2013},
while the
natural orbitals and occupation numbers of elliptically deformed
two-dimensional quantum dots are reported in~\cite{koscik_2010}. Here we
give the analytical expressions of the natural
orbitals, occupation numbers, von Neumann and R\'enyi entropies in the
strong interaction limit for any potential which depends only on the
interparticle distance.}
Our main purpose is to determine the influence of the anisotropy and the type of interparticle interaction
by looking upon the linear, von Neumann, min-entropy, max-entropy
and R\'enyi entropies as entanglement measures. We have a particular interest in which are the differences arising from a short-range interaction with respect to a long range one for which the
emergence of Wigner Molecules has been widely described (see, for example, Refs.
\cite{jauregui_1993,guclu_2008,cavaliere_2014}). With this aim, we study
two interacting potential cases for each interaction range, including one that can be exactly
solved. In the long-range interaction case we consider the inverse power and the
inverse logarithmic potential, and for the short-range interaction we solve the
screened inverse power potential and a Gaussian repulsive interaction. It is important to
emphasize that the inverse power interaction case is used to model quantum dots
\cite{kallikaos_2008} or ion traps \cite{wineland_1987} where the large
interaction regime can be achieved experimentally due to a strong interaction
between the particles or a weak confinement energy scale, for inverse square
power one gets the Calogero model, while the screened Coulomb interaction provides a simple model potential for ions and plasmas \cite{li_2012}.
The paper is organized as follows. The
model is discussed in Section \ref{section_2d2p_sys}. In Section
\ref{section_analytical_derivation}
we show the derivation of the analytical occupancies of two interacting
particles in a two-dimensional anisotropic harmonic trap, while in Section
\ref{section_entropies} we calculate the entropic entanglement measures. We
discuss the results for long- and short-range interaction potentials in Sections
\ref{section_long_range} and \ref{section_short_range} respectively. Finally, a
summary and conclusions are presented in Section
\ref{section_summary_conclusions}.
\section{Confined two-dimensional two-particle systems}
\label{section_2d2p_sys}
The physics of confined particle systems is nowadays very relevant to understand the many recent experiments conducted in cold atom traps or in quantum dots, at least in a qualitative way \cite{andrei_1998, piot_2008, kallikaos_2008, ellenberger_2006, singha_2010, kristinsdottir_2011, meyer_2009}. The models for those systems contain two contributions to the potential energy: one is given by the trap potential and the other by the interaction between the particles. For small dots, containing few electrons the trap potential can be approximated by a harmonic one \cite{cavaliere_2009}, therefore we focus here on two interacting particles in a two-dimensional anisotropic harmonic traps, and implement a method to solve the entanglement spectrum in the large interaction limit for arbitrary interaction potentials. The Hamiltonian for two particles in an anisotropic trap, in atomic units, is
\begin{equation}
\label{H_QD_2D}
H = -\frac{1}{2}\left( \nabla_1^2+\nabla_2^2\right) + \frac{1}{2}\left\lbrace (x_1^2+x_2^2)+\varepsilon^2(y_1^2+y_2^2) \right\rbrace + g V\left( r_{12} ; \left\lbrace \gamma_{i} \right\rbrace \right) \; ,
\end{equation}
\noindent where the frequency of the trap was taken equals to unity, $\varepsilon > 1$ is the anisotropy parameter, $V\left( r_{12} ; \left\lbrace \gamma_{i} \right\rbrace \right)$ denote the interaction potential as a function of the interparticle distance $r_{12}$ and some parameters $\left\lbrace \gamma_{i}
\right\rbrace$, and $g$ is the ratio between the interaction and the confinement
energy scale. By introducing the center of mass $\vec{R} = \frac{1}{2} (\vec{r}_1 + \vec{r}_2)=(X,Y)$ and relative coordinates $\vec{r} = \vec{r}_2-\vec{r}_1=(x,y)$ the Hamiltonian
(\ref{H_QD_2D}) decouple as $H = H^R + H^r$, where
\begin{eqnarray}
\label{H_cal_2D_mc}
& &H^R = -\frac{1}{4} \nabla_R^2 + \left(X^2+ \varepsilon^2 Y^2\right)\; , \\
\label{H_cal_2D_r}
& &H^r = -\nabla_r^2 + V^{eff}(x,y;\varepsilon,\left\lbrace \gamma_{i} \right\rbrace)\; ,
\end{eqnarray}
\noindent and $V^{eff}$ is the effective potential of the relative Hamiltonian given by
\begin{equation}
\label{Veff}
V^{eff}(x,y;\varepsilon,\left\lbrace \gamma_{i} \right\rbrace)=
\frac{1}{4}\left( x^2+ \varepsilon^2 y^2\right) + g V\left( \sqrt{x^2+y^2} ;
\left\lbrace \gamma_{i} \right\rbrace \right)\; .
\end{equation}
The total wave function is then the product of the center of mass wave function and the relative wave function
\begin{equation}
\label{psi_prod}
\Psi(x,y,X,Y) = \psi^R(X,Y) \psi^r(x,y) \; ,
\end{equation}
\noindent and, consequently, the Schr\"odinger equation separates into two
equations
\begin{eqnarray}
\label{Schr_eq_1}
& &H^R\psi^R(\vec{R}\,) = E^R \psi^R(\vec{R}\,)\; ,\\
\label{Schr_eq_2}
& &H^r\psi^r(\vec{r}\,) = E^r \psi^r(\vec{r}\,)\; .
\end{eqnarray}
The solutions of the center of mass equation (Eq. (\ref{Schr_eq_1})) are the
eigenfunctions of the harmonic oscillator that are invariant under particle
exchange.
The relative Hamiltonian, Eqs.~(\ref{H_cal_2D_r}) and~(\ref{Veff}), must be
analysed on a case-specific basis. However, in
the next section we present a method to obtain the large interaction strength
limit of general potentials that fulfil simple requirements.
\section{Derivation of the analytical occupancies}
\label{section_analytical_derivation}
The relative wave function may be obtained by solving the Schr\"odinger
equation in the large interaction strength regime, $g \gg 1$, by using the
harmonic approximation (HA) \cite{james_1998, balzer_2006}. In the framework of the harmonic
approximation one has to find the minima of the effective potential Eq.
(\ref{Veff}) and then the potential is replaced by its Taylor
expansion up to second order about its minima, which satisfy $\nabla V^{eff}(x,y;\varepsilon,\left\lbrace \gamma_{i} \right\rbrace) = 0 $. If the potential is repulsive, decreases monotonously
and $ V \left( r ; \left\lbrace \gamma_{i} \right\rbrace \right) \to
0 $ for $r\to\infty$, with $\varepsilon > 1$, the minima lie on the $x-$axis
and can be written as
\begin{equation}
\label{min_V_eff}
\vec{r}_{min} =\left(\pm x_0, 0 \right) \;\;\;\; \mbox{with $x_0>0$ given
by}\;\;\;\; \frac{1}{2g}=-\left.\left( \frac{1}{r} \frac{\partial V}{\partial
r}\right)\right\vert_{x_0} \,.
\end{equation}
\noindent It is important to notice that when the particles are confined in an
isotropic trap, \textit{i.e} $\varepsilon=1$, the minima degenerate into a
circle of radius $x_0$.
Within the harmonic approximation, a Hamiltonian of uncoupled oscillators is obtained
\begin{equation}
\label{H_Ha}
H^r_{HA} = -\nabla_r^2 + \frac{1}{2} \left\lbrace \omega_x^2 \left( x- x_0 \right)^2 + \frac{1}{2} \left(\varepsilon^2 -1 \right) y^2 \right\rbrace\; ,
\end{equation}
\noindent with a frequency associated to the $x$-coordinate given by
\begin{equation}
\label{wx_2}
\omega_x^2 = \frac{1}{2} \left( 1 + \frac{\left. \frac{\partial^2 V}{\partial
r^2}\right\vert_{x_0} }{\left. -\frac{1}{r} \frac{\partial V}{\partial
r}\right\vert_{x_0}} \right)\; ,
\end{equation}
\noindent where the dependence on the parameters $g$ and $\left\lbrace \gamma_{i} \right\rbrace$ is implicit in $x_0=x_0\left(g, \left\lbrace \gamma_{i} \right\rbrace \right) $.
The totally symmetric ground state wave function $\Psi ^{GS} \left( \vec{r}_1,
\vec{r}_2 \right)$ of the harmonic Hamiltonian $H^R + H^{r}_{HA}$ is a product
of Gaussians. The eigenvalues of the one-particle reduced
density matrix $\rho=\textrm{Tr}_2 \left(\left| \Psi^{GS}\right\rangle
\left\langle \Psi^{GS} \right|\right)$ are explicitly obtained (see supporting
information), and are given by
\begin{equation}
\label{auval_2D}
\Lambda_{l,\tilde{l}} = \Lambda_{l}^{x} \, \Lambda_{\tilde{l}}^{y} \;,
\end{equation}
\noindent where
\begin{equation}
\label{auval_2D_x}
\Lambda_{l}^{x} = \frac{\left(1-\zeta\left( \omega_x\right) \right)}{2 \left( 1 + e^{-\frac{x_0^2\,\omega_x}{\sqrt{2}}} \right) } \zeta\left( \omega_x\right)^{l} \;\;\;\; \mbox{,}\;\;\;\; \zeta\left( \omega_x\right) = \left( \frac{ \left( 2 \omega_x^2\right) ^{\frac{1}{4}} - 1}{\left( 2 \omega_x^2\right) ^{\frac{1}{4}} + 1}\right)^2 \;,
\end{equation}
\noindent and
\begin{equation}
\label{auval_2D_y}
\Lambda_{\tilde{l}}^{y} = \left(1-\xi(\varepsilon)\right) \xi(\varepsilon)^{\tilde{l}} \;\;\;\; \mbox{,}\;\;\;\; \xi(\varepsilon) = \left( \frac{\left( \varepsilon^2 -1 \right) ^\frac{1}{4} - \sqrt{\varepsilon}}{\left( \varepsilon^2 -1 \right) ^\frac{1}{4} + \sqrt{\varepsilon}}\right)^2 \;.
\end{equation}
\noindent where $l,\tilde{l}=0,1,2,\ldots$. Each eigenvalue, or occupancy,
is doubly degenerate due to the particle exchange symmetry.
The limiting values and behavior of $\zeta\left(\omega_x\right)$ and $\xi(\varepsilon)$ are needed to compute the entropic quantities. We note then that for $\omega_x>0$, $\zeta\left( \omega_x\right)$
is always below unity and $\zeta\left( \omega_x\right) \to 1$ when $\omega_x
\to \infty$, while for $\varepsilon>1$, $\xi(\varepsilon)$ remains below one
and $\xi(\varepsilon) \to 1$ for $\varepsilon \to 1^+$, then we must be specially careful in the isotropic confinement case (see Eq. (\ref{auval_2D_y})). For large anisotropy parameter
$\varepsilon \gg 1 $ one gets $\xi(\varepsilon) \to 0$ and the occupancies reach
the asymptotic values of the one dimensional model $\Lambda_{l}^{x}$.
\section{Entropies in the large interaction strength limit}
\label{section_entropies}
The entanglement can be measured using different entropic quantities. If
$\{\Lambda_i\}$ is the complete set of eigenvalues, then the
R\'enyi entropies are a family of such entropies defined by
\begin{equation}
\label{def_Renyi}
S^{\alpha} = \frac{1}{1-\alpha} \,\log_2 \mbox{Tr} \, \rho^{\alpha} =
\frac{1}{1-\alpha}\, \log_2 \left( \sum_i \Lambda^{\alpha}_i \right) \,,
\end{equation}
\noindent which are widely used in many-body or extended systems
\cite{Calabrese2011a,Alba2010}.
Special values of the parameter
$\alpha$ allow to recover other entropies, being the min- and max-entropy good
examples obtained by taking the limits $\alpha\rightarrow \infty$ and
$\alpha\rightarrow 0$, respectively. The min-entropy serves as a lower bound to
the entanglement measures obtained from the whole family of entropies. The
Hartley or
max-entropy, $S^{0}=\log_2 R$, only depends on the Schmidt rank $R$
of the spectrum distribution and is a measure of bipartite entanglement which
serves as a criterion for efficient classical representation of the state
\cite{amico_2014}. The distribution of the entanglement spectrum can be better
understood by computing the R\'enyi entropies for many different values of the
parameter $\alpha$ \cite{garagiola_2016}. The von Neumann entropy is given by
\begin{equation}
\label{def_vN}
S_{vN} = - \mbox{Tr} \left(\rho \log_2 \rho \right) = - \sum_i \Lambda_i \log_2
\Lambda_i \,.
\end{equation}
\noindent The von Neumann entropy has been used to study entanglement in
continuous variables systems and spin models \cite{Alba2010, osenda_2007,
pont_2010}. It can be recovered from the R\'enyi entropies in the
limit $\alpha
\to 1$. Finally, some authors use the linear entropy, defined by
\begin{equation}
\label{def_SL}
S_{L} = 1 - \mbox{Tr} \, \rho^2 = 1 - \sum_i \Lambda^{2}_i \,,
\end{equation}
\noindent since for continuous variable systems the calculation of $\mbox{Tr}
\,
\rho^2$ is reduced to a single integral. Even though, the linear entropy has no
relevant information for the systems studied in the present work, we compute it
for the sake of completeness.
Once we have obtained the occupancies it is possible to calculate the quantum entropies. Since these calculations involve geometric series in
$\zeta\left(\omega_x\right)$ and $\xi(\varepsilon)$, the limiting values must be carefully computed.
Let us start with the R\'enyi entropies defined by Eq. (\ref{def_Renyi}). It is straightforward to show that due to the separability of the wave function, the R\'enyi entropies in the large interaction strength limit are the sum of the entropy associated to $\psi_x(x_1,x_2)$ and
$\psi_y(y_1,y_2)$, then
\begin{equation}
\label{Renyi_xy}
S^{\alpha} = S^{\alpha}_{x}(\omega_x) + S^{\alpha}_{y}(\varepsilon) \,,
\end{equation}
\noindent where
\begin{equation}
\label{Renyi_2D_x}
S^{\alpha}_{x} \left( \omega_x\right) = \frac{1}{1-\alpha} \log_2 \left(
\frac{(1-\zeta\left( \omega_x\right))^ \alpha}{(1-\zeta\left( \omega_x\right)^\alpha)}
\right) + 1 \,,
\end{equation}
\noindent and
\begin{equation}
\label{Renyi_2D_y}
S^{\alpha}_{y} (\varepsilon) = \frac{1}{1-\alpha} \log_2 \left(
\frac{(1-\xi(\varepsilon))^ \alpha}{(1-\xi(\varepsilon)^\alpha)}
\right) \,.
\end{equation}
Again, due to the separability of the wave function, we can write the two-dimensional von Neumann entropy of Eq. (\ref{def_vN}) as
\begin{equation}
\label{S_VN}
S_{vN} = S_{x}^{1}\left( \omega_x\right) + S_{y}^{1}(\varepsilon) \;,
\end{equation}
\noindent where each one of the terms in the sum has the form of a
one-dimensional von Neumann entropy \cite{koscik_2015}, {\it i.e.}
\begin{equation}
\label{Sx}
S_{x}^{1}\left( \omega_x\right) = - \frac{\log_2\left( \left(1-\zeta\left( \omega_x\right) \right)^{(1-\zeta\left( \omega_x\right))} \zeta\left( \omega_x\right)^{\zeta\left( \omega_x\right)} \right)}{ \left(1-\zeta\left( \omega_x\right)\right)} + 1\;,
\end{equation}
\begin{equation}
\label{Sy}
S_{y}^{1}(\varepsilon) = - \frac{\log_2\left( \left(1-\xi(\varepsilon)\right)^{(1-\xi(\varepsilon))} \xi(\varepsilon)^{\xi(\varepsilon)} \right)}{ \left(1-\xi(\varepsilon)\right)}\;.
\end{equation}
The super-index points that the von Neumann entropy can be obtained as a limiting case of the
R\'enyi entropies when $\alpha\to 1$.
It is worth to notice that from Eq. (\ref{Renyi_2D_x}) and (\ref{Renyi_2D_y})
it is straightforward to show that the min-entropy $S^\infty$, can also be written as a two-term sum:
\begin{equation}
\label{min_entropy}
S^\infty = \lim_{\alpha\to\infty} \left( S^{\alpha}_{x}\left( \omega_x\right) + S^{\alpha}_{y}(\varepsilon) \right) = \lim_{\alpha\to\infty} S^{\alpha}_{x}\left( \omega_x\right) + \lim_{\alpha\to\infty} S^{\alpha}_{y}(\varepsilon)= S^{\infty}_{x}\left( \omega_x\right)+S^{\infty}_{y}(\varepsilon) \,.
\end{equation}
The Hartley or max-entropy in the large interaction strength limit can also be calculated as a limiting case with $\alpha\to 0$, $S^{0}=\log_2 R$, and has finite value only when the one-particle reduced density matrix has finite support.
The two-dimensional linear entropy defined by Eq. (\ref{def_SL}) gives
\begin{equation}
\label{le_final}
S_{L} = 1 - \frac{1}{2} \; \frac{1-\zeta\left(
\omega_x\right)}{1+\zeta\left( \omega_x\right)} \;
\frac{1-\xi(\varepsilon)}{1+\xi(\varepsilon)} \;.
\end{equation}
\noindent For the isotropic model $\varepsilon \to 1^{+}$, $\xi(\varepsilon)\to 1$ and the linear entropy goes to one, while for any other value of $\varepsilon$ the linear entropy remains below one.
A comment on the extension of the previous results to dimension $D$ is in
place. They can be extended if one considers $D-1$ anisotropy parameters (see
supporting information). The von Neumann and R\'enyi entropies are the sum of
$D$ terms each one associated to one cartesian coordinate and, as we
demonstrated for the two-dimensional case, the $x$-entropy term depends on the
parameters of the interaction potential through $\omega_x$ and each one of the
remaining terms depend on only one of the $D-1$ anisotropy parameters.
\begin{figure}
\begin{center}
\includegraphics[width=0.3\textwidth]{fig_1a.eps}
\includegraphics[width=0.32\textwidth]{fig_1b.eps}
\end{center}
\caption{\label{f_S_2D_omega} Both terms of the two-dimensional von Neumann (black dashed), min-entropy (magenta dash-dotted) and R\'enyi entropies with $\alpha = 0.2,\,0.4,\,0.8,\,1.5,\,2$ (red, blue, green, orange and cyan full lines respectively). (a) $S_y$ as a function of the anisotropy parameter $\varepsilon$. (b) One dimensional entropies $S_x$ as a function of the squared frequency, $\omega_x^2$.
}
\end{figure}
The two terms of the two-dimensional von Neumann, min-entropy and R\'enyi entropies
($\alpha = 0.2, \,0.4, \,0.8, \,1.5, \,2$) are depicted in Fig. \ref{f_S_2D_omega}. Let us first discuss the behavior of the entropies with respect to the anisotropy of the trap, and afterwards the influence of the interparticle interaction.
As can be appreciated in Fig. \ref{f_S_2D_omega}(a), for the isotropic model ($\varepsilon \to 1^{+}$) the entropies $S^{\alpha}_y(\varepsilon)$ diverge logarithmically, whilst for any other values
of $\varepsilon$ they remain finite. By calculating the first derivative of Eq. (\ref{Sy}) it is
straightforward to show that
\begin{equation}
\label{vNe_limit}
S_{vN} \sim -\frac{\ln(\varepsilon-1)}{\ln
16} \;\;\;\; \mbox{for}\;\; \varepsilon \sim 1^{+}\,.
\end{equation}
\noindent This asymptotic leading term is depicted in the figure as a yellow dashed line which makes the logarithmic divergence of the isotropic von Neumann entropy evident. Actually, for $\varepsilon \to 1^{+}$ the von Neumann, min-entropy and the family of R\'enyi entropies present this same behaviour. The figure also shows that for large anisotropy parameter the entropies $S^{\alpha}_{y}(\varepsilon)$ vanish. In other words, for $\varepsilon \gg 1$ the one dimensional problem is recovered and the von Neumann, R\'enyi and min-entropy reach the one dimensional values $S_{x}^{\alpha}(\omega_x)$.
The behavior of the $x$-entropies (denoted by $S^{\alpha}_x(\omega_x)$) as a function of the frequency is shown in Fig. \ref{f_S_2D_omega}(b). The figure shows that the entropies are decreasing functions of the frequency for $0< \omega_x^2 < 1/2$, and increasing functions for $\omega_x^2 > 1/2$.
Actually, the entropies diverge logarithmically for large frequencies and also
for $\omega_x \to 0$, because in these limits one gets that $\zeta(\omega_x) \to 1$.
The entropy of a given system is computed using the frequency $\omega_x$ obtained by the harmonic approximation Eq. (\ref{wx_2}). If it remains finite for large interactions parameters $g \gg 1$, then the von Neumann, min-entropy and the family of R\'enyi entropies are finite for the anisotropic model and diverge
logarithmically for the isotropic model. In the deformed or anisotropic case the
particles crystallize around the two classical minima of the relative
Hamiltonian giving rise to a Wigner molecule, while for the isotropic model
those minima degenerate into a circle, the particles are no
longer localized around discrete minima and this lack of information is
reflected in the divergence of the entanglement entropies. If the obtained frequency
increases monotonously for large interactions, the von Neumann, min-entropy and
the family of R\'enyi entropies diverge logarithmically for any anisotropy
parameter. In this sense, the behaviour of the system is defined by the
one-dimensional entropy $S_{x}^{\alpha}(\omega_x)$.
The previous analysis can be understood more qualitatively by using the
Heisenberg uncertainty principle.
The width of the Gaussian wave packet in the relative coordinate
$\psi^{r}(\vec{r})$ (ground state of the Hamiltonian Eq. (\ref{H_Ha})) goes to
zero
when the frequency increases. Actually, the relative position and momentum
uncertainty are $\Delta x^{r}_{HA} = \sqrt{\langle \left( x_2 - x_1 \right)^2
\rangle - \langle x_2 - x_1 \rangle^2} = 2^{\frac{1}{4}}/\sqrt{\omega_x}$ and
$\Delta p^{r}_{HA} = \sqrt{\omega_x}/2^{\frac{5}{4}}$, then if $\omega_{x} \to
0$ we obtain that $\Delta x^{r}_{HA} \to \infty$ and, conversely, when $\omega_{x} \to \infty$
it is straightforward to show that $\Delta p^{r}_{HA} \to \infty$.
Thus, we see that for $\omega_{x} \to \infty$ the entropy of the Wigner molecule diverge
because the position is completely determined and consequently the momentum
uncertainty diverges, we refer to this limit as \emph{strong
crystallization}. The opposite case, $\omega_{x} \to 0$ leads to a
well defined momentum state and hence we have no knowledge of the position. In
both cases the divergence in the position or momentum width leads to the
divergence of the entanglement entropies. Furthermore, the divergence of the
$y$-entropies could also be explained in a similar way: for the isotropic model
the minima degenerate into a circle and consequently the particles are no longer
localized around any definite angular positions, but the state has definite
angular momentum.
For $\omega_x^2=1/2$ the entropies have
their minimum value equal to unity. Around this point, the von Neumann,
min-entropy and R\'enyi entropies with $\alpha>1$ present an analytical
behaviour while the R\'enyi entropies with $\alpha<1$ have a non-analytical
behaviour. The von Neumann and R\'enyi entropies with $\alpha = 0.4,\,0.5,\,0.6$
and their first derivatives around the point $\omega_x^2=1/2$ are shown in Fig.
\ref{f_S_x_der} (a) and (b) respectively. It shows that the R\'enyi entropies present an infinite
derivative for $\alpha=0.4$, discontinuous derivative for $\alpha=0.5$ and a
continuous derivative with infinite second derivative for $\alpha=0.6$, while
the von Neumann entropy ($\alpha\rightarrow {}^{+}1$) is an
analytical function of the frequency.
\begin{figure}
\begin{center}
\includegraphics[height=0.29\textwidth]{fig_2a.eps}
\includegraphics[height=0.30\textwidth]{fig_2b.eps}
\end{center}
\caption{\label{f_S_x_der} (a) Entropy terms $S^{\alpha}_x(\omega_x)$ and (b) their derivatives around the point $\omega_x^2=1/2$. The von Neumann (black dashed), and R\'enyi entropies with $\alpha = 0.4,\,0.5,\,0.6$ (red, blue and green full lines respectively) are shown.
}
\end{figure}
Recent studies by Amico and co-workers in $1/2$-spin chains show the physical
implications of non-monotonous properties of the R\'enyi entropies in many-body
systems with topological order due to a truncation of the support of the reduced
density matrix \cite{amico_2013,amico_2013_2,amico_2014_2}. In
Ref.~\cite{osenda_2015} some of the present authors found that the Calogero
model in one dimension has a finite number of non-zero occupancies for a
discrete set of values of the interaction parameter, and in Ref.
\cite{garagiola_2016} we demonstrated that in those particular values of
the interaction parameter the R\'enyi entropies present a non-analytical
behaviour.
Summarizing, non-analytical behaviour of the R\'enyi entropies exposes the
finite support of the reduced density matrix. In the present case, taking
$\omega_x^2=1/2$ in Eq. (\ref{auval_2D_x}) it is straightforward
to see that for this particular frequency there is only two non vanishing
occupancies $\Lambda^{x}_{0}$ associated to the two lowest natural orbitals in
the $x$-coordinate.
In the following sections we apply our findings to study the behavior of the occupancies and entropic entanglement measures in the large interaction strength limit for different cases divided as long- or short-range potentials. From now on we calculate only the one dimensional entropy $S_{x}^{\alpha}(\omega_x)$, since the behaviour of the entropy terms $S^{\alpha}_{y}(\varepsilon)$ were already
analysed.
\section{Long-range interaction potentials}
\label{section_long_range}
In the present section we consider two long-range interactions to exemplify our results: the inverse power interaction and inverse logarithmic interaction.
\subsection{Inverse power interaction}
The inverse power potential is
\begin{equation}
\label{V_inv_power}
V^{ip}\left( r ; \beta \right) = \frac{1}{r^{2\beta}}\; .
\end{equation}
\noindent For this potential $x_0$ and $\omega_x$, Eqs. (\ref{min_V_eff}) and (\ref{wx_2}), can be obtained exactly and give
\begin{equation}
\label{min_freq_V_inv_power}
x_0=(4g\beta)^{\frac{1}{2(\beta+1)}} \;\;\;\; \mbox{and}\;\;\;\; \omega_x^2=\beta + 1\,.
\end{equation}
\noindent Thus, $x_0$ increases when increasing the interaction strength
parameter $g$, but the frequency remains invariant. For $\beta = \frac{1}{2},1$ one gets the Hook and the Calogero model respectively.
\begin{figure}
\begin{center}
\includegraphics[height=0.3\textwidth]{fig_3a.eps}
\end{center}
\caption{ \label{f_auval_2D} Occupancies as a function of $\ln
(\varepsilon-1)$ (see Eq. (\ref{auval_2D})) obtained for large interaction
strength parameter, $g \gg 1$. $\Lambda_{l,\tilde{l}}$ with $l=0$ and
$\tilde{l}=0,1,2,...,20$ from top to bottom. The dominant one-dimensional
eigenvalue is also shown (grey dashed line) \cite{koscik_2015}.}
\end{figure}
Let us start with the Calogero model. The occupancies
defined in Eq. (\ref{auval_2D}) for large interaction strength parameter
$g\gg1$ are shown as a function
of $\ln (\varepsilon-1)$ in Fig. \ref{f_auval_2D}, where the grey dashed line
is the dominant one-dimensional occupancy in the large interaction strength
limit, obtained from Eq. (\ref{auval_2D_x}), that is in agreement with the
value reported in Ref. \cite{koscik_2015}. The figure shows that for $\epsilon
\to 1^{+}$ (isotropic model) the occupancies go to zero, but note that
their sum is always equal to $1/2$ due to the mentioned
double degeneracy \cite{koscik_2010}. When the
anisotropy increases all the occupancies $\Lambda_{l,\tilde{l}}$ with
$\tilde{l}\neq 0$ present a local maximum. For fixed $l$ the value of the
anisotropy parameter at which the maximum occurs decreases when $\tilde{l}$
increases, while for fixed $\tilde{l}$ this value is the same for each
$l$. For $\varepsilon \gg 1$ the
Hamiltonian reduces to a one dimensional oscillator and the occupancies
$\Lambda_{l,0}$ reach the asymptotic values of the one dimensional model. For
values of $\varepsilon$ near $\varepsilon_c = \sqrt{5}$ the occupancies with
$\tilde{l}=0$ stabilize on the one dimensional values and those with
$\tilde{l}\neq 0$ saturate at vanishingly small values. This feature can be
explained if one takes into account that for $\varepsilon =\varepsilon_c$ the
relative Hamiltonian Eq. (\ref{H_Ha}) reduces to a harmonic oscillator in polar
coordinates around each minimum. More
generally, for arbitrary $\beta$, the one dimensional regime is reached at the value $\varepsilon_c = \sqrt{2 \left(\beta+1\right)+1} $. In this case the effective potential of the relative
Hamiltonian Eq. (\ref{Veff}) is isotropic in a small neighborhood around its minima.
For $\varepsilon > \varepsilon_c$, the largest occupancy $\Lambda_{00}$ reaches
the value $\sim 0.4853$, and the sum of
all the remaining occupancies is only $\sim 0.0147$; this means that the two natural orbitals associated to this eigenvalue are the only two that are occupied while all the others natural orbitals contribution are negligible, and consequently, the spatial wave functions are quite similar to those two natural orbitals \cite{koscik_2010}.
\begin{figure}
\begin{center}
\includegraphics[height=0.3\textwidth]{fig_4.eps}
\end{center}
\caption{\label{f_Ss} One-dimensional entropy terms $S^{\alpha}_x$ obtained for large interaction strength $g \gg 1$, as a function of the exponent of the interaction between particles, $\beta$. The von Neumann (black dashed), min-entropy (magenta dash-dotted) and R\'enyi entropies with $\alpha = 0.2,\,0.4,\,0.8,\,1.5,\,2$ (red, blue, green, orange and cyan full lines respectively) are shown.
}
\end{figure}
As we mentioned above, the dependence with $\beta$ is present only
through the entropy term $S^{\alpha}_x(\omega_x)$. The width of the Gaussian wave packet in the relative coordinate is finite, $\Delta x^{r}_{HA} = 2^{\frac{1}{4}}/(\beta + 1)^\frac{1}{4}$ and consequently the von Neumann, min-entropy and R\'enyi entropies are finite. However, the max-entropy diverges due to the infinite support of the one-particle density matrix. Notice that in the limit $\beta \to \infty$ the entropies diverge logarithmically
due to the divergence in the momentum uncertainty. This behavior can be seen in Fig. \ref{f_Ss} where the von Neumann, the min-entropy and R\'enyi entropies are depicted as a function of the parameter $\beta$. The R\'enyi entropies increases for decreasing $\alpha$, and the von Neumann entropy is a limiting case with $\alpha \to 1$. It is important to emphasize that taking the limit $\beta \to 0$ in the entropies does not result in the same entropies obtained for a system with harmonic confinement and a constant interaction (Eq. (\ref{V_inv_power}) with $\beta=0$), since this limit does not commute with the large interaction limit.
\subsection{Inverse logarithmic interaction}
The potential for inverse logarithmic interparticle interaction is
\begin{equation}
\label{V_inv_log}
V^{il}\left( r \right) = \frac{1}{\ln (r+1)}\; .
\end{equation}
\noindent In this case $x_0$ and $\omega_x$ satisfy the following equations
\begin{equation}
\label{min_freq_V_inv_log}
2g = x_0 \left( x_0+1 \right) \ln^{2} \left( x_0+1 \right) \;\;\;\; \mbox{and}\;\;\;\; \omega_x^2= \frac{1}{2} \left\lbrace 1 + \left( \frac{\frac{2}{\ln \left( x_0+1 \right)}+1}{\frac{1}{x_0}+1} \right) \right\rbrace \,.
\end{equation}
\noindent For large interaction strength parameter the value of $x_0$ increases
when $g$ increases, and consequently, the frequency goes to unity. Therefore,
for $g\gg 1$, the one-dimensional von Neumann
and R\'enyi entropies with $\alpha>0$ remain finite, but once more the
max-entropy diverges. We included two figures in the supporting
information showing the qualitative behavior
of $x_0$ and the R\'enyi entropy as a function of $\alpha$. The same
analysis performed in the inverse power interaction case can be done for the
inverse logarithmic potential.
\section{Short-range interaction potentials}
\label{section_short_range}
In the present section we consider two particles in a two-dimensional
anisotropic harmonic trap with two different short-range interactions: the screened inverse
power interaction and Gaussian repulsive interaction.
\subsection{The screened inverse power interaction}
For the screened inverse power interaction the potential is
\begin{equation}
\label{V_screened_inv_power}
V^{sip}\left( r ; \lbrace\beta,\gamma\rbrace \right) = \frac{e^{-\gamma r}}{r^{2\beta}}\; ,
\end{equation}
\noindent where $1/\gamma$ is the cut-off distance. In this case $x_0$ and $\omega_x$ are given by
\begin{equation}
\label{min_freq_V_screened_inv_power}
2g= \frac{e^{\gamma x_0} x_0^{2(1+\beta)}}{2\beta+\gamma x_0} \;\;\;\; \mbox{and}\;\;\;\; \omega_x^2= \frac{1}{2} \left( 1+ \frac{2 \beta}{2 \beta +\gamma x_0} + 2\beta +\gamma x_0\right) \,.
\end{equation}
\noindent Notice that taking $\gamma=0$ the minima and the frequency of the
inverse power interaction is recovered, and for $\beta=0$ the interaction has
exponential decay. For large interaction strength parameter $g$ the minima
and the frequency increase monotonously with $g$, and consequently, the one-dimensional von Neumann and R\'enyi entropies diverge logarithmically. As we mentioned in section \ref{section_entropies} the divergence of the entropies can be explained as arising from the momentum uncertainty $\Delta p^{r}_{HA} = \sqrt{\omega_x}/2^{\frac{5}{4}}$, which diverges when $\omega_{x} \to \infty$. Actually, the larger the $\gamma$ parameter is, the larger the frequency is and the higher the entanglement
entropies are, this behaviour is shown in Fig.
\ref{f_S_screened_inverse_power} where the one-dimensional von Neumann entropy
is depicted as a function of the interaction strength
for $\beta=1$ and $\gamma = 0,\, 1/2,\,1,\;2$.
\begin{figure}
\begin{center}
\includegraphics[height=0.3\textwidth]{fig_6.eps}
\end{center}
\caption{ \label{f_S_screened_inverse_power} One-dimensional von Neumann entropy $S_x^{1}$ as a function of the interaction strength parameter $g$, with $\beta=1$ and $\gamma = 0,\, 1/2,\,1,\;2$ (from bottom to top, black, red, blue and green lines).
}
\end{figure}
\subsection{The Gaussian repulsive interaction}
In this subsection we consider the following interaction potential
\begin{equation}
\label{V_Gaussian}
V^{gr}\left( r ; \sigma\right) = e^{-\frac{r^2}{2\sigma^2}} \; ,
\end{equation}
\noindent where $\sigma$ is the half width of the potential. In this case $x_0$ and $\omega_x$ can be found exactly
\begin{equation}
\label{min_freq_V_Gaussian}
x_0 = \sigma \sqrt{2 \ln \left( \frac{2 g}{\sigma^2} \right)} \;\;\;\; \mbox{and}\;\;\;\; \omega_x^2= \frac{1}{2} \frac{x_0^2}{\sigma^2} = \sqrt{ \ln \left( \frac{2 g}{\sigma^2} \right)} \;\;\;\; \mbox{with}\;\;\;\; g \geq \frac{\sigma^2}{2}\,.
\end{equation}
\noindent They are increasing functions of the interaction strength parameter $g$, thus, the one-dimensional von Neumann and R\'enyi entropies in the large interaction strength limit diverge
logarithmically. We interpret this divergence in the same way as for the
screened inverse power interaction. It is worth to mention that the limit
$\sigma \to 0$ does not reproduce the results of a delta interaction, which have a finite von Neumann entropy \cite{avakian_1987}, since this limit does not commute with the large interaction strength limit.
From Eq. (\ref{min_freq_V_Gaussian}) it is straightforward to show that for $g=g_c$ with
\begin{equation}
\label{g_c_Gaussian}
g_c = \frac{ \sigma^2 e^{\frac{1}{2}}}{2}\,,
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[height=0.3\textwidth]{fig_7a.eps}
\includegraphics[height=0.3\textwidth]{fig_7b.eps}
\end{center}
\caption{ \label{f_occupancies_Gaussian} One-dimensional occupancies $\Lambda_{l}^{x}$ (Eq. \ref{auval_2D_x}) with (a) $l=0$ and (b) $l=1$, for $\sigma=10$, the value $g_c$ is depicted as a gray dashed line.
}
\end{figure}
\noindent $x_0 = \sigma$ and $\omega_x^2= \frac{1}{2}$, therefore, as was
explained in section \ref{section_analytical_derivation} all
the occupancies vanish except two of them. The reduced
density matrix has then a finite support and the R\'enyi entropies
with $\alpha <1$ have a non-analytical behaviour, while the von Neumann entropy and R\'enyi entropies with $\alpha>1$ present a minimum at $g=g_c$. The behaviour of the first and second degenerate
occupancies (Eq. (\ref{auval_2D_x})) for $\sigma=10$ is depicted in Fig.
\ref{f_occupancies_Gaussian} together with the value $g_c$ as a gray dashed
line. The first two occupancies $\Lambda_0^x$ (Fig.
\ref{f_occupancies_Gaussian}(a)) reach the maximum value $1/2$ for $g=g_c$,
value at which all the others occupancies vanish as can be appreciated for
$\Lambda_1^x$ in Fig. \ref{f_occupancies_Gaussian}(b).
Summarizing, we found that two trapped particles with a Gaussian repulsive
interaction between them, have a reduced density matrix with infinite support
(infinite non-vanishing occupancies) for all the interaction
strengths but for $g=g_c$, value at which all the occupancies vanish except
two, and the support is finite. Nevertheless, it is important to notice that the
largest occupancy $\Lambda_0^x \gtrsim 0.49$ throughout the considered range of
$g$ values, and the sum of all the remaining occupancies is $\lesssim 0.01$.
Therefore, in the neighborhood of $g_c$, the two natural orbitals associated to
$\Lambda_0^x$ are the main contributions to the spatial wave function expansion.
\section{Summary and Conclusions}
\label{section_summary_conclusions}
In this work we present analytical expressions in the large interaction strength limit for the occupancies and quantum entropies for the ground state of a two-particle Wigner molecule in a two-dimensional anisotropic harmonic trap. Our main result is that one is able to determine the influence of the anisotropy
and the range of the interparticle interaction looking upon the entropic entanglement measures.
The wave function is obtained within the framework of the harmonic approximation
for large interaction strength values, and once we have the ground state
wave function, we calculate the occupancies from the Schmidt decomposition of the
reduced density matrix. We obtain doubly degenerate occupancies, and relate this to
the equivalence between particle exchange in the wave function and the exchange
between the two minima of the effective potential of the relative Hamiltonian.
The linear, von Neumann, min-entropy, max-entropy and R\'enyi entropies are
calculated exactly in terms of the occupancies as a function of the anisotropy parameter and the parameters of the interaction potential.
We found that, due to the coordinate separability of the wave function, the von Neumann, min-entropy, max-entropy and R\'enyi entropies are a sum of terms associated to each coordinate, and that only one of these
terms depends on the anisotropy parameter and the other term is associated to
the interaction potential. Consequently, the behavior of the entropies with
respect to the anisotropy parameter can be analyzed without regard of the
interaction potential, and the dependence on the interaction potential is
entirely defined by the frequency obtained by the harmonic approximation of the
one-dimensional problem. Moreover, we generalize these results to
dimensions higher than two, see details in the supporting information.
We show that when the frequency remains finite for large
interactions, then the von Neumann, min-entropy and the family of
R\'enyi entropies remain finite for the anisotropic model and diverge
logarithmically for the isotropic model. The divergence of the
entanglement measure entropies of the isotropic model can be understood as follows: in the
deformed or anisotropic case the particles locate around the two classical
minima of the relative Hamiltonian forming a Wigner molecule, while for the
isotropic model those minima degenerate into a circle, the
particles are no longer localized and this lack of
information is reflected in the divergence of the entanglement entropies. If the
frequency increases monotonously for large interactions then, the von Neumann,
min-entropy and the family of R\'enyi entropies diverge logarithmically for any
anisotropy parameter. In this sense, the influence of the interaction potential
is present only in the one-dimensional entropies.
The previous interaction-independent analysis allows us to apply
them to different interactions straightforwardly. We group the interactions
into short and long-range potentials and show the differences of the
obtained results between the groups. For long-range interaction potentials, the frequencies remain finite in the large interaction strength limit, and the von Neumann, min-entropy and R\'enyi entropies are finite. In contradistinction, for short-range interaction potentials, the frequencies increase monotonously as a function of the interaction strength and, consequently, the one-dimensional von Neumann and R\'enyi entropies diverge in the large interaction strength limit. The divergence of the entanglement entropies can be explained as arising from the momentum uncertainty divergence at large frequencies. It is important to mention that the one-dimensional von Neumann, min-entropy and R\'enyi entropies of the inverse power interaction model diverge logarithmically when the power of the inverse interaction increases (see Eqs. (\ref{Renyi_2D_x}) and (\ref{Sx})), since in this limit the interaction between the particles goes to a short range one.
We also demonstrate that when the frequency associated to the interaction potential, satisfy $\omega_x^2=1/2$, the entropies have their minimum value equal to unity. Actually, the von Neumann, min-entropy and R\'enyi entropies with $\alpha>1$ present an analytical behaviour around this point, while the R\'enyi entropies with $\alpha<1$ have a non-analytical behaviour, which exposes the finite support of the reduced density matrix. For this particular frequency there is only two non vanishing occupancies $\Lambda^{x}_{0}$. Similar features were also recently reported by Amico and co-workers for $1/2$-spin chains \cite{amico_2013,amico_2013_2,amico_2014_2}, and for the Calogero model \cite{garagiola_2016} by the present authors. We further illustrate these features showing that two trapped particles with a Gaussian repulsive interaction have a reduced density matrix with infinite support (infinite non-vanishing occupancies) for all Hamiltonian parameters except for those values that allow $\omega_x^2=1/2$, where all the occupancies vanish except two, and the reduced density matrix has finite support.
As a final remark, there is a very recent work concerning a
system of two Coulombically interacting particles confined to a $D-1$ sphere,
where the dependence of the entanglement measures on the radius of the system
and the spatial dimensionality has been investigated \cite{toranzo_2015}. Thus,
as future perspectives we would like to study the effects of the dimensionality
and the interaction strength on the entanglement of two confined particles which
interact via a general potential, taking as a starting point the results
obtained in the generalization to dimensions higher than two presented in the
second section of the supporting information.
\section*{Acknowledgments}
We acknowledge SECYT-UNC and CONICET for partial financial support of this
project. E.C. would like to thank Alvaro Cuestas for an exhaustive reading of
the manuscript.
|
\section{Introduction}
\label{sec:intro}
Hydrogen burning, the conversion of four protons into an alpha particle in the interior of stars, is the most important energy source in the universe and it is also responsible for the existence of several chemical elements. Besides the pp-chains powering e.g. our Sun, catalytic reactions cycles, like the various CNO cycles play the major role in hydrogen burning \cite{ade11,wie10}. Depending on the temperature and chemical composition of the stellar plasma, different CNO cycles can take place involving various isotopes of carbon, nitrogen, oxygen and fluorine.
The $^{17}$O(p,$\gamma$)$^{18}$F reaction, which competes with the $\alpha$-emission in $^{17}$O(p,$\alpha$)$^{14}$N \cite{str16} is the starting point of the third CNO cycle. This cycle is activated in various stellar conditions such as red giant and asymptotic giant stars and classical novae. The abundances of fluorine and the heavy oxygen isotopes are strongly related to the operation of this cycle and therefore the rates of the participating reactions must be known.
Below 1.5\,MeV the $^{17}$O(p,$\gamma$)$^{18}$F reaction is characterized by many broad and narrow resonances. Therefore, the temperature dependence of the $^{17}$O(p,$\gamma$)$^{18}$F thermonuclear reaction rate shows a complicated picture. The rate depends on the direct capture component as well as on the narrow low energy resonances and the tails of the higher energy broad resonances (see e.g. Fig.\,10 in ref.\,\cite{dil14} for the contribution of the different components to the reaction rate). An R-matrix fit to the experimental data is therefore inevitable to provide reaction rates at various temperatures for stellar models.
The first cross section measurement of $^{17}$O(p,$\gamma$)$^{18}$F was carried out several decades ago by C. Rolfs in a wide energy range between 300\,keV and 1.9\,MeV \cite{rol73}. After the turn of the century, several experimental studies were carried out mostly concentrating on the low energy region below about 500\,keV \cite{fox04,fox05,cha05,cha07,new10,kon12,hag12,sco12,dil14,buc15}. (The only exception was the work of A. Kontos \textit{et al.} \cite{kon12} which extended up to 1.6\,MeV.) The comparison of the new precise data with the results of \cite{rol73} revealed some discrepancy both in the absolute scale and the energy dependence of the cross section at the lowest energies studied by \cite{rol73}.
In most of the previous experiments the cross section of $^{17}$O(p,$\gamma$)$^{18}$F was measured with in-beam $\gamma$-spectroscopy: the prompt $\gamma$-radiation from the formed $^{18}$F nucleus was detected. The complicated level scheme of $^{18}$F (see e.g. Fig.\,1. in ref.\,\cite{buc15}) implies that the detection of many primary and secondary transitions is necessary for the cross section determination. This represents a source of uncertainty in the experiments. In order to provide the astrophysically relevant total cross section, all the transitions must be measured and care must be taken to measure even the weakest $\gamma$-lines. The angular distributions of all the $\gamma$-emissions must also be known. Moreover, in order to measure low cross sections, close target-detector geometries are typically used leading to strong true coincidence summing effects.
All these experimental difficulties can be avoided by the application of the activation method, which was used by only two experiments before at energies below 400\,keV \cite{cha05,cha07,sco12,dil14}. The reaction product of $^{17}$O(p,$\gamma$)$^{18}$F is radioactive, decays by positron emission \footnote{A weak electron capture decay branching also exist with 3\,\% probability} with a half-life of 109.77\,$\pm$\,0.05\,minutes \cite{til95}. The decay is entirely to the ground state of $^{18}$O, no $\gamma$-radiation follows thus the decay. The emission of the 511\,keV $\gamma$-radiation following the positron annihilation, on the other hand, allows the measurement of the decay by $\gamma$-detection. By measuring the $^{18}$F activity the number of reaction product and therefore the total reaction cross section can be determined directly. The activation measurement of $^{17}$O(p,$\gamma$)$^{18}$F provides therefore an independent means of cross section determination which can be used to check earlier experimental data and provide a constraint for R-matrix calculations regarding the total reaction cross section.
The aim of the present work is therefore to measure the $^{17}$O(p,$\gamma$)$^{18}$F cross section with the activation method in a wide energy range. The next section provides detailed information about the experimental technique, the results are presented in Sec.\,\ref{sec:results} while Sec.\,\ref{sec:summary} provides the summary and conclusions.
\section{Experimental procedure}
\label{sec:experiment}
\subsection{Target preparation and characterization}
\label{subsec:target}
Solid state oxygen targets were produced by anodic oxidation of tantalum disks in water enriched in $^{17}$O. With this technique Ta$_2$O$_5$ layers can be produced with well defined Ta:O ratio and the targets have high stability under beam bombardment. The anodization setup was the same as used recently by the LUNA collaboration for the low energy $^{17}$O(p,$\gamma$)$^{18}$F cross section measurements \cite{sco12,dil14}. Full details of the anodization device and the preparation procedure have been published by the LUNA collaboration \cite{cac12}, here only the most important features and the differences are summarized.
Two water samples were used for the target preparations. The isotopic abundances of the $^{16}$O, $^{17}$O and $^{18}$O isotopes, respectively, were the following: (15.5\,$\pm$\,0.6)\%, (77.8\,$\pm$\,0.6)\% and (6.7\,$\pm$\,0.2)\% (sample 1.) and (39.5\,$\pm$\,0.6)\%, (27.4\,$\pm$\,0.6)\% and (33.1\,$\pm$\,0.6)\% (sample 2.). These values are quoted by the supplier.
Applying two different anodization voltages (24V and 50V), targets with two different thicknesses were produced. Altogether seven targets were prepared from the two water samples and with the two thicknesses. Intercomparison of the different targets were done by carrying out activation at the same proton energy on targets with different isotopic composition and/or thickness.
As the determination of the number of target atoms is crucial for the precise cross section measurements, different experimental techniques were used to determine this quantity. First, the Ta:O stoichiometry ratio and the thickness of the oxide layer were measured with Rutherford Backscattering Spectrometry (RBS). The first set of RBS measurements were carried out before the start of the activation experiments at the microbeam setup installed at the 5\,MV Van de Graaff accelerator of Atomki \cite{hus16}. A 1.6\,MeV $\alpha$ beam bombarded the Ta$_2$O$_5$ targets and the scattered particles were detected by two ion implanted Si detectors positioned at 135 and 165 degrees with respect to the beam direction. Exploiting the high lateral resolution of the microbeam setup, spectra were recorded at several different positions on the target surfaces. This test proved that the thickness and stoichiometry of the targets are uniform along the whole surface of the targets. The spectra were analysed using the SIMNRA code \cite{SIMNRA} which provided the areal density of the O atoms as well as the Ta:O ratio.
A second set of RBS measurement was carried out after the activation experiments using a completely independent setup, namely the activation chamber itself (i.e. similar beam size and position to the proton beam used for the activations, see below). At the Tandetron accelerator a 10\,MeV $^{16}$O$^{4+}$ beam bombarded the Ta$_2$O$_5$ targets and a Si detector built into the activation chamber detected the backscattered ions.
Figure \ref{fig:RBS} shows typical spectra of the two RBS measurements. The measured data as well as the fits using the SIMNRA code are shown. The results of the $^{16}$O RBS measurements were in good agreement with the ones obtained with $\alpha$-RBS (see below).
The ratio of the Ta:O atoms was found to be 0.411\,$\pm$\,0.015 in agreement with the stoichiometric value of 0.4.
\begin{figure}
\includegraphics[angle=270,width=\columnwidth]{fig1a_alpha_RBS.eps}\\
\includegraphics[angle=270,width=\columnwidth]{fig1b_16O_RBS.eps
\caption{\label{fig:RBS} Relevant parts of the RBS spectra of a Ta$_2$O$_5$ target measured with $\alpha$-beam at the microprobe facility (upper panel) and with $^{16}$O$^{4+}$ beam in the activation chamber at the Tandetron accelerator.}
\end{figure}
If the Ta:O ratio is known, a totally independent target thickness value can be obtained by the measurement of the resonance profile on a suitable nuclear resonance. We have investigated the target thicknesses also by this method using both $^{17}$O and $^{18}$O isotope content of the targets. The E$_p$\,=\,1098\,keV and E$_p$\,=\,1925\,keV resonances in the $^{17}$O(p,$\gamma$)$^{18}$F and $^{18}$O(p,$\gamma$)$^{19}$F reactions, respectively, were used to measure the target profiles. A 100\,\% relative efficiency HPGe detector were placed next to the activation chamber at zero degree with its front face about 1\,cm distance from the target. The yield of the strongest transition was used for the measurement of the profiles which was the 937\,keV transition of the first excited state to the ground state in the case of $^{18}$F ($^{17}$O(p,$\gamma$)$^{18}$F reaction) and the 197 keV transition of the second excited state to the ground state in the case of $^{19}$F ($^{18}$O(p,$\gamma$)$^{19}$F reaction). The number of target atoms was obtained from the width of the target profiles using the Ta:O ratio given by the RBS measurements.
Figure\,\ref{fig:resonance} shows a typical resonance profile measured with the $^{17}$O(p,$\gamma$)$^{18}$F reaction. The target thickness obtained from the resonance profile measurement using the two reactions gave consistent results. The comparison with the RBS results, however, revealed a roughly 9\,\% systematic difference. The RBS measurements resulted in systematically higher thickness values. Table\,\ref{tab:target} summarizes the thickness results of a given target (prepared with 50\,V anodization voltage) obtained with the four measurements.
\begin{figure}
\includegraphics[angle=270,width=\columnwidth]{fig2_resonance.eps
\caption{\label{fig:resonance} Measured profile on the E$_p$\,=\,1098\,keV resonance in $^{17}$O(p,$\gamma$)$^{18}$F. The target thickness was obtained from the width of the fitted resonance profile.}
\end{figure}
The uncertainties quoted in the table are statistical only stemming form the fit of the RBS spectra and the resonance profiles. Taking into account only these errors, the two methods are in contradiction. If, however, one includes the uncertainty of the stopping power, the results can be considered to be consistent. It is difficult to quantify the uncertainty of stopping power in our experiment as the stopping of protons, $\alpha$-particles and $^{16}$O isotopes should be considered in O and Ta, and the related information \cite{SRIMerror} in the widely used SRIM code indicates uncertainties from about 2\,\% up to 6\,\%. Most likely the deviation of the thickness values has its origin in the uncertainty of the stopping power. Therefore, we have adopted the average of the two methods and assigned a conservative 6\,\% uncertainty to the number of target atoms.
\begin{table}
\caption{\label{tab:target} Results of the various thickness measurements on one of the targets. See text for details.
}
\begin{ruledtabular}
\begin{tabular}{ll}
Method & No. of O atoms \\
& [10$^{17}$ atoms/cm$^2$] \\
\hline
$\alpha$-RBS & 5.10\,$\pm$\,0.13 \\
$^{16}$O-RBS & 5.00\,$\pm$\,0.20 \\
$^{17}$O(p,$\gamma$)$^{18}$F resonance & 4.67\,$\pm$\,0.15 \\
$^{18}$O(p,$\gamma$)$^{19}$F resonance & 4.63\,$\pm$\,0.18 \\
\hline
RBS average & 5.07\,$\pm$\,0.11 \\
resonance average & 4.65\,$\pm$\,0.12 \\
\hline
adopted & 4.87\,$\pm$\,0.29 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\subsection{Activations}
\label{subsec:activations}
The activations were carried out at the new Tandetron laboratory of Atomki where a 2\,MV Tandetron accelerator manufactured by High Voltage Engineering Europa B.V. was installed in 2015. The energy calibration of the accelerator was carried out by measuring resonances in the $^{27}$Al(p,$\gamma$)$^{28}$Si reaction and the neutron thresholds in $^{7}$Li(p,n)$^{7}$Be and $^{13}$C(p,n)$^{13}$O reactions \cite{raj16}.
The Tandetron provided proton beams in the energy range between 500\,keV and 1.8\,MeV and the beam current was limited to about 5\,$\mu$A in order to avoid target degradation. The lifetime of the targets was also increased by using an off-axis target chamber where the beam spot was shifted from the target center by 6\,mm. By rotating the target between the consecutive activations, fresh or not heavily bombarded target spots could be selected. The target chamber was insulated from the rest of the beam line and served as a Faraday cup in order to determine the number of projectiles by charge integration. A secondary electron suppression voltage of -300\,V was applied behind the 4\,mm diameter entrance aperture of the chamber.
Depending on the cross section, the length of the irradiations varied between 15\,minutes and 5 hours.
Although the beam intensity during the irradiations were typically very stable, in order to follow the possible fluctuations, the beam current was recorded in multichannel scaling mode with one minute time basis. The recorded time dependence of the beam current was then used in the analysis.
\subsection{Measurement of the $^{18}$F decay}
\label{subsec:decay}
After the irradiation the target was removed from the chamber and transported to the counting laboratory where a 100\,\% relative efficiency HPGe detector equipped with full 4\,$\pi$ lead shielding was used to measure the annihilation $\gamma$-radiation of the targets. The $\gamma$-countings started typically 15 minutes after the end of the irradiation and the spectra were recorded in every 10 minutes in order to follow the $^{18}$F decay.
Since the 511\,keV annihilation radiation is present also in the laboratory background and can come from many possible sources, it is crucial to determine the background. The length of the countings was therefore typically 16 hours. Towards the end of this counting period, the activity of $^{18}$F decayed to a negligible level and therefore the 511\,keV background level could be estimated. This was always found to be consistent with the laboratory background measured without target, indicating that no long-lived positron emitter was created in the targets.
In some cases excess in the 511\,keV activity was observed at the beginning of the counting period indicating the production of some short-lived positron emitter. From its decay rate it was identified as $^{13}$N produced by the $^{12}$C(p,$\gamma$)$^{13}$N reaction on carbon impurity of the target. This identification was also supported by the fact that such a deviation from the pure $^{18}$F decay was observed mostly around 500\,keV proton energy where the $^{12}$C(p,$\gamma$)$^{13}$N reaction has high cross section due to a broad resonance at about 420\,keV \cite{bur08}.
In such cases roughly the first one hour of the counting was omitted from the analysis.
The decay of the 511\,keV activity could always be fitted well using the literature half-life of $^{18}$F. As an example, Fig.\,\ref{fig:decay} shows the decay curve measured after the irradiation at 520\,keV. The figure indicates the above discussed $^{13}$N contribution, the $^{18}$F decay fitted with the literature half-life and the laboratory background level of the 511\,keV line.
\begin{figure}
\includegraphics[angle=270,width=\columnwidth]{fig3_decay.eps
\caption{\label{fig:decay} Decay of $^{18}$F in a target irradiated with an 520\,keV proton beam. Taking into account the initial short lived positron emitter and the laboratory background, the decay can be well fitted with the 109.77\,min half-life of $^{18}$F.}
\end{figure}
In order to maximize the detection efficiency, the targets were placed in close geometry onto the detector, directly on top of the detector end cap. Since the two annihilation photons are emitted from the source at opposite directions and no other X-ray or $\gamma$-radiation follows the $^{18}$F decay, true coincidence summing effect was not present in this measurement in spite of the close source-to-detector geometry \footnote{The true coincidence of the two annihilation gammas through a Compton scattering process was observed to cause less than 0.5\,\% loss of counts from the 511\,keV peak and was therefore neglected.}. The summing effect, on the other hand, is significant in the case of any multiline calibration source which could be used for the measurement of the absolute detection efficiency. The absolute efficiency was therefore measured in the counting geometry only with single line calibration sources. Calibrated $^{7}$Be, $^{65}$Zn, $^{85}$Sr and $^{137}$Cs sources were used to obtain the efficiency curve of the detector. The $^{85}$Sr source was especially useful for the efficiency determination as it has a single $\gamma$-line at 514\,keV, very close to the relevant 511\,keV. The absolute efficiency was measured with 3\,\% uncertainty which includes also the beam spot size and target positioning effect.
\section{Experimental results}
\label{sec:results}
The cross section of $^{17}$O(p,$\gamma$)$^{18}$F was measured between proton energies of 500\,keV and 1.8\,MeV. The selection of the actual proton energies was based on the structure of the $^{17}$O(p,$\gamma$)$^{18}$F excitation function. The two broad resonances at 590 and 717\,keV proton energies were measured with fine energy steps. The low energy tails of these resonances have significant contributions to the astrophysical reaction rate of $^{17}$O(p,$\gamma$)$^{18}$F especially at higher temperatures.
The energy regions where there are no resonances were covered with fewer data points. The aim of these measurements were to fix the absolute value of the direct capture part of the cross section which again has an important contribution to the reaction rate.
In the studied energy range there are several narrow resonances at proton energies of 517, 673, 741, 826, 926, 1098, 1240, 1270 and 1345 keV. Activation runs at these energies were also carried out with the aim of confirming their existence in the (p,$\gamma$) channel and check their resonance energies. The widths of these resonances, on the other hand, are often comparable with the target thicknesses used in the present work, the determination of the strengths of these resonances was therefore not aimed.
The obtained cross section results are listed in Table\,\ref{tab:results}. The first column shows the energy of the proton beam provided by the Tandetron accelerator. Based on the accelerator calibration, this value is known with a total uncertainty of less than 0.5\,keV. The energy loss of the beam in the target layer is given in the second column. Especially near the resonances the cross section changes significantly in the energy range covered by the target thickness. An effective proton energy was therefore calculated. For this calculation it was supposed that the cross section has a linear energy dependence in the energy range of the target. The slope of the cross section was estimated based on the adjacent experimental data points and on the shape of the excitation function as measured by previous works. The effective energy was then given by the median of the yield curve within the target thickness \cite{lem08}. The uncertainty of the effective energy as listed in the table was estimated based on the slope of the cross section function. Higher energy uncertainties were assigned to the data points near the narrow resonances where the cross section changes strongly within the target thickness.
\begingroup
\begin{table*}
\caption{\label{tab:results} Measured cross section of the $^{17}$O(p,$\gamma$)$^{18}$F reaction in the present work. The quoted cross section uncertainties are statistical only. For the total uncertainty, 7.6\,\% systematic uncertainty must be added quadratically to the relative statistical uncertainties.}
\begin{ruledtabular}
\begin{tabular}{llr@{\hspace{-2mm}}c@{\hspace{-2mm}}lr@{\hspace{-2mm}}c@{\hspace{-2mm}}l||llr@{\hspace{-2mm}}c@{\hspace{-2mm}}lr@{\hspace{-2mm}}c@{\hspace{-2mm}}l}
E$_p$ & Energy loss & \multicolumn{3}{c}{E$_{p\rm ,eff.}$} & \multicolumn{3}{c||}{Cross section} & E$_p$ & Energy loss & \multicolumn{3}{c}{E$_{p\rm ,eff.}$} & \multicolumn{3}{c}{Cross section}\\
& in target\footnote{See Section \ref{subsec:target} for information about the target thicknesses} & & & & & & & & in target$^a$ \\
keV & keV & \multicolumn{3}{c}{keV} & \multicolumn{3}{c||}{$\mu$barn} & keV & keV & \multicolumn{3}{c}{keV} & \multicolumn{3}{c}{$\mu$barn}\\
\hline
500.0 & 4.38 & 497.8 & $\pm$ & 1.4 & 0.592 & $\pm$ & 0.021 & 789.7 & 7.23 & 786.0 & $\pm$ & 2.1 & 3.30 & $\pm$ & 0.1 \\
509.8 & 9.05 & 505.4 & $\pm$ & 2.5 & 0.587 & $\pm$ & 0.118 & 819.8 & 3.41 & 818.2 & $\pm$ & 1.3 & 2.36 & $\pm$ & 1.0 \\
514.7 & 4.32 & 512.6 & $\pm$ & 1.4 & 0.738 & $\pm$ & 0.120 & 824.8 & 3.40 & 823.2 & $\pm$ & 1.3 & 9.51 & $\pm$ & 0.3 \\
519.8 & 4.30 & 517.7 & $\pm$ & 1.4 & 16.3 & $\pm$ & 1.8 & 829.7 & 3.38 & 828.0 & $\pm$ & 1.3 & 12.2 & $\pm$ & 0.3 \\
524.7 & 4.28 & 522.0 & $\pm$ & 2.0 & 6.13 & $\pm$ & 0.37 & 834.7 & 3.37 & 832.9 & $\pm$ & 1.3 & 2.5 & $\pm$ & 0.20 \\
529.7 & 4.26 & 527.1 & $\pm$ & 1.8 & 1.37 & $\pm$ & 0.38 & 880.0 & 3.28 & 878.4 & $\pm$ & 1.2 & 2.06 & $\pm$ & 0.13 \\
539.7 & 4.22 & 537.7 & $\pm$ & 1.4 & 1.96 & $\pm$ & 0.23 & 919.8 & 6.69 & 916.4 & $\pm$ & 1.8 & 2.18 & $\pm$ & 0.12 \\
549.7 & 8.72 & 545.9 & $\pm$ & 2.8 & 2.93 & $\pm$ & 0.14 & 924.8 & 3.20 & 924.0 & $\pm$ & 1.6 & 1.97 & $\pm$ & 0.09 \\
559.8 & 8.64 & 556.1 & $\pm$ & 2.9 & 4.95 & $\pm$ & 0.59 & 929.7 & 6.66 & 926.4 & $\pm$ & 1.9 & 12.1 & $\pm$ & 0.2 \\
569.8 & 4.11 & 568.0 & $\pm$ & 1.6 & 10.3 & $\pm$ & 1.0 & 934.7 & 3.19 & 933.1 & $\pm$ & 1.2 & 2.65 & $\pm$ & 0.09 \\
579.7 & 8.49 & 576.7 & $\pm$ & 3.4 & 30.5 & $\pm$ & 0.8 & 999.7 & 6.43 & 996.5 & $\pm$ & 1.8 & 2.67 & $\pm$ & 0.08 \\
584.7 & 4.06 & 582.9 & $\pm$ & 1.5 & 67.8 & $\pm$ & 1.2 & 1089.7 & 6.17 & 1086.8 & $\pm$ & 1.9 & 3.55 & $\pm$ & 0.08 \\
589.7 & 4.04 & 587.7 & $\pm$ & 1.3 & 107 & $\pm$ & 0.8 & 1096.8 & 2.95 & 1095.3 & $\pm$ & 1.1 & 11.7 & $\pm$ & 2.8 \\
599.7 & 8.35 & 595.0 & $\pm$ & 2.7 & 70.7 & $\pm$ & 1.2 & 1101.8 & 6.14 & 1098.8 & $\pm$ & 1.8 & 176 & $\pm$ & 0.8 \\
609.8 & 8.28 & 604.5 & $\pm$ & 3.3 & 25.0 & $\pm$ & 0.4 & 1106.7 & 6.13 & 1103.7 & $\pm$ & 1.7 & 35.5 & $\pm$ & 1.6 \\
619.8 & 8.21 & 614.9 & $\pm$ & 2.9 & 10.5 & $\pm$ & 0.2 & 1111.7 & 6.12 & 1108.1 & $\pm$ & 2.3 & 6.03 & $\pm$ & 0.4 \\
629.7 & 3.91 & 627.7 & $\pm$ & 1.4 & 4.14 & $\pm$ & 0.34 & 1150.0 & 2.89 & 1148.6 & $\pm$ & 1.1 & 4.24 & $\pm$ & 0.1 \\
639.7 & 3.88 & 637.8 & $\pm$ & 1.3 & 4.03 & $\pm$ & 0.21 & 1224.8 & 5.86 & 1221.9 & $\pm$ & 1.7 & 5.71 & $\pm$ & 0.1 \\
649.7 & 3.84 & 647.8 & $\pm$ & 1.3 & 3.39 & $\pm$ & 0.21 & 1239.7 & 5.82 & 1237.8 & $\pm$ & 2.6 & 12.2 & $\pm$ & 1.3 \\
659.8 & 3.81 & 657.9 & $\pm$ & 1.3 & 3.78 & $\pm$ & 0.15 & 1244.7 & 5.81 & 1241.9 & $\pm$ & 1.7 & 55.7 & $\pm$ & 0.7 \\
669.8 & 7.88 & 665.8 & $\pm$ & 2.1 & 5.00 & $\pm$ & 0.51 & 1249.7 & 2.79 & 1248.1 & $\pm$ & 1.3 & 19.2 & $\pm$ & 0.9 \\
672.5 & 3.78 & 671.0 & $\pm$ & 1.7 & 138 & $\pm$ & 1.7 & 1254.7 & 5.79 & 1251.4 & $\pm$ & 2.1 & 12.7 & $\pm$ & 0.4 \\
674.8 & 3.77 & 672.8 & $\pm$ & 1.3 & 218 & $\pm$ & 3.6 & 1259.7 & 5.78 & 1256.6 & $\pm$ & 1.8 & 9.34 & $\pm$ & 0.22 \\
677.3 & 3.76 & 674.6 & $\pm$ & 1.9 & 100 & $\pm$ & 2.3 & 1264.8 & 5.77 & 1262.2 & $\pm$ & 2.0 & 8.83 & $\pm$ & 0.36 \\
679.7 & 3.76 & 676.4 & $\pm$ & 1.9 & 16.9 & $\pm$ & 0.6 & 1274.8 & 5.75 & 1271.9 & $\pm$ & 1.7 & 19.5 & $\pm$ & 0.2 \\
689.7 & 7.76 & 685.8 & $\pm$ & 2.1 & 8.72 & $\pm$ & 0.20 & 1279.8 & 2.76 & 1278.3 & $\pm$ & 1.1 & 12.0 & $\pm$ & 0.3 \\
699.7 & 3.70 & 698.0 & $\pm$ & 1.4 & 14.9 & $\pm$ & 0.93 & 1284.7 & 2.75 & 1283.3 & $\pm$ & 1.1 & 7.75 & $\pm$ & 0.10 \\
704.7 & 7.68 & 702.1 & $\pm$ & 3.3 & 26.3 & $\pm$ & 0.83 & 1299.8 & 5.70 & 1296.8 & $\pm$ & 1.7 & 5.19 & $\pm$ & 0.22 \\
709.7 & 7.65 & 707.5 & $\pm$ & 3.7 & 47.0 & $\pm$ & 1.9 & 1339.7 & 2.70 & 1338.5 & $\pm$ & 1.1 & 4.22 & $\pm$ & 0.14 \\
714.8 & 3.66 & 713.1 & $\pm$ & 1.4 & 135 & $\pm$ & 4.2 & 1345.7 & 5.62 & 1343.0 & $\pm$ & 1.6 & 23.1 & $\pm$ & 0.2 \\
717.8 & 3.65 & 716.1 & $\pm$ & 1.4 & 170 & $\pm$ & 0.8 & 1349.7 & 2.69 & 1348.1 & $\pm$ & 1.3 & 5.42 & $\pm$ & 0.17 \\
719.8 & 7.59 & 715.6 & $\pm$ & 2.4 & 178 & $\pm$ & 1.0 & 1354.7 & 2.69 & 1353.3 & $\pm$ & 1.1 & 3.98 & $\pm$ & 0.16 \\
724.8 & 7.57 & 720.3 & $\pm$ & 2.7 & 121 & $\pm$ & 1.2 & 1359.7 & 2.68 & 1358.4 & $\pm$ & 1.0 & 4.33 & $\pm$ & 0.21 \\
729.7 & 3.62 & 727.6 & $\pm$ & 1.5 & 38.6 & $\pm$ & 1.7 & 1400.0 & 5.52 & 1397.2 & $\pm$ & 1.6 & 5.25 & $\pm$ & 0.12 \\
739.7 & 7.48 & 735.3 & $\pm$ & 2.7 & 18.6 & $\pm$ & 1.0 & 1500.0 & 2.57 & 1498.7 & $\pm$ & 1.0 & 5.76 & $\pm$ & 0.08 \\
744.7 & 3.58 & 743.3 & $\pm$ & 1.6 & 12.2 & $\pm$ & 0.8 & 1600.0 & 5.18 & 1597.4 & $\pm$ & 1.5 & 7.94 & $\pm$ & 0.31 \\
749.7 & 3.57 & 747.9 & $\pm$ & 1.2 & 49.6 & $\pm$ & 0.6 & 1700.0 & 5.01 & 1697.5 & $\pm$ & 1.5 & 10.7 & $\pm$ & 0.3 \\
754.7 & 3.56 & 752.9 & $\pm$ & 1.2 & 7.66 & $\pm$ & 0.2 & 1800.0 & 2.32 & 1798.8 & $\pm$ & 1.0 & 11.8 & $\pm$ & 0.2 \\
759.7 & 7.38 & 755.9 & $\pm$ & 2.1 & 7.34 & $\pm$ & 0.3 \\
\end{tabular}
\end{ruledtabular}
\end{table*}
\endgroup
In the table only the statistical uncertainty of the cross section values is quoted. This is obtained simply from the peak integration of the 511\,keV $\gamma$-peak and the background subtraction. Typically the statistical uncertainties are between 0.5\,\% and 5\,\%. Higher statistical uncertainties can be found in the case of the lowest cross sections and for those points where based on the literature data higher cross sections were expected at a resonance but the actual resonance was found at slightly shifted energy (see below).
In order to obtain the total uncertainty of the cross section values, 7.6\,\% systematic uncertainty must be added quadratically to the relative statistical uncertainties. This systematic uncertainty is the quadratic sum of the following components: number of oxygen atoms in the target (6\,\%), $\gamma$-detection efficiency (3\,\%), number of protons hitting the target (3\,\% from charge integration), $^{17}$O enrichment (2\,\%). Uncertainties well below 1\,\% - like the uncertainty of $^{18}$F decay parameters or the measurement of irradiation and counting times - were neglected.
In order to increase the reliability of our experiments, repeated activations were carried out at a few different proton energies using different targets. The results were always in agreement within the statistical uncertainties of the measurements. In table\,\ref{tab:results} either the weighted average of these points are shown or - if targets with different widths were used - the more precise value was adopted.
\section{Discussion}
\label{sec:discussion}
As most of the previous experiments yielded partial cross sections for the various transitions in $^{18}$F measured at a given angle, it is rather difficult to compare the results of the present work with previous experiments. Above 500\,keV proton energy the only total cross section (in the form of an astrophysical S-factor figure) is provided by C. Rolfs \cite{rol73}. Although C. Rolfs studied the reaction in a wide energy range, total S-factor is only provided in the energy regions far from the resonances, i.e. below 450\,keV (outside the energy range of the present work) and above 900\,keV.
Figure\,\ref{fig:results_highE} shows the total cross section determined in the present work and that of C. Rolfs in this high energy range. The points of C. Rolfs are taken from the EXFOR \cite{EXFOR} database where they were obtained by scanning Fig. 17 of \cite{rol73}. As one can see, the data of C. Rolfs are on average a factor of 1.5 higher than the present data although the agreement becomes somewhat better at the highest energies.
In addition to the total cross section data, the partial cross section involving the first excited state to ground state transition in $^{18}$F measured by C. Rolfs is also included in the figure taken from EXFOR. Obviously, this partial cross section is lower than the total one, but it is included in the figure in order to compare the observed resonances in the high energy region. The existence of the narrow resonances observed by C. Rolfs is confirmed by the present work. There is an apparent energy shift between the two datasets, the resonances in the present work are observed at slightly higher energies than in \cite{rol73} as given in EXFOR. However, the numerical values of the resonance energies, as given in \cite{rol73b} are in reasonably good agreement with the present work. The apparent discrepancy as can be seen in the figure can therefore most likely be attributed to the digitization uncertainty of the low resolution Fig. 11 of \cite{rol73}.
\begin{figure*}
\includegraphics[angle=270,width=0.8\textwidth]{fig4_results_highE.eps
\caption{\label{fig:results_highE} Experimental cross section of the $^{17}$O(p,$\gamma$)$^{18}$F reaction in the upper part of the studied energy range. Besides the present work, the total and partial cross sections of C. Rolfs \cite{rol73} are also shown as discussed in the text. The lines through the points are only to guide the eye.}
\end{figure*}
Between 500\,keV and 900\,keV proton energies no total cross section values are available in literature making the comparison of our data with the existing database even more difficult. Figure\,\ref{fig:results_lowE} shows the measured cross section in this energy range. In order to compare at least the energy dependence of the cross section, besides the present data, partial cross sections measured by C. Rolfs \cite{rol73} and A. Kontos \textit{et al.} \cite{kon12} are also included in the figure. As in Fig.\,\ref{fig:results_highE}, the data of C. Rolfs are taken from its Fig.\,11 as compiled in EXFOR. In the case of A. Kontos \textit{et al.}, capture to the first excited state of $^{18}$F measured at 135 degrees is arbitrarily chosen. The energy dependence of the cross section is very similar in the three datasets. The partial cross section of C. Rolfs exceeds the total cross section measured in the present work. This is similar to the observation on the direct capture cross section at higher energies.
A further comparison with literature data can be made at the lowest studied energy of the present work at E$_p$\,=\,500\,keV. Several recent low energy datasets extend up to this energy and some of them quote total cross section (or S-factor) which can be compared with the present work. Table\,\ref{tab:500keV} lists the experimental (or quasi-experimental, see below) cross section values at 500\,keV proton energy. The following literature data were considered: U. Hager \textit{et al.} \cite{hag12} measured the total cross section with the DRAGON recoil separator at E$_{c.m.}$\,=\,470\,keV corresponding to 497.9\,keV proton energy which matches exactly our lowest energy. The value is taken from table\,VI of \cite{hag12}. In-beam $\gamma$-spectroscopy measurement of J.R. Newton \textit{et al.} \cite{new10} provided total cross section at 500\,keV proton energy which again coincides with our data point taking into account the energy uncertainties. The value is taken from Table\,I of \cite{new10}. A. Kontos \textit{et al.} \cite{kon12} do not provide total cross section data directly at E$_p$\,=\,500\,keV. Measured partial cross section around this energy region, however, is available and based on these data the authors provide total S-factor values in tabular form in their Table\,V. Interpolated value for E$_{c.m.}$\,=\,470\,keV (corresponding to E$_p$\,=\,500\,keV) is put into Table\,\ref{tab:500keV} keeping the 12\,\% relative experimental uncertainty.
\begin{figure*}
\includegraphics[angle=270,width=0.8\textwidth]{fig5_results_lowE.eps
\caption{\label{fig:results_lowE} Experimental cross section of the $^{17}$O(p,$\gamma$)$^{18}$F reaction in the lower part of the studied energy range. Besides the present work, the partial cross sections of C. Rolfs \cite{rol73} and A. Kontos \textit{et al.} \cite{kon12} are also shown as discussed in the text. The lines through the points are only to guide the eye.}
\end{figure*}
\begin{table}
\caption{\label{tab:500keV} Experimental cross section of $^{17}$O(p,$\gamma$)$^{18}$F at E$_p$\,=\,500\,keV from the present and previous works.
}
\begin{ruledtabular}
\begin{tabular}{ll}
Reference & Cross section at\\
& E$_p$\,=\,500\,keV [nbarn] \\
\hline
U. Hager \textit{et al.} \cite{hag12} & 585\,$\pm$\,8$_{\rm stat.}$\,$\pm$\,75$_{\rm syst.}$ \\
J.R. Newton \textit{et al.} \cite{new10} & 488\,$\pm$\,49 \\
A. Kontos \textit{et al.} \cite{kon12} & 588\,$\pm$\,71\footnote{Not purely experimental value. See text.} \\
\hline
present work & 592\,$\pm$\,21$_{\rm stat.}$\,$\pm$\,45$_{\rm syst.}$ \\
\end{tabular}
\end{ruledtabular}
\end{table}
The result of the present work at E$_p$\,=\,500\,keV is in good agreement with U. Hager \textit{et al.} \cite{hag12} and A. Kontos \textit{et al.} \cite{kon12}. The cross section of J.R. Newton \textit{et al.} \cite{new10}, on the other hand, is almost 20\,\% lower than - and therefore barely consistent with - the other three values.
\section{R-matrix analysis}
\label{sec:rmat}
AZURE2, a multichannel and multilevel R-matrix code \cite{azuma10}, was used to simultaneously fit the total cross section, measured by J. R. Newton \textit{et al.} \cite{new10}, U. Hager \textit{et al.} \cite{hag12}, A. Di Leva \textit{et al.} \cite{dil14}, M. Q. Buckner \textit{et al.} \cite{buc15} and by the present work, as well as the primary transitions, measured by A. Kontos \textit{et al.} \cite{kon12}, of the $^{17}$O(p,$\gamma$)$^{18}$F reaction. The fit using these data will be referred to as ``our'' fit in the following.
\begin{figure}
\includegraphics[width=\columnwidth]{fig6_Rmat_KontosGyurkyTotalS.eps}
\caption{\label{fig:Rmat_KontosGyurkyTotalS} Total $S$ factor obtained from an R-matrix fit made by A. Kontos \textit{et al.} \cite{kon12} is compared with our experimental data of the $^{17}$O(p,$\gamma$)$^{18}$F reaction.}
\end{figure}
First, we have compared the total $S$ factor obtained from an R-matrix fit made by A. Kontos \textit{et al.} \cite{kon12} with our experimental data. One can see in Fig.\,\ref{fig:Rmat_KontosGyurkyTotalS} that there is a good agreement between our data and the calculated one of Kontos, although some narrow resonances are omitted from their plot. Values of Kontos are obtained from Fig.\,9 of \cite{kon12} by figure digitization using the software PlotDigitizer 2.6.8 \cite{PlotDig}.
In our R-matrix fit for the determination of direct capture, because of the nice agreement with the results of Kontos, we used same {\it asymptotic normalization coefficient} (ANC) values and high energy background poles as they used. Table\,\ref{tab:rmatANC} lists these fixed ANC values. In addition, 15 MeV as the excitation energy of the background poles was selected. There are no proton scattering data to provide restrictions for the proton partial widths of the poles, so they were fixed at $\Gamma_p = 6$ MeV, close to the Wigner limit. The R-matrix radius was taken as $r_c = r_0 \times ( A^{1/3}_t+A^{1/3}_p ) = 4.46$ fm, with $r_0 = 1.25$ fm. More details about the selected values are in \cite{kon12}.
\begin{table}
\caption{\label{tab:rmatANC} Fixed ANCs based on \cite{kon12}.}
\begin{ruledtabular}
\begin{tabular}{ c c c }
Energy (keV) & $\ell$ & ANC (fm$^{-1/2}$)\\[0.05cm]
\hline\\[-0.2cm]
~937 & 0 & 6.1 \\
~937 & 2 & 1.2 \\
1121 & 2 & 2.7 \\
2523 & 0 & 1.4 \\
3062 & 0 & 4.5 \\
3062 & 2 & 1.0 \\
3839 & 0 & 4.6 \\
3839 & 2 & 0.6 \\
4115 & 0 & 2.5 \\
4115 & 2 & 1.0 \\
4652 & 2 & 1.3 \\
4964 & 0 & 3.2 \\
4964 & 2 & 0.7 \\
\end{tabular}
\end{ruledtabular}
\end{table}
Our R-matrix analysis used the data set of table\,\ref{tab:results}.
As the quoted effective energies were used, no target effect was taken into account. No normalization of datasets was applied and for physical parameters the Brune parameterization \cite{bru02} was used. The full parameter list of our R-matrix fit is provided as Supplemental Material \cite{supmat}. It contains the used datasets and the AZURE2 input file with all parameters.
The estimated dependence of the R-matrix extrapolation on the choice of the channel radius, the position of the background poles and ANC values are $\sim$\,4\,\%, $\sim$\,7\,\% and $\sim$\,15\,\%, respectively. These values are estimated from the manual variation of the above parameters around their fixed values. The uncertainty of the extrapolation of the total $S$ factor to zero energy is $\sim$\,20\,\%.
Table\,\ref{tab:rmatS0} lists the calculated contributions of all the measured transitions to the total $S$ factor at zero energy. The second column is calculated by Kontos \textit{et al.} \cite{kon12}, the third one by Di Leva \textit{et al.} \cite{dil14} and the last one comes from our fit. The error of our data is $\sim$\,15\,\% because of the uncertainty of the choice of ANC. The uncertainties are statistical only. The total $S$(0) value of Kontos and Di Leva are $5.4 \pm \text{(th.)}1.0 \pm \text{(exp.)}0.6$ keV\,b and $5.0 \pm 0.3$ keV\,b, respectively. Our total $S$ factor value at zero energy is $4.7 \pm 1.0$ keV\,b where the error is statistical only.
\begin{table*}
\caption{\label{tab:rmatS0} Calculated $S$(0) values for each $\gamma$-ray transitions measured by Kontos \textit{et al.} \cite{kon12}.}
\begin{ruledtabular}
\begin{tabular}{ c c c c}
Transition (keV) & $S$(0)$^{\text{Kontos \cite{kon12}}}$ (keV\,b) & $S$(0)$^{\text{Di Leva \cite{dil14}}}$ (keV\,b) & $S$(0)$^{\text{Present}}$ (keV\,b)\footnote{The $S$(0) of the present work was obtained by using for the fit simultaneously the partial cross sections from the literature (see text) and the total cross section presented in this paper.}\\[0.05cm]
\hline\\[-0.2cm]
R/DC $\rightarrow 937~\,$ & $1.7\pm 0.3$ & $1.48\pm 0.08$ & $1.73 \pm 0.26$\\
R/DC $\rightarrow 1121$ & $0.66\pm 0.13$ & $0.47\pm 0.05$ & $0.65 \pm 0.10$ \\
R/DC $\rightarrow 1700$ & $0.013\pm 0.002$ & & $0.013 \pm 0.002$ \\
R/DC $\rightarrow 2523$ & $0.17\pm 0.03$ & $0.12\pm 0.03$ & $0.15 \pm 0.02$ \\
R/DC $\rightarrow 3062$ & $0.66\pm 0.1~\,$ & $0.59\pm 0.03$ & $0.45 \pm 0.07$ \\
R/DC $\rightarrow 3791$ & $0.032\pm 0.005$ & $0.20\pm 0.05$ & $0.030 \pm 0.005$ \\
R/DC $\rightarrow 3839$ & $0.93\pm 0.14$ & $0.92\pm 0.04$ & $0.66 \pm 0.10$ \\
R/DC $\rightarrow 4115$ & $0.55\pm 0.08$ & $0.50\pm 0.03$ & $0.51 \pm 0.08$ \\
R/DC $\rightarrow 4652$ & $0.21\pm 0.03$ & $0.10\pm 0.03$ & $0.19 \pm 0.03$ \\
R/DC $\rightarrow 4964$ & $0.49\pm 0.07$ & $0.43\pm 0.03$ & $0.35 \pm 0.05$ \\
\end{tabular}
\end{ruledtabular}
\end{table*}
Fig.\,\ref{fig:Rmat_FitGyurkyTotalS} shows the total $S$ factor obtained from our R-matrix fit (continuous line) as well as experimental datasets of J. R. Newton \textit{et al.} \cite{new10}, U. Hager \textit{et al.} \cite{hag12}, A. Di Leva \textit{et al.} \cite{dil14}, M. Q. Buckner \textit{et al.} \cite{buc15} and present work. Narrow resonances are also included. The $\chi^2$ value of our dataset is $7.3$ without any normalization of datasets. The contribution of the direct capture to the total $S$ factor at zero energy in our fit is $S_{\text{DC}}=4.3\pm1.0$ keV\,b, where the uncertainty is statistical only. Fig.\,\ref{fig:Rmat_FitS0} shows the low energy total $S$ factor obtained from our R-matrix fit with the above experimental datasets.
\begin{figure}
\includegraphics[width=\columnwidth]{fig7_Rmat_FitGyurkyTotalS.eps}
\caption{\label{fig:Rmat_FitGyurkyTotalS} Total $S$ factor obtained from our R-matrix fit (continuous line) is compared with experimental data of the $^{17}$O(p,$\gamma$)$^{18}$F reaction. The dashed line is the contribution of the direct capture to the total $S$ factor (background poles included).}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{fig8_Rmat_FitS0.eps}
\caption{\label{fig:Rmat_FitS0} Low energy total $S$ factor obtained from our R-matrix fit is compared with experimental data of the $^{17}$O(p,$\gamma$)$^{18}$F reaction. (A.: activation; P.: primary transitions; S.: secondary transitions)}
\end{figure}
\section{Summary and conclusions}
\label{sec:summary}
In the present work the total cross section of the $^{17}$O(p,$\gamma$)$^{18}$F reaction was measured with the activation method in a wide energy range for the first time with a total uncertainty of about 10\,\%. Since this method provides cross sections which are in several aspects independent from the ones obtained with in-beam $\gamma$-spectroscopy and some systematic errors are not present, our results can be used to check the validity of previous data.
In general, our data is in good qualitative agreement with the structure of the excitation function of previous works. The possibility of the comparison of total cross sections is limited owing to the scarcity of total cross section data in the literature in the studied energy range. At energies above 900\,keV our results are on average a factor of 1.5 lower than that of C. Rolfs \cite{rol73}. Similar deviation is found at lower energies where the partial cross sections of C. Rolfs exceed substantially our total cross section. Too high values of C. Rolfs were also pointed out earlier by A. Kontos \textit{et al.} \cite{kon12} in the case of individual transitions. This observation is confirmed by the present work.
A direct comparison of our cross section data with the literature was carried out also at a single proton energy of 500\,keV. It is found that our value agrees well with that of U. Hager \textit{et al.} \cite{hag12} and A. Kontos \textit{et al.} \cite{kon12}, while the result of J.R. Newton \textit{et al.} \cite{new10} is about 20\,\% (two standard deviations) lower.
An R-matrix analysis with the AZURE2 code was performed to check the conformity of our measured total cross section dataset and to extrapolate the astrophysical $S$ factor to lower energies. In this analysis all primary transitions observed from Ref. \cite{kon12} were simultaneously fitted with some total cross section datasets, included the present one. The resulting total $S$ factor is in good agreement with previous measurements and calculations within the experimental uncertainties.
Our total cross section data can be used to constrain any future theoretical description of the $^{17}$O(p,$\gamma$)$^{18}$F reaction.
\begin{acknowledgments}
We thank A. Formicola and the LUNA collaboration for giving us access to the Ta$_2$O$_5$ target preparation device at LNGS, Italy. We also thank R. J. deBoer for all of his helpful advices regarding the use of the R-matrix code, AZURE2. This work was supported by the SROP-4.2.2.B-15/1/KONV-2015-0001 project, by the European Union, co-financed by the European Social Fund and by OTKA grants No. K108459, K120666 and K112962
\end{acknowledgments}
|
\section{Introduction}
A {\it $(\gamma,n)$-gonal pair} is a pair $(S,f)$, where $S$ is a closed Riemann surface and $f:S \to R$ is a degree $n \geq 2$ holomorphic branched cover onto a closed Riemann surface $R$ of genus $\gamma$; in this case, we say that $S$ is {\it $(\gamma,n)$-gonal}. A $(0,n)$-gonal pair is usually called
an {\it $n$-gonal pair}. In the case that the branched cover $f$ is {\it simple}, that is, each branch value of $f$ has exactly $n-1$ preimages, we talk of
a {\it simple $(\gamma,n)$-gonal pair}.
There are different notion of equivalences between $(\gamma,n)$-gonal pairs. We are interested in three of them. If
$(S_{1},f_{1}:S_{1} \to R_{1})$ and $(S_{2},f_{2}:S_{2} \to R_{2})$ are $(\gamma,n)$-gonal pairs, then we say that they are: (i) {\it topologically equivalent} if there are orientation preserving homeomorphisms $\phi:S_{1} \to S_{2}$ and $\psi:R_{1} \to R_{2}$ so that $f_{2} \circ \phi =\psi \circ f_{1}$, (ii) {\it twisted isomorphic} if we may assume $\phi$ and $\psi$ to be isomorphisms, and (iii) {\it isomorphic} if $R_{1}=R_{2}$, $\phi$ is an isomorphism and $\psi$ is the identity map.
Associated to a $(\gamma,n)$-gonal pair $(S,f)$ are the Hurwitz spaces ${\mathcal H}_{0}(S,f)$ and ${\mathcal H}(S,f)$, consisting respectively of the isomorphic classes and of the twisted isomorphic classes of the $(\gamma,n)$-gonal pairs which are topologically equivalent to $(S,f)$. If $(S',f')$ is topologically equivalent to $(S,f)$, then there are natural bijections between (i) ${\mathcal H}_{0}(S,f)$ and ${\mathcal H}_{0}(S',f')$ and (ii) ${\mathcal H}(S,f)$ and ${\mathcal H}(S',f')$.
In 1873, Clebsch \cite{Clebsch} proved that any two simple $n$-gonal pairs are topologically equivalent, in particular, for $(S,f)$ a simple $n$-gonal pair, the space ${\mathcal H}(S,f)$ is the classical Hurwitz space parametrizing twisted isomorphic classes of degree $n$ simple covers of $\widehat{\mathbb C}$.
In 1891, Hurwitz \cite{Hurwitz} proved that ${\mathcal H}(S,f)$ has the structure of a complex manifold of dimension $r-3$, where $S$ has genus $g$ and $r=2g+2n-2 \geq 3$. So, by Clebsch's result, ${\mathcal H}(S,f)$ is irreducible. As every Riemann surface of genus $g$ admits a degree $n \geq g+1$ simple branched cover (by Riemann-Roch's theorem), Severi \cite{Severi} used the above facts to prove that the moduli space ${\mathcal M}_{g}$, of closed Riemann surfaces of genus $g$, is irreducible. For the case of not necessarily simple $n$-gonal pairs $(S,f)$, Fried \cite{Fried} proved that ${\mathcal H}(S,f)$ also has the structure of a connected complex manifold of dimension $r-3$.
In the general case of an $n$-gonal pair $(S,f)$ (i.e., not necessarily simple), in \cite{Natanzon1,Natanzon2} Natanzon constructed an uniformizing pair $(\Gamma,G)$, where $G$ is a Fuchsian group isomorphic to a free group of rank $|B_{f}|-1$, where
$B_{f}$ is the set of branched values of $f$ (assuming its cardinality is at least $3$), and $\Gamma$ is a suitable finite index subgroup of it (of index equal to the degree of $f$). Using such type of uniformizations, he was able to obtain that ${\mathcal H}(S,f)$ is homeomorphic to the a quotient ${\mathbb R}^{m}/M$, where $M$ is a certain discrete group (in this case, ${\mathbb R}^{m}$ is the real structure of the Teichm\"uller space of the Riemann sphere punctured at $|B_{f}|$ points).
Let us now consider a $(\gamma,n)$-gonal pair $(S,f)$. In Section \ref{Sec:uniformizationsofpairs} we define for such a pair to be of {\it hyperbolic type} and, in such a case, we associate to it a pair $(\Gamma,G)$, called a {\it uniformization of $(S,f)$}, where $G$ is a Fuchsian group acting on the unit disc ${\mathbb D}$ containing $\Gamma$ as an index $n$ subgroup, so that there are isomorphisms $\phi:S \to {\mathbb D}/\Gamma$ and $\psi: R \to {\mathbb D}/G$ with $\pi=\psi f \phi^{-1}$ being a branched covering induced by the inclusion of $\Gamma$ inside $G$. The uniformizing pair is uniquely determined up to conjugation by holomorphic automorphisms of ${\mathbb D}$. Using such an uniformization pair, it is possible to obtain the following fact that generalizes Natanzon's above.
\vspace{0.3cm}
\noindent
\begin{theo}\label{main1}
Let $(S,f)$ be a $(\gamma,n)$-gonal pair of hyperbolic type and $(\Gamma,G)$ be a unifomizing pair of it. Then
(i) ${\mathcal H}_{0}(S,f)$ is isomorphic to the Teichm\"uller space $T({\mathbb D},G)$ of the Riemann orbifold ${\mathbb D}/G$, this being a finite dimensional simply-connected complex manifold,
and (ii) ${\mathcal H}(S,f)$ is a complex orbifold, being a quotient $T({\mathbb D},G)/M(\Gamma,G)$, where $M(\Gamma,G)$ is a suitable finite index subgroup of its holomorphic automorphisms.
\end{theo}
\vspace{0.3cm}
In \cite{BCI,BC,GWW} there is some study of $(\gamma,n)$-gonal pairs and their groups of automorphisms when $n$ is prime integer and the $n$-gonal map is a regular branched cover.
In the literature, compactifications of Hurwitz spaces of simple $n$-gonal pairs has been obtained by adding the so called stable $n$-gonal pairs (also called admissible ones by some authors; see for instance \cite{H-M}), which are certain geometric degenerations of simple $n$-gonal pairs. Next, we proceed to recall such king of degenerations in the more general case (i.e., it might be either $\gamma \geq 0$ or non-simple situation).
Let $(S,f:S \to R)$ be a fixed $(\gamma,n)$-gonal pair of hyperbolic type. If $\gamma=0$, then we also assume that its branch value set $B_{f}$ has cardinality at least $4$; otherwise there is no possible degeneration to be done (if $B_{f}$ has cardinality $3$, then $(S,f)$ is a Belyi pair which is definable over $\overline{\mathbb Q}$ as a consequence of Belyi's theorem \cite{Belyi}). Let us consider a collection ${\mathcal F}$ of pairwise disjoint simple loops $\gamma_{1}, \ldots, \gamma_{s} \subset R-B_{f}$ so that the Euler characteristic of each connected component of $R-(B_{f} \cup \gamma_{1}\cup \cdots \cup \gamma_{s})$ is negative and none of the components of $R-(\gamma_{1} \cup \cdots \cup \gamma_{s})$ is a disc with only two branch values, both with branch order equal to $2$. Let ${\mathcal G} \subset S$ be the collection of (necessarily simple) loops obtained by lifting those in ${\mathcal F}$ by $f$. Next, we proceed to identify all points belonging to the same loop in ${\mathcal G}$ to obtain a stable surface $S^{*}$ of the same genus as $S$. Similarly, by doing the same procedure to the loops of ${\mathcal F}$, we obtain an stable genus $\gamma$ orbifold $R^{*}$. The map $f$ induces a continuous map $f^{*}:S^{*} \to R^{*}$ of degree $n$. We call such a pair $(S^{*},f^{*})$ a {\it topological stable $(\gamma,n)$-gonal pair}. Now, if we provide analytically finite Riemann orbifold structures to each of the components of $R^{*}$ (that is, to the complement of its nodes), then we may lift these Riemann orbifold structures under $f^{*}$ to obtain an analytically finite Riemann orbifold structure on each component of $S^{*}$ minus their nodes. In this way, $S^{*}$ and $R^{*}$ will now carry stable Riemann orbifold structures so that $f^{*}$ still continuous of degree $n$ and its restriction to each connected component of $S^{*}$ minus its nodes is meromorphic. The resulting pairs will be called {\it stable $(\gamma,n)$-gonal pairs modelled by $(S^{*},f^{*})$}. Let us observe that the stable Riemann orbifold structures on $(S^{*},f^{*})$ depends on the spaces of orbifold structures on each of the connected components of $R^{*}$; in fact, the product space of the Teichm\"uller spaces of these components provide a parametrization of the space of stable Riemann orbifold structures of $(S^{*},f^{*})$. In particular, if each of components of $R^{*}$ are spheres with only $3$ marked points, then the structure is unique.
In Section \ref{Sec:nodedpairs} we provide uniformizations of stable $(\gamma,n)$-gonal pairs using pairs of certain noded Fuchsian groups. This uniformization permits to obtain the following fact.
\vspace{0.3cm}
\noindent
\begin{theo}\label{main2}
Let $(S,f)$ be a $(\gamma,n)$-gonal pair of hyperbolic type and let $(\Gamma,G)$ be an uniformizing pair of it. Then
\begin{enumerate}
\item[(i)] The augmented Teichm\"uller psce $NT({\mathbb D},G)$ provides a model for the partial closure of ${\mathcal H}_{0}(S,f)$
obtained by adding the corresponding equivalence classes of stable $(\gamma,n)$-gonal pairs.
\item[(ii)] The space parametrizing twisted isomorphic classes of stable $(\gamma,n)$-gonal pairs can be identified with the quotient $NT({\mathbb D},G)/M(\Gamma,G)$, which has the structure of a compact complex orbifold.
\end{enumerate}
\end{theo}
\vspace{0.3cm}
The above provides a Kleinian groups description being in a parallel point of view as the description provided in \cite{Co-Se} for the case $\gamma=0$.
An interesting situation is provided for genus zero $n$-gonal pairs as the above provides a description of degenerations of rational maps in terms of Kleinian groups which is somehow related to part of the work done in \cite{Arfeux}.
\vspace{0.3cm}
\noindent
\begin{coro}
Let $R \in {\mathbb C}(z)$ be a rational map of degree $d \geq 2$ and let $B=\{p_{1},\ldots,p_{r}\}$ be its locus of branch values. For each $p_{j}$ let $n_{j}$ be the minimum common multiple of the local degrees of $R$ at its preimages. Assume that $n_{1}^{-1}+\cdots+n_{r}^{-1}<r-2$.
Let $G$ be a Fuchsian group, acting on the unit disc ${\mathbb D}$, uniformizing the orbifold of genus zero whose conical points set is $B$ and the conical order of $p_{j}$ is $n_{j}$. Then the space of isomorphic classes of rational maps topologically equivalent to $R$ is isomorphic to the Teichm\"uller space of $G$, this being a simply-connected complex manifold of dimension $r-3$. Its partial closure obtained by adding isomorphic classes of geometrical degenerations of it is isomorphic to the augmented Teichm\"uller space of $G$. Similarly, the corresponding space of twisted isomorphic classes is a complex orbifold of dimension $r-3$ and its closure obtained by adding the twisted classes of its degenerations is isomorphic to a compact complex orbifold of same dimension.
\end{coro}
\vspace{0.3cm}
This paper is organized as follows. In Section \ref{Sec:prelim1} we review some definitions and basic facts on Kleinian groups, in particular of noded Fuchsian groups and some of its properties previously obtained in \cite{Hidalgo:nodedSchottky, Hidalgo:nodedFuchsian}. In Section \ref{Sec:prelim2} we recall the definition and some facts on the quasiconformal deformation of Kleinian groups, in particular, the Teichm\"uller space of a finitely generated Fuchsian group. In Section \ref{Sec:NodedBeltrami} we recall the concept of noded Beltrami coefficients of Kleinian groups from \cite{Hidalgo-Vasiliev} which permits to construct a partial closure of the quasiconformal deformation space of a Kleinian group. We will be mainly interested, for the purpose of this article, on the case of Fuchsian groups, but we provide the general point of view. In the final two sections we use the previous facts to describe, in terms of Fuchsian groups and noded Fuchsian groups, the unifomizations of $(\gamma,n)$-gonal and stable $(\gamma,n)$-gonal pairs, respectively.
\section{Some preliminaries on Kleinian groups}\label{Sec:prelim1}
\subsection{Analytically finite Riemann orbifolds}
An {\it analytically finite Riemann orbifold} is given by a closed Riemann surface $S$ of genus $g \geq 0$ (the underlying Riemann surface structure of the Riemann orbifold) together with a finite collection of conical points $x_{1},...,x_{r} \in S$ of orders $2 \leq m_{1} \leq m_{2} \leq \cdots \leq m_{r} \leq \infty$, respectively. If $x_{j}$ has cone order $\infty$, then this means that it represents a puncture of the orbifold. The {\it signature} of the orbifold is defined by the tuple $(g;m_{1},...,m_{r})$ and it is called {\it hyperbolic} if $2g-2+\sum_{i=1}^{r}(1-m_{i}^{-1})>0$.
If (i) $r=3$ and $g= 0$, then the orbifold is called a {\it triangular orbifold}, (ii) if $r=0$, then the orbifold is a closed Riemann surface, and (iii) if $r>0$ and $m_{j}=\infty$ for all $j$, then it is an analytically finite punctured Riemann surface (this is $S$ minus all the points $x_{j}$).
\subsection{Noded Riemann surfaces/orbifolds}
Let us consider a countable collection $\{{\mathcal O}_{j}\}_{j \in J}$ of Riemann orbifolds, that is, each ${\mathcal O}_{j}$ consists of a Riemann surface $S_{j}$ together a discrete collection of cone points $B_{j}=\{p_{ji}\} \subset S_{j}$ and integer values $n_{ji} \geq 2$ (the cone orders of the points $p_{ji}$). Let us consider a discrete collection of points ${\mathcal E} \subset \bigcup_{j \in J} (S_{j}-B_{j})$ and an order two bijective map $T:{\mathcal E} \to {\mathcal E}$.
We proceed to identify the points $q$ and $T(q)$, for each $q \in {\mathcal E}$, to obtain a space $X$; called a {\it noded Riemann orbifold}. The points obtained by the identification of the points $q$ and $T(q)$, for $q \in {\mathcal E}$, are called (i) {\it nodes} of $X$ if $T(q)=q$ and (ii) {\it phantom nodes} if $T(q)=q$. We denote the set of nodes and phantom nodes of $X$ by $N(X)$. Observe that
the points in $N(X)$ correspond to punctures on each connected component of $X-N(X)$ and that these components are given by $S_{j}-({\mathcal E} \cap S_{j})$. The set of cone points of $X$ is given by $B(X):=\bigcup_{j} B_{j}$.
In the case that every $B_{j}=\emptyset$ (that is, ${\mathcal O}_{j}$ is just a Riemann surface), we say that $X$ is a {\it noded Riemann surface}.
An isomorphism between noded Riemann orbifolds is a homeomorphism that send cone points to cone points (respecting their orders) which, restricted to the complement of the nodes, is analytic. If there is an isomorphism between two noded Riemann orbifolds/surfaces,
then we say that they are isomorphic or conformally equivalent. This, in particular, permits to talk on automorphisms of noded Riemann orbifolds/surfaces.
A noded Riemann surface which is homeomorphic to the space obtained by pinching a non-trivial simple loop on a torus is called a {\it stable Riemann surface of genus one}. A noded Riemann surface which is homeomorphic to the space obtained by the process of pinching a family (necessarily finite) of
disjoint simple closed hyperbolic geodesics on a closed Riemann surface of genus $g \geq 2$ is called a {\it stable Riemann surface of genus $g$}.
Isomorphic classes of stable Riemann surfaces are the extra points Mumford needed to add to the moduli space ${\mathcal M}_{g}$ of closed Riemann surfaces of genus $g$ to provide the Deligne-Mumford's compactification $\overline{\mathcal M}_{g}$.
Let us observe that the space obtained as the quotient $X/H$, where $X$ is a noded Riemann surface and $H$ is a (finite) group of automorphisms of $X$
is an example of a noded Riemann orbifold. In the case that $X$ is a stable Riemann surface, then $X/H$ is also called a {\it stable Riemann orbifold}.
\subsection{Kleinian groups}
A {\it Kleinian group} is just a discrete subgroup $G$ of ${\rm PSL}_{2}({\mathbb C})$ (seen as the group of M\"obius transformations acting on $\widehat{\mathbb C}$). Generalities on Kleinian/Fuchsian groups can be found, for instance, in the books \cite{Beardon,M1}. In this section we recall some of the basics we will need in the rest of this paper.
\subsubsection{The region of discontinuity}
The {\it region of discontinuity} of a Kleinian group $G$ is the (might be empty) open set $\Omega(G)$ of points over which $G$ acts properly discontinuous; the complement $\Lambda(G)=\widehat{\mathbb C}-\Omega(G)$ is its {\it limit set}. If $\Omega(G) \neq \emptyset$, then the quotient space $\Omega(G)/G$ is a union of Riemann orbifolds. By Ahlfor's finiteness theorem \cite{Ahlfors:finitud}, if $G$ is finitely generated, then such a quotient consists of a finite number of analytically finite Riemann orbifolds.
Let $\delta \in G$ be a loxodromic transformation. We say that $\delta$ is {\it primitive} if it is not a nontrivial positive power of another loxodromic transformation in $G$. We say that $\delta$ is {\it simple loxodromic} if there is a simple arc on $\Omega(G)$ which is invariant under $\delta$, we call it an {\it axis} of $\delta$, whose projection on $\Omega(G)/G$ is a simple loop or a simple arc connecting two conical points of order $2$.
\subsubsection{The extended region of discontinuity}
Let $G$ be a finitely Kleinian group with non-empty region of discontinuity.
A parabolic transformation $\eta \in G$, with fixed point $p$, is called {\it double-cusped}, if (i) any parabolic element of $G$ commuting with $\eta$ belongs to the cyclic group $\langle \eta \rangle$, and (ii) there are two tangent open discs at $p$ in $\Omega(G)$ whose union is invariant under the stabilizer of $p$ in $G$.
The {\it extended region of discontinuity} of $G$ is defined as $\Omega(G)^{ext}=\Omega(G) \cup P(G)$, where $P(G)$ is the set of fixed points of
the double-parabolic elements of $G$.
On $\Omega(G)^{ext}$ we consider its cuspidal topology; the topology
generated by the usual open sets in $\Omega(G)$ and the sets of the form
$D_{1} \cup D_{2} \cup \{p\}$, where $p \in P(G)$, and $D_{1}, D_{2} \subset
\Omega(G)$ are round discs tangent at $p$. Observe that if $G$ has no parabolic transformations,
then the extended region of discontinuity coincides with its region of
discontinuity, that is, $\Omega(G)=\Omega(G)^{ext}$.
In the cuspidal topology, the group $G$ acts as a group of homeomorphisms on $\Omega(G)^{ext}$,
keeping invariant each $\Omega(G)$ and $P(G)$, and its restriction to
$\Omega(G)$ being by holomorphic automorphisms.
By Selberg's lemma, $G$ contains a torsion-free finite index normal subgroup $K$. The finite index condition asserts that $\Omega(K)=\Omega(G)$ and $P(K)=P(G)$, in particular, $\Omega^{ext}(K)=\Omega^{ext}(G)$. The quotient $\Omega(G)^{ext}/K$ is a noded Riemann surface whose nodes corresponds one-to-one to the $K-$equivalence classes of the parabolic fixed points in $P(G)$ (those having an involution in the stabilizer produce the so called phantom nodes). By Ahlfors' finiteness theorem,
the number of components of $\Omega(G)/K=\Omega(G)^{ext}/K -N(\Omega(G)^{ext}/K)$
is finite, each one an analytically finite Riemann surface. It follows that
$\Omega(G)^{ext}/G$, with the quotient cuspidal topology, is a noded Riemann orbifold, and it contains the orbifold $\Omega(G)/G$ as a dense open subset. The points in $P(G)/G$ are the nodes (and phantom nodes) of $\Omega(G)^{ext}/G$.
\subsection{Fuchsian groups}
Let ${\mathbb D}$ be the unit disc in the complex plane and let ${\rm Aut}({\mathbb D})$ be its group of conformal automorphisms (this being a subgroup of ${\rm PSL}_{2}({\mathbb C})$).
A finitely generated Kleinian group $F$ being a subgroup of ${\rm Aut}({\mathbb D})$ is called a {\it Fuchsian group}. The Fuchsian group $F$ is called of the {\it first kind} if its limit set is all of the unit circle (so is region of discontinuity consists of two discs); otherwise, it is called of the {\it second kind} (so its region of discontinuity is connected).
As a consequence of Selberg's lemma \cite{Beardon}, $F$ has a torsion-free normal subgroup $K$ of finite index;
this is again a finitely generated Fuchsian group of the same kind as $F$.
If $F$ is of the second kind without parabolic elements, then $K$ is a Schottky group of some finite rank $g \geq 0$ and if $F$ is of the first kind without parabolic elements, then $K$ is a co-compact Fuchsian group uniformizing a closed Riemann surface of some genus $g \geq 2$ (in this case we say that $K$ is a $\pi^{g}$ group).
By classical work of Fricke and Klein, if the Fuchsian group $F$ is of the first kind and without parabolic elements, then it has a presentation of the form
$$F=\langle \alpha_{1},\ldots,\alpha_{g},\beta_{1},\ldots,\beta_{g},\delta_{1},\ldots,\delta_{r}:\prod_{j=1}^{g}[\alpha_{j},\beta_{j}] \prod_{i=1}^{r}\delta_{i}=\delta_{1}^{m_{1}}=\cdots=\delta_{r}^{m_{r}}=1\rangle,$$
where $[\alpha_{j},\beta_{j}]=\alpha_{j}\beta_{j}\alpha_{j}^{-1}\beta_{j}^{-1}$, $m_{i} \geq 2$ are integers so that
$2g-2+\sum_{i=1}^{r}(1-m_{i}^{-1})>0$. In this case, ${\mathbb D}/F$ is a compact hyperbolic Riemann orbifold of signature
$(g;m_{1},\ldots,m_{r})$; which is also called the signature of $F$. If $r=0$, then ${\mathbb D}/F$ is a closed Riemann surface of genus $g$ and its signature is denoted by $(g;-)$.
As a consequence of the uniformization theorem, every compact hyperbolic orbifold ${\mathcal O}$ is isomorphic to ${\mathbb D}/F$ for some Fuchsian group $F$.
\subsection{Noded Fuchsian groups}
A Kleinian group is called a {\it noded Fuchsian group} \cite{Hidalgo:nodedFuchsian}
if it is a geometrically finite and isomorphic to some Fuchsian group of the first kind.
Torsion-free noded Fuchsian groups come in two flavors \cite{Hidalgo:nodedSchottky,Hidalgo:nodedFuchsian}: (i) {\it noded Schottky groups of rank $g\geq 0$} (isomorphic to free groups of rank $g\geq 0$) and (ii) {\it noded $\pi^{g}$ groups} (isomorphic to the fundamental group of a closed orientable surface of genus $g \geq 2$). So, as a consequence of Selberg's lemma, every noded Fuchsian group has a finite index normal subgroup being either a noded Schottky group or a noded $\pi^{g}$ group.
\vspace{0.3cm}
\noindent
\begin{theo}[\cite{Hidalgo:nodedFuchsian}]
Noded Fuchsian groups have non-empty region of discontinuity.
\end{theo}
\vspace{0.3cm}
If $G$ is a noded Fuchsian group, then, as $G$ is isomorphic to some Fuchsian group, it cannot have rank two parabolic subgroups and all parabolics are double-cusped. As already noted above, $G$ contains a torsion-free finite index normal subgroup $K$, this being either a noded Schottky group or a noded $\pi^{g}$ group. The finite index condition asserts that $\Omega(K)=\Omega(G)$ and $P(K)=P(G)$, in particular, $\Omega^{ext}(K)=\Omega^{ext}(G)$.
Below, in Sections \ref{Sec:nrt} and \ref{Sec:nsut}, we will see that noded Riemann surface $\Omega(G)^{ext}/K$ consists of either: (i) two stable Riemann surfaces of genus $g$, if $K$ is noded $\pi^{g}$-group, or (ii) one stable Riemann surface of genus $g$, if $K$ is a noded Schottky group of rank $g$). It follows that $\Omega(G)^{ext}/G$, with the quotient cuspidal topology, is a stable Riemann orbifold, and it contains the Riemann orbifold $\Omega(G)/G$ as a dense open subset. The points in $P(G)/G$ are the nodes of $\Omega(G)^{ext}/G$ (and there are no phantom nodes).
As an example, the cyclic group $G=\langle \gamma(z)=z+1 \rangle$ is a noded Fuchsian group (in fact a noded Schottky group of rank one). In this case, $\Omega(G)^{ext}=\widehat{\mathbb C}$, $\Omega(G)={\mathbb C}$ and $\Omega(G)^{ext}/G$ is a stable Riemann surface of genus one (its node being the projection of $\infty$).
\subsection{Noded retrosection theorem}\label{Sec:nrt}
Koebe's retrosection theorem \cite{Koebe} asserts that every closed Riemann surface of genus $g$ can be uniformized by a Schottky group of rank $g$.
A retrosection theorem with nodes hold and it can be stated as follows.
\vspace{0.3cm}
\noindent
\begin{theo}[Noded retrosection theorem \cite{Hidalgo:nodedSchottky,Hidalgo:nodedFuchsian}]
\mbox{}
\begin{enumerate}
\item If $G$ is a noded Schottky group of rank $g$, then $\Omega(G)^{ext}$ is connected and
$\Omega(G)^{ext}/G$ is a stable Riemann surface of genus $g$.
\item If $S$ is a stable Riemann surface of
genus $g$, then there is a noded Schottky group $G$ of
genus $g$ such that $\Omega(G)^{ext}/G$ is conformally equivalent to $S$.
\end{enumerate}
\end{theo}
\vspace{0.3cm}
\noindent
\begin{rema}
The region of discontinuity of a Schottky group is always connected, but that of a noded Schottky group is not in general; connectivity only holds for its extended region of discontinuity.
\end{rema}
\vspace{0.3cm}
\noindent
\begin{coro}
If $G$ is a noded Fuchsian group containing a noded Schottky group as a finite index normal subgroup, then $\Omega(G)^{ext}$ is connected and $\Omega(G)^{ext}/G$ consists of an stable Riemann orbifold.
\end{coro}
\vspace{0.3cm}
\subsection{Simultaneous uniformization theorem with nodes}\label{Sec:nsut}
If $G$ is a torsion free purely loxodromic quasifuchsian group (i.e., a quasiconformal deformation of a torsion free Fuchsian group of the first kind without parabolics), then its region of discontinuity consists of two topological discs, say $D_{1}$ and $D_{2}$, and the quotients $D_{1}/G$ and $D_{2}/G$ are closed Riemann surfaces of the same genus $g \geq 2$. Bers' simultaneous uniformization theorem \cite{Bers} asserts that, given any two closed Riemann surfaces $S_{1}$ and $S_{2}$ of the same genus $g \geq 2$, then it is possible to find $G$ as above so that we may assume that $S_{j}$ is isomorphic to $D_{j}/G$, for $j=1,2$.
Let $G$ be now a noded $\pi^{g}$ group ($g \geq 2$). If $G$ is purely loxodromic, then Maskit \cite{M3} has shown that $G$ is
in fact a quasifuchsian group and we are as above. So, we are left to consider the case when $G$ contains parabolic transformations.
In \cite{Hidalgo:nodedFuchsian,M3} there provided examples of noded $\pi^{g}$ groups (with parabolic transformations) which are not quasifuchsian ones (the example in \cite{M3} is a B-group \cite{M1} and that in \cite{Hidalgo:nodedFuchsian} is a group without invariant components in its region of discontinuity). But, if consider the extended region of discontinuity, then the situation is similar to the quasifuchsian case, as seen in the following.
\vspace{0.3cm}
\noindent
\begin{theo}[Simultaneous uniformization theorem with nodes \cite{Hidalgo:nodedFuchsian, Kra-Maskit}]
\mbox{}
\begin{enumerate}
\item If $g \geq 2$ and $G$ is a noded $\pi^{g}$ group, then $\Omega(G)^{ext}$ consists
of exactly two simply-connected invariant components, and $\Omega(G)^{ext}/G$ consists of exactly
two stable Riemann surface of genus $g$.
\item If $S$ and $R$ are two stable Riemann surfaces of genus $g \geq 2$, then there is a
noded $\pi^{g}$ group $G$ such that $\Omega(G)^{ext}/G$ is isomorphic to $S \cup R$.
\end{enumerate}
\end{theo}
\vspace{0.3cm}
As a consequence, the same fact holds for those noded Fuchsian groups containing noded $\pi^{g}$ groups.
\vspace{0.3cm}
\noindent
\begin{coro}
Let $G$ be a noded Fuchsian group, containing a noded $\pi^{g}$ group as a finite index normal subgroup. Then $\Omega(G)^{ext}$ consists of two simply-connected invariant components, and $\Omega(G)^{ext}/G$ consists of exactly two stable Riemann orbifolds.
\end{coro}
\vspace{0.3cm}
The above result was proved by Abikoff in \cite{Ab1} for the case
of cusps, that is, for regular B-groups.
\section{Some preliminaries on Teichm\"uller spaces of Kleinian groups}\label{Sec:prelim2}
In this section, $G$ will be a finitely generated Kleinian group, with non-empty region of discontinuity $\Omega(G)$, and $\Delta \subset \Omega(G)$ will be a non-empty $G$-invariant collection of components of $\Omega(G)$, that is, every $\gamma \in G$ permutes the connected components of $\Delta$.
We proceed to recall the quasiconformal deformation space $T(\Delta,G)$ of $G$ supported at $\Delta$. In the case that $G$ is Fuchsian group of the first kind acting on the unite disc ${\mathbb D}$, then $T({\mathbb D},G)$ is a model for the Teichm\"uller space of the Riemann orbifold ${\mathbb D}/G$.
\subsection{Quasiconformal homeomorphisms}
Let us consider the Banach space $L^{\infty}(\widehat{\mathbb C})$ of measurable maps $\mu:\widehat{\mathbb C} \to {\mathbb C}$,
with the essential supreme norm $\|\cdot\|_{\infty}$. We denote its unit ball as
$L^{\infty}_{1}(\widehat{\mathbb C})$.
An orientation preserving homeomorphism $w:\widehat{\mathbb C} \to
\widehat{\mathbb C}$ is called a {\it quasiconformal homeomorphism} if there is some
$\mu \in L^{\infty}_{1}(\widehat{\mathbb C})$
(called a {\it complex dilatation} of $w$) such that $w$ has distributional partial
derivatives $\partial w$, $\overline{\partial} w$ in
$L^{2}_{loc}(\widehat{\mathbb C})$ satisfying the Beltrami equation
$$\overline{\partial}w(z)=\mu(z)\partial w(z), \quad \mbox{ a.e. $z \in
\widehat{\mathbb C},$ }$$
where $L^{2}_{loc}(\widehat{\mathbb C})$ means $L^{2}$ on compacts in $\widehat{\mathbb C}$.
The existence and uniqueness of quasiconformal homeomorphisms is due to Morrey \cite{Morrey}.
\vspace{0.3cm}
\begin{theo}[Measurable Riemann mapping theorem \cite{A-B,Morrey}]\label{A-B}
If $\mu \in L^{\infty}_{1}(\widehat{\mathbb C})$, then there exists, and it is unique,
a quasiconformal homeomorphism $w_{\mu}:\widehat{\mathbb C} \to \widehat{\mathbb C}$, with complex dilation $\mu$, and fixing $\infty$, $0$ and $1$. Moreover, if $\mu$ vary continuously or real-analytically or holomorphically and $z_0$ is a fixed point, then $w_{\mu}(z_0)$ varies also in the same way.
\end{theo}
\vspace{0.3cm}
\subsection{Beltrami coefficients for $G$ supported at $\Delta$}
The discreteness property of $G$ asserts that
$$L^{\infty}(\Delta,G)=
\left\{\mu \in L^{\infty}(\widehat{\mathbb C}): \; \mu(\gamma(z)) \overline{\gamma'(z)} = \mu(z) \gamma'(z), \;
{\rm a.e.}\; \Delta,\; \mbox{for all} \; \gamma \in G, \; \mu(z)=0,\; z \in \Delta^{c}\right\}$$
is a closed subspace of $L^{\infty}(\widehat{\mathbb C})$; so it is a Banach space with essential supreme norm $\|\cdot\|_{\infty}$.
Let us denote its unit ball as
$$L^{\infty}_{1}(\Delta,G)=L^{\infty}(\Delta,G) \cap L^{\infty}_{1}(\widehat{\mathbb C})=\left\{\mu \in L^{\infty}(\Delta,G); \;
\|\mu\|_{\infty} < 1\right\}.$$
The measurable functions in $L^{\infty}_{1}(\Delta,G)$ are called the
{\it Beltrami coefficients for $G$ supported in $\Delta$}.
The following lemma is a
classical result and its proof can be found, for instance, in \cite{L}.
\vspace{0.3cm}
\noindent
\begin{lemm} \label{lema1}
Let $\mu \in L^{\infty}_{1}(\Delta,G)$ and let
$w:\widehat{\mathbb C} \to \widehat{\mathbb C}$
be a quasiconformal homeomorphism with complex dilatation $\mu$.
Then $w G w^{-1}$ is a Kleinian group with the
region of discontinuity $w(\Omega(G))$. Moreover, $w(\Delta)$ is a $W G w^{-1}$-invariant collection of components.
\end{lemm}
\vspace{0.3cm}
\subsection{Teichm\"uller and moduli spaces}
Observe that, by the measurable Riemann mapping theorem, if $\mu \in L^{\infty}_{1}(\Delta,G)$ and $w_{1}, w_{2}$ are quasiconformal homeomorphisms, both with complex dilation $\mu$, then there is a M\"obius transformation $A \in {\rm PSL}_{2}({\mathbb C})$ so that $w_{2}=A w_{1}$.
Let $\mu,\nu \in L^{\infty}_{1}(\Delta,G)$ be two Beltrami coefficients for $G$ supported in $\Delta$. If $w_{\mu}$ (respectively $w_{\nu}$) is a
quasiconformal homeomorphism with complex dilation $\mu$ (respectively $\nu$), then (by Lemma \ref{lema1}) there is a natural isomorphism $\theta_{\mu}:G \to w_{\mu} G w_{\mu}^{-1}$ (respectively $\theta_{\nu}:G \to w_{\nu} G w_{\nu}^{-1}$). We say that $\mu$ and $\nu$ are {\it Teichm\"uller equivalent} (respectively, {\it isomorphic})
if there is $A \in {\rm PSL}_{2}({\mathbb C})$ with
$\theta_{\mu}(\gamma)=A \theta_{\nu}(\gamma) A^{-1}$, for all $\gamma \in G$ (respectively, $\theta_{\mu}(G)=A \theta_{\nu}(G) A^{-1}$).
The space $T(\Delta,G)$ (respectively, ${\mathcal M}(\Delta,G)$) of Teichm\"uller (respectively, isomorphic) equivalence classes of Beltrami coefficients for $G$ supported in $\Delta$ is called the {\it Teichm\"uller space} (respectively, {\it moduli space}) {\it of $G$ supported in $\Delta$}.
The {\it modular group of $G$ supported at $\Delta$} is the group $M(\Delta,G)$ given by the isotopy classes of quasiconformal homeomorphisms $w:\widehat{\mathbb C} \to \widehat{\mathbb C}$, with complex dilation in $L^{\infty}_{1}(\Delta,G)$, so that $w G w^{-1}=G$. There is the natural action
$$M(\Delta,G) \times T(\Delta,G) \to T(\Delta,G): ([w],[\mu]) \mapsto [\nu],$$
where $\nu \in L^{\infty}_{1}(\Delta,G)$ is complex dilation of the quasiconformal homeomorphism $w_{\mu} w^{-1}$.
It is well known that $T(\Delta,G)$ is a finite dimensional complex manifold (simply-connected in the case that $\Delta$ is connected and simply-connected), that $M(\Delta,G)$ acts as a discrete group of its holomorphic automorphisms and that ${\mathcal M}(\Delta,G)=T(\Delta,G)/M(\Delta,G)$ is a complex orbifold of same dimension as $T(\Delta,G)$ (generalities on this can be found, for instance, in \cite{Bers2,Kra,M5,N}).
\subsection{The Fuchsian case} \label{Sec:Fuchsian}
Let us assume that $G$ is a Fuchsian group of the first kind
acting on the unit disc ${\mathbb D}$; so the Riemann orbifold ${\mathbb D}/G$ has signature $(\gamma;m_{1},\ldots,m_{r})$. In this case $T({\mathbb D},G)$ is a simply-connected complex manifold of dimension $3\gamma-3+r$ (see, for instance, the book \cite{N}) and, moreover, this space is a model for the Teichm\"uller space of the Riemann orbifold ${\mathbb D}/G$.
If $\Gamma$ is a finite index subgroup of $G$, then the natural inclusion $L^{\infty}_{1}({\mathbb D},G) \subset L^{\infty}_{1}({\mathbb D},\Gamma)$ induces a holomorphic embedding $T({\mathbb D},G) \subset T({\mathbb D},\Gamma)$. Also, in this case, the subgroup $M(\Gamma,G)$ of $M({\mathbb D},G)$, formed by those isotopy classes of quasiconformal homeomorphisms for which $w \Gamma w^{-1}=\Gamma$, has finite index. So the complex orbifold $T({\mathbb D},G)/M(\Gamma,G)$ provides a finite degree branched cover of moduli space ${\mathcal M}({\mathbb D},G)$.
\vspace{0.3cm}
\noindent
\begin{rema}\label{obs2}
Let $G$ a finitely generated Kleinian group and let $\Delta$ a $G$-invariant collection of connected components of $\Omega(G) \neq \emptyset$.
Let $\Delta_{1},\ldots,\Delta_{r}$ be a maximal collection of non-$G$-conjugate components of $\Delta$. Let $G_{j}$ be the $G$-stabilizer of $\Delta_{j}$ and
let $G(\Delta_{j})$ be the union of all components of $\Delta$ which are $G$-conjugate to $\Delta_{j}$. As a consequence of results of Kra \cite{Kra} and Sullivan \cite{Sullivan}, it was observed by Kra and Maskit in \cite{K-M} that
$$T(\Delta,G)=T(G,G(\Delta_{1})) \times \cdots \times T(G,G(\Delta_{r}))$$
and that its universal cover space is
$$T({\mathbb D},F_{1})\times \cdots \times T({\mathbb D},F_{r}),$$
where $F_{j}$ is a Fuchsian group acting on ${\mathbb D}$ so that ${\mathbb D}/F_{j} \cong \Delta_{j}/G_{j}$. This in particular asserts that that the statement at the end of Section \ref{Sec:Fuchsian} still valid for $G$.
\end{rema}
\vspace{0.3cm}
\section{Noded quasiconformal deformation spaces of Kleinian groups}\label{Sec:NodedBeltrami}
If $G$ is a Fuchsian group of the first kind acting on the unit disc ${\mathbb D}$, many different compactifications of $T({\mathbb D},G)$ have been provided. For instance, Bers' compactification \cite{Bers3} is obtained by holomorphically embedding $T({\mathbb D},G)$ into the space of quadratic holomorphic forms on ${\mathbb D}/G$ (this being a finite dimensional complex vector space) and taking its closure in there, and Thurston's compactification \cite{Thurston1988} is obtained by taking hyperbolic lengths at simple closed geodesics, which provides a holomorphic embedding into an infinite-dimensional projective space. On Ber's compactifications it is no possible to extend continuously the action of the corresponding modular group. There is not a natural relation between these compactifications and Deligne-Mumford's compactification. A partial closure of $T({\mathbb D},G)$, called the {\it Augmented Teichm\"uller space} $\widehat{T}({\mathbb D},G)$, was constructed by Bers \cite{Bers3} (see also W. Abikoff in \cite{Abikoff}). This space is a non-compact Hausdorff space over which the modular group $M({\mathbb D},G)$ extends continuously and so that $\widehat{T}({\mathbb D},G)/M({\mathbb D},G)$ coincides with Deligne-Mumford's compactification. The added points to $\widehat{T}({\mathbb D},G)$ are certain regular b-groups (geometrically finite Kleinian groups with an invariant simply-connected connected of its region of discontinuity \cite{M1})
If $\Delta$ is a collection of $G$-invariant components of $\Omega(G)$, where $G$ is a finitely generated Kleinian group, then
in \cite{Hidalgo-Vasiliev} we have constructed a partial closure $NT(\Delta,G)$ of the Teichm\"uller space $T(\Delta,G)$ so that the boundary points correspond in a natural way to the boundary points in the Deligne-Mumford compactification of moduli space of $\Delta/G$. In the particular case that $G$ is Fuchsian of the first kind and $\Delta={\mathbb D}$, the partial closure $NT({\mathbb D},G)$ coincides with the augmented Teichm\"uller space. The extra points we add to $T(\Delta,G)$ produce, in terms of Kleinian groups, stable Riemann orbifolds and they correspond to deformations of the group $G$ by the process of approximation of double-cusped parabolic elements of $G$ by certain {\it primitive simple loxodromic} ones (see the works of Keen, Series and Maskit in \cite{KMS}, and of Maskit in \cite{M2,M4}). The deformation is produced by some boundary points of $L^{\infty}_{1}(G,\Delta)$, called {\it noded Beltrami differentials for $G$}. At the level of the Riemann orbifold $\Delta/G$ this means that we permit certain pairwise disjoint simple closed geodesics and maybe some simple geodesic arcs connecting conical values of order $2$ to degenerate to points in order to produce a finite collection of noded Riemann orbifolds. The loops and arcs which we consider in the degeneration process are the projections of appropriately chosen axes of the primitive simple loxodromic elements of the group that approach doubly-cusped parabolic transformations.
In this section we will assume that $G$ is a finitely generated Kleinian group, with non-empty region of discontinuity, and that $\Delta$ a non-empty $G$-invariant collection of connected components of its region of discontinuity, and we proceed to recall the construction of $NT(\Delta,G)$ as done in \cite{Hidalgo-Vasiliev}.
\subsection{Region of discontinuity of Beltrami forms}
For each $\mu \in L^{\infty}(\widehat{\mathbb C})$ we define its {\it region of discontinuity} $\Omega(\mu)$ as the set of all points $p \in \widehat{\mathbb C}$ for which there is
an open neighborhood $U$ of $p \in U$ so that $\|\mu|_{U}\|_{\infty} <1$.
Its complement $\Lambda(\mu)=\widehat{\mathbb C}-\Omega(\mu)$ is the {\it limit set}
of $\mu$. By the definition, the set $\Omega(\mu)$ is open and $\Lambda(\mu)$ is
compact.
If $\mu \in \overline{L^{\infty}_{1}(\Delta,G)}$, the closure of the unit ball
$L^{\infty}_{1}(\Delta,G)$ inside the Banch space $L^{\infty}(\Delta,G)$, then
in \cite{Hidalgo-Vasiliev} it was observed that both $\Omega(\mu)$ and $\Lambda(\mu)$ are $G-$invariant.
Also, by the $G$-invariance of $\Omega(\mu)$ and $\Omega(G)$, the open set $\Omega:=\Omega(\mu) \cap \Omega(G)$ is $G$-invariant. Observe that, as $\mu$ is equal to zero on the complement of $\Delta$, all connected components of $\Omega(G)-\Delta$ are necessarily contained in $\Omega$.
\subsection{Noded quasiconformal maps for $(\Delta,G)$}
If $V \subset \widehat{\mathbb C}$ is a non-empty open set, then $L^{2,1}_{loc}(V)$ denotes the complex vector space of maps $w:V \to \widehat{\mathbb C}$ with locally integrable distributional derivatives.
Let $\mu \in \overline{L^{\infty}_{1}(\Delta,G)}$, where $\Omega(\mu) \neq \emptyset$. An orientation-preserving map $w \in L^{2,1}_{loc}(\Omega(\mu))$ is a {\it noded quasiconformal map for $(\Delta,G)$} with dilatation $\mu$ if the following hold:
\begin{itemize}
\item[(i)] there is a component of $\Omega(\mu)$ homeomorphically mapped by $w$ onto its image;
\item[(ii)] $ \overline{\partial} w(z) = \mu(z) \partial w(z), \,\,\, \mbox{\rm a.e. in $\Omega(\mu)$}$;
\item[(iii)] there is a sequence $\mu_{n} \in L^{\infty}_{1}(\Delta,G)$,
converging to $\mu$ almost everywhere in $\Omega(\mu)$;
\item[(iv)] there is a sequence $w_{n}:\widehat{\mathbb C} \to \widehat{\mathbb C}$ of
quasiconformal homeomorphisms with complex dilatations $\mu_{n}$, converging to $w$
locally uniformly in $\Omega(\mu)$.
\end{itemize}
\vspace{0.3cm}
In the above definition there are many properties to be checked, but the
following existence result is classical and a proof can be find in \cite{Hidalgo-Vasiliev}.
\vspace{0.3cm}
\noindent
\begin{prop}\label{Prop2}
Let $\mu \in \overline{L^{\infty}_{1}(\Delta,G)}$ be such that $\Omega(\mu) \neq \emptyset$, $\Omega_{1}$ be a connected component
of $\Omega(\mu)$ and $x_{1}, x_{2}, x_{3} \in \Omega_{1}$ three different points. Then
there is a noded quasiconformal map $w:\Omega(\mu) \to \widehat{\mathbb C}$, for $(\Delta,G)$, with dilatation $\mu$,
fixing the points $x_{1}$, $x_{2}$ and $x_{3}$,
which is a homeomorphism when restricted to $\Omega_{1}$.
\end{prop}
\vspace{0.3cm}
\noindent
\begin{rema}
Let us observe that the noded quasiconformal map $w$ for $(\Delta,G)$ in Proposition \ref{Prop2} might be constant on some other connected component of $\Omega(\mu)$.
\end{rema}
\vspace{0.3cm}
\subsection{Noded family of arcs}
A countable collection ${\mathcal F}_{\mu}=\{ \alpha_{1},\alpha_{2},...\}$ of pairwise disjoint simple arcs
(including end points)
is called a {\it noded family of arcs associated with $\mu \in \overline{L^{\infty}_{1}(G,\Delta)}$}, if the
following properties hold:
\begin{itemize}
\item[(1)] $\alpha_{n}^{*} \subset \Delta$, where $\alpha_{n}^{*}$ denotes
$\alpha_{n}$ minus both extremes;
\item[(2)] the spherical diameter of
$\alpha_{n}$ goes to $0$ as $n$ goes to $\infty$;
\item[(3)] $\Lambda(\mu)=\overline{\cup_{j=1}^{\infty}\alpha_{j}}$;
\item[(4)] $\Omega(\mu) \subset \widehat{\mathbb C}$ is a dense subset;
\item[(5)] the group $G_{n}=\{g \in G; g(\alpha_{n})=\alpha_{n}\}$, is either
a cyclic loxodromic group or a ${\mathbb Z}_{2}-$extension of a cyclic
loxodromic group.
\end{itemize}
\vspace{0.3cm}
\noindent
\begin{rema}
There are exist examples of $\mu \in \overline{L^{\infty}_{1}(G,\Delta)}$ for which there is no associated noded family of arcs. In \cite{Hidalgo-Vasiliev} there are constructed examples in the positive direction.
\end{rema}
\vspace{0.3cm}
\subsection{Noded Beltrami coefficients for $G$ supported in $\Delta$}
By the stereographic projection, we may see the Riemann sphere as the unit sphere in ${\mathbb R}^{3}$. This allows us to consider the spherical metric and work with the spherical diameter of a subset of $\widehat{\mathbb C}$.
An element $\mu \in \overline{L^{\infty}_{1}(\Delta,G)}$ will be called a
{\it noded Beltrami coefficient for $G$ supported in $\Delta$} if:
\begin{enumerate}
\item[(I)] $\mu$ has a noded family of arcs ${\mathcal F}_{\mu}=\{ \alpha_{1},\alpha_{2},...\}$, and
\item[(II)] there is a continuous map $w:\widehat{\mathbb C} \to \widehat{\mathbb C}$, called a {\it noded quasiconformal deformation of $G$ with complex dilatation $\mu$}, satisfying the following properties:
\begin{enumerate}
\item[(1)] $w$ is injective in $\widehat{\mathbb C}-{\mathcal F}_{\mu}$;
\item[(2)] $w:\Omega(\mu) \to \widehat{\mathbb C}$ is a noded quasiconformal map for $(\Delta,G)$ with complex dilatation $\mu$;
\item[(3)] the restriction of $w$ to each arc $\alpha_{i}$ is a constant
$p_{i}$, where $p_{i} \neq p_{j}$ for $i \neq j$.
\end{enumerate}
\end{enumerate}
We denote by $L^{\infty}_{{\rm noded}}(\Delta,G)$ the subset of $\overline{L^{\infty}_{1}(\Delta,G)}$ consisting of the noded Beltrami coefficients for $G$ supported at $\Delta$. Clearly, $L^{\infty}_{1}(\Delta,G) \subset L^{\infty}_{{\rm noded}}(\Delta,G)$.
\vspace{0.3cm}
\noindent
\begin{rema}
Let $\mu \in L^{\infty}_{{\rm noded}}(\Delta,G)$, with noded family of arcs ${\mathcal F}$, and let
$w:\widehat{\mathbb C} \to \widehat{\mathbb C}$ be a
noded quasiconformal deformation of $G$ with the complex dilatation $\mu$. Then,
the following statements are true
\begin{itemize}
\item[(1)] if $\alpha \in {\mathcal F}_{\mu}$, then
both end points are the fixed points
of a loxodromic element of $G$. Such a loxodromic element keeps
the connected component of $\Omega(G)$ containing $\alpha$
invariant;
\item[(2)] if $\Lambda(\mu) \neq \emptyset$, then
$\Lambda(G) \subset \Lambda(\mu)$. This is a consequence of Proposition E.4 in \cite[page 96]{M1};
\item[(3)] $w(\Omega(u)) \cap w(\Lambda(\mu))=\emptyset$. Indeed, if there
were points $p_{1} \in \Omega(\mu)$ and $p_{2} \in \Lambda(\mu)$, such that
$w(p_{1})=w(p_{2})=q$, then by continuity of $w$ we could find two disjoint open
sets $U \subset \Omega(\mu)$ and $V$, $p_{1} \in U$, and $p_{2} \in V$, such that
$w(U)=w(V)$. The density property of $\Omega(\mu)$ asserts that there
are points $q_{1} \in U$ and $q_{2} \in V \cap \Omega(\mu)$ for which
$w(q_{1})=w(q_{2})$, contradicting the injectivity of the map $w$ restricted to
$\Omega(\mu)$;
\item[(4)] As a consequence of the definition of noded Beltrami differential, we have $w(\widehat{\mathbb C})=\widehat{\mathbb C}$, that is, the map
$w$ is surjective.
\end{itemize}
\end{rema}
\vspace{0.3cm}
The importance of the noded quasiconformal deformations of Kleinian groups is
reflected in the following result.
\vspace{0.3cm}
\noindent
\begin{theo}[\cite{Hidalgo-Vasiliev}]\label{Teo1}
Let $\mu \in L^{\infty}_{{\rm noded}}(\Delta,G)$ and
$w:\widehat{\mathbb C} \to \widehat{\mathbb C}$ be a
noded quasiconformal deformation for $G$ with the complex dilatation $\mu$. Then
there is a unique Kleinian group $\theta(G)$
and a unique isomorphism of groups $\theta:G \to \theta(G)$ such
that $w \circ \gamma = \theta(\gamma) \circ w$. Moreover, the region of discontinuity of
$\theta(G)$ is $w(\Omega(\mu) \cap \Omega(G))$.
\end{theo}
\vspace{0.3cm}
The proof of the above theorem provided in \cite{Hidalgo-Vasiliev} also permit to have the following more general situation.
\vspace{0.3cm}
\noindent
\begin{coro}[\cite{Hidalgo-Vasiliev}]\label{Cor1}
Let $\mu \in \overline{L^{\infty}_{1}(\Delta,G)}$
and let $w:\Omega(\mu) \to w(\Omega(\mu))$ be a
homeomorphism with complex dilatation $\mu$.
Suppose that there is a sequence $w_{n}$ of quasiconformal homeomorphisms
with corresponding complex dilatations $\mu_{n} \in L^{\infty}_{1}(\Delta,G)$,
converging locally uniformly to $w$ in $\Omega(\mu)$. Then
there exist a group $\theta(G)$ of M\"obius transformations and
an isomorphism of groups $\theta:G \to \theta(G)$,
such that $w \circ \gamma = \theta(\gamma) \circ w$.
\end{coro}
\vspace{0.3cm}
The following result relates two noded quasiconformal deformations with the same dilation similar as the situation for the quasiconformal ones.
\vspace{0.3cm}
\noindent
\begin{prop}\label{Lema2}
Let $\mu \in L^{\infty}_{noded}(\Delta,G)$ and let $w_{1}$ and $w_{2}$ be noded
quasiconformal deformations of $G$ with the complex dilatation $\mu$. Then there exists
an orientation preserving homeomorphism
$T:\widehat{\mathbb C} \to \widehat{\mathbb
C}$ whose restriction $T:w_{1}(\Omega(\mu)) \to w_{2}(\Omega(\mu))$ is a conformal mapping, such that
$T \circ w_{1} = w_{2}$.
\end{prop}
\begin{proof} The construction of $T$ is given as follows.
\begin{itemize}
\item[(3.1)] If $x \in w_{1}(\Omega(\mu))$, then set
$T(x)=w_{2}(w_{1}^{-1}(x))$.
\item[(3.2)] If $x \in w_{1}(\alpha_{n})$, then
set $T(x)=w_{2}(\alpha_{n})$.
\item[(3.3)] If $x \in
\Lambda(\mu)-\cup_{n}\alpha_{n}([0,1])$, then set $T(x)=w_{2}(w_{1}^{-1}(x))$.
\end{itemize}
\end{proof}
\vspace{0.3cm}
\noindent
\begin{rema}
Let $\mu \in L^{\infty}_{{\rm noded}}(\Delta,G)$, with associated noded family of arcs ${\mathcal F}$,
and let
$w:\widehat{\mathbb C} \to \widehat{\mathbb C}$ be a
noded quasiconformal deformation of $G$ with complex dilatation $\mu$. If
$\theta(G)$ and $\theta:G \to \theta(G)$ are as in Theorem \ref{Teo1}, then the following are easy to see.
\begin{itemize}
\item[(1)] If $p$ belongs to one of the arcs in ${\mathcal F}$,
then $w(p)$ is necessarily a doubly-cusped parabolic fixed point of $\theta(G)$.
\item[(2)] If $p \in \Lambda(G)$ is a loxodromic fixed point in $G$,
which does not belong to any arc of ${\mathcal F}$, then $w(p)$ is again a
loxodromic fixed point of $\theta(G)$.
\item[(3)] If $p$ is a rank two parabolic fixed point of $G$, then $w(p)$ is
again a rank two parabolic fixed point in $\theta(G)$.
\item[(4)] If $p$ is a doubly-cusped parabolic fixed point of $G$, then $w(p)$
is again doubly-cusped in $\theta(G)$.
\end{itemize}
\end{rema}
\vspace{0.3cm}
The previous remark permits to see the following fact.
\vspace{0.3cm}
\noindent
\begin{theo}\label{Teo2}
Under the hypothesis of Theorem \ref{Teo1}, if $G$ is
geometrically finite, then $\theta(G)$ is also geometrically finite.
\end{theo}
\vspace{0.3cm}
The above, applied to Fuchsian groups of the first kind, permits to obtain the following with respect to noded Fuchsian groups.
\vspace{0.3cm}
\noindent
\begin{coro}\label{coro4}
Let $G$ be a Fuchsian group of the first kind. Let $\mu \in L^{\infty}_{noded}({\mathbb D},G)$, let $w:\widehat{\mathbb C} \to \widehat{\mathbb C}$ be a noded quasiconformal deformation for $G$ with the complex dilation $\mu$ and let $\theta:G \to \theta(G)$ such that $w \circ \gamma=\theta(\gamma) \circ w$, for $\gamma \in G$. Then $\theta(G)$ is a noded Fuchsian group.
\end{coro}
\vspace{0.3cm}
\subsection{Topological Realizations}\label{Sec:toporeal}
A simple loop $\alpha \subset S=\Delta/G$ (or a simple arc connecting two branch values of order $2$) is called {\it pinchable} if (a) its lifting on $\Omega(G)$
consists of pairwise disjoint simple arcs, and (b) each of these components is
stabilized by a primitive loxodromic transformation in $G$.
The stabilizer of a component of the lifting of a pinchable
loop $\alpha$ is a cyclic group generated by a primitive loxodromic
transformation, and the stabilizer of a component of the lifting of a
pinchable arc is a ${\mathbb Z}/2{\mathbb Z}$-extension of a cyclic group
generated by a primitive loxodromic transformation. We say that such a cyclic
group (or ${\mathbb Z}/2{\mathbb Z}$-extension) is defined by the pinchable
loop $\alpha$.
If $\beta_{1}$ and $\beta_{2}$ are two components of the lifting of a pinchable
$\alpha$, then their stabilizers are conjugate in $G$.
We say that a
collection $\{\alpha_{1},...,\alpha_{m}\}$ of pairwise disjoint pinchable
loops or arcs is admissible if they define non-conjugate
groups in $G$ for $i \neq j$.
\vspace{0.3cm}
\noindent
\begin{rema}\label{obsefuch}
In the particular case that $G$ is a Fuchsian group of the first kind and $\Delta={\mathbb D}$, then every homotopically non-trivial loop in $\Delta/G$ avoiding the branch locus is pinchable. Moreover, every collection ${\mathcal F}$ of pairwise disjoint pinchable loops is admissible. Also, if $\Gamma$ is a finite index subgroup of $G$ and $Q:{\mathbb D}/\Gamma \to {\mathbb D}/G$ is a branched cover induced by the inclusion $\Gamma < G$, then the collection $Q^{-1}({\mathcal F})$ is admissible for $\Gamma$.
\end{rema}
\vspace{0.3cm}
For each admissible collection ${\mathcal F}=\{\alpha_{1},...,\alpha_{m}\}$ of pinchable
loops on $S$ we define the following equivalence relation on the Riemann sphere
$\widehat{\mathbb C}$. Two points $p, q \in \widehat{\mathbb C}$
are equivalent if either:
\begin{itemize}
\item[(1)] $p=q$; or
\item[(2)] there is a component $\widetilde{\alpha_{j}}$ of the lifting of
some pinchable loop or arc $\alpha_{j}$, such that $p,q \in
\widetilde{\alpha_{j}} \cup \{a,b\}$, where $a$ and $b$ are the endpoints of
$\widetilde{\alpha_{j}}$ (that is, the fixed points of a primitive
loxodromic transformation in the stabilizer of
$\widetilde{\alpha_{j}}$ in $G$)
\end{itemize}
The set of equivalence classes for such an equivalence relation is
topologically the Riemann sphere. In fact, let us denote by $\widetilde{\mathcal F}$ the collection of all arcs (including their endpoints), as considered in (2) above, and let us consider the collection of continua given by the collection of arcs in $\widetilde{\mathcal F}$ as points and also each of the points in the complement of $\widetilde{\mathcal F}$. The discreteness of $G$ asserts that such collections of points is a semi-continuous collection of points, as defined in \cite{Moore}, and the result now follows from \cite[Thm2]{Moore}.
Let us denote by
$P:\widehat{\mathbb C} \to \widehat{\mathbb C}$ the natural continuous
projection defined by the above relation. As a consequence of the results in \cite{M4} we have the following fact.
\vspace{0.3cm}
\noindent
\begin{prop} \label{propo3}
There exists an orientation preserving homeomorphism
$Q:\widehat{\mathbb C} \to \widehat{\mathbb C}$, such that the map
$Q \circ P:\widehat{\mathbb C} \to \widehat{\mathbb C}$ is a noded quasiconformal
deformation of the group $G$.
\end{prop}
\vspace{0.3cm}
As a consequence of the above, the geometrically finite Kleinian groups constructed in \cite{KMS} are obtained by noded quasiconformal deformations of the suitable Kleinian groups. The following topological realization was seen in \cite{Hidalgo-Vasiliev}.
\vspace{0.3cm}
\noindent
\begin{prop}[\cite{Hidalgo-Vasiliev}] \label{propo4}
Let $G$ be a Kleinian group and ${\mathcal F}$ be an admissible
collection of pinchable loops. Denote by
$P:\widehat{\mathbb C} \to \widehat{\mathbb C}$ the continuous projection naturally induced by the equivalence relation defined by $\mathcal F$. Then,
there is a discrete group $\theta(G)$ of orientation preserving homeomorphisms of
the Riemann sphere and there is an isomorphism $\theta:G \to \theta(G)$, such that $P \gamma =
\theta(\gamma) P$, for all $\gamma \in G$.
\end{prop}
\vspace{0.3cm}
\noindent
\begin{rema}
Results of \cite{M4} and \cite{Ohshika}
assert that, in Proposition \ref{propo4}, if $G$ is geometrically finite, then
$\theta(G)$ is also geometrically finite. In fact, it can be shown that $\theta$ and $\theta(G)$ are as obtained in Theorem \ref{Teo1}, so the geometrically finiteness also follows from Theorem \ref{Teo2}.
\end{rema}
\subsection{The Noded Teichm\"uller space of $G$ supported in $\Delta$}\label{construccion}
\subsubsection{}
We may extend the Teichm\"uller equivalence relation, given previously on $L^{\infty}_{1}(\Delta,G)$, to the whole $L^{\infty}_{{\rm noded}}(\Delta,G)$ as follows. Let $\mu,\nu \in L^{\infty}_{{\rm noded}}(\Delta,G)$ and $w_{\mu}$, $w_{\nu}$ be associated noded quasiconformal deformations for $G$, respectively.
Theorem \ref{Teo1} asserts the existence of
isomorphisms $$\theta_{\mu}:G \to G_{\mu} \quad \mbox{and} \quad
\theta_{\nu}:G \to G_{\nu}\;,$$ where
$G_{\mu}$ and $G_{\nu}$ are Kleinian groups
such that $w_{\mu} g = \theta_{\mu}(g) w_{\mu}$ and
$w_{\nu} g = \theta_{\nu}(g) w_{\nu}$ for all $g \in G$.
We say that $\mu$ and $\nu$ are {\it noded Teichm\"uller
equivalent} if there is an orientation preserving homeomorphism
$A:\widehat{\mathbb C} \to \widehat{\mathbb C}$, such that
\begin{itemize}
\item[(1)] $A(w_{\mu}(\Omega(\mu)))=w_{\nu}(\Omega(\nu))$;
\item[(2)] $A:w_{\mu}(\Omega(\mu)) \to w_{\nu}(\Omega(\nu))$ is conformal;
\item[(3)] $\theta_{\nu}(g)=A \theta_{\nu}(g) A^{-1}$ for all $g \in G$.
\end{itemize}
If in the above we replace (3) by $\theta_{\nu}(G)=A \theta_{\nu}(G) A^{-1}$, then we will say that $\mu$ and $\nu$ are {\it isomorphic}.
The {\it noded Teichm\"uller space of $G$ supported in $\Delta$} is the set $NT(\Delta,G)$ of the noded Teichm\"uller equivalence classes of noded Beltrami coefficients
for $G$ supported in $\Delta$. If $\Delta=\Omega(G)$, then we denote it by $NT(G)$.
\vspace{0.3cm}
\noindent
\begin{rema} If $\mu, \nu \in L^{\infty}_{{\rm noded}}(\Delta,G)$ are noded
Teichm\"uller equivalent and $\mu \in L^{\infty}_{1}(\Delta,G)$, then (1) and
(2) in above definition assert that $\nu \in L^{\infty}_{1}(\Delta,G)$ and
that they are Teichm\"uller equivalent. Moreover,
the inclusion $L^{\infty}_{1}(\Delta,G) \subset L^{\infty}_{{\rm noded}}(\Delta,G)$ induces, under the above equivalence relation,
the inclusion $T(\Delta,G) \subset NT(\Delta,G)$.
\end{rema}
\vspace{0.3cm}
The modular group $M(\Delta,G)$ extends naturally to act on $NT(\Delta,G)$
$$M(\Delta,G) \times NT(\Delta,G) \to NT(\Delta,G): ([w],[\mu]) \mapsto [\nu],$$
where $\nu \in L^{\infty}_{noded}(\Delta,G)$ is complex dilation of the noded quasiconformal deformation $w_{\mu} w^{-1}$. The quotient space
$N{\mathcal M}(\Delta,G)=NT(\Delta,G)/M(\Delta,G)$ is the space of isomorphic classes of noded Beltrami coefficients for $G$ supported at $\Delta$. It provides Deligne-Mumford's compactification of the moduli space ${\mathcal M}(\Delta,G)$.
\vspace{0.3cm}
\noindent
\begin{rema} In the setting of representation of groups, $NT(G)$ corresponds to
consider, in the representation space
Hom($G$,PGL($2,{\mathbb C}$)), the faithful representations of $G$ which are
geometrically represented by noded quasiconformal deformations.
This is a natural generalization for the deformation space of $G$,
on which one considers the geometric representations given by
quasiconformal homeomorphisms of $G$.
\end{rema}
\vspace{0.3cm}
\subsection{The case of Fuchsian groups: Augmented Teichm\"uller space}\label{Ex5}
Let us assume in here that $G$ is a Fuchsian group of the first kind without parabolic elements, keeping the unit disc $\Delta_{1}={\mathbb D}$ invariant, and set
$\Delta_{2}=\widehat{\mathbb C}-\overline{\Delta_{1}}$. In this case, $\Omega(G)=\Delta_{1} \cup \Delta_{2}$ and
the Riemann orbifolds $\Delta_{1}/G$ and $\Delta_{2}/G$ are of signature $(g;n_{1},\stackrel{r}{\ldots},n_{r})$, where $2 \leq n_{1} \leq \cdots \leq n_{r}<\infty$ and
$n_{1}^{-1}+\cdots+n_{r}^{-1}<2g+r-2$.
Associated to $G$ we have the Teichm\"ullers spaces $T(\Delta_{1},G)$, $T(\Delta_{2},G)$ and $T(G)=T(\Omega(G),G)$. These are
simply connected complex manifolds, the first two isomorphic and of complex dimensions $3g-3+r$, and the last one of complex dimension $6g-6+2r$. In fact, there is an isomorphism
$$\theta:T(G) \to T(\Delta_{1},G) \times T(\Delta_{2},G): [\mu] \mapsto ([\mu_{1}],[\mu_{2}]),$$
where $\mu_{i}$ is defined as $\mu$ on $\Delta_{i}$ and zero on its complement.
Associated to the above three spaces are the corresponding partial closures given by the noded Teichm\"uller spaces $NT(\Delta_{1},G)$, $NT(\Delta_{2},G)$ and $NT(G)$. Each $NT(\Delta_{j},G)$ can be identified with the Augmented Teichm\"uller
space of $G$ as defined in \cite{Ab1,Ab2,Bers3}. The isomorphism $\theta$ extends continuously to an isomorphism
$$\theta:NT(G) \to NT(\Delta_{1},G) \times NT(\Delta_{2},G).$$
The Teichm\"uller space $T(\Delta_{j},G)$ has associated the Weil-Petersson (WP) metric, which is K\"ahler, has negative sectional curvature, is not complete and for every pair of points there is a unique geodesic connecting them\cite{Ahlfors,Royden,Tromba,Wolpert1,Wolpert2}. It is known that the WP metric completion of $T(\Delta_{j},G)$ is the augmented Teichm\"uller space $NT(\Delta_{j},G)$ \cite{Masur} (but it is a non-locally compact space).
Results on the geometry of geodesics on these spaces is given in \cite{Wolpert3}.
The modular group $M(\Delta_{j},G)$ acts a group of WP orientation-preserving isometries (these are all these isometries if $(g,r) \neq (1,2)$ \cite{Wolpert3}).
It was observed by Masur that $NT(\Delta_{j},G)/M(\Delta_{j},G)$ is the quotient WP metric completion $\overline{\mathcal M}(\Delta_{j},G)$ (this being the Deligne-Mumford's compactification) of the moduli space ${\mathcal M}(\Delta_{j},G)$ (the space of isomorphic classes of Riemann surfaces of genus $g$ with $r$ punctures).
In \cite{HV} it was proved that if $M$ is a finite index subgroup of the modular group $M({\mathbb D},G)$, then the quotient spaces $NT(\Delta_{1},G)/M$ and $NT(\Delta_{2},G)/M$ are compact complex orbifolds. In particular, $NT(\Omega(G)/M$ is also a compact complex orbifold.
\vspace{0.3cm}
\noindent
\begin{rema}
Now, following Remark \ref{obs2} and the above, for every pair $(\Delta,G)$, where $G$ is a finitely generated Kleinian group, and $M$ being of finite index in $M(\Delta,G)$, it holds that $NT(\Delta,G)/M$ is a compact complex orbifold.
\end{rema}
\vspace{0.3cm}
\section{Uniformizations of $(\gamma,n)$-gonal pairs: proof of Theorem \ref{main1}} \label{Sec:uniformizationsofpairs}
Let us consider a $(\gamma,n)$-gonal pair $(S,f)$ and
let us denote by $B_{f}\subset R$ the locus of branch values of $f$.
\subsection{The signature of $(S,f)$}
For each point in $B_{f}$ we define its {\it branch order} as the minimum common multiple of the local degrees at all its preimages under $f$.
Let us write $B_{f}=\{p_{1},\ldots,p_{r}\}$ so that $n_{j}$ is the branch order of $p_{j}$, where $2 \leq n_{1} \leq n_{2} \leq \cdots \leq n_{r}$.
The {\it signature} of $(S,f)$ is then defined by the tuple $(\gamma;n_{1},\ldots,n_{r})$. We assume, from now on, that $(S,f)$ is of {\it hyperbolic type}, that is,
$n_{1}^{-1}+\cdots+n_{r}^{-1}<2\gamma+r-2$ (in particular, if $\gamma=0$, then $r \geq 3$). If $S$ has genus $g$, then Riemann-Hurwitz formula asserts the equality
$$2(g-1-n(\gamma-1))=\sum_{j=1}^{r}\left(n-\# f^{-1}(p_{j})\right).$$
Topologically equivalent $(\gamma,n)$-gonal pairs have the same signature, but the converse is in general false.
\subsection{A pair of Riemann orbifolds}
For each point $q \in f^{-1}(p_{j})$ set
$m_{q}=n_{j}/d_{q}$, where $d_{q}$ is the local degree of $f$ at $q$, and let $N_{f}$ be the set of of points $q \in f^{-1}(B_{f})$ with $m_{q}>1$.
Let $R^{orb}$ be the Riemann orbifold whose underlying Riemann surface structure is $R$, its cone point set is $B_{f}$ and the cone order of each $p_{j}$ is $n_{j}$. Similarly, let $S^{orb}$ be the Riemann orbifold whose underlying Riemann surface structure is $S$ and its cone points are given by the points $q \in N_{f}$ with cone order $m_{q}>1$.
\subsection{An uniformizing pair}
As a consequence of the uniformization theorem: (i) there is a co-compact Fuchsian group $G$ acting on the unit disc ${\mathbb D}$ and an orbifold isomorphism $\psi:R^{orb} \to {\mathbb D}/G$; that is, $G$ has signature $(\gamma;n_{1},\ldots,n_{r})$, and (ii) there is an index $n$ subgroup $\Gamma$ and an orbifold isomorphism $\phi:S^{orb} \to {\mathbb D}/\Gamma$ so that $\psi f \phi^{-1}:{\mathbb D}/\Gamma \to {\mathbb D}/G$ is being induced by the inclusion $\Gamma < G$.
The constructed pair of Fuchsian groups $(\Gamma,G)$ is uniquely determined by the pair $(S,f)$, up to conjugation by elements in ${\rm Aut}({\mathbb D})$, the group of holomorphic automorphisms of ${\mathbb D}$; we say that $(\Gamma,G)$ {\it uniformizes} $(S,f)$. The following fact follows almost immediately from the definitions.
\vspace{0.3cm}
\noindent
\begin{lemm}\label{lema2}
Two pairs of Fuchsian groups $(\Gamma_{1},G_{1})$ and $(\Gamma_{2},G_{2})$ uniformize topologically equivalent $(\gamma,n)$-gonal pairs (of hyperbolic type) if and only if there is an orientation-preserving homeomorphism $u:{\mathbb D} \to {\mathbb D}$ so that $u G_{1} u^{-1}=G_{2}$ and $u \Gamma_{1} u^{-1}=\Gamma_{2}$ (as the orbifolds are compact, we may assume $u$ to be a quasiconformal homeomorphism). They correspond to twisted isomorphic pairs if $u \in {\rm Aut}({\mathbb D})$ and to isomorphic ones if $G_{1}=G_{2}$ and $u \in G_{2}$.
\end{lemm}
\vspace{0.3cm}
\noindent
\begin{rema}
Let us recall that a holomorphic automorphism of $(S,f)$ is a holomorphic automorphism $T$ of $S$ so that $f=f T$. In this context,
the group of holomorphic automorphisms of $(S,f)$ is naturally identified with the quotient $N_{G}(\Gamma)/\Gamma$, where $N_{G}(\Gamma)$ is the normalizer of $\Gamma$ in $G$.
\end{rema}
\vspace{0.3cm}
As a consequence of Lemma \ref{lema2} one obtains that ${\mathcal H}_{0}(S,f)$ can be identified with
the Teichm\"uller space $T({\mathbb D},G)$ of the orbifold ${\mathbb D}/G$, which is a simply-connected complex manifold of dimension $3\gamma+ r-3$ \cite{K-M,L,N}.
\vspace{0.3cm}
As every point in ${\mathcal H}(S,f)$ is uniformized by a quasiconformal deformation of $(\Gamma,G)$, is a consequence of the measurable Riemann mapping theorem \cite{A-B,Morrey} that the Hurwitz space ${\mathcal H}(S,f)$ is connected and it can be identified with
$T({\mathbb D},G)/M(\Gamma,G)$, where $M(\Gamma,G)$ is a subgroup of the modular group $M({\mathbb D},G)$ (the group of holomorphic automorphisms of $T({\mathbb D},G)$) induced by the isotopic classes of those quasiconformal self-homeomorphisms of ${\mathbb D}$ normalizing both $G$ and also $\Gamma$. The above, in particular, asserts that ${\mathcal H}(S,f)$ is a connected complex orbifold of dimension $3\gamma+r-3$ whose orbifold fundamental group is isomorphic to $M(\Gamma,G)$.
\vspace{0.3cm}
\noindent
\begin{rema}
As the subgroup $M(\Gamma,G)$ has finite index in the modular group $M({\mathbb D},G)$, there is a finite degree branched covering ${\mathcal H}(S,f) \to {\mathcal M}({\mathbb D},G)$, where
${\mathcal M}({\mathbb D},G)$ is the moduli space of the orbifold ${\mathbb D}/G$ (this being the moduli space of an $r$-punctured Riemann surface of genus $\gamma$). For instance, if $\Gamma$ is a characteristic subgroup of $G$, then $M(\Gamma,G)=M({\mathbb D},G)$, so ${\mathcal H}(S,f)={\mathcal M}({\mathbb D},G)$.
\end{rema}
\vspace{0.3cm}
\section{Uniformizations of stable $(\gamma,n)$-gonal pairs: Proof of Theorem \ref{main2}}\label{Sec:nodedpairs}
In this section we proceed to construct uniformizations of an stable $(\gamma,n)$-gonal pair using noded Fuchsian groups. These uniformization permits to provide a compactification $\overline{\mathcal H}(S,f)$, for $(S,f)$ a $(\gamma,n)$-gonal pair of hyperbolic type, this being a complex orbifold containing ${\mathcal H}(S,f)$ as an open dense suborbifold.
\subsection{}
Let us recall that each stable $(\gamma,n)$-gonal pair $(S^{*},f^{*})$ is obtained by considering a suitable $(\gamma,n)$-gonal pair $(S,f:S \to R)$ of hyperbolic type and a suitable collection ${\mathcal F}=\{\gamma_{1}, \ldots, \gamma_{s} \}$ of pairwise disjoint simple loops in $R-B_{f}$, where $B_{f}$ is the set of branch values of $f$. If $r$ is the cardinality of $B_{f}$, then for $\gamma=0$ we also assume that $r \geq 4$ (otherwise, there is no possible degeneration to make).
The collection ${\mathcal F}$ satisfies that the Euler characteristic of each connected component of $R-(B_{f} \cup \gamma_{1}\cup \cdots \cup \gamma_{s})$ is negative and none of the components of $R-(\gamma_{1} \cup \cdots \cup \gamma_{s})$ is a disc with only two branch values, both with branch order equal to $2$.
By the pinching process of the loops in ${\mathcal F}$ and those in $f^{-1}({\mathcal F})$ (as indicated in the introduction), we obtain a topological stable $(\gamma,n)$-gonal pair homeomorphic to $(S^{*},f^{*})$. To obtain $(S^{*},f^{*})$ we need to provide to each component of the nodes of a suitable analytically finite Riemann surface.
\subsection{}
Let us consider a pair $(\Gamma,G)$ of Fuchsian groups uniformizing $(S,f)$, that is, there are orbifold isomorphisms
$$\phi:S^{orb} \to {\mathbb D}/\Gamma, \quad \psi:R^{orb} \to {\mathbb D}/G$$
so that $ \pi:=\psi f \phi^{-1}$ is branched cover induced by the inclusion of $\Gamma$ in $G$. Recall that the pair $(\Gamma,G)$ is unique up to conjugation by elements of ${\rm Aut}({\mathbb D})$.
\subsection{}
The collection of simple loops $\psi({\mathcal F}) \subset {\mathbb D}/G$ lifts under $\pi$ to a collection ${\mathcal G}$ of simple arcs in ${\mathbb D}$ which is pinchable for $G$, so also for the subgroup $\Gamma$ (see Remark \ref{obsefuch}).
\subsection{}
Set $\overline{\mathbb D}=\{z \in {\mathbb C}: |z|>1\} \cup \{\infty\}$ and let us consider three different points $a,b,c \in \overline{\mathbb D}$.
\subsection{}
Now, as a consequence of Propositions \ref{propo3} and \ref{propo4}, there is a $ \mu \in L^{\infty}_{noded}({\mathbb D},G)$ with limit set being the closure of the collection ${\mathcal G}$. If $w_{\mu}:\widehat{\mathbb C} \to \widehat{\mathbb C}$ is a noded quasiconformal deformation for $G$ with the complex dilation $\mu$ and fixing the points $a,b,c$, then Theorem \ref{Teo1} asserts the existence of a unique Kleinian group $\theta(G)$ and a unique isomorphism of groups $\theta:G \to \theta(G)$ such
that $w_{\mu} \circ \gamma = \theta(\gamma) \circ w_{\mu}$. Moreover, the region of discontinuity for the action of
$\theta(G)$ on the Riemann sphere is $w_{\mu}(\Omega(\mu) \cap \Omega(G))=w_{\mu}(\Omega(\mu) \cap {\mathbb D}) \cup w_{\mu}(\overline{\mathbb D})$. Observe that $w_{\mu}(\overline{\mathbb D})$ is a quasidisc.
Corollary \ref{coro4} asserts that $\theta(G)$ (and so $\theta(\Gamma)$) is a noded Fuchsian group. By Proposition \ref{propo3}, the pair $(\theta(\Gamma),\theta(G))$, when restricted to $w_{\mu}(\Omega(\mu) \cap {\mathbb D})$, provides a stable $(\gamma,n)$-gonal pair modelled by $(S^{*},f^{*})$. Also, the same pair, when restricted to $w_{\mu}(\overline{\mathbb D})$ provides the $(\gamma,n)$-gonal pair $(S,f)$.
Up to post-composition by a suitable quasiconformal homeomorphisms, we may assume that the stable $(\gamma,n)$-gonal pair we obtain is the original one, i.e., $(S^{*},f^{*})$;
we say that the pair $(\theta(\Gamma),\theta(G))$ {\it uniformizes} it.
\subsection{}
For $t \in {\mathbb D}$, we may consider $t\mu \in L^{\infty}_{1}({\mathbb D},G)$ and the quasiconformal homeomorphism
$w_{t\mu}:\widehat{\mathbb C} \to \widehat{\mathbb C}$ with complex dilation $t\mu$ and fixing the points $a,b,c$. If $G_{t}=w_{t\mu} G w_{t\mu}^{-1}$ and
$\Gamma_{t}=w_{t\mu} \Gamma w_{t\mu}^{-1}$, then $(\Gamma_{t},G_{t})$ provides a continuous family of $(\gamma,n)$-gonal pairs $(S_{t},f_{t})$ converging to $(S^{*},f^{*})$.
\subsection{}
The above construction asserts that the Hurwitz space $\overline{\mathcal H}_{0}(S,f)$, parametrizing isomorphic classes of stable $(\gamma,n)$-gonal pairs, modelled by degeneration of a the $(\gamma,n)$-gonal pair $(S,f)$, can be identified with the noded Teichm\"uller space $NT({\mathbb D},G)$.
Similarly, the Hurwitz space $\overline{\mathcal H}(S,f)$, parametrizing twisted isomorphic classes of stable $(\gamma,n)$-gonal pairs, modelled by degeneration of a the $(\gamma,n)$-gonal pair $(S,f)$, can be identified with the quotient space $NT({\mathbb D},G)/M(\Gamma,G)$. As $M(\Gamma,G)$ has finite index in the modular group $M({\mathbb D},G)$, it follows from the results of \cite{HV} that $NT({\mathbb D},G)/M(\Gamma,G)$ carries a structure of a complex orbifold, being a finite branched cover of the Deligne-Mumford's compactification space of ${\mathbb D}/G$.
\vspace{0.3cm}
\noindent
\begin{rema}
The forgetful map (of finite degree)
$${\mathcal H}(S,f) \to {\mathcal M}_{g}: [(S',f')] \mapsto [S']$$ extends to the corresponding forgetful map (also of finite degree)
$$\overline{\mathcal H}(S,f) \to \widehat{\mathcal M}_{g}: [(S',f')] \mapsto [S']$$
\end{rema}
\vspace{0.3cm}
|
\section{Introduction}\label{Section: Introduction}
\subsection{Motivation and objectives}
In linear algebra, it is a classical exercise to determine whether two given matrices, $A,B\in\text{Mat}_{n\times n}(\mathbb{C})$, are conjugate, ie. whether $B=QAQ^{-1}$ for some invertible $Q\in\text{Mat}_{n\times n}(\mathbb{C})$. More generally, one can fix $A\in\text{Mat}_{n\times n}(\mathbb{C})$ and consider the set of all $B\in\text{Mat}_{n\times n}(\mathbb{C})$ that are conjugate to $A$. This set is precisely the conjugacy class of $A$, and, as we shall see, it is the paradigm example of a complex adjoint orbit. Despite this classical motivation, complex adjoint orbits are central to modern research in algebraic geometry \cite{Beauville,BeauvilleSingular,Fu,FuNamikawa,Hesselink,Kaledin,Kraft}, differential geometry \cite{BielawskiOnthe,BielawskiReducible,Biquard,Santa-Cruz,Kovalev,KronheimerInstantons,KronheimerSemisimple}, Lie theory \cite{Collingwood,Vogan,Sommers,Sommers2,Graham2}, and geometric representation theory \cite{Chriss,Jantzen,Springer,Borho}
This article is an introduction to complex adjoint orbits in algebraic geometry, Lie theory, and, to a lesser extent, the other fields mentioned above. It is written with a view to meeting several key objectives. Firstly, it should be accessible to a fairly broad mathematical audience and require only introductory algebraic geometry and Lie theory as prerequisites. Secondly, it aims to build a clear, intuitive, and reasonably self-contained foundation for more advanced topics in which complex adjoint orbits play a role. Thirdly, it is designed to motivate and outline some of these more advanced topics (ex. hyperk\"{a}hler geometry, Landau-Ginzburg models, and symplectic singularities). Our final objective is to emphasize a few contextually appropriate arguments that, while possibly well-known to experts, do not seem to appear in the research literature. The arguments in question are the proofs of Proposition \ref{Proposition: Stabilizer structure}, Lemma \ref{Lemma: K-stabilizer}, Proposition \ref{Proposition: Diffeomorphism with the cotangent bundle}, Corollary \ref{Corollary: Equivariant homotopy equivalence}, Proposition \ref{Proposition: Intersection of opposite parabolics}, Theorem \ref{Theorem: Equivariant projective compactification}, and Proposition \ref{Proposition: Nilpotent orbits form a connected component}. While these are new proofs, at least to the author's knowledge, we emphasize that the underlying results themselves are well-known.
\subsection{Structure of the article}
This article is organized as follows. Section \ref{Section: Preliminaries on algebraic group actions} addresses some of the relevant background on actions of complex algebraic groups, focusing largely on the structure of algebraic group orbits. It includes a number of examples, some of which are designed to give context and build intuition for later sections. Section \ref{Section: Preliminaries on Lie theory} subsequently discusses a few critical topics in Lie theory. Firstly, in the interest of clarity and consistency in later sections, \ref{Subsection: The basic objects} fixes various objects that arise when one studies representations of a complex semisimple group (ex. a maximal torus, a Borel subgroup, weights, roots, etc.). Secondly, \ref{Subsection: semisimple and nilpotent elements} and \ref{Subsection: The Jacobson-Morozov Theorem and sl2-triples} review the basics of semisimple and nilpotent elements in a complex semisimple Lie algebra.
Section \ref{Section: General adjoint orbits} specializes the previous sections to examine the adjoint orbits of a complex semisimple group. Adjoint orbits are defined in \ref{Subsection: Definitions and conventions}, and \ref{Subsection: Dimension and regularity} then introduces \textit{regular} adjoint orbits. Next, \ref{Subsection: Some first results} presents some results relating Jordan decompositions to the study of adjoint orbits. Section \ref{Subsection: The adjoint quotient} discusses adjoint orbits in the context of the \textit{adjoint quotient}, concluding with some of Kostant's foundational results (see Theorem \ref{Theorem: Kostant's theorem}). The topics become more manifestly geometric in \ref{Subsection: Geometric features}, which mentions the role of adjoint orbits in hyperk\"{a}hler and holomorphic symplectic geometry.
Section \ref{Section: Semisimple orbits} is devoted to semisimple adjoint orbits and their features. These orbits are introduced in \ref{Subsection: Definitions and characterizations}, where it is also shown that semisimple orbits are exactly the closed orbits of the adjoint action (see Theorem \ref{Theorem: Semisimple equivalent to closed}). Section \ref{Subsection: Stabilizer descriptions} develops a uniform description of a semisimple element's stabilizer under the adjoint action of a complex semisimple group (see Proposition \ref{Proposition: Stabilizer structure} and Corollary \ref{Corollary: Stabilizers of semisimple elements are reductive}). The discussion becomes more esoteric in Sections \ref{Subsection: Fibrations over partial flag varieties} and \ref{Subsection: Equivariant projective compactifications}. The former shows semisimple orbits to be explicitly diffeomorphic to cotangent bundles of partial flag varieties $G/P$ (see Proposition \ref{Proposition: Diffeomorphism with the cotangent bundle}), which is consistent with results in hyperk\"{a}hler geometry (see Remark \ref{Remark: Connection to hyperkahler geometry}). Section \ref{Subsection: Equivariant projective compactifications} reviews a procedure for compactifying semisimple orbits (see Theorem \ref{Theorem: Equivariant projective compactification}), and subsequently gives some context from the homological mirror symmetry program (see Remark \ref{Remark: Connection to mirror symmetry}).
Section \ref{Section: Nilpotent orbits} is exclusively concerned with nilpotent adjoint orbits and their properties. It begins with \ref{Subsection: Definitions and first results}, in which nilpotent orbits are defined and then characterized in several equivalent ways. Also included is a proof that there are only finitely many nilpotent orbits of a given complex semisimple group (see Theorem \ref{Theorem: Finitely many nilpotent orbits}). Section \ref{Subsection: The closure order on nilpotent orbits} studies the set of nilpotent orbits as a poset, partially ordered by the closure order. In particular, it recalls Gerstenhaber and Hesselink's classical description of this poset in Lie type $A$ (see Example \ref{Example: Nilpotent orbits and the dominance order}). Section \ref{Subsection: The regular nilpotent orbit} shows there to be a unique maximal element in the poset of nilpotent orbits (see Proposition \ref{Proposition: Unique maximal}), called the \textit{regular nilpotent orbit}. We also recall Kostant's construction of a standard representative for this orbit (see Proposition \ref{Proposition: Nilpotent orbit representative}). Next, \ref{Subsection: The minimal nilpotent orbit} introduces $\mathcal{O}_{\text{min}}$, the minimal (non-zero) nilpotent orbit of a complex simple group. This leads to a discussion of nilpotent orbit projectivizations in \ref{Subsection: Orbit projectivizations}, where $\mathbb{P}(\mathcal{O}_{\text{min}})$ and its properties are discussed in detail. Some emphasis is placed on the role of $\mathbb{P}(\mathcal{O}_{\text{min}})$ in quaternionic-K\"{a}hler geometry. Lastly, \ref{Subsection: Orbit closures and singularities} outlines how nilpotent orbit closures can arise in geometric representation theory, and in the study of symplectic singularities and resolutions.
\subsection{Acknowledgements}
The author is grateful to Roger Bielawski, Steven Rayan, and Markus R\"{o}ser for many fruitful discussions, and to the Institute of Differential Geometry at Leibniz Universit\"{a}t Hannover for its hospitality.
\subsection{General conventions}
While many of our conventions will be declared in Sections \ref{Section: Preliminaries on algebraic group actions} and \ref{Section: Preliminaries on Lie theory}, the following are some things we can establish in advance.
\begin{itemize}
\item We will discuss several objects that, strictly speaking, make sense only when a base field has been specified (ex. vector spaces, Lie algebras, representations, algebraic varieties, etc.). Unless we indicate otherwise, we will always take this field to be $\mathbb{C}$.
\item If $X$ is a set with an equivalence relation $\sim$, then $[x]\in X/{\sim}$ will denote the equivalence class of $x\in X$. The equivalence relations of interest to us will usually come from the (left) action of a group $G$ on a set $X$, in which $x_1\sim x_2$ if and only if $x_2=g\cdot x_1$ for some $g\in G$. In such cases, we will denote the quotient set $X/{\sim}$ by $X/G$.
\item Many of the spaces we will consider can be viewed both as algebraic varieties with their Zariski topologies, and as manifolds with their analytic topologies. Accordingly, if one of our statements has topological content, the reader should by default understand the underlying topology to be the Zariski topology. Only when a space does not have a clear algebraic variety structure (ex. a compact Lie group or an orbit thereof), or when manifold-theoretic terms are being used (ex. diffeomorphism, holomorphic, homotopy-equivalence, equivariant cohomology) should we be understood as referring to the analytic topology.
\end{itemize}
\section{Preliminaries on algebraic group actions}\label{Section: Preliminaries on algebraic group actions}
In what follows, we review some pertinent concepts from the theory of group actions on algebraic varieties. This review is designed to serve two principal purposes: to clearly establish some of our major conventions, and to give motivation/context for what is to come. However, it is not intended to be a comprehensive overview of algebraic group actions. For this, we refer the reader to \cite[Chapt. II--IV]{Humphreys}.
\subsection{Definitions and examples}
Let $G$ be a connected linear algebraic group with Lie algebra $\mathfrak{g}$ and adjoint representation $\Adj:G\rightarrow\GL(\mathfrak{g})$. By $G$-\textit{variety}, we shall mean an algebraic variety $X$ equipped with an algebraic $G$-action, namely a variety morphism\\ $G\times X\rightarrow X$, $(g,x)\mapsto g\cdot x$, satisfying the usual properties of a left $G$-action:
\begin{itemize}
\item[(i)] $e\cdot x=x$ for all $x\in X$, where $e\in G$ is the identity element,
\item[(ii)] $(g_1g_2)\cdot x=g_1\cdot (g_2\cdot x)$ for all $g_1,g_2\in G$ and $x\in X$.
\end{itemize}
A morphism of $G$-varieties is then appropriately defined to be a $G$-equivariant morphism of algebraic varieties, ie. a variety morphism $\varphi:X\rightarrow Y$ satisfying $\varphi(g\cdot x)=g\cdot\varphi(x)$ for all $g\in G$ and $x\in X$. This induces the definition of a $G$-variety isomorphism, which is precisely a $G$-equivariant isomorphism of algebraic varieties.
The literature contains many examples of $G$-varieties and related ideas, some of which we mention below.
\begin{example}\label{Example: Left-multiplication} Let $G=\GL_n(\mathbb{C})$ act by left multiplication on the vector space of $n\times n$ matrices $\text{Mat}_{n\times n}(\mathbb{C})$, so that $g\cdot A=gA$, $g\in GL_n(\mathbb{C}),A\in\text{Mat}_{n\times n}(\mathbb{C}).$ In this way, $\text{Mat}_{n\times n}(\mathbb{C})$ is a $\GL_n(\mathbb{C})$-variety.
\end{example}
\begin{example}\label{Example: Conjugation example}Alternatively, one has the conjugation action of $\GL_n(\mathbb{C})$ on $\text{Mat}_{n\times n}(\mathbb{C})$, ie. $g\cdot A=gAg^{-1}$, $g\in GL_n(\mathbb{C}),A\in\text{Mat}_{n\times n}(\mathbb{C}).$ This, too, renders $\text{Mat}_{n\times n}(\mathbb{C})$ a $\GL_n(\mathbb{C})$-variety.
\end{example}
\begin{example}\label{Example: partial flag variety in type A}
By a \textit{flag in } $\mathbb{C}^n$, we shall mean a sequence $$V_{\bullet}=(\{0\}\subsetneq V_1\subsetneq V_2\subsetneq\ldots\subsetneq V_k\subsetneq\mathbb{C}^n)$$ of subspaces of $\mathbb{C}^n$. Now, suppose that $d_{\bullet}=(d_1,d_2,\ldots,d_k)$ is a strictly increasing sequence of integers with $1\leq d_i\leq n-1$ for all $i=1,\ldots,k$. We denote by $\text{Flag}(d_{\bullet},\mathbb{C}^n)$ the set of all flags in $\mathbb{C}^n$ whose constituent subspaces have dimensions $d_1,d_2,\ldots,d_k$, namely
$$\text{Flag}(d_{\bullet};\mathbb{C}^n)=\{V_{\bullet}:\dim(V_i)=d_i,\text{ }i=1,\ldots,k\}.$$ This is a smooth projective variety and it carries a $\GL_n(\mathbb{C})$-variety structure in which $\GL_n(\mathbb{C})$ acts via
$$g\cdot V_{\bullet}=(\{0\}\subsetneq gV_1\subsetneq gV_2\subsetneq\ldots\subsetneq gV_k\subsetneq\mathbb{C}^n),\quad g\in\GL_n(\mathbb{C}),V_{\bullet}\in\text{Flag}(d_{\bullet},\mathbb{C}^n).$$
\end{example}
\begin{example}
Let $X$ be a $G$-variety and $H\subseteq G$ a closed subgroup. One can restrict the $G$-action to $H$ in the sense that $H\times X\rightarrow X$, $(h,x)\mapsto h\cdot x$, $h\in H,x\in X$ defines an $H$-variety structure on $X$.
\end{example}
\begin{example}\label{Example: restriction to an invariant subvariety}
Let $X$ be a $G$-variety and $Y\subseteq X$ a locally closed subvariety.\footnote{Recall that a subset of $X$ is defined to be locally closed if it is expressible as the intersection of an open and a closed subset of $X$. Equivalently, a subset is locally closed if and only if the subset is open in its closure.} If $g\cdot y\in Y$ for all $g\in G$ and $y\in Y$, we will refer to $Y$ as being $G$-\textit{invariant} (or simply \textit{invariant} if the action is clear from context). In this case, the map $G\times Y\rightarrow Y$, $(g,y)\mapsto g\cdot y$, is a $G$-variety structure on $Y$.
\end{example}
\begin{example}\label{Example: Diagonal action}
Let $X$ and $Y$ be $G$-varieties. The product $X\times Y$ then carries a canonical $G$-variety structure, defined by the ``diagonal'' $G$-action $$g\cdot(x,y)=(g\cdot x,g\cdot y),\quad g\in G,\text{ }(x,y)\in X\times Y.$$
\end{example}
\begin{example}\label{Example: action from representation}
Let $V$ be a $G$-\textit{representation}, meaning for us that $V$ is a finite-dimensional vector space together with a morphism of algebraic groups $\rho:G\rightarrow\GL(V)$ (ex. $V=\mathfrak{g}$ and $\rho=\Adj$). In this case, the map $G\times V\rightarrow V$, $(g,v)\mapsto\rho(g)(v)$, gives $V$ the structure of a $G$-variety. Note that Examples \ref{Example: Left-multiplication} and \ref{Example: Conjugation example} feature $G$-varieties induced by representations.
\end{example}
\begin{example}\label{Example: Isotropy representations}
As a counterpart to the last example, one can use $G$-varieties to induce representations. To this end, let $X$ be a $G$-variety and for each $g\in G$ let $\phi_g:X\rightarrow X$ be the automorphism defined by $\phi_g(x)=g\cdot x$. Now suppose $H\subseteq G$ is a closed subgroup and $x\in X$ is a point satisfying $h\cdot x=x$ (ie. $\phi_h(x)=x$) for all $h\in H$. It follows that the differential of $\phi_h$ at $x$ is a vector space automorphism $d_x(\phi_h):T_xX\rightarrow T_xX$ of the tangent space $T_xX$. Furthermore, one can verify that $H\rightarrow\GL(T_xX)$, $h\mapsto d_x(\phi_h)$, is an $H$-representation.
\end{example}
\begin{example}\label{Example: Descent to projective space} Let $\rho:G\rightarrow\GL(V)$ be a $G$-representation, and let $\mathbb{P}(V)=(V\setminus\{0\})/\mathbb{C}^*$ denote the projectivization of $V$. The formula $g\cdot[v]=[\rho(g)(v)]$ gives a well-defined action of $G$ on $\mathbb{P}(V)$, rendering the latter a $G$-variety.
\end{example}
\begin{example}\label{Example: Induced action on functions}
Let $X$ be an affine $G$-variety and $\mathbb{C}[X]$ the algebra of regular functions on X. This algebra carries a canonical action of $G$ by algebra automorphisms, in which $g\in G$ acts on $f\in\mathbb{C}[X]$ by $(g\cdot f)(x):=f(g^{-1}\cdot x)$, $x\in X$.
\end{example}
\begin{example}\label{Example: Quotient by a closed subgroup} Let $H$ be a closed subgroup of $G$. The set of left $H$-cosets $X=G/H=\{gH:g\in G\}$ is naturally an algebraic variety. Furthermore, when $G$ acts by left multiplication, $G/H$ is a $G$-variety.
\end{example}
A few remarks are in order. Firstly, one may summarize Examples \ref{Example: action from representation} and \ref{Example: Descent to projective space} together as the statement that a $G$-representation $V$ induces canonical $G$-variety structures on both $V$ and $\mathbb{P}(V)$. We will always understand $V$ and $\mathbb{P}(V)$ to be $G$-varieties in exactly this way when $V$ is a $G$-representation. Secondly, we shall always view a variety $G/H$ as carrying the left-multiplicative action of $G$.
\subsection{Generalities on orbits}\label{Subsection: Generalities on orbits}
Let $X$ be a $G$-variety. One may associate to each point $x\in X$ the subset \begin{equation}\label{Equation: Orbit definition}
\mathcal{O}_G(x):=\{g\cdot x:g\in G\},
\end{equation}
called the $G$-\textit{orbit of} $x$. Should the underlying group $G$ be clear from context, we will sometimes suppress it and write $\mathcal{O}(x)$ in place of $\mathcal{O}_G(x)$. Now, let us call $\mathcal{O}\subseteq X$ a $G$-\textit{orbit} if $\mathcal{O}=\mathcal{O}(x)$ for some $x\in X$. In this case, $\overline{\mathcal{O}}$ shall denote the closure of $\mathcal{O}$ in $X$. The orbit $\mathcal{O}$ is open in $\overline{\mathcal{O}}$ (ie. is locally closed), so that $\mathcal{O}$ inherits from $X$ the structure of a subvariety. Furthermore, $\mathcal{O}$ is known to be a smooth subvariety of $X$.
\begin{example}Let $\GL_n(\mathbb{C})$ act by left multiplication on $\text{Mat}_{n\times n}(\mathbb{C})$. If $A,B\in\text{Mat}_{n\times n}(\mathbb{C})$, then $B=gA$ for some $g\in\GL_n(\mathbb{C})$ if and only if $A$ and $B$ are equivalent under row operations, which holds if and only if their reduced echelon forms coincide. It follows that the $\GL_n(\mathbb{C})$-orbits are parametrized by the set of $n\times n$ matrices in reduced echelon form. In particular, the orbit of matrices with reduced echelon form $I_n$ is $\mathcal{O}(I_n)=\GL_n(\mathbb{C})\subseteq\text{Mat}_{n\times n}(\mathbb{C})$. This orbit is open and dense.
\end{example}
\begin{example}\label{Example: Conjugation}
Let $\GL_n(\mathbb{C})$ act by conjugation on $\text{Mat}_{n\times n}(\mathbb{C})$, so that the $\GL_n(\mathbb{C})$-orbits are precisely the conjugacy classes of matrices in $\text{Mat}_{n\times n}(\mathbb{C})$. These classes are indexed by the equivalence classes of $n\times n$ matrices in Jordan canonical form, where two such matrices are equivalent if they agree up to a reordering of Jordan blocks along the diagonal. To build some algebro-geometric intuition for these orbits, we specialize to three cases.
\begin{itemize}
\item Suppose that $A\in\text{Mat}_{n\times n}(\mathbb{C})$ is diagonal with pairwise distinct eigenvalues\\ $\lambda_1,\ldots,\lambda_n\in\mathbb{C}$. If $B\in\text{Mat}_{n\times n}(\mathbb{C})$, then $B\in\mathcal{O}(A)$ if and only if the characteristic polynomial of $B$ has the form $$\det(\lambda I_n-B)=(\lambda-\lambda_1)(\lambda-\lambda_2)\cdots(\lambda-\lambda_n).$$ Equating corresponding coefficients of $\lambda^k$ for each $k\in\{0,1,\ldots,n-1\}$, we obtain $n$ polynomial equations in the entries of $B$. From this, it follows that $\mathcal{O}(A)$ is a closed (hence affine) subvariety of $\text{Mat}_{n\times n}(\mathbb{C})$.
\item Let $N_1\in\text{Mat}_{n\times n}(\mathbb{C})$ be the matrix consisting of a single nilpotent Jordan block. If $B\in\text{Mat}_{n\times n}(\mathbb{C})$, then $B\in\mathcal{O}(N_1)$ if and only $B$ has a minimal polynomial of $\lambda^n$, ie.
$$\mathcal{O}(N_1)=\{B\in\text{Mat}_{n\times n}(\mathbb{C}):B^n=0,\text{ }B^{n-1}\neq 0\}.$$ The conditions $B^n=0$ and $B^{n-1}\neq 0$ define closed and open subsets of $\text{Mat}_{n\times n}(\mathbb{C})$, respectively, and $\mathcal{O}(N_1)$ is the intersection of these subsets. We thus see explicitly that $\mathcal{O}(N_1)$ is locally closed. Also, one can show that $\overline{\mathcal{O}(N_1)}=\{B\in\text{Mat}_{n\times n}(\mathbb{C}):B^n=0\}$, the locus of all nilpotent matrices.
\item Let $N_2\in\text{Mat}_{n\times n}(\mathbb{C})$ be a nilpotent matrix in Jordan canonical form having a single $2\times 2$ Jordan block and all remaining blocks of dimension $1\times 1$. To describe $\mathcal{O}(N_2)$, note that the Jordan canonical form of a nilpotent $B\in\text{Mat}_{n\times n}(\mathbb{C})$ has only $1\times 1$ and $2\times 2$ blocks if and only if $B^2=0$. To ensure the existence of exactly one $2\times 2$ block, one must impose the condition $\text{rank}(B)=1$. Equivalently, one requires that $B\neq 0$ and that all $2\times 2$ minors of $B$ vanish. It follows that \begin{align*}\mathcal{O}(N_2) & =\{B\in\text{Mat}_{n\times n}(\mathbb{C}):B^2=0,\text{ rank}(B)=1\}\\ & = \{B\in\text{Mat}_{n\times n}(\mathbb{C}):B^2=0,\text{ every } 2\times 2\text{ minor of }B=0,\text{ }B\neq 0\}.
\end{align*}
As with $\mathcal{O}(N_1)$, we explicitly see that $\mathcal{O}(N_2)$ is the intersection of a closed subset of $\text{Mat}_{n\times n}(\mathbb{C})$ (defined by $B^2=0$ and the vanishing of the $2\times 2$ minors) with an open subset (defined by $B\neq 0$). Furthermore, it is not difficult to see that $0\in\overline{\mathcal{O}(N_2)}$, and that $\mathcal{O}(N_2)\cup\{0\}$ is the closed subset mentioned in the previous sentence. We conclude that $\overline{\mathcal{O}(N_2)}=\mathcal{O}(N_2)\cup\{0\}$.
\end{itemize}
\end{example}
\begin{example}\label{Example: second partial flag variety in type A}
Refer to Example \ref{Example: partial flag variety in type A}, in which $\GL_n(\mathbb{C})$ acts on $\text{Flag}(d_{\bullet};\mathbb{C}^n)$. One can verify that this action is transitive, so that $\text{Flag}(d_{\bullet};\mathbb{C}^n)$ is the unique $\GL_n(\mathbb{C})$-orbit. However, let us restrict the $\GL_n(\mathbb{C})$-action to the Borel subgroup $B$ of upper-triangular matrices. To describe the $B$-orbits, let $w\in S_n$ be a permutation and consider the point
$$V_{\bullet}^w:=(\{0\}\subsetneq V_{1}^w\subsetneq V_2^w\subsetneq\ldots\subsetneq V_{k}^w\subsetneq\mathbb{C}^n)\in \text{Flag}(d_{\bullet};\mathbb{C}^n),$$ where $V_i^w:=\text{span}\{e_{w(1)},e_{w(2)},\ldots,e_{w(d_i)}\}$, $i\in\{1,2,\ldots,k\}$. It turns out that $w\mapsto\mathcal{O}_B(V_{\bullet}^w)$ is a surjection from $S_n$ to the collection of $B$-orbits. Furthermore, $\mathcal{O}_B(V_{\bullet}^w)$ is isomorphic to affine space and called a \textit{Schubert cell}, alluding to the fact that the $B$-orbits constitute a cell decomposition of $\text{Flag}(d_{\bullet};\mathbb{C}^n)$. The $B$-orbit closures are the well-studied \textit{Schubert varieties}. For further details in this direction, we refer the reader to \cite[Chapt. 10]{Fulton}.
\end{example}
\subsection{Orbits as quotient varieties}\label{Subsection: Orbits as quotient varieties}
Let $X$ be a $G$-variety and $\mathcal{O}\subseteq X$ an orbit. Since $\mathcal{O}$ is locally closed and $G$-invariant, it inherits from $X$ the structure of a $G$-variety (cf. Example \ref{Example: restriction to an invariant subvariety}). To better understand this structure, fix a point $x\in\mathcal{O}$ and consider its $G$-stabilizer
\begin{equation}\label{Equation: Stabilizer definition}
C_G(x):=\{g\in G:g\cdot x=x\}.
\end{equation}
This is a closed subgroup of $G$ and one has a well-defined $G$-variety isomorphism
\begin{equation}\label{Equation: Orbit-stabilizer isomorphism}
G/C_G(x)\xrightarrow{\cong}\mathcal{O},\quad [g]\mapsto g\cdot x.
\end{equation}
Note that \eqref{Equation: Orbit-stabilizer isomorphism} is just the Orbit-Stabilizer Theorem in an algebro-geometric context.
One can use \eqref{Equation: Orbit-stabilizer isomorphism} to yield a convenient description of the tangent space $T_x\mathcal{O}$. Indeed, since \eqref{Equation: Orbit-stabilizer isomorphism} sends the identity coset $[e]\in G/C_G(x)$ to $x$, it determines an isomorphism of tangent spaces,
\begin{equation}\label{Equation: Vector space isomorphism}
T_{[e]}(G/C_G(x))\xrightarrow{\cong} T_x\mathcal{O}.
\end{equation}
Also, the differential of $G\rightarrow G/C_G(x)$ at $e\in G$ is surjective with kernel the Lie algebra of $C_G(x)$. Let us denote this Lie algebra by $C_{\mathfrak{g}}(x)$, so that we have $T_{[e]}(G/C_G(x))\cong\mathfrak{g}/C_{\mathfrak{g}}(x)$ as vector spaces. The isomorphism \eqref{Equation: Vector space isomorphism} then takes the form
\begin{equation}\label{Equation: Tangent space description}
T_x\mathcal{O}\cong\mathfrak{g}/C_{\mathfrak{g}}(x).
\end{equation}
It turns out that \eqref{Equation: Tangent space description} is more than an isomorphism of vector spaces. To see this, note that $T_x\mathcal{O}$ is a $C_G(x)$-representation by virtue of $x$ being fixed by $C_G(x)$ (see Example \ref{Example: Isotropy representations}). At the same time, $\mathfrak{g}$ carries the adjoint representation of $G$, and one can restrict this to a representation of $C_G(x)$. One notes that $C_{\mathfrak{g}}(x)$ is a $C_G(x)$-invariant subspace, making the quotient $\mathfrak{g}/C_{\mathfrak{g}}(x)$ a $C_G(x)$-representation. With this discussion in mind, one can verify that \eqref{Equation: Tangent space description} is actually an isomorphism of $C_G(x)$-representations.
\begin{example}\label{Example: Stabilizer in a flag variety}
Recall the action of $\GL_n(\mathbb{C})$ on $\text{Flags}(d_{\bullet};\mathbb{C}^n)$ from Example \ref{Example: partial flag variety in type A}. Following the notation of Example \ref{Example: second partial flag variety in type A}, the $\GL_n(\mathbb{C})$-stabilizer of the flag $V^e_{\bullet}$ is
$$C_{\GL_n(\mathbb{C})}(V^e_{\bullet})=\{(A_{ij})\in\GL_n(\mathbb{C}):A_{ij}=0\text{ for all }r\in\{1,\ldots,k\}, i>d_r, j\in\{d_{r-1},\ldots,d_r\}\}.$$ More simply, a matrix $A\in\GL_n(\mathbb{C})$ is in $C_{\GL_n(\mathbb{C})}(V^e_{\bullet})$ if and only if it is block upper-triangular with blocks of dimensions $d_1\times d_1, (d_2-d_1)\times (d_2-d_1),\ldots,(d_k-d_{k-1})\times (d_k-d_{k-1}),(n-d_k)\times (n-d_k)$ along the diagonal (read from top to bottom). Now, since the action of $\GL_n(\mathbb{C})$ on $\text{Flags}(d_{\bullet};\mathbb{C}^n)$ can be shown to be transitive, \eqref{Equation: Orbit-stabilizer isomorphism} gives an isomorphism
$$\GL_n(\mathbb{C})/C_{\GL_n(\mathbb{C})}(V^e_{\bullet})\cong \text{Flags}(d_{\bullet};\mathbb{C}^n).$$
\end{example}
\begin{example}
Recall Example \ref{Example: Conjugation}, and let $A,N_1\in\text{Mat}_{n\times n}(\mathbb{C})$ be as introduced there. One can verify that $C_{\GL_n(\mathbb{C})}(A)$ is the maximal torus of diagonal matrices in $\GL_n(\mathbb{C})$, while $C_{\GL_n(\mathbb{C})}(N_1)$ can be seen to have the following description:
$$C_{\GL_n(\mathbb{C})}(N_1)=\{(B_{ij})\in\GL_n(\mathbb{C}):B_{ij}=0\text{ for }i>j, \text{ }B_{ij}=B_{(i+1)(j+1)}\text{ for }1\leq i\leq j\leq n-1\}.$$
In other words, a matrix $B\in\GL_n(\mathbb{C})$ is in $C_{\GL_n(\mathbb{C})}(N_1)$ if and only if $B$ is upper-triangular and any two entires of $B$ lying on the same diagonal must be equal. Now, let $U\subseteq C_{\GL_n(\mathbb{C})}(N_1)$ be the unipotent subgroup of matrices having only $1$ along the main diagonal. This subgroup is complementary to the centre $Z(\GL_n(\mathbb{C}))$, and one has an internal direct product decomposition
$C_{\GL_n(\mathbb{C})}(N_1)=Z(\GL_n(\mathbb{C}))\times U$.
\end{example}
\subsection{The closure order}\label{Subsection: The closure order}
The set of orbits in a $G$-variety $X$ turns out to have a canonical partial order, based fundamentally on the following proposition.
\begin{proposition}\label{Proposition: orbit boundary}
If $\mathcal{O}\subseteq X$ is a $G$-orbit, then $\overline{\mathcal{O}}$ is a union of $\mathcal{O}$ and orbits having dimensions strictly less than $\dim(\mathcal{O})$.
\end{proposition}
\begin{proof}
Choose $x\in\mathcal{O}$ and note that $\mathcal{O}$ is the image of the morphism $G\rightarrow X$, $g\mapsto g\cdot x$. Since $G$ is irreducible, so too is the image $\mathcal{O}$. It follows that $\overline{\mathcal{O}}$ is also irreducible. Appealing to the fact that $\mathcal{O}$ is locally closed (ie. open in $\overline{\mathcal{O}}$), we see that $\overline{\mathcal{O}}\setminus\mathcal{O}$ is a closed subvariety of $\overline{\mathcal{O}}$, each of whose irreducible components then necessarily has dimension strictly less than $\dim(\overline{\mathcal{O}})=\dim(\mathcal{O})$ (see \cite[Sect. 3.2]{Humphreys}). Of course, as $\overline{\mathcal{O}}\setminus\mathcal{O}$ is $G$-invariant, it is a union of $G$-orbits. Each of these orbits is also irreducible and therefore has dimension at most that of an irreducible component of $\overline{\mathcal{O}}\setminus\mathcal{O}$ to which it belongs. Our previous discussion implies that the dimension of such an orbit is strictly less than $\dim(\mathcal{O})$.
\end{proof}
Proposition 2 implies that each $G$-orbit $\mathcal{O}$ is the unique orbit in $\overline{\mathcal{O}}$ having dimension $\dim(\mathcal{O})$.
With this in mind, suppose that $\mathcal{O}_1,\mathcal{O}_2\subseteq X$ are $G$-orbits satisfying $\mathcal{O}_1\subseteq\overline{\mathcal{O}_2}$ and $\mathcal{O}_2\subseteq\overline{\mathcal{O}_1}$. These inclusions show $\mathcal{O}_1$ and $\mathcal{O}_2$ to have the same dimension. Since both orbits lie in $\overline{\mathcal{O}_1}$, it follows that $\mathcal{O}_1=\mathcal{O}_2$ must hold. It is thus not difficult to see that
\begin{equation}\label{Equation: closure order}
\mathcal{O}_1\leq\mathcal{O}_2\Longleftrightarrow\mathcal{O}_1\subseteq\overline{\mathcal{O}_2}
\end{equation}
defines a partial order on the set of $G$-orbits in $X$, which we shall call the \textit{closure order}.
\begin{example}
Recall the setup of Example \ref{Example: Conjugation}, in which $\GL_n(\mathbb{C})$ acts by conjugation on $\text{Mat}_{n\times n}(\mathbb{C})$ and $A,N_1,N_2\in\text{Mat}_{n\times n}(\mathbb{C})$ are three fixed matrices. Having shown $\mathcal{O}(A)$ to be closed in $\text{Mat}_{n\times n}(\mathbb{C})$, we see that $\mathcal{O}(A)$ is a minimal element in the closure order. Secondly, remembering that $\overline{\mathcal{O}(N_1)}$ is the locus of all nilpotent $n\times n$ matrices, the $G$-orbits in $\overline{\mathcal{O}(N_1)}$ are the conjugacy classes of nilpotent matrices. These conjugacy classes are therefore precisely the orbits $\mathcal{O}\subseteq\text{Mat}_{n\times n}(\mathbb{C})$ satisfying $\mathcal{O}\leq\mathcal{O}(N_1)$. Finally, having seen in Example \ref{Example: Conjugation} that $\overline{\mathcal{O}(N_2)}=\mathcal{O}(N_2)\cup\{0\}$, we conclude that $\{0\}$ and $\mathcal{O}(N_2)$ are the unique orbits $\mathcal{O}$ satisfying $\mathcal{O}\leq\mathcal{O}(N_2)$.
\end{example}
\begin{example}
Refer to Examples \ref{Example: partial flag variety in type A} and \ref{Example: second partial flag variety in type A}, where $\GL_n(\mathbb{C})$ and its subgroup $B$ act on $\text{Flag}(d_{\bullet};\mathbb{C}^n)$. We will assume that $d_{\bullet}=(1,2,\ldots,n-1)$, in which case $w\mapsto\mathcal{O}_B(V_{\bullet}^w)$ is a bijection from $S_n$ to the set of $B$-orbits. The latter set has the closure order, and there is a unique partial order on $S_n$ for which our bijection is a poset isomorphism. One thereby obtains the well-studied \textit{Bruhat order} on $S_n$, which we now describe. Given $i\in\{1,\ldots,n-1\}$, let $s_i\in S_n$ be the transposition interchanging $i$ and $i+1$. The $n-1$ transpositions $s_i$ generate $S_n$, so that each $w\in S_n$ has an expression $w=s_{i_1}s_{i_2}\cdots s_{i_k}$. One calls such an expression \textit{reduced} if its length $k$ is minimal. Now given $v,w\in S_n$, the Bruhat order is defined as follows: $v\leq w$ if whenever $w=s_{i_1}s_{i_2}\cdots s_{i_k}$ is a reduced expression for $w$, it is possible to delete some of the $s_{i_j}$ so that the resulting expression is a reduced expression for $v$.
\end{example}
\section{Preliminaries on Lie theory}\label{Section: Preliminaries on Lie theory}
The following is a brief review of Lie-theoretic topics central to this article, with a strong emphasis on the structure of semisimple algebraic groups and their Lie algebras. Like the previous section, this review is more concerned with establishing clear conventions than being comprehensive. References for the relevant Lie theory include \cite[Chapt. III--VIII]{Humphreys} and \cite[Chapt. 20--29]{Tauvel}.
\subsection{The basic objects}\label{Subsection: The basic objects}
In what follows, we introduce several Lie-theoretic objects to be understood as fixed for the rest of this article. To begin, let $G$ be a connected, simply-connected semisimple linear algebraic group having Lie algebra $\mathfrak{g}$ and exponential map $\exp:\mathfrak{g}\rightarrow G$. Let $$\Adj:G\rightarrow\GL(\mathfrak{g}),\quad g\mapsto\Adj_g,\text{ }g\in G$$ and $$\adj:\mathfrak{g}\rightarrow\mathfrak{gl}(\mathfrak{g}),\quad x\mapsto\adj_x,\text{ }x\in\mathfrak{g}$$ denote the adjoint representations of $G$ and $\mathfrak{g}$, respectively. Recall that $\Adj_g$ is a Lie algebra automorphism of $\mathfrak{g}$ for all $g\in G$, a fact we will use repeatedly. Note also that $\adj_x(y)=[x,y]$ for all $x,y\in\mathfrak{g}$, from which it follows that the kernel of $\adj$ is the centre of $\mathfrak{g}$. This centre is trivial by virtue of the fact that $\mathfrak{g}$ is semisimple. In other words, $\adj$ is injective.
Now, suppose that $T\subseteq B\subseteq G$, with $T$ a maximal torus of $G$ and $B$ a Borel subgroup of $G$. One then has inclusions $\mathfrak{t}\subseteq\mathfrak{b}\subseteq\mathfrak{g}$, where $\mathfrak{t}$ and $\mathfrak{b}$ are the Lie algebras of $T$ and $B$, respectively. Let $X^*(T)$ be the additive group of weights (ie. algebraic group morphisms $T\rightarrow\mathbb{C}^*$), with the sum of $\lambda_1,\lambda_2\in X^*(T)$ defined by $(\lambda_1+\lambda_2)(t)=\lambda_1(t)\lambda_2(t)$ for all $t\in T$. The map sending $\lambda\in X^*(T)$ to its differential at the identity $e\in T$ includes $X^*(T)$ into $\mathfrak{t}^*$ as an additive subgroup. Where appropriate, we will use this inclusion to regard a weight as belonging to $\mathfrak{t}^*$.
Given a weight $\alpha\in X^*(T)$, let us set $$\mathfrak{g}_{\alpha}:=\{x\in\mathfrak{g}:\Adj_t(x)=\alpha(t)x\text{ for all }t\in T\}.$$ If we regard $\alpha$ as belonging to $\mathfrak{t}^*$, then $\mathfrak{g}_{\alpha}$ admits the following alternative description:
$$\mathfrak{g}_{\alpha}=\{x\in\mathfrak{g}:\adj_h(x)=\alpha(h)x\text{ for all }h\in\mathfrak{t}\}.$$
One knows that $\mathfrak{g}_0=\mathfrak{t}$, while any non-zero weight $\alpha$ satisfying $\mathfrak{g}_{\alpha}\neq\{0\}$ is called a \textit{root}. In this case, $\mathfrak{g}_{\alpha}$ is one-dimensional and called a \textit{root space}. Furthermore, letting $\Delta\subseteq X^*(T)$ denote the set of roots, we have the following decomposition of $\mathfrak{g}$: \begin{equation}\label{Equation: Root space decomposition}
\mathfrak{g}=\mathfrak{t}\oplus\bigoplus_{\alpha\in\Delta}\mathfrak{g}_{\alpha}.
\end{equation}
One can check that $[\mathfrak{g}_{\alpha},\mathfrak{g}_{\beta}]\subseteq\mathfrak{g}_{\alpha+\beta}$ for all $\alpha,\beta\in\Delta$. In particular, if $\alpha+\beta$ is neither zero nor a root, then $[\mathfrak{g}_{\alpha},\mathfrak{g}_{\beta}]=\{0\}$.
Our choice of $B$ gives rise to several distinguished subsets of $\Delta$. The positive roots $\Delta_{+}$ are those $\alpha\in\Delta$ for which $\mathfrak{g}_{\alpha}\subseteq\mathfrak{b}$, so that
$$\mathfrak{b}=\mathfrak{t}\oplus\bigoplus_{\alpha\in\Delta_{+}}\mathfrak{g}_{\alpha}.$$ The negative roots $\Delta_{-}\subseteq\Delta$ are defined by negating the positive roots, ie. $\Delta_{-}:=-\Delta_{+}$. We then have a disjoint union $\Delta=\Delta_{+}\cup\Delta_{-}$. Finally, the simple roots $\Pi\subseteq\Delta$ are positive roots with the property that every $\beta\in\Delta$ can be written as \begin{equation}\label{Equation: Root decomposition}\beta=\sum_{\alpha\in\Pi}c_{\alpha}\alpha\end{equation} for unique coefficients $c_{\alpha}\in\mathbb{Z}$. These coefficients are either all non-negative (in which case $\beta\in\Delta_{+}$) or all non-positive (in which case $\beta\in\Delta_{-}$). Furthermore, the simple roots are known to form a basis of $\mathfrak{t}^*$.
It will be beneficial to note that the weight lattice carries a particular partial order, defined in terms of the simple roots. Given $\beta,\gamma\in X^*(T)$, one has \begin{equation}\label{Equation: Partial order on the weight lattice}\beta\leq\gamma\Longleftrightarrow\gamma-\beta=\sum_{\alpha\in\Pi}c_{\alpha}\alpha\end{equation} for some strictly non-negative integers $c_{\alpha}\in\mathbb{Z}$.
The Lie algebra $\mathfrak{g}$ carries a distinguished symmetric bilinear form $\langle\cdot,\cdot\rangle:\mathfrak{g}\otimes\mathfrak{g}\rightarrow\mathbb{C}$, called the \textit{Killing form} and defined by
$$\langle x,y\rangle:=\trace(\adj_x\circ\adj_y),\quad x,y\in\mathfrak{g}.$$ This form is $G$-invariant in the sense that $\langle\Adj_g(x),\Adj_g(y)\rangle=\langle x,y\rangle$ for all $g\in G$ and $x,y\in\mathfrak{g}$. Furthermore, the Killing form is non-degenerate both on $\mathfrak{g}$ and when restricted to a bilinear form on $\mathfrak{t}$. This non-degeneracy gives rise to vector space isomorphisms $\mathfrak{g}\cong\mathfrak{g}^*$ and $\mathfrak{t}\cong\mathfrak{t}^*$, under which the Killing form corresponds to bilinear forms on $\mathfrak{g}^*$ and $\mathfrak{t}^*$, respectively. In an abuse of notation, we will also use $\langle\cdot,\cdot\rangle$ to denote these forms. Each simple root $\alpha\in\Pi$ then determines a \textit{simple coroot} $h_{\alpha}\in\mathfrak{t}$, defined by the following property:
$$\phi(h_{\alpha})=2\frac{\langle\alpha,\phi\rangle}{\langle\alpha,\alpha\rangle}$$
for all $\phi\in\mathfrak{t}^*$. One can check that $h_{\alpha}\in[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]$ and that the simple coroots form a basis of $\mathfrak{t}$.
Now let $W:=N_G(T)/T$ be the Weyl group and $s_{\alpha}\in W$ the reflection associated to $\alpha\in\Delta$, so that $W$ is generated by the simple reflections $s_{\alpha},\alpha\in\Pi$. Note that $W$ acts on $\mathfrak{t}$ by $w\cdot x=\Adj_g(x)$, $w\in W$, $x\in\mathfrak{t}$, where $g\in N_G(T)$ is any representative of $w$. There is an induced action on $\mathfrak{t}^*$ defined by $$(w\cdot\phi)(x)=\phi(w^{-1}\cdot x),\quad w\in W, \text{ }\phi\in\mathfrak{t}^*, \text{ }x\in\mathfrak{t},$$
under which the reflections act as follows:
$$s_{\alpha}\cdot\phi=\phi-2\frac{\langle\alpha,\phi\rangle}{\langle\alpha,\alpha\rangle}\alpha,\quad\alpha\in\Delta,\text{ }\phi\in\mathfrak{t}^*.$$
It is known that $X^*(T)$ and $\Delta$ are invariant under this $W$-action.
Now recall that a closed subgroup $P\subseteq G$ is called \textit{parabolic} if $P$ contains a conjugate of $B$. If $P$ contains $B$ itself, one calls it a \textit{standard parabolic} subgroup. To construct such a subgroup, consider a subset $S\subseteq\Pi$ and let $\Delta_{S}$ (resp. $\Delta_{S}^+$, $\Delta_{S}^{-}$) denote the collection of roots (resp. positive roots, negative roots) expressible as $\mathbb{Z}$-linear combinations of the elements of $S$. It follows that \begin{equation}\label{Equation: Lie algebra of standard parabolic}\mathfrak{p}_S:=\mathfrak{b}\oplus\bigoplus_{\alpha\in\Delta_{S}^{-}}\mathfrak{g}_{\alpha}\end{equation}
is a Lie subalgebra of $\mathfrak{g}$, and we shall let $P_S\subseteq G$ denote the corresponding closed connected subgroup of $G$. Note that $P_S$ contains $B$, by construction. Moreover, it turns out that $S\mapsto P_S$ defines a bijective correspondence between the subsets of $\Pi$ and the standard parabolic subgroups of $G$. The inverse associates to a standard parabolic subgroup $P$ the subset $\Pi_P:=\{\alpha\in\Pi:\mathfrak{g}_{-\alpha}\subseteq\mathfrak{p}\}$, where $\mathfrak{p}$ is the Lie algebra of $P$.
We will later benefit from recalling a particular Levi decomposition of a standard parabolic subgroup $P_S$ (cf. \cite[Sect. 30.2]{Humphreys}). One has \begin{equation}\label{Equation: Levi decomposition of parabolic} P_S=U_S\rtimes L_S,\end{equation} where $U_S$ is the unipotent radical of $P_S$ and $L_S$ is a (reductive) Levi factor having Lie algebra
\begin{equation}\label{Equation: Lie algebra of Levi factor}
\mathfrak{l}_S:=\mathfrak{t}\oplus\bigoplus_{\alpha\in\Delta_S}\mathfrak{g}_{\alpha}=\left(\bigoplus_{\alpha\in\Delta_S^{-}}\mathfrak{g}_{\alpha}\right)\oplus\mathfrak{t}\oplus\left(\bigoplus_{\alpha\in\Delta_S^+}\mathfrak{g}_{\alpha}\right).\end{equation}
The Lie algebra of $U_{S}$ is
\begin{equation}\label{Equation: Lie algebra of unipotent radical}
\mathfrak{u}_{S}:=\bigoplus_{\alpha\in\Delta_{+}\setminus\Delta_{S}^+}\mathfrak{g}_{\alpha}.
\end{equation}
We now specialize some of the above Lie-theoretic generalities to the case of $G=\SL_n(\mathbb{C})$, a recurring example in this article.
\begin{example}\label{Example: The type A setup}
Let $G=\SL_n(\mathbb{C})$, whose Lie algebra is $\mathfrak{g}=\mathfrak{sl}_n(\mathbb{C})=\{x\in\mathfrak{gl}_n(\mathbb{C}):\trace(x)=0\}$ with Lie bracket the commutator of matrices. The adjoint representations are given by
\begin{equation}\label{Equation: Type A adjoint representation}
\Adj_g(x)=gxg^{-1},\quad\adj_x(y)=[x,y]=xy-yx
\end{equation}
for all $g\in\SL_n(\mathbb{C})$ and $x,y\in\mathfrak{sl}_n(\mathbb{C})$. Also, the Killing form $\langle\cdot,\cdot\rangle$ is given by
\begin{equation}\label{Equation: Type A Killing form}
\langle x,y\rangle=2n\trace(xy),\quad x,y\in\mathfrak{sl}_n(\mathbb{C}).
\end{equation}
One may take $T$ and $B$ to be the subgroups of diagonal and upper-triangular matrices in $\SL_n(\mathbb{C})$, respectively. It follows that $\mathfrak{t}$ and $\mathfrak{b}$ are the Lie subalgebras of diagonal and upper-triangular matrices in $\mathfrak{sl}_n(\mathbb{C})$, respectively. Now, for each $i\in\{1,\ldots,n\}$, let $t_i\in X^*(T)$ be the weight defined by
$$t_i:T\rightarrow\mathbb{C}^*,\quad\begin{bmatrix} a_1 & 0 & \ldots & 0\\ 0 & a_2 & \ldots & 0\\ \vdots & \vdots & \ddots & \vdots\ \\ 0 & 0 & \ldots & a_n\end{bmatrix}\mapsto a_i.$$ We then have $\Delta=\{t_i-t_j:1\leq i,j\leq n,i\neq j\}$, $\Delta_{+}=\{t_i-t_j:1\leq i<j\leq n\}$, $\Delta_{-}=\{t_i-t_j:1\leq j<i\leq n\},$ and $\Pi=\{t_i-t_{i+1}:1\leq i\leq n-1\}$. If $\alpha=t_i-t_j\in\Delta$, then $\mathfrak{sl}_n(\mathbb{C})_{\alpha}=\text{span}_{\mathbb{C}}\{E_{ij}\}$ where $E_{ij}\in\mathfrak{sl}_n(\mathbb{C})$ has an entry of $1$ in position $(i,j)$ and 0 elsewhere.
Recall that the Weyl group identifies with $S_n$ in such a way that $w\in S_n$ acts on $h\in\mathfrak{t}$ by permuting entries along the diagonal.
\end{example}
\subsection{Semisimple and nilpotent elements}\label{Subsection: semisimple and nilpotent elements}
It will be prudent to recall some of the fundamentals concerning semisimple and nilpotent elements in $\mathfrak{g}$. We begin with the official, Lie-theoretic definitions of these terms.
\begin{definition}\label{Definition: semisimple and nilpotent}
We call a point $x\in\mathfrak{g}$ \textit{semisimple} (resp. \textit{nilpotent}) if $\adj_x:\mathfrak{g}\rightarrow\mathfrak{g}$ is semisimple (resp. nilpotent) as a vector space endomorphism.
\end{definition}
At first glance, there might appear to be other legitimate definitions. Suppose, for instance, that $\mathfrak{g}$ is explicitly presented as a subalgebra of $\mathfrak{gl}_n(\mathbb{C})$ for some $n$. Elements of $\mathfrak{g}$ are then $n\times n$ matrices, and one has the usual notions of ''semisimple" and ''nilpotent" for square matrices. Fortunately, these turn out to coincide with Definition \ref{Definition: semisimple and nilpotent} (see \cite[Cor. 20.4.3]{Tauvel}).
\begin{theorem}\label{Theorem: equivalence of semisimple / nilpotent}
If $\mathfrak{g}$ is a subalgebra of $\mathfrak{gl}_n(\mathbb{C})$, then $x\in\mathfrak{g}$ is semisimple (resp. nilpotent) in the sense of Definition \ref{Definition: semisimple and nilpotent} if and only if $x$ is semisimple (resp. nilpotent) as a matrix.
\end{theorem}
\begin{example}\label{Example: sl2 triple}
Suppose that $G=\SL_2(\mathbb{C})$, whose Lie algebra $\mathfrak{g}=\mathfrak{sl}_2(\mathbb{C})$ has the usual generators \begin{equation}\label{Equation: Usual sl2 generators}X=\begin{bmatrix}0 & 1\\ 0 & 0\end{bmatrix}, \quad H=\begin{bmatrix}1 & 0\\ 0 & -1\end{bmatrix},\quad Y=\begin{bmatrix}0 & 0\\ 1 & 0\end{bmatrix}.\end{equation} By Theorem \ref{Theorem: equivalence of semisimple / nilpotent}, it is immediate that $H$ is a semisimple element of $\mathfrak{sl}_2(\mathbb{C})$, while $X$ and $Y$ are nilpotent elements. The algebra $\mathfrak{sl}_2(\mathbb{C})$ and triple of elements $(X,H,Y)$ will play a substantial role in \ref{Subsection: The Jacobson-Morozov Theorem and sl2-triples}.
\end{example}
The relationship between semisimple (resp. nilpotent) elements of $\mathfrak{g}$ and semisimple (resp. nilpotent) matrices extends beyond Theorem \ref{Theorem: equivalence of semisimple / nilpotent}. For instance, recall that if an $n\times n$ matrix is both semisimple and nilpotent, it is necessarily zero. At the same time, if $x\in\mathfrak{g}$ is both semisimple and nilpotent, so too is the endomorphism $\text{ad}_x:\mathfrak{g}\rightarrow\mathfrak{g}$. It follows that $\adj_x=0$, and the injectivity of $\text{ad}:\mathfrak{g}\rightarrow\mathfrak{gl}(\mathfrak{g})$ implies $x=0$. For a second comparison, note that an $n\times n$ matrix is semisimple (resp. nilpotent) if and only if it is conjugate to a diagonal (resp. strictly upper-triangular) matrix. Dually, $x\in\mathfrak{g}$ is semisimple (resp. nilpotent) if and only if there exists $g\in G$ satisfying $\Adj_g(x)\in\mathfrak{t}$ (resp. $\Adj_g(x)\in\bigoplus_{\alpha\in\Delta_{+}}\mathfrak{g}_{\alpha}$). Thirdly, one knows that every $n\times n$ matrix $A$ gives rise to unique $n\times n$ matrices $S$ and $N$ such that $S$ is semisimple, $N$ is nilpotent, $A=S+N$, and $SN-NS=0$. The analogous fact is that for every $x\in\mathfrak{g}$, there exist unique elements $x_s,x_n\in\mathfrak{g}$ for which $x_s$ is semisimple, $x_n$ is nilpotent, $x=x_s+x_n$, and $[x_s,x_n]=0$. The expression $x=x_s+x_n$ is called the \textit{Jordan decomposition} of $x$, while $x_s$ and $x_n$ are called the semisimple and nilpotent parts of $x$, respectively.
Let us turn to a more geometric discussion. Denote by $\mathfrak{g}_{\text{ss}}$ and $\mathcal{N}$ the subsets of semisimple and nilpotent elements of $\mathfrak{g}$, respectively. The latter is called the \textit{nilpotent cone}, in reference to the fact that $\mathcal{N}$ is invariant under the dilation action of $\mathbb{C}^*$ on $\mathfrak{g}$. In the interest of preparing for later sections of this article, let us mention a few preliminary facts concerning $\mathfrak{g}_{ss}$ and $\mathcal{N}$. Firstly, since the semisimple endomorphisms are open and dense in $\mathfrak{gl}(\mathfrak{g})$, it follows that $\mathfrak{g}_{\text{ss}}$ is an open dense subvariety of $\mathfrak{g}$. One analogously sees that $\mathcal{N}$ is a closed subvariety of $\mathfrak{g}$. Secondly, noting that $\Adj_g$ is a Lie algebra automorphism for each $g\in G$, we conclude that $x\in\mathfrak{g}$ is semisimple (resp. nilpotent) if and only if $\Adj_g(x)$ is semisimple (resp. nilpotent). This is equivalent to the observation that both $\mathfrak{g}_{\text{ss}}$ and $\mathcal{N}$ are invariant under the adjoint representation.
\begin{example}
Let $G=\SL_2(\mathbb{C})$. By Theorem \ref{Theorem: equivalence of semisimple / nilpotent}, $A\in\mathfrak{sl}_2(\mathbb{C})$ is nilpotent if and only if it is nilpotent as a matrix. Equivalently, the characteristic polynomial $\det(\lambda I_2-A)$ is $\lambda^2$, which is the case if and only if $\det(A)=0$ (the condition $\trace(A)=0$ being automatic). In other words, we have
$$\mathcal{N}=\left\{\begin{bmatrix} x & y \\ z & -x\end{bmatrix}\in\mathfrak{sl}_2(\mathbb{C}):x^2+yz=0\right\}.$$
\end{example}
\subsection{The Jacobson-Morozov Theorem and $\mathfrak{sl}_2(\mathbb{C})$-triples}\label{Subsection: The Jacobson-Morozov Theorem and sl2-triples}
Let $(X,H,Y)$ be the triple of matrices in $\mathfrak{sl}_2(\mathbb{C})$ from Example \ref{Example: sl2 triple}. These matrices form a $\mathbb{C}$-basis of $\mathfrak{sl}_2(\mathbb{C})$ and satisfy the following bracket relations:
\begin{equation}\label{Equation: sl2 relations}
[X,Y]=H,\quad [H,X]=2X, \quad [H,Y]=-2Y.
\end{equation}
More generally, one calls a triple $(x,h,y)$ of points in $\mathfrak{g}$ an $\mathfrak{sl}_2(\mathbb{C})$-\textit{triple} if \eqref{Equation: sl2 relations} holds in $\mathfrak{g}$ when $X$, $H$, and $Y$ are replaced with $x$, $h$, and $y$, respectively. Such triples are in bijective correspondence with the Lie algebra morphisms $\phi:\mathfrak{sl}_2(\mathbb{C})\rightarrow\mathfrak{g}$, where $\phi\in\Hom(\mathfrak{sl}_2(\mathbb{C}),\mathfrak{g})$ identifies with $(\phi(X),\phi(H),\phi(Y))$. Since any $\phi\in\Hom(\mathfrak{sl}_2(\mathbb{C}),\mathfrak{g})$ is either zero or injective (a consequence of $\mathfrak{sl}_2(\mathbb{C})$ having no non-trivial ideals), there are exactly two possibilities for an $\mathfrak{sl}_2(\mathbb{C})$-triple $(x,h,y)$. Either $x=h=y=0$ or $\text{span}_{\mathbb{C}}\{x,h,y\}$ is a subalgebra of $\mathfrak{g}$ isomorphic to $\mathfrak{sl}_2(\mathbb{C})$.
\begin{example}\label{Example: Triple}
Let $\alpha\in\Pi$ be a simple root and $h_{\alpha}\in[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]$ the corresponding simple coroot (as defined in \ref{Subsection: The basic objects}). If $e_{\alpha}\in\mathfrak{g}_{\alpha}$ and $e_{-\alpha}\in\mathfrak{g}_{-\alpha}$ are chosen so that $h_{\alpha}=[e_{\alpha},e_{-\alpha}]$, then one can show $(e_{\alpha},h_{\alpha},e_{-\alpha})$ to be an $\mathfrak{sl}_2(\mathbb{C})$-triple.
\end{example}
\begin{example}\label{Example: Kostant triple}
In the interest of preparing for discussions to come (particularly \ref{Subsection: The regular nilpotent orbit}), we now consider a more complicated example (cf. \cite[Lemma 5.2]{Kostant}). To this end, since $\Pi$ is a basis of $\mathfrak{t}^*$, we may define $\{\epsilon_{\alpha}\}_{\alpha\in\Pi}$ to be the basis of $\mathfrak{t}$ satisfying
\begin{equation}\label{Equation: Dual basis}
\alpha(\epsilon_{\beta}) =
\begin{cases}
2 & \text{if } \alpha=\beta \\
0 & \text{if } \alpha\neq\beta
\end{cases}
\end{equation}
for all $\alpha,\beta\in\Pi$. Now consider the semisimple element
\begin{equation}\label{Equation: Regular semisimple element}h:=\sum_{\alpha\in\Pi}\epsilon_{\alpha}\in\mathfrak{t}.\end{equation} Noting that $\{h_{\alpha}\}_{\alpha\in\Pi}$ is a basis of $\mathfrak{t}$ (see \ref{Subsection: The basic objects}), there exist unique coefficients $c_{\alpha}\in\mathbb{C}$, $\alpha\in\Pi$, such that
$$h=\sum_{\alpha\in\Pi}c_{\alpha}h_{\alpha}.$$ We now define $\xi,\eta\in\mathfrak{g}$ by
$$\xi:=\sum_{\alpha\in\Pi}e_{\alpha}$$ and
$$\eta:=\sum_{\alpha\in\Pi}c_{\alpha}e_{-\alpha},$$ respectively, where for each $\alpha\in\Pi$, $e_{\alpha}\in\mathfrak{g}_{\alpha}$ and $e_{-\alpha}\in\mathfrak{g}_{-\alpha}$ are chosen such that $h_{\alpha}=[e_{\alpha},e_{-\alpha}]$ (as in Example \ref{Example: Triple}). Furthermore, we claim that $(\xi,h,\eta)$ is an $\mathfrak{sl}_2(\mathbb{C})$-triple. To see this, first note that \begin{equation}\label{Equation: Technical calculation}[\xi,\eta]=\sum_{\alpha,\beta\in\Pi}c_{\beta}[e_{\alpha},e_{-\beta}].\end{equation} If $\alpha,\beta\in\Pi$ are distinct, then the discussion of simple roots in \ref{Subsection: The basic objects} implies that $\alpha-\beta$ cannot be a root. Since $\alpha-\beta$ is also non-zero in this case, $[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\beta}]=\{0\}$ (see \ref{Subsection: The basic objects}). In particular, $[e_{\alpha},e_{-\beta}]=0$ when $\alpha$ and $\beta$ are distinct. At the same time, we know that $[e_{\alpha},e_{-\alpha}]=h_{\alpha}$ for all $\alpha\in\Pi$. Hence \eqref{Equation: Technical calculation} can be written as $$[\xi,\eta]=\sum_{\alpha\in\Pi}c_{\alpha}h_{\alpha}=h.$$ Now computing $[h,\xi]$, we have
$$[h,\xi]=\sum_{\alpha\in\Pi}[h,e_{\alpha}]=\sum_{\alpha\in\Pi}\alpha(h)e_{\alpha}.$$ The description \eqref{Equation: Regular semisimple element} of $h$ implies that $\alpha(h)=2$ for all $\alpha\in\Pi$, so that
$$[h,\xi]=\sum_{\alpha\in\Pi}2e_{\alpha}=2\xi.$$ The verification of $[h,\eta]=-2\eta$ is similar, since $-\alpha(h)=-2$ for all $\alpha\in\Pi$. It follows that $(\xi,h,\eta)$ is indeed an $\mathfrak{sl}_2(\mathbb{C})$-triple.
\end{example}
We will benefit from understanding $\mathfrak{sl}_2(\mathbb{C})$-triples in the context of the \textit{Jacobson-Morozov Theorem}, a crucial result that will feature prominently in \ref{Subsection: Some first results}. To motivate this theorem, suppose that $(x,h,y)$ is the $\mathfrak{sl}_2(\mathbb{C})$-triple corresponding to a non-zero (ie. injective) $\phi\in\Hom(\mathfrak{sl}_2(\mathbb{C}),\mathfrak{g})$. It follows that $\phi$ defines an isomorphism with its image $\text{span}_{\mathbb{C}}\{x,h,y\}$, so that $\phi$ sends nilpotent elements of $\mathfrak{sl}_2(\mathbb{C})$ to nilpotent elements of $\text{span}_{\mathbb{C}}\{x,h,y\}$. In particular, $x=\phi(X)$ is nilpotent in $\text{span}_{\mathbb{C}}\{x,h,y\}$. If one chooses a Lie algebra embedding $\mathfrak{g}\subseteq\mathfrak{gl}_n(\mathbb{C})$\footnote{Since $\mathfrak{g}$ is semisimple, such an embedding always exists. For instance, one may identify $\mathfrak{g}$ with its image under the (injective) adjoint representation $\adj:\mathfrak{g}\rightarrow\mathfrak{gl}(\mathfrak{g})\cong\mathfrak{gl}_n(\mathbb{C})$, $n=\dim(\mathfrak{g})$.}, then $\text{span}_{\mathbb{C}}\{x,h,y\}$ becomes a subalgebra of $\mathfrak{gl}_n(\mathbb{C})$ and Theorem \ref{Theorem: equivalence of semisimple / nilpotent} implies that $x$ is nilpotent matrix. By applying Theorem \ref{Theorem: equivalence of semisimple / nilpotent} again, this time to the subalgebra $\mathfrak{g}\subseteq\mathfrak{gl}_n(\mathbb{C})$, we see that $x$ is nilpotent in $\mathfrak{g}$. This is clearly also true when $\phi=0$ (in which case $x=0$), so we have shown $x\in\mathfrak{g}$ to be nilpotent whenever it appears as the first element in an $\mathfrak{sl}_2(\mathbb{C})$-triple. For a converse, one has the Jacobson-Morozov Theorem.
\begin{theorem}[The Jacobson-Morozov Theorem; \text{cf. \cite[Thm. 3.7.1]{Chriss}}]
If $x\in\mathfrak{g}$ is nilpotent, then there exist $h,y\in\mathfrak{g}$ such that $(x,h,y)$ is an $\mathfrak{sl}_2(\mathbb{C})$-triple.
\end{theorem}
\section{General adjoint orbits}\label{Section: General adjoint orbits}
Equipped with Section \ref{Section: Preliminaries on Lie theory}, we can apply parts of Section \ref{Section: Preliminaries on algebraic group actions} to study a distinguished class of algebraic group orbits -- adjoint orbits. Our overarching objective is to give the background and context necessary for the more specialized discussions of semisimple and nilpotent adjoint orbits (Sections \ref{Section: Semisimple orbits} and \ref{Section: Nilpotent orbits}, respectively). Nevertheless, we will see that general adjoint orbits are interesting objects of study in Lie theory and geometry (see \ref{Subsection: Geometric features}, for example).
\subsection{Definitions and conventions}\label{Subsection: Definitions and conventions}
Following the framework of Example \ref{Example: action from representation}, the adjoint representation $\Adj:G\rightarrow\GL(\mathfrak{g})$ induces a $G$-variety structure on $\mathfrak{g}$. One calls the resulting action the \textit{adjoint action} and its orbits \textit{adjoint orbits}.
\begin{example}\label{Example: Type A adjoint orbits}
Let $G=\SL_n(\mathbb{C})$. Referring to \eqref{Equation: Type A adjoint representation}, we see that the adjoint orbits of $\SL_n(\mathbb{C})$ are precisely the conjugacy classes of the traceless $n\times n$ matrices. Note that the term ``conjugacy class'' is unambiguous, since two $n\times n$ matrices are conjugate under $\SL_n(\mathbb{C})$ if and only if they are conjugate under $\GL_n(\mathbb{C})$. With this point in mind, Example \ref{Example: Conjugation} describes several adjoint $\SL_n(\mathbb{C})$-orbits.\footnote{To ensure that $\mathcal{O}(A)$ from Example \ref{Example: Conjugation} is an adjoint $\SL_n(\mathbb{C})$-orbit, one must impose the extra condition $\trace(A)=0$.}
\end{example}
Now, recall the discussion of stabilizers from \ref{Subsection: Orbits as quotient varieties}. Given $x\in\mathfrak{g}$, we shall always understand $C_G(x)$ as referring to the $G$-stabilizer of $x$ under the adjoint action, ie. $$C_G(x):=\{g\in G:\Adj_g(x)=x\},\quad x\in\mathfrak{g}.$$ The following fact about its Lie algebra, $C_{\mathfrak{g}}(x)\subseteq\mathfrak{g}$, will be used repeatedly and (in the interest of parsimony) without explicit indication that it is being used.
\begin{lemma}[\text{cf. \cite[Sect. 2.2]{Jantzen}}]
If $x\in\mathfrak{g}$, then $C_{\mathfrak{g}}(x)=\{y\in\mathfrak{g}:[x,y]=0\}$.
\end{lemma}
\subsection{Dimension and regularity}\label{Subsection: Dimension and regularity}
In the interest of later sections, we now take a moment to discuss the dimensions of adjoint orbits. One might begin with the observation that each adjoint orbit can have dimension at most $\dim(G)$. However, this dimension bound turns out to be highly sub-optimal. To improve it, we note that $\dim(C_G(x))\geq\text{rank}(G)$ for all $x\in\mathfrak{g}$ (see \cite[Sect. 19.7, Sect. 29.3]{Tauvel}\footnote{Strictly speaking, this reference bounds the dimension of $C_{\mathfrak{g}}(x)$ instead of $C_G(x)$. This is, of course, equivalent to the dimension inequality stated here.}). It follows that
$$\dim(\mathcal{O}(x))=\dim(G/C_G(x))=\dim(G)-\dim(C_G(x))\leq\dim(G)-\text{rank}(G)$$
for all $x\in\mathfrak{g}$, so that $\dim(G)-\text{rank}(G)$ is our new upper bound on adjoint orbit dimensions. This bound is sharp in the sense that $\mathfrak{g}$ always contains adjoint orbits of dimension $\dim(G)-\text{rank}(G)$. We defer the proof of this fact to \ref{Subsection: Stabilizer descriptions} (see Corollary \ref{Corollary: The existence of regular orbits}), which is completely independent of the present section.
One calls an adjoint orbit $\mathcal{O}\subseteq\mathfrak{g}$ \textit{regular} if $\dim(\mathcal{O})=\dim(G)-\text{rank}(G)$.
\subsection{Some first results}\label{Subsection: Some first results}
We now gather some initial results on adjoint orbits, emphasizing connections to semisimple/nilpotent elements and the Jordan decomposition (discussed in \ref{Subsection: semisimple and nilpotent elements}). The facts we establish here will be essential to Sections \ref{Section: Semisimple orbits} and \ref{Section: Nilpotent orbits}.
\begin{proposition}\label{Proposition: Dilating the nilpotent part}
If $\mathcal{O}\subseteq\mathfrak{g}$ is an adjoint orbit and $x\in\mathcal{O}$ has Jordan decomposition $x=x_s+x_n$, then $x_s+ax_n\in\mathcal{O}$ for all $a\in\mathbb{C}^*$.
\end{proposition}
\begin{proof}
Since $[x_s,x_n]=0$, it follows that $x_n\in C_{\mathfrak{g}}(x_s)$. However, $C_{\mathfrak{g}}(x_s)$ is known to be a reductive subalgebra of $\mathfrak{g}$ (see \cite[Prop. 20.5.13]{Tauvel}), and one can then deduce that $x_n\in [C_{\mathfrak{g}}(x_s),C_{\mathfrak{g}}(x_s)]$ (see \cite[Prop. 20.5.14]{Tauvel}). Also, the fact that $[C_{\mathfrak{g}}(x_s),C_{\mathfrak{g}}(x_s)]$ is reductive implies that $[C_{\mathfrak{g}}(x_s),C_{\mathfrak{g}}(x_s)]$ is a semisimple subalgebra. The element $x_n$ is nilpotent in $[C_{\mathfrak{g}}(x_s),C_{\mathfrak{g}}(x_s)]$, as one can see by embedding $\mathfrak{g}$ (hence also $[C_{\mathfrak{g}}(x_s),C_{\mathfrak{g}}(x_s)]$) into $\mathfrak{gl}_n(\mathbb{C})$ and applying Theorem \ref{Theorem: equivalence of semisimple / nilpotent}. By the Jacobson-Morozov Theorem, there exists $h\in [C_{\mathfrak{g}}(x_s),C_{\mathfrak{g}}(x_s)]$ for which $[h,x_n]=2x_n$. Now for all $c\in\mathbb{C}$, we have
\begin{equation}\label{Equation: First dilation invariance}\Adj_{\exp(ch)}(x)=\exp(\adj_{ch})(x)=\exp(\adj_{ch})(x_s)+\exp(\adj_{ch})(x_n)
\end{equation}
where $\exp(\adj_{ch})$ denotes the exponential of the endomorphism $\adj_{ch}:\mathfrak{g}\rightarrow\mathfrak{g}$. Observe that $ch\in C_{\mathfrak{g}}(x_s)$, so that $\exp(\adj_{ch})(x_s)=x_s$. Secondly, the condition $[ch,x_n]=2cx_n$ gives $\exp(\adj_{ch})(x_n)=e^{2c}x_n$. These two observations allow one to write \eqref{Equation: First dilation invariance} as
$$\Adj_{\exp(ch)}(x)=x_s+e^{2c}x_n.$$
In particular, $x_s+e^{2c}x_n\in\mathcal{O}$ for all $c\in\mathbb{C}$. To prove the proposition as stated, choose $c\in\mathbb{C}$ such that $e^{2c}=a$.
\end{proof}
By letting $a\rightarrow 0$ in Proposition \ref{Proposition: Dilating the nilpotent part}, one obtains the following corollary.
\begin{corollary}\label{Corollary: Semisimple part in closure}
If $\mathcal{O}\subseteq\mathfrak{g}$ is an adjoint orbit and $x\in\mathcal{O}$ has Jordan decomposition $x=x_s+x_n$, then $x_s\in\overline{\mathcal{O}}$.
\end{corollary}
Our next results will make extensive use of $G$-invariant polynomials on $\mathfrak{g}$. To be more precise, note that the adjoint action induces an action of $G$ on $\mathbb{C}[\mathfrak{g}]$ (cf. Example \ref{Example: Induced action on functions}), the algebra of regular (ie. polynomial) functions on $\mathfrak{g}$. Let $\mathbb{C}[\mathfrak{g}]^G\subseteq\mathbb{C}[\mathfrak{g}]$ denote the subalgebra of all $G$-invariant polynomial functions, meaning
$$\mathbb{C}[\mathfrak{g}]^G:=\{f\in\mathbb{C}[\mathfrak{g}]:g\cdot f=f\text{ for all }g\in G\}.$$ Analogously, the $W$-action on $\mathfrak{t}$ (see \ref{Subsection: The basic objects}) gives rise to a $W$-action on the polynomial functions $\mathbb{C}[\mathfrak{t}]$, and to a subalgebra $\mathbb{C}[\mathfrak{t}]^W\subseteq\mathbb{C}[\mathfrak{t}]$ of $W$-invariant polynomial functions. One can verify that $G$-invariant polynomials on $\mathfrak{g}$ restrict to $W$-invariant polynomials on $\mathfrak{t}$, and it turns out that this restriction process defines an algebra isomorphism \begin{equation}\label{Equation: Chevalley restriction}
\mathbb{C}[\mathfrak{g}]^G\xrightarrow{\cong}\mathbb{C}[\mathfrak{t}]^W
\end{equation}
(see \cite[Thm. 3.1.38]{Chriss}).
Moreover, $\mathbb{C}[\mathfrak{g}]^G$ (hence also $\mathbb{C}[\mathfrak{t}]^W$) is known to be freely generated as a commutative $\mathbb{C}$-algebra by $r=\text{rank}(G)$ homogeneous polynomials (see \cite[Sect. 3.3]{KostantPoly}).
\begin{proposition}\label{Proposition: Invariant functions have constant values}
If $x,y\in\mathfrak{g}$ are semisimple, then $\mathcal{O}(x)=\mathcal{O}(y)$ if and only if $f(x)=f(y)$ for all $f\in\mathbb{C}[\mathfrak{g}]^G$.
\end{proposition}
\begin{proof}
If $\mathcal{O}(x)=\mathcal{O}(y)$, then $y=\Adj_g(x)$ for some $g\in G$ and $f(y)=f(\Adj_g(x))=f(x)$ for all $f\in\mathbb{C}[\mathfrak{g}]^G$. Conversely, suppose that $f(x)=f(y)$ for all $f\in\mathbb{C}[\mathfrak{g}]^G$. Since $x$ and $y$ are semisimple, there exist $h_x,h_y\in\mathfrak{t}$ such that $\mathcal{O}(h_x)=\mathcal{O}(x)$ and $\mathcal{O}(h_y)=\mathcal{O}(y)$ (see the discussion in \ref{Subsection: semisimple and nilpotent elements}). Also, repeating the argument from the first sentence of our proof, we see $f(h_x)=f(x)$ and $f(h_y)=f(y)$ for all $f\in\mathbb{C}[\mathfrak{g}]^G$. We may therefore assume that $x,y\in\mathfrak{t}$ when proving our converse. It then follows from \eqref{Equation: Chevalley restriction} that $f(x)=f(y)$ for all $f\in\mathbb{C}[\mathfrak{t}]^W$. Using Lemma 34.2.1 of \cite{Tauvel}, we conclude that $y=w\cdot x$ for some $w\in W$. After lifting $w$ to a representative $g\in N_G(T)$, this statement becomes $y=\Adj_g(x)$. Hence $\mathcal{O}(x)=\mathcal{O}(y)$.
\end{proof}
Focusing now on nilpotent elements and adjoint orbits, we have the following.
\begin{proposition}\label{Proposition: Characterizations of nilpotent elements}
If $x\in\mathfrak{g}$, then the following conditions are equivalent.
\begin{itemize}
\item[(i)] $x$ is nilpotent.
\item[(ii)] For all $a\in\mathbb{C}^*$, $ax\in\mathcal{O}(x)$.
\item[(iii)] $0\in\overline{\mathcal{O}(x)}$.
\item[(iv)] For all $f\in\mathbb{C}[\mathfrak{g}]^G$, $f(x)=f(0)$.
\end{itemize}
\end{proposition}
\begin{proof}
We will verify the implications (i)$\Rightarrow$(ii)$\Rightarrow$(iii)$\Rightarrow$(iv)$\Rightarrow$(i).
\noindent\underline{(i)$\Rightarrow$(ii):} Note that $x_s=0$ and $x_n=x$, so that (ii) follows immediately from Proposition \ref{Proposition: Dilating the nilpotent part}.
\noindent\underline{(ii)$\Rightarrow$(iii):} Since $ax\in\mathcal{O}(x)$ for all $a\in\mathbb{C}^*$, letting $a\rightarrow 0$ establishes that $0\in\overline{\mathcal{O}(x)}$.
\noindent\underline{(iii)$\Rightarrow$(iv):} If $f\in\mathbb{C}[\mathfrak{g}]^G$, then $f(\Adj_g(x))=f(x)$ for all $g\in G$. In other words, $f$ takes the constant value $f(x)$ on $\mathcal{O}(x)$. By continuity, $f$ actually takes the constant value $f(x)$ on $\overline{\mathcal{O}(x)}$. Since $0\in\overline{\mathcal{O}(x)}$, we see that $f(x)=f(0)$.
\noindent\underline{(iv)$\Rightarrow$(i):} For each positive integer $n$, consider the function $\phi_n\in\mathbb{C}[\mathfrak{g}]$ defined by $\phi_n(y)=\trace((\adj_y)^n)$. To see that $\phi_n\in\mathbb{C}[\mathfrak{g}]^G$, suppose that $g\in G$ and $y,z\in\mathfrak{g}$. We have $\adj_{\Adj_g(y)}(z)=[\Adj_g(y),z]=\Adj_g([y,(\Adj_{g})^{-1}(z)])$, meaning that $\adj_{\Adj_g(y)}=\Adj_g\circ\adj_y\circ(\Adj_g)^{-1}$. Hence, $$\phi_n(\Adj_g(y))=\trace((\adj_{\Adj_g(y)})^n)=\trace(\Adj_g\circ(\adj_y)^n\circ(\Adj_g)^{-1})=\trace((\adj_y)^n)=\phi_n(y),$$ and we conclude that $\phi_n\in \mathbb{C}[\mathfrak{g}]^G$. Now by hypothesis, $$\trace((\adj_x)^n)=\phi_n(x)=\phi_n(0)=0$$ for all positive integers $n$. It follows that $\adj_x$ is a nilpotent endomorphism, as desired.
\end{proof}
\subsection{The adjoint quotient}\label{Subsection: The adjoint quotient}
Adjoint orbits feature prominently in a very natural and well-studied fibration whose total space is $\mathfrak{g}$. To construct it, one can appeal to the basics of geometric invariant theory. The inclusion $\mathbb{C}[\mathfrak{g}]^G\hookrightarrow\mathbb{C}[\mathfrak{g}]$ corresponds to a $G$-invariant surjective map of affine varieties,
\begin{equation}\label{Equation: The adjoint quotient}
\Phi:\mathfrak{g}\rightarrow\Spec(\mathbb{C}[\mathfrak{g}]^G),
\end{equation}
(see \cite[Sect. 1.4.1, Sect. 1.4.2]{Schmitt}),
called the \textit{adjoint quotient}. Note $\Spec(\mathbb{C}[\mathfrak{g}]^G)$ refers to the maximal ideal spectrum of $\mathbb{C}[\mathfrak{g}]^G$, while $G$-invariance is the condition that $\phi\circ\Adj_g=\phi$ for all $g\in G$.
One can say a great deal more about \eqref{Equation: The adjoint quotient} by invoking the discussion of $\mathbb{C}[\mathfrak{g}]^G$ in \ref{Subsection: Some first results}. Indeed, recall that $\mathbb{C}[\mathfrak{g}]^G$ is freely generated by $r:=\text{rank}(G)$ homogeneous generators, $\chi_1,\chi_2,\ldots,\chi_r\in\mathbb{C}[\mathfrak{g}]^G$. We may therefore identify $\Spec(\mathbb{C}[\mathfrak{g}]^G)$ with $\mathbb{C}^r$ and re-write $\Phi$ as the map
\begin{equation}\label{Equation: The second adjoint quotient}
\Phi:\mathfrak{g}\rightarrow\mathbb{C}^r,\quad \xi\mapsto(\chi_1(\xi),\chi_2(\xi),\ldots,\chi_r(\xi)).
\end{equation}
Kostant studied this form of the adjoint quotient in detail (see \cite{KostantPoly}). More specifically, note that the $G$-invariance of $\Phi$ is equivalent to each fibre of $\Phi$ being a union of adjoint orbits. With this in mind, Kostant gave the following description of each fibre's decomposition into adjoint orbits (see \cite[Thm. 0.6, Thm. 0.7]{KostantPoly}).
\begin{theorem}\label{Theorem: Kostant's theorem}
If $z\in\Spec(\mathbb{C}[\mathfrak{g}]^G)$, then $\Phi^{-1}(z)$ is a union of finitely many adjoint orbits and contains a unique regular orbit. This regular orbit is open and dense in $\Phi^{-1}(z)$.
\end{theorem}
\begin{corollary}
If $\mathcal{O}\subseteq\mathfrak{g}$ is an adjoint orbit, then $\overline{\mathcal{O}}$ is a union of finitely many adjoint orbits.
\end{corollary}
\begin{proof}
Choose a point $x\in\mathcal{O}$. Since the fibres of $\Phi$ are $G$-invariant, it follows that $\mathcal{O}$ belongs to $\Phi^{-1}(\Phi(x))$. This fibre is closed, meaning that $\overline{\mathcal{O}}\subseteq\Phi^{-1}(\Phi(x))$. Theorem \ref{Theorem: Kostant's theorem} implies that $\Phi^{-1}(\Phi(x))$ is a union of finitely many adjoint orbits, so that the same must be true of $\overline{\mathcal{O}}$.
\end{proof}
\subsection{Geometric features}\label{Subsection: Geometric features}
We now turn our attention to some more inherently geometric aspects of adjoint orbits. However, to avoid the lengthy digressions needed to properly motivate certain definitions, we will sometimes not define the geometric structures under consideration (ex. Poisson and hyperk\"{a}hler structures). In such cases, we will give suitable references to the research literature.
Consider the \textit{coadjoint representation} $\Adj^*:G\rightarrow\GL(\mathfrak{g}^*)$, $g\mapsto\Adj_{g}^*$, defined as follows:
$$(\Adj_g^*(\phi))(x):=\phi(\Adj_{g^{-1}}(x)),\quad g\in G,\text{ }\phi\in\mathfrak{g}^*\text{ },x\in\mathfrak{g}.$$
This representation gives $\mathfrak{g}^*$ the structure of a $G$-variety, and the resulting orbits are called \textit{coadjoint orbits}. Moreover, $\mathfrak{g}^*$ is known to carry a canonical holomorphic Poisson structure whose symplectic leaves are precisely the coadjoint orbits (see \cite[Sect. 1.2, 1.3]{Chriss}). It follows that every coadjoint orbit inherits a holomorphic symplectic form, namely a closed, non-degenerate, holomorphic $2$-form (see \cite[Sect. 1.1]{Chriss} for further details on symplectic forms). One call this particular symplectic form the \textit{Kirillov-Kostant-Souriau form}.
Now, recall that the Killing form induces a vector space isomorphism $\mathfrak{g}\cong\mathfrak{g}^*$. Noting that the Killing form is $G$-invariant (see \ref{Subsection: The basic objects}), one can verify that our isomorphism is $G$-equivariant. Hence, in addition to identifying $\mathfrak{g}$ with $\mathfrak{g}^*$, the isomorphism identifies adjoint orbits with coadjoint orbits. We may therefore transfer the above-mentioned geometric structures on $\mathfrak{g}^*$ and coadjoint orbits to $\mathfrak{g}$ and adjoint orbits, respectively. The Lie algebra $\mathfrak{g}$ then has a canonical holomorphic Poisson structure and its symplectic leaves are exactly the adjoint orbits. In particular, every adjoint orbit $\mathcal{O}\subseteq\mathfrak{g}$ carries a preferred holomorphic symplectic form, $\omega_{\mathcal{O}}\in\Omega^2(\mathcal{O})$. To describe it, fix $x\in\mathcal{O}$ and identify the tangent space $T_x\mathcal{O}$ with $\mathfrak{g}/C_{\mathfrak{g}}(x)$ via \eqref{Equation: Tangent space description}, so that the restriction of $\omega_{\mathcal{O}}$ to $x$ is a bilinear form on $\mathfrak{g}/C_{\mathfrak{g}}(x)$. We then have
\begin{equation}\label{Equation: The canonical symplectic form}
\omega_{\mathcal{O}}\vert_x:\mathfrak{g}/C_{\mathfrak{g}}(x)\otimes\mathfrak{g}/C_{\mathfrak{g}}(x)\rightarrow\mathbb{C},\quad\omega_{\mathcal{O}}\vert_x([y],[z]):=\langle x,[y,z]\rangle\end{equation}
for all $[y],[z]\in\mathfrak{g}/C_{\mathfrak{g}}(x)$.
Thus far, our treatment of adjoint orbits and geometry has been classical. However, there is a more modern and elaborate framework within which to appreciate the canonical symplectic forms described above -- \textit{hyperk\"{a}hler geometry}. It turns out that each adjoint orbit is a hyperk\"{a}hler manifold (see \cite{Dancer} for a definition) whose underlying holomorphic symplectic form is the canonical one we have just described. This result was proved by Kronheimer \cite{KronheimerInstantons,KronheimerSemisimple} in important special cases and by Biquard \cite{Biquard} and Kovalev \cite{Kovalev} in full generality. The key is to identify each adjoint orbit with a certain moduli space of solutions to \textit{N\"{a}hm's equations} (see \cite[Sect. 2.2]{Bielawski}), known to be a hyperk\"{a}hler manifold, and then transfer the hyperk\"{a}hler structure over to the orbit. It is nevertheless often difficult to describe this induced hyperk\"{a}hler structure in explicit terms (see \cite[Sect 2.3]{Bielawski}). This is one of several unresolved issues that have made and continue to make the hyperk\"{a}hler geometry of adjoint orbits an active area of research (see \cite{BielawskiReducible,Santa-Cruz,Villumsen,Kobak,BielawskiOnthe}).
\section{Semisimple Orbits}\label{Section: Semisimple orbits}
Informally speaking, Example \ref{Example: Type A adjoint orbits} shows adjoint orbits to be generalizations of matrix conjugacy classes. In what follows, we will study the adjoint orbits that generalize the conjugacy classes of diagonalizable matrices -- the \textit{semisimple orbits}.
\subsection{Definitions and characterizations}\label{Subsection: Definitions and characterizations}
Recall from \ref{Subsection: semisimple and nilpotent elements} that the subvariety of semisimple elements $\mathfrak{g}_{\text{ss}}\subseteq\mathfrak{g}$ is invariant under the adjoint representation of $G$. One may rephrase this $G$-invariance in the following way: an adjoint orbit contains a semisimple element if and only if it consists of semisimple elements. The adjoint orbits satisfying these equivalent conditions are called the \textit{semisimple orbits}.
\begin{definition}\label{Definition: Semisimple orbit}
An adjoint orbit $\mathcal{O}\subseteq\mathfrak{g}$ is called a \textit{semisimple orbit} if $\mathcal{O}\cap\mathfrak{g}_{\text{ss}}\neq\emptyset$, or equivalently $\mathcal{O}\subseteq\mathfrak{g}_{\text{ss}}$.
\end{definition}
\begin{example}\label{Example: Type A semisimple orbits}
Let $G=\SL_n(\mathbb{C})$. Using Example \ref{Example: Type A adjoint orbits} and Theorem \ref{Theorem: equivalence of semisimple / nilpotent}, we see that the semisimple orbits in $\mathfrak{sl}_n(\mathbb{C})$ are precisely the conjugacy classes of semisimple (ie. diagonalizable) $n\times n$ matrices having zero trace.
\end{example}
While we shall take Definition \ref{Definition: Semisimple orbit} to be our definition of a semisimple orbit, there are alternatives. To motivate one of these, recall that Example \ref{Example: Conjugation} shows the conjugacy class of a diagonal $n\times n$ matrix with pairwise distinct eigenvalues to be closed in $\text{Mat}_{n\times n}(\mathbb{C})$. Now suppose that the matrix in question lies in $\mathfrak{sl}_n(\mathbb{C})$. Its $\SL_n(\mathbb{C})$-adjoint orbit is semisimple (see Example \ref{Example: Type A semisimple orbits}), and the argument from Example \ref{Example: Conjugation} shows the orbit to be closed in $\mathfrak{sl}_n(\mathbb{C})$. At the same time, the matrices $N_1$ and $N_2$ from Example \ref{Example: Conjugation} belong to $\mathfrak{sl}_n(\mathbb{C})$. Their $\SL_n(\mathbb{C})$-adjoint orbits are not semisimple (see Example \ref{Example: Type A semisimple orbits} again) and the arguments from Example \ref{Example: Conjugation} show these orbits to be non-closed in $\mathfrak{sl}_n(\mathbb{C})$. Our findings here are an instance of a general fact: an adjoint orbit is closed if and only if it is semisimple. Before proving this, we present the following lemma.
\begin{lemma}\label{Lemma: Unique semisimple orbit in closure}
If $\mathcal{O}\subseteq\mathfrak{g}$ is an adjoint orbit, then there exists a unique semisimple orbit belonging to $\overline{\mathcal{O}}$.
\end{lemma}
\begin{proof}
Choose a point $x\in\mathcal{O}$. For existence, Corollary \ref{Corollary: Semisimple part in closure} implies that $x_s\in\overline{\mathcal{O}}$. It follows that $\mathcal{O}(x_s)$ is a semisimple orbit lying in $\overline{\mathcal{O}}$. As for uniqueness, suppose that $f\in\mathbb{C}[\mathfrak{g}]^G$. Since $f(\Adj_g(x))=f(x)$ for all $g\in G$, we see that $f$ is constant-valued on $\mathcal{O}$. By continuity, $f$ is actually constant-valued on $\overline{\mathcal{O}}$. In particular, any two semisimple $y,z\in\overline{\mathcal{O}}$ must satisfy $f(y)=f(z)$. By Proposition \ref{Proposition: Invariant functions have constant values}, $\mathcal{O}(y)=\mathcal{O}(z)$.
\end{proof}
\begin{theorem}\label{Theorem: Semisimple equivalent to closed}
An adjoint orbit $\mathcal{O}\subseteq\mathfrak{g}$ is semisimple if and only if it is closed in $\mathfrak{g}$.
\end{theorem}
\begin{proof}
Assume that $\mathcal{O}$ is semisimple and let $y\in\overline{\mathcal{O}}$ be given. It follows that $\mathcal{O}(y)\subseteq\overline{\mathcal{O}}$, or equivalently $\overline{\mathcal{O}(y)}\subseteq\overline{\mathcal{O}}$. Using this together with Corollary \ref{Corollary: Semisimple part in closure}, we see that $y_s\in\overline{\mathcal{O}}$. The orbit $\mathcal{O}(y_s)$ is then a semisimple orbit belonging to $\overline{\mathcal{O}}$, and Lemma \ref{Lemma: Unique semisimple orbit in closure} implies that $\mathcal{O}=\mathcal{O}(y_s)$. We already know that $\mathcal{O}(y)\subseteq\overline{\mathcal{O}}$ (ie. $\mathcal{O}(y)\leq\mathcal{O}$), while Corollary \ref{Corollary: Semisimple part in closure} gives $\mathcal{O}(y_s)\subseteq\overline{\mathcal{O}(y)}$ (ie. $\mathcal{O}=\mathcal{O}(y_s)\leq\mathcal{O}(y)$). Hence $\mathcal{O}(y)=\mathcal{O}$. This implies that $y\in\mathcal{O}$, and we conclude that $\mathcal{O}$ is closed.
Conversely, assume that $\mathcal{O}$ is closed in $\mathfrak{g}$ and choose a point $x\in\mathcal{O}$. Corollary \ref{Corollary: Semisimple part in closure} implies that $x_s\in\overline{\mathcal{O}}=\mathcal{O}$, meaning that $\mathcal{O}=\mathcal{O}(x_s)$ is a semisimple orbit.
\end{proof}
\begin{corollary}\label{Corollary: Semisimple orbits are affine}
If $\mathcal{O}\subseteq\mathfrak{g}$ is a semisimple orbit, then $\mathcal{O}$ is an affine variety.
\end{corollary}
\begin{proof}
This follows from the fact that $\mathcal{O}$ is closed in the affine variety $\mathfrak{g}$.
\end{proof}
\subsection{Stabilizer descriptions}\label{Subsection: Stabilizer descriptions}
Recall from \ref{Subsection: semisimple and nilpotent elements} that $x\in\mathfrak{g}$ is semisimple if and only if $\Adj_g(x)\in\mathfrak{t}$ for some $g\in G$. It follows that an adjoint orbit is semisimple if and only if it is of the form $\mathcal{O}(h)$ for $h\in\mathfrak{t}$. Now given $h_1,h_2\in\mathfrak{t}$, one can show that $\mathcal{O}(h_1)=\mathcal{O}(h_2)$ if and only if $h_1$ and $h_2$ belong to the same $W$-orbit in $\mathfrak{t}$. We conclude that $[h]\mapsto \mathcal{O}(h)$ defines a bijection from $\mathfrak{t}/W$ to the set of semisimple orbits. At the same time, each equivalence class in $\mathfrak{t}/W$ has a unique representative belonging to \begin{equation}\label{Equation: Fundamental domain}\mathcal{D}:=\{h\in\mathfrak{t}:\forall\alpha\in\Pi,\text{ Re}(\alpha(h))\geq 0\text{ and }\text{Re}(\alpha(h))=0\Longrightarrow\text{Im}(\alpha(h))\geq 0\}
\end{equation}
(see \cite{Collingwood}, Section 2.2). Our parametrization of semisimple orbits therefore takes the form $h\mapsto\mathcal{O}(h)$, $h\in\mathcal{D}$.
Given $h\in\mathcal{D}$, we have the following instance of the orbit-stabilizer isomorphism \eqref{Equation: Orbit-stabilizer isomorphism}:
\begin{equation}\label{Equation: semisimple orbit-stabilizer isomorphism}
\mathcal{O}(h)\cong G/C_G(h).
\end{equation}
In the interest of using \eqref{Equation: semisimple orbit-stabilizer isomorphism} to study semisimple orbits, we turn our attention to the structure of $C_G(h)$. To this end, set
$$\Pi(h):=\{\alpha\in\Pi:\alpha(h)=0\}.$$ Recall that $\Pi(h)$ determines a standard parabolic subgroup $P_{\Pi(h)}$ with Levi decomposition $P_{\Pi(h)}=U_{\Pi(h)}\rtimes L_{\Pi(h)}$, as explained in \ref{Subsection: The basic objects}.
\begin{proposition}\label{Proposition: Stabilizer structure}
If $h\in\mathcal{D}$, then $C_G(h)=L_{\Pi(h)}$.
\end{proposition}
\begin{proof}
To begin, note that $P_{\Pi(h)}$ is connected (see \cite[Thm. 28.4.2]{Tauvel}). It follows that the quotient $P_{\Pi(h)}/U_{\Pi(h)}\cong L_{\Pi(h)}$ is also connected. At the same time, $C_G(h)$ is known to be connected (see \cite[Thm. 2.3.3]{Collingwood}). Proving $C_G(h)=L_{\Pi(h)}$ therefore amounts to establishing the equality of these groups' respective Lie algebras $C_{\mathfrak{g}}(H)$ and $\mathfrak{l}_{\Pi(h)}$. To this end, as $\mathfrak{t}$ is abelian and $h\in\mathfrak{t}$, we have $\mathfrak{t}\subseteq C_{\mathfrak{g}}(h)$. It follows that $C_{\mathfrak{g}}(h)$ decomposes as a direct sum of $\mathfrak{t}$ and certain root spaces. To determine these root spaces note that $\alpha\in\Delta$ satisfies $\mathfrak{g}_{\alpha}\subseteq C_{\mathfrak{g}}(h)$ if and only if $\alpha(h)=0$. Hence, $$C_{\mathfrak{g}}(h)=\mathfrak{t}\oplus\bigoplus_{\substack{\alpha\in\Delta\\ \alpha(h)=0}}\mathfrak{g}_{\alpha}.$$ Using this together with \eqref{Equation: Lie algebra of Levi factor}, we are reduced to proving that \begin{equation}\label{Equation: Reduced to proving}
\Delta_{\Pi(h)}=\{\alpha\in\Delta:\alpha(h)=0\}.
\end{equation}
The inclusion $\Delta_{\Pi(h)}\subseteq\{\alpha\in\Delta:\alpha(h)=0\}$ clearly holds. Conversely, suppose that $\alpha\in\Delta$ satisfies $\alpha(h)=0$. There exist simple roots $\alpha_1,\ldots,\alpha_k\in\Pi$ and non-zero integers $n_1,\ldots,n_k\in\mathbb{Z}\setminus\{0\}$ such that \begin{equation}\label{Equation: Linear combination of simple roots}\alpha=\sum_{i=1}^kn_i\alpha_i.\end{equation} Since $\alpha(h)=0$, we have \begin{equation}\label{Equation: Root sum}
0=\sum_{i=1}^k n_i\alpha_i(h).
\end{equation} In particular, $0=\sum_{i=1}^k n_i\text{Re}(\alpha_i(h))$. Now, recall from \eqref{Equation: Fundamental domain} that $\text{Re}(\alpha_i(h))\geq 0$ for all $i$, and note that all $n_i$ are strictly positive or all $n_i$ are strictly negative. It follows that $\text{Re}(\alpha_i(h))=0$ for all $i$. Noting that $h\in\mathcal{D}$, it must be true that $\text{Im}(\alpha_i(h))\geq 0$ for all $i$. Also, taking imaginary parts in \eqref{Equation: Root sum} gives
$0=\sum_{i=1}^k n_i\text{Im}(\alpha_i(h))$.
Again using the fact that all $n_i$ are strictly positive or all $n_i$ are strictly negative, it follows that $\text{Im}(\alpha_i(h))=0$ for all $i$. We conclude that $\alpha_i(h)=0$ (ie. $\alpha_i\in\Pi(h)$) for all $i$, which by \eqref{Equation: Linear combination of simple roots} implies $\alpha\in\Delta_{\Pi(h)}$. Having shown \eqref{Equation: Reduced to proving} to hold, our proof is complete.
\end{proof}
Proposition \ref{Proposition: Stabilizer structure} gives rise to the following more uniform description of the stabilizers of semisimple elements.
\begin{corollary}\label{Corollary: Stabilizers of semisimple elements are reductive}
If $x\in\mathfrak{g}$ is semisimple, then $C_G(x)$ is a Levi factor of a parabolic subgroup of $G$.
\end{corollary}
\begin{proof}
Since $x$ is semisimple, $x=\Adj_g(h)$ for some $g\in G$ and $h\in\mathcal{D}$. It follows that $C_G(x)=gC_G(h)g^{-1}$, which by Proposition \ref{Proposition: Stabilizer structure} is precisely $gL_{\Pi(h)}g^{-1}$. This is a Levi factor of the parabolic subgroup $gP_{\Pi(h)}g^{-1}$.
\end{proof}
We conclude this section with a proof of the existence of regular adjoint orbits, as promised in \ref{Subsection: Dimension and regularity}.
\begin{corollary}\label{Corollary: The existence of regular orbits}
There exist regular adjoint orbits in $\mathfrak{g}$.
\end{corollary}
\begin{proof}
Let $\{\epsilon_{\alpha}\}_{\alpha\in\Pi}$ and $h\in\mathfrak{t}$ be as defined in Example \ref{Example: Kostant triple}. Since $\alpha(h)=2$ for all $\alpha\in\Pi$, we see that $h\in\mathcal{D}$ and $\Pi(h)=\emptyset$. It follows that $\mathfrak{l}_{\Pi(h)}=\mathfrak{t}$ (see \eqref{Equation: Lie algebra of Levi factor}), which together with the connectedness of $L_{\Pi(h)}$ (established in the proof of Proposition \ref{Proposition: Stabilizer structure}) implies that $L_{\Pi(h)}=T$. Now the orbit-stabilizer isomorphism \eqref{Equation: semisimple orbit-stabilizer isomorphism} and Proposition \ref{Proposition: Stabilizer structure} give
$$\dim(\mathcal{O}(h))=\dim(G/T)=\dim(G)-\dim(T)=\dim(G)-\text{rank}(G).$$ In other words, $\mathcal{O}(h)$ is a regular adjoint orbit.
\end{proof}
\subsection{Fibrations over partial flag varieties}\label{Subsection: Fibrations over partial flag varieties}
Consider a point $h\in\mathcal{D}$, to be regarded as fixed throughout \ref{Subsection: Fibrations over partial flag varieties}, and recall the variety isomorphism $O(h)\cong G/C_G(h)$ in \eqref{Equation: semisimple orbit-stabilizer isomorphism}. Proposition \ref{Proposition: Stabilizer structure} allows us to write the right-hand-side of this isomorphism as $G/L_{\Pi(h)}$, and the inclusion $L_{\Pi(h)}\subseteq P_{\Pi(h)}$ gives rise to a fibration $U_{\Pi(h)}\rightarrow G/L_{\Pi(h)}\rightarrow G/P_{\Pi(h)}$. In other words, we have a canonical fibre bundle
\begin{equation}\label{Equation: Affine bundle}
U_{\Pi(h)}\rightarrow \mathcal{O}(h)\xrightarrow{\pi_h} G/P_{\Pi(h)},
\end{equation}
where $\pi_h(\text{Ad}_g(h))=[g]\in G/P_{\Pi(h)}$ for all $g\in G$.
The fibre $U_{\Pi(h)}$ is unipotent, so that the exponential map restricts to a variety isomorphism $\mathfrak{u}_{\Pi(h)}\xrightarrow{\cong} U_{\Pi(h)}$ (see \cite[Chapt. VIII, Thm. 1.1]{Hochschild}). It follows that the fibres of \eqref{Equation: Affine bundle} are affine spaces, raising the issue of whether $\mathcal{O}(h)$ is isomorphic to the total space of a vector bundle over $G/P_{\Pi(h)}$. Perhaps surprisingly, one cannot expect an isomorphism on the level of algebraic varieties. The total space of a vector bundle over $G/P_{\Pi(h)}$ contains $G/P_{\Pi(h)}$ as the zero-section, and one knows $G/P_{\Pi(h)}$ to be irreducible (since $G$ is irreducible) and projective (see \cite[Cor. 28.1.4]{Tauvel}). It follows that the zero-section is a closed, irreducible, projective subvariety. However, the only a such subvarieties of the affine variety $\mathcal{O}(h)$ are the singletons. A similar argument precludes the existence of a biholomorphism, as a compact, connected, complex submanifold of the Stein manifold $\mathcal{O}(h)$ is necessarily a point.\footnote{For the definitions and results necessary to make this argument, we refer the reader to \cite[Chapt. V, Sect. 1]{Fritzsche}.} Nevertheless, it turns out that there is a diffeomorphism between $\mathcal{O}(h)$ and (the total space of) $T^*(G/P_{\Pi(h)})$ that respects fibrations over $G/P_{\Pi(h)}$. We devote the balance of \ref{Subsection: Fibrations over partial flag varieties} to an explicit construction of this diffeomorphism.
\begin{remark}\label{Remark: Connection to hyperkahler geometry}
By invoking well-known facts from hyperk\"{a}hler geometry, one can deduce that $\mathcal{O}(h)$ and $T^*(G/P_{\Pi(h)})$ are diffeomorphic without having to construct a diffeomorphism. To this end, recall that $\mathcal{O}(h)$ is a hyperk\"{a}hler manifold (see \ref{Subsection: Geometric features}). Among other things, it follows that the underlying smooth manifold $\mathcal{O}(h)$ carries a family of complex structures. This family turns out to include the usual complex structure on $\mathcal{O}(h)$ (ie. the one it inherits as a subvariety of $\mathfrak{g}$), as well as one in which $\mathcal{O}(h)$ becomes biholomorphic to $T^*(G/P_{\Pi(h)})$ (see \cite[Thm. 2]{Biquard}). One concludes that $\mathcal{O}(h)$ and $T^*(G/P_{\Pi(h)})$ must be diffeomorphic. While this approach is elegant, it is difficult to extract an explicit diffeomorphism $\mathcal{O}(h)\cong T^*(G/P_{\Pi(h)})$ from the supporting arguments.
\end{remark}
Let $K\subseteq G$ be a maximal compact subgroup having $K\cap T$ as a maximal real torus. We shall assume the Lie algebra $\mathfrak{k}$ of $K$ to have the form
\begin{equation}\label{Equation: Lie algebra of maximal compact subgroup}
\mathfrak{k}=(\mathfrak{k}\cap\mathfrak{t})\oplus\left(\bigoplus_{\alpha\in\Delta_{+}}\text{span}_{\mathbb{R}}\{e_{\alpha}-e_{-\alpha}\}\right)\oplus\left(\bigoplus_{\alpha\in\Delta_{+}}\text{span}_{\mathbb{R}}\{i(e_{\alpha}+e_{-\alpha})\}\right),
\end{equation}
where $e_{\alpha}\in\mathfrak{g}_{\alpha}\setminus\{0\}$ and $e_{-\alpha}\in\mathfrak{g}_{-\alpha}\setminus\{0\}$ for each $\alpha\in\Delta_+$ (cf. the proof of Theorem 6.11 in \cite{Knapp}).
Now note that $K$ acts transitively on $G/P_{\Pi(h)}$ by left multiplication (see \cite[Chapt. 5, Sect. 8]{Atiyah}), and that the $K$-stabilizer of the identity coset $[e]\in G/P_{\Pi(h)}$ is $K\cap P_{\Pi(h)}=:K_{\Pi(h)}$. By the orbit-stabilizer isomorphism \eqref{Equation: Orbit-stabilizer isomorphism}, this time for compact Lie group actions, we have a $K$-equivariant diffeomorphism.
\begin{equation}\label{Equation: Equivariant diffeomorphism}
K/K_{\Pi(h)}\xrightarrow{\cong} G/P_{\Pi(h)}.
\end{equation}
We will later need the following result concerning $K_{\Pi(h)}$.
\begin{lemma}\label{Lemma: K-stabilizer}
We have $K_{\Pi(h)}\subseteq L_{\Pi(h)}$.
\end{lemma}
\begin{proof}
We begin by showing $K_{\Pi(h)}$ to be connected. To this end, consider the fibration $K_{\Pi(h)}\rightarrow K\rightarrow K/K_{\Pi(h)}$. The base is simply-connected by virtue of being diffeomorphic to $G/P_{\Pi(h)}$, a simply-connected space (see \cite[Chapt. 3, Sect. 1, Prop. 7]{Akhiezer}). The total space $K$ is homotopy-equivalent to $G$ (see \cite[Appendix A, Thm. 1.1]{Whitehead}), so that the connectedness of $K$ follows from that of $G$. Using these two observations together with the long exact sequence of homotopy groups associated to our fibration (see \cite[Chapt. IV, Cor. 8.6]{Whitehead}), we see that $K_{\Pi(h)}$ is also connected. It will therefore suffice to prove that $\mathfrak{k}_{\Pi(h)}\subseteq\mathfrak{l}_{\Pi(h)}$, where $\mathfrak{k}_{\Pi(h)}$ and $\mathfrak{l}_{\Pi(h)}$ are the Lie algebras of $K_{\Pi(h)}$ and $L_{\Pi(h)}$, respectively. The former Lie algebra is $\mathfrak{k}\cap\mathfrak{p}_{\Pi(h)}$, which via \eqref{Equation: Lie algebra of standard parabolic} and \eqref{Equation: Lie algebra of maximal compact subgroup} can be shown to coincide with $$(\mathfrak{k}\cap\mathfrak{t})\oplus\left(\bigoplus_{\alpha\in\Delta_{\Pi(h)}^{+}}\text{span}_{\mathbb{R}}\{e_{\alpha}-e_{-\alpha}\}\right)\oplus\left(\bigoplus_{\alpha\in\Delta_{\Pi(h)}^{+}}\text{span}_{\mathbb{R}}\{i(e_{\alpha}+e_{-\alpha})\}\right).$$ Since $\mathfrak{t}\subseteq\mathfrak{l}_{\Pi(h)}$ and $\mathfrak{g}_{\alpha}\oplus\mathfrak{g}_{-\alpha}\subseteq\mathfrak{l}_{\Pi(h)}$ for all $\alpha\in\Delta_{\Pi(h)}^+$ (see \ref{Equation: Lie algebra of Levi factor}), it follows that $\mathfrak{k}_{\Pi(h)}\subseteq\mathfrak{l}_{\Pi(h)}$.
\end{proof}
Let us return to the main discussion. Note that the pullback of $T^*(G/P_{\Pi(h)})$ under \eqref{Equation: Equivariant diffeomorphism} is a smooth complex vector bundle $E\rightarrow K/K_{\Pi(h)}$, so that we have the commutative diagram \begin{equation}\label{Equation: Commutative diagram}
\begin{tikzcd}[column sep=30pt]
E \ar{d} \ar{r}{\cong} & T^*(G/P_{\Pi(h)}) \ar{d} \\
K/K_{\Pi(h)} \ar{r}{\cong} & G/P_{\Pi(h)}
\end{tikzcd}
\end{equation}
One can show $E$ to be a $K$-equivariant vector bundle, meaning that $K$ acts smoothly on $E$ via a lift of the left multiplication action on $K/K_{\Pi(h)}$ (hence $k\in K$ sends the fibre over $x\in K/K_{\Pi(h)}$ to the fibre over $k\cdot x$), and secondly that restricting the action of $k\in K$ to the fibre over $x\in K/K_{\Pi(h)}$ gives a vector space isomorphism with the fibre over $k\cdot x$. Since the identity coset $[e]\in K/K_{\Pi(h)}$ is fixed by $K_{\Pi(h)}$, this second condition implies that the fibre over $[e]$ in any $K$-equivariant vector bundle is a complex $K_{\Pi(h)}$-representation. In fact, this turns out to define a bijective correspondence between isomorphism classes of (smooth) complex $K_{\Pi(h)}$-representations and isomorphism classes of smooth $K$-equivariant complex vector bundles over $K/K_{\Pi(h)}$. The inverse process assigns to each representation $\varphi:K_{\Pi(h)}\rightarrow\GL(V)$ the \textit{associated bundle} $K\times_{K_{\Pi(h)}}V$, defined as follows. The group $K_{\Pi(h)}$ acts freely on $K\times V$ via $g\cdot(k,v)=(kg^{-1},\varphi(g)v)$, $g\in K_{\Pi(h)}$, $k\in K$, $v\in V$. One defines $K\times_{K_{\Pi(h)}}V$ to be the quotient manifold $(K\times V)/K_{\Pi(h)}$, equipped with the projection to $K/K_{\Pi(h)}$ given by $[(k,v)]\mapsto[k]$. The fibres of this projection are naturally complex vector spaces, and the $K$-equivariant structure comes from the left-multiplicative action of $K$ on the first factor of $K\times_{K_{\Pi(h)}}V$.
With the above discussion in mind, it is natural to realize $E$ as an associated bundle. To this end, $U_{\Pi(h)}$ is a normal subgroup of $P_{\Pi(h)}$ by virtue of the former being the unipotent radical of the latter. It follows that $\mathfrak{u}_{\Pi(h)}$ is invariant under $P_{\Pi(h)}$ (hence also the subgroup $K_{\Pi(h)}$), acting through the adjoint representation of $G$. In particular, $\mathfrak{u}_{\Pi(h)}$ is a $K_{\Pi(h)}$-representation.
\begin{lemma}\label{Lemma: The fibre of E over [e]}
There is an isomorphism $E\cong K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}$ of $K$-equivariant vector bundles over $K/K_{\Pi(h)}$.
\end{lemma}
\begin{proof}
Given our discussion of the correspondence between isomorphism classes of $K$-equivariant vector bundles and those of $K_{\Pi(h)}$-representations, it will suffice to prove that $\mathfrak{u}_{\Pi(h)}$ is isomorphic to the fibre of $E$ over $[e]\in K/K_{\Pi(h)}$ as a $K_{\Pi(h)}$-representation. The latter is the dual of the tangent space to $[e]\in G/P_{\Pi(h)}$, which by \eqref{Equation: Tangent space description} (with $\mathcal{O}=G/P_{\Pi(h)}$) is isomorphic to $(\mathfrak{g}/\mathfrak{p}_{\Pi(h)})^*$. One can identify $(\mathfrak{g}/\mathfrak{p}_{\Pi(h)})^*$ with $\{\phi\in\mathfrak{g}^*:\phi\vert_{\mathfrak{p}_{\Pi(h)}}=0\}$, which under our isomorphism $\mathfrak{g}^*\cong\mathfrak{g}$ corresponds to $\{x\in\mathfrak{g}:\langle x,y\rangle=0\text{ for all }y\in\mathfrak{p}_{\Pi(h)}\}$. This last subspace can be seen to coincide with $\mathfrak{u}_{\Pi(h)}$, as desired. We leave it as an exercise for the reader to verify that each of the above-constructed isomorphisms is an isomorphism of $P_{\Pi(h)}$ (hence also $K_{\Pi(h)}$-) representations.
\end{proof}
Using Lemma \ref{Lemma: The fibre of E over [e]}, we may present \eqref{Equation: Commutative diagram} as
\begin{equation}\label{Equation: Second commutative diagram}
\begin{tikzcd}[column sep=30pt]
K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)} \ar{d} \ar{r}{\cong} & T^*(G/P_{\Pi(h)}) \ar{d} \\
K/K_{\Pi(h)} \ar{r}{\cong} & G/P_{\Pi(h)}.
\end{tikzcd}
\end{equation}
\begin{proposition}\label{Proposition: Diffeomorphism with the cotangent bundle}
There is a $K$-equivariant diffeomorphism $T^*(G/P_{\Pi(h)})\xrightarrow{\cong}\mathcal{O}(h)$ for which the following diagram commutes:
\begin{equation}\label{Equation: Diffeomorphism with cotangent bundle}
\begin{tikzcd}[column sep=10pt]
T^*(G/P_{\Pi(h)}) \ar{rr}{\cong} \ar{dr} & & \mathcal{O}(h) \ar{dl}{\pi_h} \\
& G/P_{\Pi(h)} &
\end{tikzcd}.
\end{equation}
\end{proposition}
\begin{proof}
We will construct a $K$-equivariant diffeomorphism $K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}\xrightarrow{\cong}\mathcal{O}(h)$ making
\begin{equation}\label{Equation: Third commutative diagram}
\begin{tikzcd}[column sep=30pt]
K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)} \ar{d} \ar{r}{\cong} & \mathcal{O}(h) \ar{d}{\pi_h} \\
K/K_{\Pi(h)} \ar{r}{\cong} & G/P_{\Pi(h)}
\end{tikzcd}
\end{equation}
commute. By reversing the horizontal arrows in \eqref{Equation: Second commutative diagram} and combining with \eqref{Equation: Third commutative diagram}, we will have constructed the desired isomorphism $T^*(G/P_{\Pi(h)})\xrightarrow{\cong}\mathcal{O}(h)$. Accordingly, let \\ $K\times_{K_{\Pi(h)}}(P_{\Pi(h)}/L_{\Pi(h)})$ denote the quotient of $K\times (P_{\Pi(h)}/L_{\Pi(h)})$ by the $K_{\Pi(h)}$-action $g\cdot (k,[p])=(kg^{-1},[gp])$, $g\in K_{\Pi(h)}$, $k\in K$, $p\in P_{\Pi(h)}$. The proof of Lemma 2.4 in \cite{Azad} shows the map
\begin{equation}\label{Equation: Biswas diffeomorphism}
K\times_{{K_{\Pi(h)}}}(P_{\Pi(h)}/L_{\Pi(h)})\rightarrow G/L_{\Pi(h)},\quad [(k,[p])]\mapsto[kp]
\end{equation}
to be a (well-defined) diffeomorphism. In the interest of re-writing \eqref{Equation: Biswas diffeomorphism}, recall that $L_{\Pi(h)}=C_G(h)$ (see Proposition \ref{Proposition: Stabilizer structure}). One therefore has the orbit-stabilizer isomorphism \begin{equation}\label{Equation: Second orbit-stabilizer isomorphism}
G/L_{\Pi(h)}\xrightarrow{\cong}\mathcal{O}(h),\quad [g]\mapsto\Adj_g(h).\end{equation} As for the domain of \eqref{Equation: Biswas diffeomorphism}, note that the Levi factorization \eqref{Equation: Levi decomposition of parabolic} implies $U_{\Pi(h)}\rightarrow P_{\Pi(h)}/L_{\Pi(h)}$, $u\mapsto[u]$, is a variety isomorphism. Identifying $U_{\Pi(h)}$ with $\mathfrak{u}_{\Pi(h)}$ via the exponential map, this isomorphism becomes \begin{equation}\label{Equation: Useful isomorphism}
\mathfrak{u}_{\Pi(h)}\rightarrow P_{\Pi(h)}/L_{\Pi(h)},\quad x\mapsto[\exp(x)].\end{equation} Since $K_{\Pi(h)}\subseteq L_{\Pi(h)}$ (see Lemma \ref{Lemma: K-stabilizer}) and $\exp(\Adj_g(x))=g\exp(x)g^{-1}$ for all $g\in K_{\Pi(h)}$ and $x\in\mathfrak{u}_{\Pi(h)}$, one can show \eqref{Equation: Useful isomorphism} to be $K_{\Pi(h)}$-equivariant. This allows us to replace $P_{\Pi(h)}/L_{\Pi(h)}$ in \eqref{Equation: Biswas diffeomorphism} with $\mathfrak{u}_{\Pi(h)}$, giving rise to the diffeomorphism
\begin{equation}\label{Equation: Second useful isomorphism}
K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}\rightarrow K\times_{K_{\Pi(h)}}(P_{\Pi(h)}/L_{\Pi(h)}),\quad [(k,x)]\mapsto[(k,[\exp(x)])].\end{equation}
Now composing the three diffeomorphisms \eqref{Equation: Biswas diffeomorphism}, \eqref{Equation: Second orbit-stabilizer isomorphism}, and \eqref{Equation: Second useful isomorphism} (in the only order that makes sense), we obtain
\begin{equation}\label{Equation: Final diffeomorphism}
K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}\rightarrow\mathcal{O}(h),\quad [(k,x)]\mapsto\Adj_{k\exp(x)}(h).
\end{equation}
This diffeomorphism is clearly $K$-equivariant. To see that it makes \eqref{Equation: Third commutative diagram} commute, note that the composite map $K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}\rightarrow K/K_{\Pi(h)}\rightarrow G/P_{\Pi(h)}$ sends $[(k,x)]\in K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}$ to $[k]\in G/P_{\Pi(h)}$. On the other hand, the composite map $K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}\rightarrow\mathcal{O}(h)\rightarrow G/P_{\Pi(h)}$ sends $[(k,x)]\in K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}$ to $[k\exp(x)]\in G/P_{\Pi(h)}$. However, since $x\in\mathfrak{u}_{\Pi(h)}\subseteq\mathfrak{p}_{\Pi(h)}$, we must have $\exp(x)\in P_{\Pi(h)}$. It follows that $[k\exp(x)]=[k]$ in $G/P_{\Pi(h)}$, so that \eqref{Equation: Third commutative diagram} indeed commutes. This completes the proof.
\end{proof}
Proposition \ref{Proposition: Diffeomorphism with the cotangent bundle} has strong implications for the topology (both equivariant and non-equivariant) of a semisimple orbit.
\begin{corollary}\label{Corollary: Equivariant homotopy equivalence}
There is a $K$-equivariant homotopy equivalence between $\mathcal{O}(h)$ and the compact group orbit $\mathcal{O}_K(h)=K\cdot h$.
\end{corollary}
\begin{proof}
Note that $K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}$ is $K$-equivariantly homotopy equivalent to its zero-section. Since \eqref{Equation: Final diffeomorphism} is a $K$-equivariant diffeomorphism, it follows that the image of the zero-section under \eqref{Equation: Final diffeomorphism} is $K$-equivariantly homotopy equivalent to $\mathcal{O}(h)$. The zero-section is given by $\{[(k,0)]\in K\times_{K_{\Pi(h)}}\mathfrak{u}_{\Pi(h)}:k\in K\}$, and its image under \eqref{Equation: Final diffeomorphism} is the $K$-orbit $\mathcal{O}_K(h)\subseteq\mathcal{O}(h)$. This completes the proof.
\end{proof}
An essentially analogous version of Corollary \ref{Corollary: Equivariant homotopy equivalence} is as follows. The cotangent bundle projection $T^*(G/P_{\Pi(h)})\rightarrow G/P_{\Pi(h)}$ is a $K$-equivariant (in fact, $G$-equivariant) homotopy equivalence, so that \eqref{Equation: Diffeomorphism with cotangent bundle} shows $\pi_h:\mathcal{O}(h)\rightarrow G/P_{\Pi(h)}$ to be a $K$-equivariant homotopy equivalence. In principle, one can use this fact to compute various topological invariants of $\mathcal{O}(h)$. Indeed, since $G/P_{\Pi(h)}$ is simply-connected (as discussed in the proof of Lemma \ref{Lemma: K-stabilizer}), so too is $\mathcal{O}(h)$. Secondly, the cohomology ring $H^*(\mathcal{O}(h);\mathbb{Z})$ is isomorphic to $H^*(G/P_{\Pi(h)};\mathbb{Z})$, which has a well-studied description in terms of Schubert calculus (see \cite{Bernstein}). This result has an equivariant counterpart, namely that the $L$-equivariant cohomology rings $H_L^*(\mathcal{O}(h);\mathbb{Z})$ and $H_L^*(G/P_{\Pi(h)};\mathbb{Z})$ are isomorphic for any closed subgroup $L\subseteq K$ (see \cite{Brion} for details on equivariant cohomology). If one takes $L$ to be the maximal torus $K\cap T$ of $K$, then $H_L^*(G/P_{\Pi(h)};\mathbb{Z})$ is describable via equivariant Schubert calculus (see \cite{Graham,Mihaleca}) and (if one works over $\mathbb{C}$ instead of $\mathbb{Z}$) Goresky-Kottwitz-MacPherson (GKM) theory (see \cite{Guillemin}).
\subsection{Equivariant projective compactifications}\label{Subsection: Equivariant projective compactifications}
While semisimple orbits are affine (see Corollary \ref{Corollary: Semisimple orbits are affine}), each turns out to admit a $G$-equivariant projective compactification (ie. a projective $G$-variety $X$ and a $G$-variety isomorphism between the orbit and an invariant open dense subvariety of $X$). To properly construct this compactification, however, we will need a few preliminary results. Let $w_0\in W$ denote the longest element of the Weyl group (see \cite[Chapt. 20]{Bump}). It is known that $w_0\cdot\Pi=-\Pi$, so that $S\mapsto S^{\vee}:=-w_0\cdot S$ defines an involution on the collection of subsets $S\subseteq\Pi$. With this in mind, we have the following lemma.
\begin{lemma}\label{Lemma: Root set equality}
For $S\subseteq\Pi$, we have $w_0\cdot\Delta_{S^{\vee}}^{-}=\Delta_{S}^+$.
\end{lemma}
\begin{proof}
Let us write $S=\{\alpha_1,\ldots,\alpha_n\}$, so that $S^{\vee}=\{\gamma_1,\ldots,\gamma_n\}$ with $\gamma_i=-w_0\cdot\alpha_i$ for all $i$. Given $\beta\in\Delta_S^+$, we have $\beta=\sum_{i=1}^nk_i\alpha_i$ with $k_1,\ldots,k_n\in\mathbb{Z}_{\geq 0}$. It follows that
$$w_0\cdot\beta=\sum_{i=1}^n(-k_i)(-w_0\cdot\alpha_i)=\sum_{i=1}^n(-k_i)\gamma_i.$$ Since $w_0\cdot\beta$ is a root, this implies that $w_0\cdot\beta\in\Delta_{S^{\vee}}^{-}$. Using the fact that $w_0^2=e$, we have $\beta\in w_0\cdot\Delta_{S^{\vee}}^{-}$.
Conversely, if $\beta\in w_0\cdot\Delta_{S^{\vee}}^{-}$, then $\beta=w_0\cdot\left(\sum_{i=1}^nk_i\gamma_i\right)$ with $k_1,\ldots,k_n\in\mathbb{Z}_{\leq 0}$. We conclude that
$$\beta=w_0\cdot\left(\sum_{i=1}^nk_i(-w_0\cdot\alpha_i)\right)=\sum_{i=1}^n(-k_i)\alpha_i.$$
Noting that $\beta$ is a root, this implies that $\beta\in\Delta_{S}^+$.
\end{proof}
Given $S\subseteq\Pi$, the subset $S^{\vee}\subseteq\Pi$ determines a standard parabolic subgroup $P_{S^{\vee}}\subseteq G$ (as discussed in \ref{Subsection: The basic objects}). Now consider the parabolic subgroup $P_S^*\subseteq G$ defined by
\begin{equation}\label{Equation: Dual parabolic}
P_S^*:=w_0P_{S^{\vee}}(w_0)^{-1},
\end{equation}
where $w_0P_{S^{\vee}}(w_0)^{-1}:=gP_{S^{\vee}}g^{-1}$ for any representative $g\in N_G(T)$ of $w_0$.
\begin{proposition}\label{Proposition: Intersection of opposite parabolics}
For $S\subseteq\Pi$, we have $P_S\cap P_{S}^*=L_S$.
\end{proposition}
\begin{proof}
Since $L_S$ and $P_S\cap P_{S}^*$ are connected (the former by the beginning of the proof of Proposition \ref{Proposition: Stabilizer structure} and the latter by \cite[Prop. 2.1]{Digne}), it suffices to establish the equality of their respective Lie algebras. The latter Lie algebra is precisely $\mathfrak{p}_S\cap w_0\cdot\mathfrak{p}_{S^{\vee}}$, where $w_0\cdot\mathfrak{p}_{S^{\vee}}$ denotes the adjoint action of a lift of $w_0$ to $N_G(T)$. This intersection contains $\mathfrak{t}$ and therefore decomposes as a direct sum of $\mathfrak{t}$ and the root spaces belonging to both $\mathfrak{p}_{S}$ and $w_0\cdot\mathfrak{p}_{S^{\vee}}$. To identify these root spaces, note that \begin{equation}\label{Equation: First root space decomposition}
\mathfrak{p}_S=\mathfrak{t}\oplus\left(\bigoplus_{\alpha\in\Delta_+}\mathfrak{g}_{\alpha}\right)\oplus\left(\bigoplus_{\alpha\in\Delta_{S}^{-}}\mathfrak{g}_{\alpha}\right)\end{equation} and
\begin{equation}\label{Equation: Second root space decomposition}
w_0\cdot\mathfrak{p}_{S^{\vee}}=\mathfrak{t}\oplus\left(\bigoplus_{\alpha\in w_0\cdot\Delta_{+}}\mathfrak{g}_{\alpha}\right)\oplus\left(\bigoplus_{\alpha\in w_0\cdot\Delta_{S^{\vee}}^{-}}\mathfrak{g}_{\alpha}\right).\end{equation}
Since $w_0\cdot\Delta_{+}=\Delta_{-}$ (see \cite[Prop. 20.14]{Bump} and $w_0\cdot\Delta_{S^{\vee}}^{-}=\Delta_{S}^+$ (by Lemma \ref{Lemma: Root set equality}), we may re-write \eqref{Equation: Second root space decomposition} to obtain
\begin{equation}\label{Equation: Third root space decomposition}
w_0\cdot\mathfrak{p}_{S^{\vee}}=\mathfrak{t}\oplus\left(\bigoplus_{\alpha\in \Delta_{S}^+}\mathfrak{g}_{\alpha}\right)\oplus\left(\bigoplus_{\alpha\in\Delta_{-}}\mathfrak{g}_{\alpha}\right).
\end{equation}
Using \eqref{Equation: First root space decomposition} and \eqref{Equation: Third root space decomposition}, we conclude that
$$\mathfrak{p}_{S}\cap w_0\cdot\mathfrak{p}_{S^{\vee}}=\mathfrak{t}\oplus\left(\bigoplus_{\alpha\in\Delta_S^+}\mathfrak{g}_{\alpha}\right)\oplus\left(\bigoplus_{\alpha\in\Delta_S^-}\mathfrak{g}_{\alpha}\right)=\mathfrak{l}_S,$$ the Lie algebra of $L_S$. This completes the proof.
\end{proof}
Returning to the discussion of equivariant projective orbit compactifications, fix $h\in\mathcal{D}$ and consider its semisimple orbit $\mathcal{O}(h)$. Also, let $X$ be any $G$-variety. By means of \eqref{Equation: Second orbit-stabilizer isomorphism}, a $G$-variety isomorphism between $\mathcal{O}(h)$ and an invariant subvariety $Y\subseteq X$ is equivalent to a $G$-variety isomorphism between $G/L_{\Pi(h)}$ and $Y$. Now by the general orbit-stabilizer theory developed in \ref{Subsection: Orbits as quotient varieties}, specifying the latter isomorphism is equivalent to specifying a point $x\in X$ with $C_G(x)=L_{\Pi(h)}$. In this case, $G/L_{\Pi(h)}$ (hence $\mathcal{O}(h)$) is equivariantly identified with the $G$-orbit of $x$ in $X$.
In light of the above, we seek a projective $G$-variety $X$ and a point $x\in X$ with $C_G(x)=L_{\Pi(h)}$. To this end, consider the diagonal $G$-action (see Example \ref{Example: Diagonal action}) on the projective variety $X=G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^*}$. Consider also the pair of identity cosets $x=([e],[e])\in X$. The $G$-stabilizer of $x$ is $P_{\Pi(h)}\cap P_{\Pi(h)}^*$, which by Proposition \ref{Proposition: Intersection of opposite parabolics} is precisely $L_{\Pi(h)}$. Appealing to the previous paragraph, the following is a $G$-variety isomorphism between $\mathcal{O}(h)$ and the $G$-orbit $G\cdot([e],[e])\subseteq G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^*}$:
\begin{equation}\label{Equation: First inclusion}
\mathcal{O}(h)\xrightarrow{\cong}G\cdot([e],[e]),\quad \text{Ad}_g(h)\mapsto ([g],[g]),\quad g\in G.
\end{equation}
\begin{theorem}\label{Theorem: Equivariant projective compactification}
If $h\in\mathcal{D}$, then \eqref{Equation: First inclusion} makes $G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^{*}}$ a $G$-equivariant projective compactification of $\mathcal{O}(h)$.\footnote{For the precise meaning of ``$G$-equivariant projective compactification'', see the beginning of \ref{Subsection: Equivariant projective compactifications}.}
\end{theorem}
\begin{proof}
It remains only to prove that $G\cdot([e],[e])$ is open and dense in $G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^{*}}$. However, since $G\cdot([e],[e])$ is open in its closure (see \ref{Subsection: Generalities on orbits}), it will suffice to prove that $\overline{G\cdot([e],[e])}=G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^{*}}$. Now note that $G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^{*}}$ is irreducible, so that it has no proper irreducible closed subvarieties of codimension zero (see \cite[Sect. 3.2]{Humphreys}). We are therefore reduced to proving that $\overline{G\cdot([e],[e])}$ and $G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^{*}}$ have equal dimensions. The dimension of the former equals that of its open subvariety $G\cdot([e],[e])$, which by \eqref{Equation: First inclusion} equals $\dim(\mathcal{O}(h))$. Proposition \ref{Proposition: Diffeomorphism with the cotangent bundle} implies $\dim(\mathcal{O}(h))=2\dim(G/P_{\Pi(h)})$, so that we have \begin{equation}\label{Equation: First dimension}\dim(\overline{G\cdot([e],[e])})=2\dim(G/P_{\Pi(h)}).\end{equation} At the same time, we have \begin{equation}\label{Equation: Second dimension}
\dim(G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^{*}})=\dim(G/P_{\Pi(h)})+\dim(G/{P_{\Pi(h)}^{*}}).\end{equation} In light of \eqref{Equation: First dimension} and \eqref{Equation: Second dimension}, it will be enough to prove that $\dim(G/P_{\Pi(h)})=\dim(G/{P_{\Pi(h)}^*})$ (or equivalently, $\dim(P_{\Pi(h)})=\dim(P_{\Pi(h)}^*)$). To this end, Lemma \ref{Lemma: Root set equality} implies that $\Delta_{\Pi(h)}^+$ and $\Delta_{\Pi(h)^{\vee}}^{-}$ have the same cardinality. Of course, since $\Delta_{\Pi(h)}^{-}=-\Delta_{\Pi(h)}^+$, this is equivalent to $\Delta_{\Pi(h)}^{-}$ and $\Delta_{\Pi(h)^{\vee}}^-$ having the same cardinality. Now \eqref{Equation: Lie algebra of standard parabolic} implies that $\dim(\mathfrak{p}_{\Pi(h)})=\dim(\mathfrak{p}_{\Pi(h)^{\vee}})$, which one can re-write as $\dim(P_{\Pi(h)})=\dim(P_{\Pi(h)^{\vee}})$. Finally, since $P_{\Pi(h)}^{*}$ was defined to be a conjugate of $P_{\Pi(h)^{\vee}}$ (see \eqref{Equation: Dual parabolic}), the dimensions of $P_{\Pi(h)^{\vee}}$ and $P_{\Pi(h)}^{*}$ must also agree. Hence $\dim(P_{\Pi(h)})=\dim(P_{\Pi(h)}^*)$, as desired.
\end{proof}
\begin{remark}\label{Remark: Connection to mirror symmetry}
The compactification in Theorem \ref{Theorem: Equivariant projective compactification} turns out to have some context in the homological mirror symmetry program (see \cite{Givental,Hori} for an introduction to this program). Vaguely speaking, an instance of homological mirror symmetry is an equivalence between symplecto-geometric data on a smooth manifold (called the $A$-side of the symmetry) and algebro-geometric data on a projective variety (called the $B$-side of the symmetry). In \cite{Ballico}, the authors interpret this as a symmetry between so-called $A$-side Landau-Ginzburg (LG) models and $B$-side LG models (see \cite[Sect. 1]{Ballico} for further details). Building on an existing interpretation of a semisimple orbit $\mathcal{O}(h)$ as an $A$-side LG model (explained in \cite[Sect. 1]{Ballico} and based on \cite[Thm. 2.2]{GGSM}), the authors explain how one would make $G/P_{\Pi(h)}\times G/{P_{\Pi(h)}^{*}}$ into the corresponding $B$-side model.
\end{remark}
\section{Nilpotent Orbits}\label{Section: Nilpotent orbits}
Recall that we introduced semisimple orbits as generalizing the conjugacy classes of diagonalizable matrices. We now discuss the appropriate generalization of a nilpotent conjugacy class -- a \textit{nilpotent orbit}.
\subsection{Definitions and first results}\label{Subsection: Definitions and first results}
Recall that the nilpotent cone $\mathcal{N}\subseteq\mathfrak{g}$ is invariant under the adjoint action of $G$ (see \ref{Subsection: semisimple and nilpotent elements}). It follows that $\mathcal{N}$ is a union of those adjoint orbits $\mathcal{O}$ for which $\mathcal{O}\cap\mathcal{N}\neq\emptyset$ (ie. $\mathcal{O}$ contains a nilpotent element). Such adjoint orbits will be called \textit{nilpotent orbits}.
\begin{definition}
An adjoint orbit $\mathcal{O}\subseteq\mathfrak{g}$ is called a \textit{nilpotent orbit} if $\mathcal{O}\cap\mathcal{N}\neq\emptyset$, or equivalently $\mathcal{O}\subseteq\mathcal{N}$.
\end{definition}
While this is generally taken to be the definition of a nilpotent orbit, there are several equivalent descriptions. Indeed, one can rephrase Proposition \ref{Proposition: Characterizations of nilpotent elements} in the language of adjoint orbits as follows.
\begin{proposition}\label{Proposition: Characterizations of nilpotent orbits}
If $\mathcal{O}\subseteq\mathfrak{g}$ is an adjoint orbit, then the following conditions are equivalent.
\begin{itemize}
\item[(i)] $\mathcal{O}$ is a nilpotent orbit.
\item[(ii)] $\mathcal{O}$ is invariant under the dilation action of $\mathbb{C}^*$ on $\mathfrak{g}$.
\item[(iii)] $0\in\overline{\mathcal{O}}$.
\item[(iv)] For all $f\in\mathbb{C}[\mathfrak{g}]^G$, $f$ takes the constant value $f(0)$ on $\mathcal{O}$.
\end{itemize}
\end{proposition}
\begin{example}\label{Example: Type A nilpotent orbits}
Let $G=\SL_n(\mathbb{C})$. By Theorem \ref{Theorem: equivalence of semisimple / nilpotent} and Example \ref{Example: Type A adjoint orbits}, the nilpotent orbits in $\mathfrak{sl}_n(\mathbb{C})$ are exactly the conjugacy classes of the nilpotent $n\times n$ matrices (which necessarily have zero trace). These conjugacy classes are in turn indexed by the nilpotent $n\times n$ matrices in Jordan canonical form, modulo permutations of Jordan blocks along the diagonal. Of course, each of these equivalence classes of nilpotent Jordan matrices has a unique representative whose block heights, read from top to bottom, form a non-decreasing sequence. The sequences one obtains in this way are precisely the partitions of $n$, ie. the positive integer sequences $\lambda=(\lambda_1,\lambda_2,\ldots,\lambda_k)$, $1\leq k\leq n$, satisfying $\lambda_1\geq\lambda_2\geq\ldots\geq\lambda_k$ and $\lambda_1+\lambda_2+\ldots+\lambda_k=n$. It follows that the nilpotent orbits in $\mathfrak{sl}_n(\mathbb{C})$ are explicitly parametrized by the partitions of $n$. In particular, there are only finitely many nilpotent $\SL_n(\mathbb{C})$-orbits.
\end{example}
This last point about finiteness turns out to hold in much greater generality. More precisely, it turns out there are only finitely many nilpotent $G$-orbits. The proof of this fact, however, is far more complicated than what appears in Example \ref{Example: Type A nilpotent orbits}. One approach uses ideas related to the Springer resolution and Steinberg variety (see \cite[Sect. 3.3]{Chriss} for more details). Another strategy is to classify nilpotent orbits using so-called ``weighted Dynkin diagrams'' and appeal to the fact that there are only finitely many of the latter (see \cite[Chapt. 3]{Collingwood}). Here, we will see the finiteness of nilpotent orbits as a direct consequence of Kostant's work on the adjoint quotient.
\begin{theorem}\label{Theorem: Finitely many nilpotent orbits}
There are only finitely many nilpotent orbits.
\end{theorem}
\begin{proof}
Recall the description of the adjoint quotient given in \eqref{Equation: The second adjoint quotient}. It will suffice to prove that $\Phi^{-1}(0)=\mathcal{N}$, since Theorem \ref{Theorem: Kostant's theorem} will then show $\mathcal{N}$ to be a union of finitely many adjoint orbits. To this end, Proposition \ref{Proposition: Characterizations of nilpotent elements} implies that $x\in\mathfrak{g}$ is nilpotent if and only if $f(x)=f(0)$ for all $f\in\mathbb{C}[\mathfrak{g}]^G$. Letting $\chi_1,\chi_2,\ldots,\chi_r\in\mathbb{C}[\mathfrak{g}]^G$ be the homogeneous generators mentioned in \ref{Subsection: The adjoint quotient}, this is equivalent to $\chi_j(x)=0$ for all $j=1,\ldots,n$, ie. $\Phi(x)=0$. The preceding argument establishes that $\Phi^{-1}(0)=\mathcal{N}$, as desired.
\end{proof}
\subsection{The closure order on nilpotent orbits}\label{Subsection: The closure order on nilpotent orbits}
In what follows, we will see that nilpotent orbits are profitably studied in the context of the closure order (see \ref{Subsection: The closure order}) on the set of adjoint $G$-orbits. To help formulate this, recall that a subset $Q$ of a poset is called a ``connected component'' if it is maximal with respect to the following property: if $p,q\in Q$, then there exists a sequence $r_1,\ldots,r_k\in Q$ with $r_1=p$, $r_k=q$ and for all $j\in\{1,\ldots,k-1\}$, $r_j\leq r_{j+1}$ or $r_{j}\geq r_{j+1}$ (ie. $p$ and $q$ are ``connected'' by a sequence of comparisons in $Q$). It turns out that the nilpotent orbits form a ``connected component'' in the poset of adjoint orbits.
\begin{proposition}\label{Proposition: Nilpotent orbits form a connected component}
The nilpotent orbits form a connected component in the poset of adjoint orbits.
\end{proposition}
\begin{proof}
To begin, $\{0\}$ is clearly a nilpotent orbit and Proposition \ref{Proposition: Characterizations of nilpotent orbits} (iii) implies that $\{0\}\leq\mathcal{O}$ for all nilpotent orbits $\mathcal{O}$. In other words, all nilpotent orbits are ``connected'' to $\{0\}$ in the above-defined sense. In particular, any two nilpotent orbits are connected by a sequence of comparisons with other nilpotent orbits. It now remains only to show maximality, namely that if a nilpotent orbit $\mathcal{O}$ and an adjoint orbit $\Theta$ satisfy $\Theta\leq\mathcal{O}$ or $\mathcal{O}\leq\Theta$, then $\Theta$ is a nilpotent orbit. To this end, suppose $\mathcal{O}$ and $\Theta$ are nilpotent and adjoint orbits, respectively. If $\Theta\leq\mathcal{O}$, then by definition $\Theta\subseteq\overline{\mathcal{O}}$. Since $\mathcal{N}$ is a closed subvariety of $\mathfrak{g}$ (see \ref{Subsection: semisimple and nilpotent elements}) containing $\mathcal{O}$, we must have $\overline{\mathcal{O}}\subseteq\mathcal{N}$. Hence $\Theta\subseteq\mathcal{N}$, meaning that $\Theta$ is a nilpotent orbit. For the other case, assume that $\mathcal{O}\leq\Theta$. It follows by definition that $\mathcal{O}\subseteq\overline{\Theta}$, and therefore also $\overline{\mathcal{O}}\subseteq\overline{\Theta}$. Since $0\in\overline{\mathcal{O}}$ (see Proposition \ref{Proposition: Characterizations of nilpotent orbits} (iii)), $0\in\overline{\Theta}$ and Proposition \ref{Proposition: Characterizations of nilpotent orbits} shows that
$\Theta$ is a nilpotent orbit.
\end{proof}
\begin{remark}
There is no immediate analogue of Proposition \ref{Proposition: Nilpotent orbits form a connected component} for semisimple orbits. For one thing, it is possible for $\mathcal{O}\leq\Theta$ to hold when $\mathcal{O}$ is a semisimple orbit and $\Theta$ is a non-semisimple adjoint orbit. Indeed, let $\Theta$ be any non-semisimple adjoint orbit. Lemma \ref{Lemma: Unique semisimple orbit in closure} implies that there exists a unique semisimple orbit $\mathcal{O}$ satisfying $\mathcal{O}\subseteq\overline{\Theta}$. By construction, $\mathcal{O}\leq\Theta$.
\end{remark}
\begin{example}\label{Example: Nilpotent orbits and the dominance order}
Let us describe the poset of nilpotent orbits in $\mathfrak{sl}_n(\mathbb{C})$. To this end, recall that Example \ref{Example: Type A nilpotent orbits} uses the partitions of $n$ to explicitly parametrize the nilpotent $\SL_n(\mathbb{C})$-orbits. It follows that the closure order on these nilpotent orbits corresponds to a partial order on the partitions of $n$, which Gerstenhaber \cite[Chapt. 1, Thm. 2]{Gerstenhaber} and Hesselink \cite[Thm. 3.10]{Hesselink} showed to coincide with the \textit{dominance order} on partitions. Two partitions $\lambda=(\lambda_1,\lambda_2,\ldots,\lambda_k)$ and $\mu=(\mu_1,\mu_2,\ldots,\mu_{\ell})$ of $n$ satisfy $\lambda\leq\mu$ in the dominance order precisely when for all $p\in\{1,\ldots,n\}$, $\lambda_1+\lambda_2+\ldots+\lambda_p\leq\mu_1+\mu_2+\ldots+\mu_p$, where we define $\lambda_j:=0$ for $j>k$ and $\mu_{j}:=0$ for $j>\ell$.
\end{example}
\subsection{The regular nilpotent orbit}\label{Subsection: The regular nilpotent orbit}
Having discussed nilpotent orbits in general, we turn our attention to a few particularly notable orbits. For the first of these, recall that the proof of Theorem \ref{Theorem: Finitely many nilpotent orbits} shows $\mathcal{N}$ to be a fibre of the adjoint quotient. It then follows from Theorem \ref{Theorem: Kostant's theorem} that $\mathcal{N}$ contains a unique regular adjoint orbit. This amounts to the existence of a unique regular nilpotent orbit, which we shall denote by $\mathcal{O}_{\text{reg}}$. Alternatively, one can characterize $\mathcal{O}_{\text{reg}}$ via the closure order as follows.
\begin{proposition}\label{Proposition: Unique maximal}
In the poset of nilpotent orbits with the closure order, $\mathcal{O}_{\emph{reg}}$ is the unique maximal element.
\end{proposition}
\begin{proof}
Let $\mathcal{O}$ be any nilpotent orbit. Since $\mathcal{O}_{\text{reg}}$ is dense in $\mathcal{N}$ (by Theorem \ref{Theorem: Kostant's theorem}), we have $\mathcal{O}\subseteq\mathcal{N}=\overline{\mathcal{O}_{\text{reg}}}$. Hence $\mathcal{O}\leq\mathcal{O}_{\text{reg}}$, as desired.
\end{proof}
\begin{example}
Recall that Example \ref{Example: Type A nilpotent orbits} used partitions of $n$ to index nilpotent $\SL_n(\mathbb{C})$-orbits, and Example \ref{Example: Nilpotent orbits and the dominance order} in turn mentioned the correspondence between the dominance order on partitions of $n$ and the closure order on nilpotent $\SL_n(\mathbb{C})$-orbits. Furthermore, one readily sees that the partition $\lambda=(n)$ is the unique maximal element in the dominance order. It follows from Proposition \ref{Proposition: Unique maximal} that $\mathcal{O}_{\text{reg}}\subseteq\mathfrak{sl}_n(\mathbb{C})$ is indexed by $\lambda=(n)$, meaning that $\mathcal{O}_{\text{reg}}$ is the conjugacy class of the $n\times n$ nilpotent Jordan block.
\end{example}
This last example raises the question of whether, in general, $\mathcal{O}_{\text{reg}}$ has any standard representatives. Kostant \cite{Kostant} answered this question in the affirmative, but we will need to recall a few facts about the representation theory of $\mathfrak{sl}_2(\mathbb{C})$ before elaborating on his argument. Let $\phi:\mathfrak{sl}_2(\mathbb{C})\rightarrow\mathfrak{gl}(V)$ be an $\mathfrak{sl}_2(\mathbb{C})$-representation, ie. $V$ is a finite-dimensional complex vector space and $\phi$ is a Lie algebra morphism. One knows such a representation to be completely reducible, meaning that $$V=\bigoplus_{k=1}^n V_k$$ for irreducible $\mathfrak{sl}_2(\mathbb{C})$-subrepresentations $V_k\subseteq V$. Now let $X,H,Y\in\mathfrak{sl}_2(\mathbb{C})$ be the usual generators, as defined in \eqref{Equation: Usual sl2 generators}. Each subrepresentation $V_k$ decomposes into eigenspaces of $\phi(H)$, which are known to be one-dimensional. The eigenvalues of $\phi(H):V_k\rightarrow V_k$ are integers and form a string as follows: if $\lambda_k$ is the largest eigenvalue, then $\lambda_k\geq 0$ and $-\lambda_k,-\lambda_k+2,-\lambda_k+4,\ldots,\lambda_k-2,\lambda_k$ is a complete list of the eigenvalues. Finally, the last facts we will need are the following decompositions of the kernels of $\phi(X):V\rightarrow V$ and $\phi(H):V\rightarrow V$:
\begin{equation}\label{Equation: Kernel decomposition}\ker(\phi(X))=\bigoplus_{k=1}^nV_k(\lambda_k),
\end{equation}
\begin{equation}\label{Equation: Second kernel decomposition}\ker(\phi(H))=\bigoplus_{k=1}^nV_k(0),
\end{equation}
where $V_k(\lambda_k)$ and $V_k(0)$ denote the $\lambda_k$ and $0$-eigenspaces of $\phi(H):V_k\rightarrow V_k$, respectively.\footnote{Note that $0$ need not be an eigenvalue, so that $V_k(0)=\{0\}$ may hold.}
\begin{proposition}\label{Proposition: Nilpotent orbit representative}
Recalling the notation used in Example \ref{Example: Kostant triple}, the nilpotent element $\xi=\sum_{\alpha\in\Pi}e_{\alpha}$ belongs to $\mathcal{O}_{\emph{reg}}$.
\end{proposition}
\begin{proof}
This proof will begin with two intermediate claims. For the first, recall that Example \ref{Example: Kostant triple} includes $\xi$ in an explicit $\mathfrak{sl}_2(\mathbb{C})$-triple $(\xi,h,\eta)$. We claim that every eigenvalue of $\adj_h:\mathfrak{g}\rightarrow\mathfrak{g}$ is an even integer. However, since each summand appearing in the root space decomposition \eqref{Equation: Root space decomposition} of $\mathfrak{g}$ is $\adj_h$-invariant, it will suffice to establish that $\adj_h$ acts on each summand with only even eigenvalues. Now, since $\adj_h(x)=0$ for all $x\in\mathfrak{t}$, we see that $0$ is the unique eigenvalue of $\adj_h:\mathfrak{t}\rightarrow\mathfrak{t}$. For the root space summands, suppose $\beta\in\Delta$. Note that for all $x\in\mathfrak{g}_{\beta}$, we have $\adj_h(x)=\beta(h)x$. To see that the eigenvalue $\beta(h)$ is an even integer, let us write $\beta=\sum_{\alpha\in\Pi}c_{\alpha}\alpha$ for $c_{\alpha}\in\mathbb{Z}$, $\alpha\in\Pi$. Example \ref{Example: Kostant triple} shows that $\alpha(h)=2$ for all $\alpha\in\Pi$, meaning that $\beta(h)=2\sum_{\alpha\in\Pi}c_{\alpha}$ is indeed an even integer. Hence every eigenvalue of $\adj_h:\mathfrak{g}\rightarrow\mathfrak{g}$ is an even integer.
Our second claim is that $C_{\mathfrak{g}}(h)$ and $C_{\mathfrak{g}}(\xi)$ have the same dimension. To see this, we restrict the adjoint representation $\adj:\mathfrak{g}\rightarrow\mathfrak{gl}(\mathfrak{g})$ to a representation of the subalgebra $\mathfrak{a}:=\text{span}\{\xi,h,\eta\}\cong\mathfrak{sl}_2(\mathbb{C})$ on $\mathfrak{g}$. One can now think of $\mathfrak{g}$ as an $\mathfrak{sl}_2(\mathbb{C})$-representation in which $X,H,Y\in\mathfrak{sl}_2(\mathbb{C})$ are identified with $\xi,h,\eta\in\mathfrak{a}$, respectively. Now, decompose $\mathfrak{g}$ into irreducible $\mathfrak{a}$-subrepresentations,
\begin{equation}\label{Equation: Decomposition into sl2 reps}
\mathfrak{g}=\bigoplus_{k=1}^nV_k.
\end{equation}
Since the eigenvalues of $\adj_h:\mathfrak{g}\rightarrow\mathfrak{g}$ are even integers, the same is necessarily true of the eigenvalues of $\adj_h$ on each irreducible subrepresentation $V_k$. The string description of the latter eigenvalues (explained just before this proposition) implies that $0$ occurs as an eigenvalue in each $V_k$. Since $V_k(0)$ is then one-dimensional for each $k$, \eqref{Equation: Second kernel decomposition} implies that $\dim(\ker(\adj_h))$ equals $n$, the number of irreducible summands in \eqref{Equation: Decomposition into sl2 reps}. A similar argument uses \eqref{Equation: Kernel decomposition} to show that $\dim(\ker(\adj_{\xi}))=n$. Of course, as $\ker(\adj_h)=C_{\mathfrak{g}}(h)$ and $\ker(\adj_{\xi})=C_{\mathfrak{g}}(\xi)$, we have actually shown that $\dim(C_{\mathfrak{g}}(h))=\dim(C_{\mathfrak{g}}(\xi))$.
We may now give a direct proof of our proposition, which is equivalent to the statement that $\mathcal{O}(\xi)=\mathcal{O}_{\text{reg}}$. Now, as $\xi$ is nilpotent and $\mathcal{O}_{\text{reg}}$ is the unique regular nilpotent orbit, this is equivalent to $\mathcal{O}(\xi)$ being a regular adjoint orbit. This is in turn equivalent to $\dim(C_G(\xi))=\text{rank}(G)$, since $\dim(\mathcal{O}(\xi))=\dim(G/C_G(\xi))=\dim(G)-\dim(C_G(\xi))$. At the same time, we know $\dim(C_G(\xi))=\dim(C_{\mathfrak{g}}(\xi))=\dim(C_{\mathfrak{g}}(h))=\dim(C_G(h))$. Our task is therefore to show that $\dim(C_{G}(h))=\text{rank}(G)$. However, the proof of Corollary \ref{Corollary: The existence of regular orbits} explains that $C_G(h)=L_{\Pi(h)}=T$ for our choice of $h$, so that the $\dim(C_{G}(h))=\dim(T)=\text{rank}(G)$. This completes the proof.
\end{proof}
\begin{remark}
Kostant found a curious relationship between the decomposition \eqref{Equation: Decomposition into sl2 reps} and the topology of $G$. To formulate it, note that the last two paragraphs in the proof of Proposition \ref{Proposition: Nilpotent orbit representative} can be used to show that \eqref{Equation: Decomposition into sl2 reps} has exactly $r:=\text{rank}(G)$ irreducible summands. If $d_1,d_2,\ldots,d_r$ are their respective dimensions, then the Poincar\'{e} polynomial of $G$ factors as follows:
\begin{equation}\label{Equation: Poincare polynomial}P_G(t)=\prod_{j=1}^r(1+t^{d_j})
\end{equation}
(see \cite[Cor. 8.7]{Kostant}).\footnote{Strictly speaking, Kostant's result is that \eqref{Equation: Poincare polynomial} is the Poincar\'{e} polynomial of the adjoint group of $\mathfrak{g}$. The adjoint group is just the quotient of $G$ by its (finite) centre, and one can use this fact to show that \eqref{Equation: Poincare polynomial} must also be the Poincar\'{e} polynomial of $G$.}
\end{remark}
\subsection{The minimal nilpotent orbit}\label{Subsection: The minimal nilpotent orbit}
Having studied $\mathcal{O}_{\text{reg}}$ as the unique maximal element in the poset of nilpotent orbits (see Proposition \ref{Proposition: Unique maximal}), it is natural to wonder about an analogous result at the opposite extreme --- the existence of a unique minimal nilpotent orbit. However, a first investigation yields underwhelming results. The singleton $\{0\}$ is the unique minimal nilpotent orbit, since Proposition \ref{Proposition: Characterizations of nilpotent orbits} implies that $\{0\}\subseteq\overline{\mathcal{O}}$ for all nilpotent orbits $\mathcal{O}$. It is therefore a natural next step to study the minimal non-zero nilpotent orbits. Interestingly, there turns out to be a unique such orbit when $G$ is simple. At the same time, $G$ being simple implies the existence of a unique $\theta\in\Delta$ with the property of being maximal among all roots in the partial order \eqref{Equation: Partial order on the weight lattice} (see \cite[Thm. 18.9.2]{Tauvel}). We then have the following theorem, which is essentially a restatement of Theorem 4.3.3 in \cite{Collingwood}.
\begin{theorem}\label{Theorem: Existence of a minimal nilpotent orbit}
If $G$ is simple and $e_{\theta}\in\mathfrak{g}_{\theta}\setminus\{0\}$, then $\mathcal{O}(e_{\theta})$ is the unique minimal element in the poset of non-zero nilpotent orbits.
\end{theorem}
One denotes the above orbit by $\mathcal{O}_{\text{min}}$ and calls it the \textit{minimal nilpotent orbit}. Note that being minimal in the poset of non-zero nilpotent orbits amounts to the condition $\overline{\mathcal{O}_{\text{min}}}=\mathcal{O}_{\text{min}}\cup\{0\}$.
\begin{example}
Let $G=\SL_n(\mathbb{C})$, a simple group. Recall from Example \ref{Example: Nilpotent orbits and the dominance order} the poset isomorphism between nilpotent $\SL_n(\mathbb{C})$-orbits in the closure order and partitions of $n$ in the dominance order. While $\lambda=(1,1,\ldots,1)$ (which corresponds to the orbit $\{0\}$) is clearly the unique minimal partition of $n$, $\mu=(2,1,1,\ldots,1)$ is seen to be minimal once $\mu$ is excluded. It follows that $\mu$ corresponds to $\mathcal{O}_{\text{min}}\subseteq\mathfrak{sl}_n(\mathbb{C})$, or equivalently that $\mathcal{O}_{\text{min}}$ is the conjugacy class of a nilpotent Jordan canonical form matrix having a single $2\times 2$ Jordan block and $0$ for all other entries.
\end{example}
Theorem \ref{Theorem: Existence of a minimal nilpotent orbit} assumes $G$ is simple, so one might legitimately ask whether a unique minimal non-zero nilpotent orbit exists when $G$ is semisimple. The following example shows this to be false in general.
\begin{example}
Assume that $G$ is simple, in which case the product $G\times G$ is semisimple but not simple. The latter group has Lie algebra $\mathfrak{g}\oplus\mathfrak{g}$, and the adjoint representation $\Adj:G\times G\rightarrow\GL(\mathfrak{g}\oplus\mathfrak{g})$ is as follows:
$$\Adj_{(g_1,g_2)}(x_1,x_2)=(\Adj_{g_1}(x_1),\Adj_{g_2}(x_2))$$ for all $(g_1,g_2)\in G\times G$ and $(x_1,x_2)\in\mathfrak{g}\oplus\mathfrak{g}$, where (via a slight abuse of notation) $\Adj_{g_1}(x_1)$ and $\Adj_{g_2}(x_2)$ come from the adjoint representation of $G$. Now, let $\theta\in\Delta$ and $e_{\theta}\in\mathfrak{g}_{\theta}\setminus\{0\}$ be as discussed above. It follows that the adjoint $G\times G$-orbits of $(e_{\theta},0)$ and $(0,e_{\theta})$ are $$\mathcal{O}_{G\times G}(e_{\theta},0)=\mathcal{O}_{\text{min}}\times\{0\}\quad\text{and}\quad\mathcal{O}_{G\times G}(0,e_{\theta})=\{0\}\times\mathcal{O}_{\text{min}},$$
respectively, where $\mathcal{O}_{\text{min}}\subseteq\mathfrak{g}$ is the minimal nilpotent $G$-orbit. Taking closures, we have
$$\overline{\mathcal{O}_{G\times G}(e_{\theta},0)}=\overline{\mathcal{O}_{\text{min}}}\times\{0\}=(\mathcal{O}_{\text{min}}\cup\{0\})\times\{0\}=\mathcal{O}_{G\times G}(e_{\theta},0)\cup\{(0,0)\}$$ and
$$\overline{\mathcal{O}_{G\times G}(0,e_{\theta})}=\{0\}\times\overline{\mathcal{O}_{\text{min}}}=\{0\}\times(\mathcal{O}_{\text{min}}\cup\{0\})=\{(0,0)\}\cup\mathcal{O}_{G\times G}(0,e_{\theta}).$$ These calculations establish two things. Firstly, since the closures of $\mathcal{O}_{G\times G}(e_{\theta},0)$ and $\mathcal{O}_{G\times G}(0,e_{\theta})$ contain $(0,0)$, each orbit is nilpotent (see Proposition \ref{Proposition: Characterizations of nilpotent orbits}). Secondly, for each orbit, $\{(0,0)\}$ is the unique orbit that is strictly smaller in the closure order. Equivalently, $\mathcal{O}_{G\times G}(e_{\theta},0)$ and $\mathcal{O}_{G\times G}(0,e_{\theta})$ are minimal among the non-zero nilpotent orbits in the closure order. It is nevertheless clear that $\mathcal{O}_{G\times G}(e_{\theta},0)\neq\mathcal{O}_{G\times G}(0,e_{\theta})$.
\end{example}
\subsection{Orbit projectivizations}\label{Subsection: Orbit projectivizations}
Recall from Proposition \ref{Proposition: Characterizations of nilpotent orbits} that one can characterize nilpotent orbits as those adjoint orbits which are invariant under the dilation action of $\mathbb{C}^*$ on $\mathfrak{g}$. We will see that this fact gives rise to a special geometry enjoyed only by the nilpotent orbits. Indeed, given a non-zero nilpotent orbit $\mathcal{O}\subseteq\mathfrak{g}$, we may define its \textit{projectivization} to be the quotient
$$\mathbb{P}(\mathcal{O}):=\mathcal{O}/\mathbb{C}^*.$$ Note that $\mathbb{P}(\mathcal{O})$ is, by construction, a subset of the projective space $\mathbb{P}(\mathfrak{g})=(\mathfrak{g}\setminus\{0\})/\mathbb{C}^*$. The latter is a $G$-variety with action map
$$G\times\mathbb{P}(\mathfrak{g})\rightarrow\mathbb{P}(\mathfrak{g}),\quad (g,[x])\mapsto[\Adj_g(x)],\quad g\in G, \text{ }[x]\in\mathbb{P}(\mathfrak{g})$$
(cf. Example \ref{Example: Descent to projective space}), and it is not difficult to check that $\mathbb{P}(\mathcal{O})$ is a $G$-orbit in $\mathbb{P}(\mathfrak{g})$. As such, $\mathbb{P}(\mathcal{O})$ is a smooth locally closed subvariety of $\mathbb{P}(\mathfrak{g})$ (see \ref{Subsection: Generalities on orbits}). However, $\mathbb{P}(\mathcal{O})$ is seldom closed.
\begin{proposition}\label{Proposition: Unique closed orbit}
Assume that $G$ is simple and let $\mathcal{O}\subseteq\mathfrak{g}$ be a non-zero nilpotent orbit. Then $\mathbb{P}(\mathcal{O})$ is closed in $\mathbb{P}(\mathfrak{g})$ if and only if $\mathcal{O}=\mathcal{O}_{\emph{min}}$.
\end{proposition}
\begin{proof}
The nilpotent cone $\mathcal{N}\subseteq\mathfrak{g}$ is $\mathbb{C}^*$-invariant, and one may define $$\mathbb{P}(\mathcal{N}):=(\mathcal{N}\setminus\{0\})/\mathbb{C}^*.$$ It is not difficult to see that $\mathbb{P}(\mathcal{N})$ is a $G$-invariant closed subvariety of $\mathbb{P}(\mathfrak{g})$ whose $G$-orbits are precisely the projectivizations of the non-zero nilpotent orbits. It will therefore be equivalent to prove that $\mathbb{P}(\mathcal{O}_{\text{min}})$ is the unique closed orbit in $\mathbb{P}(\mathcal{N})$.\footnote{Strictly speaking, this will only show $\mathbb{P}(\mathcal{O})$ to be closed if and only if $\mathbb{P}(\mathcal{O})=\mathbb{P}(\mathcal{O}_{\text{min}})$. We leave it as an exercise to check that $\mathbb{P}(\mathcal{O})=\mathbb{P}(\mathcal{O}_{\text{min}})$ if and only if $\mathcal{O}=\mathcal{O}_{\text{min}}$.} Now, remember that $\overline{\mathbb{P}(\mathcal{O})}$ is a union of $\mathbb{P}(\mathcal{O})$ and the $G$-orbits lying below $\mathbb{P}(\mathcal{O})$ in the closure order. Showing $\mathbb{P}(\mathcal{O}_{\text{min}})$ to be the unique closed orbit is therefore equivalent to showing it to be the unique minimal orbit in $\mathbb{P}(\mathcal{N})$. We shall do the latter.
Let $\mathcal{O}\subseteq\mathfrak{g}$ be any non-zero nilpotent orbit. Since $\mathcal{O}_{\text{min}}$ is the unique minimal non-zero nilpotent orbit in the closure order, it must be true that $\mathcal{O}_{\text{min}}\subseteq\overline{\mathcal{O}}\setminus\{0\}$. Also, because $\mathcal{N}$ is closed in $\mathfrak{g}$ and contains $\mathcal{O}$, $\overline{\mathcal{O}}\setminus\{0\}$ is the closure of $\mathcal{O}$ in $\mathcal{N}\setminus\{0\}$. The continuity of the quotient map $\pi:\mathcal{N}\setminus\{0\}\rightarrow\mathbb{P}(\mathcal{N})$ therefore implies $\pi(\overline{\mathcal{O}}\setminus\{0\})\subseteq\overline{\pi(\mathcal{O})}$, and we have
$$\mathbb{P}(\mathcal{O}_{\text{min}})=\pi(\mathcal{O}_{\text{min}})\subseteq\pi(\overline{\mathcal{O}}\setminus\{0\})\subseteq\overline{\pi(\mathcal{O})}=\overline{\mathbb{P}(\mathcal{O})}.$$
In other words, $\mathbb{P}(\mathcal{O}_{\text{min}})$ belongs to the closure of $\mathbb{P}(\mathcal{O})$ for all non-zero nilpotent orbits $\mathcal{O}$. We conclude that $\mathbb{P}(\mathcal{O}_{\text{min}})$ is the unique minimal orbit in $\mathbb{P}(\mathcal{N})$, as desired.
\end{proof}
Seeking to build on Proposition \ref{Proposition: Unique closed orbit}, we will assume $G$ to be simple for the duration of \ref{Subsection: Orbit projectivizations}.
\begin{remark}
Recall that the image of a projective variety under a morphism is necessarily closed. If $\mathbb{P}(\mathcal{O})$ is projective, we may take this morphism to be the inclusion $\mathbb{P}(\mathcal{O})\hookrightarrow\mathbb{P}(\mathfrak{g})$ and conclude that $\mathbb{P}(\mathcal{O})$ is closed in $\mathbb{P}(\mathfrak{g})$. Conversely, if $\mathbb{P}(\mathcal{O})$ is closed in $\mathbb{P}(\mathfrak{g})$, then the former is necessarily projective. These last two sentences explain that $\mathbb{P}(\mathcal{O})$ is projective if and only if it is closed in $\mathbb{P}(\mathfrak{g})$, which by Proposition \ref{Proposition: Unique closed orbit} is equivalent to $\mathcal{O}=\mathcal{O}_{\text{min}}$. In particular, $\mathbb{P}(\mathcal{O})$ is generally not projective.
\end{remark}
While Proposition \ref{Proposition: Unique closed orbit} is an important first step in understanding $\mathbb{P}(\mathcal{O}_{\text{min}})$, one can say a great deal more. Indeed, as $\mathbb{P}(\mathcal{O}_{\text{min}})$ is a $G$-orbit, it must be $G$-equivariantly isomorphic to $G/P$ for some closed subgroup $P\subseteq G$. Furthermore, since $\mathbb{P}(\mathcal{O}_{\text{min}})$ is projective, $P$ is necessarily a parabolic subgroup of $G$ (see \cite[Cor. 28.1.4]{Tauvel}). This subgroup contains a conjugate of $B$, meaning that $P$ is conjugate to a standard parabolic subgroup $P_S$ (as defined in \ref{Subsection: The basic objects}). One can use this to verify that $G/P\cong G/P_S$, so that $\mathbb{P}(\mathcal{O}_{\text{min}})$ and $G/P_S$ are $G$-equivariantly isomorphic for some subset $S\subseteq\Pi$. The following proposition shows that one can take $S$ to be
$$\Pi(\theta):=\{\alpha\in\Pi:\langle\alpha,\theta\rangle=0\},$$ the set of simple roots that are orthogonal to the maximal root $\theta$.
\begin{proposition}\label{Proposition: Description of P(Omin) as a partial flag variety}
If $G$ is simple, then there is a $G$-variety isomorphism $G/P_{\Pi(\theta)}\cong\mathbb{P}(\mathcal{O}_{\emph{min}})$.
\end{proposition}
\begin{proof}
Fix $e_{\theta}\in\mathfrak{g}_{\theta}\setminus\{0\}$ and recall that $e_{\theta}\in\mathcal{O}_{\text{min}}$ (see Theorem \ref{Theorem: Existence of a minimal nilpotent orbit}). It follows that $[e_{\theta}]\in\mathbb{P}(\mathcal{O}_{\text{min}})$, so that $\mathbb{P}(\mathcal{O}_{\text{min}})$ is the $G$-orbit of $[e_{\theta}]$ in $\mathbb{P}(\mathfrak{g})$. We claim that $C_G([e_{\theta}])=P_{\Pi(\theta)}$, which together with the orbit-stabilizer isomorphism \eqref{Equation: Orbit-stabilizer isomorphism} will prove the proposition.
Since $G$ is simple, its adjoint representation is irreducible (see \cite[Sect. 0.11]{HumphreysConjugacy}) and $\theta$ is the maximal weight of this representation. It follows that $\Adj(b)(e_{\theta})$ is a non-zero multiple of $e_{\theta}$ for each $b\in B$ (see \cite[Sect. 31.3]{Humphreys}). This shows that $B\subseteq C_G([e_{\theta}])$, meaning that $C_G([e_{\theta}])$ is a standard parabolic subgroup of $G$. In other words, $C_G([e_{\theta}])=P_S$ for some subset $S\subseteq\Pi$, and we must show $S=\Pi(\theta)$. Now by the correspondence between subsets of $\Pi$ and standard parabolic subgroups (see \ref{Subsection: The basic objects}), $S$ is the set of simple roots $\alpha$ for which $\mathfrak{g}_{-\alpha}$ belongs to $C_{\mathfrak{g}}([e_{\theta}])$, the Lie algebra of $C_G([e_{\theta}])$. One can show that \\
$C_{\mathfrak{g}}([e_{\theta}])=\{\xi\in\mathfrak{g}:[\xi,\mathfrak{g}_{\theta}]\subseteq \mathfrak{g}_{\theta}\}$, so that $\alpha\in\Pi$ belongs to $S$ if and only if $[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\theta}]\subseteq\mathfrak{g}_{\theta}$. Furthermore, as $[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\theta}]\cap\mathfrak{g}_{\theta}=\{0\}$, it follows that $\alpha$ belongs to $S$ if and only if $[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\theta}]=\{0\}$. Hence, showing $S=\Pi(\theta)$ amounts to proving that $\alpha\in\Pi$ satisfies $[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\theta}]=\{0\}$ if and only if $\langle\alpha,\theta\rangle=0$.
Assume that $[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\theta}]=\{0\}$.
Since $\theta$ is the maximal root, $\alpha+\theta$ is not a root and we have
$[\mathfrak{g}_{\alpha},\mathfrak{g}_{\theta}]\subseteq\mathfrak{g}_{\alpha+\theta}=\{0\}$. Applying the Jacobi identity, we see that
\begin{equation}\label{Equation: Three Root Spaces}
[[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\alpha}],\mathfrak{g}_{\theta}]=\{0\}.
\end{equation} Letting $h_{\alpha}\in[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\alpha}]$ be the coroot associated with $\alpha$, \eqref{Equation: Three Root Spaces} gives
$$0=[h_{\alpha},e_{\theta}]=\theta(h_{\alpha})e_{\theta}=2\frac{\langle\alpha,\theta\rangle}{\langle\alpha,\alpha\rangle}e_{\theta}.$$ It follows that $\langle\alpha,\theta\rangle=0$.
Conversely, assume that $\langle\alpha,\theta\rangle=0$. If $[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\theta}]\neq\{0\}$, then $\mathfrak{g}_{\theta-\alpha}\neq\{0\}$ and $\theta-\alpha$ is a root. Since the set of roots is invariant under the action of $W$ on $\mathfrak{t}^*$, $s_{\alpha}\cdot(\theta-\alpha)$ is also a root. However, $$s_{\alpha}\cdot(\theta-\alpha)=s_{\alpha}\cdot\theta-s_{\alpha}\cdot\alpha=\theta-2\frac{\langle\alpha,\theta\rangle}{\langle\alpha,\alpha\rangle}\alpha+\alpha=\theta+\alpha,$$ and this contradicts the maximality of $\theta$. We conclude that $[\mathfrak{g}_{-\alpha},\mathfrak{g}_{\theta}]=\{0\}$.
\end{proof}
\begin{example}\label{Example: Description of P(Omin) in type A}
One can use Proposition \ref{Proposition: Description of P(Omin) as a partial flag variety} to give a more classical description of $\mathbb{P}(\mathcal{O}_{\text{min}})$ when $G=\SL_n(\mathbb{C})$. To do this, recall the $\SL_n(\mathbb{C})$-specific discussion of the maximal torus, Borel subgroup, roots, simple roots, etc. given in Example \ref{Example: The type A setup}. It turns out that $\theta=t_1-t_n$ is the maximal root, which by the Killing form description \eqref{Equation: Type A Killing form} implies that \\ $\Pi(\theta)=\{t_i-t_{i+1}:2\leq i\leq n-2\}$. The set $\Delta_{\Pi(\theta)}^{-}$ of negative roots expressible as linear combinations of the roots in $\Pi(\theta)$ is $\Delta_{\Pi(\theta)}^{-}=\{t_i-t_j:2\leq j<i\leq n-1\}$. Using \eqref{Equation: Lie algebra of standard parabolic} to compute $\mathfrak{p}_{\Pi(\theta)}$, one finds the corresponding parabolic subgroup $P_{\Pi(\theta)}\subseteq\SL_n(\mathbb{C})$ to be
\begin{equation}\label{Equation: Type A parabolic for P(Omin)}P_{\Pi(\theta)}=\{(A_{ij})\in\SL_n(\mathbb{C}):A_{i1}=0\text{ for all }i\geq 2\text{ and }A_{nj}=0\text{ for all }j\leq n-1\}.
\end{equation}
In other words, a matrix $A\in\SL_n(\mathbb{C})$ belongs to $P_{\Pi(\theta)}$ if and only if it is block upper-triangular with blocks of dimensions $1\times 1$, $(n-2)\times (n-2)$, and $1\times 1$ along the diagonal (read from top to bottom).
Now, recall that Example \ref{Example: Stabilizer in a flag variety} computed the $\GL_n(\mathbb{C})$-stabilizer of the flag $V_{\bullet}^e\in\text{Flag}(d_{\bullet};\mathbb{C}^n)$. Note that the $\SL_n(\mathbb{C})$-stabilizer of $V_{\bullet}^e$ is obtained by intersecting $\SL_n(\mathbb{C})$ with the $\GL_n(\mathbb{C})$-stabilizer. Furthermore, when $d_{\bullet}=(1,n-1)$, it is not difficult to see that $C_{\SL_n(\mathbb{C})}(V_{\bullet}^e)$ coincides with \eqref{Equation: Type A parabolic for P(Omin)}. The action of $\SL_n(\mathbb{C})$ on $\text{Flag}(1,n-1;\mathbb{C}^n)$ is transitive, so that we have
$\text{Flag}(1,n-1;\mathbb{C}^n)\cong\SL_n(\mathbb{C})/C_{\SL_n(\mathbb{C})}(V_{\bullet}^e)=\SL_n(\mathbb{C})/P_{\Pi(\theta)}$. Proposition \eqref{Proposition: Description of P(Omin) as a partial flag variety} identifies $\SL_n(\mathbb{C})/P_{\Pi(\theta)}$ with $\mathbb{P}(\mathcal{O}_{\text{min}})$, meaning that we have an $\SL_n(\mathbb{C})$-variety isomorphism
$$\mathbb{P}(\mathcal{O}_{\text{min}})\cong\text{Flag}(1,n-1;\mathbb{C}^n).$$
\end{example}
\begin{remark}
Using similar ideas, one can obtain classical descriptions of $\mathbb{P}(\mathcal{O}_{\text{min}})$ in other Lie types. In type $D_n$, for instance, $\mathbb{P}(\mathcal{O}_{\text{min}})$ identifies with the Grassmannian of isotropic $2$-dimensional subspaces in $\mathbb{C}^{2n}$ (see \cite[Sect. 4.2]{CrooksRayan})
\end{remark}
On a somewhat different note, $\mathbb{P}(\mathcal{O}_{\text{min}})$ turns out to feature prominently in \textit{quaternionic-K\"{a}hler geometry} (see \cite{Swann} for a brief introduction). To develop this point, let $X$ be a complex manifold and $E$ a holomorphic subbundle of $TX$. One then has a short exact sequence of holomorphic bundles on $X$,
$$0\rightarrow E\rightarrow TX\xrightarrow{\theta} F\rightarrow 0,$$ where $F:=TX/E$ and $\theta$ is the quotient map. Let $\mathcal{E}$ and $\mathcal{F}$ denote the sheaves of sections of $E$ and $F$, respectively, and consider the following pairing:
\begin{equation}\label{Equation: Pairing}\mathcal{E}\otimes \mathcal{E}\rightarrow \mathcal{F},\quad(s_1,s_2)\mapsto \theta([s_1,s_2]).\footnote{Note that $s_1$ and $s_2$ may be regarded as vector fields via the inclusion $E\subseteq TX$, so that $[s_1,s_2]$ makes sense.}\end{equation}
One calls $E$ a \textit{complex contact structure} if $E$ has co-rank one in $TX$ and the pairing \eqref{Equation: Pairing} is non-degenerate (on both local and global sections of $E$). Once endowed with a complex contact structure, $X$ is called a \textit{complex contact manifold}.
Now, let $\mathcal{O}\subseteq\mathfrak{g}$ be a non-zero nilpotent orbit and recall the canonical symplectic form $\omega_{\mathcal{O}}\in\Omega^2(\mathcal{O})$ defined in \eqref{Equation: The canonical symplectic form}. One can use $\omega_{\mathcal{O}}$ to induce a canonical complex contact structure on $\mathbb{P}(\mathcal{O})$ (see \cite[Prop. 2.2, Rmk. 2.3, Sect. 2.4]{Beauville}\footnote{Note that some of these are phrased in the language of coadjoint orbits, rather than adjoint orbits. However, recall from \ref{Subsection: Geometric features} that one can identify coadjoint orbits with adjoint orbits.}), so that $\mathbb{P}(\mathcal{O})$ is a complex contact manifold. In particular, $\mathbb{P}(\mathcal{O}_{\text{min}})$ belongs to the class of Fano contact manifolds (see \cite[Def. 1.3, Def. 2.5]{LeBrunSalamon}), which are of interest to quaternionic-K\"{a}hler geometers (see \cite[Sect. 1]{LeBrunSalamon}). Moreover, it is in the quaternionic-K\"{a}hler context that LeBrun and Salamon conjectured the following classification of Fano contact manifolds: each is isomorphic, as a complex contact manifold, to the projectivization of the minimal nilpotent orbit of a simple algebraic group. This conjecture is known to be true in a number of cases (see \cite{Beauville}), and counter-examples have not yet been found.
\begin{remark}
When $\mathbb{P}(\mathcal{O}_{\text{min}})$ is described in classical terms, as in Example \ref{Example: Description of P(Omin) in type A}, one should expect a similarly classical description of its complex contact structure. Such a description is given in \cite[Sect. 4.2]{CrooksRayan} in Lie type $D_n$, when $\mathbb{P}(\mathcal{O}_{\text{min}})$ is viewed as the Grassmannian of isotropic $2$-dimensional subspaces of $\mathbb{C}^{2n}$.
\end{remark}
\subsection{Orbit closures and singularities}\label{Subsection: Orbit closures and singularities}
It turns out that nilpotent orbit closures have been studied extensively in the literature, both intrinsically and as parts of broader research programs. We now briefly survey some of this literature, giving references where necessary in order to avoid long digressions.
Recall that the closure of $\mathcal{O}_{\text{reg}}$ is the nilpotent cone $\mathcal{N}$. This is singular variety and there exists a canonical resolution of singularities $\mu:T^*(G/B)\rightarrow\mathcal{N}$ \footnote{By ``resolution of singularities'', we mean that $\mu$ is a proper surjective morphism with the property of being an isomorphism over the smooth locus of $\mathcal{N}$.}, called the \textit{Springer resolution} (see \cite[Def. 3.2.4]{Chriss}\footnote{Note that this reference substitutes $G/B$ with the variety of all Borel subalgebras of $\mathfrak{g}$.}). While fundamental to the study of singularities in $\mathcal{N}$, $\mu$ also plays a significant role in adjacent subjects. In geometric representation theory, one constructs a representation of the Weyl group on the Borel-Moore homology of each fibre of $\mu$ (see \cite[Sect. 3.4--3.6]{Chriss}). In symplectic geometry, $T^*(G/B)$ and $\mathcal{O}_{\text{reg}}$ have canonical symplectic forms (see \cite[Sect. 1.1]{Chriss} for the former and \ref{Subsection: Geometric features} for the latter) and $\mu$ is an example of a \textit{symplectic resolution} (see \cite[Sect. 1.3]{Fu} for a definition, and all of \cite{Fu} for an introduction to the subject). This turns out to be an instance of a more general occurrence: for certain nilpotent orbits $\mathcal{O}\subseteq\mathfrak{g}$, one can find a parabolic subgroup $P\subseteq G$ and a symplectic resolution $T^*(G/P)\rightarrow\overline{\mathcal{O}}$ (see \cite[Sect. 5.1]{Fu}\footnote{This reference focuses on resolutions of normalizations of orbit closures, but many closures happen to be normal (see \cite{Kraft}, for example).}). These nilpotent orbit closures thereby give examples of \textit{symplectic singularities} (see \cite[Def. 1.1]{BeauvilleSingular}), which have received a great deal of recent attention (see \cite{BeauvilleSingular,FuNamikawa,Ginzburg,Kaledin}).
In addition to the above, nilpotent orbits closures are sometimes studied through their intersections with certain affine-linear subsets of $\mathfrak{g}$, often called \textit{Slodowy slices} (see \cite[Prop. 3.7.15]{Chriss}). This study was initiated by the works of Brieskorn \cite{Brieskorn} and Slodowy \cite{Slodowy}, which show the intersection of $\overline{\mathcal{O}_{\text{reg}}}=\mathcal{N}$ with a particular Slodowy slice to contain a Kleinian surface singularity when $\mathfrak{g}$ is of type $ADE$ (see \cite[Sect. 6.4]{Slodowy} for further details). This result is interesting in many contexts, and in particular plays a role in Kronheimer's study of nilpotent orbits in hyperk\"{a}hler geometry (see \cite[Prop. 2(a), Prop. 2(b)]{KronheimerInstantons}).
\bibliographystyle{acm}
|
\section{Introduction}
\label{intro}
A nova eruption occurs when a white dwarf has accreted enough material from a companion star to trigger thermonuclear runaway on its surface (e.g., Starrfield et al. 1972).
Observations with modern telescopes have revealed the presence of strong shocks in nova systems. In particular, the discovery of GeV $\gamma$-rays from V407 Cyg indicated that in at least some systems, the shocks were strong enough to accelerate particles to very high energies (Abdo et al. 2010). In the case of V407 Cyg, the shocks were the result of the nova ejecta colliding with a dense stellar wind from a Mira variable companion (Abdo et al. 2010). Much to the surprise of the nova community, further observations with the \emph{Fermi Gamma-ray Space Telescope} discovered $\gamma$-ray emission from systems with main sequence companions (e.g., Ackermann et al. 2014).
While observations across the electromagnetic spectrum have contributed to the discovery of shocks in novae, radio observations have proven to be particularly valuable. Resolved radio images can reveal non-spherical structure and provide evidence for multiple outflows, as in the case of V959 Mon (Chomiuk et al. 2014). Early monitoring of the radio emission, even when the ejecta are too small to resolve, has also prove to be an effective probe of shocks in novae. The shocks can manifest as synchrotron emission with non-thermal brightness temperatures in spectacular ways as in V1723 Aql (Weston et al. 2016a), or more subtle ways as in V5589 Sgr (Weston et al. 2016b). When the radio synchrotron emission is bright enough, the shocks can be directly imaged with very long baseline interferometry, as in V959 Mon (Chomiuk et al. 2014) and RS Oph (Rupen et al. 2008; Sokoloski et al. 2008).
Observations with modern X-ray telescopes have also proven to be useful in characterizing shocks in novae. Those novae embedded within the wind of a giant companion typically emitting bright, hard X-rays early in their evolution, such as RS Oph (Bode et al. 2006; Nelson et al. 2008) and V745 Sco (Page et al. 2015; Orio et al. 2015). Novae in systems with main sequence companions are also capable of producing strong X-rays, but typically later in their evolution Such was the case in both V382 Vel (Mukai \& Ishida 2001) and V5589 Sgr (Weston et al. 2016b).
On 2015 February 11, the nova V1535 Sco was discovered in the constellation Scorpius. It was reported in vsnet-alert 18276 by P. Schmeer and in CBET 4078 by T. Kojina. Early optical and near infrared follow-up indicated the nova was fading rapidly (Walter 2015). The combination of rapid fading and a possible bright near-infrared counterpart indicated that the companion star could be an M giant, in which case the nova could be embedded in the wind from a giant companion (Walter 2015). Srivastava et al. (2015) later argued that the companion could be a K giant based on pre-eruption 2MASS photometry of a likely counterpart (2MASS J17032617-3504178). Spectroscopic observations indicated it was a He/N type nova (Walter 2015; Srivastava et al. 2015). Early X-ray and radio observations (Nelson et al. 2015) indicated the presence of strong shocks in the ejecta, which was further evidence that the ejecta were expanding into a dense medium. The nova was originally given the designation PNV J1703260-3504140, then known as Nova Sco 2015, and eventually given the official designation V1535 Sco. Srivastava et al. (2015) also noted that narrowing of the H-line profiles in the infrared indicated the presence of a deceleration shock, and estimated the total mass ejected to be between $4.5\times10^{-6}$ and $2.6\times10^{-4}$ $\epsilon$ M$_{\odot}$ (where $\epsilon$ is the filling factor of the ejecta).
During the 2015 outburst, no \textit{Fermi Gamma-ray Space Telescope} staring mode observations were scheduled and the nova was not reported as a detection from the survey mode data. However, it must be noted that \textit{Fermi} is generally less sensitive to transients at the Galactic Center due to a combination of the high background, and the north-south ``rocking'' profile of the survey mode results in the Galactic Center having minimal exposure (Acero et al. 2015). It is therefore possible that V1535 Sco produced $\gamma$-rays like other novae (e.g., V407 Cyg and V745 Sco; Abdo et al. 2010; Cheung et al. 2014), but none were detectable due to these limiting factors. Recent work by Morris et al. (2017) argues that all novae produce $\gamma$-ray emission, but we mainly detect the nearby ones with \emph{Fermi}.
In this paper, we present our multi-wavelength observations of V1535 Sco made during 2015. These observations were made with the Karl J. Jansky Very Large Array (VLA), the Very Long Baseline Array (VLBA), the \textit{Swift} X-ray Telescope (XRT), and the Small \& Moderate Aperture Research Telescope System (SMARTS).
We present our observations and data reduction methods in Section 2. Our knowledge about the distance to V1535 Sco is discussed in Section 3. In Section 4, we present results from the multi-wavelength observations. We discuss our findings in Section 5. Final conclusions are summarized in Section 6. Throughout this paper, we use the initial optical detection of the nova on 2015 February 11.837 (MJD 57064.837) as Day 0.0. We also use the term ``embedded nova'' to refer to any nova embedded in the wind material from its companion star (e.g. Chomiuk et al. 2012; Mukai et al. 2014).
\section{Observations and Data Reduction}
\subsection{VLA Observations}
We began monitoring V1535 Sco with the VLA within 64 hours of its discovery in the optical. The first epoch only included observations at C-band (4.0 - 8.0 GHz) and Ka-band (26.5 - 40.0 GHz). All following epochs also included observations at L-band (1.0-2.0 GHz) and Ku-band (12.0 - 18.0 GHz). To maximize spectral coverage, all bands were split into upper and lower sidebands. For the first 8 epochs, we used 8-bit sampling for all frequencies. This gave us a total bandwidth of 1.0 GHz at L-band, and 2.0 GHz at all other bands. For the ninth epoch (Day 93.463) we used 3-bit sampling for Ku-band and Ka-band, resulting in total bandwidths of 6.0 GHz and 8.0 GHz, respectively. For the final epoch (Day 122.363), we used 3-bit sampling for C-band, Ku-band, and Ka-band, giving total bandwidths of 4.0 GHz, 6.0 GHz, and 8.0 GHz, respectively. 10 epochs were observed under the program VLA/13B-057, and one epoch (Day 93.463) was observed under the program S61420. The nova was detected during the first epoch (Day 2.7) and remained detectable at multiple frequencies for nearly 100 days. We ceased observations once the nova had faded to the point where it was no longer detected at all frequencies.
\begin{center}
\begin{deluxetable*}{ccccccc}
\tablewidth{0 pt}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.025in}
\tablecaption{ \label{radiotab}
VLA Observations}
\tablehead{UT Date & Day\tablenotemark{1} & VLA & Central Frequency & Time on Source & Flux Density & Uncertainty \\
& & Config & (GHz) & (minutes) & (mJy) & (mJy) }
\startdata
2015-02-14 & 2.663 & B & 4.55 & 13.6 & 4.13 & 0.13 \\
& & & 7.38 & 13.6 & 2.786 & 0.085 \\
& & & 28.2 & 11.3 & 0.819 & 0.071 \\
& & & 36.5 & 11.3 & 0.675 & 0.090 \\
2015-02-18 & 6.663 & B & 13.5 & 11.9 & 0.416 & 0.021 \\
& & & 17.4 & 11.9 & 0.344 & 0.024 \\
& & & 28.2 & 11.8 & 0.295 & 0.053 \\
& & & 36.5 & 11.8 & 0.376 & 0.074 \\
2015-02-19 & 7.663 & B & 1.26 & 15.6 & 1.57 & 0.090 \\
& & & 1.74 & 15.6 & 1.21 & 0.081 \\
& & & 4.55 & 13.6 & 0.650 & 0.026 \\
& & & 7.38 & 13.6 & 0.439 & 0.018 \\
2015-02-24 & 12.763 & B & 13.5 & 12.5 & 0.385 & 0.020 \\
& & & 17.4 & 12.5 & 0.456 & 0.024 \\
& & & 28.2 & 12.4 & 0.785 & 0.062 \\
& & & 36.5 & 12.4 & 0.845 & 0.087 \\
2015-03-01 & 17.663 & B & 1.26 & 15.6 & \textless0.317 & 0.10 \\
& & & 1.74 & 15.6 & \textless0.317 & 0.066 \\
& & & 4.55 & 13.6 & 0.221 & 0.019 \\
& & & 7.38 & 13.6 & 0.192 & 0.017 \\
2015-03-07 & 23.663 & B & 13.5 & 12.0 & 0.585 & 0.024 \\
& & & 17.4 & 12.0 & 0.755 & 0.031 \\
& & & 28.2 & 11.9 & 1.15 & 0.074 \\
& & & 36.5 & 11.9 & 1.68 & 0.11 \\
2015-03-10 & 26.563 & B & 1.26 & 15.6 & \textless0.374 & 0.074 \\
& & & 1.74 & 15.6 & \textless0.271 & 0.054 \\
& & & 4.55 & 13.5 & 0.455 & 0.023 \\
& & & 7.38 & 13.5 & 0.536 & 0.021 \\
2015-03-25 & 41.583 & B & 13.5 & 12.0 & 0.267 & 0.015 \\
& & & 17.4 & 12.0 & 0.299 & 0.017 \\
& & & 28.2 & 12.0 & 0.348 & 0.044 \\
& & & 36.5 & 12.0 & \textless0.351 & 0.055 \\
2015-03-25 & 41.633 & B & 1.26 & 15.6 & \textless0.609 & 0.112 \\
& & & 1.74 & 15.6 & \textless0.271 & 0.060 \\
& & & 4.55 & 13.8 & 0.222 & 0.017 \\
& & & 7.38 & 13.8 & 0.253 & 0.014 \\
2015-04-07 & 54.563 & B & 13.5 & 12.5 & 0.457 & 0.028 \\
& & & 17.4 & 12.5 & 0.399 & 0.022 \\
& & & 28.2 & 12.4 & 0.323 & 0.047 \\
& & & 36.5 & 12.4 & 0.276 & 0.062 \\
2015-04-08 & 55.733 & B & 1.26 & 15.7 & 0.87 & 0.10 \\
& & & 1.74 & 15.7 & 0.661 & 0.060 \\
& & & 4.55 & 13.7 & 0.483 & 0.024 \\
& & & 7.38 & 13.7 & 0.374 & 0.018 \\
2015-04-18 & 65.513 & B & 1.26 & 16.6 & 0.444 & 0.082 \\
& & & 1.74 & 16.6 & 0.258 & 0.057 \\
& & & 4.55 & 14.6 & 0.256 & 0.019 \\
& & & 7.38 & 14.6 & 0.228 & 0.014 \\
2015-04-19 & 66.493 & B & 13.5 & 12.0 & 0.209 & 0.017 \\
& & & 17.4 & 12.0 & 0.167 & 0.023 \\
& & & 28.2 & 12.0 & 0.173 & 0.049 \\
& & & 36.5 & 12.0 & \textless0.289 & 0.083 \\
2015-05-01 & 78.463 & B & 1.26 & 15.6 & \textless0.32 & 0.11 \\
& & & 1.74 & 15.6 & \textless0.378 & 0.071 \\
& & & 4.55 & 13.6 & 0.170 & 0.019 \\
& & & 7.38 & 13.6 & 0.139 & 0.015 \\
2015-05-01 & 78.503 & B & 13.5 & 12.1 & 0.129 & 0.017 \\
& & & 17.4 & 12.1 & 0.164 & 0.022 \\
& & & 28.2 & 12.0 & 0.174 & 0.053 \\
& & & 36.5 & 12.0 & \textless0.241 & 0.075 \\
2015-05-16 & 93.463 & B$\rightarrow$BnA & 1.26 & 8.0 & \textless0.531 & 0.177 \\
& & & 1.74 & 8.0 & \textless0.336 & 0.078 \\
& & & 4.55 & 8.0 & 0.0727 & 0.023 \\
& & & 7.38 & 8.0 & 0.0708 & 0.017 \\
& & & 13.5 & 8.0 & 0.0518 & 0.018 \\
& & & 16.5 & 8.0 & \textless0.0859 & 0.021 \\
& & & 29.5 & 9.0 & \textless0.149 & 0.036 \\
& & & 35.0 & 9.0 & \textless0.135 & 0.045 \\
2015-06-14 & 122.363 & BnA$\rightarrow$A & 1.26 & 8.0 & \textless0.78 & 0.12 \\
& & & 1.74 & 8.0 & \textless0.233 & 0.078 \\
& & & 5.0 & 8.0 & \textless0.0806 & 0.020 \\
& & & 7.0 & 8.0 & \textless0.111 & 0.022 \\
& & & 13.5 & 8.0 & \textless0.106 & 0.016 \\
& & & 16.5 & 8.0 & \textless0.120 & 0.021 \\
& & & 29.5 & 9.0 & \textless0.264 & 0.045 \\
& & & 35.0 & 9.0 & \textless0.176 & 0.059
\enddata
\tablenotetext{1}{We take the time of initial detection 2015 February 11.837 UT (MJD 57064.837) to be Day 0.0}
\end{deluxetable*}
\end{center}
For all epochs, we used the absolute flux calibrator 3C286. Complex gains calibrators were J1626-2951 for both L-band and C-band, and J1650-2943 for both Ku-band and Ka-band. All VLA data were calibrated with the NRAO VLA calibration pipeline, version 1.3.1, which uses CASA\footnote{http://www.casa.nrao.edu} version 4.2.2. The pipeline script was executed on either a dedicated desktop or an NRAO Lustre node. Once calibrated, the data were exported to AIPS\footnote{http://www.aips.nrao.edu} for additional flagging. The fully-flagged data were imaged with Difmap (Sheperd 1997). The images were then imported back into AIPS and flux densities were measured using the task \verb|JMFIT|. The L-band and C-band data from 2015 April 18 were also calibrated in AIPS in order to look for polarization. Also, the Ku-band and Ka-band data from 2015 March 25 were re-calibrated in AIPS to check the results of the CASA pipeline. Our VLA results are presented in Table~\ref{radiotab}. For non-detections, we calculated the upper limit as the flux density at the nova location plus 3 times the image root mean square (rms). Our uncertainties are calculated by adding the image rms and an absolute flux density uncertainty in quadrature. Perley \& Butler (2013) report that the VLA absolute flux density calibration is stable to within $1\%$ for 1 to 20 GHz, and within $3\%$ for 20 to 50 GHz. Because V1535 Sco is far from 3C286, our uncertainty in the flux density calibration will be higher. We adopt an absolute flux density uncertainty of $3\%$ for 1 to 20 GHz, and $5\%$ for above 20 GHz.
\begin{center}
\begin{deluxetable*}{cccccc}
\tablewidth{0 pt}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.025in}
\tablecaption{ \label{vlbatab}
VLBA Observations}
\tablehead{UT Date & Day\tablenotemark{1} & Central Frequency & Time on Source & Flux Density & Uncertainty \\
& & (GHz) & (minutes) & (mJy) & (mJy) }
\startdata
2015-02-19 & 7.663 & 4.87 & 177.3 & 0.477 & 0.056 \\
2015-02-24 & 12.763 & 4.87 & 177.3 & \textless0.278 & 0.041 \\
\enddata
\tablenotetext{1}{We take the time of initial detection 2015 February 11.837 UT (MJD 57064.837) to be Day 0.0}
\end{deluxetable*}
\end{center}
\begin{center}
\begin{deluxetable*}{ccccccc}
\tablewidth{0 pt}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.025in}
\tablecaption{ \label{swiftab1}
\emph{Swift}/XRT Observations}
\tablehead{Obs ID & Day\tablenotemark{1} & Exposure & Total Counts & 0.3-1.0 keV & 1-10 keV & Total \\
& & (s) & & Count Rate & Count Rate & Count Rate }
\startdata
00033634002 & 4.16 & 4084 & 591 & 0.0062 & 0.1578 & 0.1640 \\
00033634003 & 11.04 & 3485 & 198 & 0.0555 & 0.0878 & 0.1428 \\
00033634004 & 13.94 & 3407 & 431 & 0.0886 & 0.0553 & 0.1429 \\
00033634005 & 17.93 & 4669 & 676 & 0.1342 & 0.0329 & 0.1666 \\
00033634006 & 24.75 & 2959 & 252 & 0.0925 & 0.0174 & 0.1091 \\
00033634007 & 31.51 & 4378 & 46 & 0.0102 & 0.0067 & 0.0161 \\
00033634008 & 38.77 & 4568 & 27 & 0.0021 & 0.0054 & 0.0075 \\
00033634009 & 45.55 & 2138 & \nodata & \textless0.0011 & \textless0.0011 & \textless0.0082 \\
00033634010 & 49.08 & 2575 & 5 & 0.0008 & 0.0021 & 0.0024 \\
00033634011 & 53.07 & 4714 & 12 & 0.0016 & 0.0025 & 0.0041 \\
00033634012 & 60.03 & 4845 & 13 & 0.0013 & 0.0015 & 0.0027 \\
00033634013 & 66.71 & 2979 & 7 & 0.0007 & 0.0026 & 0.0028 \\
00033634014 & 73.86 & 4934 & \nodata & \textless0.0005 & \textless0.0002 & \textless0.0036 \\
00033634015 & 80.32 & 4526 & \nodata & \textless0.0002 & \textless0.0001 & \textless0.0024 \\
\enddata
\tablenotetext{1}{We take the time of initial detection 2015 February 11.837 UT (MJD 57064.837) to be Day 0.0}
\end{deluxetable*}
\end{center}
\subsection{VLBA Observations}
We had 2 epochs of VLBA observations under the program VLBA/15A-269, both in C-band with a central frequency of 4.87 GHz and a total bandwidth of 256 MHz. For both epochs, we used J1709-3525 as the phase reference source. The bright sources J1656-3302 and J1713-3418 were also observed as a means of gauging the successful calibration of the nova. For the first epoch, 3 antennas were not usable: Hancock, North Liberty, and Mauna Kea. For the second epoch, Hancock and Mauna Kea could again not be used. The total time on source for each epoch was approximately 177.3 minutes (or 2.96 hours).
Both epochs were calibrated using standard routines in AIPS, including the new bandpass correction routine described by Walker (2014). Images were made in Difmap. The nova was detected in the first epoch, but not in the second epoch (see Table~\ref{vlbatab}). Schedules for further epochs were submitted, but conflict with a top-priority VLBA program prevented them from being observed. The position of the compact source detected in the first VLBA epoch was RA 17$^{\text{h}}$03$^{\text{m}}$26$^{\text{s}}$.17218 $\pm$0$^{\text{s}}$.00002, DEC -35\arcdeg04\arcmin17.87267\arcsec $\pm$0.00071\arcsec.
\subsection{Swift Observations}
We obtained 14 epochs of \emph{Swift} XRT observations. The exposure times ranged from 2.5 to 4.9 ks. The nova was detected by the XRT in 11 of the 14 observations. Details of the observations are given in Table~\ref{swiftab1}.
We divided the \emph{Swift} XRT detections into soft and hard bands based on the photon energy, $E$. We chose $E$\textless1.0 keV to be soft and $E$\textgreater1.0 keV to be hard. Our decision to use these designations is based on previous observations of novae which indicate which indicate that the $E$\textless1.0 keV band can be dominated by the super-soft X-rays, while the $E$\textgreater1.0 keV X-rays are likely to be from shocks (e.g., Mukai et al. 2008; Mukai et al. 2014). The X-ray emission was initially hard. The soft emission increased over the first $\sim$20 days while the hard emission decreased over the same time period. We had no detections with the XRT after Day 67. Our \emph{Swift} XRT results are plotted in Figure~\ref{swift_all}.
\begin{figure}
\includegraphics[width=3.3in]{sco15_SwiftALL3.pdf}
\caption{Top panel: Total count rate from \emph{Swift} XRT. Second panel: Count rate in the 0.3-1.0 keV energy range. Third panel: Count rate in the 1.0-10.0 keV energy range. Bottom panel: Hardness ratio (1.0-10 keV / 0.3-1.0 keV). The dotted line indicates a ratio of 1.0.}
\label{swift_all}
\end{figure}
\begin{figure}
\includegraphics[trim=10 10 10 10,clip,width=3.5in]{Sco15_20150213_SMARTS.pdf}
\caption{Example optical spectrum of V1535 Sco obtained with the SMARTS telescope. This spectrum was from Day 1.565. Note the high velocity outliers on H$\alpha$ (6563 \AA).}
\label{smarts_exspec}
\end{figure}
\begin{figure}
\includegraphics[trim=0 10 10 20,clip,width=3.2in]{v1535sco_ha.pdf}
\caption{The H$\alpha$ line over the first 16.5 days of nova eruption. Note the rapid narrowing of the line as the fastest material becomes optically thin.}
\label{smarts_ha}
\end{figure}
\subsection{SMARTS Photometric and Spectroscopic Observations}
SMARTS monitors novae in the optical and near infrared and provides the results to the public via the Stony Brook/SMARTS Spectroscopic Atlas of (mostly) Southern Novae\footnote{www.astro.sunysb.edu/fwalter/SMARTS/NovaAtlas/atlas.html}. Photometric observations of V1535 Sco began on 2015 February 13.35 using the ANDICAM\footnote{http://www.astronomy.ohio-state.edu/ANDICAM/detectors.html} dual-channel imager on the SMARTS 1.3m telescope.
Spectroscopic measurements of the ejecta began on 2015 February 13.5 as part of the SMARTS observation campaign using the CHIRON echelle spectrometer on the 1.5m SMARTS telescope (Tokovinin et al. 2013). As an example, we show the spectrum from Day 1.565 in Figure~\ref{smarts_exspec}. There is P Cyg absorption present on the H$\beta$ (4861 \AA), H$\gamma$ (4340 \AA), \ion{He}{1} (4471, 4921,\& 5879 \AA), \ion{Fe}{2} (5169 \AA), and \ion{O}{1} (7774 \& 8446 \AA) lines. This is unusual for a He/N nova, as they typically do not have much mass loss.
The H$\alpha$ line width measurements are presented in Table~\ref{smarts_hatab}, and we show the evolution of the H$\alpha$ line from Day 1.565 to Day 16.548 in Figure~\ref{smarts_ha}. Because we are concerned with the outermost edge of the main component of the ejecta, we estimate the maximum velocity as the full width at 3$\sigma$ (FW3$\sigma$). We convert the H$\alpha$ FWHM velocities to FW3$\sigma$ velocities by assuming the core of the spectral line has a Gaussian shape, then we multiply by a factor of $1.5(2 \ln 2)^{-1/2}$ in order to get the $3\sigma$ velocities. Our velocity evolution is shown in Figure~\ref{sco15_vnr}. While the measured velocity of the ejecta appears to decrease with time, we must emphasize that what is actually being measured here is the velocity of the H$\alpha$ emitting region, which does not necessarily correspond to the outer edge of the ejecta.
\begin{figure}
\includegraphics[trim=0 10 10 25,clip,width=3.2in]{v1535sco_hab.pdf}
\caption{The evolution of the blue-shifted high velocity H$\alpha$ outlier over the first 16.5 days of the nova eruption. The vertical line is at -4850 km s$^{-1}$. Note that by Day 6.5 the \ion{Fe}{2} lines at 6455.8 and 6456.4 \AA\, dominate this region of the spectrum.}
\label{smarts_hab}
\end{figure}
The H$\alpha$ (6563 \AA) line shows high velocity outliers (both red and blue shifted; see Table~\ref{smarts_hawings}). These outliers do not appear to decelerate as long as they are present. In fact, the high velocity outliers appear to accelerate throughout the first week of the nova's evolution, although this could be the result of blending with lines from \ion{He}{1} (6678 \AA) and \ion{Fe}{2} (6455.8 \& 6456.4 \AA) which are not detectable until later in the nova's evolution. We show the evolution of the high velocity outliers in Figures~\ref{smarts_hab} (blue-shifted) and \ref{smarts_har} (red-shifted). These outliers could be explained by a shell of high velocity material. However, the narrowness of the outlier lines would require that the shell had a very small thickness, and such a fast-moving thin shell should become diffuse (and thus undetectable) much faster than we see here. Another explanation is a bipolar outflow, where the narrowness of these high velocity outliers would point to a relatively well-collimated outflow. It would also likely have a fairly small inclination angle for us to detect such large velocities. Such well-collimated bipolar outflows have been seen in several other novae. Examples of embedded novae with bipolar outflows are U Sco (L\'{e}pine et al. 1999) and RS Oph (Rupen et al. 2008; Sokoloski et al. 2008). Non-embedded novae are also known to have bipolar outflows, such as V1494 Aql (Iijima \& Esenoglu 2003), V475 Sct (Kawabata et al. 2006), V445 Pup (Woudt et al. 2009), V5668 Sgr (Bannerjee et al. 2016), and possibly T Pyx (Chesneau et al. 2011).
\begin{figure}
\includegraphics[trim=0 10 10 25,clip,width=3.2in]{v1535sco_har.pdf}
\caption{The evolution of the red-shifted high velocity H$\alpha$ outlier over the first 16.5 days of the nova eruption. The vertical line is at 4550 km s$^{-1}$. Note that the peak of the outlier clearly moves to higher velocities through Day 6.5. Also, note that the \ion{He}{1} line at 6678 \AA\, appears on Day 4.5 and dominates this region of the spectrum by Day 7.5.}
\label{smarts_har}
\end{figure}
The high velocity outliers in Figures~\ref{smarts_exspec}, \ref{smarts_hab}, and \ref{smarts_har} are likely to be H$\alpha$. We see marginal evidence for similar outliers with identical velocities around the H$\beta$ line. Were these the [\ion{N}{2}] lines, the 6482 \AA\, line would be blueshifted by approximately 1150 km s$^{-1}$, while the 6668 \AA\, line would be redshifted by only about 300 km s$^{-1}$. Low velocity \ion{Fe}{2} and \ion{He}{1} lines appear in these regions around day 5 and strengthen as the ejecta evolve. Lacking other plausible identifications, we identify the outliers as H$\alpha$ ejected at high velocities.
Using the H$\alpha$ line width as the approximate velocity evolution for the optically-thick component of the ejecta, we can use $r(t_{1})=\int_{0}^{t_{1}}v(t)dt$ to determine the approximate radial size of the ejecta at any given time $t_{1}$. We perform this integration numerically using the trapezoidal approximation. First, we interpolate our velocity measurements to give a smoother function to integrate over. We assume that the velocity is constant between Day 0 and our first spectroscopic measurement. This is likely an oversimplification, although it may not be drastically far from the average velocity over this time, as the ejecta must first be accelerated to some (unknown) maximum velocity, then decelerate to the velocity we measured with our first spectroscopic observation. Our resulting radial size of the H$\alpha$ emitting region is shown in the bottom panel of Figure~\ref{sco15_vnr}.
The radial size of the H$\alpha$ emitting region can serve as a minimum size of the ejecta. A more realistic approximation of the radial size of the ejecta is to assume that the ejecta expand at roughly the velocity from the first observation: $\sim$1659 km s$^{-1}$. However, neither of these simple models account for the high velocity outliers seen in the early spectra.
\begin{deluxetable}{ccccc}
\tablewidth{0 pt}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.025in}
\tablecaption{ \label{smarts_hatab}
SMARTS H$\alpha$ Spectral Measurements}
\tablehead{Day\tablenotemark{1} & FWHM & FWHM Velocity & FW3$\sigma$ & FW3$\sigma$ Velocity \\
& (\AA) & (km s$^{-1}$) & (\AA) & (km s$^{-1}$) }
\startdata
1.565 & 47.5 & 1302 & 60.5 & 1659 \\
2.491 & 37.3 & 1023 & 47.5 & 1304 \\
4.493 & 36.0 & 987 & 45.8 & 1257 \\
5.495 & 36.7 & 1006 & 46.7 & 1282 \\
6.495 & 34.1 & 936 & 43.5 & 1193 \\
7.484 & 32.8 & 899 & 41.8 & 1145 \\
10.478 & 26.0 & 713 & 33.1 & 908 \\
11.505 & 20.9 & 574 & 26.7 & 731 \\
13.528 & 14.5 & 397 & 18.4 & 505 \\
14.505 & 13.1 & 359 & 16.7 & 458 \\
15.483 & 11.3 & 311 & 14.4 & 397 \\
16.548 & 10.5 & 288 & 13.4 & 366 \\
17.482 & 9.2 & 252 & 11.7 & 322 \\
18.456 & 8.6 & 237 & 11.0 & 302 \\
19.525 & 8.2 & 225 & 10.5 & 287 \\
20.53 & 7.7 & 212 & 9.9 & 271 \\
21.498 & 7.3 & 200 & 9.3 & 255 \\
22.472 & 6.8 & 186 & 8.6 & 237 \\
23.451 & 6.2 & 170 & 7.9 & 217 \\
25.477 & 5.6 & 154 & 7.2 & 196 \\
26.445 & 5.2 & 141 & 6.6 & 180 \\
27.507 & 4.7 & 129 & 6.0 & 164 \\
28.462 & 4.4 & 121 & 5.6 & 154 \\
29.435 & 4.1 & 112 & 5.2 & 143 \\
31.469 & 3.7 & 101 & 4.7 & 129 \\
32.432 & 3.5 & 96 & 4.5 & 122 \\
33.416 & 3.3 & 91 & 4.2 & 116 \\
34.432 & 3.1 & 85 & 3.9 & 108 \\
35.44 & 3.1 & 86 & 4.0 & 110 \\
36.42 & 3.1 & 84 & 3.9 & 107 \\
37.432 & 3.0 & 83 & 3.9 & 106 \\
38.413 & 3.0 & 82 & 3.8 & 105 \\
39.398 & 2.9 & 79 & 3.7 & 101 \\
50.435 & 2.7 & 75 & 3.5 & 96 \\
51.468 & 2.7 & 74 & 3.5 & 95 \\
52.412 & 2.7 & 74 & 3.5 & 95 \\
53.364 & 2.7 & 74 & 3.4 & 94 \\
54.463 & 2.7 & 73 & 3.4 & 93 \\
54.505 & 2.6 & 72 & 3.4 & 92 \\
55.389 & 2.6 & 72 & 3.4 & 92 \\
56.447 & 2.7 & 73 & 3.4 & 93 \\
57.438 & 2.7 & 73 & 3.4 & 93 \\
58.371 & 2.7 & 74 & 3.4 & 94 \\
59.398 & 2.6 & 72 & 3.3 & 91 \\
60.343 & 2.5 & 69 & 3.2 & 88 \\
61.382 & 2.5 & 69 & 3.2 & 88 \\
62.381 & 2.5 & 69 & 3.2 & 88 \\
64.495 & 2.5 & 69 & 3.2 & 88 \\
65.371 & 2.5 & 69 & 3.2 & 87 \\
66.37 & 2.5 & 67 & 3.1 & 86 \\
67.413 & 2.4 & 67 & 3.1 & 85 \\
68.407 & 2.4 & 66 & 3.1 & 84 \\
69.473 & 2.4 & 64 & 3.0 & 82 \\
70.434 & 2.3 & 64 & 3.0 & 82 \\
71.4 & 2.3 & 64 & 3.0 & 81 \\
72.422 & 2.3 & 64 & 3.0 & 81 \\
74.341 & 2.3 & 63 & 2.9 & 80 \\
76.393 & 2.3 & 63 & 2.9 & 80 \\
78.453 & 2.4 & 66 & 3.0 & 84 \\
80.477 & 2.2 & 61 & 2.8 & 78 \\
82.36 & 2.3 & 62 & 2.9 & 79 \\
86.377 & 2.3 & 63 & 2.9 & 80 \\
88.287 & 2.3 & 64 & 3.0 & 81 \\
91.302 & 2.3 & 63 & 2.9 & 80 \\
93.322 & 2.3 & 62 & 2.9 & 79 \\
98.299 & 2.5 & 68 & 3.2 & 87 \\
100.274 & 2.4 & 66 & 3.1 & 84 \\
103.232 & 2.4 & 66 & 3.0 & 84 \\
105.25 & 2.3 & 64 & 3.0 & 82 \\
109.298 & 2.4 & 66 & 3.0 & 84 \\
111.376 & 2.4 & 66 & 3.1 & 84 \\
116.206 & 2.4 & 66 & 3.1 & 84 \\
119.308 & 2.4 & 66 & 3.0 & 84 \\
120.457 & 2.4 & 65 & 3.0 & 83 \\
\enddata
\tablenotetext{1}{We take the time of initial detection 2015 February \\11.837 UT (MJD 57064.837) to be Day 0.0}
\end{deluxetable}
\begin{figure}
\includegraphics[width=3.5in]{sco15_SMARTS_VelRad.pdf}
\caption{{\bf Upper panel:} Velocity profile from the H$\alpha$ line width of V1535 Sco. The dashed blue line shows the interpolation of the data used to estimate the radius of the H$\alpha$ emitting region. {\bf Bottom panel:} Integrated radial size of the H$\alpha$ emitting region.}
\label{sco15_vnr}
\end{figure}
\begin{deluxetable}{ccccc}
\tablewidth{0 pt}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.025in}
\tablecaption{ \label{smarts_hawings}
SMARTS H$\alpha$ Shoulders}
\tablehead{Day\tablenotemark{1} & $\lambda_{blue}$ & Blue-shifted Velocity & $\lambda_{red}$ & Red-shifted Velocity \\
& (\AA) & (km s$^{-1}$) & (\AA) & (km s$^{-1}$) }
\startdata
1.565 & 6458.25 & 4778 & 6658.67 & 4377 \\
2.491 & 6458.09 & 4785 & 6660.00 & 4414 \\
4.493 & 6456.03 & 4880 & 6665.90 & 4707 \\
5.495 & 6455.45 & 4906 & 6668.08 & 4807 \\
6.495 & 6455.5 & 4904 & 6669.98 & 4893 \\
7.484 & 6456.3 & 4867 & 6670.75 & 4929 \\
10.478 & 6456.52 & 4857 & 6671.3 & 4954 \\
11.505 & 6456.91 & 4839 & \nodata & \nodata \\
\enddata
\tablenotetext{1}{We take the time of initial detection 2015 February \\11.837 UT (MJD 57064.837) to be Day 0.0}
\end{deluxetable}
\section{The Distance to V1535 Sco}
The distance to V1535 Sco is highly uncertain. Srivastava et al. (2015) apply the maximum magnitude rate of decline (MMRD) relations from della Valle \& Livio (1995) and Downes \& Duerbeck (2000) to get distance estimates of $13.7\pm0.4$ and $14.7\pm3.8$ kpc, respectively.
However, Munari et al. (2017) reported a distance to V1535 Sco of approximately 9.7 kpc, despite using the same MMRD method as Srivastava et al. (2015). This indicates the MMRD distance may be highly dependent on the data used, as Munari et al. (2017) performed all their observations on a single telescope and Srivastava et al. (2015) obtained their data from multiple telescopes via the American Association of Variable Star Observers (AAVSO). Furthermore, substantial uncertainty has been shed on the MMRD method in general by the work of Kasliwal et al. (2011) and Cao et al. (2012). The Srivastava et al. MMRD distances seem unlikely as they place the nova on the opposite side of the Galactic Bulge, which should make it extremely reddened and obscured due to the intervening material.
We were able to constrain the distance to V1535 Sco by using pre-outburst photometry to determine a spectral type for the companion. The source 2MASS1703261-350417 is within $\leq 0.1$" of V1535 Sco, and has $J=13.423 \pm 0.037$, $H = 12.500 \pm 0.033$, and $K=12.190 \pm 0.041$. Without correcting for reddening, we find near-IR colors of $J-K = 1.233 \pm 0.078$ and $H-K = 0.310 \pm 0.074$.
Using the derived reddening from Srivastava et al. (2015) of $E(B-V) = 0.72 \pm 0.05$ we can compute the necessary color corrections. The reddening conversion functions of Schlafly \& Finkbeiner (2011) give $E(J-K) = 0.413\times E(B-V)$, $E(H-K) = 0.15\times E(B-V)$ and $A_J = 0.723\times E(B-V)$. Using these, we find the dereddened colors to be $(J-K)' = 0.936 \pm 0.098$, $(H-K)' = 0.202 \pm 0.081$, and $J' = 12.902 \pm 0.073$, where the prime signifies that it is dereddened. Using the spectral class system of Covey et al.(2007), the colors are consistent with a spectral class from K3 III to M0 III which, in turn, give an absolute $J$ magnitude between $-3.92$ to $-4.75$. This gives a distance modulus of $m-M = 15.212 \pm 1.253$, or a distance between 6.2 and 19.6 kpc. If we instead use $E(B-V) = 0.96$, as reported in Munari et al. (2017), our distance range is 5.8 to 14.0 kpc.
Note that, although we cannot definitively rule out the companion being a main sequence star using this method, if it were a main sequence progenitor system the distance modulus would be $<8$ mags. Since the peak brightness of the nova was $m_V \approx 9.5$, the absolute magnitude at peak would be $M_V > 1.0$, which is far too dim for a nova. Therefore, we can reasonably rule out a main sequence companion. We therefore find it more likely that the companion star is an evolved star in the range between subgiant and giant.
Because the progenitor distance is consistent with being located at the Galactic Center, and that is also where we expect most novae to occur, we will assume a distance of $\sim8.5$ kpc for V1535 Sco. This distance agrees well with estimates in Munari et al. (2017) based on both MMRD and the Buscombe \& de Vaucouleurs (1955) method of estimating the absolute magnitude 15 days after optical maximum.
\section{Results}
\subsection{VLA Light Curve and Spectral Indices}
The full VLA light curves for V1535 Sco are presented in Figure~\ref{vlalc}. The nova was initially detected at relatively high flux densities, faded substantially, and then rose again to secondary peak around Day 25. This secondary peak is unusual for novae, but not unprecedented (e.g., Taylor et al. 1987; Krauss et al. 2011; Eyres et al. 2009; Weston et al. 2016; Finzell et al. 2017). After Day 27, the source faded again until about Day 56 when there is a tertiary peak. This is highly unusual. After the tertiary peak, the nova fades until it is no longer detectable.
The radio light curve for V1535 Sco is significantly different from the radio light curve of a non-embedded nova. In a ``typical'' non-embedded nova, the radio emission is delayed from the optical emission by up to $\sim$2 weeks, and then slowly rises to a maximum over several months. This is due to the fact that non-embedded novae are typically \emph{thermal} sources at radio wavelengths. Therefore, their flux density is directly related to the size of the ejecta. As the ejecta expand, the size of the emitting surface area increases and so does the flux density (e.g., Hjellming 1996; Seaquist \& Bode 2008). In contrast, the radio light curves for embedded novae are dominated by synchrotron radiation in the first $\sim$ weeks, and therefore have high flux densities early and fade as the ejecta cool (e.g., O'Brien et al. 2006).
\begin{figure}
\includegraphics[width=3.5in]{sco15_lightcurveSL3.pdf}
\caption{Full light curves for the VLA observations. Initial detection was 2015 February 11.837 UT (MJD 57064.837), which is used as $t=0$. Non-detections are indicated with a downward-pointing triangle.}
\label{vlalc}
\end{figure}
\begin{figure*}
\includegraphics[width=7.0in]{sco15_SI5d2.pdf}
\caption{Spectral energy distribution of VLA observations plotted for each epoch. Non-detections are indicated by red downward-pointing triangles. The fit to the same-day data is shown as a dashed cyan line, and the spectral index value is given in cyan.}
\label{vlasi}
\end{figure*}
We measured the spectral index $\alpha$ (using $S_{\nu}\propto\nu^{\alpha}$) for each of our VLA observations. We fit the flux densities to a power law using the non-linear least squares \verb|curve_fit| function in the \emph{SciPy} package of python, weighted by the 1$\sigma$ uncertainties. We ignored upper limit values for this fit, using only solid detections. If two observations were separated by less than 24 hours, they were combined to make a single spectral index measurement. Our calculated spectral indices are presented in Table~\ref{vlasitab}, including the associated 1$\sigma$ uncertainties from the variance-covariance matrix.
The evolution of the spectral index is shown graphically in Figure~\ref{vlasi}. During the first week of its evolution, the nova's radio spectral index rises toward lower frequencies, an indicator of optically thin synchrotron emission. The spectral index then switches to rising toward higher frequencies. This is typical of sources emitting via optically thick thermal bremsstrahlung. To our knowledge, such a dramatic switch between a synchrotron-like $\alpha$ to a thermal-like $\alpha$ has never been seen before. However, the spectral index of V1324 Sco showed evidence of being negative at early times followed by a positive slope later (Finzell et al. 2017), and Eyres et al. (2009) reported the presence of both positive and negative spectral indices present simultaneously for RS Oph. At the tertiary peak around Day 56, the spectral index again rises toward lower frequencies and appears to be optically thin synchrotron. This apparent switch from a thermal-like to synchrotron-like spectral index is unprecedented. As the nova fades, its spectral index flattens as expected for an optically thin thermal bremsstrahlung source.
The evolution of V1535 Sco's radio spectral index is very different from the majority of classical novae. Typically, a the radio emission begins with a positive (rising toward higher frequencies) spectral index while the ejecta is optically thick. As the ejecta becomes optically thin, the radio spectral index flattens to value around -0.1. This transition from strongly positive to flat begins with the highest frequencies first and progresses through the lower frequencies as the ejecta dims overall. Even in novae where shocks are detected such as V959 Mon, the spectral index at early times is still usually positive (Chomiuk et al. 2014). A negative spectral index, especially during the first few weeks of a nova's evolution, usually occurs in systems with evolved companions producing a strong wind for the nova ejecta to shock against, as in RS Oph (e.g., Eyres et al. 2009).
\begin{deluxetable}{cccc}
\tablewidth{0 pt}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.025in}
\tablecaption{ \label{vlasitab}
VLA Spectral Indices}
\tablehead{UT Date & Day\tablenotemark{1} & Freq. Range & Spectral Index \\
& & (GHz) & }
\startdata
2015-02-14 & 2.663 & 4.5\textendash 36.5 & -0.88$\pm$0.02 \\
2015-02-18 & 6.663 & 13.5\textendash 36.5 & -0.36$\pm$0.22 \\
2015-02-19 & 7.663 & 1.26\textendash 7.38 & -0.71$\pm$0.02 \\
2015-02-24 & 12.762 & 13.5\textendash 36.5 & 0.87$\pm$0.10 \\
2015-03-01 & 17.663 & 4.55\textendash 7.38 & -0.29$\pm$0.34\tablenotemark{\dag} \\
2015-03-07 & 23.663 & 13.5\textendash 36.5 & 1.01$\pm$0.07 \\
2015-03-10 & 26.563 & 4.55\textendash 7.38 & 0.34$\pm$0.14\tablenotemark{\dag} \\
2015-03-25 & 41.583 \& 41.633 & 4.55\textendash 28.2 & 0.21$\pm$0.04 \\
2015-04-07 & 54.563 & 13.5\textendash 36.5 & -0.49$\pm$0.02\\
2015-04-08 & 55.733 & 1.26\textendash 7.38 & -0.43$\pm$0.05 \\
2015-04-18 \& 19 & 65.513 \& 66.493 & 1.26\textendash 28.2 & -0.24$\pm$0.13 \\
2015-05-01 & 78.463 \& 78.503 & 4.55\textendash 28.2 & -0.06$\pm$0.13 \\
2015-05-16 & 93.463 & 4.55\textendash 13.5 & -0.32$\pm$0.15 \\
\enddata
\tablenotetext{1}{We take the time of initial detection 2015 February 11.837 UT \\(MJD 57064.837) to be Day 0.0}
\tablenotetext{\dag}{Only 2 data points available for fit}
\end{deluxetable}
\subsection{VLBA Compact Emission}
Our VLBA detection on 2015 February 19 (Figure~\ref{vlba1}) was simultaneous with VLA L-band and C-band observations. Using Difmap, we modelled the VLBA detection as a circular Gaussian in the \emph{uv}-plane. The resulting total flux density is 0.477 mJy, with an off-source image rms of 0.044 mJy/beam, and a diameter of 6.52 mas. The resolving beam for this observation was 2.5 x 12.4 mas, so this could indicate a resolved source.
\begin{figure}
\includegraphics[width=3.3in]{SCO15_VLBA0219C_SHIFT_cntr.pdf}
\caption{4.87 GHz VLBA image of V1535 Sco on 2015 February 19 (Day 7.6). The gray ellipse in the lower left corner indicates the shape and size of the restoring beam. The apparent extension to the southeast is likely an imaging artifact resulting from imperfectly calibrated data. The image is 0.1 arcsec on a side.}
\label{vlba1}
\end{figure}
The brightness temperature (T$_{B}$) in K is given by the Rayleigh-Jeans relation (e.g., Rohlfs \& Wilson 2006)
\begin{equation}
S_{\nu} = \frac{2\, k_{B}\, T_{B}\, \Omega}{\lambda^{2}}
\end{equation}
where $S_{\nu}$ is the flux density in erg s$^{-1}$ cm$^{-2}$ Hz$^{-1}$ at frequency $\nu$, $\lambda$ is the observing wavelength in cm, $k_{B}$ is the Boltzmann constant in erg K$^{-1}$, and $\Omega$ is the solid angle of the source. For a resolved source with a circular Gaussian shape, $\Omega = \pi\theta_{D}^{2}(4\ln2)^{-1}$, where $\theta_{D}$ is the angular diameter of the source in radians. Therefore, the brightness temperature is calculated using
\begin{equation}
T_{B} = \frac{2\, \ln2\, S_{\nu}\, \lambda^{2}}{\pi\, k_{B}\, \theta_{D}^{2}}
\end{equation}
for our source.
This gives a brightness temperature of 5.8$\times$10$^{5}$ K. This is significantly lower than the 10$^{7}$ K $T_{B}$ of the synchrotron components in RS Oph (Rupen et al. 2008). However, we emphasize that our observation was made with lower resolution due to the drop-out of the Mauna Kea antenna. It is possible that our single structure is actually composed of 2 or more components that are too compact to resolve with our configuration. Modelling the source as two point sources results in a similar total flux density (0.445 mJy) and image rms (0.04 mJy/beam) with the same reduced $\chi^{2}$ (1.09) as the circular Gaussian model. If we make the assumption that the angular size of the components is equal to the minor axis of the restoring beam (2.5 mas), the resulting $T_{B}$ estimates are 1.8$\times$10$^{6}$ K and 1.9$\times$10$^{6}$ K. These are still lower than the $T_{B}$ for RS Oph, but are consistent with the $T_{B}$ for the compact components in V959 Mon (Chomiuk et al. 2014).
The simultaneous VLA flux density at 4.55 GHz was 0.650 mJy. Because the VLBA is only sensitive to high brightness temperature (i.e., compact) emission, and the spectral index from the VLA observations of the nova ejecta was negative (i.e., larger flux density at lower frequencies), we conclude that the VLBA emission is most likely from synchrotron radiation. The compact emission on this day dominated the total emission at the 73\% level. The excess VLA radio flux on this day may be due to thermal bremsstrahlung emission from the shocked plasma or the pre-shocked, cooler outflow.
Using the VLBA flux density of the single circular Gaussian model component, we estimate the magnetic field strength using the revised equipartition formula from Beck \& Krause (2005). This formula requires knowledge of the path length through which the radiation propagates $l$, the filling factor of the emitting material $f$, the inclination angle of the field $i$, and the proton-to-electron number density ratio, $K_{0}$:
\begin{equation}
B\propto \left(\frac{K_{0}}{l\, f\, \cos(i)^{(1-\alpha)}}\right)^{1/(3-\alpha)}
\end{equation}
where $\alpha$ is the radio spectral index ($S_{\nu}\propto\nu^{\alpha}$).
To make our estimate, we will assume a filling factor of 1, and an inclination angle of 0 (i.e., the field is face-on). For $l$, we assume the emitting region was relatively thin, $\sim10\%$ of the radius of the H$\alpha$ emitting region (see Figure~\ref{sco15_vnr}) at the time of the observation, which gives us $l\sim9\times10^{12}$ cm. We used the typical values for $K_{0}$ of 40 and 100. The resulting magnetic field strengths were 0.14 G for $K_{0}=40$ and 0.18 G for $K_{0}=100$. If we assume that the path length is $\sim10\%$ of the radius of an ejecta shell expanding at a constant velocity of 1659 km s$^{-1}$ (the maximum velocity measured from the H$\alpha$ linewidth), we get $l\sim10^{13}$ cm and a magnetic field of 0.13 G for $K_{0}=40$ and 0.17 G for $K_{0}=100$. If we assume that the path length is $\sim10\%$ of the radius of an eject expanding at a constant velocity of 4782 km s$^{-1}$ (the median velocity from the high velocity outliers in the H$\alpha$ spectra over the first 5.5 days), we get $l\sim3\times10^{12}$ cm and a magnetic field of 0.10 G for $K_{0}=40$ and 0.13 G for $K_{0}=100$. These values are all larger than, but still comparable to, the magnetic field strengths estimated for RS Oph of $0.03(1+K_{0})^{2/7}$ (Rupen et al. 2008). Assuming the same values of $K_{0}$ as we used, the RS Oph magnetic field strength was between 0.087 G and 0.11 G. It should also be noted that Rupen et al. (2008) estimated the magnetic field for RS Oph over 20 days after the eruption, whereas we are estimating the field for V1535 Sco only 7.7 days after eruption. It is possible that the magnetic field declines over time as the ejecta expand and the shocks dissipate.
\subsection{Optical and Near-Infrared Photometry}
\begin{figure}
\includegraphics[trim=30 20 0 20,clip,width=3.5in]{SMARTS_NovaeCompare.pdf}
\caption{Comparison of photometric measurements for two novae known to have red giant companions with V1535 Sco. The left hand column shows the optical and near infrared light curves. The right hand column shows the difference in magnitude from the first observation. The first row is RS Oph (2006 eruption), the second row is V745 Sco (2014 eruption), and the bottom row is V1535 Sco.}
\label{smarts_comp}
\end{figure}
The optical and near-infrared light curve for V1535 Sco is not consistent with light curves for other novae known to have red giant companions (see Figure~\ref{smarts_comp}). In particular, when one considers the difference between the peak magnitude and the current magnitude (right hand side of Figure~\ref{smarts_comp}), V1535 Sco is clearly an outlier. In other novae with red giant companions, the companion star begins to dominate the difference light curve at longer wavelengths relatively early, while the shorter wavelengths are still dominated by the nova ejecta. V1535 Sco, on the other hand, does not follow this behavior. Instead, all wavelengths fade together for the duration of the light curve. This indicates that the companion star is unlikely to be a red giant. However, the presence of synchrotron emission in the radio (Sections 4.1 and 4.2), the hard X-ray emission, and the rapid fading of the optical light curve all indicate the nova ejecta were shocking against an external medium.
\begin{figure}
\includegraphics[width=3.25in]{v1535sco_sed170402.pdf}
\caption{The fit to the SED on 2017 April 2 with an active accretion disk and a K5~III
star. Model details are described in the text. Filled circles are the observed
data. The open white circles are the dereddened observations; open red circles
are the data after subtracting the K5 III SED. The green and blue dashed
curves are the fits to the accretion disk and the star, respectively. The red
dashed curve is the dereddened fit to the accretion disk. The fit underpredicts
the observed $J$ and $H$ fluxes.}
\label{smarts_sedfit}
\end{figure}
One possible explanation is that the companion star is not an M giant, but a K giant. We tested this hypothesis by fitting our optical and near infrared fluxes with a model that combined a hot accretion disk and a stellar spectral energy distribution for a K giant star.
The quiescent unreddened optical colors of V1535 Sco are close to 0. We modelled the optical-infrared (OIR) spectral energy distribution (SED) as the sum of a mid-K giant plus an active accretion disk. The accretion disk
is constructed following the Bertout, Basri, \& Bouvier (1988) formalism.
The optically thick disk is constructed of a series of annuli, each of which
emits as a black body at a temperature set by its distance from the star.
We set the inner edge of the accretion disk to the radius of a white dwarf;
exact details are not important since the boundary layer and the inner edge
of the disk emit in the ultraviolet. On the red side of the peak, the accretion disk
follows a power law F$_{\lambda} \sim \lambda^{-2.3}$.
With 7 OIR fluxes, the model is under-constrained. We fix the extinction A$_V$ to
be 2.16~mag and set the distance to 8.5~kpc. We constrain the white dwarf to be
hot. We constrain the donor star to be K5~III, with VJHK colors taken from
Koornneef (1983). Free parameters in the fit are the magnitude of the donor
star and the mass accretion rate. An example of a fit is shown in Figure~\ref{smarts_sedfit}.
We fit the ten SEDs obtained 2016 April 25 through 2017 April 02. Over that
year the donor star $K$ faded by 0.15$\pm$0.07 mag. This fading could be a
cooling after irradiation by the nova, or it could be part of a longer term
variation. The mean fit $K$ of 12.14 is consistent with the 2MASS $K$ magnitude.
Meanwhile the mass accretion rate has dropped linearly by a factor of 6, from
12 to 2$\times$10$^{-8}$M$_\odot$~yr$^{-1}$. Numbers should be taken with a
grain of salt. The mass accretion rate and the assumed distance are strongly
correlated where the inferred mass accretion rate scales as the cube of the inferred distance.
Using our fits and assuming a distance of 8.5 kpc, we find that the absolute $J$ magnitude for V1535 Sco in quiescence should be -1.2. Covey et al. (2007) gives the absolute $J$ magnitude for a K3~III star as -1.13, which is remarkably close for such a simple model. Munari et al. (2017) also conclude that the companion in V1535 Sco is consistent with a K3-4 III giant.
\subsection{Optical and Near Infrared Spectroscopy}
\begin{figure}
\includegraphics[width=3.5in]{nsco15_ha_Walter_crop.pdf}
\caption{Negative image of the H$\alpha$ region of the spectrum as a function of time. Darker regions indicate more emission, brighter regions indicate absorption.}
\label{walterha}
\end{figure}
Srivastava et al. (2015) presented measurements of the ejecta velocity for V1535 Sco based on the Pa$\beta$ emission line at 1.2818 $\mu$m starting 7 days after the detection and ending 40 days after the detection. Their velocities are given as full width at half maximum (FWHM) of the line. As with our H$\alpha$ measurements, we approximate the FW3$\sigma$ velocities by applying the same calculation as described in Section 2.4. We then fit the line-width-vs-time data with a power law using the non-linear least squares \verb|curve_fit| technique in the \emph{SciPy} package of python.
Our resulting fit is somewhat different from that given in Srivastava et al. (2015). They report a power law of $t^{-1.13\pm0.17}$ whereas our fit gives $t^{-0.89\pm0.14}$. However, we should note that we are uncertain as to the methods used by Srivastava et al. (2015) because they do not describe their fitting procedure in detail. To compare the Pa$\beta$ results with our more densely sampled H$\alpha$ data, we fit our H$\alpha$ FW3$\sigma$ measurements from Day 7.48 to Day 39.4 using the same \emph{SciPy} method and find a power law of $t^{-1.49\pm0.05}$, which nearly agrees within uncertainties with the Srivastava et al. (2015) value. Of course, this simple power-law fit completely ignores the complicated behavior observed during the first week following the explosion (see Figure~\ref{sco15_vnr}).
The SMARTS optical spectroscopic monitoring of V1535 Sco also revealed blue-shifted H$\alpha$ absorption features. These features become noticeable around Day 15, after the H$\alpha$ line has narrowed significantly (see Figure~\ref{walterha}). The blue-shifted features have an initial velocity of $\sim500$ km s$^{-1}$, but slow to $\sim50$ km s$^{-1}$ by Day 247. These absorption features indicate the presence of a large amount of cool, neutral material ahead of the H$\alpha$ emitting region. It is possible that this neutral material was pre-shocked wind material being swept up by the nova ejecta in a ``snowplow'' fashion.
\subsection{X-rays}
We fit the \emph{Swift} data using the \verb|XSpec| version 12.8.2 package. The spectra were best fit using a combination of a black body (BB) and a thermal plasma (apec). The apec model includes contributions from free-free continuum and lines (Smith et al. 2001). Only the first 5 detections had high enough signal-to-noise to successfully model the emission. The results are given in Table~\ref{swiftab2}. The Hydrogen column density (N$_{\rm H}$) appears to be fairly constant throughout the first 18 days, with a deviation on Day 11 which is most likely the fitting software getting stuck in a local minimum. This constant N$_{\rm H}$ is quite different from the sharply decreasing N$_{\rm H}$ observed in the embedded novae RS Oph (Bode et al. 2006) and V745 Sco (Page et al. 2015). The black body temperature appears to decrease by only a factor of 2 over the first 25 days, with a minor fluctuation around Day 17.9. The hot plasma temperature, on the other hand, dramatically cools by at least 2 orders of magnitude during the same time period. This is likely due to the shocks expanding and becoming radiatively efficient.
The $E(B-V)$ of 0.96 reported by Munari et al. (2017) implies N$_{\rm H}\approx5\times10^{21}$ cm$^{-2}$ (Predehl \& Schmitt 1995). Our values of N$_{\rm H}$ are nearly twice this, indicating a local enhancement from a possible stellar wind. Interestingly, our N$_{\rm H}$ values are quite similar to those measured for 2 recurrent novae known to have red giant companions: V745 Sco and RS Oph. In RS Oph, the values for N$_{\rm H}$ were $\sim10^{23}$ cm$^{-2}$ just 0.16 days after eruption, but fell to $\sim10^{22}$ cm$^{-2}$ by Day 3.23 (Page et al. 2015) and were $\sim 8\times10^{21}$ cm$^{-2}$ by Day 10 (Orio et al. 2015). In RS Oph, Nelson et al. (2008) reported N$_{\rm H}\sim10^{22}$ cm$^{-2}$ 14 days after eruption. We stress that this does not directly indicate the presence or absence of a red giant, especially as the kT values for V1535 Sco are similar to those that Mukai \& Ishida (2001) found for V382 Vel, which is known to have a main sequence companion. However, it is important to note that V1535 Sco was detected by \emph{Swift} only 5 days after its discovery in the optical, while V382 Vel was not detected in X-rays until 20 days after optical discovery. The delay in X-rays in V382 Vel is often seen in non-embedded novae, and is thought to be due to the fact that it takes time for multiple outflows to be ejected and then collide. It may therefore be more appropriate to compare the X-ray emission from V382 Vel to the later (\textgreater Day 49) X-rays from V1535 Sco.
Overall, the X-ray emissions are somewhat puzzling. On the one hand, the flat N$_{\rm H}$ for several days is different from the behavior in known embedded novae. On the other hand, there appears to be local absorption in the system on levels similar to known embedded novae. There are two other major points to consider. First, the X-ray emission in V1535 Sco was hard early in its evolution, which is much more typical of embedded novae. Second, those early X-rays were very bright. Using our assumed distance of 8.5 kpc, the early kT of 40 keV, and N$_{\rm H}\approx 10^{22}$ cm$^{-2}$, we find an X-ray luminosity of $\sim4\times10^{35}$ erg s$^{-1}$, which is much higher than any non-embedded nova we are aware of. For comparison, the X-ray luminosity for V382 Vel was $7\times10^{34}$ erg s$^{-1}$ (Mukai \& Ishida 2001), while RS Oph as over $10^{36}$ erg s$^{-1}$ (Mukai et al. 2008). While the X-ray data do not provide proof that the nova was embedded in wind material from the companion star, it is at least a plausible explanation for the interesting behavior.
\section{Discussion}
Bringing together all of our various observations and examining them together leads to some interesting conclusions. Figure~\ref{rxofig} shows information from all of our observing wavelengths with important times (e.g., the VLBA detection and the second radio synchrotron emission event) indicated. The initial negative radio spectral index combined with hard X-rays indicates synchrotron radiation produced by strong shocks. This is expected for a nova exploding into a thick wind from a giant companion. The transition from a negative to positive radio spectral index corresponds with both a softening of the X-rays while the total X-ray count rate remains relatively constant. This indicates the shocks are weakening and the plasma is cooling. At the same time, the radio thermal photosphere is expanding and begins to dominate the radio emission.
\begin{center}
\begin{deluxetable*}{cccccccc}
\tablewidth{0 pt}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.025in}
\tablecaption{ \label{swiftab2}
Best-Fit Model Parameters from \emph{Swift} Observations}
\tablehead{Obs\tablenotemark{1} & Day\tablenotemark{2} & N$_{\rm H}$ & kT$_{BB}$ & norm$_{BB}$ & kT$_{apec}$ & norm$_{apec}$ & cstat/DoF \\
ID & & ($10^{21}$ cm$^{-2}$) & (eV) & & (keV) & & }
\startdata
2 & 4.16 & 9.5$\pm$1.9 & 67$\pm$11 & 349$^{+14210}_{-341}$ & \textgreater41 & 0.021$\pm$0.003 & 373.39/377 \\
3 & 11.04 & 3$\pm$2 & 45$^{+16}_{-11}$ & 0.0037$^{+0.1359}_{-0.0035}$ & 3.8$^{+2.7}_{-1.3}$ & 0.0027$^{+0.0007}_{-0.0005}$ & 214.39/121 \\
4 & 13.94 & 9.6$\pm$1.9 & 35$\pm$4 & 5.9$^{+57.7}_{-5.3}$ & 1.0$\pm$0.09 & 0.0037$^{+0.0011}_{-0.0009}$ & 151.98/159 \\
5 & 17.93 & 9.61$\pm$2.4 & 45$^{+8}_{-4}$ & 0.74$^{+7.5}_{-0.68}$ & 1.013$^{+0.22}_{-0.17}$ & 0.0018$^{+0.0006}_{-0.0004}$ & 203.73/138 \\
6 & 24.75 & 18.4$^{+5.9}_{-8.3}$ & 31$^{+17}_{-45}$ & 4264$^{+21089}_{-4240}$ & 0.27$^{+0.65}_{-0.09}$ & 0.026$^{+0.200}_{-0.025}$ & 80.10/88 \\
\enddata
\tablenotetext{1}{All Observation ID's are 0003363400X}
\tablenotetext{2}{We take the time of initial detection 2015 February 11.837 UT (MJD 57064.837) to be Day 0.0}
\end{deluxetable*}
\end{center}
The ejecta cool and fade for several weeks until around Day 49 when the X-ray count rate appears to increase slightly (going from a non-detection to detection again) and the radio spectral index again shows evidence of synchrotron emission. There are two possible explanations for this second episode of synchrotron emission. First, the ejecta encountered another dense medium to shock against. Using the fit to the velocity from H$\alpha$ emission lines and integrating over time, the H$\alpha$ emitting region had a radius of approximately 13 AU by this time. If we assume that the ejecta that were emitting synchrotron radiation at this time was instead travelling at a velocity of 1659 km s$^{-1}$ for 49 days, that would be at a radius of approximately 47 AU. If we assume that the synchrotron-emitting material was from the ejecta producing the high velocity outliers in the H$\alpha$ spectra and travelled with a constant velocity of 4782 km $^{-1}$, it would have travelled 135 AU by Day 49. It is possible that the ejecta caught up to material ejected during a previous nova eruption. If we assume that the slowest material from a previous ejecta were moving at 50 km $^{-1}$, the previous eruption would have occurred 1.2 years ago for a distance of 13 AU, 4.5 years ago for a distance of 47 AU, or 12.8 years ago for a distance of 137 AU. Of these three, the 12.8 years seems the most likely, as it is hard to imagine the astronomy community could have missed multiple outbursts of this nova in the past 10 years. However, a recurrence time of 12.8 years would still require that this system is a very young recurrent nova undergoing only its second or third outburst, or that we have missed a very large number of its previous eruptions.
The second possibility is that the second synchrotron event is the result of shocks within the ejecta itself. We already have evidence for two different outflows from the width of the H$\alpha$ line and the high velocity outliers (see Section 2.4). It is not hard to imagine an interaction between these two outflows, or possibly even a third outflow such as a hot, fast wind driven by thermonuclear burning on the surface of the white dwarf but hidden below the optical photosphere of the bulk of the ejecta.
\begin{figure*}
\includegraphics[trim=20 80 20 80,clip,width=7.5in]{Sco15_VLA-Swift-SMARTS-size.pdf}
\caption{{\bf Top:} VLA light curve at 36.5, 7.38, and 1.26 GHz. {\bf 2nd:} \emph{Swift} total count rate. {\bf 3rd:} \emph{Swift} hardness ratio (1-10 keV/0.3-1.0 keV). {\bf 4th:} Optical photometry from the SMARTS program. {\bf Bottom:} Estimate of the radius of the ejecta {\bf shell}: blue -- velocity from H$\alpha$ line widths; red -- constant velocity of 1659 km s$^{-1}$; black -- constant velocity of 4782 km s$^{-1}$. {\bf Vertical Lines:} Solid = VLBA detection; Dashed = VLBA non-detection; Dotted = secondary maximum in the radio light curve; Dash-dotted = period of apparent second synchrotron bump in VLA spectral indices.}
\label{rxofig}
\end{figure*}
\subsection{Expected Brightness Temperature for VLBA Source}
Using the radial size, we can estimate what we expect the brightness temperature of the ejecta to be at the time of the VLBA observation using
\begin{equation}
T_{B} \approx \frac{S_{\nu}\, c^{2}\, D^{2}}{2\, \pi\, k_{B}\, \nu^{2}\, R^{2}}
\end{equation}
where $R$ is the radius of the spherically expanding ejecta and $D$ is the distance to the source (e.g., Seaquist \& Bode 2008). We will start by assuming the size of the radio emitting region is the same as the H$\alpha$ emitting region. From our fit to the H$\alpha$ line widths presented in Section 2.4, we derive a radius for the H$\alpha$ emitting region of $8.98\times10^{13}$ cm on Day 7.663. Using only the VLBA flux of 0.477 mJy, and assuming a distance of 8.5 kpc, this gives $T_{B} \approx 1.8 \times 10^{7}$ K. Alternatively, the radius could be the distance travelled by ejecta moving with a velocity of 1659 km s$^{-1}$, which gives is $1.10\times10^{14}$ cm. Using the VLBA flux and a distance of 8.5 kpc, this material would have a $T_{B} \approx 1.2 \times 10^{7}$ K. Both of these are significantly higher than the estimate of $5.8\times10^{5}$ K using the distance-free formula for $T_{B}$ in Section 4.2. To get the largest possible radius, we could assume the ejecta were travelling at 4782 km s$^{-1}$ for 7.663 days, giving a radius of $3.17\times10^{14}$ cm and $T_{B} \approx 1.4 \times 10^{6}$ K. The discrepancy between the observational and expected $T_{B}$ can indicate several possibilities: first, the ejecta were unresolved by the VLBA; second, the nova is much closer than 8.5 kpc; third, the ejecta shape is highly non-spherical; or some combination of these reasons. We should note that the expected $T_{B}$ using the largest possible radius agrees well with the observational $T_{B}$ assuming 2 point sources. However, using the largest possible radius would only give us a single sphere with an angular size of 2.5 mas at 8.5 kpc (which just happens to be the minor axis of the VLBA restoring beam for our observation), not the 2 point sources modelled in Section 4.2, so we cannot reconcile the measurement with the observation quite so easily.
From Section 3, recall that the closest the nova can be is 5.8 kpc. At this distance, the brightness temperature would be $8.3 \times 10^{6}$ K for a radius of $8.98\times10^{13}$ cm, $5.5 \times 10^{6}$ K for a radius of $1.10\times10^{14}$ cm, or $6.7 \times 10^{5}$ K for a radius of $3.17\times10^{14}$ cm. The first two are still larger than our value of $5.8 \times 10^{5}$ K we calculated for the single-component model in Section 4.2, although they are comparable to the two-component model. The final one, using our maximum possible radius, agrees well with the single-component model $T_{B}$ in Section 4.2. However, recall that we find it more likely that the high velocity outliers in the H$\alpha$ spectra represent a bipolar outflow, not an expanding spherical shell.
Note that this approach to estimating what the brightness temperature should be at the time of the VLBA detection assumes that the entire surface of the ejecta shell is emitting. If the emission is coming from compact, jet-like structures as in RS Oph (Rupen et al. 2008; Sokoloski et al. 2008), the value for $R$ would be smaller and the estimate for $T_{B}$ would be even larger. Placing the nova at further distances also increases the estimate for $T_{B}$.
Because placing the nova closer does not make the values of $T_{B}$ match, it is more likely that the VLBA did not fully resolve the compact emitting region on Day 7.663. Because of the highly-elliptical shape of the restoring beam, it is especially likely that there is unresolved structure in the north-south direction. However, the image does look like there is some resolved structure in the east-west direction. We find it likely that the compact component is non-spherical, possibly bi-lobed, with more extension in the east-west direction.
\subsection{X-ray and Radio Emission Measures Near Day 25}
We used the emission measures for both radio and X-ray to investigate whether the secondary maximum in the radio light curve that occurs between Day 23.663 and 26.563 can be explained by thermal radio emission from the X-ray emitting plasma. For the X-rays, the emission measure depends on the volume of the emitting region:
\begin{equation}
EM_{x} = \int n_{i}\, n_{e}\, dV
\end{equation}
where $n_{i}$ and $n_{e}$ are the number densities of ions and electrons, respectively, and $V$ is the volume of emitting material.
Generally it is assumed that $n_{e} \approx n_{i} = constant$ so that it simplifies to $EM_{x} \approx n_{e}^{2}V$. To make a further simplification, the emitting volume is assumed to be thin compared to the total extent of the ejecta so that it can be approximated as a uniform slab: $V \approx \pi r^{2} dl$, where $r$ is the radius of the ejecta (we are assuming the emitting region can be approximated as a cylinder) and $dl$ is the thickness of the emitting region. The X-ray emission measure can be determined from the hot plasma parameters in the fits to the \textit{Swift} data:
\begin{equation}
EM_{x} = (4 \times 10^{14})\, \pi\, D^{2}\, norm_{apec}
\end{equation}
where $D$ is the distance to the nova. Assuming a distance of 8.5 kpc (placing the nova approximately at the Galactic Center) and using $norm_{apec}$ from Day 24.75 (see Table~\ref{swiftab2}) gives $EM_{x} \approx 2.25 \times 10^{58}$ cm$^{-3}$. Assuming the ejecta travel with a constant velocity of 1659 km s$^{-1}$, we can determine the approximate radius for the shell on Day 24.75 was $3.5 \times 10^{14}$ cm. If we assume that the thickness of the emitting shell is $\sim10\%$ of its radius, we can solve for the number density of the emitting electrons: $n_{e} \approx 4.0 \times 10^{7}$ cm$^{-3}$ emitting at a temperature of approximately $10^{6}$ K (from kT$_{apec}$ in Table~\ref{swiftab2}). Note that if we assume a larger distance to the nova, both $EM_{x}$ and $n_{e}$ increase.
The radio (path length) emission measure is given by:
\begin{equation}
EM_{r} = n_{e}^{2}\, dl
\end{equation}
plugging in the same $dl$ (but this time in pc) used for $EM_{x}$ and using $n_{e}$ derived from the X-ray data gives $EM_{r} = 1.8 \times 10^{10}$ cm$^{-6}$ pc. The radio emission measure is also part of the equation for optical depth, $\tau_{\nu}$ (Rohlfs \& Wilson 2006):
\begin{equation}
\tau_{\nu} = 8.235\times10^{-2}\, T_{e}^{-1.35}\, \nu^{-2.1}\, EM_{r}\, a(\nu,T)
\end{equation}
Therefore, another way to calculate the radio emission measure is given by:
\begin{equation}
EM_{r} = 12.143\, T_{e}^{1.35}\, \nu^{2.1}\, \tau_{\nu}\, a(\nu,T)^{-1}
\end{equation}
where $T_{e}$ is the electron temperature in K, $\nu$ is the observing frequency in GHz, $\tau_{\nu}$ is the optical depth at the observing frequency, and $a(\nu,T)$ is a correction term usually assumed to be 1. In order for the emission to be optically thick at the observed frequency, $\tau_{\nu}\gtrsim1$. Because we are trying to determine if the radio flux can be explained by the X-ray plasma, we set $T_{e} \approx 10^{6}$ K from the kT$_{apec}$ value from the fits to the Swift data on Day 24.75 (see Table~\ref{swiftab2}). We use our highest observing frequency of 36.5 GHz because it provides the tightest constraint. This gives us $EM_{r} \approx 2.92 \times 10^{12}$ cm$^{-6}$ pc, which is more than two orders of magnitude larger than expected from the X-ray electron number density used above. Therefore, under the reasonable assumptions we have made, the X-ray emitting plasma cannot account for the thermal radio emission detected around Day 24. It is more likely that there is also a warm ($\sim10^{4}$K) ionized ejecta that begins to dominate the radio emission at this time.
Unfortunately, the uncertainty on $norm_{apec}$ from Day 24.75 is very high. If we use the $norm_{apec}$ value from Day 17.93 with its much lower uncertainty and assume that the number density remains mostly constant until Day 24, we get $n_{e} = 1.7\times10^{7}$ cm$^{-3}$, which would imply $EM_{r} = 2.4\times10^{9}$ cm$^{-6}$ pc. This makes the discrepancy between the two methods for determining the radio emission measure even greater, further strengthening our argument that the X-ray emitting plasma cannot account for the radio flux density observed during the second radio maximum.
\section{Conclusions}
V1535 Sco showed peculiar behavior at nearly every wavelength we observed. The radio emission started bright, faded quickly, and then had two re-brightening events. The spectral index for early radio emission and the second radio re-brightening event were consistent with optically thin synchrotron emission. The X-ray emission began promptly with a hard spectrum. There was a re-brightening in the X-rays which appeared to correspond to the second re-brightening event in the radio. The lack of connection between the first radio re-brightening event and the X-ray emission implies that are are at least two emitting components in the ejecta: one shock-heated plasma and one thermal bremsstrahlung. The optical observations indicate that the nova was discovered post-peak, and faded very fast. Optical spectroscopy also indicated the presence of two outflows: 1) a relatively slow ($\sim$1659 km s$^{-1}$) outflow; and 2) a fast ($\sim$4782 km s$^{-1}$), possibly bipolar outflow. Spectral monitoring of the H$\alpha$ line at late times also indicated the presence of a neutral, dense surrounding cloud of emitting material.
The early hard X-rays combined with the radio synchrotron emission and the detection with the VLBA strongly support the existence of strong shocks very early in the evolution of the nova. Such early strong shocks are most easily explained by the presence of a dense wind from the companion star. The estimated magnetic field from the VLBA detection was between 0.10 G and 0.18 G, which is larger than but comparable to the magnetic field strength found by Rupen et al. (2008) for RS Oph, a system known to have a red giant companion.
There was evidence for a second shock around Day 50. The radio spectral index at this time changed from being consistent with optically thick thermal bremsstrahlung, to being synchrotron-like. The X-ray emission also showed an increase at this time. We posit that this second shock was the result of collisions between multiple outflows within the ejecta, but the presence of a shell of dense material (possibly from a previous eruption) has not been completely ruled out.
The nova had strong hard X-ray emission early in its evolution, but no detectable $\gamma$-rays. However, it should be noted that V1324 Sco is the only nova \emph{Fermi} has detected which has distance comparable to V1535 Sco (Finzell et al. 2015). We also note that V745 Sco, with its red giant companion and presumably denser wind, was only marginally detected by \emph{Fermi} (Cheung et al. 2014). It is very likely that V1535 Sco produced at least some $\gamma$-ray emission, but it was simply too far away for \emph{Fermi} to detect.
To date, only a handful of Galactic novae with red giant companions are known, including the recurrent novae RS Oph and V745 Sco. However, recent studies on the nova population in M31 indicate that there may be many novae with red giants that are not detected, and they may constitute $\sim$30\% of the nova eruptions in the Milky Way (Williams et al. 2016). There is particular interest in discovering more of these novae because they are possible progenitors to Type Ia supernovae (e.g., Starrfield et al. 2012). In fact, Dilday et al. (2012) claim that the supernova PTF 11kx originated from such a system.
Novae that occur in a system with a red giant companion are often referred to as ``symbiotic novae''. However, the term ``symbiotic novae'' is also used to describe a particularly slow and long-lasting class of novae where nuclear burning on the surface of the white dwarf is sustained for several years, and sometimes decades such as PU Vul, HM Sge, and AG Peg (e.g., Iben Jr. \& Fujimoto 2008). In order to avoid confusion and better describe the general class of thermonuclear novae with evolved companions, our collaboration has adopted the term ``embedded nova'' to refer to any nova that is embedded in the dense wind of its post-main sequence companion star, regardless of the duration of its optical maximum (e.g. Chomiuk et al. 2012; Mukai et al. 2014).
We find some evidence that V1535 Sco is an embedded nova. The early strong, hard X-rays and non-thermal radio emission argue for the presence of a pre-existing dense material to shock against. Also, the X-ray emission showed signs of absorption beyond what is expected for the measured $E(B-V)$, which also points to the nova being embedded in some pre-existing cloud. The fast optical decline is also consistent with an embedded nova. However, the long-term behavior of the optical and near-infrared light curves were unlike other novae known to have red giant companions. We also did not observe the rapid decrease in N$_{\rm H}$ that previous X-ray observations of novae with known red giant companions have reported. We produced a model of the quiescent system as a K giant with an accretion disk which is in good agreement with the observed optical and infrared magnitudes. We therefore find that the companion star in this system is producing a significant stellar wind and is most likely a K giant, specifically a K3-5 III, which is in agreement with Munari et al. (2017). Several symbiotic binary systems with K giant stars are known, and they are often referred to as ``yellow symbiotics" (e.g., Baella et al. 2016).
\acknowledgements
We thank the anonymous referee for their thoughtful criticism of this manuscript which led to significant improvements. We are grateful to C.~C. Cheung for his insight into the operations of the \textit{Fermi Gamma-ray Space Telescope}. We also appreciate the useful conversations about novae, shocks, and accretion at the Stellar Remnants at the Junction meeting and the Conference on Shocks and Particle Acceleration in Novae and Supernovae.
We thank the NRAO for the generous allocation of VLA and VLBA time for our observations. We also thank Neil Gehrels and the \textit{Swift} mission for the generous allocation of target-of-opportunity time to observe V1535 Sco.
We acknowledge support from NASA award NNX14AQ36G.
T.~N. was supported in part by NASA award NNX13A091G.
J.~L.~S. and J.~H.~S.~W. were funded in part by NSF award AST-1211778.
The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. Access to SMARTS has been made possible by generous support from the office of the Provost of Stony Brook University. This research made use of \emph{SciPy}, an open-source library of numerical routines for scientific computing available at http://www.scipy.org; \emph{APLpy}, an open-source plotting package for Python hosted at http://aplpy.github.com; and \emph{Astropy}, a community-developed core Python package for Astronomy (AstroPy Collaboration, 2013) available at http://www.astropy.org. Many of the figures in this manuscript were made using the matplotlib 2D graphics package for Python.\\
\facility{Karl G. Jansky VLA, VLBA, Swift, SMARTS}
\software{AIPS, difmap, CASA, XSpec, SciPy, APLpy, AstroPy, matplotlib, IDL}
|
\section{Introduction}
\label{sec:introduction}
Many modern machine learning techniques require large amounts of training data to reach their full potential.
However, annotated data is hard and expensive to obtain, notably in specialized domains where only experts whose time is scarce and precious can provide reliable labels. Active learning (AL) aims to ease the data collection process by automatically deciding which instances an annotator should label to train an algorithm as quickly and effectively as possible.
Over the years many AL strategies have been developed for various classification tasks, without any one of them clearly outperforming others in all cases.
Consequently, a number of meta-AL approaches have been proposed to automatically select the best strategy. Recent examples include bandit algorithms~\cite{Baram04,Hsu15, Chu16b} and reinforcement learning approaches~\cite{Ebert12}.
A common limitation of these methods is that they cannot go beyond combining pre-existing hand-designed heuristics.
Besides, they require reliable assessment of the classification performance which is problematic because the annotated data is scarce.
In this paper, we overcome these limitations thanks to two features of our approach.
First, we look at a whole continuum of AL strategies instead of combinations of pre-specified heuristics.
Second, we bypass the need to evaluate the classification quality from application-specific data because we rely on experience from previous tasks instead.
More specifically, we formulate Learning Active Learning ({\textsc{LAL}}{}) as a regression problem.
Given a trained classifier and its output for a specific sample without a label, we predict the reduction in generalization error that can be expected by adding the label to that point.
In practice, we show that we can train this regression function on synthetic data by using simple features, such as the variance of the classifier output or the predicted probability distribution over possible labels for a specific datapoint.
Furthermore, if a sufficiently large annotated set can be provided initially, the regressor can be trained on it instead of on synthetic data.
The resulting AL strategy is then tailored to the particular problem at hand, and can be used to further extend the initial dataset.
We show that {\textsc{LAL}}{} works well on real data from several
different domains such as biomedical imaging, economics, molecular
biology and high energy physics.
This query selection strategy outperforms competing methods without requiring hand-crafted heuristics and at a comparatively low computational cost.
\section{Related work }
\label{sec:related}
The extensive development of AL in the last decade has resulted in various AL strategies.
They include uncertainty sampling~\cite{Tong02,Joshi09,Settles10,Yang15}, query-by-committee~\cite{GiladBachrach05,Iglesias11}, expected model change~\cite{Settles10,Sznitman10,Vezhnevets12}, expected error or variance minimization~\cite{Joshi12, Hoi06} and information gain~\cite{Houlsby11}.
Among these, uncertainty sampling is both simple and computationally efficient.
This makes it one of the most popular strategies in real applications.
In short, it suggests labeling samples that are the most uncertain, i.e., closest the classifier's decision boundary.
The above methods work very well in cases such as the ones depicted in the top row of Fig.~\ref{fig:synthetic}, but often fail in the more difficult ones of the bottom row~\cite{Baram04}.
Among AL methods, some cater to specific classifiers, such as those relying on Gaussian Processes~\cite{Kapoor07}, or to specific applications, such as natural language processing~\cite{Tong02,Olsson09}, sequence labeling tasks~\cite{Settles08b}, visual recognition~\cite{Luo04,Long15a}, semantic segmentation~\cite{Vezhnevets12}, foreground-background segmentation~\cite{Konyushkova15}, and preference learning \cite{singla2016a, Maystre17}.
Moreover, various query strategies aim to maximize different performance metrics, as evidenced in the case of multi-class classification~\cite{Settles10}.
However, there is no one algorithm that consistently outperforms all others in all applications~\cite{Settles08b}.
Meta-learning algorithms have been gaining in popularity in recent years~\cite{tamar2016, Santoro16}, but few AL scenarios tackle the problem of learning AL strategies.
\citet{Baram04} combine several known heuristics with the help of a bandit algorithm.
This is made possible by the maximum entropy criterion, which estimates the classification performance without labels.
\citet{Hsu15} improve it by moving the focus from datasamples as arms to heuristics as arms in the bandit and use a new unbiased estimator of the test error.
\citet{Chu16b} go further and transfer the bandit-learnt combination of AL heuristics between different tasks.
Another approach is introduced by~\citet{Ebert12}. It involves balancing exploration and exploitation in the choice of samples with a Markov decision process.
The two main limitations of these approaches are as follows.
First, they are restricted to combining already existing techniques and second, their success depends on the ability to estimate the classification performance from scarce data.
The data-driven nature of {\textsc{LAL}}{} helps to overcome these limitations.
Sec.~\ref{sec:experiments} shows that it outperforms several baselines including those of \citet{Hsu15} and \citet{Kapoor07}.
The method of \citet{Hsu15} is chosen as a our main baseline because it is a recent example of meta AL and is known to outperform several benchmarks.
\section{Towards data-driven active learning}
\label{sec:motivation}
In this section we briefly introduce the active leaning framework along with uncertainty sampling (US), the most frequently-used AL heuristic.
Then, we motivate why a data-driven approach can improve AL strategies and how it can deal with the situations where US fails.
We selected US as a representative method because it is popular and widely applicable, however the behavior that we describe is not specific to this strategy.
\subsection{Active learning (AL)}
\label{sec:standardAL}
Given a machine learning model and a pool of unlabeled data,
the goal of AL is to select which data should be annotated in order to learn the model as quickly as possible.
In practice, this means that instead of asking experts to annotate all the data, we select iteratively and adaptively which datapoints should be annotated next
In this paper we are interested in classifying datapoints from a target dataset $\mathcal{Z} = \{(x_1, y_1),\ldots,(x_N, y_N)\}$, where $x_i$ is a $D$-dimensional feature vector and $y_i \in \{0,1\}$ is its binary label.
We choose a probabilistic classifier $f$ that can be trained on some $\mathcal{L}_t \subset \mathcal{Z}$ to map features to labels, $f_t(x_i) = \hat{y}_i$, through the predicted probability $p_t(y_i=y \mid x_i)$.
The standard AL procedure unfolds as follows.
\begin{enumerate}
\item The algorithm starts with a small labeled training dataset $\mathcal{L}_t \subset \mathcal{Z}$ and large pool of annotated data $\mathcal{U}_t = \mathcal{Z} \setminus \mathcal{L}_t$ with $t=0$.
\item A classifier $f_t$ is trained using $\mathcal{L}_t$.
\item A query selection procedure picks an instance $x^* \in \mathcal{U}_t$ to be annotated at the next iteration.
\item $x^*$ is given a label $y^*$ by an oracle. The labeled and unlabeled sets are updated.
\item $t$ is incremented, and steps \num{2}--\num{5} iterate until the desired accuracy is achieved or the number of iterations has reached a predefined limit.
\end{enumerate}
\paragraph{Uncertainty sampling (US)}
\label{sec:uncertainty}
US has been reported to be successful in numerous scenarios and settings and despite its simplicity, it often works remarkably well~\cite{Tong02,Joshi09,Settles10,Yang15,Konyushkova15,Mosinska16}.
It focuses its selection on samples which the current classifier is the least certain about.
There are several definitions of maximum uncertainty but one of the most widely used ones is to select a sample $x^*$ that maximizes the entropy $\mathcal{H}$ over the predicted classes:
\begin{equation}
x^* = \argmax_{{x_i} \in \mathcal{U}_t} \mathcal{H}[p_t(y_i=y \mid x_i)]
\; .
\label{eq:uncertainty}
\end{equation}
\subsection{Success, failure, and motivation}
\label{subsec:motivational-example}
We now motivate the need for LAL by presenting two toy examples.
In the first one, US is empirically observed to be the best greedy approach, but in the second it makes suboptimal decisions.
Let us consider simple two-dimensional datasets $\mathcal{Z}$ and $\mathcal{Z}'$ drawn from the same distribution
with an equal number of points in each class (Fig.~\ref{fig:exampleGauss}, left).
The data in each class comes from a Gaussian distribution with a different mean and the same variance.
We can initialize the AL procedure of Sec.~\ref{sec:standardAL} with one sample from each class and its respective label: $\mathcal{L}_0=\{(x_1,0), (x_2,1) \} \subset \mathcal{Z}$ and $\mathcal{U}_0 = \mathcal{Z} \setminus \mathcal{L}_0$.
Here we train a simple logistic regression classifier $f$ on $\mathcal{L}_0$ and then test it on $\mathcal{Z}'$.
If~ $|\mathcal{Z}'|$ is large, the test error can be considered as a good approximation of the generalization error:
$\ell_0 = \sum_{(x',y') \in \mathcal{Z}'}\ell(\hat{y},y')$, where $\hat{y}=f_0(x')$.
Let us try to label every point $x$ from~ $\mathcal{U}_0$ one by one, form a new labeled set $\mathcal{L}_x = \mathcal{L}_0 \cup (x,y)$ and check what error a new classifier $f_x$ yields on $\mathcal{Z}'$, that is, $\ell_x = \sum_{(x',y') \in \mathcal{Z}'} \ell (\hat{y},y')$, where $\hat{y}=f_x(x')$.
The difference between errors obtained with classifiers constructed on $\mathcal{L}_0$ and $\mathcal{L}_x$ indicates how much the addition of a new datapoint $x$ reduces the generalization error: $\delta_x = \ell_0-\ell_x$.
We plot $\delta_x$ for the $0 / 1$ loss function, averaged over \num{10000} experiments as a function of the predicted probability $p_0$ (Fig.~\ref{fig:exampleGauss}, left).
By design, US would select a datapoint with probability of class \num{0} close to \num{0.5}.
We observe that in this experiment, the datasample with $p_0$ closest to \num{0.5} is indeed the one that yields the greatest error reduction.
\input{figs/exampleGauss}
In the next experiment, the class \num{0} contains twice as many datapoints as the other class, see Fig.~\ref{fig:exampleGauss}, right.
As before, we plot the average error reduction as a function of $p_0$ in Fig.~\ref{fig:exampleGauss} (right).
We observe this time that the value of $p_0$ that corresponds to the largest expected error reduction is different from \num{0.5} and thus the choice of US becomes suboptimal.
Also, the reduction in error is no longer symmetric for the two classes.
The more imbalanced the two classes are, the further from the optimum the choice made by US is.
In complex realistic scenario, there are many other factors such as label noise, outliers or shape of distribution that further compound the problem.
Although query selection procedures can take into account statistical properties of the datasets and classifier, there is no simple way to foresee the influence of all possible factors.
Thus, in this paper, we suggest Learning Active Learning ({\textsc{LAL}}{}).
It uses properties of classifiers and data to predict the potential error reduction.
We treat the query selection problem by using a regression model; this perspective enables us to construct new AL strategies in a flexible way.
For instance, in the example of Fig.~\ref{fig:exampleGauss} (right) we expect {\textsc{LAL}}{} to learn a model that automatically adapts its selection to the relative prevalence of the two classes without having to explicitly state such a rule.
\section{Monte-Carlo LAL}
\label{sec:approach}
Our approach to AL is data-driven and can be formulated as a regression problem.
Given a {\it representative} dataset with ground truth, we simulate an online learning procedure using a Monte-Carlo approach.
We propose two versions of AL strategies.
When building the first one, {\textsc{LALindependent}}{}, we incorporate unused labels individually and at random to retrain the classifier.
Our goal is to correlate the change in test performance with the properties of the classifier and of newly added datapoint.
To build the {\textsc{LALiterative}}{} strategy, we further extend our method by a sequential procedure to account for selection bias caused by AL.
We formalize our LAL procedure in the remainder of the section.
\subsection{Independent LAL}
\label{sec:monteCarlo}
Let the {\it representative} dataset be split into a training $\mathcal{D}$ and a testing set $\mathcal{D}'$.
Let $f$ be a classifier with a given training procedure.
We start collecting data for the regressor by splitting $\mathcal{D}$ into a labeled set $\mathcal{L}_{\tau}$ of size $\tau$ and an unlabeled set~ $\mathcal{U}_{\tau}$ containing the remaining points (Alg.~\ref{alg:lal-construct-data} {\textsc{DataMonteCarlo}}{}).
We then train a classifier $f$ on $\mathcal{L}_{\tau}$, resulting in a function $f_{\tau}$ that we use to predict class labels for elements $x'$ from the test set $\mathcal{D}'$ and estimate the test classification loss $\ell_{\tau}$.
We characterize the classifier state by $K$ parameters $\phi_\tau = \{ \phi^1_\tau,\ldots,\phi^K_\tau \}$, which are specific to the particular classifier type and are sensitive to the change in the training dataset while being relatively invariant to the stochasticity of the optimization procedure.
For example, they can be the parameters of the kernel function if $f$ is kernel-based, the average depths of the trees if $f$ is a random forest, or prediction variability if $f$ is an ensemble classifier.
The above steps are summarized in lines \num{3}--\num{5} of Alg.~\ref{alg:lal-construct-data}.
Next, we randomly select a new datapoint $x$ from $\mathcal{U}_\tau$ which is characterized by $R$ parameters $\psi_x = \{ \psi^1_x,\ldots,\psi^R_x \}$.
For example, they can include the predicted probability to belong to class $y$, the distance to the closest point in the dataset or the distance to the closest labeled point.
We form a new labeled set $\mathcal{L}_x = \mathcal{L}_{\tau} \cup \{x\}$ and retrain $f$ (lines \num{7}--\num{13} of Alg.~\ref{alg:lal-construct-data}).
The new classifier $f_x$ results in the test-set loss
$\ell_x$.
Finally, we record the difference between previous and new loss $\delta_x =
\ell_{\tau} - \ell_x$ which is associated to the learning state in which it was received.
The learning state is characterized by a vector
$ \xi_\tau^x = \begin{bmatrix} \phi^1_\tau & \cdots & \phi^K_\tau & \psi^1_x & \cdots & \psi^R_x \end{bmatrix} \in \mathbb{R}^{K+R}$,
whose elements depend both on the state of the current classifier $f_\tau$ and on the datapoint $x$.
\input{algs/randLALdata}
To build an AL strategy {\textsc{LALindependent}}{} we repeat the {\textsc{DataMonteCarlo}}{} procedure for $Q$ different initializations $\mathcal{L}_\tau^1, \mathcal{L}_\tau^2,\ldots,\mathcal{L}_\tau^Q$ and $T$ various labeled subset sizes $\tau = 2,\ldots,T+2$ (Alg.~\ref{alg:lal-monte-carlo} lines \num{4} and \num{5}).
For each initialization $q$ and iteration $\tau$, we sample $M$ different datapoints $x$ each of which yields classifier/datapoint state pairs with an associated reduction in error (Alg.~\ref{alg:lal-construct-data}, line \num{13}).
This results in a matrix $\Xi \in \mathbb{R}^{(QMT)\times(K+R)}$ of observations $\xi$ and a vector $\Delta \in \mathbb{R}^{QMT}$ of labels $\delta$ (Alg.~\ref{alg:lal-monte-carlo}, line \num{9}).
Our insight is that observations $\xi$ should lie on a smooth manifold and that similar states of the classifier result in similar behaviors when annotating similar samples.
From this, a regression function can predict the potential error reduction of annotating a specific sample in a given classifier state.
Line \num{10} of {\textsc{buildLALindependent}}{} algorithm looks for a mapping $g: \xi \rightarrow \delta$, which is not specific to the dataset $\mathcal{D}$, and thus can be used to detect samples that promise the greatest increase in classifier performance in other target domains $\mathcal{Z}$.
The resulting {\textsc{LALindependent}}{} strategy greedily selects a datapoint with the highest potential in error reduction at iteration $t$ by taking the maximum of the value predicted by the regressor $g$:
\begin{equation}
x^* =\argmax_{x \in \mathcal{U}_t} g(\phi_t, \psi_x).
\label{eq:lal-monte-carlo}
\end{equation}
\subsection{Iterative LAL}
\label{sec:approach-iterative}
For any AL strategy at iteration $t>0$, the labeled set $\mathcal{L}_t$ consists of samples selected at previous iterations, which is clearly {\it not} random.
However, in Sec.~\ref{sec:monteCarlo} the dataset $\mathcal{D}$ is split into $\mathcal{L}_\tau$ and $\mathcal{U}_\tau$ randomly no matter how many labeled samples $\tau$ are available.
To account for this, we modify the approach of Section~\ref{sec:monteCarlo} in Alg.~\ref{alg:lal-iterative} {\textsc{buildLALiterative}}{}.
Instead of partitioning the dataset $\mathcal{D}$ into $\mathcal{L}_\tau$ and $\mathcal{U}_\tau$ randomly, we suggest simulating the AL procedure which selects datapoints according to the strategy learnt on the previously collected data (Alg.~\ref{alg:lal-iterative}, line \num{10}).
It first learns a strategy $\mathcal{A}(g_2)$ based on a regression function $g_2$ which selects the most promising \nth{3} datapoint when \num{2} random points are available.
In the next iteration, it learns a strategy $\mathcal{A}(g_3)$ that selects \nth{4} datapoint given \num{2} random points and \num{1} selected by $\mathcal{A}(g_2)$ etc.
In this way, samples at each iteration depend on the samples at the previous iteration and the sampling bias of AL is represented in the data $\Xi, \Delta$ from which the final strategy {\textsc{LALiterative}}{} is learnt.
The resulting strategies {\textsc{LALindependent}}{} and {\textsc{LALiterative}}{} are both reasonably fast during the online steps of AL.
The offline part, generating a datasets to learn a regression function, can induce a significant computational cost depending on the parameters of the algorithm.
For this reason, {\textsc{LALindependent}}{} is preferred to {\textsc{LALiterative}}{} when an application-specific strategy is needed.
\input{algs/randLAL-AL}
\section{Experiments}
\label{sec:experiments}
\paragraph{Implementation details}
\label{sec:lalimplementation}
We test AL strategies in two possible settings:
\begin{enumerate*}[label={\alph*)}]
\item {\it cold start}, where we start with one sample from each of two classes and
\item {\it warm start}, where a larger dataset of size $N_0 \ll N$ is available to train the initial classifier.
\end{enumerate*}
The {\it warm start scenario} is largely overloooked in the litterature, but we believe it has a significant practical interest.
Learning a classifier for a real-life application with AL rarely starts from scratch, but a small initial annotated set is provided to understand if a learning based approach is applicable at all.
While a small set is good to provide an initial insight, a real working prototype still requires much more training.
In this situation, we can benefit from the available training data to learn a specialized AL strategy for an application.
In {\it cold start} we take the representative dataset to be a \num{2}D synthetic dataset where class-conditional data distributions are Gaussian.
In most of the experiments, we use Random Forest (RF) classifiers for $f$ and a RF regressor for $g$.
The state of the learning process consists of the following features:
\begin{enumerate*}[label={\alph*)}
\item predicted {\it probability} $p(y=0 | \mathcal{L}_t, x)$;
\item {\it proportion} of class \num{0} in $\mathcal{L}_t$;
\item {\it out-of-bag} cross-validated accuracy of $f_t$;
\item variance of {\it feature importances} of $f_t$;
\item {\it forest variance} computed as variance of trees' predictions on $\mathcal{U}_t$;
\item average {\it tree depth} of the forest;
\item {\it size} of $\mathcal{L}_t$.
\end{enumerate*}
For additional implementational details, including examples of the synthetic datasets, parameters of the data generation algorithm and features in the case of GP classification, we refer to the supplementary materials.
\paragraph{Baselines and protocol}
We compare the three versions of our approach:
\begin{enumerate*}[label={\alph*)}
\item {\bf LAL-independent-2D}{}, {\textsc{LALindependent}}{} strategy trained on a synthetic dataset of {\it cold start};
\item {\bf LAL-iterative-2D}{}, {\textsc{LALiterative}}{} strategy trained on a synthetic dataset of {\it cold start};
\item {\bf LAL-independent-WS}{}, {\textsc{LALindependent}}{} strategy trained on {\it warm start} representative data;
\end{enumerate*}
against the following baselines:
\begin{enumerate*}[label={\alph*)}
\item {\bf Rs}{}, random sampling;
\item {\bf Us}{}, uncertainty sampling;
\item {\bf Kapoor}{}~\cite{Kapoor07}, an algorithm that balances exploration and exploitation by incorporating mean and variance estimation of the GP classifier;
\item {\bf ALBE}{}~\cite{Hsu15}, a recent example of meta-AL that adaptively uses a combination of strategies, including ~\cite{Huang10}, {\bf Us}{} and {\bf Rs}{}.
\end{enumerate*}
In all AL experiments we select samples from a training set and report the classification performance on an independent test set.
We repeat each experiment \num{50}--\num{100} times with random permutations of training and testing splits and different initializations.
Then we report the average test performance as a function of the number of labeled samples.
The performance metrics are task-specific and include classification accuracy, IOU~\cite{Everingham10}, dice score~\cite{Gordillo13}, AMS score~\cite{adam15}, as well as area under the ROC curve.
\subsection{Synthetic data}
\input{figs/experimentsSynthetic}
\paragraph{Two-Gaussian-clouds experiments}
In this dataset we test our approach with two classifiers: RF and Gaussian Process classifier (GPC).
Due to the the computational cost of GPC, it is only tested in this experiment.
We generate \num{1000} new unseen synthetic datasets as shown in the top row of Fig.~\ref{fig:synthetic}.
In both cases the proposed {\textsc{LAL}}{} strategies selects datapoints that help to construct better classifiers faster than {\bf Rs}{}, {\bf Us}{}, {\bf Kapoor}{} and {\bf ALBE}{}.
\paragraph{XOR-like experiments}
XOR-like datasets are known to be challenging for many machine learning methods and AL is not an exception.
It was reported in~\citet{Baram04} that various AL algorithms struggle with tasks such as those depicted in the bottom row of Fig.~\ref{fig:synthetic}, namely {\it Checkerboard \num{2 x 2}}, {\it Checkerboard \num{4 x 4}}, and the {\it Banana} dataset from~\citet{Ratsch01}.
As previously observed, {\bf Us}{} loses to {\bf Rs}{} in these cases.
{\bf ALBE}{} does not suffer from such adversarial conditions as much as {\bf Us}{}, but {\bf LAL-iterative-2D}{} outperforms it on {\it Checkerboard \num{2 x 2}} and {\it Checkerboard \num{2 x 2}} and matches its performance on the {\it Banana} dataset.
\subsection{Real data}
We now turn to real data from domains where annotating is hard because it requires special training to do so correctly:
\begin{enumerate*}[label={\alph*)}
\item {\it Striatum}, \num{3}D Electron Microscopy stack of rat neural tissue, the task is to detect and segment mitochondria~\cite{Lucchi12, Konyushkova15};
\item {\it MRI}, brain scans obtained from the BRATS competition~\cite{Menze14}, the task is to segment brain tumor in T1, T2, FLAIR, and post-Gadolinium T1 MR images;
\item {\it Credit card}~\cite{Dal15}, a dataset of credit card transactions made in \num{2013} by European cardholders, the task is to detect fraudulent transactions;
\item {\it Splice}, a molecular biology dataset with the task of detecting splice junctions in DNA sequences~\cite{Lorena02};
\item {\it Higgs}, a high energy physics dataset that contains measurements simulating the ATLAS experiment~\cite{adam15}, the task is to detect the Higgs boson in the noise signal.
\end{enumerate*}
Additional details about the above datasets including sizes, dimensionalities and preprocessing techniques can be found in the supplementary materials.
\paragraph{Cold Start AL}
Top row of Fig.~\ref{fig:realcold} depicts the results of applying {\bf Rs}{}, {\bf Us}{}, {\bf LAL-independent-2D}{}, and {\bf LAL-iterative-2D}{} on the {\it Striatum, MRI}, and {\it Credit card} datasets.
Both {\textsc{LAL}}{} strategies outperform {\bf Us}{}, with {\bf LAL-iterative-2D}{} being the best of the two.
Considering that the {\textsc{LAL}}{} regressor was learned using a simple synthetic \num{2}D dataset, it is remarkable that it work effectively on such complex and high-dimensional tasks.
Due to the high computational cost of {\bf ALBE}{}, we downsample {\it Striatum} and {\it MRI} datasets to \num{2000} datapoints (referred to as {\it Striatum mini} and {\it MRI mini}).
Downsampling was not possible for the {\it Credit card} dataset due to the sparsity of positive labels (\num{0.17}\%).
We see in the bottom row of Fig.~\ref{fig:realcold} that {\bf ALBE}{} performs even worse than {\bf Us}{}.
We ascribe this to the lack of labeled data, which {\bf ALBE}{} needs to estimate classification accuracy (see Sec.~\ref{sec:related}).
\input{figs/experimentsRealCold}
\paragraph{Warm Start AL}
In Fig.~\ref{fig:realwarm} we compare {\bf LAL-independent-WS}{} on the {\it Splice} and {\it Higgs} datasets by initializing {\textsc{buildLALindependent}}{} with \num{100} and \num{200} datapoints from the corresponding tasks.
We tested {\bf ALBE}{} on the {\it Splice} dataset, however in the {\it Higgs} dataset the number of iterations in the experiment is too big for it.
{\bf LAL-independent-WS}{} outperforms other methods with {\bf ALBE}{} delivering competitive performance---yet, at a high computational cost---only at the end of AL.
\input{figs/experimentsRealWarm}
\subsection{Analysis of LAL strategies and time comparison}
\label{sec:analysis}
To better understand LAL strategies, we show in Fig.~\ref{fig:featimport} (left) the relative importance of the features of the regressor $g$ for {\textsc{LALiterative}}{}.
As expected, both classifier state parameters and datapoint parameters influence the AL selection.
In order to understand what kind of selection {\textsc{LALindependent}}{} and {\textsc{LALiterative}}{} do, we record the predicted probability of the chosen datapoint $p(y^*=0 | \mathcal{D}_t, x^*)$ in \num{10} {\it cold start} experiments with the same initialization on the {\it MRI} dataset.
Fig.~\ref{fig:featimport}(right) shows the histograms of these probabilities for {\bf Us}{}, {\bf LAL-independent-2D}{} and {\bf LAL-iterative-2D}{}.
LAL strategies have high variance and modes different from \num{0.5}.
Not only does the selection by LAL strategies differ significantly from standard US, but also the independent and iterative approaches differ from each other.
\input{figs/alalysisFeatImpHist}
\paragraph{Computational costs}
While collecting synthetic data can be slow, it must only be done {\it once, offline,} for all applications.
Collecting data offline for {\it warm start}, that is application specific, took us approximately \num{2.7}h and \num{1.9}h for {\it Higgs} and {\it Splice} datasets respectively.
By contrast, the online user-interaction part is fast: it simply consists of learning $f_t$, extracting learning state parameters and evaluating the regressor $g$.
The LAL run time depends on the parameters of the random forest regressor which are estimated via cross-validation (discussed in the supplementary materials).
Run times of a python-based implementation with \num{1} core are given in Tab.~\ref{tab:time} for a typical parameter set ($\pm$ \num{20}\% depending on exact parameter values).
Real-time performance can be attained by parallelising and optimising the code, even in applications with large amounts of high-dimensional data.
\input{figs/analysisTime}
\section{Conclusion}
In this paper we introduced a new approach to AL that is driven by data: Learning Active Learning.
We found out that Learning Active Learning from simple 2D data generalizes remarkably well to challenging new domains.
Learning from a subset of application-specific data further extends the applicability of our approach.
Finally, LAL demonstrated robustness to the choice of type of classifier and features.
\section*{Acknowledgements}
This project has received funding from the European Union’s Horizon 2020 Research and Innovation Programme under Grant Agreement No. 720270 (HBP SGA1).
We would like to thank Carlos Becker and Helge Rhodin for their comments on the text, and Lucas Maystre for his discussions and attention to details.
|
\section{Introduction}\label{S:1}
Following Ozawa \cite{ozawa1,ozawa2}, the \textbf{representativity} $r(L)$ of a link $L\subset S^3$ is:
\[r(L)=\max_{F\in\mathcal{F}_L}\min_{X\in\mathcal{X}_F}|\partial X\cap L|,\]
where $\mathcal{F}_L$ is the set of positive genus, closed surfaces $F\subset S^3$ containing $L$, and a ``closed surface'' is compact and connected without boundary; $\mathcal{X}_F$ is the set of compressing disks for $F$ in $S^3$; and $|\partial X\cap L|$ is the number of connected components (i.e. points) of $\partial X\cap L$. This notion extends the earlier notion of representativity from graph theory. In 2011 \cite{pardon}, Pardon applied representativity (although he did not use this term) to answer a question posed by Gromov \cite{gromov} in 1983 regarding knot distortion. The distortion of an embedded circle $\gamma$ in $\mathbb{R}^3$ is defined to be:
\[\delta(\gamma)=\sup_{p,q\in\gamma}\frac{d_\gamma(p,q)}{d_{\mathbb{R}^3}(p,q)},\]
where $d_\gamma$ is arclength along $\gamma$, and $d_{\mathbb{R}^3}$ is Euclidean distance in $\mathbb{R}^3$. Gromov asked whether there exists a uniform upper bound on distortion for all isotopy classes of knots, or at least for torus knots. Specifically, Gromov asked, does every isotopy class of knots have a representative with distortion less than, say, 100? To answer this question, Pardon showed that every knot isotopy class $K$ satisfies:
\[\delta(K):=\min_{ \text{representatives }\gamma \text{ of }K}\delta(\gamma)\geq\frac{1}{160} r(K),\]
where $r(K)$ denotes representativity. In particular, since the representativity of any $(p,q)$ torus knot is $r(T_{p,q})= \text{min}\{p,q\}$ (more to the point and easier to check is $r(T_{p,q})\geq\min\{p,q\}$), so that $\delta(T_{p,q})\to\infty$ as $p,q\to\infty$, Pardon was able to answer Gromov's question in the negative. Current work of Blair, Campisi, Taylor and Tomova \cite{bt} provides a lower bound for $\delta(K)$ in terms of distance and bridge numbers. Our main result implies that their lower bound improves Pardon's lower bound in the case of alternating knots with sufficiently large distance:
\begin{maintheorem}
Every non-trivial, non-split alternating link $L$ has representativity $r(L)=2$.
\end{maintheorem}
As another application, the main theorem gives a new proof of the fact that the only alternating torus links are the 2-braids $T_{2,q}$, again since $r(T_{p,q})=\min\{p,q\}$. As $T_{3,4}$ is almost-alternating, the main theorem does not extend to this class of links. We return to this example in \textsection\ref{S:5}, suggesting a possible approach to the characterization of almost-alternating links with representativity $\geq 3$.
To prove the main theorem, we employ the crossing ball structures introduced by William Menasco \cite{men}. Roughly, the idea is to insert a ball $C_t$ at each crossing of a given diagram $D$ on $S^2$, to perturb $L$ to lie on $(S^2\setminus C)\cup\partial\nu C$, where $C=\bigsqcup C_t$, and then to isotope a given closed surface $F\supset L$ (fixing $L$ and the crossing ball structure $S^2\cup C$) so as to minimize its intersections with $C$ and $S^2$ away from $L$.
We show that whenever $F$ is essential (incompressible and $\partial$-incompressible in the link exterior $S^3\setminus \text{int}(\nu L)$), there exists an isotopy of $F$ which produces a standard tube near some crossing (cf. Figure \ref{Fi:Smoothing}).
\begin{ctlemma}
Given a non-trivial, connected, reduced alternating diagram of a link $L$ and a closed, essential surface $F\supset L$,
there exists an isotopy of $F$ after which some crossing has a standard \textbf{tube}.
\end{ctlemma}
This lemma not only provides a compressing disk for $F$ whose boundary intersects $L$ in two points; it also provides a possible inductive move, in the tradition of \cite{gab,ak}, albeit one still awaiting application.
\textbf{Thank you} to Seungwon Kim for sharing this problem during a visit to Iowa; to Colin Adams for introducing the author to Menasco's techniques (in particular to ``bigon moves''); to Maggy Tomova and Ryan Blair for helpful discussions around early and final versions of this paper, respectively; and to Charlie Frohman for patient and inspired coaching.
\section{Initial setup}\label{S:2}
\subsection{Link diagrams and crossing balls}\label{S:21}
A {link diagram} $D\subset S^2$ is the image of an immersion of one or more circles in $S^2\subset S^3$ in which all self-intersections are double-points at which the two intersecting arcs are transverse in $S^2$ and are labeled with opposite normal directions relative to $S^2$. Thus, a link diagram $D$ can also be seen as a smoothly embedded 4-valent graph in $S^2$ with over-under information at each vertex. An embedding of the underlying link $L$ can be obtained by perturbing the two arcs of $D$ near each crossing point in the indicated normal directions.
A link diagram $D$ is called \textbf{alternating} if, for each edge of $D$ (seen as a 4-valent graph), the crossing points at its two ends are labeled with {opposite} normal directions. A link is called {alternating} if it has an alternating diagram. A link diagram $D$ is called \textbf{reduced} if it lacks \textit{nugatory} crossings, which means that every crossing point is incident to four distinct components of $S^2\setminus D$.
A link $L$ is called \textbf{split} if $S^3\setminus L$ is reducible, i.e. if there is an embedded 2-sphere in $S^3\setminus L$ which does not bound a ball in $S^3\setminus L$.
Let $D$ be a diagram of a link $L$ (in \textsection\ref{S:3}-\ref{S:4}, we will assume further that $D$ is non-trivial, connected, reduced, and alternating) with crossing points $c_t$, $t=1,\hdots, n$. Insert small, mutually disjoint (closed) \textbf{crossing balls} $C_t$, $t=1,\hdots, n$, centered at the respective crossing points $c_t$. Denote $C=\bigsqcup_{t=1}^nC_t$. Construct an embedding of $L$ in $(S^2\setminus C)\cup\partial C$ by perturbing the two arcs in which $D$ intersects each crossing ball $C_t$ in their indicated normal directions from $S^2\cap C_t$ to $\partial C_t$, while fixing $D\cap S^2\setminus \text{int}(C)$.
Call each resulting component of $L\cap S^2$ an \textbf{edge} of $L$---note that $L\cap S^2=L\cap D=D\setminus \text{int}(C)$---and call each component of $L\cap\partial C$ an \textbf{overpass} or an \textbf{underpass} of $L$, according to which side of $S^2$ it lies on. Near each crossing, this looks like Figure \ref{Fi:crossingball} (center).
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{CrossingBallPartsHorizB.pdf}
\caption{A link near a crossing ball (center), with $S^+$ (left) and $S^-$ (right).}
\label{Fi:crossingball}
\end{center}
\end{figure}
\subsection{The regular neighborhood $\nu L$}\label{S:22}
Let $\nu L\subset S^3$ be a {closed} regular neighborhood of the link $L$, viewed as (the total space of) a disk-bundle $\pi:\nu L\to L$ for which the restrictions $\pi|_{\partial C}$, $\pi|_{S^2\setminus \text{int}(C)}$ are also bundle maps onto their images (each with fiber a closed interval).
Thus, each component of $\nu L$ intersects $(S^2\setminus C)\cup\partial C$ in an annular neighborhood of its core; and, for each point $p\in L\cap S^2\cap\partial C$, the boundary of the disk $\pi^{-1}(p)$ is a meridian on $\partial\nu L$ which consists of an arc of $\partial\nu L\cap\partial C\cap S^\pm$ glued to an arc of $\partial\nu L\cap S^\mp\setminus C$ at the two points of $\pi^{-1}(p)\cap\partial\nu L\cap S^+\cap S^-$
Let $B^+$ and $B^-$ be the two components into which $S^2\cup C\cup\nu L$ cuts $S^3$; i.e., $B^\pm$ are the
closures of the components of $S^3\setminus (S^2\cup C\cup\nu L)$.
Let $S^+=\partial B^+$ and $S^-=\partial B^-$. Near each crossing, $S^+$ and $S^-$ appear as in Figure \ref{Fi:crossingball} (left and right).
As a quick exercise, check that $S^+\cup S^-=(S^2\setminus (C\cup\nu L))\cup\partial (C\cup\nu L)$ and $S^+\cap S^-=S^2\setminus \text{int}(C\cup\nu L)$.
Use this setup to extend the terminology of edges, overpasses, and underpasses from $L$ to $\partial\nu L$ as follows.
Call each component of $\pi^{-1}(L\cap\partial C)\cap\partial\nu L\cap S^+$ an \textbf{overpass} of $\partial\nu L$,
call each component of $\pi^{-1}(L\cap\partial C)\cap\partial\nu L\cap S^-$ an \textbf{underpass} of $\partial\nu L$, and
call each component of $\pi^{-1}(L\cap S^2)\cap\partial\nu L$ an \textbf{edge-pair}; each edge-pair is the union of two \textbf{edges} of $\partial\nu L$, one a component of $\pi^{-1}(L\cap S^2)\cap S^+$, the other a component of $\pi^{-1}(L\cap S^2)\cap S^-$.
This terminology gives $\partial\nu L$ the following cell decomposition (cf. Figures \ref{Fi:crossingball}, \ref{Fi:EdgeBlank}).
\begin{itemize}
\item
The 0-cells are the points of $\partial\nu L\cap S^2\cap\partial C
, eight on the boundary of each crossing ball.
\item
There are several types of (closures of) 1-cells:
\begin{itemize}
\item arcs of $\partial\nu L\cap S^+\cap S^-$, two running along each edge-pair \textit{of $\partial\nu L$};
\item arcs of $\partial\nu L\cap\partial C\cap S^+$, two along each overpass of $\partial\nu L$;
\item arcs of $\partial\nu L\cap\partial C\cap S^-$, two along each underpass of $\partial\nu L$; and
\item arcs of $\pi^{-1}(L\cap C\cap S^2)\cap S^\pm$, eight near each crossing ball $C_t$:
\begin{itemize}
\item four in $\partial\nu L\cap\partial C_t$,
\item two in $S^+\cap\partial\nu L\setminus\partial C$, joining the overpass of $\partial\nu L$ at $C_t$ with edges of $\partial\nu L$, and
\item two in $S^-\cap\partial\nu L\setminus\partial C$, joining the underpass of $\partial\nu L$ at $C_t$ with edges of $\partial\nu L$.
\end{itemize}
\end{itemize}
\item The
2-cells' closures are overpasses, underpasses, and edges of $\partial\nu L$; plus components of $\partial\nu L\cap C$.
\end{itemize}
\begin{figure}
\begin{center}
\includegraphics[height=1.4in]{EdgeBlankB.pdf}
\hspace{.25in}
\includegraphics[height=1.7in]{Edge3DI.pdf}
\caption{Two views of an edge of $\partial\nu L\cap S^+$ from an alternating link diagram. }\label{Fi:EdgeBlank}
\end{center}
\end{figure}
\subsection{Essentiality and minimal complexity for the closed surface $F\supset L$}\label{S:23}
So far, the link $L\subset (S^2\setminus C)\cup\partial C$ follows a
link diagram $D\subset S^2$; $\nu L$ is a closed regular neighborhood of $L$ in $S^3$, seen as (the total space of) a disk-bundle $\pi:\nu L\to L$; and $B^\pm$ are the components into which $S^2\cup C\cup\nu L$ cuts $S^3$, with $S^\pm=\partial B^\pm$. Now also
let $F$ be any closed surface in $S^3$ of positive genus that contains $L$. (Recall that a ``closed surface'' is assumed to be compact and connected without boundary).
Fixing $L\subset F$, $S^2$, and $C$, isotope $F$ so that:
{\begin{itemize}
\item $F$ is transverse to $S^+$ and $S^-$;
\item the restriction $\pi|_F$ is a bundle map (so that $F\cap\nu L$ is a regular neighborhood of $L$ in $F$ and each component of $F\cap\partial\nu L$ is the image of a section of $\pi:\nu L\to L$); and
\item $F\cap C\cap\partial\nu L=\varnothing$ (so that $F$ intersects each overpass and underpass of $\partial\nu L$ in precisely two arcs).
\end{itemize}}
Perform such isotopy so as to \textbf{minimize lexicographically} the numbers of components of $F\cap\partial C\setminus\nu L$
and of $F\cap S^+\cap S^-$, i.e. to minimize the \textbf{complexity} $\left(|F\cap\partial C\setminus\nu L|,|F\cap S^+\cap S^-|\right)$, where bars count connected components. Note that the first and last conditions above ensure that $F\cap\partial C\setminus\nu L$ will consist only of simple closed curves, since the endpoints of any arc of $F\cap\partial C\setminus\nu L$ would lie on $\partial C\cap\partial\nu L$.
The surface $F$ is said to be \textbf{compressible} in the link exterior $E=S^3\setminus \text{int}(\nu L)$ if there is a disk $X\subset E$ with $ X\cap F=\partial X$ a simple closed curve (a ``circle'') that does not bound a disk in $F$. The surface $F$ is compressible in $E$ if and only if $r(F,L)=0$. In particular, every split link $L$ has representativity $r(L)=0$.
The surface $F$ is said to be \textbf{$\partial$-compressible} in the link exterior $E$ if there is an arc $\alpha$ on $\partial\nu L$
which is parallel to $F$ through $E$---say, through a disk $X$, with $\beta$ denoting the arc $\partial X\cap F$---but not through $\partial\nu L$. In this case, $\alpha$ is also parallel through a bigon $Y\subset\nu L$ to an arc $\alpha'\subset F\cap\nu L$ which intersects $L$ in one point. Gluing $X$ and $Y$ along $\alpha$ produces a compressing disk $Z$ for $F$ in $S^3$ whose boundary $\partial Z=\alpha'\cup\beta$ intersects $L$ in one point. Thus, if $F$ is $\partial$-compressible in $E$, then $r(F,L)\leq 1$; conversely as well.
In particular, the trivial {knot} $L$ has representativity $r(L)=1$, since any positive genus, closed surface containing $L$ must be compressible or $\partial$-compressible. A closed surface $F\supset L$ is called \textbf{essential} if it is neither compressible nor $\partial$-compressible in the link exterior.
Any non-trivial \textit{knot} $L$ is contained in a positive genus, closed surface $F$ with $r(F,L)=2$, namely the {interpolating surface} for any incompressible Seifert surface for $L$, or more generally for any algebraically essential spanning surface for $L$. (A spanning surface $V$ for $L$ is an embedded surface with boundary $\partial V=L$ in $S^3$; the interpolating surface for a spanning surface $V$ is the boundary of a regular neighborhood of $V$ in the link exterior; and $V$ is called algebraically essential if its interpolating surface is incompressible and $\partial$-incompressible in the link exterior.) Is this also true of non-split \textit{links} $L$? Is $r(L)\geq 2$ \textit{if} and only if $L$ is non-split and non-trivial? This is true for all non-split, non-trivial links with an algebraically essential, \textit{connected} spanning surface, using the interpolating surface. Does every non-split link have such a span?
At least in the alternating case, the answer is yes (in fact, \textit{all} spanning surfaces for non-split alternating links are connected \cite{ak}). Thus, an alternating link $L$ in a reduced alternating diagram $D$ obeys:
\begin{align*}
r(L)=0&\iff~L \text{ is split }\iff~D \text{ is disconnected;}\\
r(L)=1&\iff~L \text{ is the unknot }\iff~D \text{ is the trivial diagram.}\\
\end{align*}
The main theorem states that whenever $L$ is alternating, non-trivial, and non-split, $r(L)=2$.
\subsection{A preliminary consequence of the initial setup}\label{S:24}
The three results in \textsection\ref{S:24} assume the following setup from \textsection\ref{S:21}-\ref{S:23}.
A link $L\subset (S^2\setminus C)\cup\partial C$ follows a non-trivial, connected link diagram $D\subset S^2$ with $n$ crossings, and $\nu L$ is a regular neighborhood of $L$, seen as (the total space) of a disk-bundle $\pi:\nu L\to L$.
Balls $B^\pm$ are the closures of the two components of $S^3\setminus (S^2\cup C\cup\nu L)$, with $S^\pm=\partial B^\pm$ and $S^+\cap S^-=S^2\setminus \text{int}(C\cup\nu L)$.
And a closed essential surface $F\supset L$ has been isotoped, subject to the conditions that $F$ is transverse to $S^+$ and $S^-$, $\pi|_F$ is a bundle map, and $F\cap C\cap\partial\nu L=\varnothing$, so as to minimize its complexity $\left(|F\cap\partial C\setminus\nu L,F\cap S^+\cap S^-|\right)$. This setup implies:
%
\begin{prop}\label{P:Disks}
All components of $F\setminus (S^+\cup S^-\cup\nu L)$ are disks.
\end{prop}
\begin{proof}
The $n+2$ components of $S^3\setminus (S^+\cup S^-\cup\nu L)$---namely $B^+$, $B^-$, and $C_t\setminus \text{int}(\nu L)$, $t=1,\hdots, n$---are all topological 3-balls, and their boundaries are the 2-spheres $S^+$, $S^-$, and $\partial (C_t\setminus\nu L)$. These spheres intersect $F$ transversally, hence in circles. Each such circle bounds a disk in the corresponding ball.
The incompressibility of $F$ implies that each of these circles also bounds a disk in $F$. Minimality then implies that this disk in $F$ must lie entirely in the appropriate ball, as claimed. Specifically, each circle of $F\cap S^\pm$ bounds a disk of $F\cap B^\pm$, and each circle of $F\cap (\partial C_t\setminus \text{int}(\nu L))$ bounds a disk of $F\cap C_t\setminus \text{int}(\nu L)$.
\end{proof}
\section{Technical conveniences}\label{S:broader}
Throughout \textsection\ref{S:broader}, maintain all setup from \textsection\ref{S:21}-\ref{S:23}, but replace the assumption that the complexity of $F$ is minimized with the assumption that all components of $F\setminus (S^+\cup S^-\cup\nu L)$ are disks. Proposition \ref{P:Disks} implies that this setting is more general than the initial setup.
\subsection{Preliminary consequences in the broader setting -- arcs and balls.}
\begin{prop}\label{P:Arcs}
All components of $F\cap\partial\nu L\cap S^\pm$, $F\cap\partial C\cap S^\pm$, and $F\cap S^+\cap S^-$ are arcs.
\end{prop}
%
\begin{proof}
Every component of $\nu L$ contains an overpass and an underpass, or else $L$ would be a split link, and $F$ (assumed to be connected) would be compressible, contrary to assumption. Therefore, each component of $F\cap\partial\nu L$, (the image of) a section of $\pi:\nu L\to L$ by assumption, must intersect $S^+\cap S^-$. Hence, no component of $F\cap\partial\nu L\cap S^\pm$ is a circle; instead, each
must be an arc.
All components of $\partial C\cap S^+$, $\partial C\cap S^-$, and $S^+\cap S^-$ are disks. If $F$ intersected one of these disks in a circle, $\gamma$, then all components of $F\setminus (S^+\cup S^-\cup\nu L)$ are disks, $\gamma$ would bound disks of $F$ in both components of $F\setminus (S^+\cup S^-\cup\nu L)$ whose boundaries contain $\gamma$. This contradicts the assumption that $F$ is connected.
\end{proof}
The conclusion of Proposition \ref{P:Arcs}, implies that the number of components of $F\cap S^+\cap S^-$ equals half the number of points of $F\cap\partial (S^+\cap S^-
$, which will be more convenient to count. Also, the conclusions of Propositions \ref{P:Disks} and \ref{P:Arcs} together imply that $F\setminus \text{int}(\nu L)$ has the following cell decomposition:
\begin{itemize}
\item The 0-cells are the points of $F\cap\partial (S^+\cap S^-)$.
\item There are three types of (closures of) 1-cells: arcs of $F\cap S^+\cap S^-$, $F\cap\partial C\cap S^\pm$, and $F\cap\partial\nu L\cap S^\pm$.
\item The (closures of the) 2-cells are the components of $F\cap B^+$, $F\cap B^-$, and $F\cap C\setminus\nu L$.
\end{itemize}
\begin{prop}\label{P:Balls}
All components of $B^\pm\setminus F$ are 3-balls
\end{prop}
\begin{proof}
Let $A$ be a component of $B^\pm\setminus F$, and let $X$ be a component of $\partial A$. Since $F$ is connected, $X$ must contain (at least) one component $Y$ of $S^\pm\setminus F$.
The boundary of $Y$ is a union of circles of $F\cap S^\pm$, each of which bounds a disk of $F\cap B^\pm$, by assumption. The union of these disks and $Y$ is a closed surface and therefore equals $X$; computing euler characteristic reveals that $X$ is a 2-sphere in $B^\pm$ and therefore bounds a 3-ball in $B^\pm$. This 3-ball must be $A$; otherwise, since each component of $\partial A$ is a 2-sphere which bounds a 3-ball in $B^\pm$, gluing each of these 3-balls to $A$ along $\partial A$ would yield a closed 3-manifold (compact and connected without boundary) contained in $B^\pm$.
\end{proof}
\subsection{Bigon moves}\label{S:bigon}
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{Bigon3DB.pdf}
\caption{Bigon moves will often prove useful, even when they increase the complexity of $F$.}
\label{Fi:bigon}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{MoveC.pdf}
\caption{A bigon move pushes an arc $\beta\subset F\cap B^\pm$ past a parallel arc $\alpha\subset S^\pm\setminus\pi^{-1}(C)$, provided $\alpha$ is not parallel to $F$ through $S^\pm\setminus C$ and
$|\alpha\cap S^+\cap S^-|=1$.}
\label{Fi:move}
\end{center}
\end{figure}
Several proofs in \textsection\ref{S:3}-\ref{S:4} will use an isotopy move which pushes an arc $\beta\subset F\cap B^\pm$ past a parallel arc $\alpha\subset S^\pm\setminus F$ through a disk $Z$ with $Z\cap F=\beta\subset\partial Z=\alpha\cup\beta$. One type of this ``bigon move'' is illustrated in Figure \ref{Fi:bigon}; all three types are diagrammed in Figure \ref{Fi:move}. More precisely, a bigon move follows an arc $\alpha\subset S^\pm$ that:
\begin{itemize}
\item intersects $F$ precisely on its endpoints, which lie on the same circle $\gamma$ of $F\cap S^\pm$;
\item is not parallel in $S^\pm\setminus C$ to $F$;
\item is disjoint from $\pi^{-1}(C)$, i.e. from crossing balls and over/underpasses; and
\item intersects $S^+\cap S^-$ in exactly one component.
\end{itemize}
Think of $\alpha$, which initially is not ``part of the diagram,'' as a marker which joins two points that lie on the same circle $\gamma$ of $F\cap S^\pm$, but ``not obviously'' so, locally.
Because the circle
$\gamma$ bounds a disk $Y\subset F\cap B^\pm$, there is an arc $\beta\subset Y$ with the same endpoints as $\alpha$. The circle $\alpha\cup\beta$ lies on the boundary of some component of $B^\pm\setminus F$, a ball,
and therefore bounds a disk $Z\subset B^\pm$ whose interior is disjoint from $F$. The disk $Z$ is a bigon in the sense that $\partial Z=\alpha\cup\beta$.
\begin{bigonmove}\label{BigonMove}
Given an arc $\alpha\subset S^\pm$ that satisfies the four conditions above, it is possible to isotope $F$ near a parallel arc $\beta\subset F\cap B^\pm$ through a bigon $Z$ past $\alpha$.
\end{bigonmove}
All bigon moves take place away from the crossing balls, and thus preserve both the number of components of $F\cap\partial C\setminus\nu L$ and the fact that each of these components bounds a disk of $F\cap C\setminus \nu L$.
Not all bigon moves, however, preserve the fact that all components of $F\setminus (S^+\cup S^-\cup \nu L)$ are disks. At least:
\begin{lemma}\label{L:Move}
Performing a bigon move preserves the fact that all components of $F\setminus (S^+\cup S^-\cup \nu L)$ are disks whenever $\alpha\cap\partial\nu L\neq\varnothing$ and whenever the complexity of $F$ is minimized.
\end{lemma}
\begin{proof}
As noted, any bigon moves preserves the fact that all components of $F\cap C\setminus\nu L$ are disks.
In order for a bigon move to upset the fact that all components of $F\cap B^+$ and $F\cap B^-$ are disks, a necessary condition is that the endpoints of $\alpha$ lie on the same circle of $F\cap S^\pm$ and also on the same circle of $F\cap S^\mp$. For this to be the case, both endpoints of $\alpha$ must lie in $S^+\cap S^-$, as must all of $\alpha$, since $\alpha\cap S^+\cap S^-$ is connected by assumption. Thus, the condition $\alpha\cap\partial\nu L\neq\varnothing$ in the statement of the Lemma suffices as claimed.
Suppose instead that, with the complexity of $F$ minimized, a bigon move follows an arc $\alpha\subset S^+\cap S^-$ whose endpoints lie both on the same circle $\gamma=\partial Y$ of $F\cap S^\pm$ \textit{and} on the same circle $\gamma'=\partial Y'$ of $F\cap S^\mp$. Letting $Y$, $Y'$ denote the disks of $F\cap B^+$, $F\cap B^-$ respectively bounded by $\gamma$, $\gamma'$, there are arcs $\beta\subset Y$, $\beta'\subset Y'$ with the same endpoints as $\alpha$.
Further, $\alpha\cup\beta$, $\alpha\cup\beta'$ respectively bound disks $Z\subset B^\pm$, $Z'\subset B^\mp$ whose interiors are disjoint from $F$.
Gluing $Z$ and $Z'$ along $\alpha$ produces a disk $Z\cup Z'$ with boundary $\beta\cup\beta'\subset F\setminus\nu L$ whose interior is disjoint from $F$; since $F$ is incompressible in the link exterior, $\beta\cup\beta'$ must bound a disk $X\subset F\setminus\text{int}(\nu L)$.
Since $L$ is non-split, the 2-sphere $Z\cup Z'\cup X\subset S^3\setminus\nu L$ bounds a 3-ball $W\subset S^3\setminus\nu L
, through which $X$ is parallel to $Z\cup Z'$.
Since the complexity of $F$ is minimized, the disk $X$, like $Z\cup Z'$, must be disjoint from $C$ and must intersect $S^+\cap S^-$ in a single arc, $\delta$. From $X\cap C=\varnothing$, it follows that $\partial W\cap C=\varnothing$. This and the fact that $W\cap\nu L=\varnothing$ imply that $W$ is disjoint from $C\cup \nu L$, and in particular from $\partial(S^+\cap S^-)$. Therefore, contrary to assumption, $W$ intersects $S^+\cap S^-$ in a single disk through which $\alpha$ is parallel to the arc $\delta\subset F$:
\[
\pushQED{\qed}
\partial(W\cap S^+\cap S^-)=\left(\partial W\cap (S^+\cap S^-)\right)\cup\left(W\cap\partial(S^+\cap S^-)\right)=\alpha\cup \delta.
\qedhere\]
\end{proof}
Note that a bigon move fixes the number of components of $F\cap S^+\cap S^-$---which equals half the number of points of $F\cap\partial(S^+\cap S^-)$---if and only if $\alpha\cap\partial\nu L=\varnothing$; otherwise, a bigon move increases this number, and thus the complexity of $F$.
In particular:
\begin{lemma}\label{L:MoveCase}
If sequence of bigon moves begins with the complexity of $F$ minimized, and if all bigon moves in this sequence along arcs disjoint from $\partial\nu L$ precede all other bigon moves in this sequence, then this sequence of bigon moves preserves the fact that all components of $F\setminus(S^+\cup S^-\cup\nu L)$ are disks.
\end{lemma}
\begin{proof}
Because bigon moves along arcs disjoint from $\partial\nu L$ fix the complexity of $F$, all such moves in this sequence are performed while the complexity of $F$ is still minimized. Satisfying the second sufficient condition from Lemma \ref{L:Move}, these bigon moves preserve the fact that all components of $F\setminus(S^+\cup S^-\cup\nu L)$ are disks. All remaining bigon moves follow arcs that intersect $\partial\nu L$, meeting the first sufficient condition from Lemma \ref{L:Move}. Therefore, these moves too preserve the fact that all components of $F\setminus(S^+\cup S^-\cup\nu L)$ are disks.
\end{proof}
\subsection{A key lemma}\label{S:lemma}
Several proofs in \textsection\ref{S:3}-\ref{S:4} will use the following lemma.
\begin{lemma}\label{L:inessential}
If $\alpha\subset\partial\nu L\cap S^\pm$ is an arc whose endpoints $\partial\alpha=\alpha\cap F$ lie on distinct circles of $F\cap\partial\nu L$, then these endpoints also lie on distinct circles of $F\cap S^\pm$.
\end{lemma}
\begin{proof}
Let $\alpha\subset\partial\nu L\cap S^\pm$ be an arc whose endpoints $\partial\alpha=\alpha\cap F$ lie the same circle $\gamma$ of $F\cap S^\pm$. We claim that these endpoints must also lie on the same circle of $F\cap\partial\nu L$.
Since $\gamma$ bounds a disk $Y\subset F\cap B^\pm$, there is an arc $\beta\subset Y$ with the same endpoints as $\alpha$. The circle $\alpha\cup\beta$ lies on the boundary of some component of $B^\pm\setminus F$, a ball, and therefore bounds a disk $Z\subset B^\pm$ whose interior is disjoint from $F$.
The arc $\alpha\subset\partial\nu L$ is parallel in the link exterior through $Z$ to $F$; $\partial$-incompressibility implies that $\alpha$ must also be parallel in $\partial\nu L$ to $F$, and in particular that the endpoints of $\alpha$ must lie on the same component of $F\cap\partial\nu L$, as claimed.
\end{proof}
The following special case is particularly noteworthy:
\begin{lemma}\label{L:inessentialCase}
The two arcs of $F\cap\partial\nu L$ traversing each over/underpass lie on distinct circles of $F\cap S^\pm$.
\end{lemma}
\begin{proof}
This follows immediately from Lemma \ref{L:inessential}.
\end{proof}
\subsection{Crossing tubes}\label{S:ct}
Say that $F$ has a standard \textbf{tube} near a crossing ball $C_t$ if there are two arcs $\alpha_1,\alpha_2\subset F\cap S^+\cap S^-$ such that (1) for $r=1,2$, there is an isotopy of $(\alpha_r,\partial\alpha_r)$ through $(S^+\cap S^-\setminus F, S^+\cap S^-\cap\partial\nu L)$ to $(\partial C_t,S^+\cap S^-\cap\partial\nu L\cap\partial C_t)$---i.e., for $r=1,2$, $\alpha_r$ is parallel through $S^+\cap S^-$ to $C_t$, allowing the endpoints to slide along $S^+\cap S^-\cap\partial\nu L$---and (2) the endpoints of $\alpha_1$, $\alpha_2$ are also endpoints of the four arcs of $F\cap\partial\nu L\cap S^\pm$ that traverse the overpass and underpass at $C_t$---i.e. these endpoints are (among) the points of $F\cap\partial\nu L\cap S^+\cap S^-$ closest to $C_t$ in each of the four directions along $\partial\nu L$, in the sense of the disk-bundle $\pi:\nu L\to L$.
\begin{figure}
\begin{center}
\includegraphics[height=1.75in]{SmoothingB.pdf}
\hspace{.15in}
\raisebox{.875in}{$\begin{matrix}
\raisebox{-.075in}{\includegraphics[width=.55in]{SmoothingKeyNuL.pdf}}:&\partial\nu L\\~\\
\includegraphics[width=.5in]{SmoothingKeyFNuLSPlus.pdf}:&F\cap\partial\nu L\cap S^+\\~\\
\includegraphics[width=.5in]{SmoothingKeyFSPlusSMinus.pdf}:&F\cap S^+\cap S^-\\~\\
\includegraphics[width=.5in]{SmoothingKeyFNuLSMinus.pdf}:&F\cap\partial\nu L\cap S^-\\
\end{matrix}$}
\caption{A \textbf{tube} near a crossing ball $C_t$ features two arcs $\alpha_1,\alpha_2\subset F\cap S^+\cap S^-$ parallel through $S^+\cap S^-\setminus F$ to $C_t$ whose endpoints are (among) the points of $F\cap\partial\nu L\cap S^+\cap S^-$ closest to $C_t$ in each direction along $\partial\nu L$. In a minimal crossing link diagram, such a tube gives a compressing disk $Z$ for $F$ in $S^3$ with $|\partial Z\cap L|=2$ (cf. Lemma \ref{L:sufficient}), implying that $r(F,L)\leq 2$. Compressing $F$ along $Z$ changes $\nu L$, $F$ as shown.}
\label{Fi:Smoothing}
\end{center}
\end{figure}
Up to symmetry, this appears as in Figure \ref{Fi:Smoothing}---the arcs $\alpha_1$, $\alpha_2$ must be in opposite quadrants relative to $C_t$, not adjacent ones. The reason for this is that the only other possibility (cf. Figure \ref{Fi:TwoInnermosts}, second from left) contradicts the essentiality of $F$ in the link exterior, using Lemma \ref{L:inessential}, specifically Lemma \ref{L:inessentialCase}.
A crossing tube sets up a surgery move on $F$, $L$ as follows.
Each of the two arcs $\alpha_1$, $\alpha_2$ associated with a crossing tube near $C_t$ has one endpoint on an edge-pair incident to the overpass at $C_t$ and the other on an edge-pair incident to the underpass at $C_t$. The two endpoints on edge-pairs incident to the overpass at $C_t$ can be joined by an arc $\beta_1\subset\partial\nu L\cap (S^-\cup C_t)\setminus F$; likewise, the two endpoints on edge-pairs incident to the underpass at $C_t$ can be joined by an arc $\beta_2\subset\partial\nu L\cap (S^+\cup C_t)\setminus F$.
The circle $\alpha_1\cup\beta_1\cup\alpha_2\cup\beta_2$ bounds a disk $X$ whose interior lies in $S^3\setminus (F\cup\nu L)$. Both $\beta_1$, $\beta_2$ are parallel through disks $Y_1,Y_2\subset\nu L$ to arcs $\beta_1',\beta_2'\subset F$ each of which intersects $L$ in a single point. Thus, the disk $Z:=X\cup Y_1\cup Y_2$ satisfies $Z\cap F=\partial Z=\alpha_1\cup\beta_1'\cup\alpha_2\cup\beta_2'$ with $|\partial Z\cap L|=2$. Moreover, the arcs $\beta_1'$, $\beta_2'$ can be isotoped so that the two points (total) in which they intersect $L$ are the endpoints of the vertical crossing arc in $C_t$, one on the overpass and the other on the underpass.
The surgery move associated to the crossing tube near $C_t$ (cf. Figure \ref{Fi:Smoothing}) consists of (1) cutting $F$ along $\partial Z$, while cutting $L$ at the two points of $\partial Z\cap L$, and (2) gluing in two parallel copies of the disk $Z$, while joining each pair of endpoints of $L$ on the boundary of the glued-in copy of the disk $Z$ with an arc in that disk. This surgery move is a compression of $F$ in $S^3$ unless $\partial Z$ bounds a disk in $F$, in which case the surgery move yields a surface with two components, one of them a sphere. The effect of the surgery move on $L$ is the same as one of two possible ``smoothings'' near $C_t$, in the traditional sense from skein relations. After the surgery move, the resulting link is again embedded in the resulting surface.
\begin{lemma}\label{L:sufficient}
Given a crossing tube in a minimal crossing diagram of a non-split link, the associated surgery move is
a compression---the boundary of the surgery disk $Z$ does not bound a disk in $F$.
\end{lemma}
In particular, a crossing tube in a reduced alternating link diagram contains a ``genuine'' compressing disk $Z$ for $F$ in $S^3$ with $|\partial Z\cap L|=2$. This lemma will round off the proof of the main theorem in \textsection\ref{S:4}.
\begin{proof}
Let $C_t$ be the crossing ball with the tube in question. Construct the disk $Z$ as before, and suppose for contradiction that $\partial Z$ bounds a disk $Y$ in $F$.
Since $|\partial Y\cap L|=|\partial Z\cap L|=2$ and $L$ is non-split, $Y\cap L$ consists of a single arc, call it $\delta$. From $(Y\cup Z)\cap L=\delta$ it follows that the two 2-spheres on the boundary of a thin regular neighborhood of $Y\cup Z$ intersect $L$ in a total of two points; hence, one of these 2-spheres is disjoint from $L$. Since $L$ is non-split, this implies that the disks $Y$ and $Z$ are parallel through a ball $W$ in $S^3$ \textit{whose interior is disjoint from $L$}.
The arc $\delta=Y\cap L$ is parallel through the ball $W$ to any arc $\delta'$ in $Z$ that joins the overpass and underpass of $L$ at $C_t$. Taking $\delta'$ to be the vertical arc in $C_t$ that joins the overpass and underpass of $L$, isotope the arc $\delta\subset L$ to $\delta'$, while fixing the rest of $L$. This isotopy eliminates all crossings incident to $\delta$, including the one at $C_t$, without creating any new ones. We assumed this was impossible.
\end{proof}
\subsection{Height}\label{S:height}
In the broader setting of \textsection\ref{S:broader}, construct graphs $G^\pm$ (they will be trees) whose vertices correspond to the components of $S^\pm$ cut along $F$ and whose edges correspond to the components of $F\cap S^\pm$, such that
the edge corresponding to each circle $\gamma\subset F\cap S^\pm$ joins the vertices corresponding to the two components of $S^\pm$ cut along $F$ whose boundaries contain $\gamma$.
The two ``sides of $\gamma$ in $S^\pm$'' refer to the two components of $S^\pm$ cut along $\gamma$ and correspond to the two components of the graph $G^\pm_\gamma$ obtained from $G^\pm$ by deleting the interior of the edge associated with $\gamma$. Define the \textbf{height} of $\gamma$ on each side to be the \textit{maximum} edge-length among all (simple) paths in the corresponding component of $G^\pm_\gamma$ that start at the (appropriate endpoint of the) edge associated with $\gamma$.
Thus, innermost circles of $F\cap S^\pm$ have height 0 (to one side), circles which enclose (to one side) innermost circles and no others have height 1, and so on. Figure \ref{Fi:T34} shows an example---a torus containing the knot $T_{3,4}$ in an almost-alternating diagram, with the associated graph $G^+$.
\begin{prop}\label{P:Height1Exists}
Either all circles of $F\cap S^\pm$ have height 0 (to one side), or $F\cap S^\pm$ contains a circle with height 1 (to one side).
\end{prop}
\begin{proof}
Let $\gamma$ be a circle of $F\cap S^\pm$ with height $k\geq 2$ (to one side)---if no such $\gamma$ exists, there is nothing to prove. Consider a maximal edge-length path in $G^\pm_{\gamma}$ that starts at the appropriate end of the edge associated with $\gamma$. This path traverses $k$ edges,
the $r^ \text{th}$ of which corresponds to a circle of $F\cap S^\pm$ with height $k-r$.
\end{proof}
\section{Consequences of minimality and essentiality}\label{S:3}
To review the setup and preliminary results, a link $L\subset (S^2\setminus C)\cup\partial C$ follows a
link diagram $D\subset S^2$ with $n$ crossings; $\nu L$ is a regular neighborhood of $L$, seen as (the total space) of a disk-bundle $\pi:\nu L\to L$; balls $B^\pm$ are the closures of the two components of $S^3\setminus (S^2\cup C\cup\nu L)$, with $S^\pm=\partial B^\pm$ and $S^+\cap S^-=S^2\setminus \text{int}(C\cup\nu L)$; and a closed surface $F$, which contains $L$ and is \textbf{essential} (i.e. incompressible and $\partial$-incompressible in the link exterior, i.e. $r(F,L)\geq 2$), has been isotoped such that:
\begin{itemize}
\item The restriction $\pi|_F$ is a bundle map;
\item $F$ is transverse to $S^+$, $S^-$;
\item $F\cap C\cap\partial\nu L=\varnothing$;
\end{itemize}
and, subject to these conditions, the \textbf{complexity} of $F$ is \textbf{minimized}, specifically:
\begin{itemize}
\item The numbers of components of $F\cap\partial C\setminus\nu L$ and of $F\cap S^+\cap S^-$ have lexicographically been minimized.
\end{itemize}
With this initial setup, \textsection\ref{S:24} showed that all components of $F\cap B^+$, $F\cap B^-$ and $F\cap C$ are disks; all components of $F\cap\partial\nu L\cap S^\pm$, $F\cap\partial C\cap S^\pm$, and $F\cap S^+\cap S^-$ are arcs; and all components of $B^+\setminus F$ and $B^-\setminus F$ are balls. Actually, the last two statements follow from the first one, using the essentiality of $F$. In the somewhat more general setting where this first condition replaces the assumption of minimal complexity, \textsection\ref{S:bigon}-\ref{S:height} established such technical conveniences as bigon moves, crossing tubes, and height.
Section \ref{S:3} delimits which local configurations are consistent with the initial setup, where $F$ is essential and its complexity is minimized. Most results address the arcs of $F\cap\partial\nu L\cap S^\pm$, $F\cap\partial C\cap S^\pm$, and $F\cap S^+\cap S^-$, and many extend to the case where $F$ is incompressible but $\partial$-compressible in the link exterior. (One way to extend the proofs, roughly, is to isotope $F$ so as to push any $\partial$-compressing disks into $\nu L$, and then to slide these disks along the link away from the local area under consideration.)
Several of the proofs require disrupting minimality by creating new components of $F\cap S^+\cap S^-$, usually through a sequence of bigon moves. There are several valid reasons to do this. In the proof of Lemma \ref{L:EdgeCrossingLoop}, a \textit{temporary} increase in complexity enables the removal of a component of $F\cap\partial C\setminus\nu L$, lessening the complexity of $F$, a contradiction. One case in the proof of Lemma \ref{L:TwoInnermosts} increases complexity in order to reveal a $\partial$-compressing disk, thus contradicting essentiality. The other case in the proof of Lemma \ref{L:TwoInnermosts}---and the proof of the crossing tube lemma in \textsection\ref{S:4}---increase complexity in order to procure a crossing tube. In such cases, Lemma \ref{L:Move} will confirm that the bigon moves, while disrupting minimality, preserve at least the fact that all components of $F\cap B^+$, $F\cap B^-$ and $F\cap C$ are disks, and thus the more general setting of \textsection\ref{S:bigon}-\ref{S:height}. In particular, this will validate further bigon moves and applications of Lemma \ref{L:inessential}.
\subsection{Local possibilities, regardless of alternatingness}
\label{S:31}%
Assume throughout \textsection\ref{S:31} that $D$ is a diagram of a non-trivial, non-split link $L$, and that $L$ is contained in a closed surface $F\subset S^3$ (compact and connected without boundary). Establish all setup from \textsection\ref{S:21}-\ref{S:23}. Assume in particular that $F$ is essential and its complexity $\left(|F\cap\partial C\setminus\nu L|,|F\cap S^+\cap S^-|\right)$ has been minimized.
\begin{prop}\label{P:EdgeLoopB}
No arc of $F\cap\partial\nu L\cap S^\pm $ has both endpoints on the same component of $\partial\nu L\cap S^+\cap S^-$.
\end{prop}
\begin{prop}\label{P:CrossingLoopAA}
No arc of $F\cap\partial C\cap S^\pm$ has both endpoints on the same component of $\partial C\cap S^+\cap S^-$.
\end{prop}
\begin{figure}
\begin{center}
\includegraphics[height=1.75in]{EdgeLoopB.pdf}
\hspace{.5in}
\includegraphics[height=1.75in]{CrossingLoopAA.pdf}
\caption{No arc of $F\cap\partial\nu L\cap S^\pm$ is parallel in $\partial\nu L$ to $S^+\cap S^-$ (left), and no arc of $F\cap\partial C\cap S^\pm$ is parallel in $\partial C$ to $S^+\cap S^-$ (right).}
\label{Fi:EdgeLoop}
\end{center}
\end{figure}
See Figure \ref{Fi:EdgeLoop}. Recall that $F\cap\partial\nu L\cap C=\varnothing$ by assumption.
%
\begin{proof}[Proof of Propositions \ref{P:EdgeLoopB}, \ref{P:CrossingLoopAA}]
Any arc $\alpha_1$ of $F\cap\partial\nu L\cap S^\pm$ with endpoints on the same component of $\partial\nu L\cap S^+\cap S^-$ must be parallel through a disk $X\subset\partial\nu L\cap S^\pm$ to an arc $\beta_1\subset\partial\nu L\cap S^+\cap S^-$; but then, isotoping $F$ near $\alpha_1$ through $X$ past $\beta_1$ (Figure \ref{Fi:EdgeLoop}, left) would reduce the complexity of $F$, contrary to assumption.
Likewise, any arc $\alpha_2$ of $F\cap\partial C\cap S^\pm$ with endpoints on the same component of $\partial C\cap S^+\cap S^-$ must be parallel through a disk $Y\subset\partial C\cap S^\pm$ to an arc $\beta_2\subset\partial C\cap S^+\cap S^-$; but then, isotoping $F$ near $\alpha_2$ through $Y$ past $\beta_2$ (Figure \ref{Fi:EdgeLoop}, right) would reduce the complexity of $F$, contrary to assumption.
\end{proof}
%
\begin{lemma}\label{L:saddle}
Every component of $F\cap C\setminus\nu L$ is a disk whose boundary consists of four arcs, alternately on $S^+\cap\partial C$ and $S^-\cap\partial C$, none of which is parallel in $\partial C\setminus\nu L$ to $\partial C\cap S^+\cap S^-$.
\end{lemma}
%
\begin{figure}
\begin{center}
\includegraphics[height=1.75in]{CrossingSaddleC.pdf}
\caption{An arbitrary crossing ball: (left) disjoint from $F$, (center) intersecting $F$ in a single component, (right) intersecting $F$ in at least two components
(cf. Lemma \ref{L:saddle}). }
\label{Fi:CrossingSaddle}
\end{center}
\end{figure}
That is, each component of $F\cap C\setminus\nu L$ looks like a saddle, as in Figure \ref{Fi:CrossingSaddle} (center), and each crossing ball looks like one of the pictures in Figure \ref{Fi:CrossingSaddle}, depending on the number of components in which it intersects $F$.
\begin{proof}
Lemma \ref{L:saddle} is an immediate consequence of Propositions \ref{P:Disks}, \ref{P:Arcs}, and \ref{P:CrossingLoopAA}.
\end{proof}
%
\begin{prop}\label{P:EdgeLoopE}
No arc of $F\cap S^+\cap S^-$ has both endpoints on the same component of $\partial\nu L\cap S^+\cap S^-$.
\end{prop}
\begin{prop}\label{P:CrossingLoopE}
No arc of $F\cap S^+\cap S^-$ has both endpoints on the same crossing ball.
\end{prop}
%
\begin{figure}
\begin{center}
\includegraphics[height=1.75in]{EdgeLoop2.pdf}
\hspace{.5in}
\includegraphics[height=1.75in]{CrossingLoopE.pdf}
\caption{No arc of $F\cap S^+\cap S^-$ is parallel in $S^+\cap S^-$ to $\partial\nu L$ (left) or to $\partial C$ (right).}
\label{Fi:CrossingLoop}
\end{center}
\end{figure}
\begin{proof}[Proof of Propositions \ref{P:EdgeLoopE}, \ref{P:CrossingLoopE}]
If both endpoints of some arc of $F\cap S^+\cap S^-$ were on the same component of $\partial\nu L\cap S^+\cap S^-$ then, applying Proposition \ref{P:EdgeLoopB}, an outermost such arc in $S^+\cap S^-$ would appear as left in Figure \ref{Fi:CrossingLoop}, contradicting the assumed $\partial$-incompressibility of $F$ in the link exterior, e.g. by Lemma \ref{L:inessential}.
Suppose instead that $\alpha_0$ is an arc of $F\cap S^+\cap S^-$ with both endpoints on the same crossing ball $C_t$, and assume that $\alpha_0$ is outermost in $S^+\cap S^-$, i.e. parallel through $S^+\cap S^-\setminus F$ to $C_t$, as in Figure \ref{Fi:CrossingLoop}. Push $F$ near $\alpha_0$ through $S^+\cap S^-$ past $\partial C_t$. This attaches two saddle-shaped disks in the interior of $C_t$, lessening the complexity of $F$, contrary to assumption.
\end{proof}
\begin{lemma}\label{L:EdgeCrossingLoop}
No arc $\alpha_0\subset F\cap S^+\cap S^-$ has
one endpoint on a crossing ball and the other on an incident edge of $\partial\nu L$.
\end{lemma}
\begin{proof}
Suppose that $\alpha_0$ is an arc of $F\cap S^+\cap S^-$ with one endpoint on a crossing ball $C_t$ and the other on an incident edge of $\partial\nu L$.
Assume that $\alpha_0$ is outermost in $S^+\cap S^-$, i.e. parallel through $S^+\cap S^-\setminus F$ to $\nu L\cup C$. Consider the circle of $F\cap\partial\nu L$ that contains an endpoint of $\alpha_0$. Moving along this circle from that endpoint toward $C_t$, there is at most one more point on $S^+\cap S^-$, by Proposition \ref{P:EdgeLoopB} and the assumption that $\alpha_0$ is outermost. There are thus two cases up to symmetry (cf. Figure \ref{Fi:EdgeCrossingLoop}).
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{EdgeCrossingLoopE1.pdf}
\includegraphics[width=6.5in]{EdgeCrossingLoopE2.pdf}
\caption{No arc $\alpha_0$ of $F\cap S^+\cap S^-$ has one endpoint on a crossing ball and the other on an incident edge of $\partial\nu L$.}
\label{Fi:EdgeCrossingLoop}
\end{center}
\end{figure}
In either case, begin with a sequence of (up to) three bigon moves,
through the arcs labeled $\alpha_1$, $\alpha_2$, $\alpha_3$ in Figure \ref{Fi:EdgeCrossingLoop}, in that order. Any of the arcs $\alpha_r$ can be parallel in $S^\pm\setminus C$ to $F$; in this case, omit the bigon move along $\alpha_r$.
In all cases, this sequence of bigon moves fits the hypotheses of Lemma \ref{L:MoveCase} and thus preserves the fact that all components of $F\setminus(S^+\cup S^-\cup\nu L)$ are disks.
Now an arc
$\beta\subset\partial\nu L\cap F$ is parallel through a disk $Y\subset\partial\nu L$ with $Y\cap F=\beta$ to a second arc $\beta'\subset\partial\nu L\cap S^\pm$ with $\beta'\cap F=\partial\beta'$.
This arc $\beta'$ is parallel to an arc $\beta''\subset F\cap B^\pm$ through a disk $Z\subset B^\pm$ with $Z\cap F=\beta''$.
Further, $\beta''$ is parallel to $\beta$ through a disk $X\subset F$, which contains an entire disk of $F\cap C\setminus\nu L$.
Finally, the 2-sphere $X\cup Y\cup Z$ bounds a ball $W$ in the link exterior.
Isotope $(X,\beta)$ through $(W,Y)$ to $(Z,\beta')$, while fixing $\beta''=\partial X\cap\partial Z$. This removes the disk of $X\cap C$ and thus a component of $F\cap C\cap\partial\setminus\nu L$. Since bigon moves always fix $|F\cap\partial C\setminus\nu L|$, this contradicts the initial assumption that the complexity of $F$ was minimized.
\end{proof}
\subsection{Consequences of alternatingness.}
\label{S:32}
Maintain all setup from \textsection\ref{S:31}, with the additional assumption that $D$ is alternating. That is, assume throughout \textsection\ref{S:32} that $D$ is a non-trivial, connected, alternating diagram of a link $L$, and that $L$ is contained in a closed essential surface $F\subset S^3$ (compact and connected without boundary) whose complexity $\left(|F\cap\partial C\setminus\nu L|,|F\cap S^+\cap S^-|\right)$ has been minimized.
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{EdgeTypesC.pdf}
\caption{The types of edge-pairs of $\partial\nu L$ when $D$ is alternating (cf. Lemma \ref{L:EdgeTypes}).}
\label{Fi:EdgeTypes}
\end{center}
\end{figure}
\begin{lemma}\label{L:EdgeTypes}
Every edge-pair of $\partial\nu L$ appears as in Figure \ref{Fi:EdgeTypes}.
\end{lemma}
\begin{proof}
This follows immediately from Propositions \ref{P:Arcs} and \ref{P:EdgeLoopB} and the alternatingness of $D$.
\end{proof}
\newpage
\begin{prop}\label{P:InnermostCrossingBall}
If $\gamma$ is an innermost circle of $F\cap S^\pm$, then $\gamma\cap\partial C=\varnothing$.
\end{prop}
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{InnermostCrossingBallB.pdf}
\caption{Every innermost circle $\gamma_0\subset F\cap S^\pm$ is disjoint from $\partial C$ (cf. Proposition \ref{P:InnermostCrossingBall}).}\label{Fi:InnermostCrossingBall}
\end{center}
\end{figure}
\begin{proof}
Let $\gamma_0$ be an innermost circle of $F\cap S^+$ (\textsc{wlog}), and suppose $\gamma_0\cap\partial C\neq\varnothing$. Then the disk $X\subset S^+\setminus F$ with $\partial X=\gamma_0$ intersects $\partial C$; let $Y$ be a component of $X\cap\partial C$. There are now three cases, using Lemma \ref{L:saddle} (cf. Figure \ref{Fi:InnermostCrossingBall}): in one case (left), $\partial Y$ contains two arcs of $\gamma_0\cap\partial C$. Otherwise $\partial Y$ contains an arc on $\partial\nu L$, along either the underpass (center) or the overpass (right) at $C_t$.
If $\partial Y$ contains more than one arc of $\gamma_0\cap C_t$ (left in Figure \ref{Fi:InnermostCrossingBall}), then a bigon move yields an arc of $F\cap S^+\cap S^-$ with both endpoints on $C_t$. This contradicts the minimality of $F\cap\partial C\setminus\nu L$ (recall Figure \ref{Fi:CrossingLoop}, right, and the proof of Proposition \ref{P:CrossingLoopE}).
In both remaining cases (center and right), a bigon move yields an arc of $F\cap S^+\cap S^-$ with one endpoint on $C_t$ and the other on an incident edge of $\partial\nu L$. Crucial to the initial bigon moves in Figure \ref{Fi:InnermostCrossingBall}, center and right, is that every edge of $\partial\nu L$ contains endpoints of $F\cap S^+\cap S^-$, by alternatingness. Next, perform the sequence of bigon moves from the top of Figure \ref{Fi:EdgeCrossingLoop} (omitting any trivial ones as usual), to set up the final isotopy move from that sequence. This final move is valid, since the preceding sequence of bigon moves meets the conditions of Lemma \ref{L:MoveCase}, and it removes a component of $F\cap C\cap\partial\setminus\nu L$. Since bigon moves always fix $|F\cap\partial C\setminus\nu L|$, this contradicts the initial assumption that the complexity of $F$ was minimized.
\end{proof}
\begin{prop}\label{P:InnermostPass}
If $\gamma$ is an innermost circle of $F\cap S^\pm$, so that $\gamma$ bounds a disk $X\subset S^\pm$, then at least one component of $\gamma\cap\partial\nu L$ traverses an over/underpass.
\end{prop}
\begin{proof}
Lemma \ref{L:EdgeTypes}
implies that the only arcs of $F\cap\partial\nu L$ with endpoints on the same component of $\nu L\cap S^2\setminus C$ look like the arcs of this type in Figure \ref{Fi:EdgeTypes}, up to reflection---there is one far left in Figure \ref{Fi:EdgeTypes}, one second from right, and two far right. If such an arc lies on an innermost circle $\gamma$ of $F\cap S^+$ (\textsc{wlog}), then, since $\gamma$ is innermost and $\gamma\cap\partial C=\varnothing$ by Proposition \ref{P:InnermostCrossingBall}, $\gamma$ must traverse the overpass at the crossing where the edge of $\partial\nu L$ containing this arc of $\gamma$ meets an underpass. This is evident in Figure \ref{Fi:EdgeTypes}, using Lemma \ref{L:inessential}.
\end{proof}
\begin{lemma}\label{L:TwoInnermosts}
If both circles of $F\cap S^\pm$ traversing a given over/underpass have height 0 (to one side), i.e. are innermost, then $F$ can be isotoped to have a standard tube near that crossing.
\end{lemma}
\begin{figure}
\begin{center}
\includegraphics[width=3in]{TwoInnermostsC1B.pdf}\hspace{.45in}
\includegraphics[width=3in]{TwoInnermostsC2.pdf}
\caption{If both circles of $F\cap S^\pm$ traversing a given over/underpass, say at $C_t$, have height zero, then $F$ can be isotoped to have a standard tube near $C_t$.}
\label{Fi:TwoInnermosts}
\end{center}
\end{figure}
\begin{proof}
Let $C_t$ be the ball at the crossing in question. Recall that the two arcs traversing the overpass (\textsc{wlog}) at $C_t$ lie on distinct circles of $F\cap S^+$, by Lemma \ref{L:inessentialCase}. Consider the two edge-pairs of $\partial\nu L$ that are incident to the underpass at $C_t$. Because $D$ is alternating, each edge-pair contains endpoints of $F\cap S^+\cap S^-$. Consider the endpoint on each edge-pair that is nearest to $C_t$ in $\partial\nu L$---recall that $F\cap\partial\nu L$ is (the image of) a section of the disk-bundle $\nu L\to L$. Up to symmetry, there are two cases, depending on which sides of $D$ these two points lie on, relative to each other (cf. Figure \ref{Fi:TwoInnermosts}).
If these two points lie in adjacent quadrants near $C_t$ (left in Figure \ref{Fi:TwoInnermosts}), perform two bigon moves (unless the associated arc is parallel in $S^\pm\setminus C$ to $F$). Since each (possible) bigon move follows an arc with an endpoint on $\partial\nu L$, Lemma \ref{L:Move} implies that these moves preserve the fact that $S^+\cup S^-\cup \nu L$ cuts $F$ into disks. Moreover, they produce a diagram in which an arc $\alpha\subset\partial\nu L\cap S^-$ has endpoints on the same circle of $F\cap S^-$ but on distinct circles of $F\cap\partial\nu L$, which Lemma \ref{L:inessential} states is impossible.
Therefore, these two points must lie in opposite quadrants $F\setminus D$ near $C_t$ (right in Figure \ref{Fi:TwoInnermosts}). In this case, a pair of bigon moves (to be omitted if trivial) immediately fashions a standard tube near $C_t$. Lemma \ref{L:MoveCase} implies that these moves preserve the fact that $S^+\cup S^-\cup \nu L$ cuts $F$ into disks.
\end{proof}
\section{Main results}\label{S:4}
%
\begin{ctlemma}
Given a non-trivial, connected, reduced alternating diagram of a link $L$ and a closed, essential surface $F\supset L$,
there exists an isotopy after which $F$ has a standard {tube} near some crossing.
\end{ctlemma}
\begin{proof}
As in \textsection\ref{S:2}, let $L\subset (S^2\setminus C)\cup\partial C$ follow a reduced alternating diagram, with $\nu L$ a closed regular neighborhood of $L$ seen as (the total space of) a disk-bundle $\pi:\nu L\to L$, $B^\pm$ the components of $S^3\setminus (S^2\cup \text{int}(C\cup\nu L))$, and $S^\pm=\partial B^\pm$. Let $F$ be a closed, essential surface containing $L$. Fixing $L\subset F$, $S^2$, and $C$, isotope $F$---subject to the requirements that $F\pitchfork S^+,S^-$; $\pi|_F$ is a bundle map; and $F\cap C\cap\partial\nu L=\varnothing$---so as to minimize lexicographically the numbers of components of $F\cap\partial C\setminus\nu L$ and $F\cap S^+\cap S^-$.
Consider $F\cap S^+$. If all circles have height 0, apply Lemma \ref{L:TwoInnermosts} at any overpass, done. Otherwise, by Proposition \ref{P:Height1Exists}, there exists a circle $\gamma_1$ of $F\cap S^+$ with height 1. Let $\gamma_0$ be any (innermost) circle enclosed by $\gamma_1$. Apply Proposition \ref{P:InnermostPass} to consider an overpass which $\gamma_0$ traverses. Let $\gamma$ denote the other circle of $F\cap S^+$ traversing this overpass. Note that $\gamma\neq\gamma_0$ by Lemma \ref{L:inessential}, specifically Lemma \ref{L:inessentialCase}. If $\gamma$ has height 0, then Lemma \ref{L:TwoInnermosts} completes the proof. Otherwise, $\gamma$ must equal $\gamma_1$.
See Figure \ref{Fi:FinalMove}.
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{FinalMoveE.pdf}
\caption{The final sequence of moves in the proof of the crossing tube lemma.}
\label{Fi:FinalMove}
\end{center}
\end{figure}
Next, consider the circle $\gamma'$ of $F\cap S^+$ from Figure \ref{Fi:FinalMove}, which must exist and be distinct from $\gamma_0$, due to Lemmas \ref{L:EdgeTypes}, \ref{L:inessential}, and the assumption that $\gamma_0$ has height 0.
If $\gamma'$ also has height 0, then Lemma \ref{L:inessential}
implies that the edge of $\partial\nu L$ in question must appear as in Figure \ref{Fi:EdgeTypes}, second from right; thus, $\gamma_0$ and $\gamma'$ must traverse a common overpass, completing the proof, using Lemma \ref{L:TwoInnermosts}.
Otherwise, $\gamma'=\gamma_1=\gamma$. This allows the sequence of isotopy moves shown in Figure \ref{Fi:FinalMove}, yielding the desired crossing tube. (Again, omit either isotopy move if the associated arc is parallel in $S^\pm\setminus C$ to $F$; and Lemma \ref{L:MoveCase} applies since each bigon move is along an arc with an endpoint on $\partial\nu L$.)
\end{proof}
\begin{maintheorem}
Every non-split, non-trivial alternating link $L$ has representativity $r(L)=2$.
\end{maintheorem}
\begin{proof}
Beginning with a {reduced} alternating diagram of $L$, apply the crossing tube lemma to obtain a standard tube at some crossing. Then apply Lemma \ref{L:sufficient} to conclude that the crossing tube contains a disk $Z$ with $Z\cap F=\partial Z$, such that $\partial Z$ intersects $L$ in two points and does not bound a disk in $F$.
\end{proof}
\section{Conclusion}\label{S:5}
\begin{figure}
\begin{center}
\includegraphics[height=3.5in]{AlmostAltTorusB.pdf}
\caption{A torus $F$ containing the knot $T_{3,4}$ in an almost-alternating diagram, with the
graph $G^+$. There are two circles of $F\cap S^+$ with height 2, two with height 1, and six
with height 0.}\label{Fi:T34}
\end{center}
\end{figure}
Figure \ref{Fi:T34} shows a torus $F$ containing the knot $L=T_{3,4}$, with $L$ in an almost-alternating diagram. The torus is difficult to visualize directly from this diagram, but its homeomorphism type is straightforward to verify by computing euler characteristic. As $r(L)\geq r(F,L)=3>2$,
this example illustrates how the arguments leading up to the crossing tube lemma break down without alternatingness. The configuration satisfies the conclusions of all Propositions and Lemmas from \textsection\ref{S:31}, but escapes those from \textsection\ref{S:32} (regarding edge types and innermost possibilities).
If an almost-alternating link $L$ has representativity $r(L)\geq 3$, then there can be no crossing tube in any diagram with minimal crossing number. Thus, after applying the setup and results from \textsection2, \textsection3.1, all non-innermost circles of $F\cap S^+$ and $F\cap S^-$ must (properly enclose or) be incident to the ``de-alternator.'' Otherwise, a local application of the proof of the crossing tube lemma gives $r(F)\leq 2$. Likewise, at least one of the two circles traversing each over/underpass must be incident to (or properly enclose) the de-alternator. Perhaps it is feasible to list explicitly which almost alternating links have representativity $\geq 3$.
|
\section{Introduction}
On 14$^{th}$ September 2015 the analysis, by the LIGO and Virgo Collaborations, of the data collected by the two LIGO interferometers during the science run O1, allowed to detect for the first time the gravitational waves (GW) emitted by a binary black hole merger \cite{abbot:detection}, followed on 26$^{th}$ December 2015 by the detection of a second event again associated to a binary black hole merger\cite{abbot:detection2}. These two events have started the era of gravitational waves astronomy. In fact, other sources of GWs are expected to emit signals potentially detectable by current interferometers. Among these there are spinning neutron stars (NS) asymmetric with respect to the rotation axis, which are expected to emit continuous, nearly monochromatic signals (CW). In the standard case of an asymmetric star rotating around one of its principal axis of inertia, the signal frequency is twice the rotational frequency of the star.
The expected signal amplitude is very small but one can exploit the signal long duration, at least in principle, to build-up the signal-to-noise ratio to detectable levels.
Depending on the degree of knowledge of the NS parameters, different types of searches can be done. \textit{Targeted searches} for CW from known pulsars are the most sensitive
\cite{jasi:gw_known_pulsars} \cite{matt:bayes} \cite{krolak:fstat}. They are based on the application of matched filtering and require a very accurate knowledge of the source parameters:
right ascension $\alpha$, the declination $\delta$ and the rotational frequency of the NS $f_{rot}$ with its derivatives $\dot{f}_{rot}$, $\ddot{f}_{rot},...$. The knowledge of the source
parameters
comes from electromagnetic observations at different wavelengths. On the other extreme, \textit{blind searches}, which are based on hierarchical approaches, are able to explore a wide volume in the parameter space at the cost of a lower sensitivity \cite{astone:allsky}\cite{powerflux}\cite{fstat}\cite{skyhough}\cite{einathome}.
\textit{Narrow-band searches} are an extension of targeted searches, in which the position of the source is assumed to be accurately known while the rotational parameters are slightly uncertain. This type of search can still be based on matched filtering but, of course, is computationally heavier with respect to targeted searches. In general, narrow-band searches allow to take into account a possible mismatch between the GW rotational parameters and those inferred from electromagnetic observations. For instance, the GW signal could be emitted by the core of the NS which may have a slightly different rotational frequency with respect to the magnetosphere. Secondly, in order to make a targeted search for a given object updated ephemeris covering the time span of the data at hand are needed. If they are not available, a fully coherent search based on wrong ephemeris could introduce a phase error with a consequent loss of signal-to-noise ratio \cite{ashton:snr_logg}. In recent years an analysis pipeline able to make a fully coherent search over a fraction of Hertz and tens of spin-down values has been developed \cite{rob:method} and applied to interferometric detector data \cite{rob:obs}.
In this paper we present an improved, and computationally cheap, algorithm for narrow-band searches which is also suited to coherently explore a wider volume
of the parameter space. The paper is organized as follows. In section \ref{sec:II} we summarize the main features of CW signals. In section \ref{sec:III} we
describe the new narrow-band pipeline, stressing the improvements with respect to the previous implementation. In section \ref{sec:IV} we describe the
validation tests of the algorithm, done using both software and hardware simulated CW signals. In this section we also discuss the computational load of the pipeline. In section \ref{sec:V} we present a case study discussing the search for CW from the central compact object G353.6-0.7 in Virgo detector VSR4 data. Finally, in Sec. \ref{sec:VI} we discuss the future perspectives for the application of this new pipeline.
\section{The signal}
\label{sec:II}
The CW signal emitted by an asymmetric spinning neutron star rotating around a principal axis of inertia can be written, following the formalism first introduced in \cite{pia:articolo},
as:
\begin{equation}
h(t)= H_0 ( H^+ A_+ (t) + H^\times A_\times (t)) e^{2 \pi i f_{gw} (t)}
\label{eq:Hgrande}
\end{equation}
where taking the real part is understood.
The complex amplitudes $H^+, H^\times$ are given by:
\begin{eqnarray*}
H^+ =\frac{\cos(2 \psi) - i \eta \sin (2 \psi)}{\sqrt{1+\eta^2}} \qquad H^\times =\frac{\sin(2 \psi) - i \eta \cos (2 \psi)}{\sqrt{1+\eta^2}} \\
\end{eqnarray*}
where $\eta$ is the ratio of the polarization ellipse semi-minor to semi-major axis and the polarization angle $\psi$ is defined as the direction of the major axis with respect to the celestial parallel of the source (measured counter-clockwise).
The parameter $\eta$ takes the values from -1 to 1 (in particular, $\eta=\pm 1$ if the wave is circularly polarized clockwise or counter-clockwise and $\eta=0$ if the wave is linearly polarized).
The functions $A_+ (t),\, A_\times (t)$ are the detector \textit{sidereal responses} and encode the interferometer response to the GW. The signal at
the detector is not monochromatic, i.e. the frequecy $f_{gw}=2f_{rot}$ in Eq. (\ref{eq:Hgrande}) is not constant. In fact it is modulated by some effects such as the \textit{Romer Delay},
due to
the detector motion, and the source intrinsic spin-down, due to energy loss from the source. All these effects must be properly taken into account in order to increase the signal to noise ratio in the analysis.
It can be shown that the waveform defined by Eq. \ref{eq:Hgrande} is equivalent to the GW signal expressed in the more standard formalism of \cite{jasi:gw_known_pulsars}, given the following relations:
\begin{equation}
\eta=-\frac{2\cos \iota}{1+\cos^2 \iota},
\label{eq:etaiota}
\end{equation}
where $\iota$ is the angle between the line of sight to the source and the star rotation axis, and
\begin{equation}
H_0=h_0 \sqrt{ \frac{1+6 \cos ^2 \iota + \cos^4 \iota}{4}}
\end{equation}
with
\begin{equation}
h_0=\frac{1}{r} \frac{4 \pi^2 G }{c^4} I_{zz} f_{gw}^2 \epsilon
\label{eq:GW_amplitude}
\end{equation}
being $d,I_{zz}$ and $\epsilon$ respectively the distance, the NS's moment of inertia with respect to the rotation axis and the {\it ellipticity}, which measures the star degree of
asymmetry.
Given a source with measured rotation frequency $f_{rot}$, frequency derivative $\dot{f}_{rot}$ and distance $d$, the GW signal amplitude can be constrained assuming that all the
rotational energy is lost via gravitational radiation. This absolute upper limit, called {\it spin-down limit}, is given by:
\begin{equation}
h_{sd}=8.06 \cdot 10^{-19} I_{38} \bigg(\frac{1kpc}{d} \bigg) \bigg(\frac{\dot{f}_{rot}}{Hz/s} \bigg)^{1/2} \bigg(\frac{Hz}{f_{rot}} \bigg)^{1/2}
\label{eq:sd_limit}
\end{equation}
being $I_{38}$ the star moment of inertia in units of $10^{38} kg \, m^2$. The corresponding spin-down limit on the star "fiducial" ellipticity
\cite{jasi:gw_known_pulsars} is
\begin{equation}
\epsilon_{sd}=0.307 \bigg( \frac{h_{sd}}{10^{-24}}\bigg) f_{rot}^{-2} I_{38}^{-1} \bigg(\frac{d}{1 kpc} \bigg)
\end{equation}
Eq. \ref{eq:sd_limit} can be re-written using the age parameter $\tau=\frac{f_{rot}}{\dot{f}_{rot}}$ and assuming that the NS was born with a spin much higher than the one that we observe
now:
\begin{equation}
h_{sd,\tau} \simeq 1.24 \times 10^{-24} \left( \frac{3.3 kpc}{d} \right) \sqrt{\left( \frac{I_{zz}}{10^{45} g cm^2} \right) \left( \frac{300 yr}{\tau} \right) }
\label{eq:sdlimit}
\end{equation}
For a given source, even in absence of a detection, establishing an upper limit below the spin-down limit is an important milestone, as it allow us to put a non-trivial
constraint on the fraction of rotational energy lost through GWs.
\section{Narrow-band search}
\label{sec:III}
In this section we remind the main general features of a narrow-band search, and discuss what are the main improvements we have introduced with respect to the original pipeline firstly described in \cite{rob:method}. In a narrow-band search the source position in the sky is assumed to be known and the analysis is run over a grid built in the frequency/spin-down space. The number of points in the grid obviously depends on the range of frequency, $\Delta f$, and spin-down, $\Delta \dot{f}$, that we want to explore, and on the grid steps which, for a fully coherent analysis, depend only on the total observation time $T_{obs}$. More precisely, the number of frequency bins is given by the ratio between the frequency range $\Delta f$ and the frequency bin width $\delta f=\frac{1}{T_{obs}}$:
\begin{equation}
n_f = \Delta f T_{obs} \approx 3 \cdot 10^7 \bigg( \frac{\Delta f}{1 Hz} \bigg) \bigg( \frac{T_{obs}}{1 yr}\bigg),
\end{equation}
while the number of spin-down bins is given by the ratio of the spin-down range $\Delta \dot{f}$ and the spin-down bin width $ \delta \dot{f}=1/ T^2 _{obs}$\footnote{The spin-down bin is defined as that spin-down value such that during the observation time $T_{obs}$ the frequency variation due to it is within a frequency bin.} :
\begin{equation}
n_{\dot{f}} = \Delta \dot{f} T^{2}_{obs} \approx 900 \bigg( \frac{\Delta \dot{f}}{10^{-12} Hz} \bigg) \bigg( \frac{T_{obs}}{1 yr}\bigg)^2
\end{equation}
Similarly, the number of values for the second order spin-down term is given by
\begin{equation}
n_{\ddot{f}} = \Delta \ddot{f} T^{3}_{obs}/2 \approx 0.16 \bigg( \frac{\Delta \ddot{f}}{10^{-23} Hz} \bigg) \bigg( \frac{T_{obs}}{1 yr}\bigg)^3
\end{equation}
For reasonably small ranges of the second order spin-down parameter $ n_{\ddot{f}}=1$, then the total number of points explored in the parameter space is then $n_{tot}=n_f n_{\dot{f}}$, which scales as $T_{obs}^3$.
The flow charts of the old and the new pipelines are shown if Fig. \ref{fig:flowchart}. Some of the core parts of the analysis, highlighted by dashed boxes, have a different implementation in the two cases. Let us start briefly describing what are the main steps of the old pipeline.
\begin{figure}[h!]
\hspace*{1.7cm}
\includegraphics[scale=0.35]{flow_chart.eps}
\caption{Top Panel: Flowchart of the old pipeline In particular the procedures that have been improved are indicated in a dashed box. More details on the pipeline can be found in \cite{rob:method}. Bottom Panel: Flowchart of the new pipeline, the procedures indicated in the dasehed box will be described in this paper.}
\label{fig:flowchart}
\end{figure}
For a given detector and source position, the barycentric corrections are applied by introducing a new time variable defined as
\begin{equation}
t'= t+ \Delta_{R} + \Delta_{E} -\Delta_{S}
\label{eq:tprimo}
\end{equation}
where $\Delta_{R},\, \Delta_{E},\, \Delta_{S}$ are, respectively, the Romer, Einstein and Shapiro delay. Seen in the new time variable $t'$, which is the time
in the Solar System Barycenter, a signal in the data would not be subject to the barycentric modulation anymore. It is important to stress that the time $t'$
does not depend on the frequency, hence the barycentric corrections are computed just once and hold for the whole frequency band. The spin-down correction is
done using the same technique, re-defining the time as:
\begin{equation}
\tau= t'+ \frac{\dot{f}}{2f}(t'-t_0')+ \frac{\dot{f}}{2f}(t'-t_0')^2
\end{equation}
It is clear that the time $\tau$ depends explicitly on the frequency and spin-down values.
Seen in the non-uniformly sampled time variable $\tau$, the signal is now monochromatic apart from the amplitude and phase modulation encoded in the sidereal responses.
Once the corrections have been done the data are down-sampled at 1 Hz. Now, for each frequency and spin-down bin of the grid, a set of five discrete Fourier Transforms (DFT) is evaluated in order to get the so-called {\it 5-vectors}. Given a time series, a 5-vector consists of the Fourier components at the five frequencies at which the power of a CW signal is split by the sidereal modulation:
\begin{equation}
G_k(f)=\sum_{i=0}^{N_{s}} s(t_i)e^{- 2 \pi i(f-k F_{\oplus})t_i}dt
\end{equation}
where $F_{\oplus}$ is the Earth's sidereal frequency and $k_i=-2,-1,0,1,2$. See \cite{pia:articolo} for more details. Then, two matched filters are computed in the frequency domain by making the scalar product between the data 5-vector $\vec{X}$ and the normalized sidereal response 5-vectors $\vec{A}_{+ / \times} / |\vec{A}_{+ / \times}|^2 $:
\begin{equation}
\widehat{H}^{+/ \times} (f) =\vec{X} (f) \cdot \frac{\vec{A}_{+ / \times}}{|\vec{A}_{+ / \times}|^2}
\end{equation}
where the "$\cdot$" implies taking the complex conjugate of the second term.
The quantities $\widehat{H}^{+/ \times}$ are estimators of the CW complex amplitudes, defined in Eq.\ref{eq:Hgrande}, and are used to build the detection statistic:
\begin{equation}
S=|\vec{A}_{+}|^4 |\widehat{H}^+|^2 +|\vec{A}_\times|^4 |\widehat{H}^\times|^2
\end{equation}
Once we have obtained a value of the detection statistic for each frequency and spin-down bin, we select the most interesting candidates for the follow-up. Typically, this is done by setting a threshold on the detection statistic, corresponding to a small p-value on the noise-only distribution (e.g. 1\%, after taking into account the number of trials), and taking the candidates with a detection statistic above the threshold \cite{rob:obs}.
The computational cost of the pipeline described so far is dominated by the computation of a large number of DFTs, one for each frequency and spin-down bin.
\subsection{Main features of the new pipeline}
In the new pipeline, the barycentric corrections are done exactly in the same way as in the old pipeline.
Let us now describe which are the main improvements in the rest of the pipeline, which allow to dramatically reduce the computational load of the analysis and, at the same time, also improve the sensitivity. They can be summarised as: \textit{a)} application of the spin-down correction in phase and after the down-sampling ; \textit{b)} use of the Fast Fourier Transform algorithm (FFT) to compute 5-vectors over the frequency grid; \textit{c)} FFT interpolation at half-bins in order to reduce the sensitivity loss due to the discretization.
\subsubsection{Spin-down correction}
The spin-down correction is now applied in phase rather than in time. The frequency evolution of the signal emitted by a spinning NS can be expressed using a Taylor expansion of the frequency with respect to time:
\begin{equation}
f(t)= f(t_0)+ \sum_{k\ge } \frac{1}{k!} \frac{d^k f}{dt^k} \bigg{|}_{t=t_0} (t-t_0)^k
\label{eq:sd}
\end{equation}
Theoretically, we can have any value for the order $k$ but, practically, for most known pulsars only the first 2-3 terms are measurable.
The corresponding phase evolution is given by
\begin{equation}
\phi_{sd} (t)= 2 \pi \int_{t_0}^{t} \sum_{k\ge 1} \frac{1}{k!} \frac{d^k f}{dt^k} \bigg{|}_{t=t_0} (t-t_0)^k dt = 2 \pi \sum_{k\ge 1} \frac{1}{(k+1)!} \frac{d^k f}{dt^k} \bigg{|}_{t=t_0} (t-t_0)^{k+1}
\label{eq:sd_correction}
\end{equation}
The frequency variation, due to the spin-down of a possible signal into the data, can be removed by multiplying the data by the factor $e^{-i 2 \pi \phi_{sd} (t)} $. In this way the signal would become
\begin{equation}
h'(t)=h(t)\cdot e^{- 2 \pi i \phi_{sd}}=H_0(H^{+}A_{+}(t)+H^{\times}A_{\times}(t))e^{i (f_{gw} t + \phi_0)}
\end{equation}
which is monochromatic apart from the amplitude and phase modulation due to the sidereal responses. T0he spin-down corrections are applied after the down-sampling, thus reducing the computational cost of the algorithm.
\subsubsection{Replacing the Discrete Fourier Trasform with the Fast Fourier Trasform }
As shown in previous sections, we previously computed the DFT in every frequency bin in order to identify the 5-vectors. The computational cost of the DFT is the main bottleneck if it must be performed a large number of times.
To overcome this problem we have modified the original pipeline (written in Matlab) replacing the computation of the DFT for each frequency bin with an FFT.
The FFT is computed in Matlab using the Turkey-Cooley algorithm \cite{turkey:FFT}. Before proceeding, let us see how the DFT and the FFT are related. The value
of the DFT of a time series with samples $x_j$, at the frequency $f_k$, is given by:
\begin{equation}
\tilde{x}_{DFT}(f_k)=\sum_{j=0}^{N-1} x_{j} e^{-2 \pi i j \delta t f_k} \delta t
\end{equation}
where $i$ the imaginary unit and $ \delta t $ is the sampling time. With some algebraic manipulations we find
\begin{equation}
e^{-2 \pi i \delta t f_k}\tilde{x}_{DFT}(f_k)=\sum_{j=1}^{N} x_{j-1} e^{-2 \pi i j \delta t f_k} \delta t
\label{eq:DFT}
\end{equation}
The FFT is defined in Matlab as
\begin{equation}
\tilde{x}_{FFT}(k)=\sum_{j=1}^{N} x_j e^{-\frac{2 \pi i}{N}(j-1)(k-1)}=\sum_{j=1}^{N} x_j e^{-2 \pi i \delta t \delta f(j-1)(k-1)}
\end{equation}
where $\delta f$ is the frequency bin. After further manipulations we obtain
\begin{equation}
\tilde{x}_{FFT}(k)=e^{2 \pi i \delta t \delta f (k-1)} \sum_{j=1}^{N} x_j e^{-2 \pi i \delta t \delta f j (k-1)}
\label{eq:FFT}
\end{equation}
Given that $k \cdot \delta f=f_k$ and $j \cdot \delta t=t_j$ are the time samples, we have:
\begin{eqnarray*}
& \tilde{x}_{FFT}(k=1)=\tilde{x}_{FFT}(f=0)=\sum_{j=1}^{N} x_j \\
&\tilde{x}_{FFT}(k=2)=\tilde{x}_{FFT}(f=\delta f)=e^{2 \pi i \delta t \delta f} \sum_{j=1}^{N} x_j e^{-2 \pi i t_i \delta f }
\end{eqnarray*}
and so on. By comparing Eq. \ref{eq:DFT} and Eq. \ref{eq:FFT} we obtain the simple relation:
\begin{equation}
\tilde{x}_{DFT} (f)=\tilde{x}_{FFT} (f) \cdot \delta t
\end{equation}
Hence, the following relations for the complex amplitude estimators follow:
\begin{equation}
\widehat{H}^{+/ \times}_{FFT}(f)=\widehat{H}^{+/ \times}_{DFT}(f)
\end{equation}
while for the detection statistic we have
\begin{equation}
S_{FFT}(f_k)\cdot \delta t^4=S_{DFT}(f_k)
\label{eq:rel_dt}
\end{equation}
In order to fully exploit the use of the FFT in the computation of 5-vectors, a suitable grid in frequency must be built. More precisely, the sidereal angular frequency of the Earth is chosen to be an integer multiple of the frequency bin. In this way the 5-vector components correspond to integer frequency bins of the FFT. This is done by properly adjusting the observation time in such a way that the frequency bin is a sub-multiple of the sidereal angular frequency. Typically, this implies cutting away a few hours of data with a completely negligible effect on the search sensitivity.
The use of the FFT, together with this adjusted frequency grid, allows to compute the entire ensemble of 5-vectors over a given frequency range with only one application of the FFT algorithm, thus providing a huge gain in computational speed.
We have done several checks of consistency between the DFT-based and the FFT-based algorithm.
A first check consisted in comparing the estimators $\widehat{H}^{+/ \times}$ computed in the two cases on simulated Gaussian noise. The standard deviation of the mismatch, defined as
$\frac{Q_{DFT}}{Q_{FFT}}-1$, where $Q$ is the real or imaginary part of $\widehat{H}^{+/ \times}$, is up to a few units per $\sim 10^{-4}$ , see Fig. \ref{fig:H}.
\begin{figure}[h]
\hspace*{-0.8cm}
\includegraphics[scale=0.32]{ratios_H.eps}
\caption{The plots show the ratio of the amplitude estimators, computed respectively with the $DFT$ and $FFT$, as a function of the frequency bin on simulated
Gaussian noise data covering 4 months ( for both the real and imaginary part). Top left: ratio of the real part of the $\times$ estimators. Top right: ratio of
the imaginary part of the $\times$ estimators. Bottom left: ratio of the real part of the $+$ estimators. Bottom right: ratio of the imaginary part of the $+$
estimators. The standard deviation is indicated in the boxes.}
\label{fig:H}
\end{figure}
\clearpage
The next check consisted in comparing the values of the detection statistic obtained using the DFT and the FFT algorithms. We have tested the relation given by Eq. \ref{eq:rel_dt} for sampling rates $\delta t$ different from $1s$ proving that it is valid apart from a relative error smaller than $\sim 10^{-5}$, see Fig. \ref{fig:DS_dt}.
\begin{figure}[h!]
\hspace*{-2.25cm}
\includegraphics[scale=0.38]{ratio_S_di_dt.eps}
\caption{Ratio of the detection statistics computed with the DFT and the FFT as a funcition of the frequency bin, using simulated Gaussian data at various sampling times $\delta t$. Top Panel: $\delta t=0.5 s $. Central Panel: $\delta t=1.0 s $. Bottom Panel: $\delta t=2.5 s $. As expected from Eq. \ref{eq:rel_dt}, the variance of the mismatch is a function of the sampling rate, with higher sampling rates providing a smaller difference.}
\label{fig:DS_dt}
\end{figure}
\subsubsection{FFT interpolation}
Due to the discretizzation of the frequency, it is possible that the signal's frequency falls between two frequency grid points. This produces a loss of signal-to-noise ratio of up to $ \sim 36 \%$. In order to reduce this loss a standard and cheap method consists in estimating the FFT values at half bin using the ``interbinning'' approximation, see e.g. \cite{ransom:fourier}:
\begin{equation}
\tilde{x}_{FFT,k+\frac{1}{2}} \simeq \frac{1}{4} \pi (\tilde{x}_{FFT,k} - \tilde{x}_{FFT,{k+1}})
\end{equation}
where the index $k$ denotes the $k$-th frequency bin. We have added this feature to the narrow-band search pipeline. Interbinning has some weak points, like
the fact it introduces a correlation between half and integer bins. An alternative, which avoids correlations but double the time series length, and then the
computational cost, would consist in using zero-padding, adding a number of zeros equal to the original data length. As discussed in \cite{ransom:fourier},
however, interbinning can be usefully used in the detection stage, eventuallys relying on more sophisticate, and computationally heavier, techniques to make deeper searches in the follow-up stage.
\section{Full pipeline tests}
\label{sec:IV}
In this section we describe two kinds of tests done using the new narrow-band search pipeline. The first one consisted in analyzing Virgo VSR4 data searching for a few simulated CW signals directly injected into the interferometer (called {\it hardware injections}). The aim of the test was to verify if we are able to detect these signals and accurately estimate their parameters. The second test consisted in running a narrow-band search for the Crab pulsar again in Virgo VSR4 data. This analysis has been already done, on the same data set, using the old pipeline \cite{rob:obs}. Here we mainly want to make a comparison of the pipelines speed.
\subsection{Hardware injections}
Hardware injections are artificial signal injected during detector observation runs. For the pipeline's test we have chosen three hardware injections, which parameters are shown in Tab. \ref{tabl2}.
\begin{table}[h!]
\begin{center}
\caption{Parameters of the hardware injections used to test the algorithm. $H_0$ is the signal amplitude, $f$ is the signal frequency, $\dot{f}$ is the spin-down, $\alpha$ the right ascension, $\delta$ the declination, $\eta$ the ratio between the polarization ellipse semi-minor and semi-major axes and $\psi$ is the wave polarization angle.}
\label{tabl2}
\hspace*{-1.0cm}
\begin{tabular}{cccccccc}
\hline
HI name&$f~$[Hz]&$\dot{f}~$[Hz/s]&$H_0$&$\eta$&$\psi$[deg]&$\alpha$ [deg]&$\delta$ [deg]\\ \hline
Pulsar2&$575.16357300$&$1.37 \cdot 10^{-13}$&$5.2844 \cdot 10^{-24}$&$0.997$&$77.292$&$215.256166$&$3.443989$ \\
Pulsar3&$108.85715940$&$1.46 \cdot 10^{-17}$&$8.2961 \cdot 10^{-24}$&$0.160$&$25.439$&$178.372574$&$−33.436602$ \\
Pulsar8&$192.23705881$&$8.65 \cdot 10^{-9}$&$8.0673 \cdot 10^{-24}$&$-0.147$&$9.7673$&$351.389582$&$−33.418516$ \\
\hline
\end{tabular}
\end{center}
\end{table}\newline
For the tests we have used the entire VSR4 run of Virgo. The run duration is about $T_{obs}\approx 8.14 \cdot 10^{6} s$. Therefore, we have the following bin width for the frequency, first order spin-down and second order spin-down, respectively:
\begin{eqnarray*}
& \delta f =\frac{1}{T_{obs}} \simeq 1.23 \cdot 10^{-7} \, \textrm{Hz} \qquad \delta \dot{f} =\frac{1}{T_{obs}^2} \simeq 1.51 \cdot 10^{-14} \, \textrm{Hz}/s \\ & \qquad \qquad \qquad \delta \ddot{f} =\frac{2}{T_{obs}^3} \simeq 3.70 \cdot 10^{-21} \, \textrm{Hz}/s^2
\end{eqnarray*}\newline
We have ran the analysis in a small volume of the frequency / spin-down space around the parameters of each injected signal, see Tab \ref{tablII}. In all the
cases, the loudest value of the detection statistic corresponds to the frequency and spin-down bins nearest to the injected parameters.
\Table{\label{tablII} Parameter space explored around three hardware injections. $\Delta f$ is the frequency range and $n_f$ the corresponding number of bins,
$\Delta \dot{f}$ is the spin-down range and $n_{\dot{f}}$ the corresponding number of bins. Finally, $n_{tot}=n_f\cdot n_{\dot f}$ is the total number of points.}
\br
Name&$\Delta f$ [Hz]&$\Delta \dot{f}$ [Hz/s] &$n_f$ &$n_{\dot{f}}$ &$n_{tot}$ &\\
\mr
Pulsar2&$0.02$&$2.6 \cdot 10^{-13}$&$1.3 \cdot 10^5$&$17$&$2.2 \cdot 10^6$ &\\
Pulsar3&$0.04$&$3.0 \cdot 10^{-13}$&$2.6 \cdot 10^5$&$20$&$5.2 \cdot 10^6$&\\
Pulsar8&$0.20$&$2.2 \cdot 10^{-13}$&$1.3 \cdot 10^6$&$15$&$ 1.9 \cdot 10^7$&\\
\br
\end{tabular}
\end{indented}
\end{table}
From the loudest candidate we have estimated the signal parameters, obtaining values in good agreement with the injected ones. In fact, they are the best estimation we can obtain with the given search grid. In Tab. \ref{tabl3} we report the test results for the three considered hardware injections. Note, in particular, that the frequency error is always smaller than 1 bin, while the error in the estimation of the spin-down is always zero because the grid we have used for this test was centered at the exact spin-down of each signal.
\Table{\label{tabl3} Recovered parameters for the hardware injection analysis: $\epsilon_f$ is the mismatch, in units of frequency bin, between the injected
frequency and the detected frequency; $\epsilon_{\dot{f}}$ is the the mismatch between the injected spin-down value and the detected spin-down value; $\epsilon_{H_0}$ is the ratio of the estimated and injected amplitude; $\epsilon_\eta$ the normalized relative error on the parameter $\eta$ and $\epsilon_\psi$ the normalized relative error on the parameter $\psi$.}
\br
&\centre{5}{Relative errors}\\
&\crule{6}\\
Name&$\epsilon_f$&$\epsilon_{\dot{f}}$ &$\epsilon_{H_0}$ &$\epsilon_\eta$ &$\epsilon_\psi$ &\\
\mr
Pulsar2&$0.0373$&$0$&$1.041$&$0.0307$&/$^{\rm a}$&\\
Pulsar3&$0.1835$&$0$&$1.100$&$0.01$&$0.0085$&\\
Pulsar8&$0.0465$&$0$&$0.9833$&$0.0115$&$0.0225$&\\
\br
\end{tabular}
\item[]$^{\rm a}$ Pulsar\_2 is nearly circularly polarized and the polarization angle $\psi$ is then ill-defined and, consequently, badly estimated.
\end{indented}
\end{table}
\subsection{Computational cost}
We have performed a narrow-band analysis for the Crab pulsar, similar to the one described in \cite{rob:obs} and compared the corresponding computational costs. The analysis in \cite{rob:obs} covered a frequency range $\Delta f=0.02 \, \textrm{Hz} $ and a spin-down range $\Delta \dot{f}=2.49 \cdot 10^{-13} \, \textrm{Hz}/s $.
For testing purposes we have repeated the narrow-band searches with the improved pipeline covering a larger frequency and spin-down range.\\
Specifically, we have used a frequency range of $\Delta f= 0.1 \,\textrm{Hz}$ and a spin-down range of $ \Delta \dot{f}=6.40 \cdot 10^{-13} \, \textrm{Hz}/s$ around the "central" values.
The corresponding number of points in the parameter space is given, for both the old and the new pipeline, in Tab. \ref{tabl4}.
\Table{\label{tabl4} Explored volume in the parameter space for the narrow-band searches performed for the Crab pulsar with the "old" and new pipelines. $\Delta f$ is the explored frequency range, $\Delta \dot{f}$ the explored spin-down range, $n_f$ and $n_{\dot{f}}$ are the corresponding number of explored bins, and $n_{tot}$ is the total number of points in the parameter space.}
\br
Pipeline&$\Delta f$ [Hz]&$ \Delta \dot{f}$[Hz/s]&$n_{f}$&$n_{\dot{f}}$&$n_{tot}$\\
\mr
Old&$0.02$&$2.49 \cdot 10^{-13}$&$1.6 \cdot 10^{5}$&$33$&$5.8 \cdot 10^{6}$\\
New&$0.1$&$6.40 \cdot 10^{-13}$&$8.16 \cdot 10^{5}$&$165$&$1.34 \cdot 10^{8}$\\
\br
\end{tabular}
\end{indented}
\end{table}
For both the old and the new analysis we consider only one single value for the second order spin-down. This is due to the fact that the VSR4 run is short enough that the frequency variation due to second order spin-down is smaller than one frequency bin.
In the current test the number of points in the parameter space is $n_{tot}^{new}/n_{tot}^{old} \approx 26 $ times bigger than in the old analysis. The total
computational time for the entire analysis, using the new pipeline, was of about 10 hours with a single job. The old analysis was split in 33 jobs, corresponding to the number of spin-down corrections, each long about 10 hours. The analyses have been carried on machines with similar characteristics. This means that the new method is nearly 3 order of magnitudes faster than the previous one at fixed parameters space. This huge speed increase makes the pipeline suited also for wider band searches, like those for neutron stars with known position and more uncertain rotational parameters. This are called {\it directed searches}. An example of such kind of search is discussed in the next section.
\section{Application to a directed search}
\label{sec:V}
Semi-coherent directed searches have been done in the past for the central compact object in Cassiopeia A \cite{abadie:casA} and recently for other few supernova remnants \cite{aasi:SNR}.
In order to show the possible extension to this kind of search of the pipeline described in this paper, we have performed a fully coherent directed search on Virgo VSR4 data for the central compact object in a supernova remnant, which rotational parameters are unknown.
\subsection{Target description}
The central compact object (CCO) XMM UJ173203.3-344518 was discovered by XMM-Newton in 2007 \cite{tian:HESS} \cite{abra:hess} as a point-like X-ray source near the center of the TEV-emitting supernova remnant HESS J1731-347, a.k.a. G353.6-0.7.
The supernova remnant has a large angular size of $ \approx 0.5 \, \textrm{deg} $, from which a distance of 3.2 kpc and an estimated age of about 27 kyr have
been deduced \cite{tian:HESS2010}.
The celestial coordinates of the central compact object are $\alpha$=263.04166 deg and $\delta$= -34.7667 deg. No pulsation was observed hence we have no estimation of the rotational frequency and spin-down values of the neutron star.
We can estimate the indirect spin-down limit for the amplitude of the gravitational waves using Eq. \ref{eq:sdlimit}. For a distance of 3.2 kpc and an age of 27 kyr we obtain:
\begin{equation}
h_{0,in} \approx 1.4 \cdot 10^{-25}
\label{eq:gw_G353}
\end{equation}
Considering the number of points in the parameter space, see next section, the estimated sensitivity of the search is \cite{rob:method}
\begin{equation}
h_{sens}(f) \approx 25 \sqrt{\frac{S_{n}(f)}{T_{obs}}}
\label{eq:blabla}
\end{equation}
Given the typical VSR4 noise level and duration, the sensitivity given by Eq.(\ref{eq:blabla}) is about 3 times larger than the indirect spin-down limit in the frequency range of the best detector sensitivity.
Although this search has a limited scientific value, we have used it to show the feasibility of directed searches with the pipeline we have developed.
\subsection{Selection of the parameter space}
We have chosen a 20 Hz frequency band from 116 Hz to 136 Hz, in which the noise level is relatively stationary.
To define the spin-down range we have used the approximate relation which connects the frequency derivative to the frequency, the neutron star age $\tau$ and the braking index $n$:
\begin{equation}
\dot{f} \approx -\frac{f}{(n-1) \tau}
\end{equation}
Following the same approach used by \cite{wette:cassiopea} and \cite{abadie:casA}, we have considered a range of braking indexes between $n_{min}=3$ and $n_{max}=7$ which correspond, respectively, to magnetic dipole \cite{aasi:SNR} and {\it r-modes} emission \cite{owen:rmode}.
As a consequence, the spin-down range is given by
\begin{equation}
-\frac{f}{(n_{min}-1) \tau} \leq \dot{f} \leq - \frac{f}{(n_{max}-1) \tau}
\label{eq:sd_range}
\end{equation}
For the second order spin-down we have used the definition of braking index, obtaining a range
\begin{equation}
\frac{\dot{f}^2 n_{min}}{f} \leq \ddot{f} \leq \frac{\dot{f}^2 n_{max}}{f}
\label{eq:ssd_range}
\end{equation}
Given VSR4 run duration, $T_{obs}\simeq 8.14 \cdot 10^6 s$, the resulting bin widths are
$\delta f=1.2285 \cdot 10^{-7} \, \textrm{Hz} \ $, $ \delta \dot{f}=1.509 \cdot 10^{-14} \textrm{Hz}/s$,
$ \delta \ddot{f}=3.70 \cdot 10^{-21} \textrm{Hz}/s^2$.
corresponding to the following number of points in the parameter space:
$n_f \approx 1.6 \cdot 10^{8},~n_{\dot{f}} \approx 3300,~n_{\ddot{f}} =1,~n_{tot} \approx 4.8 \cdot 10^{11}$.
For practical reasons the search band has been divided in ten 2 Hz sub-bands later down-sampled at 2 Hz. This part of the analysis took bout 18 hours on a 8-core CPU.
The band 116-136 Hz is populated by several known disturbances. The polluted frequencies have not been taken into account in the computation of the detection statistic. This disturbances are listed in Tab. \ref{table5}. The very noisy band from 130 Hz to 132 Hz has been excluded from the analysis since the beginning.
\Table{\label{table5} Known noise lines in the frequency band 116-136 Hz of VSR4 data. These bands have not been considered in the computation of the detection statistic.}
\br
\textbf{Frequency} [Hz] & \textbf{Width} [Hz] & \textbf{Origin} \\
\mr
116.5 & 0.015& Laser calibration line \\
120.006 & 0.0055 & Harmonic of 10 Hz \\
123.343 & 0.0085 & Harmonic of 10.2872 Hz \\
133.622 & 0.0085 & Harmonic of 10.2872 Hz \\
\br
\end{tabular}
\end{indented}
\end{table}
In Tab. \ref{table6} the main parameters of the search for each 2-Hz band are reported. The total observation time is not long enough to appreciate more than one second order spin-down bin.
\Table{\label{table6}Search parameter space for XMM UJ173203.3-344518. The number of frequency bins explored in each 2 Hz sub-band is $n_{f}=1.63 \cdot 10^7$.}
\br
\centre{1}{\textbf{Frequency}} &&& \centre{1}{\textbf{Spin-down}} &&& \textbf{$n_{\dot{f}}$} && \centre{1}{$\ddot{f} $} \\
\textbf{band} [Hz] &&&\textbf{band} [Hz/s $\cdot 10^{-11} $] &&& && [Hz$/s^2 \cdot 10^{-23}$] \\
\mr
116-118 &&& $[-6.9246;-2.2691]$ &&& 3083 && \centre{1}{ $9.0464$} \\
118-120 &&& $[-7.0419;-2.3082]$ &&& 3134 && \centre{1}{$8.8956$} \\
120-122 &&& $[-7.1593;-2.3473]$ &&& 3186 && \centre{1}{$8.7498$} \\
122-124 &&& $[-7.2767;-2.3864]$ &&& 3238 && \centre{1}{$8.6086$} \\
124-126 &&& $[-7.3940;-2.4256]$ &&& 3290 && \centre{1}{$8.4720$} \\
126-128 &&& $[-7.5414;-2.4647]$ &&& 3342 && \centre{1}{$8.3396$} \\
128-130 &&& $[-7.6288;-2.5429]$ &&& 3394 && \centre{1}{$8.2113$} \\
130-132 &&& \centre{1}{-} &&& \centre{1}{-} && \centre{1}{-} \\
132-134 &&& $[-7.7461;-2.5429]$ &&& 3497 && \centre{1}{$7.9662$} \\
134-136 &&& $[-7.9800;-2.6212]$ &&& 3549 && \centre{1}{$7.8490$} \\
\br
\end{tabular}
\end{indented}
\end{table}\newline
The corresponding value of the frequency second derivative has been obtained, for each band of 2 Hz, taking the mean value of $\ddot{f}$ predicted from the braking index relation using breaking indexes n=3,5,7. The total explored parameter space consists of $n_{tot} \simeq 4.8 \cdot 10^{11} $ points.
The detection statistic has been computed for each point in the parameter space and we have taken the local maxima of the detection statistic every 0.001 Hz and over the full spin-down range. Keeping into account the "look-elsewhere" effect, as was done in \cite{rob:obs}, we have selected as candidates those local maxima in the frequency/spin-down plane with a value of the detection statistic corresponding to an overall p-value of 1\% or less on the noise only distribution.
\subsection{Candidates}
The selected candidates are plotted in Fig. \ref{fig:candidate}.
\begin{figure}[h!]
\includegraphics[width=1.\textwidth]{candidates_belli.eps}
\caption{Candidates with a value of the detection statistic over the threshold corresponding to a p-value equal to 1\% .}
\label{fig:candidate}
\end{figure}
They are grouped in three different sets.
Two of them are due to instrumental disturbances, although they are not in the official noise lines list. In particular, the set of candidates near 130 Hz has
been identified by the noise lines hunting tool NoEMi\footnote{Noise Frequency Event Miner (NoEMi) is a tool able to indetify on a daily basis, the frequency
lines observed in the Virgo science data looking for coincident lines present in different environment sensors \cite{accadia:NOEMI} \cite{accadia:NOEMI2}} as
due to an harmonic of the 10 Hz comb, the same that generated the 120.006 Hz line listed in Tab. \ref{table5}. These harmonics have several side-bands
extending in a range of $\sim 0.4 Hz$ around the central frequency, which are the likely source of this set of candidates.
The set of candidates at 133.648 Hz has been associated by NoEMi to an harmonic of the 10.2872 Hz comb, also shown in Tab. \ref{table5}. Although the nominal frequencies corresponding to this line have been removed before the computation of the detection statistic, the application of the Doppler correction shifts the line frequency pushing some candidates outside the vetoed band. There is just one remaining candidate, indicated as \textbf{Candidate 1} in Fig. \ref{fig:candidate}, which cannot be associated to any known instrumental disturbance. This candidate has the following parameters:
\begin{eqnarray}
& f= 116.6296417 \, \textrm{Hz} \qquad \dot{f}=-5.01608 \cdot 10^{-11} \, \textrm{Hz}/s \\
& \qquad \qquad \qquad \ddot{f}= 9.20232 \cdot 10^{-23} \, \textrm{Hz}/s^2
\label{eq:param}
\end{eqnarray}
The overall associated p-value is 0.006 which, in the case of noise with Gaussian distribution, would correspond to $\sim$2.5$\sigma$. No known noise line was found neither by
NoEMi
neither in the noise spectra at the candidate's frequency. However, this is not an outstanding candidate, and could be simply due to a noise fluctuation. Moreover, a detection at this level of sensitivity would require some non trivial explanation for why the signal strain amplitude is above the spin-down limit. Nevertheless, we have decided to proceed with some basic follow-up steps in order to confirm or reject the candidate. In Fig. \ref{fig:candidate1_out} the data power spectrum around the candidate is plotted,showing a region of excess power around the candidate frequency. In Fig. \ref{fig:candidate1} a zoom of the power spectrum is shown. The five dashed vertical lines correspond to the frequencies at which the signal power would be split due to the sidereal motion of the detector. There are indeed peaks at frequencies near the expected ones, but no clear correspondence.
\begin{figure}
\centering
\subfloat[][Data power spectrum around Candidate 1. The green dashed line corresponds to the frequency of the candidate 1. \label{fig:candidate1_out}]
{\includegraphics[width=.5\textwidth]{candidate1_faraway.eps}} \qquad
\subfloat[][Zoom of the power spectrum in Fig. \ref{fig:candidate1_out} around Candidate 1. The green dashed vertical lines are the expected frequencies of a
CW signal associated with Candidate 1. \label{fig:candidate1}]
{\includegraphics[width=.5\textwidth]{candidate1.eps}}
\end{figure}
As a further check, we have performed a targeted search for Candidate 1 in the data of the second science run of Virgo (VSR2). VSR2 extended from July-7th-2009 21h UTC to January-8th-2010, 22h UTC, two years before VSR4. Considering that VSR2 run has a sensitivity comparable to that of VSR4 in the explored frequency region, if \textbf{Candidate 1} arises form a continuous gravitational wave signal then it should be detectable also in this data set. The rotational parameters of the candidate given in Eq. \ref{eq:param}, which refer to the starting time of VSR4, have been shifted to the VSR2 starting epoch by taking into account the spin-down, obtaining the following parameters:
\begin{eqnarray*}
& f= 116.63265826 \textrm{Hz} \dot{f}=-5.0166 \cdot 10^{-11} \textrm{Hz}/s \\
& \qquad \qquad \quad \ddot{f}= 9.20232 \cdot 10^{-23} \textrm{Hz}/s^2
\end{eqnarray*}
A full coherent targeted search has been done using these parameters and the power spectrum of the corrected time series has been computed to look for the presence of significant peaks.
\begin{figure}
\centering
\subfloat[][Power spectrum of VSR2 data around the frequency of Candidate 1 after having performed a fully coherent search targeted at Candidate 1's parameters. No significant peak is visible. \label{fig:candidateVSR2}]
{\includegraphics[width=.5\textwidth]{candidateVSR2_faraway.eps}} \qquad
\subfloat[][Zoom of the power spectrum in Fig. \ref{fig:candidateVSR2} around Candidate 1 in VSR2 data. The green dashed vertical lines are the expected frequencies of a signal associated with Candidate 1. \label{fig:candidateVSR2_faraway}]
{\includegraphics[width=.5\textwidth]{candidateVSR2.eps}}
\end{figure}
The power spectrum is shown in Fig. \ref{fig:candidateVSR2}. The data appear quite clean and no significant peak is visible. The corresponding value of the detection statistic is fully compatible with noise. We conclude it is very unlikely \textbf{Candidate 1} is due to a GW signal.
\newpage
\section{Conclusions}
\label{sec:VI}
The search for continuous graviational wave signal emitted by known spinning neutron stars, such as pulsars, typically assumes that the gravitational wave
frequency is at a known fixed ratio with respect to the star rotational frequency, thus implying that the gravitational signal is "phase locked" to the
electromagnetic one. Fully coherent searches, based on matched filtering, in principle offer the best sensitivity. However, they assume signal parameters are
known with high accuracy. Due to our ignorance of the mechanisms at the base of both the electromagnetic and gravitational emission in spinning neutron stars,
we cannot exclude a small mismatch between the gravitational wave frequency and that inferred from electromagnetic observations. For long observation times,
even a very small discrepancy can prevent signal detection. In order to take into account such possible mismatch, it is very important to perform narrow-band
searches with the best possible sensitivity, that, in layman's terms means to apply matched filtering on several points of the source parameter space. The computational load of the analysis rapidly increases with the searched volume and the observation time. For these reasons past narrow-band searches were limited to frequency bands of a few milliHertz, and to few spin-down values around the central search values inferred from EM observation \cite{rob:obs}.
In this paper we have presented an improved pipeline for narrow-band searches of continuous gravitational wave signals emitted by asymmetric spinning neutron stars. The new pipeline is about three orders of magnitude faster than the old version allowing to perform, in few hours of CPU time on a standard workstation, a full coherent search of months of data over a frequency band of $\mathcal{O}(1$ Hz$)$ and several hundreds of spin-down values.
In the future, the use of many-core processors, like GPU, could boost the analysis further. Currently, this novel pipeline allowed us to make a narrow-band
search on Advanced LIGO O1 data for 11 known pulsars for some of which no updated ephemeris were available. The results of this analysis will be described
elsewhere \cite{simone:O1_narrow}.
We have also shown the possibility to extend this algorithm to directed searches of sources, like central compact objects in supernova remnants such as G353.6-0.7, for which the position is accurately known while the rotational parameters are uncertain. Despite with VSR4 data it was not possible to beat the spin-down limit, it will become possible using the advanced detectors data such as the first obervational run of LIGO (O1). Thus making narrow-band searches very important to study this type of astrophysical sources.
\section*{References}
|
\section{Introduction}
\citet{fanaroff74} introduced the first classification scheme for
extragalactic radio sources with large-scale structures (i.e., greater than
$\sim$15-20 kpc in size) based on the ratio $R_{FR}$ of the distance between
the regions of highest surface brightness on opposite sides of the central
host galaxy to the total extent of the source up to the lowest brightness
contour in the radio images. Radio sources with $R_{FR}<$0.5 were placed in
Class I (i.e., the edge-darkened FR~Is) and sources with $R_{FR}>$0.5 in Class
II (i.e., the edge-brightened FR~IIs). This morphology-based classification
scheme was found to be linked to their intrinsic power; Fanaroff and Riley
found that all sources in their sample with luminosity at 178 MHz smaller than
2$\times$10$^{25}$ W Hz$^{-1}$ sr$^{-1}$ (for a Hubble constant of 50
$\rm{\,km \,s}^{-1}$\ Mpc$^{-1}$) were classified as FR~I while the brighter sources all were
FR~II. The luminosity distinction between FR classes is fairly sharp at 178
MHz; their separation is even cleaner in an optical-radio luminosity plane,
implying that the FR~I/FR~II dichotomy depends on both optical and radio
luminosity \citep{ledlow96}.
The two Fanaroff-Riley classes do not instead correspond to a division from
the point of view of the optical spectroscopic properties of their hosts.
\citet{laing94} defined low and high excitation galaxies (LEGs and HEGs) based
on the ratios of the diagnostic optical emission lines in a scheme similar to
that adopted to distinguish LINERs and Seyferts in radio quiet AGN
(\citealt{kewley06}). While all FR~Is for which a reliable classification can
be obtained are LEGs, both LEGs and HEGs are found among the FR~IIs (e.g.,
\citealt{buttiglione10}). Buttiglione et al. find that LEGs and HEGs also differ
from other points of view. While LEGs do not show prominent broad lines,
they are observed in $\sim 30\%$ of HEGs; narrow emission lines are a factor
$\sim$ 10 brighter in HEGs than in LEGs at the same radio luminosity. Also, HEGs
show bluer colors than LEGs \citep{smolcic09}. \citet{baldi08} identify
compact knots in the UV images of 3C HEG radio galaxies, a morphological
evidence of recent star formation extending over 5$-$20 kpc; conversely, LEGs
hosts are usually red, passive galaxies. These results suggest that the radio
galaxies belonging to the two spectroscopic classes correspond to different
manifestation of the radio loud AGN phenomenon
\citep{hardcastle07,buttiglione10,best12}.
The recent multiwavelength large-area surveys are a unique tool to further
explore the connection between the morphological and spectroscopic classes of
radio galaxies, providing us with large samples of radio emitting AGN
extending to lower luminosities than in previous studies.
In \citet{capetti16} we created a catalog of 219 edge-darkened FR~I radio
galaxies called FRI{\sl{CAT}}. We found that the FRI{\sl{CAT}}\ hosts are remarkably homogeneous,
as they are all luminous red early-type galaxies (ETGs) with large black hole
masses and spectroscopically classified as LEGs. All these properties are
shared by the hosts of more powerful FR~Is in the 3C sample. They do not show
significant differences from the point of view of their colors with respect to
the general population of massive ETGs. The presence of an active nucleus (and
its level of activity) does not appear to affect the hosts of FR~Is.
We now extend this study to the population of edge-brightened FR~II radio
galaxies with the main aim of comparing the properties of FR~Is and FR~IIs
by also considering their spectroscopic classification.
This paper is organized as follows. In Sect.\ 2 we present the selection
criteria of the sample of FR~IIs. The radio and optical properties of the
selected sources are presented in Sect.\ 3. Sect.\ 4 is devoted to results
and conclusions.
Throughout the paper we adopt the same cosmology parameters used in
\citet{capetti16}, i.e., $H_0=67.8 \, \rm km \, s^{-1} \, Mpc^{-1}$,
$\Omega_{\rm M}=0.308$, and $\Omega_\Lambda=0.692$ \citep{ade16}.
For our numerical results, we use c.g.s. units unless stated
otherwise. Spectral indices $\alpha$ are defined in the usual convention on
the flux density, $S_{\nu}\propto\,\nu^{-\alpha}$. The SDSS magnitudes are in
the AB system and are corrected for the Galactic extinction; {\em WISE}
magnitudes are instead in the Vega system and are not corrected for extinction
since, as shown by, for example, \citet{dabrusco14}, such correction affects
mostly the magnitude at 3.4 $\mu$m of sources lying at low Galactic latitudes
(and by less than $\sim$3\%).
\section{Sample selection}
\label{sample}
We searched for FR~II radio galaxies in the sample of 18,286 radio sources
built by Best \& Heckman (2012; hereafter the BH12 sample) by limiting our
search to the subsample of objects in which, according to these authors, the
radio emission is produced by an active nucleus. They cross-matched the
optical spectroscopic catalogs produced by the group from the Max Planck
Institute for Astrophysics and Johns Hopkins University \citep{bri04,tre04}
based on data from the data release 7 of the Sloan Digital Sky Survey
(DR7/SDSS; \citealt{abazajian09}),\footnote{Available at {\tt
http://www.mpa-garching.mpg.de/SDSS/}.} with the National Radio Astronomy
Observatory Very Large Array Sky Survey (NVSS; \citealt{condon98}) and the
Faint Images of the Radio Sky at Twenty centimeters survey (FIRST;
\citealt{becker95,helfand15}) adopting a radio flux density limit of 5 mJy in
the NVSS. We focused on the sources with redshift $z < 0.15$.
We adopted a purely morphological classification based on the radio
structure shown by the FIRST images. We visually inspected the FIRST
images of each source and preserved those with an edge brightened
morphology in which at least one of the emission peaks lies at a
distance of at least 30 kpc from the center of the optical host. The
30 kpc radius corresponds to 11$\farcs$4 for the farthest objects; the
$z<0.15$ redshift limit ensures that all the selected sources are well
resolved with the 5$\arcsec$ resolution of the FIRST images. The three
authors performed this analysis independently and we included only the
sources for which a FR~II classification is proposed by at least two
of us. We allowed for the presence of diffuse emission leading to X-,
Z-, or C-shaped morphologies, but not extending at larger distances
with respect to the emission peaks, thus excluding wide angle tail
sources \citep{owen76}. Most of these sources are double, i.e., they
do not show nuclear radio emission; the lack of this precise position
reference requires a further check of the original optical
identifications. We discarded three objects in which the
identification of the host is not secure.
The resulting sample, to which we refer as FRII{\sl{CAT}}, is formed by 122 FR~IIs
whose FIRST images are presented in the Appendix. Their main properties
are presented in Table \ref{tab}, where we report the SDSS name, redshift, and
NVSS 1.4 GHz flux density (from BH12). The [O~III] line flux, the r-band SDSS
AB magnitude, $m_r$, the Dn(4000) index (see Section 3 for its definition),
and the stellar velocity dispersion $\sigma_*$ are instead from the MPA-JHU
DR7 release of spectrum measurements. The concentration index $C_r$ was
obtained for each source directly from the SDSS database. For sake of clarity,
uncertainties are not shown in the table; we estimated a median uncertainty of
0.09 on $C_r$, of 0.03 on Dn(4000), of 0.005 magnitudes on $m_r$, and of 10
$\rm{\,km \,s}^{-1}$\ on $\sigma_*$. We also list the resulting radio and line luminosity and
the black hole masses estimated from the stellar velocity dispersion and the
relation $\sigma_* - M_{\rm BH}$ of \citet{tremaine02}. The uncertainty in the
$M_{\rm BH}$ value is dominated by the spread of the relation used (rather
than by the errors in the measurements of $\sigma_*$) resulting in an
uncertainty of a factor $\sim$ 2. Finally, we give the classification (from
BH12) into LEGs and HEGs based on the optical emission line ratios in their
SDSS spectra.
\section{FRII{\sl{CAT}}\ hosts and radio properties}
\label{hosts}
\subsection{Hosts properties}
The majority (107) of the selected FR~IIs are classified as LEG, but there
are also 14 HEG and just one source that cannot be classified spectroscopically
because of the lack of emission lines, namely J1446+2142.
The distribution of absolute magnitude of the FRII{\sl{CAT}}\ hosts covers the range
$-20 \gtrsim M_r \gtrsim -24$ (see Fig. \ref{mhist}, left panel). The
distribution of black hole masses (Fig. \ref{mhist}, right panel) is rather
broad. Most sources have $8.0 \lesssim \log M_{\rm BH} \lesssim 9.0 M_\odot$,
but a tail toward smaller values, down to $M_{\rm BH} \sim10^{6.5} M_\odot$,
that includes $\sim 15\%$ (13 LEGs and 4 HEGs) of the sample.
The FR~II HEGs are less luminous, overall, and harbor less massive black holes
with respect to the FR~II LEGs; the medians of their distributions are
$<M_r({\rm HEG})> = -21.97 \pm 0.17$, $<M_r({\rm LEG})> = -22.62 \pm 0.06$,
$<\log M_{\rm BH}({\rm HEG})> = 8.21 \pm 0.11$, and $<\log M_{\rm BH}(LEG)> =
8.46 \pm 0.04$, respectively. The comparison between the FRII{\sl{CAT}}\ LEG sources
and the FRI{\sl{CAT}}\ hosts ($< M_r({\rm FR~I}) > = -22.69 \pm 0.03$ and $<\log M_{\rm
BH}({\rm FR~I})> = 8.55 \pm 0.02$) indicates that only marginal differences
(and not statistically significant) are present between the medians of these
distributions.
The Dn(4000) spectroscopic index, defined according to
\citet{balogh99} as the ratio between the flux density measured on the
red side and blue side of the Ca~II break, is an indicator of the
presence of young stars or of nonstellar emission. Low redshift
($z < 0.1$) red galaxies have Dn(4000)$= 1.98 \pm 0.05$, which is a
value that decreases to $= 1.95 \pm 0.05$ for $0.1 < z < 0.15$
galaxies \citep{capetti15}.
The concentration index $C_r$, which is defined as the ratio of the
radii including 90\% and 50\% of the light in the $r$ band, can be
used for a morphological classification of galaxies, in which
early-type galaxies have higher values of $C_r$ than late-type
galaxies. Two thresholds have been suggested to define ETGs: a more
conservative value at $C_r \gtrsim$ 2.86 \citep{nakamura03,shen03} and
a more relaxed selection at $C_r \gtrsim$ 2.6
\citep{strateva01,kauffmann03b,bell03}. \citet{bernardi10} found that
the second threshold of the concentration index corresponds to a mix
of E+S0+Sa types, while the first mainly selects ellipticals galaxies,
removing the majority of Sas, but also some Es and S0s.
In Fig. \ref{crdn} we show the concentration index $C_r$ versus the Dn(4000)
index (left panel) for the FRII{\sl{CAT}}\ sources. More than $\sim$90\% of the LEG
FR~II hosts lie in the region of high $C_r$ and Dn(4000) values, indicating
that they are red ETGs. The HEG FR~II are still ETGs, but they show generally
lower values of Dn(4000).
We also consider the $u-r$ color of the galaxies, obtained from SDSS
imaging and thus referred to the whole galaxy rather than just the
3\arcsec\ circular region covered by the SDSS spectroscopic aperture. In
Fig.~\ref{crdn} (central panel) we show the $u-r$ color versus the absolute
r-band magnitude $M_r$ of the hosts. As already found by considering the
Dn(4000) index, the FR~II HEGs show bluer color than the FR~II LEGs; in this
latter class, only two sources have a $u-r$ color that is smaller than the threshold
separating red and blue ETGs \citep{schawinski09}.
In Fig. \ref{crdn}, right panel, we show the comparison of the
{\em{WISE}} mid-IR colors of FRII{\sl{CAT}}\ and FRI{\sl{CAT}}\ sources; the associations
with the {\em{WISE}} catalog are computed by adopting a 3\farcs3
search radius \citep{dabrusco13}. All but one of the (J1121+5344)
FRII{\sl{CAT}}\ sources have a {\em{WISE}} counterpart, but 25 of these sources
are undetected in the $W3$ band. The LEG FR~IIs have mid-IR colors
similar to those of the FRI{\sl{CAT}}\ sources; their mid-IR emission is
dominated by their host galaxies since they fall in the same region of
elliptical galaxies \citep{wright10}. Only five LEGs have
$W2-W3 > 2.5$, exceeding the highest value measured for FR~Is.
Conversely, HEGs reach mid-IR colors as high as $W2-W3$=4.3, colors similar to
those of Seyfert and starburst galaxies (e.g., \citealt{stern05}). Their red
colors are likely due to a combination of star-forming regions and/or emission
from hot dust within a circumnuclear dusty torus.
Overall, we found 10 LEG FR~IIs whose properties do not conform with
the general behavior of their class, for example, showing blue colors
or being associated with small black holes. In some cases, this is due
to relatively large errors particularly in the measurement of
$\sigma_*$, a possible uncertain identification of their spectroscopic
class, or a substantial contribution from a bright nonthermal
nucleus. However, there are three objects (namely J0755+5204,
J1158+3006, and J1226+2538) for which we obtain estimates of the black
hole mass of $\log M_{\rm BH} \sim 6.5 - 6.8$; based on their $C_r$
value these objects are late-type galaxies and two of them also show
blue optical colors (and red mid-IR colors). These properties are all
typical of radio quiet AGN. This contrasts with the observed radio
power ($\log \nu L_r \sim 40.5 - 40.9$) and morphology.
\begin{figure*}
\includegraphics[width=9.5cm]{mrhist.ps}
\includegraphics[width=9.5cm]{mbhhist.ps}
\caption{Distributions of the $r$-band absolute magnitude (left) and black
hole masses (right). The black histograms are for FRII{\sl{CAT}}\ (cyan for the HEG
FRII{\sl{CAT}}\ subsample), the red histograms for the FRI{\sl{CAT}}. The FRI{\sl{CAT}}\ histograms are all scaled by the relative number of FR~I and FR~II, i.e., by 122/219.}
\label{mhist}
\end{figure*}
\begin{figure*}
\includegraphics[width=6.2cm]{crdn.epsi}
\includegraphics[width=6.2cm]{ur.epsi}
\includegraphics[width=6.2cm]{wise.epsi}
\caption{Left: concentration index $C_r$ vs. Dn(4000) index for the FRII{\sl{CAT}}\ and
FRI{\sl{CAT}}\ samples indicated by black and red dots, respectively. The HEG FR~II are
represented by cyan circles. Center: absolute $r$-band magnitude, $M_r$,
vs. $u-r$ color. Right: {\em{WISE}} mid-IR colors.}
\label{crdn}
\end{figure*}
\subsection{Radio properties}
\begin{figure*}
\includegraphics[width=9.5cm]{lrhist.ps}
\includegraphics[width=9.5cm]{ledlow.ps}
\caption{Left panel: radio luminosity distribution of the FRII{\sl{CAT}}\ (black, cyan
for the HEGs) and FRI{\sl{CAT}}\ sources (red). The dotted vertical line indicates the
transition power between FR~I and FR~II reported by \citet{fanaroff74}.
Right panel: radio luminosity (NVSS) vs. host absolute magnitude, $M_r$, for
FRII{\sl{CAT}}\ and FRI{\sl{CAT}}\ (black and red, respectively; cyan for the HEG FR~II). The
dotted line shows the separation between FR~I and FR~II reported by
\citet{ledlow96} to which we applied a correction of 0.34 mag to account for
the different magnitude definition and the color transformation between the
SDSS and Cousin systems.}
\label{lr}
\end{figure*}
The radio luminosity at 1.4 GHz of the FRII{\sl{CAT}}\ covers the range $L_{1.4}$ =
$\nu_{\rm r} l_{\rm r}$ $\sim 10^{39.5} - 10^{42.5}$ $\>{\rm erg}\,{\rm s}^{-1}$ (Fig. \ref{lr},
left panel), reaching a radio power almost two orders of magnitude lower than
the FR~IIs in the 3C sample. The HEGs are brighter than LEGs (with median of $\log
L_{1.4} = 41.37$ and 40.76, respectively) and LEGs are brighter than the
FRI{\sl{CAT}}\ sources by a factor $\sim$3; 90\% of the FRII{\sl{CAT}}\ fall below the
separation between FR~Is and FR~IIs originally reported by \citet{fanaroff74}
which translates, with our adopted cosmology and by assuming a spectral index
of 0.7 between 178 MHz and 1.4 GHz, into $L_{1.4} \sim 10^{41.6}$ $\>{\rm erg}\,{\rm s}^{-1}$.
Similarly, we find that $\sim$75\% of the FRII{\sl{CAT}}\ sources (and including also
four HEGs) are located {\sl below} the dividing line in the optical-radio
luminosity plane defined by \citet{ledlow96};\footnote{We shifted the dividing
line to the right of the diagram to include a correction of 0.12 mag to
scale our total host magnitude to the M$_{24.5}$ used by these authors, and
an additional 0.22 mag to convert the Cousin system into the SDSS system
\citep{fukugita96}.} see Fig. 3, right panel.
FR~IIs show a large spread in both radio and [O~III] line luminosities
(see Fig. \ref{lrlo3}). In this plane, the FR~II LEGs cover
essentially the same region of the FR~Is with just a tail toward
higher power both in line and in radio; no correlation is seen between
these two quantities. The FR~II HEGs generally have higher ratios
between $L_{\rm[O~III]}$ and $L_{1.4}$ than LEGs, which is an effect
already found in the 3C sample \citep{buttiglione10}. The HEGs in
FRII{\sl{CAT}}\ are mostly located above the correlation defined by the 3C
HEGs. A linear fit including both samples is indeed shallower (with a
slope of 0.91) than that obtained from the 3C sources alone (whose
slope is 1.15).
\subsection{Comparison with previous works}
As discussed in Sect. 2, we decided to maintain the traditional
morphological visual classification into edge-brightened FR~IIs and
edge-darkened FR~Is rather than adopting quantitative methods such as those
used by \citet{lin10}. The comparison of our classification with that
proposed by these authors indicates that, among the 96 sources in common,
$\sim 80\%$ of the FRII{\sl{CAT}}\ sources are classified as class {\sl a} in their
nomenclature, i.e., sources with two hot spots on either side of the
galaxy. Most of the remaining objects fall in class {\sl b}, in which the
emission peak is coincident with the host galaxy; however, the inspection of
these FR~II radio sources does not show any clear distinguishing feature
from those in the main class, other than having a relatively brighter
central source in addition to the two lobes.
The main drawback of our scheme, based on the subjective visual
inspection of radio images, is the relatively high fraction of
sources of uncertain classification. The rather strict criteria
adopted for a positive classification as FR~I in \citet{capetti16}
and here for the FR~II enabled us to select only 219 FR~Is and 122
FR~II; more than half of the 714 radio galaxies extended more than
30 kpc cannot be allocated to any FR class. On the other hand, this
strategy allows us to select samples that are very uniform from a
morphological point of view and that are optimally suited for our
main purpose, i.e., the comparison of the properties of the two
classes.
More recently, \citet{miraghaei17} performed an analysis on the same initial
sample, with an apparently similar selection strategy based on visual
inspection. However, the resulting sample of both FR~I and FR~II differ
significantly from those we obtained with only $\sim$ 25\% of objects in
common for both classes, even restricting the comparison to the same range of
redshift, $0.03 < z < 0.15$. This mismatch is likely due to the different
requirements (based, e.g., on linear instead of angular sizes and different
radio flux limits); most importantly they considered only sources
corresponding to multicomponents objects in FIRST and this rejects most of
the edge darkened sources we included in FRI{\sl{CAT}}. Overall, their results do not
strongly differ from ours, probably because (leaving aside the HEGs) the
properties of low z radio AGN are very homogeneous regardless of their radio
morphology. However, for example, we do not find significative differences in
the $C_r$ values between FRI and FRII hosts.
\section{Discussion and conclusions}
The properties of the FRII{\sl{CAT}}\ sources differ between those
spectroscopically identified as LEGs or HEGs. The HEGs have lower
optical luminosities, smaller black hole masses, and higher radio
luminosities with respect to LEGs, although a substantial overlap
between the two classes exist for all these quantities. The clearest
differences are related to the ratio between line and radio
luminosities and to their colors; HEGs are bluer in the optical and
redder in the mid-IR. These results confirm the conclusions of
previous studies (e.g.,
\citealt{baldi08,buttiglione10,baldi10b,best12}).
The population of the LEG FR~IIs included in the FRII{\sl{CAT}}\ is remarkably
uniform. They are all luminous red ETGs with large black hole masses ($M_{\rm
BH} \gtrsim 10^8 M_\odot$); only $\sim$ 10\% of the LEG\ FR\ IIs depart from
this general description. All these properties are shared with the hosts of
the FRI{\sl{CAT}}\ sources. The distributions of $M_{\rm BH}$ and $M_{r}$ differ with a
statistical significance higher than 95\% according to the Kolmogoroff-Smirnov
test; this is due to the presence of a tail of low $M_{\rm BH}$ sources,
reaching value as low as 3$\times 10^6 M_\odot$. However, the median of
$M_{\rm BH}$ and $M_{r}$ differ only marginally by less than 0.1 dex. Even
the median radio luminosity of LEG FR~II is just a factor $\sim$3 higher than
that measured in FR~I. Apparently, the difference in radio morphology between
edge-brightened and edge-darkened radio sources does not translate into a
clear separation between the nuclear and host properties, while the
spectroscopic classes, LEG and HEG, do.
The FRII{\sl{CAT}}\ sample unveils a population of FR~IIs of much lower radio
power with respect to those obtained at high radio flux thresholds,
extending it downward by two orders of magnitude. The correspondence
of the morphological classification of FR~I and FR~II with a
separation in radio power that is observed, for example, in the 3C
sample, disappears. This conclusion is in line with previous results
\citep{best09,lin10,wing11}. A radio source produced by a low power
jet can be edge brightened or edge darkened and the outcome is not due
to differences in the optical properties of the host galaxy.
Nonetheless, \citet{capetti16} find that the connection between radio
morphology and host properties is preserved in FR~Is; there is a
well-defined threshold of radio power above which an edge darkened
radio source does not form and this limit has a strong positive
dependence on the host luminosity. This effect was originally seen by
\citet{ledlow96} but partly lost in subsequent studies; we believe
that we recover it because of the stricter criteria we adopted for the
selection of FR~Is.
It can be envisaged that brighter galaxies are associated with denser and more
extended hot coronae that are able to disrupt more powerful jets. But the
large population of low power FR~IIs indicates that the situation is more
complex; there is a large overlap of radio power between FR~Is and FR~IIs and
radio power is believed to be a robust proxy for the jet power (e.g.,
\citealt{willott99,birzan04}). Apparently, jets of the same power expanding in
similar galaxies can form both FR~I and FR~II. This indicates that the optical
properties of the host and radio luminosity are not the only parameters
driving the evolution of low power radio sources. Further studies of, for example,
the X-rays properties and the larger scale environment are needed to clarify
this issue.
\begin{figure}
\includegraphics[width=9.5cm]{lrlo3.ps}
\caption{Radio (NVSS) vs. [O~III] line luminosity of the FRII{\sl{CAT}}\ (black),
FRI{\sl{CAT}}\ (red), 3C-FR~I (green), and 3C-FR~II samples (blue) The green line
(blue) shows the linear correlation between these two quantities derived
from the FR~Is (FR~IIs) of the 3C sample from \citep{buttiglione10}.
The dashed blue-cyan line is instead the linear fit on both the 3C
and the FRII{\sl{CAT}}\ HEGs.}
\label{lrlo3}
\end{figure}
\bibliographystyle{aa}
|
\section{Introduction}
There are many sensor technologies proposed for the use
in the vertex detector for experiments at the International
Linear Collider (ILC)~\cite{behnke}.
Fine pixel CCD (FPCCD) is one of the
candidates~\cite{sugimoto05}.
Due to its small pixel size of $\sim 5\ \mu$m,
FPCCD vertex detector can offer very good
impact parameter resolution and excellent two-track separation
capability. FPCCD sensors will be read out
in $\sim 200$~ms between bunch trains.
Because there is no beam crossing during the readout,
FPCCD sensor option is completely free from the RF noise
caused by the very short bunches of the beam.
One drawback of FPCCD sensors is relatively
poor radiation immunity, particularly large charge transfer
inefficiency (CTI) due to radiation induced trap levels.
CTI of FPCCD due to radiation damage is a function of temperature.
A simple simulation of CTI based on Shockley-Read-Hall theory
shows that around $-40^\circ$C is the optimal operating temperature
as shown in Figure~\ref{fig:hcti}.
To achieve this cooling temperature, two-phase CO$_2$ cooling system
seems most suited.
\begin{figure}
\centerline{\includegraphics[width=0.55\columnwidth]{hcti.eps}}
\caption{Results of a simple simulation of CTI with trap levels of 0.17~eV
and 0.42~eV based on Shockley-Read-Hall theory.
With fat-zero charge injection of 100 electrons and notch channel,
$\sim -40^\circ$C
is the optimum operation temperature.}
\label{fig:hcti}
\end{figure}
\section{Advantages of two-phase CO$_2$ cooling system}
In cooling systems using two-phase coolant,
the cooling temperature is controlled by controlling
the pressure of the two-phase coolant.
Because the heat load is consumed only for evaporation of the
coolant, the cooling temperature is constant along
the cooling tube.
Perfluorocarbon (PFC) such as C$_2$F$_6$ has been
also used as two-phase coolant for detector cooling systems,
for example ATLAS SCT~\cite{attree}.
Compared with PFC, CO$_2$ has several
advantages.
Table~\ref{tab:coolant} shows properties of CO$_2$ and
PFC.
Latent heat of CO$_2$ is much larger than that of PFC.
Because the pressure of two-phase CO$_2$ is higher than
that of PFC at the same temperature,
CO$_2$ has less temperature drop due to
pressure drop along the cooling tube, and less
evaporated gas volume than those of PFC. Thanks to
these properties, we can use thinner cooling tube
for two-phase CO$_2$ than PFC. Outer diameter of 2~mm
or less is good enough for the cooling tube of FPCCD
vertex detector.
Because heat sources of FPCCD vertex detector
locate mainly at both ends of ladders (on-chip amplifiers and
front-end ASICs), cooling at both ends of ladders
through the ladder base made of carbon fiber reinforced plastic
(CFRP) sheet and endplates on which cooling tube is attached
is an attractive solution.
The support shell including the end-plate is enclosed in a
cryostat made of heat insulating material.
The return line of CO$_2$ will be used to cool the electronics
(clock drivers and data compression circuits) placed
outside the cryostat.
The power consumption of the FPCCD vertex detector will be
less than 100~W inside the cryostat and about 200~W/side
for the electronics outside the cryostat.
Additional material budget
due to attached cooling tube of 2~mm$\phi$ made of
Titanium would be only 0.3\%$X_0$ if averaged over
the end-plate.
Gas cooling is a possible alternative.
However, if we try to cool this vertex detector with
cold air or nitrogen gas, the flow rate will be
quite large. As a consequence, vibration of
the ladders could be caused.
Much thicker transfer tube than two-phase CO$_2$ cooling
is necessary for gas cooling,
which requires more dead space between forward tracking discs (FTD)
and the beam pipe.
In addition, constant temperature cooling is almost impossible
in case of gas cooling.
\begin{table}
\centerline{\begin{tabular}{|l|l|l|l|}
\hline
& CO$_2$ & C$_2$F$_6$ & C$_3$F$_8$ \\
\hline
Latent heat @$-40^\circ$C [J/g] & 321 & $\sim 100$ & $\sim 110$ \\
Triple point [$^\circ$C] & $-56.4$ & $-97.2$ & $-160$ \\
Critical point [$^\circ$C] & 31.1 & 19.7 & 71.9 \\
Pressure @$-40^\circ$C [MPa] & 1 & $\sim 0.5$ & $\sim 0.1$ \\
Global warming potential & 1 & 12200 & 8830\\
\hline
\end{tabular}}
\caption{Properties of several kinds of two-phase coolant.}
\label{tab:coolant}
\end{table}
\section{Gas-compressor based cooling system}
Detector cooling systems using liquid pump for
circulation of two-phase CO$_2$ have been
developed by several groups~\cite{delli,beuzekom}.
In such a
``pump-based system'', CO$_2$ is liquefied
at temperature below the cooling temperature.
The temperature of circulating CO$_2$ is
below or at cooling temperature.
Therefore, very tight thermal insulation
is required for the whole system, including
the transfer tubes and the liquid pump.
An expensive low-temperature chiller is
necessary for liquefaction if the cooling temperature
as low as $-40^\circ$C has to be achieved.
A sophisticated ``two-phase accumulator''
has to be adopted for pressure (and temperature)
control of the two-phase CO$_2$ for the
pump-based system.
Our R\&D goal is to develop a two-phase CO$_2$
cooling system using a gas compressor,
rather than a liquid pump, for
circulation of CO$_2$.
Figure~\ref{fig:schema} shows
schematic diagrams of a pump-based system and
a gas-compressor based system.
In the gas-compressor based system,
CO$_2$ gas is liquefied by a condenser at near room temperature
after compression.
The liquid CO$_2$ is transferred to a
heat exchanger near the detector through a
transfer tube.
The long transfer tube between the liquefier plant
and the heat exchanger can be at near room temperature.
At the heat exchanger, the liquid CO$_2$ is cooled down
to the detector cooling temperature by the
returning two-phase CO$_2$.
Then, the pressure of the CO$_2$ is decreased
by a needle valve (or a capillary tube).
The cooling is achieved basically by the latent heat
(evaporation) of the returning
two-phase CO$_2$,
rather than the Joule-Thomson effect.
The two-phase CO$_2$ is completely evaporated
by a heater, and goes back to the liquefier plant.
The pressure of the two-phase CO$_2$ is controlled
by a back pressure valve in the liquefier plant.
The gas-compressor system has several advantages
over the pump system. Because CO$_2$ is liquefied
at near room temperature, we don't need an expensive
low temperature chiller. Cooling water near room temperature
is enough for the liquefaction. In case of ILC experiment,
such cooling water must be available in the detector hall.
We don't need strict thermal insulation for long transfer tubes
between the liquefier plant and the detector.
Flexible transfer tubes
off-the-shelf can be used for this purpose.
These tubes can be placed on the cable chain
supposed to be used for push-pull operation of
ILC detectors~\cite{behnke}.
\begin{figure}
\centerline{\includegraphics[width=0.75\columnwidth]{schema.eps}}
\caption{Schematic diagrams of a pump-based cooling system (top)
and a gas-compressor based cooling system (bottom).}
\label{fig:schema}
\end{figure}
\section{Construction of a prototype}
Several prototypes of the two-phase CO$_2$ cooling system
have been constructed at KEK~\cite{sugimoto12,sugimoto12b,sugimoto13}.
The latest prototype consists of three units;
a liquefier unit, a flow meter unit, and a cooling unit.
Figure~\ref{fig:prototype} shows
a simplified schematic diagram of the prototype system.
Three units are connected with metal-core flexible tubes
($1/4$ inch for liquid, $3/8$ inch for gas).
As a gas compressor, Haskel gas booster AGD-7 is used.
This compressor is a reciprocating type, and driven
by compressed air. The exhausted air is used for
cooling of gas cylinder of the compressor.
\begin{figure}
\centerline{\includegraphics[width=0.85\columnwidth]{prototype.eps}}
\caption{Schematic diagrams of the prototype cooling system.}
\label{fig:prototype}
\end{figure}
The phase diagram of the system in an ideal operating
condition is shown in
Figure~\ref{fig:phasediagram}.
The expected cooling power at $-40^\circ$C is 200~J/g.
The actual cooling power has been measured using dummy load
by looking at the dry-out point where the temperature of
CO$_2$ starts increasing.
The two-phase CO$_2$ with flow rate of 1.4~g/s
at cooling temperature of $-40^\circ$C
has dried out with 170~W dummy load power, which is
significantly less than the expected cooling power
of 280~W (=200~J/g$\times$1.4~g/s).
The reason of this deficit of the cooling power
is presumably heat load of the transfer tube
and other low temperature part.
This measurement was done at the ambient temperature
of $27^\circ$C, while the temperature of the liquid
CO$_2$ was $15^\circ$C.
\begin{figure}
\centerline{\includegraphics[width=0.98\columnwidth]{phasediagram.eps}}
\caption{Phase diagram (p-H curve) of the cooling system in an ideal operation
condition.}
\label{fig:phasediagram}
\end{figure}
Pressure drop in the metal core flexible tubes
was one of concerns because inner wall of the tubes
has corrugated structure. The pressure drop has
actually measured for several values of the flow rate.
The measured pressure drop is reasonably low.
At low flow rate ($< 1.4$~g/s), the pressure drop
is less than resolution of the digital pressure gauge.
At high flow rate ($> 2$~g/s), the pressure drop
is dominated by the flow meter for liquid,
and still less than the resolution of the pressure gauge
for gas.
\section{Further R\&D}
\subsection{Pressure control}
In the prototype cooling system, pressure of
the two-phase CO$_2$ is manually controlled by a
back-pressure valve. This method is however somewhat
unstable and time consuming for adjustment.
To overcome this disadvantage, we plan to
replace the manual back-pressure valve with
an automatic pressure controller.
We have tested a commercially available pressure
controller, Bronkhorst P-702CV. Using this controller,
the back pressure can be controlled by external
voltage setting.
We have studied this pressure controller
by inserting it in the return gas line between
the cooling unit and the liquefier unit.
The back-pressure valve in the liquefier unit
was set at the lowest pressure.
Figure~\ref{fig:pcontrol} shows the result of the
measurement of the setting voltage, pressure
and temperature of the two-phase CO$_2$,
and flow rate of CO$_2$.
It can be seen that quick and stable control
of the pressure and the temperature is
achieved.
\begin{figure}[t]
\centerline{\includegraphics[width=0.77\columnwidth]{pcontrol.eps}}
\caption{Setting voltage for the pressure controller,
and the response of pressure and temperature of two-phase CO$_2$
to the setting voltage. The flow rate is also shown.}
\label{fig:pcontrol}
\end{figure}
\subsection{O-ring material}
O-rings made of elastomer are used for the gas compressor
and safety valves in our system.
Degradation of the O-rings called as explosive decompression (ED)
was seen in these O-rings, and caused gas leak.
Explosive decompression, also called as rapid gas decompression,
is a mechanism of degradation in elastomer due to rapid
decompression of gaseous media.
At high pressure environment, CO$_2$ gas immigrates into
elastomer. When the pressure is reduced suddenly,
CO$_2$ dissolved inside the elastomer comes out
as micro bubbles, expands, and damage the
elastomer from inside.
In order to mitigate the risk of gas leak due to ED,
we can replace the safety valves with metal-seal
safety valves. For the gas compressor, we should
find out better material for O-rings.
We have constructed O-ring ED test apparatus,
and plan to test several kinds of O-rings with
different Shore durometer hardness and
different materials such as Kalrez, Chemraz,
and so on.
In the worst case, frequent overhaul of the gas compressor
would be the solution.
\subsection{Other R\&D issues}
In the present prototype system, a very massive stainless-steel plate
heat exchanger is used. We plan to develop a very low mass
heat exchanger which can be placed inside the detector.
As a candidate, a heat exchanger made of double-layer tube
will be studied.
The size of the liquefier unit for the present prototype
is quite large ($\sim$1~m$\times$1~m$\times$2~m).
We would like to develop a more compact liquefier unit.
The gas booster used in the liquefier is quite noisy.
Sound insulation should be implemented in the next prototype
of the liquefier unit.
The present prototype is controlled manually.
The remote control system using a programmable logic
controller (PLC) is one of the R\&D issues.
On the detector side, thermal contact between the cooling tube
and the end-plate of the vertex detector, and between
the end-plate and the ladders has to be studied.
\section{Summary}
We have successfully developed a prototype of
two-phase CO$_2$ cooling system using a
gas compressor for CO$_2$ circulation.
Cooling power of the system has been measured
at the cooling temperature of $-40^\circ$C,
and a satisfactory results have been obtained
for FPCCD vertex detector cooling.
On the other hand, degradation problem of
O-rings exists, and has to be solved.
There are still many R\&D issues to be
accomplished to realize the practical
cooling system.
\section*{Acknowledgments}
This work is supported by Grant-in-Aid No.24540312
and No.16H03992 by Japan Society for
Promotion of Science (JSPS).
This work is also supported by KEK
Detector Technology Project.
\begin{footnotesize}
|
\section{Introduction}
In potential theory the notion of subharmonic functions, $\mathcal{SH}$, is of fundamental importance, and in pluripotential theory the notion of plurisubharmonic functions, $\mathcal{PSH}$, is of the same importance. In 1985,
Caffarelli et al.~\cite{CHS} proposed a model that makes it possible to study the common properties of potential and pluripotential theories, as well as the transition between them. It also gives a splendid tool in geometric constructions. The core focus of the Caffarelli-Nirenberg-Spruck framework is what is known today as $m$-subharmonic functions, $\mathcal{SH}_m$. These functions are considered in
complex space, and on different types of complex manifolds. If $n$ is the underlying dimension, then it holds that
\[
\mathcal{PSH}=\mathcal{SH}_n \subset\cdots\subset \mathcal{SH}_1=\mathcal{SH}\, .
\]
To mention a few references related to the Caffarelli-Nirenberg-Spruck model~\cite{Blocki_weak, huang, Li, phong, WanWang}.
Let $\Omega\subset \mathbb C^n$ be a bounded domain, and let $f$ be a real-valued function defined on the topological boundary $\partial \Omega$. It is well-known that one can not always extend $f$ to the inside to a $m$-subharmonic function. This is not possible even in the cases $m=1$, and $m=n$. The aim of this paper is to find a characterization of the functions $f$ that have this classical extension property, but in the process we shall also be interested in when this extension can be approximated in neighborhoods of $\bar\Omega$. The first obstruction is that $\Omega$ is only
assumed to be a bounded domain. This does not yield a satisfying amount of $m$-subharmonic functions. Therefore, we assume
that there exists at least one non-constant and negative $m$-subharmonic function $\psi:\bar{\Omega}\to\RE$ such that for any $c\in\RE$ the set $\{x\in\Omega:\psi(x)<c\}$ is relatively compact in $\Omega$ (see Definition~\ref{def_msubkomp} for the meaning of being $m$-subharmonic on $\bar\Omega$). A bounded domain in $\C^n$ that satisfies this condition is called $P_m$-hyperconvex. More about this in Section~\ref{sec Pmhxdomains}.
Inspired by the work of Poletsky~\cite{Po,PO2}, and Poletsky and Sigurdsson~\cite{PS}, we use ideas from the theory of function algebras defined on a compact set. In the mentioned references, the authors use the beautiful and
intricate holomorphic disk-theory. Within the Caffarelli-Nirenberg-Spruck framework there are no Poletsy disks except in the case $m=n$. Therefore, we uses the idea of duality between functions and its corresponding Jensen measures. In Section~\ref{sec Jensen and msh}, we introduce and study necessary properties of $m$-subharmonic functions defined on a compact set in $\C^{n}$, and with the help of those results we arrive in Section~\ref{sec_extention} at the following theorem:
\bigskip
\noindent \textbf{Theorem~\ref{ext_in_pm_hyp}.} \emph{Let $\Omega$ be a bounded $P_m$-hyperconvex domain in $\mathbb C^n$, $1\leq m\leq n$, and let $f$ be a real-valued function defined on $\partial \Omega$. Then the following are equivalent:}
\medskip
\begin{enumerate}\itemsep2mm
\item \emph{there exists $F\in \mathcal{SH}_m(\bar \Omega)$ such that $F=f$ on $\partial \Omega$;}
\item $f\in \mathcal{SH}_m(\partial \Omega)$.
\end{enumerate}
\medskip
\noindent\emph{Furthermore, if $f$ is continuous on $\partial \Omega$, then the function $F$ can be chosen to be continuous on $\bar{\Omega}$.}
\bigskip
Theorem~\ref{ext_in_pm_hyp} is the first result of this kind within the Caffarelli-Nirenberg-Spruck model. It should be emphasized that
this is not only the classical Dirichlet problem that in the case $m=1$, can be traced back to the work of Brelot, Lebesgue, Perron, Poincar\'{e}, Wiener, and others. This since $F\in \mathcal{SH}_m(\bar \Omega)$, and by Theorem~\ref{cont}, these functions can be characterize by approximation
on neighborhoods of $\bar \Omega$. If $m=n$, then Theorem~\ref{ext_in_pm_hyp} was proved in~\cite{HP}.
A natural question that arises from Theorem~\ref{ext_in_pm_hyp} is how to decided wether a function $u$ is in $\mathcal{SH}_m(\bar \Omega)$ or not. From Theorem~\ref{thm_pmboundary} it follows that under the assumption that $\Omega\subset\C^n$ is a bounded open set, and that $u\in\mathcal{SH}_m(\bar \Omega)$, then $u\in \mathcal{SH}_m(\Omega)$, and $u\in\mathcal{SH}_m(\partial\Omega)$. The converse statement is not always true. But if we assume that $\Omega$ is $P_m$-hyperconvex, then we prove in Theorem~\ref{cor_msubbdvalue} that
\[
u\in\mathcal{SH}_m(\bar \Omega) \quad\Leftrightarrow\quad u\in\mathcal{SH}_m(\Omega) \text{ and } u\in\mathcal{SH}_m(\partial\Omega)\, .
\]
This justify further the study of the geometry of domains that admits a negative exhaustion function that belongs to $\mathcal{SH}_m(\bar \Omega)$. This is done in
Section~\ref{sec Pmhxdomains}. We end this note
with some concluding remarks on uniform approximation of $m$-subharmonic functions (Section~\ref{sec approx}).
\bigskip
Background information on potential theory can be found in~\cite{armitage,doob,landkof}, and for more information about pluripotential theory in~\cite{demailly_bok,K}. A beautiful treatise on subharmonic and plurisubharmonic functions is the monograph~\cite{hormander} written by H\"ormander. Definition and basic properties of $m$-subharmonic functions can be found in~\cite{SA}.
\bigskip
One concluding remark is in order. There are well-developed axiomatic, and algebraic, potential theories that could
have been deployed in connection with this paper. We have chosen not to do so, and leave it for others to
draw full benefits of these abstract models in order to learn more about the Caffarelli-Nirenberg-Spruck framework on compact sets. We want to mention the references~\cite{ArsoveLeutwiler, BlietnerHansen3, Constantinescu, Gamelin}.
\section{Jensen measures and $m$-subharmonic functions}\label{sec Jensen and msh}
In this section we will define the class of $m$-subharmonic function defined on a compact set, $X\subset \mathbb C^n$, and we will prove some properties of such functions. Among other things, we shall show that these functions are closely connected to approximation by $m$-subharmonic functions defined on strictly larger domains. But, first we need some notions and definitions. Let $\mathcal{SH}_m^o(X)$ denote the set of functions that are the restriction to $X$ of functions that are $m$-subharmonic and continuous on some neighborhood of $X$. Furthermore, let $\mathcal{USC}(X)$ be the set of upper semicontinuous functions defined on $X$. Next, we define a class of Jensen measures.
\begin{definition}\label{def_JzmK} Let $X$ be a compact set in $\C^n$, $1\leq m\leq n$, and let $\mu$ be a non-negative regular Borel measure defined on $X$ with $\mu(X)=1$. We say that $\mu$ is a \emph{Jensen measure with barycenter} $z\in X$ \emph{w.r.t.} $\mathcal{SH}_m^o(X)$ if
\[
u(z)\leq \int_{X} u \, d\mu \qquad\qquad \text{for all } u \in \mathcal{SH}_m^o(X)\, .
\]
The set of such measures will be denoted by $\mathcal{J}_z^m(X)$.
\end{definition}
\begin{remark}
If $X_1\subset X_2$ are compact sets in $\mathbb C^n$, then for every $z\in X_1$ it holds
\[
\mathcal{J}_z^m(X_1)\subset \mathcal{J}_z^m(X_2)\, .
\]
\end{remark}
We shall need the following convergence result in $\mathcal{J}_z^m(X)$. It is obtained in a standard way using the Banach-Alaoglu theorem, and therefore the proof is omitted.
\begin{theorem} \label{thm_convmeasures}
Let $X$ be a compact set in $\C^n$. Let $\{z_k\} \subset X$ be a sequence that is converging to $z$, as $k\to\infty$. For each $k$, let $\mu_k \in \mathcal{J}_{z_k}^m(X)$. Then there is a subsequence $\{\mu_{k_j}\}$, and a measure $\mu \in \mathcal{J}_z^m(X)$ such that $\mu_{k_j}$ converges weak-$^\ast$ to $\mu$.
\end{theorem}
Using the Jensen measures in Definition~\ref{def_JzmK} we shall now define what it means for a function to be $m$-subharmonic on a compact set.
\begin{definition} \label{def_msubkomp}
Let $X$ be a compact set in $\C^n$. An upper semicontinuous function $u$ defined on $X$ is said to be \emph{$m$-subharmonic on $X$}, $1\leq m\leq n$, if
\[
u(z) \leq \int_X u \, d\mu\, , \ \text { for all } \ z \in X \ \text { and all }\ \mu \in \mathcal{J}_z^m(X)\, .
\]
The set of $m$-subharmonic defined on $X$ will be denoted by $\mathcal{SH}_m(X)$.
\end{definition}
\begin{remark}
By definition, we see that $\mathcal{SH}_m^o(X) \subset \mathcal{SH}_m(X)$.
\end{remark}
It is easy to see that $m$-subharmonic functions on compact sets share a lot of basic properties with $m$-subharmonic functions on open sets. Some of these properties are listed below.
\begin{theorem} \label{thm_basicprop2}
Let $X$ be a compact set in $\mathbb C^n$, and $1\leq m\leq n$. Then
\medskip
\begin{enumerate}\itemsep2mm
\item if $u,v \in \mathcal{SH}_m(X)$, then $su+kv \in \mathcal{SH}_m(X)$ for $s,k\geq 0$;
\item if $u,v \in \mathcal{SH}_m(X)$, then $\max\{u,v\}\in \mathcal{SH}_m(X)$;
\item if $u_j\in \mathcal{SH}_m(X)$ is a decreasing sequence, then $u=\lim_{j\to \infty} u_j\in \mathcal{SH}_m(X)$, provided $u(z)>-\infty$ for some point $z\in X$;
\item if $u\in \mathcal{SH}_m(X)$ and $\gamma:\mathbb R\to\mathbb R$ is a convex and nondecreasing function, then $\gamma\circ u\in\mathcal{SH}_m(X)$.
\end{enumerate}
\end{theorem}
\begin{proof}
Properties $(1)$ and $(2)$ follows by Definition \ref{def_msubkomp}. To prove $(3)$, let $u_j\searrow u$. Then we
have that $u\in \mathcal{USC}(\bar\Omega)$. For $z\in X$, $\mu\in \mathcal J_z^m(X)$, we have by the monotone convergence theorem that
\[
u(z)=\lim_{j\to \infty}u_j(z)\leq \lim_{j\to \infty}\int u_j\,d\mu=\int\lim_{j\to \infty}d\mu=\int u\,d\mu\, .
\]
Part $(4)$ is a consequence of the Jensen inequality.
\end{proof}
The set $\mathcal{SH}_m^o(X)$ is a convex cone of continuous functions containing the constants, and separating points, and therefore we can apply the techniques of Choquet theory to get the following two versions of Edwards' duality theorem.
Generalizations of Edwards' Theorem can be found in~\cite{GogusPerkinsPoletsky}.
\begin{theorem}\label{thm_edwards}
Let $X$ be a compact subset in $\mathbb C^n$, $1\leq m\leq n$, and let $\phi$ be a real-valued lower semicontinuous function defined
on $X$. Then we have
\begin{enumerate}
\item[(a)]
\[
\sup\left \{\psi(z):\psi \in \mathcal{SH}_m^o(X) , \psi \leq \phi \right\}=\inf\left \{\int \phi \, d\mu : \mu \in \mathcal{J}_z^m(X)\right\}\, ,
\ \text {and}
\]
\item[(b)]
\begin{multline*}
\sup\left\{\psi (z) : \psi \in \mathcal{SH}_m(X)\cap \mathcal{C}(X), \psi \leq \phi\right\}\\=\sup\left\{\psi (z) : \psi \in \mathcal{SH}_m(X), \psi \leq \phi\right\}
=\inf\left\{\int \phi \, d\mu: \mu \in \mathcal{J}_z^m(X)\right\}\, .
\end{multline*}
\end{enumerate}
\end{theorem}
\begin{proof}
Part (a) is the direct consequence of Edwards' Theorem, and the proof of part (b) is postponed until after Theorem~\ref{cont} is proved.
\end{proof}
One important reason to study $m$-subharmonic functions on compact sets is that they are connected to approximation. In the case $m=1$, Theorem~\ref{cont} goes back to Debiard and Gaveau~\cite{DebiardGaveau}, and Bliedtner and Hansen~\cite{BlietnerHansen1,BlietnerHansen2}(see also~\cite{perkins,perkins2}). In the case $m=n$, part $(a)$, was shown by Poletsky in \cite{Po}, and part $(b)$ in~\cite{CHP} . In
Section~\ref{sec approx}, we shall have some concluding remarks in connection with this type of approximation.
\begin{theorem}\label{cont} Let $X \subset \C^n$ be a compact set, and $1 \leq m \leq n$.
\begin{itemize}\itemsep2mm
\item[$(a)$] Let $u\in \mathcal {USC}(X)$. Then $u \in \mathcal{SH}_m(X)\cap \mathcal{C}(X)$ if, and only if, there is a sequence $u_j \in \mathcal{SH}_m^o(X)$ such that $u_j \nearrow u$ on $X$.
\item[$(b)$] Then $u \in \mathcal{SH}_m(X)$ if, and only if, there is a sequence $u_j \in \mathcal{SH}_m^o(X)$ such that $u_j \searrow u$.
\end{itemize}
\end{theorem}
\begin{proof} \emph{Part $(a)$:} Let $u \in \mathcal{SH}_m(X)\cap \mathcal{C}(X)$. Since the Dirac measure $\delta_z$ is in $\mathcal{J}_z^m(X)$,
we have that
\[
u(z)=\inf\left\{\int u \, d\mu: \mu \in \mathcal{J}_z^m(X)\right\}\, .
\]
Theorem \ref{thm_edwards} part (a), yields now that
\[
u(z)=\inf\left \{\int u \, d\mu: \mu \in \mathcal{J}_z^m(X)\right\}=\sup\left\{\varphi(z): \varphi \in \mathcal{SH}_m^o(X), \varphi \leq u\right\}\, .
\]
Since the functions in $\mathcal{SH}_m^o(X)$ are continuous, Choquet's lemma (see e.g. Lemma~2.3.4 in~\cite{K}) says that
there exists a sequence $u_j \in \mathcal{SH}_m^o(X)$ such that $u_j \nearrow u$.
Now assume that there exists a sequence
$u_j \in \mathcal{SH}_m^o(X)$ such that $u_j \nearrow u$. Then $u$ can be written as the supremum of continuous functions.
Hence, $u$ is lower semicontinuous. Thus, $u$ is continuous. Let $z \in X$, and $\mu \in \mathcal{J}_z^m(X)$, then
\[
u(z)=\lim_j u_j(z)\leq \lim_j \int u_j \, d\mu =\int \lim_j u_j \, d\mu =\int u \, d\mu\, .
\]
By Definition~\ref{def_msubkomp} we know that $u \in \mathcal{SH}_m(X)$.
\bigskip
\emph{Part $(b)$:} First assume that $u$ is the pointwise limit of a decreasing sequence of functions $u_j \in \mathcal{SH}_m^o(X)$.
Then we have that $u\in \mathcal {USC}(X)$. Let $z \in X$ and $\mu \in \mathcal{J}_z^m(X)$, then it follows that
\[
u(z)=\lim_j u_j(z)\leq \lim_j \int u_j \, d\mu= \int\lim_j\,d\mu=\int u \, d\mu\, .
\]
Hence $u \in \mathcal{SH}_m(X)$.
For the converse, assume that $u \in \mathcal{SH}_m(X)$. We now want to show that there is a sequence of
functions $u_j \in \mathcal{SH}_m^o(X)$ such that $u_j \searrow u$ on $X$. We begin by showing that for every $f \in \mathcal{C}(X)$ with $u < f$
on $X$, we can find $v \in \mathcal{SH}_m^o(X)$ such that $u < v \leq f$. Let
\[
F(z)=\sup\{\varphi(z): \varphi \in \mathcal{SH}_m^o(X), \varphi \leq f\} \, .
\]
From Theorem \ref{thm_edwards} part (a) it follows now that
\[
F(z)=\inf\left\{\int f \, d\mu : \mu \in \mathcal{J}_z^m(X)\right\}\, .
\]
From the Banach-Alaoglu theorem we know that $\mathcal{J}_z^m(X)$ is weak-$^\ast$ compact, hence for all $z \in X$ we can find
$\mu_z \in \mathcal{J}_z^m(X)$ such that
\[
F(z)=\int f \, d\mu_z\, .
\]
We have
\[
F(z)=\int f \, d\mu_z > \int u \, d\mu_z \geq u(z)\, .
\]
Hence, $u < F$. By the construction of $F$ we know that for every given $z \in X$, there exists a function $v_z \in \mathcal{SH}_m^o(X)$ such
that $v_z \leq F$ and $u(z)<v(z) \leq F(z).$ Since the function $u-v_z$ is upper semicontinuous, then the set
\[
U_z=\{y\in X: u(y)-v_z(y)<0\}
\]
is open in $X$. By assumption $X$ is compact, and therefore there are finitely many points $z_1,\ldots,z_k$ with corresponding functions
$v_{z_1},\ldots,v_{z_k}$, and open sets $U_{z_1},\ldots, U_{z_k}$, such that $u < v_{z_j}$ on $U_{z_j}$. Furthermore,
\[
X=\bigcup_{j=1}^k U_{z_j}\, .
\]
The function $v=\max\{v_{z_1},\ldots,v_{z_k}\}$ is in $\mathcal{SH}_m^o(X)$, and $u < v \leq f$. We are now ready to prove that $u$ can be
approximated as in the statement in the theorem. The function $u$ is upper semicontinuous, and therefore it can be approximated with a decreasing
sequence $\{f_j\}$ of continuous functions. We can then find $v_1 \in \mathcal{SH}_m^o(X)$ such that $u < v_1 \leq f_1$. If we now assume that we
can find a decreasing sequence of functions $\{v_1,\ldots,v_k\}$ such that $v_j \in \mathcal{SH}_m^o(X)$, and $u < v_j$ for $j=1,\ldots,k$, then we can
find a function $v_{k+1} \in \mathcal{SH}_m^o(X)$ such that $u<v_{k+1}$ and $v_{k+1} \leq \min\{f,v_k\}$. Now the conclusion of the theorem follows by
induction.
\end{proof}
\begin{remark} In Theorem~\ref{cont} part $(a)$ we have uniform approximation on $X$. One can assume that the decreasing sequence in Theorem~\ref{cont}
part $(b)$ is smooth. This follows from a standard diagonalization argument.
\end{remark}
\begin{proof}[Proof of Theorem~\ref{thm_edwards} part (b).]
Let us define the following families of probability measures defined on $X$
\[
\begin{aligned}
&\mathcal{M}_z^m(X)=\left \{\mu: \, u(z)\leq \int u\, d\mu\, , \,\, \forall u\in \mathcal{SH}_m(X)\cap \mathcal C(X)\right \}, \\
&\mathcal{N}_z^m(X)=\left \{\mu: \, u(z)\leq \int u\, d\mu\, , \,\, \forall u\in \mathcal{SH}_m(X) \right \}\, . \\
\end{aligned}
\]
We have
\[
\mathcal{N}_z^m(X)\subset \mathcal{M}_z^m(X)\subset \mathcal{J}_z^m(X)\, ,
\]
since $\mathcal{SH}_m^o(X)\subset\mathcal{SH}_m(X)\cap \mathcal C(X)\subset \mathcal{SH}_m(X)$. On the other hand, let $z\in X$, $\mu \in \mathcal{J}_z^m(X)$, and let $\varphi\in \mathcal{SH}_m(X)$, then by Theorem~\ref{cont} part (b) there exists a decreasing sequence $u_j\in \mathcal{SH}_m^o(X)$ such that $u_j\to \varphi$, when $j\to \infty$, and then we have
\[
\varphi(z)=\lim_{j\to \infty}u_j(z)\leq \lim_{j\to \infty}\int u_j\,d\mu=\int\lim_{j\to \infty}u_j\,d\mu=\int\varphi\, d\mu\, .
\]
Hence, $\mu\in \mathcal{N}_m(X)$, and therefore $\mathcal{J}_z^m(X)\subset \mathcal{N}_z^m(X)$.
\end{proof}
A direct consequence of Theorem~\ref{cont} part $(b)$ is the following corollary.
\begin{corollary}\label{cor}
If $X_1\subset X_2$ are compact sets in $\mathbb C^n$, then $\mathcal{SH}_m(X_2)\subset \mathcal{SH}_m(X_1)$.
\end{corollary}
\begin{proof}
To see this take $u\in \mathcal{SH}_m(X_2)$, then by Theorem~\ref{cont} part $(b)$ there exists a sequence $u_j\in \mathcal{SH}_m^o(X_2)$ decreasing to $u$. Since $u_j$ belongs also to $\mathcal{SH}_m^o(X_1)$ then $u\in \mathcal{SH}_m(X_1)$.
\end{proof}
In Theorem~\ref{localization_Pm}, we shall need the following localization theorem. The case $m=n$ is Gauthier's localization theorem from~\cite{Gauthier}. For the proof of the following theorem, and later sections we need to recall the following definition. A function
$u$ is said to be \emph{strictly $m$-subharmonic} on $\Omega$ if for every $p \in \Omega$ there exists a constant $c_p >0$ such that $u(z)-c_p |z| ^2$ is $m$-subharmonic in a neighborhood of $p$.
\begin{theorem}\label{localization}
If $X \subset \C^n$ is a compact set, then $u \in \mathcal{SH}_m(X)\cap \mathcal{C}(X)$ if, and only if, for each $z \in X$, there is a neighborhood
$B_z$ such that $u |_{X\cap \bar B_z}\in \mathcal{SH}_m(X\cap \bar B_z)\cap \mathcal{C}(X\cap \bar B_z)$.
\end{theorem}
\begin{proof} This proof is inspired by~\cite{Gauthier}. First we see that the restriction of a function $u \in \mathcal{SH}_m(X)\cap \mathcal C(X)$ to $X \cap \bar B$ is $m$-subharmonic on that set. This follows from Corollary~\ref{cor}. Now we show the converse statement. Since $X$ is compact there exists a finite open covering $\{B_j\}$ of $X$. Assume that $u|_{X \cap \bar B_j} \in \mathcal{SH}_m(X \cap \bar B_j)\cap \mathcal C(X \cap \bar B_j)$ for all $j$. For every $j$, we can find compact sets $K_{j,k}$ such that $K_{j,k} \subset B_k$ and
\[
\partial B_j \cap X \subset \bigcup_{k \neq j}K_{j,k}\, .
\]
Let $K_k = \bigcup_j K_{j,k}$, and note that $K_k \subset B_k$. Set
\[
d_k = \mbox{dist}(K_k, \partial B_k)\, .
\]
For every $k$ there exists a function $\chi_k$ that is smooth on $\C^n$, $-1 \leq \chi_k \leq 0$, $\chi_k(z)=0$ when $\mbox{dist}(z, K_k) \leq \frac{d_k}{2}$, and $\chi_k=-1$ outside of $B_k$. Choose an arbitrary constant $c > 0$. The function $|z|^2$ is strictly $m$-subharmonic, so there exists a constant $\eta^0_k>0$ such that for every $0 < \eta_k<\eta_k^0$, the function $\eta_k \chi_k +c|z|^2$ is $m$-subharmonic and continuous on an open set $V_k$, $B_k \Subset V_k$. Choose a sequence $\{\varepsilon_j\}$ of positive numbers such that
\begin{equation}\label{eq:etaepsilon}
2 \max_{z \in \bar B_j} \varepsilon_j < \min_{z \in \bar B_j} \eta_j\, ,
\end{equation}
for every $z \in X$. The reason for this will be clear later. By the assumption that $u|_{X \cap \bar B_j} \in \mathcal{SH}_m(X \cap \bar B_j)\cap \mathcal C(X \cap \bar B_j)$ for every $j$, Theorem~\ref{cont} part (b) says that there exist open sets $U_j$, $(X \cap B_j) \Subset U_j \Subset V_j$ and functions $u_j \in \mathcal{SH}_m(U_j)\cap \mathcal C(U_j)$ such that
\begin{equation} \label{eq:epsilon_j}
|u-u_j|<\varepsilon_j \ \text{on} \ X \cap \bar B_j\, .
\end{equation}
For $z \in (U_j \setminus X)\cup (X\cap \bar B_j)$ set
\[
f_j(z)=u_j(z) + \eta_j \chi_j(z) + c|z|^2\, ,
\]
and elsewhere set $f_j=-\infty.$ Now define the function
\[
v(z)=\max_j f_j(z)\, .
\]
It remains to show that $v$ approximates $u$ uniformly on $X$, and that $v \in \mathcal{SH}_m^o(X)$. For $z \in X$ we have
\begin{multline}\label{exp}
|u(z)-v(z)| = |u(z)-\max_j f_j(z)|\\
= \left|u(z)-\max_{z \in X\cap \bar B_j}(u_j(z)+\eta_j \chi_j + c|z|^2)\right|\\
\leq \max_{z \in X\cap \bar B_j} \eta_j + |u(z)-\max_{z \in X\cap \bar B_j} u_j(z)| + c|z|^2\, .
\end{multline}
By choosing the constants $c, \eta_j, \varepsilon_j$ in the right order and small enough, then the right-hand side of~(\ref{exp}) can be made arbitrary small. Hence, $v$ approximates $u$ uniformly on $X$.
To prove that $v \in \mathcal{SH}_m^o(X)$, first take $z \in X$ that does not lie on the boundary of any $B_j$. The functions $f_k$, that are not $- \infty$ at $z$, are finitely many and they are continuous and $m$-subharmonic in a neighborhood of $z$. If $z \in \partial B_j \cap X$, then there exists a $k$ such that $z \in (X \cap K_k) \subset (X \cap B_k)$. For this $j$ and $k$ we have
\begin{multline*}
f_j(z)=u_j(z)+\eta_j \chi_j + c|z|^2
= u_j(z) - \eta_j +c|z|^2\\
=\bigl(u_j(z)-u_k(z)\bigr)+\bigl(u_k(z)+\eta_k 0+c|z|^2\bigr) - \eta_j\\
=f_k(z)+\bigl(u_j(z)-u_k(z)\bigr)-\eta_j \leq f_k(z)
\end{multline*}
where the last inequality follows from assumption (\ref{eq:etaepsilon}) together with (\ref{eq:epsilon_j}) (that makes sure that $|u_j(z)-u_k(z)|<\varepsilon_j+ \varepsilon_k$). This means that locally, near $z$, we can assume that the function $v$ is the maximum of functions $f_k$, $k \neq j$, where the functions $f_k$ are continuous and $m$-subharmonic in a neighborhood of $z$. This concludes the proof.
\end{proof}
As an immediate consequence we get the following gluing theorem for $m$-sub\-har\-mo\-nic functions on compact sets.
\begin{corollary}
Let $\omega \Subset \Omega$ be open sets, let $u\in \mathcal{SH}_m(\bar \omega)\cap \mathcal C(\bar \omega)$, $v\in \mathcal{SH}_m(\bar\Omega)\cap\mathcal C(\bar\Omega)$ and $u(z)\leq v(z)$ for $z\in \partial \omega$. Then the function
\[
\varphi=\begin{cases}
v, \, \text { on } \, \bar\Omega\setminus \omega,\\
\max\{u,v\}, \; \text { on } \, \omega,
\end{cases}
\]
belongs to $\mathcal{SH}_m(\bar \Omega)\cap\mathcal C(\bar \Omega)$.
\end{corollary}
\begin{proof}
Let $\varepsilon >0$ and define
\[
\varphi_{\varepsilon}=\begin{cases}
v+\varepsilon, \, \text { on } \, \bar\Omega\setminus \omega,\\
\max\{u,v+\varepsilon\}, \; \text { on } \, \omega.
\end{cases}
\]
Then by Theorem~\ref{localization} we get that $\varphi_{\varepsilon}\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal C(\bar \Omega)$ and $\varphi_{\varepsilon}\searrow \varphi$, as $\varepsilon \to 0$. By Theorem~\ref{thm_basicprop2} we conclude that $\varphi\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal C(\bar \Omega)$.
\end{proof}
Let us now look at a bounded domain $\Omega$ in $\mathbb C^n$. We want to investigate what the connection is between $\mathcal{SH}_m(\bar \Omega)$ and $\mathcal{SH}_m(\Omega)$. It is easy to show that $\mathcal{SH}_m(\bar \Omega) \subset \mathcal{SH}_m(\Omega)$. Using Definition \ref{def_msubkomp} we know that a function $\varphi \in \mathcal{SH}_m(\Omega)\cap \mathcal{USC}(\bar \Omega)$ is in $\mathcal{SH}_m(\bar \Omega)$ if $\varphi(z) \leq \int \varphi \, d\mu$ for all $\mu \in \mathcal{J}_z^m(\bar \Omega)$ where $z \in \bar \Omega$. In the same way as in \cite{HP} we can show that it is enough to look at the measures in $\mathcal{J}_z^m(\bar \Omega)$ for $z \in \partial \Omega$.
\begin{theorem} \label{thm_pmboundary}
Let $\Omega$ be a bounded open set in $\C^n$, and $1 \leq m \leq n$.
\medskip
\begin{enumerate}\itemsep2mm
\item[$(1)$] If $\varphi \in \mathcal{SH}_m(\bar \Omega)$, then $\varphi \in \mathcal{SH}_m(\Omega)$ and $\varphi \in \mathcal{SH}_m(\partial\Omega)$.
\item[$(2)$] If $\varphi \in \mathcal{SH}_m(\Omega)\cap \mathcal{USC}(\partial \Omega)$, and
\[
\varphi(z) \leq \int \varphi \, d\mu\, , \ \text { for all } \ z \in \partial\Omega \ \text { and all }\ \mu \in \mathcal{J}_z^m(\bar \Omega)\, ,
\]
then $\varphi \in \mathcal{SH}_m(\bar \Omega).$
\end{enumerate}
\end{theorem}
\begin{proof}
\emph{Part} $(1):$ By Theorem~\ref{cont} part $(b)$ there exists a sequence $\varphi_j\in \mathcal{SH}_m^o(\bar \Omega)$ decreasing to $\varphi$.
Then $\varphi_j\in \mathcal{SH}_m(\Omega)$, so $\varphi \in \mathcal{SH}_m(\Omega)$. The fact that
$\varphi \in \mathcal{SH}_m(\partial\Omega)$ follows from Corollary~\ref{cor}.
\bigskip
\emph{Part} $(2):$ By Theorem \ref{cont} part $(b)$ we want to prove that there is a decreasing sequence of functions
$\varphi_j$ in $\mathcal{SH}_m^o(\bar \Omega)$ such that $\varphi_j \rightarrow \varphi$ on $\bar \Omega$. Since $\varphi$
is upper semicontinuous we can find $\{u_j\} \subset \mathcal{C}(\bar \Omega)$ such that $u_j \searrow \varphi$ on $\bar \Omega$.
We are going to show that we can find functions $\{v_j\} \in \mathcal{SH}_m^o(\bar \Omega)$ such that $v_j \leq u_j$ and $v_j(z) \searrow \varphi(z)$
for every $z \in \partial \Omega$. From this it will follow that the functions
\[
\varphi_j=\begin{cases}
\max\{\varphi(z),v_j(z)\} & \text{if} \ z \in \bar \Omega\\
v_j(z) & \text{otherwise}\\
\end{cases}
\]
will belong to $\mathcal{SH}_m^o(\bar \Omega)$, and $\varphi_j \searrow \varphi$ on $\bar \Omega$.
To construct the approximating sequence $\{v_j\}$ define first
\[
F_j(z):=\sup\left \{v(z): v \in \mathcal{SH}_m^o(\bar \Omega), v \leq u_j\right \}=\inf\left \{\int u_j \, d\mu: \mu \in \mathcal{J}_z^m(\bar \Omega)\right\}.
\]
Since $\mathcal{J}_z^m(\bar \Omega)$ is compact in the weak$^\ast$-topology we can, for all $z\in \bar \Omega$ find $\mu_z \in \mathcal{J}_z^m(\bar \Omega)$ such that $F_j(z)=\int u_j \, d\mu_z$. We know, by the construction of $F_j$, that $F_j \leq u_j$, and
\[
F_j(z)=\int u_j \, d\mu_z > \int \varphi \, d\mu_z \geq \varphi(z) \ \text{ for all }z \in \partial \Omega\, .
\]
By the construction of $F_j$ we know that for every $z\in \partial \Omega$ we can find $v_z \in \mathcal{SH}_m^o(\bar \Omega)$ such that $v_z \leq F_j$ and $\varphi(z)< v_z(z)\leq F_j(z)$. The function $\varphi - v_z$ is upper semicontinuous and therefore the set
\[
U_z=\{w \in \partial \Omega: \varphi(w)-v_z(w)<0\}
\]
is open in $\partial \Omega$. It now follows from the compactness of $\partial \Omega$ that there are finitely many points $z_1,\ldots,z_k$ with corresponding functions $v_{z_1},\ldots,v_{z_k}$ and open sets $U_{z_1},\ldots,U_{z_k}$ such that $\varphi < v_{z_j}$ in $U_{z_j}$ and $\partial \Omega=\cup_{j=1}^kU_{z_j}$. The function $v_j=\max\{v_{z_1},\ldots,v_{z_k}\}$ belongs to $\mathcal{SH}_m^o(\bar \Omega)$ and $\varphi(z)< v_j(z) \leq u_j(z)$ for $z \in \partial \Omega$. This completes the proof.
\end{proof}
\section{$P_m$-hyperconvex domains}\label{sec Pmhxdomains}
Assume that $\Omega\subset\C^n$ is a bounded open set, and let $1 \leq m \leq n$. Theorem~\ref{cont}
give rise to the question of how to decide if $u$ is in $\mathcal{SH}_m(\bar \Omega)$.
From Theorem~\ref{thm_pmboundary} it follows that if $u\in\mathcal{SH}_m(\bar \Omega)$, then
$u\in \mathcal{SH}_m(\Omega)$, and $u\in\mathcal{SH}_m(\partial\Omega)$. The converse statement is not true, not even
under the assumption that $\Omega$ is $m$-hyperconvex (see Definition~\ref{def_mhx}). But if we assume that $\Omega$ admits a negative exhaustion function in $\mathcal{SH}_m(\bar \Omega)$ (notice here that $\bar\Omega$ is a compact set), then we shall prove
in Theorem~\ref{cor_msubbdvalue} that
\[
u\in\mathcal{SH}_m(\bar \Omega) \quad\Leftrightarrow\quad u\in\mathcal{SH}_m(\Omega) \text{ and } u\in\mathcal{SH}_m(\partial\Omega)\, .
\]
First we shall recall the definition of a $m$-hyperconvex domain.
\begin{definition}\label{def_mhx}
Let $\Omega$ be a domain in $\C^n$, and $1 \leq m \leq n$. We say that $\Omega$ is \emph{$m$-hyperconvex} if it admits an exhaustion function that is negative and in $\mathcal{SH}_m(\Omega)$.
\end{definition}
Let us now make a formal definition of $P_m$-hyperconvex domains.
\begin{definition}\label{def_Pmhx}
Let $\Omega$ be a domain in $\C^n$, and let $1 \leq m \leq n$. We say that $\Omega$ is \emph{$P_m$-hyperconvex} if it admits an exhaustion function that is negative, and in $\mathcal{SH}_m(\bar \Omega)$.
\end{definition}
From Theorem~\ref{thm_pmboundary} it follows that a $P_m$-hyperconvex domain is also $m$-hyper\-con\-vex. The
converse is not true. The case $m=n$ was studied in~\cite{HP}, and discussed in~\cite{PW}. A $P_n$-hyperconvex domain is $P_m$-hyperconvex for every $m=1,\ldots,n$, and as observed in~\cite{HP}, the notion of $P_n$-hyperconvexity is strictly weaker than the notion of \emph{strict hyperconvexity} that has been studied and used by
for example Bremermann~\cite{bremermann}, and Poletsky~\cite{PO3}. Furthermore, a $P_m$-hyperconvex domain is fat in the sense $\Omega=(\bar{\Omega})^{\circ}$.
It is straight forward to see that if $\Omega_1$ and $\Omega_2$ are $P_m$-hyperconvex domains in $\C^n$, then
$\Omega_1 \cap \Omega_2$ is $P_m$-hyperconvex in $\C^n$, and $\Omega_1 \times \Omega_2$ is $P_m$-hyperconvex in $\mathbb C^{2n}$.
\bigskip
As in the case of $m$-hyperconvex domains, we have in Theorem~\ref{thm_charPmhx} several nice characterizations of $P_m$-hyperconvex domains in terms of the barrier functions, and Jensen measures. The property that a domain is (globally) $P_m$-hyperconvex if, and only if, it is locally $P_m$-hyperconvex we leave to Theorem~\ref{localization_Pm}.
\begin{theorem}\label{thm_charPmhx}
Let $\Omega$ be a bounded domain in $\C^n$. Then the following assertions are equivalent:
\medskip
\begin{enumerate}\itemsep2mm
\item $\Omega$ is $P_m$-hyperconvex in the sense of Definition~\ref{def_Pmhx};
\item $\Omega$ admits a negative exhaustion function that is in $\mathcal{SH}_m(\bar \Omega)\cap \mathcal{C}(\bar \Omega)$;
\item $\partial \Omega$ has a weak barrier at every point $z_0\in\partial \Omega$ that is in $\mathcal{SH}_m(\bar \Omega)$, i.e. there exists a function $u\in\mathcal{SH}_m(\bar \Omega)$, such that $u<0$ on $\Omega$ and
\[
\lim_{z\to z_0\atop z\in\Omega} u(x)=0\, ;
\]
\item for every $z\in \partial \Omega$, and every $\mu \in \mathcal{J}_z^m(\bar \Omega)$, we have that $\operatorname{supp} (\mu) \subseteq \partial \Omega$;
\item $\Omega$ admits a continuous negative exhaustion function which is $m$-subharmonic on $\bar \Omega$, smooth and strictly $m$-subharmonic on $\Omega$.
\end{enumerate}
\end{theorem}
\begin{proof}
$(1) \Rightarrow (4):$ Assume that $\Omega$ is $P_m$-hyperconvex, then there exists a negative exhaustion function $\psi \in \mathcal{SH}_m(\bar \Omega)$. Take $z \in \partial \Omega$ and let $\mu \in \mathcal{J}_z^m(\bar \Omega)$, then
\[
0=\psi(z)\leq \int \psi \, d\mu \leq 0\, .
\]
Since $\psi <0 $ on $\Omega$, we have that $\operatorname{supp}(\mu) \subseteq \partial \Omega$.
\medskip
$(2) \Rightarrow (1):$ Follows by Definition \ref{def_Pmhx}.
\medskip
For the implications $(4)\Rightarrow (3)$, $(4) \Rightarrow (2)$, and $(4)\Rightarrow (1)$, assume that for all $w \in \partial \Omega$, the every measures $\mu \in \mathcal{J}_w^m(\bar \Omega)$ satisfy $\operatorname{supp} (\mu) \subseteq \partial \Omega$. Let $z\in \Omega$, $r>0$ be such that $B(z,r)\Subset \Omega$ and let
\[
u(z)= \sup\{\varphi(z): \varphi \in \mathcal{SH}_m(\bar\Omega)\cap \mathcal{C}(\bar \Omega), \varphi \leq 0, \varphi \leq -1 \ \text{on} \ B(z,r)\}\, .
\]
Then $u$ is lower semicontinuous, and by Theorem~\ref{thm_edwards} part (b), we have that
\[
u(z)=\inf\left\{\int -\chi_{B(z,r)} \, d\mu : \mu \in \mathcal{J}_z^m(\Omega)\right\}=-\sup\left\{\mu(B(z,r)): \mu \in \mathcal{J}_z^m(\Omega)\right\}\, .
\]
We shall prove that $\lim_{\xi\to \partial \Omega}u(\xi)=0$. Assume the contrary, i.e. that there is a point $z \in \partial \Omega$ such that $\liminf_{\xi \rightarrow z}u(\xi)<0$. Then we can find a sequence $z_n \rightarrow z$ such that $u(z_n)<-\varepsilon$ for every $n$. We can find corresponding measures $\mu_n \in \mathcal{J}_{z_n}^m(\bar\Omega)$ such that $\mu_n(B(z,r))>\varepsilon$. By Theorem~\ref{thm_convmeasures} we can (by passing to a subsequence) assume that $\mu_n$ converges weak-$^\ast$ to a measure $\mu \in \mathcal{J}_z^m(\bar \Omega)$. Then, using Lemma~2.3 in \cite{CCW}, we have that
\[
\mu(\overline {B(z,r)})=\int \chi_{\overline {B(z,r)}} \, d\mu \geq \limsup_{n \rightarrow \infty} \int \chi_{\overline {B(z,r)}} \, d\mu_n =\limsup_{n \rightarrow \infty} \mu_n(\overline {B(z,r)}) >\varepsilon\geq 0.
\]
This contradicts the assumption that $\mu \in \mathcal{J}_z^m(\bar \Omega)$ only has support on the boundary. It remains to show that $u \in \mathcal{SH}_m(\bar \Omega)\cap \mathcal{C}(\bar \Omega)$.
We have that $u^*\in \mathcal {SH}_m(\Omega)\cap \mathcal {USC}(\bar \Omega)$ and $\lim_{\xi\to \partial \Omega} u^*(\xi)=0$ so by the generalized Walsh theorem (Proposition~3.2 in~\cite{Blocki_weak}) we get that $u^*\in \mathcal C(\bar\Omega)$. This means that $u=u^*$ and $u$ is a continuous function. Finally Theorem~\ref{thm_pmboundary} gives us that $u\in \mathcal {SH}_m(\bar \Omega)$. Note that $u$ is a continuous exhaustion function for $\Omega$.
\medskip
$(3) \Rightarrow (4):$ Let $z \in \partial \Omega$ and assume that there exists a function $\varphi \in \mathcal{SH}_m(\bar \Omega)$, $\varphi \neq 0$ such that $\varphi \leq 0$ and $\varphi(z)=0$. Let $\mu \in \mathcal{J}_z^m(\bar \Omega)$, then
\[
0=\varphi(z)\leq \int \varphi \, d\mu \leq 0\, .
\]
Hence $\operatorname{supp}(\mu) \subseteq \partial \Omega$.
\medskip
The proof of equivalence (1)$\Leftrightarrow$(5) is postponed to Corollary~\ref{smooth}.
\end{proof}
\section{An extension theorem}\label{sec_extention}
In this section we shall prove the extension theorem discussed in the introduction (Theorem~\ref{ext_in_pm_hyp}).
We provide also two new characterizations of $P_m$-hyperconvex domains (Corollary~\ref{cor2} and Theorem~\ref{localization_Pm}), and finally we prove that for a $P_m$-hyperconvex domain $\Omega$ one can find a continuous $m$-subharmonic exhaustion function on $\bar \Omega$, which is strictly $m$-subharmonic and smooth in $\Omega$ (Corollary~\ref{smooth}).
We shall need the following lemma.
\begin{lemma}\label{lem}
Assume that $\Omega$ is a $P_m$-hyperconvex domain in $\C^n$, $1 \leq m \leq n$, and let $U$ be an open neighborhood of $\partial \Omega$. If $f\in\mathcal{SH}_m(U)\cap C^{\infty}(U)$ is a smooth function in some neighborhood of $\partial \Omega$, then there is a function $F \in \mathcal{SH}_m(\bar \Omega) \cap \mathcal{C}(\bar \Omega)$ such that $F=f$ on $\partial \Omega$.
\end{lemma}
\begin{proof}
Let $\psi\in \mathcal {SH}_m(\bar \Omega)\cap \mathcal C(\bar \Omega)$ be an exhaustion function for $\Omega$ (see Theorem~\ref{thm_charPmhx}). Let $U$ be an open set such that $\partial \Omega \subseteq U$ and $f\in \mathcal{SH}_m(U)\cap \mathcal C^{\infty}(U)$, and let $V$ be an open set such that $\partial \Omega \subseteq V \Subset U$. Moreover, let $K\subseteq \Omega$ be a compact set such that $\bar\Omega \subseteq K \cup U$ and $\partial K \subseteq V$.
Since $\Omega$ is also $m$-hyperconvex there exists a smooth and strictly $m$-subharmonic exhaustion function $\varphi$ for $\Omega$ (see~\cite{ACH}). Let $M>1$ be a constant large enough so that for all $z\in K$
\[
\varphi(z) - 1 > M \psi(z)\, .
\]
From Theorem~\ref{cont} part $(a)$ there exists an increasing sequence $\psi_j\in \mathcal{SH}_m(\Omega_j)\cap \mathcal \mathcal{C}(\bar \Omega_j)$, where $\bar \Omega \subseteq \Omega_j\Subset \Omega \cup V$, and such that $\psi_j\to \psi$ uniformly on $\bar \Omega$ so that
\[
\psi-\psi_j<\frac {1}{Mj}\, .
\]
Let us define
\[
\varphi_j :=
\begin{cases}
\max\left\{\varphi-\frac{1}{j}, M \psi_j\right\}, & \text{ if } z \in \Omega\, ,\\
M\psi_j, & \text{ if } z \in \Omega_j \setminus \Omega\, .
\end{cases}
\]
Note that the function $\varphi_j$ is $m$-subharmonic and continuous on $\Omega_j$, and $\varphi_j=\varphi-\frac{1}{j}$ on $K$. Next let $g$ be a smooth function such that $g=1$ on $V$, and $\operatorname{supp}(g)\subseteq U$. Since $\varphi_j$ is strictly $m$-subharmonic on the set where $g$ is non-constant, we can choose a constant $C$ so large that the function
\[
F_j:=C\varphi_j+gf
\]
belongs to $\mathcal{SH}_m^o(\bar\Omega)$. Observe that
\[
\max \{\varphi,M\psi\}-\max\left\{\varphi-\frac{1}{j}, M\psi_j\right\} \leq \frac{1}{j}\, ,
\]
and define
\[
F:=C\max\{\varphi,M\psi\}+gf.
\]
Then we have that
\[
F \geq F_j \geq F-\frac{1}{j}\, ,
\]
and therefore, by uniform convergence, we get that $F\in \mathcal{SH}_m(\bar\Omega)\cap \mathcal{C}(\bar\Omega)$. Furthermore, for $z \in \partial \Omega$, we have that
\[
0 \leq f(z)-F_j(z)=-C\varphi_j(z)=-CM\psi_j(z)\to -CM\psi (z)=0\,\text{ as } j\to \infty\, ,
\]
and we see that $F=f$ on $\partial \Omega$.
\end{proof}
Now we state and prove the main theorem of this section.
\begin{theorem}\label{ext_in_pm_hyp}
Let $\Omega$ be a bounded $P_m$-hyperconvex domain in $\mathbb C^n$, $1 \leq m \leq n$, and let $f$ be a real-valued function defined on $\partial \Omega$. Then the following are equivalent:
\medskip
\begin{enumerate}\itemsep2mm
\item there exists $F\in \mathcal{SH}_m(\bar \Omega)$ such that $F=f$ on $\partial \Omega$;
\item $f\in \mathcal{SH}_m(\partial \Omega)$.
\end{enumerate}
Furthermore, if $f$ is continuous on $\partial \Omega$, then the function $F$ can be chosen to be continuous on $\Omega$.
\end{theorem}
\begin{proof} (1)$\Rightarrow$(2): Follows immediately from Corollary~\ref{cor}.
\bigskip
(2)$\Rightarrow$(1): Let $f\in \mathcal{SH}_m(\partial \Omega)$, then by Theorem~\ref{cont} part $(b)$ there exists a decreasing sequence $u_j\in\mathcal{SH}_m^o(\partial \Omega)$ of smooth functions such that $u_j\to f$, $j\to \infty$. By assumption $\Omega$ is in particular a regular domain, and therefore there is a sequence of harmonic functions $h_j$ defined on $\Omega$, continuous on $\bar \Omega$ such that $h_j=u_j$ on $\partial \Omega$. Define
\[
Sh_j=\sup\left \{v\in \mathcal{SH}_m(\bar \Omega): v\leq h_j\right\}\, ,
\]
then by Theorem~\ref{thm_edwards} we have that
\[
Sh_j=\sup\left \{v\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal \mathcal{C}(\bar \Omega): v\leq h_j\right\}\, .
\]
Hence, $Sh_j$ is lower semicontinuous. Next we shall prove that in fact $Sh_j$ is continuous. By Lemma~\ref{lem} there exists $H_j\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal C (\bar \Omega)$ such that $H_j=h_j$ on $\partial \Omega$. This implies that $H_j\leq Sh_j\leq (Sh_j)^*$, so $(Sh_j)^*=h_j=H_j$ on $\partial \Omega$. Note also that for all $z\in \partial \Omega$, and all $\mu\in \mathcal J_z^m(\bar \Omega)$ it holds that $\operatorname{supp}(\mu)\subseteq \partial \Omega$ by Theorem~\ref{thm_charPmhx} and then
\[
\int_{\bar \Omega}(Sh_j)^*\,d\mu=\int_{\partial \Omega}(Sh_j)^*\,d\mu=\int_{\partial \Omega}H_j\,d\mu=\int_{\bar \Omega}H_j\,d\mu\geq H_j(z)=(Sh_j)^*(z)\, ,
\]
and therefore by Theorem~\ref{thm_pmboundary} $(Sh_j)^*\in \mathcal{SH}_m(\bar \Omega)$, so $(Sh_j)^*=Sh_j$ and finally $Sh_j\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal{C}(\bar \Omega)$. Now let
\[
F=\lim_{j\to \infty}Sh_j\, .
\]
Observe that $F=f$ on $\partial \Omega$, and $F\in \mathcal{SH}_m(\bar \Omega)$, since it is the limit of a decreasing sequence $Sh_j\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal C(\bar \Omega)$.
\bigskip
To prove the last statement of this theorem assume that $f\in \mathcal C(\partial \Omega)$. Let $h$ be a harmonic function on $\Omega$ that is continuous on $\bar \Omega$ with boundary values $f$. As in the previous part of the proof define
\[
Sh=\sup\left \{v\in \mathcal{SH}_m(\bar \Omega): v\leq h\right\}=\sup\left \{v\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal C(\bar \Omega): v\leq h\right\}\, ,
\]
so $Sh$ is lower semicontinuous. Furthermore, since $Sh\leq Sh_j$, then $(Sh)^*\leq (Sh_j)^*=Sh_j$ and
\[
(Sh)^*\leq \lim_{j\to \infty}Sh_j=F\leq Sh\, ,
\]
we have that $(Sh)^*=Sh=F\in \mathcal{SH}_m(\bar \Omega)\cap\mathcal C(\bar \Omega)$.
\end{proof}
Earlier we saw that if $\Omega$ is $P_m$-hyperconvex and if $z \in \partial \Omega$, then the measures in $\mathcal{J}_z^m(\bar \Omega)$ only have support on $\partial \Omega$. Following the line of \cite{HP} we will now see that, when $\Omega$ is $P_m$-hyperconvex, we actually have that $\mathcal J_z^m(\bar \Omega)=\mathcal J_z^m(\partial \Omega)$ for $z \in \partial \Omega$. This gives us another characterization of $P_m$-hyperconvex domains.
\begin{corollary}\label{cor2}
Let $\Omega$ be a bounded domain in $\C^n$. Then $\Omega$ is $P_m$-hyperconvex if, and only if, for all $z\in \partial \Omega$ we have $\mathcal J_z^m(\bar \Omega)=\mathcal J_z^m(\partial \Omega)$.
\end{corollary}
\begin{proof}
First assume that $\Omega$ is $P_m$-hyperconvex. It is clear that $\mathcal J_z^m(\partial \Omega)\subseteq\mathcal J_z^m(\bar \Omega)$. To prove the converse inclusion take $z\in \partial \Omega$, $\mu \in \mathcal J_z^m(\bar\Omega)$ and $f\in \mathcal{SH}_m^o(\partial \Omega)$, then $f\in \mathcal{SH}_m(\partial \Omega)\cap \mathcal C(\partial \Omega)$ and by Theorem~\ref{ext_in_pm_hyp} there exists $F\in \mathcal{SH}_m(\bar\Omega)$ such that $F=f$ on $\partial \Omega$. For $z\in \partial \Omega$ and $\mu\in \mathcal J_z^m(\bar \Omega)$ we have $\operatorname{supp}(\mu)\subseteq \partial \Omega$ and
\[
f(z)=F(z)\leq \int_{\bar\Omega}F\,d\mu=\int_{\partial \Omega}F\,d\mu=\int_{\partial\Omega}f\,d\mu,
\]
which means that $\mu\in \mathcal J_z^m(\partial \Omega)$.
For the converse implication assume that for all $z\in \partial \Omega$ we have $\mathcal J_z^m(\bar \Omega)=\mathcal J_z^m(\partial \Omega)$, then for all $z\in \partial \Omega$ and all $\mu \in \mathcal J_z^m(\bar \Omega)$ we have $\operatorname{supp}(\mu)\subseteq \partial \Omega$ so by Theorem~\ref{thm_charPmhx} $\Omega$ is $P_m$-hyperconvex.
\end{proof}
On $P_m$-hyperconvex domains, we can now characterize the functions $u \in \mathcal{SH}_m(\bar \Omega)$ as those functions that are in $\mathcal{SH}_m(\Omega)$ and $u|_{\partial \Omega} \in \mathcal{SH}_m(\partial \Omega)$.
\begin{theorem}\label{cor_msubbdvalue}
Let $\Omega$ be a bounded $P_m$-hyperconvex domain in $\mathbb C^n$. Then $u\in \mathcal{SH}_m(\bar \Omega)$ if, and only if, $u\in \mathcal{SH}_m(\Omega)$, and $u\in \mathcal{SH}_m(\partial \Omega)$.
\end{theorem}
\begin{proof}
It follows from Theorem~\ref{thm_pmboundary} and Corollary~\ref{cor2}.
\end{proof}
As a corollary we obtain that for $P_m$-hyperconvex domains the exhaustion function can be chosen to be strictly $m$-subharmonic and smooth, as it was announced in Theorem~\ref{thm_charPmhx}.
\begin{corollary}\label{smooth}
Let $\Omega$ be a bounded $P_m$-hyperconvex domain in $\mathbb C^n$. Then $\Omega$ admits a continuous negative exhaustion function which is $m$-subharmonic on $\bar \Omega$, smooth and strictly $m$-subharmonic on $\Omega$.
\end{corollary}
\begin{proof}
Since $\Omega$ is also $m$-hyperconvex then there exists a negative exhaustion function $\varphi\in \mathcal {SH}_m(\Omega)\cap \mathcal C^{\infty}(\Omega)$, which is strictly $m$-subharmonic on $\Omega$. Now it follows from Theorem~\ref{cor_msubbdvalue} that $\varphi \in \mathcal {SH}_m(\bar\Omega)$.
\end{proof}
Finally, we can prove that if a domain is locally $P_m$-hyperconvex then it is globally $P_m$-hyperconvex.
\begin{theorem}\label{localization_Pm}
Let $\Omega$ be a bounded domain in $\mathbb C^n$ such that for every $z\in \partial \Omega$ there exists a neighborhood $U_z$ such that $\Omega\cap U_z$ is $P_m$-hyperconvex, then $\Omega$ is $P_m$-hyperconvex.
\end{theorem}
\begin{proof} Assume that $\Omega$ is locally $P_m$-hyperconvex. Then it is also locally $m$-hyper\-con\-vex. By Theorem~3.3 in~\cite{ACH}, we know that $\Omega$ must be globally $m$-hyperconvex. Thus, there exists $\psi \in \mathcal{SH}_m(\Omega)\cap \mathcal{C}(\bar \Omega)$, $\psi \not\equiv 0,$ such that $\psi|_{\partial \Omega}=0$. We shall now show that $\psi \in \mathcal{SH}_m(\bar \Omega)$. Thanks to Theorem~\ref{localization} it is enough to show that for every $z \in \bar \Omega$ there is a ball $B_z$ such that $\psi|_{\bar \Omega \cap \bar B_z} \in \mathcal{SH}_m(\bar \Omega \cap \bar B_z)$.
For $z \in \Omega$ there exists $r>0$ such that $B(z,r)\Subset \Omega$ and then $\psi|_{B(z,r)}\in \mathcal {SH}_m(B(z,r))\cap \mathcal C(\bar B(z,r))$. Since,
\begin{equation}\label{ball}
\mathcal J^m_z(\bar B(z,r))=\mathcal J^m_z(\partial B(z,r))=\{\delta_z\}\, ,
\end{equation}
we have that $\psi\in\mathcal {SH}_m(\partial B(z,r))$ and therefore by Corollary~\ref{cor_msubbdvalue} we have that $\psi\in\mathcal {SH}_m(\bar B(z,r))$.
Now it is sufficient to look at $z \in \partial \Omega$. Fix $z_0 \in \partial \Omega$, and a small ball $B_{z_0}$ around $z_0$. Without loss of generality assume that $B_{z_0} \Subset U_{z_0}$ such that $\Omega \cap B_{z_0}$ is $P_m$-hyperconvex. Once again, by Corollary \ref{cor_msubbdvalue} it is enough to show that $\psi \in \mathcal{SH}_m\bigl(\partial (\Omega \cap B_{z_0})\bigr)$, i.e. for every $z \in \partial (\Omega \cap B_{z_0})$, and every $\mu \in \mathcal{J}_{z}^m(\partial (\Omega \cap B_{z_0}))$, it holds
\begin{equation}\label{eq:locally}
\psi(z)\leq \int \psi \, d\mu\, .
\end{equation}
Suppose that $z \in \partial \Omega \cap B_{z_0}\setminus \partial B_{z_0}$. First we shall show that if $\mu \in \mathcal{J}_z^m(\partial(\Omega \cap B_{z_0}))$, then $\mu$ has support on $\partial \Omega$ and therefore condition $(\ref{eq:locally})$ will be fulfilled. Since $\Omega \cap U_{z_0}$ is $P_m$-hyperconvex, it has an exhaustion function $\varphi \in \mathcal{SH}_m(\bar \Omega \cap \bar {U}_{z_0})$, and especially $\varphi \in \mathcal{SH}_m\left(\partial(\bar \Omega \cap \bar {B}_{z_0})\right)$. Let $\mu \in \mathcal{J}_z^m\left(\partial(\Omega \cap B_{z_0})\right)$, then we have
\[
0=\varphi(z)\leq \int \varphi \, d\mu \leq 0\, ,
\]
which means that $\mu$ has support where $\varphi=0$, i.e. on $\partial \Omega$.\\
Next, suppose that $z \in \bar\Omega \cap \partial B_{z_0}$. We claim that
\[
\mathcal{J}_{z}^m\left(\partial(\Omega \cap B_{z_0})\right)=\{\delta_z\}\, ,
\]
and this makes that $(\ref{eq:locally})$ holds. From (\ref{ball}) and from Theorem~\ref{ext_in_pm_hyp} we know that for every $z \in \bar \Omega \cap \partial B_{z_0}$ there exists a function $\varphi \in \mathcal {SH}_m(\bar B_{z_{0}})\subseteq\mathcal{SH}_m\left(\bar\Omega \cap \bar B_{z_0}\right)$ such that $\varphi(z)=0$ and $\varphi(\xi)<0$ for every $\xi \neq z$. By the same argument as above, we see that $\mathcal{J}_{z}^m\left(\partial(\Omega \cap B_{z_0})\right)=\{\delta_z\}$.
\end{proof}
\section{Some concluding remarks on approximation}\label{sec approx}
Approximation is a central part of analysis. The type of approximation needed depends obviously on the situation at hand. In connection with Theorem~\ref{cont} one can ask the following question. Let $\Omega\subseteq\C^n$ be a bounded and open set, and let $1\leq m\leq n$. Under what assumptions on $\Omega$ do we have that
\begin{equation}\label{question}
\mathcal{SH}_m(\Omega) \cap \mathcal C(\bar \Omega)=\mathcal{SH}_m(\bar\Omega) \cap \mathcal C(\bar \Omega)?
\end{equation}
In the case when $m=1$, this type of theorem can be traced back to the work of Walsh~\cite{Walsh}, Keldysh~\cite{Keldych}, and
Deny~\cite{deny, deny2}, where they considered harmonic functions. In the harmonic case, some call this theorem the approximation theorem of Keldysh-Brelot after the contributions~\cite{Keldych,Brelot, Brelot2}. For subharmonic functions this type of approximation is included in the inspiring work of Bliedtner and Hansen~\cite{BlietnerHansen2} (see also~\cite{BlietnerHansen1,Hansen}). The articles mentioned are in a very general
setting. For us here it suffice to mention:
\begin{theorem}\label{thm_question}
Let $\Omega$ be a bounded domain in $\RE^n$. The following assertions are then equivalent:
\begin{enumerate}\itemsep2mm
\item for each $u$ in $\mathcal{SH}(\Omega) \cap \mathcal C(\bar \Omega)$ and each $\varepsilon>0$ there is a function $v$ in $\mathcal{SH}(\bar\Omega) \cap \mathcal C(\bar \Omega)$ such that $|u-v|<\varepsilon$ on $\bar\Omega$;
\item the sets $\RE^n\backslash\bar{\Omega}$, and $\RE^n\backslash \Omega$, are thin at the same points of $\bar\Omega$.
\end{enumerate}
\end{theorem}
\noindent For further information on the case $m=1$ we refer to the inspiring survey written by Hedberg~\cite{Hedberg} (see also~\cite{gardiner}).
If we look at the other end case of the Caffarelli-Nirenberg-Spruck model, when $m=n$, and we are in the world of pluripotential, then our approximation question bear resemblance with the so called Mergelyan approximation of holomorphic
function. Therefore, some call~(\ref{question}) the $\mathcal{PSH}$-Mergelyan property (see e.g.~\cite{H}).
The first positive result for the $\mathcal{PSH}$-Mergelyan property is due to Sibony. In 1987, he proved in~\cite{S} that every smoothly bounded and pseudoconvex domain has this property. Later Forn\ae ss and Wiegerinck~\cite{FW} generalized this in their beautiful paper to arbitrary domains with $C^1$-boundary. Recently, Persson and Wiegerinck~\cite{PW} proved that a domain of which the boundary is continuous with the possible exception of a countable set of boundary points, has the $\mathcal{PSH}$-Mergelyan property (this generalize~\cite{avelin2,H}). Furthermore, in~\cite{PW} they constructed very enlightening examples that show that there can be no corresponding Theorem~\ref{thm_question} in the case $m=n$.
\bigskip
At this point there is no satisfactory answer to question~(\ref{question}) within the Caffarelli-Nirenberg-Spruck framework that covers
the knowledge of the end cases $m=1$, and $m=n$. Even so, in Theorem~\ref{bm-reg} we give a family of bounded domains that satisfies~(\ref{question}), and we prove several characterizations of this type of domains. Obviously, there are domains that satisfies~(\ref{question}), and is not included in Theorem~\ref{bm-reg}. For further information, and inspiration, on approximation we refer to~\cite{gardiner,Gauthier2} and the references therein.
\begin{theorem} \label{bm-reg}
Assume that $\Omega$ is a bounded domain in $\C^n$, $n\geq 2$, $1\leq m\leq n$. Then the following assertions are equivalent:
\medskip
\begin{enumerate}\itemsep2mm
\item for every continuous function $f:\partial\Omega\to \RE$ we have that
\[
\operatorname{PB}^m_f\in\mathcal{SH}_m(\bar{\Omega})\cap \mathcal C(\bar {\Omega})\, ,
\text{ and } \operatorname{PB}^m_f=f \text{ on } \partial \Omega\, ,
\]
where
\[
\operatorname{PB}^m_f(z)=\sup\Bigg\{v(z): v\in\mathcal{SH}_m(\bar \Omega),\; v(\xi)\leq f(\xi)\, , \;\; \forall \xi\in\partial\Omega\Bigg\}\, ;
\]
\item $\partial\Omega$ has a strong barrier at every point $z_0\in\partial\Omega$ that is $m$-subharmonic on $\bar{\Omega}$, i.e. there exists a $m$-subharmonic function $u:\bar{\Omega}\to\RE$ such that
\[
\lim_{x\to y_0\atop x\in\Omega} u(x)=0\, ,
\]
and
\[
\limsup_{x\to y\atop x\in\Omega} u(x)<0 \qquad \text{ for all } y\in\bar{\Omega}\backslash\{y_0\}\, ;
\]
\item $\Omega$ admits an exhaustion function $\varphi$ that is negative, continuous, $m$-subharmonic on $\bar{\Omega}$, smooth on $\Omega$, and such that
\[
\left(\varphi(z)-|z|^2\right)\in \mathcal{SH}_m(\bar{\Omega})\, ;
\]
\item for every $z\in \partial \Omega$ we have that $\mathcal{J}_z^m(\bar \Omega)=\{\delta_z\}$.
\end{enumerate}
\end{theorem}
\begin{proof}
$(1)\Rightarrow(4):$ Fix $z\in \partial \Omega$, $\mu\in \mathcal J_z^m(\bar \Omega)$, and let $f$ be a real-valued continuous function defined on $\partial \Omega$ such that $f(z)=0$ and $f(\xi)<0$ for $\xi \neq z$. Then it holds that
\[
0=\operatorname{PB}^m_f(z)\leq \int\operatorname{PB}^m_f\,d\mu\leq 0\, ,
\]
and therefore it follows that $\operatorname{supp}(\mu)\subseteq \{z\}$, Thus, $\mu=\delta_z$.
\medskip
$(4)\Rightarrow(1):$ First note that it follows from (4) that every continuous functions defined on the boundary $\partial\Omega$ is $m$-subharmonic on $\partial\Omega$ in the sense of Definition~\ref{def_msubkomp}. Let $f\in \mathcal C(\partial \Omega)$, then by Theorem~\ref{ext_in_pm_hyp} and Theorem~\ref{thm_charPmhx} there exists a function $F\in \mathcal {SH}_m(\bar \Omega)\cap \mathcal C(\bar \Omega)$ such that $F=f$ on $\partial \Omega$. Let us define
\[
\textbf{S}_f(z)=\sup\Bigg\{v(z): v\in\mathcal{SH}_m(\Omega),\; \varlimsup_{\zeta\rightarrow\xi \atop \zeta\in\Omega}v(\zeta)\leq f(\xi)\, , \;\; \forall \xi\in\partial\Omega\Bigg\}\, .
\]
For $z\in \Omega$, we have that
\[
F(z)\leq \operatorname{PB}^m_f(z)\leq \textbf{S}_f(z)\, ,
\]
and then
\[
\lim_{\zeta\to \xi}\textbf{S}_f(\zeta)=f(\xi)\,\qquad \text{for all } \xi\in \partial \Omega\, .
\]
Thanks to the the generalized Walsh theorem (Proposition~3.2 in~\cite{Blocki_weak}) we get that $\textbf{S}_f\in \mathcal {SH}_m(\Omega)\cap \mathcal C(\bar \Omega)$, and by Theorem~\ref{cor_msubbdvalue}, $\textbf{S}_f\in \mathcal {SH}_m(\bar \Omega)$. Hence,
$\operatorname{PB}^m_f=\textbf{S}_f$ and the proof is finished.
\medskip
$(1)\Rightarrow(2):$ Fix $z\in \partial \Omega$, and let $f$ be a continuous function on $\partial \Omega$ such that $f(z)=0$ and $f(\xi)<0$ for $\xi \neq z$. Then the function $\operatorname{PB}^m_f$ is a strong barrier at $z$.
\medskip
$(2)\Rightarrow(4):$ Fix $z\in \partial \Omega$, $\mu\in \mathcal J_z^m(\bar \Omega)$, and let $u_z$ be a strong barrier at $z$. Then it holds
that
\[
0=u_z(z)\leq \int u_z\,d\mu\leq 0\, ,
\]
and therefore $\operatorname{supp}(\mu)\subseteq \{z\}$. Thus, $\mu=\delta_z$.
\medskip
$(4)\Rightarrow(3):$ Let $\mathcal J_z^{c,m}$ be the class of Jensen measures defined by continuous $m$-subharmonic functions on $\Omega$ (see~\cite{ACH}), i.e. $\mu\in \mathcal J_z^{c,m}$ if
\[
u(z)\leq \int_{\bar{\Omega}} u \, d\mu\, , \text{ for all } u \in \mathcal{SH}_m(\Omega) \cap \mathcal C(\bar \Omega)\, .
\]
Let $z\in \partial \Omega$, and note that $\mathcal {SH}_m^o(\bar \Omega)\subseteq \mathcal{SH}_m(\Omega) \cap \mathcal C(\bar \Omega)$, so
\[
\mathcal J_z^{c,m}\subseteq \mathcal J_z^m(\bar \Omega)=\{\delta_z\}\, .
\]
Therefore by Theorem~4.3 in~\cite{ACH}, there exits an exhaustion function $\varphi$ that is negative, smooth, $m$-subharmonic on $\Omega$, continuous on $\bar \Omega$, and such that
\[
\left(\varphi(z)-|z|^2\right)\in \mathcal{SH}_m(\Omega)\, .
\]
Condition (4) implies that every continuous function defined on the boundary is also $m$-subharmonic. This means that $(\varphi(z)-|z|^2)\in \mathcal{SH}_m(\partial\Omega)$. Finally, Theorem~\ref{cor_msubbdvalue} gives us that $(\varphi(z)-|z|^2)\in \mathcal {SH}_m(\bar \Omega)$.
\medskip
$(3)\Rightarrow(2):$ Condition (3) implies that for $z\in\partial\Omega$ we have that
\[
\varphi(z)-|z|^2=-|z|^2\in \mathcal {SH}_m(\partial \Omega)\, .
\]
Take $z_0\in \partial \Omega$, and note that
\[
-|z-z_0|^2=-|z|^2+z\bar z_0+\bar zz_0-|z_0|^2\in \mathcal {SH}_m(\partial \Omega)\, .
\]
Theorem~\ref{ext_in_pm_hyp} and Theorem ~\ref{thm_charPmhx} imply that there exists $F\in \mathcal {SH}_m(\bar \Omega)\cap \mathcal C(\bar \Omega)$ such that $F=-|z-z_0|^2$ on $\partial \Omega$. The function $F$ is a strong barrier at $z_0$.
\end{proof}
|
\section*{INTRODUCTION}
The Dicke model is one of the simplest models that captures the effects of the interaction between several atoms and an electromagnetic mode of radiation. Since its original proposal \cite{Dicke}, it has been the subject of extensive investigations and was shown to possess a remarkable effect, namely, the existence of thermal~\cite{Hepp,Wang}, quantum~\cite{Emary} and excited state~\cite{Perez} quantum phase transitions at specific values of the temperature, coupling strength and excitation energy, respectively. Following those lines, there has been a plethora of studies discussing its spectral properties. This includes the cases of small $N$ \cite{Braak}, where integrability analyses have been done, and large but finite $N$~\cite{Chen}, with accompanying discussions about the relations between the aforementioned transitions and with chaos~\cite{Bastarrachea,Bastarrachea2,Bastarrachea3,Bastarrachea4}. The small coupling strength near the resonant regime has also been thoroughly investigated, resulting in the integrable Tavis-Cummings model \cite{Tavis} after performing a rotating-wave approximation. Several studies of the dynamics have been done with a variety of motivations, e.g., entanglement, collapses and revivals \cite{Alvermann, Agarwal}, behavior near the classical limit~\cite{Bakemeier}, and open system problems~\cite{Fuchs}. The dynamics in the small qubit frequency regime has also been explored \cite{Agarwal}, but mainly for low coupling strengths. In the experimental frontier, it has been implemented in a variety of scenarios \cite{Baumann1,Baumann2,Baden} and quantum simulations have also been considered, primarily for one qubit \cite{Ballester,Simon,Langford,Braumuller}, but also for larger $N$ \cite{Mezzacapo, Lamata}. Furthermore, current experimental trends are directioned towards reaching increasingly large values of the coupling \cite{Niemczyk,Pol, Yoshihara,Yoshihara2,Pol2}, such that it is worthwhile to analyze the large-coupling region and identify interesting effects. Recently, efforts to classify the quantum Rabi model in a variety of coupling regimes have been carried out~\cite{MikelRabi}.
In this Article, we study the dynamics of the Dicke model in the regime of small but nonzero qubit frequency $\omega_0$ for a finite number $N$ of atoms. The ensuing separation of timescales naturally leads to the use of an adiabatic approximation, a technique previously employed in the one qubit case \cite{Larson1, Irish} and for large qubit frequency \cite{Liberti,Relano} in the language of the Born-Oppenheimer approximation. This allows us to separate fast oscillating behavior at the frequency of the mode, $\omega$, from secular effects induced by the finiteness of $\omega_0$ and keep only first-order corrections. For a fixed and small value of the qubit frequency, there are two discernible regimes showing distinct behavior depending on the value of $g$. The first one is the deep strong coupling regime, introduced originally for the quantum Rabi model \cite{Casanova} and extended now to this multiqubit quantum Rabi model, which shows effects not present for the $N=1$ atomic case. The second one is an intermediate coupling regime where the atomic dynamics is nontrivial, governed by the Lipkin-Meshkov-Glick Hamiltonian \cite{Lipkin,Tsomokos}, yet the bosonic mode decouples for each number state. At the same time, both behaviors become mixed in the intermediate regime.
\section*{RESULTS}
{\it{The model}.--} We consider now the Dicke Hamiltonian,
\begin{equation}
H=\omega_0 J_z+\omega a^{\dagger}a+gJ_x(a+a^{\dagger}),
\label{DickeHamiltonian}
\end{equation}
where $\omega_0$ is the qubit frequency, $\omega$ is the photon frequency, and $g$ is the coupling constant. $J_z$ and $J_x$ are collective atomic operators, defined as $2J_z=\sum_i \sigma_z^i$, $2J_x=\sum_i\sigma_x^i$. Both of them, together with $J_y$, satisfy the angular momentum commutation relations. We also introduce $\hat{J}^2$, whose eigenvalues are $J(J+1)$ and work in the symmetric subspace with $2J=N
$, where $N$ is the number of atoms. Likewise, $a$ and $a^{\dagger}$ are photon annihilation and creation operators, respectively.
In the regimes we are interested in, both rotating and counterrotating terms are important, and we must keep both to account for the dynamics. Instead of neglecting either of them, we perform a displacement transformation on $H$~\cite{Chen,Bastarrachea5}, and its associated time evolution operator,
\begin{align*}
U_H&=D(\lambda)U_{H'}D(-\lambda),\\
H'&=D(-\lambda)HD(\lambda).\numberthis
\end{align*}
By choosing $\lambda=\frac{g}{\omega}J_x$, such that it does not commute with $J_z$, the resulting $H'$ is
\begin{equation}
H'=\omega_0 e^{-\frac{gJ_x}{\omega}(a^{\dagger}-a)}J_z e^{\frac{gJ_x}{\omega}(a^{\dagger}-a)}+\omega a^{\dagger}a-\frac{g^2}{\omega}J_x^2.
\end{equation}
The reason for this transformation is that, if $\omega_0 = 0$, the Hamiltonian would be diagonal, as is apparent from the previous equation. If we introduce ladder operators with respect to $J_x$, $J_{\pm}=J_z\mp iJ_y$, and use the ensuing commutation relations, $J_xJ_{\pm}=J_{\pm}(J_x\pm 1)$, the Hamiltonian takes the simpler form,
\begin{equation}
H'=\frac{\omega_0}{2} \left(e^{\frac{g}{\omega}(a-a^{\dagger})}J_-+{\rm H.c.}\right)+\omega a^{\dagger}a-\frac{g^2}{\omega}J_x^2.
\end{equation}
Since $\omega\gg\omega_0$, we move to an interaction picture generated by the $\omega a^{\dagger}a$ term,
\begin{gather}
U_H=e^{\frac{gJ_x}{\omega}(a^{\dagger}-a)}e^{-i\omega t\, a^{\dagger}a }U_{H_2}e^{-\frac{gJ_x}{\omega}(a^{\dagger}-a)},
\raisetag{-.5em}\label{EvolOpt}
\end{gather}
where $U_{H_2}$ is the time evolution operator associated with the following Hamiltonian,
\begin{equation}
H_2=\frac{\omega_0}{2} \left[ e^{\frac{g}{\omega}(ae^{-i\omega t}-a^{\dagger}e^{i\omega t})}J_-+{\rm H.c.} \right] - \frac{g^2}{\omega}J_x^2.\label{Hamilt}
\end{equation}
Further manipulations depend on the magnitude of $\frac{g^2}{\omega^2}$. It should be noted that that the Hamiltonian of Eq.(\ref{Hamilt}) is structurally very similar to the trapped ion Hamiltonian before doing the vibrational rotating wave approximation. In consequence, it is amenable to the same methods of analysis~\cite{Vogel}, a fact we will exploit in the following sections.
{\it Deep Strong Coupling Regime.--}
$g\gtrapprox\omega$ in this scenario so $g^2/\omega$, the prefactor of the $J_x^2$ term in Eq.~(\ref{Hamilt}), is much larger than $\omega_0$, the prefactor of the other term. Thus, the former induces evolution on a faster time scale than the latter, but, at the same time, $g^2/\omega$ remains comparable to $\omega$ itself. Therefore, it is convenient to move to an interaction picture with respect to the $J_x^2$ term as well. The resulting time evolution operator and Hamiltonian are
\begin{eqnarray}
U_H=&&e^{\frac{gJ_x}{\omega}(a^{\dagger}-a)}e^{-i\left(\omega a^{\dagger}a-\frac{g^2J_x^2}{\omega}\right)t}U_{H_3}e^{-\frac{gJ_x}{\omega}(a^{\dagger}-a)},
\end{eqnarray}
\begin{equation}
H_3=\frac{\omega_0}{2} e^{\frac{g}{\omega}\left(ae^{-i\omega t}-a^{\dagger}e^{i\omega t}\right)}e^{i\frac{g^2}{\omega}(2J_x+1)t}J_-+{\rm H.c.}\label{h3Eq}
\end{equation}
We point out that, in Eq.~(\ref{h3Eq}), the resulting dynamics depends strongly on the magnitude $\frac{g^2}{\omega^2}$.
{\it i) Resonant behavior.}
In the case of $g=\omega\sqrt{k}$ where $k$ is an integer, each $J_x$ eigenstate will generate a resonance with a different term in the displacement factor, as shown in the following table for the specific case of $J=2$, namely, four qubits,
\begin{center}
\begin{tabular}{c||c|c|c|c|c}
$J_x$ & $-2$ & $-1$ & $0$ & $1$ & $2$ \\ \hline\hline
$(2J_x+1)k\omega$& $-3k\omega$ & $-k\omega$ & $k\omega$ & $3k\omega$& $5k\omega$
\end{tabular}
\end{center}
For example, a $3k\omega$ term will resonate with an $a^{3k}$ term. For this purpose, it is convenient to perform a Fourier expansion of the displacement operator,
\begin{align*}
e^{\beta(ae^{-i\omega t}-a^{\dagger}e^{i\omega t})}&=\Omega_{\hat{n}}^0(\beta)+\sum_{m>0}^{\infty}\Omega_{\hat{n}}^m(\beta)a^m e^{-im\omega t}+(-1)^me^{im\omega t}a^{\dagger}{}^m\Omega_{\hat{n}}^m(\beta),\\[1.5em]
\Omega_{\hat{n}}^m(\beta)&=\beta^me^{-\frac{\beta^2}{2}}L_{\hat{n}}^m(\beta^2)\frac{\hat{n}!}{(\hat{n}+m)!},\numberthis
\end{align*}
where $L_n^m$ is an associated Laguerre polynomial. The first-order effective Hamiltonian is obtained by averaging with respect to $\omega$,
\begin{gather}
\begin{align*}
\braket{H_3}=\frac{\omega_0}{2} \biggl[ \Omega_{\hat{n}}^ka^k(P_{0}J_-P_{1})+\Omega_{\hat{n}}^{3k}a^{3k}(P_{1}J_-P_{2})-a^{\dagger}{}^{3k}\Omega_{\hat{n}}^{3k}(P_{-2}J_-P_{-1})-a^{\dagger}{}^{k}\Omega_{\hat{n}}^k(P_{-1}J_-P_{0})+{\rm H.c.}\biggr] ,
\end{align*}\numberthis
\raisetag{-.5em}
\end{gather}
where the $P_m$ are projectors onto the $J_x=m$ state. Note that the expected $a^{5k}$ term does not appear because $P_{2}J_-=0$. Nothing can be lowered to the $J_x=2$ eigenstate. Using states $\ket{m,n}$ for which $J_x=m$ and $a^{\dagger}a=n$, the dynamics can be seen to generate the following dispersive chain of connected states,
\begin{align*}
\ket{-2,n+4k}\leftrightarrow& \ket{-1,n+k}\leftrightarrow \ket{0,n}\leftrightarrow\ket{1,n+k}\\&\leftrightarrow \ket{2,n+4k}.\numberthis
\end{align*}
The Hilbert space can thus be divided into subspaces as shown in Fig. \ref{Cadenas}(c).
\begin{figure}
{\includegraphics[width=0.5\textwidth]{DispersiveChains.pdf}}
\caption{Schematic representation of the Hilbert space of the system. On each table, each square represents a basis state labeled by the eigenvalues of $J_x$ (left column) and of $a^{\dagger}a$ (top row). Solid lines indicate the chains of states that are coupled by the effective Hamiltonian for different values of $g$ and $J$: (a) $g=\omega$, $J=\frac{3}{2}$, (b) $g$ non resonant, $J=\frac{3}{2}$, and (c) $g=\omega$, $J=2$. Black dots represent decoupled states.}\label{Cadenas}
\end{figure}
For large $n$, these chains emerge and are connected as displayed in Fig.~\ref{Cadenas}. To verify these analytical results, we did numerical simulations of Eq.~(\ref{DickeHamiltonian}) for $N=2$, $g=\sqrt{5}\omega$ and $\omega_0=0.1\omega$, as shown in Fig.~\ref{DSCResonance}(a). We also computed the cases for $N=8$, $g=1$, $\omega_0=0.1\omega$ and $\omega_0=0.01\omega$, see Figs.~\ref{DSCResonance}(b) and~\ref{DSCResonance}(c). To achieve this, we plot
\begin{equation}
P(t)=|\bra{J_x=0,n=0}U(t)\ket{J_x=0,n=0}|^2.
\label{P(t)}
\end{equation}
\begin{figure}
\includegraphics[width=0.48\textwidth]{Resonance.pdf}
\caption{Evolution of $P(t)$ for (a) $J=1$ (2 atoms), $g=\sqrt{5}\omega$, $\omega_0=0.1\omega$; (b) $J=4$ (8 atoms), $g=\omega$, $\omega_0=0.1\omega$ and (c) $J=4$ (8 atoms), $g=\omega$, $\omega_0=0.01\omega$ showing the existence of resonances. Numerical simulation is orange and theoretical approximation is dashed black.}\label{DSCResonance}
\end{figure}
We point out that, for $\omega_0=0.1\omega$, a complete depopulation of the initial state takes place in a short time. For smaller $\omega_0$, the approximation works better but the time span required to depopulate the initial state increases accordingly. In the $N=8$, $\omega_0=0.1\omega$ case, the approximation captures qualitatively the correct behavior, but higher-order secular effects become apparent as the theoretical approximation lags behind the simulations. Peaks are distorted as well due to micromotion effects.
This same resonant behavior is present for $J$ equal to a half-integer. In the specific case of $J=\frac{3}{2}$, the corresponding table reads
\begin{center}
\begin{tabular}{c||c|c|c|c}
$J_x$ & $-3/2\hspace{0.1cm}$ & $-1/2\hspace{0.1cm}$ & $1/2$ & $3/2$ \\ \hline\hline
$(2J_x+1)k\omega$& $-2k\omega$ & $0k\omega$ & $2k\omega$ & $4k\omega$
\end{tabular}
\end{center}
In this case, the $J_x=-\frac{1}{2}\leftrightarrow J_x=\frac{1}{2}$ transition involves no change in the photon number, as shown in Fig.~\ref{Cadenas}(a). When considering higher values of $J$, the dispersive chains start growing but remain finite.
{\it ii) Off-resonant behavior.}
When moving away from these resonances, the system responds differently depending on whether $J$ is an integer or a half-integer. If $J$ is an integer, the chain dynamics are progressively suppressed as the detuning increases and the system reaches a dispersive regime. The resonances described in the previous sections have a width of order $\omega_0$, such that they are sharper if $\omega_0$ is smaller. In the specific case of $J=1$, i.e., for two qubits, the minimum value of $P$ for a $k$ resonance can be easily calculated,
\begin{align*}
P_{\rm min}&=\frac{(g^2-k\omega^2)^2}{(g^2-k\omega^2)^2+4\omega^2\omega_0^2(\Omega_0^k)^2 k!}\\[1em]
&=\frac{(g^2-k\omega^2)^2}{(g^2-k\omega^2)^2+4\omega^2\omega_0^2(\frac{g}{\omega})^{2k}\frac{e^{-\frac{g^2}{\omega^2}}}{k!}} \, . \numberthis
\end{align*}
\\
It is Lorentzian in $g^2$, rather than $g$, such that the peak is not symmetric with respect to the resonance.
\begin{figure}
\includegraphics[width=0.48\textwidth]{gsqrt3.pdf}
\caption{Evolution of $P(t)$ for $J=2$ (4 atoms) and $\omega_0=0.1\omega$, with (a) $g=(\sqrt{3}+0.03)\omega$ and (b) $g=(\sqrt{3}+0.07)\omega$. Numerical simulation is shown in orange and theoretical approximation in dashed black.}\label{DSCOResonance}
\end{figure}
Figs.~\ref{DSCOResonance}(a) and~\ref{DSCOResonance}(b) show that as we move away from the resonance, the value of $P_{\rm min}$ departs from $0$ and starts to increase, while the oscillations experience a frequency shift. The new frequency for the $k^{th}$ resonance and $J=2$ reads
\begin{equation}
\frac{1}{2}\sqrt{(g^2-k\omega^2)^2+4\omega^2\omega_0^2\left(\frac{g}{\omega}\right)^{2k}\frac{e^{-\frac{g^2}{\omega^2}}}{k!}}.
\end{equation}
\begin{figure}
\includegraphics[width=0.48\textwidth]{g12.pdf}
\caption{ (a) Evolution of $P(t)$ for $g=1.2\omega$ and $\omega_0=0.01\omega$, for 800 cycles of the mode frequency. (b) Evolution of $|\bra{J_z=0,n=0}U(t)\ket{J_z=0,n=0}|^2$ for $g=1.2\omega$ with $\omega_0=0$ (black) and $\omega_0=0.01\omega$ (orange). The plot shows the evolution from $\omega t=(2\pi)120$ to $\omega t=(2\pi)140$.}\label{DSCOResonanceII}
\end{figure}
Figure~\ref{DSCOResonanceII} shows that for a sufficiently detuned $g$, most of the unveiled dynamics is almost suppressed.
If $J$ is a half-integer, the physical description changes radically. As can be seen from the table for $J=3/2$, the $J_x=-1/2\leftrightarrow J_x=1/2$ transition is not suppressed by the rapidly oscillating terms, since $g^2(2J_x+1)/\omega=0$ for the $J_x= -1/2$ entry, independently of the value of $g$. The effective Hamiltonian in these conditions reads
\begin{equation}
\braket{H_3}=\frac{\omega_0}{2}\Omega_{\hat{n}}^0\left(P_{-\frac{1}{2}}J_-P_{\frac{1}{2}}+P_{\frac{1}{2}}J_+P_{-\frac{1}{2}}\right).
\end{equation}
This encompasses the case of the single-qubit quantum Rabi model ($J=1/2$) and accounts for the distorted peaks reported in Ref.~\cite{Casanova}. Graphically, the dispersive chains of Fig. \ref{Cadenas}(a) collapse into Fig. \ref{Cadenas}(b).
{\it Photon-number-dependent Regime.--}
A different behavior arises when $g\ll\omega$, such that $g^2\approx\omega\omega_0$. Then, both terms in Eq.~(\ref{Hamilt}) are comparable, and it is not useful to go into the interaction picture generated by the $J_x^2$ term. Instead, the first-order effective Hamiltonian is obtained by averaging directly,
\begin{equation}
\braket{H_2}=\omega_0\Omega_{\hat{n}}^0J_z-\frac{g^2}{\omega}J_x^2\label{EffHamilt2},
\end{equation}
and the result, as has been pointed out before \cite{Larson2, Morrison, Unanyan}, is a Lipkin-Meshkov-Glick Hamiltonian. Within this approximation, Fock states remain unchanged, while each Fock state induces a different evolution on the atomic internal states, characterised by the $\Omega_{\hat{n}}^0$ coefficient accompanying $\omega_0$. The evolution of the Hamiltonian of Eq.~(\ref{DickeHamiltonian}) changes the number of photons due to the displacement operators in Eq.~(\ref{EvolOpt}), albeit very weakly since $\frac{g}{\omega}$ is now a small quantity. This is verified in Figs. \ref{Atomic}(a) and \ref{Atomic}(b). Furthermore, the effective Hamiltonian is a good approximation even for $40$ atoms as shown in Figs. \ref{Atomic}(c) and \ref{Atomic}(d), while still displaying nontrivial physics. In a way, the dispersive chains of the previous section have become vertical in the tables of Fig. \ref{Cadenas}. If the initial state contains several Fock components, the $\Omega_{\hat{n}}^0$ term originates the dispersive dynamics that gives rise to the collapses and revivals reported in Ref.~\cite{Agarwal}.
\begin{figure}
\includegraphics[width=0.48\textwidth]{QPT.pdf}
\caption{For $g^2J=0.01\omega^2$ and $\omega_0=0.01\omega$, we show (a) the probability of finding one or less photons at time $t$ for $J=4$ (8 atoms), (b) the probability of finding two or less photons at time $t$ for $J=20$ (40 atoms), (c) $P(t)$ for $J=4$, (d) $P(t)$ for $J=20$. Orange curves are numerical simulations and theoretical approximation is plotted in dashed black.}\label{Atomic}
\end{figure}
It should be noted that this regime is close to the critical coupling of the Dicke quantum phase transition. To analyse this, it is convenient to discuss thoroughly the approximations that led into Eq.~(\ref{EffHamilt2}). The separation of timescales that justifies the averaging requires that the resulting effective dynamics does not occur at a rate comparable to the fast timescale. The fastest effective dynamics occur at frequencies of the order of $2J\omega_0$ and $\frac{g}{\omega}^2J^2$, since they are at first glance the largest energy differences that occur in $\braket{H_2}$. Thus, reasonable upper bounds are $2\omega_0J\lessapprox0.1\omega$ and $\frac{g}{\omega}^2J^2\lessapprox0.1\omega$.
As an example, let us consider now 100 atoms. The upper bounds for $\omega_0$ and $g$ are then $10^{-3}\omega$ and $3\times 10^{-3}\omega$, respectively. If we take $\omega_0=10^{-6}\omega$, then the critical coupling can be calculated,
\begin{equation}
g_{\rm crit}=\sqrt{\frac{\omega_0\omega}{N}}=10^{-4}\omega,
\end{equation}
which is within the approximation.
\section*{DISCUSSION}
In this work, we have developed a useful framework to analyze the dispersive dynamics of the low-qubit frequency region of the Dicke model. This can be phrased in terms of effective Hamiltonians that generate dynamics in isolated dispersive chains of the system Hilbert space, whose detailed structure depends on the value of the coupling.
Several remarks are in order. (i) The quantity plotted in most graphs, P(t), defined in Eq.~(\ref{P(t)}), is particulary convenient for showing the possible discrepancies between the complete and the effective evolutions because the state $\ket{J_x=0,n=0}$ is invariant under the displacement operators of Eq.~(\ref{EvolOpt}). However, a general initial state will be affected by such transformations and the complete evolution is bound to be complicated because the displacement parameter depends on a state's value of $J_x$. This would mask the simple nature of the effective dynamics. (ii) While the Lipkin-Meshkov-Glick model resulting from the adiabatic elimination of the photon field is well known, the results of this paper consider it as a special case of the dispersive chain dynamics in which transitions among spin states are not accompanied by changes in photon number. Thus, effects that have been discussed in this setting may be investigated in the context of more complex chains. (iii) Corrections to the effective dynamics described in the paper are of two kinds: there are secular corrections that appear as additional contributions to the effective Hamiltonians and there are small (of order $\frac{\omega_0}{\omega}$) periodic (with frequency $\omega$) coherent transitions to other states.
|
\section{Introduction}
\subsection{Énoncé du résultat principal}
On dit qu'un entier $n$ est $y$-friable si, et seulement si, son plus grand facteur premier noté $P^+(n)$ est $\leq y$. Nous adopterons la convention $P^+(1)=1$.
Les travaux d'Hildebrand et Tenenbaum~\cite{HT} complétés par ceux de Saias~\cite{Saias} permettent d'évaluer asymptotiquement $\Psi(x,y):=\card\{ S(x,y)\}$ où $S(x,y):=\{ n\leq x\,:\, P^+(n)\leq y\}$ dans un large domaine en $y$.
Déterminer la densité des entiers $y$-friables dans une suite polynomiale définie par $\{ Q(n)\}$ où $Q$ est un polynôme de $\ZZ[X]$ est une question naturelle. Malheureusement seul le cas des polynômes de degré $1$ a été pour l'instant résolu. Nous pourrons nous reporter à~\cite{FT,LB98} pour l'uniformité en fonction des coefficients.
Dans le cas du degré $\geq 2$, même une majoration du bon ordre de grandeur semble intéressant et permettrait d'importantes applications.
Lorsque $Q$ est un polynôme de $\ZZ[X]$, posons $$\Psi_Q(x,y):=\card\{ n\leq x\,:\, P^+(Q(n))\leq y\}.$$
Puisque l'on peut trivialement se ramener au cas des polynômes sans facteur carré, nous factorisons
\begin{equation}\label{factorisationQ}
Q(X):=\prod_{j=1}^r Q_j(X)
\end{equation}
où $Q_j$ sont des polynômes irréductibles non proportionnels de $\ZZ[X]$ de degré $d_j$. Martin~\cite{Martin} a conjecturé l'équivalent asymptotique
$$\Psi_Q(x,y)\sim x\prod_{j=1}^r \frac{\Psi (x^{d_j},y)}{x^{d_j}}$$ dans un large domaine en $y$.
En désignant par $\rhod$ la fonction de Dickmann et en utilisant~\cite{Hild86}, lorsque $y\geq \exp\{ (\log\log x)^{3/5+\varepsilon}\} $, nous espérons donc
$$\Psi_Q(x,y)\sim x\prod_{j=1}^r \rhod(d_ju)$$
où $u:=(\log x)/(\log y)$.
Lorsque~$(d_1,r)\neq (1,1)$, cette relation n'a été établie que certains cas, et lorsque~$u$ est restreint à de petits intervalles bornés. Nous renvoyons à la section 4.3 du survol~\cite{Granville}.
C'est donc à l'aune de cette prévision que l'on pourra mesurer la qualité des majorations que nous établirons.
Ainsi, nous conjecturons
\begin{equation}\label{eq:psiQ-conj}
\Psi_Q(x,y)=x\rho(u)^{\sum_{j=1}^r d_j+o(1)}
\end{equation}
lorsque $u$ tend vers l'infini dans un large domaine en $y$.
Lorsque $Q$ est un polynôme en plusieurs variables, il est possible d'obtenir certaines estimations asymptotiques et nous renvoyons le lecteur intéressé aux articles de~\cite{BBDT,Lachand2, Lachand3}. Dans le cas d'un polynôme univarié, qui nous intéresse ici, des minorations du bon ordre de grandeur~$\Psi_Q(x, y)\gg x$ sont établies par Dartyge, Martin et Tenenbaum~\cite{DMT} pour des petites valeurs de~$u$. L'objectif du présent travail est de montrer des majorations de~$\Psi_Q(x, y)$ qui approchent aussi près que possible la taille conjecturée~\eqref{eq:psiQ-conj}.
Une conséquence de nos résultats simple à énoncer est la suivante.
\begin{theoreme}\label{th1simple} Soit $Q$ un polynôme de $\ZZ[X]$ fixé avec la factorisation \eqref{factorisationQ} et $d_1\leq d_2\leq \cdots\leq d_r$. Soit~$\psi:\RR_+\to\RR_+$ telle que~$\lim_{x\to+\infty}\psi(x)=+\infty$. Nous avons
$$\Psi_Q(x,y)\ll x\rhod(u)^{c_Q+o(1)}$$
lorsque~$x$ tend vers l'infini,~$(\log x)^{\psi(x)} \leq y\leq x$ et
$$ \def\arraystretch{1.2}
\begin{array}{rll}
c_Q = {}& \kern-0.7em \tfrac85 & \qquad (d_1=d_2=1,\, r\geq 2), \\
c_Q = {}& \kern-0.7em \tfrac32 & \qquad (d_1=1,\, d_2\geq 2), \\
c_Q = {}& \kern-0.7em 1+\tfrac{25}{178} & \qquad (d_1=2) \\
c_Q = {}& \kern-0.7em 1 & \qquad (d_1\geq 3,~\text{\cite{Khmyrova}}).
\end{array} $$
\end{theoreme}
Nous notons que~$\frac{25}{178} = \frac1{7,\!12} \approx 0,\!1404$. Le dernier cas permet de retrouver un résultat de Khmyrova~\cite{Khmyrova}, précisé ultérieurement par Timofeev~\cite{Timofeev}.
Chacune de ces majorations repose sur une inégalité essentiellement du type
\begin{equation}
1_{P^+(n)\leq y}(n)\leq \sum_{\substack{d\mid n\\ P^+(d)\leq y\\ R/y<d\leq R }}1\label{eq:majo-basique-crible}
\end{equation}
pour tout $R\geq n$. Le choix du paramètre $R$ doit être optimisé en fonction des résultats d'équirépartition à notre disposition. Dans le premier cas, il s'agira d'une rapide adaptation des récents travaux du second auteur, dans le deuxième, du théorème de Harper de type Bombieri--Vinogradov pour les entiers friables. Dans le cas quadratique, nous développons une amélioration d'un résultat d'équirépartition d'Iwaniec, alors que dans le dernier cas, nous choisirons $R=x$ sans pouvoir utiliser de résultats spécifiques d'équirépartition.
Notons aussi que d'après~\cite{GY-Sunits}, un polynôme~$Q$ avec la factorisation~\eqref{factorisationQ} et de degré au moins~$2$, ne peut pas prendre des valeurs trop friables~: en effet, il découle du corollaire~4 de~\cite{GY-Sunits} que
$P^+(Q(n))\gg ( \log_2n)(\log_3n)/\log_4n$
lorsque $n$ tend vers l'infini\footnote{Ici $\log_k$ désigne la $k$-ième itérée du logarithme.}. Ainsi, par exemple, la $y$-friabilité de $n$ et celle de $n+1$ ne sont pas indépendantes lorsque $y$ prend des très petites valeurs par rapport à la taille de $n$.
\subsection{Applications du cas linéaire}
À titre d'illustration de l'efficacité de notre méthode dans le cas de deux facteurs linéaires, nous appliquons ces estimations à l'étude de l'ensemble~$\cC$ des entiers divisibles par le carré de leur plus grand facteur premier,
$$ \cC = \{n\geq 1:\ P^+(n)^2 | n\}. $$
Il est aisé d'établir que lorsque~$x$ tend vers l'infini,
$$ |\cC \cap [1, x]| = \sum_p \Psi\Big(\frac{x}{p^2}, p\Big) = x\e^{-(1+o(1))\sqrt{2(\log x)\log\log x}}. $$
L'ensemble des entiers~$n$ tels que~$(n, n+1)\in\cC^2$ est en revanche beaucoup plus délicat à étudier. Nous établissons la majoration suivante, qui améliore~\cite{DKDL}.
\begin{theoreme}\label{th:dkdl}
Lorsque~$x$ tend vers l'infini, l'estimation
\begin{equation}
|\{n\leq x:\ (n, n+1)\in\cC^2\}| \leq x \e^{-(c+o(1))\sqrt{2(\log x)\log\log x}}\label{eq:majo-DKDL}
\end{equation}
a lieu avec~$c=4/\sqrt{5} \approx 1,\!789 $.
\end{theoreme}
\begin{remarque}
Dans~\cite{DKDL}, les auteurs obtiennent~$c=25/24 \approx 1,\!042$. Conjecturellement, la valeur optimale attendue est~$c=2$. Cela correspond à l'heuristique que les événements~$n\in\cC$ et~$n+1\in\cC$ surviennent de façon statistiquement indépendante.
\end{remarque}
Une autre application concerne la minoration du nombre d'entiers friables dans les petits intervalles établie au Théorème 5 de \cite{G16}. Lorsque $\varepsilon\in \mo{]}0,1/6\mc{[}$, il existe $u_0=u_0(\varepsilon)$ tel que lorsque $x\geq 2 $, $u_0\leq u\leq (\log x)^{1/6-\varepsilon}$ et $ \rhod(u)^{-(3+\varepsilon)}\leq h\leq \sqrt{x}$, l'on ait
$$ \Psi(x+h\sqrt{x},x^{1/u})-\Psi(x ,x^{1/u})\geq \rhod(u)^2\frac{h\sqrt{x}}{(\log x)^3}. $$
La méthode repose de manière cruciale sur des majorations de cardinal d'ensembles d'entiers~$n$ tels que les valeurs de deux formes affines en~$n$ soient simultanément friables. Les majorations obtenues au Théorème 1.1 avec un exposant $c_Q = 8/5$ dans le cas $r=2$, $d_1=d_2=1$ permettent d'obtenir des résultats intéressants à ce sujet et de remplacer l'exposant $3+ \varepsilon$ par $3/2+\varepsilon$ dans la minoration de $h$ ci-dessus. La valeur conjecturelle $c_Q=2$ fournirait un exposant~$1/2+ \varepsilon$.
\subsection{Niveau de répartition de~$\{n^2-D\}$}
Notre valeur~$c_Q = 1+25/178$ pour~$Q$ quadratique est obtenue grâce à une amélioration d'un résultat d'Iwaniec~\cite{Iwaniec} (voir aussi~\cite{RJLO}) concernant le niveau de répartition des valeurs d'un polynôme quadratique. Ce résultat étant susceptible d'avoir un intérêt intrinsèque, nous le reportons ici.
\begin{theoreme}\label{th:exporep-quad-intro}
Soient~$\eta>0$,~$x\geq 1$, $Q\leq x^{1+25/178-\eta}$, et~$(\lambda(q))$ une fonction arithmétique bornée et ``bien factorisable'' au sens de~\cite[page~199]{FoIw}. Soit~$D\in\ZZ$ qui n'est pas un carré d'entier, et~$V:\RR\to\CC$ une fonction lisse à support compact inclus dans~$\RR_+^\ast$. Alors
\begin{equation}
\begin{aligned}
\sum_{q\leq Q} \lambda(q) \Big(\ssum{n\in\NN \\ q|n^2-D} V\Big(\frac nx\Big) - x{\hat V}(0) \frac{\rhoQ(q)}{q}\Big) \ll_{\eta, V, D} x^{1-\eta/3}
\end{aligned} \label{eq:exporep-quad-intro}
\end{equation}
avec~${\hat V}(0) = \int_\RR V(t)\dd t$ et~$\rhoQ(q)=|\{\nu\mod{q}, \nu^2\equiv D\mod{q}\}|$.
\end{theoreme}
Le gain~$25/178 = 1/7,\!12$ améliore le gain correspondant~$1/15$ de~\cite{Iwaniec}, ainsi que le gain~$1/9$ qui découlerait d'une utilisation de la conjecture~$R^\ast$ de Hooley sur les sommes d'exponentielles incomplètes (\textit{cf.}~\cite[page~185]{Iwaniec}). Pour obtenir cette amélioration, nous adaptons un argument de T\'o{th}~\cite{Toth}, en utilisant une récente extension~\cite{D-Kuz} par le second auteur des majorations de sommes d'exponentielles de Deshouillers-Iwaniec~\cite{DI}.
Les résultats qui sous-tendent le Théorème~\ref{th:exporep-quad-intro} s'appliquent également au problème de Tchebychev de minorer la fonction~$P_D(x) = P^+(\prod_{x<n\leq 2x}(n^2-D))$, pour l'historique duquel nous référons à~\cite{Hooley-Pplus} (voir aussi~\cite{Dartyge}). Le dernier résultat en date sur cette question précise, du à Deshouillers et Iwaniec~\cite{DI-Pplus}, implique en particulier que
$$ P_D(x) \geq x^{1,2024} $$
pour tout~$x$ suffisamment grand. D'un autre côté, la conjecture de Selberg sur les valeurs propres du Laplacien hyperbolique (\textit{cf.} \cite{DI}, section 1.3) permettrait d'obtenir, pour tout~$\ee>0$ suffisamment petit, la valeur $\sqrt{3/2}-\ee \geq 1,2247$ en exposant. Nos résultats permettent de rendre explicite, dans les arguments de Deshouillers-Iwaniec~\cite{DI-Pplus}, la dépendance vis-à-vis de la conjecture de Selberg.
\begin{corollaire}\label{cor:DI-P}
Pour~$\theta\in[0, 1/4]$, définissons~$\kappa(\theta)\in[1, 2]$ comme l'unique réel satisfaisant
$$ \int_1^{\kappa(\theta)} \frac{t\dd t}{1-2\theta t} = \frac1{4(1-2\theta)}. $$
Pour tout~$\ee>0$ et~$D\in\ZZ$ qui n'est pas un carré d'entier, nous avons
$$ P_D(x) \gg_{\ee, D} x^{\kappa(\theta)-\ee} $$
pour tout~$\theta\geq0$ qui est admissible pour la conjecture de Ramanujan-Selberg. En particulier,~$\theta=7/64$ convient~\cite{Kim}; nous avons donc pour~$x$ suffisamment grand
\begin{equation}
P_D(x) \geq x^{1,2182}.\label{eq:Pplus-incond}
\end{equation}
\end{corollaire}
\begin{merci}
Les auteurs prennent plaisir à remercier Adam Harper, Cécile Dartyge et Gérald Tenenbaum pour des discussions sur le présent travail.
|
\section{Introduction}
Respiratory mucus is found in the conducting airways covering the ciliated epithelium. The mucus is typically split into two layers, the periciliary layer between the cilia and the top layer forming a viscoelastic gel \cite{Button2012}. The mucus layer protects the epithelium from inhaled particles and foreign materials due to its sticky nature. Accumulation of these materials is avoided as a result of the coordinated beating of the cilia the so-called mucociliary clearance. The mucus together with the mucociliary escalator of the conducting airways is a very efficient clearance mechanism also preventing efficient drug delivery across this barrier.\\
This respiratory mucus, composed from mucin macromolecules, carbohydrates, proteins, and sulphate bound to oligosaccharide side chains\cite{Fuloria2000,Henning2008} forms a biological gel with unique properties\cite{Schuster2013}. The interaction of all kind of inhaled drugs and drug carriers with this layer and the penetration potential in and through the mucus is of outmost importance for possible therapeutic approaches.\\
Clearly, for drug delivery purposes the biochemistry of penetrating objects plays an important role but also the rheological behavior of the mucus layer. The rheological properties of mucus have been already investigated in many studies, most of them focusing on human tracheal mucus \cite{King1977,JeanneretGrosjean1988,Rubin1990,Zayas1990,Gerber2000} but they also include the examination of cystic fibrosis sputum\cite{Dawson2003,Forier2013,Forier2014}, cervicovaginal mucus\cite{Lai2009b}, gastropod pedal mucus\cite{Ewoldt2007}, as well as pig intestinal mucus\cite{Macierzanka2011}. An excellent overview on the rheological studies is given by Lai et al.\cite{Lai2009}. Since typically only small amounts of mucus are available for experiments, microscopic methods like magnetic microrheometry with test beads of the size of $50\,\mathrm{\mu m}$ to $150\,\mathrm{\mu m}$ were already applied in the 1970's \cite{King1977}. Multiple particle tracking (MPT) has evolved to one of the most favored methods in context with the microrheological characterization of biological fluids in general and of mucus in particular\cite{Oelschlaeger2008}. Still, the number of microrheological studies where the viscoelastic moduli are determined from the Brownian fluctuations spectrum of colloidal probes remain limited \cite{Lai2009b}. One important observation in this particular study of Lai et al. was that the viscosity observed using a $1\,\mathrm{\mu m}$ sized colloidal probe is much smaller than the results obtained on the macroscale. The results were interpreted with a model that assumes that the colloidal probe used can diffuse almost freely through the polymeric mucin network. In consequence, the influence of a variety of particle coatings has been examined extensively during the past decade with the goal to optimize particle transport through this natural barrier\cite{Dawson2003, Lai2007,Lai2009b,Macierzanka2011,Yang2011,Froehlich2014}. Only recently, it was shown by use of active microrheology and cryogenic-scanning-electron-microscopy (CSEM) \cite{Kirch2012b} that mucus should have a porous structure on the micron scale. The active manipulation of immersed particles offers a deeper insight into the material properties of mucus, especially into the strength of its scaffold. A further step was to demonstrate, that passive immersed particles show a very heterogenous diffusion behavior, ranging from particles firmly sticking to the supposed scaffold and particles moving almost freely in an viscous environment \cite{Murgia2016}. However, so far, studies utilizing optically trapped microparticles have been scarce although they are able to greatly enhance our understanding of material properties. They enable the mapping of pore sizes and, by taking the local mobility of particles into account, allow to distinguish in an unambiguous way between a weak and a strong confinement. By utilizing strong optical traps, the rigidity of the mucus mesh can be probed in order to determine which forces the material is able to resist to.
In this study, we will first use a sophisticated linear response theory based on the Kramers-Kronig relation in order to obtain the microscopic complex loss and storage modulus. Due to the heterogeneity of the mucus, these values show a significant scattering, especially if compared to our model gel, a hydroxyethylcellulose gel (HEC). While the mucins in the mucus form the gel network by non-covalent interchain interactions, the HEC is a classical hydrogel without any covalent interchain interactions. Therefore one might expect certain differences, but an explanation for the cause of the large heterogeneity of the mucus is still missing. Additionally we compare our microscopic data to results obtained by macroscopic oscillatory shear rheometry. The results from the microscopic and macroscopic measurements are in perfect agreement for the HEC gel, while there is a huge difference for the mucus that seems to be much stiffer on the macroscopic scale. The CSEM images allow to hypothesize a foam like structure for the mucus with a comparable rigid scaffold and pores with "walls" that are filled with a solution of low viscosity and elasticity, compared to the mesh like structure of HEC. By evaluating the volume percentage of the pores compared to the scaffold we can estimate its elastic module by use of a foam model. Clearly, the biochemistry of penetrating objects plays an important role in the diffusional properties of the mucus but we will show that it has also unique viscoelastic properties that differ strongly from synthetic gels. We postulate that both aspects need to be considered for drug delivery to the airways using particulate carriers.
\section{Materials \& Methods}
\subsection{Sample gels}
All our experiments on mucus were performed with native respiratory horse mucus. It was obtained during bronchoscopy from the distal region of four healthy horses and stored at $193\,\mathrm{K}$ until use. According to earlier studies, such storage conditions are not known to influence the material properties\cite{Gastaldi2000}. As a synthetic model gel for comparison, a $1\,\%$ (w/w) hydroxyethylcellulose gel (HEC; Natrosol 250 HHX Pharm, Ashland Aqualon Functional Ingredients) was chosen because it had similar viscoelastic moduli on the microscale. For the microrheology two kinds of particles were used, polymethylmethacrylate (PMMA) beads with a size of $4\,\mathrm{\mu m}$ and melamin resin beads with a size of $5\,\mathrm{\mu m}$ (Sigma-Aldrich). A Gene Frame (art.-no. AB-0576, ABgene, Epsom, United Kingdom) was used in microrheology as a sample cell to handle the low sample volume of $25\,\mathrm{\mu l}$.\\
In preparation of the experiments, HEC was dissolved in water and shaken gently for 24 hours. For the microrheology, approximately $2-4\,\mathrm{\mu l}$ of each particle suspension (solid content: $10\,\%$) were mixed with $100\,\mathrm{\mu l}$ of sample resulting in particle concentrations of less than $1\,\%$. Thus, hydrodynamic interactions between multiple particles are negligible. These samples were vortexed for about 5 minutes before use to make sure that the beads were distributed homogeneously. Afterwards, a Gene Frame was filled with the respective amount of sample and sealed airtight using a coverslip. No additional preparation of the samples was necessary for experiments in the cone and plate rheometer. All experiments in both setups were performed at room temperature.
\subsection{Macrorheology}
A rotational Mars II (Thermo Scientific GmbH, Karlsruhe, Germany) was used to perform the small and large amplitude oscillatory shear (SAOS and LAOS) experiments. With SAOS experiments the linear response of the material is tested, whilst LAOS experiments are used to characterize the nonlinear properties. First strain amplitude sweeps were performed in order to determine the region of linear response and the nonlinear properties of both materials and then a frequency sweep in the linear range was performed. The rheometer was equipped with a cone and plate geometry with a cone angle of $0.5^\circ$ for the measurements on mucus and a second geometry with an angle of $2^\circ$ in case of the HEC gel. In case of mucus, this enabled us to perform measurements on volumes as small as $500\,\mathrm{\mu l}$ with an acceptable signal-to-noise ratio. In case of HEC, bigger sample volumes were available so using the more sensitive $2^\circ$ geometry was a feasible option.
\subsection{Microrheology}
The optical tweezers setup described in Ref. \cite{Ziehl2009b,Kirch2012b} was used to perform passive microrheology. Particle positions in the focus of the laser beam were recorded with a high speed camera (HiSpec 2G; Fastec Imaging) at a frame rate of $16\,\mathrm{kHz}$. The recorded picture series were analyzed using a particle tracking algorithm based on the cross-correlation of successive images\cite{Ziehl2009b}. The complex shear modulus $G^*$ was then determined by applying a method proposed by Schnurr\cite{Schnurr1997}. For this purpose, the Langevin equation describing the interaction of the confined bead with its surroundings is recast in frequency-space in such a way that particle displacements $\tilde{x}$ and the Brownian random force $\tilde{F}_r$ are linked by the susceptibility or compliance $\tilde{\alpha}$
\begin{align}
\tilde{x}(\omega)&=\tilde{\alpha}^*(\omega)\tilde{F}_r(\omega)\,,
\end{align}
where
\begin{align}
\tilde{\alpha}^*(\omega)&=\frac{1}{k-i\omega\tilde{\zeta}(\omega)}\,.
\end{align}
The susceptibility is a function of the trap stiffness $k$ and the frequency-dependent friction coefficient $\zeta$. It is a complex quantity whose imaginary part is related to the power spectral density of particle displacements $\left<\left|\tilde{x}(\omega)\right|^2\right>$ by the fluctuation-dissipation-theorem\cite{Landau}
\begin{align}
\left<\left|\tilde{x}(\omega)\right|^2\right>&=\frac{2k_BT}{\omega}\alpha''(\omega)
\end{align}
with Boltzmann's constant $k_B$ and the temperature $T$. The Kramers-Kronig-relations allow the determination of the real part of the compliance by computing the principal value integral
\begin{align}
\tilde{\alpha}'(\omega)&=\frac{2}{\pi}P\mspace{-16.0mu}\int\limits_{0}^{\infty}\frac{\omega\tilde{\alpha}''(\omega)-\epsilon\tilde{\alpha}''(\epsilon)}{\epsilon^2-\omega^2}d\epsilon\,.
\end{align}
The function contained within the integral encompasses two poles at $\epsilon=\pm\omega$ which are excluded from integration by the means of the principal value integral indicated by the letter ``P'' in the integration symbol. Finally, the relation of the compliance and the complex shear modulus $G^*$ is given by
\begin{align}
\tilde{G}^*(\omega)&=\frac{1}{6\pi R_c}\cdot\frac{1}{\tilde{\alpha}^*(\omega)}\,,
\end{align}
where $R_c$ is the particle radius. The dependence of the complex shear modulus on the particle size given in this equation is the general one which arises due to the increasing drag force when choosing larger spheres. However, it does not include additional influences like for example caging effects of the spheres in pockets of a porous material like mucus. Such size dependencies which are caused by inhomogeneous structures within a fluid can be explicitly studied by varying the particle size (see for example \cite{Lai2009b}). This was not conducted in our study, though.\\
Just as in case of the macrorheologic shear modulus, the microrheologic shear modulus as well is composed of the elastic contribution $G'$ and the viscous contribution $G''$, where $G^*=G'+iG''$. However, due to the presence of the optical trap, there is an additional elastic contribution $G'_{trap}=k/6\pi R_c$ which has to be subtracted from the measured $G'$ in order to gain the actual sample properties. While it is possible to perform an online calibration of the trap stiffness in Newtonian fluids this is not possible in complex fluids like mucus. Thus, separate measurements with colloids in water were performed beforehand in a separate sample cell for this purpose using both the equipartition and the drag force method\cite{Capitanio2002}. Typically, the stiffness ranged between $3\,\mathrm{pN/\mu m}$ and $8\,\mathrm{pN/\mu m}$. Due to experimental restrictions in terms of the duration of a measurement as well as the influence of a translational drift a frequency of $1\,\mathrm{Hz}$ was chosen as the lower frequency cutoff. Hence, the microrheologic shear modulus is only given starting from a frequency of $1\,\mathrm{Hz}$. There is an upper frequency cutoff as well which is defined by the Nyquist sampling theorem as half of the recording frequency, i.\,e. $8\,\mathrm{kHz}$ in our case. In order to minimize aliasing errors, which may be caused due to the Fourier-transforms, we chose a value of $3.5\,\mathrm{kHz}$ well below the Nyquist frequency as the upper cutoff, instead.
\subsection{Cryo-SEM}
Cryo-SEM images were taken as described in Ref. \cite{Kirch2012b}. Sample gels were filled in a thin dialysis capillary and immediately frozen in liquid
propane to only allow formation of amorphous water and circumvent formation of crystalline water. Capillaries were cut to expose the brim to sublimation of the amorphous water inside the gels. Finally the surface of the dry polymer scaffold was sputter-coated with platinum and samples were transferred into the SEM (DSM 982 Gemini; Zeiss) and imaged at $-120^\circ\,\mathrm{C}$ ($5\,\mathrm{keV}$, $5\,\mathrm{mm}-6\,\mathrm{mm}$ working distance).
Additional CSEM measurements were performed with a JSM-7500F SEM (JEOL, Tokyo, Japan) equipped with an Alto 2500 Cryo transfer system (Gatan, Abingdon, UK). Respiratory horse mucus was placed between two metal freezing tubes (Gatan, Abingdon, UK) and the samples were frozen by plunging into liquid nitrogen. Inside the cryo transfer system the upper tube was knocked off to create a fracture surface and sublimation was performed for 15 min at 178$^\circ$K. Samples were sputter-coated with platinum at 133$^\circ$K, transferred to the SEM cryo-stage and imaged at 133$^\circ$K and 5 kV acceleration voltage (working distance 8.0 mm).
CSEM images were analyzed by ImageJ 1.48v software (National Institutes of Health, USA) to determine the fraction of pore volume in the mucus. The relation of pore area to measured surface area at the brim was assumed to correspond to the relation of pore volume to mucus volume. Image contrast and brightness was adjusted appropriately and a threshold was set to distinguish the inside of the pores from the pore walls (Fig.\ref{MucusCSEM2} b). Pore areas were determined by the program using the \textit{Analyze Particles} function (Fig.\ref{MucusCSEM2} c) ). The sum of the pore areas was related to the total image area. 6 images with an overall area of 1458 $\mu m^2$ were analyzed.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{MucusCSEM2.png}
\caption{Pore size analysis of CSEM images: The original images (a) were processed and a threshold was set (b). The determined pore areas are displayed in light blue (c).}
\label{MucusCSEM2}
\end{figure*}
\section{Results}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{HECMucusOT.png}
\caption{Averaged results of 11 independent microrheological measurements for (a) the elastic modulus $G'$ and (b) the viscous modulus $G''$ for the $1\,\mathrm{\%\,(w/w)}$ solution of hydroxyethylcellulose (HEC). Respective average results of 17 independent measurements are depicted for (c) the elastic and (d) the viscous modulus in mucus. Measurements were performed at different locations within each sample which resulted in values within the depicted shaded regions.}
\label{HECMucusOT}
\end{figure*}
The shear modules from the microrheological measurements are shown in Fig. \ref{HECMucusOT}. Data sets were recorded by confining particles in the focus of the optical tweezers at different locations within the bulk of the sample. The average values of more than 10 measurements are depicted by symbols while the regions in which all values are distributed are drawn as shaded areas. Both the elastic and the viscous shear modulus of mucus and the HEC gel are in the range from $1\,\mathrm{Pa}$ to $30\,\mathrm{Pa}$. In case of HEC (Fig. \ref{HECMucusOT}(a) and (b)), the shear modulus shows a limited variance when switching locations within the sample, but for the case of mucus (Fig. \ref{HECMucusOT}(c) and (d)), this variability is significantly enhanced, especially in the intermediate frequency range. For mucus, both viscous and elastic shear moduli increase monotonically and reach a plateau eventually. These results agree with earlier observations\cite{Kirch2012b}. The HEC data sets cannot be compared directly to that former study since in the present study a higher concentration of $1\,\%$ was chosen to give a better representation of the microrheologic properties of mucus. Nonetheless, besides the larger scatter for the mucus, the results for both the passive microrheology of mucus and of the HEC gel in our actual study are quite comparable, i.\,e. the absolute values are very similar, they lay in the same order of magnitude and even their functional behavior in our accessible frequency domain is almost indistinguishable.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{HECMucusAmpSweep.png}
\caption{Lissajous plots of LAOS sweeps for (a) HEC and for (c) native mucus. (b) and (d): the respective shear moduli vs. strain amplitude. Different symbols indicate different independent measurements.}
\label{HECMucusAmpSweep}
\end{figure*}
A completely different result is found in the macrorheology. Results from large amplitude oscillatory shear (LAOS) experiments are shown as shear stress versus shear strain plots, i.\,e. Lissajous plots, together with the respective shear modulus versus strain amplitude (Fig. \ref{HECMucusAmpSweep}). While the Lissajous plots for HEC gels are always elliptic within the examined strain amplitude range (Fig. \ref{HECMucusAmpSweep}(a)), this is not the case for mucus (Fig. \ref{HECMucusAmpSweep}(c)). Instead of ellipses, the curves deform into parallelograms when exceeding a strain amplitude of $\gamma=100\,\%$. While the response of a linear viscoelastic material typically has the shape of an ellipse in a Lissajous plot \cite{Ewoldt2007}, deviations indicate a non-linear response which is the case for mucus. This is also confirmed by the shear modulus versus strain amplitude plots (Fig. \ref{HECMucusAmpSweep}(b) and (d)). While in case of HEC both the elastic and the viscous modulus only show weak changes up to strain amplitudes of $\gamma=300\,\%$, in case of mucus a significant decrease becomes apparent for both. The onset of this decrease in $G'$ can already be observed at $\gamma=30\,\%$. When exceeding a value of $100\,\%$, it additionally becomes apparent in $G''$. This nonlinear behavior is an indication of the particular behavior of mucus. However, in order to avoid higher harmonics in the small amplitude oscillatory (SAOS) linear response measurements, the shear strain has to be kept below this onset of nonlinearity. For the HEC model gel, the critical shear amplitude is $\gamma \approx 300\,\%$ and for the mucus $\gamma \approx 20\,\%$. Thus, for HEC a constant strain amplitude of $25\,\%$ and for the mucus a much lower value of $1\,\%$ for the frequency sweep was used.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{HECMucusFreqSweep.png}
\caption{Shear modulus from the macro- and microrheology for (a) the HEC gel and (b) the native mucus. Error bars for SAOS experiments are standard deviations from different measurements. In case of HEC, a fit with a two-component Maxwell model is shown as lines.}
\label{HECMucusFreqSweep}
\end{figure*}
After completion of the amplitude sweep, a series of frequency sweeps was performed with the same sample. Using the strain amplitudes determined during the amplitude sweep, frequencies between $10^{-2}\,\mathrm{Hz}$ and $10^1\,\mathrm{Hz}$ were applied stepwise with five repetitions each to reduce the influence of noise while keeping the total duration of the experiment as short as possible. A short measurement duration was important to avoid evaporation of the samples. For both the HEC gel and mucus, the average of three of these sweeps is shown in Fig. \ref{HECMucusFreqSweep}. In the measured frequency range from $10^{-2}\,\mathrm{Hz}$ to $5\,\mathrm{Hz}$ we find a monotonous increase in the moduli for the HEC gel but for the mucus already a roughly constant plateau is observed. Furthermore, the HEC gel shows a viscous behaviour at low frequencies while the mucus has a higher elastic modulus for all frequencies. This is most likely a consequence of the strong non-covalent interchain interactions of the mucins. In the same graph, we plot the averaged data from the microrheology (Fig. \ref{HECMucusOT}). Here, the most striking differences between mucus and the HEC gel becomes apparent. For the HEC, we observe a continuous transition from the macro- to the microrheologic data. It is even possible to fit the combined SAOS and microrheology data approximately with the two-component Maxwell fluid model that consists of a viscoelastic contribution for the polymeric part and a Newtonian contribution for the solvent. Deviations from the model occur for $G'$ at frequencies below $1\,\mathrm{Hz}$. In principle one could improve the agreement between the fit and the data by incorporating more relaxation times but the additional physical insight will be limited. One crossover frequency between elastic and viscous part is visible at $6\,\mathrm{Hz}$ and a second crossover might be present above $4\,\mathrm{kHz}$, however, it can not be verified in the scope of our experiments since the relevant frequencies lie outside of the accessible spectrum. Thus, the HEC gel behaves mostly as a viscoelastic fluid below $6\,\mathrm{Hz}$ and as a viscoelastic solid above this value.\\
In case of mucus in Fig. \ref{HECMucusFreqSweep}b, no such smooth transition from the macro- to the microrheologic data set is observed. A significant gap between the results gained by both experiments is present which encompasses three to four orders of magnitude. The SAOS data sets indicate that $G'$ and $G''$ are only weakly dependent on the frequency within the probed frequency range. A slightly more pronounced frequency dependence is observed for the microrheology data. However, all values of the viscous and elastic modulus remain between $1\,\mathrm{Pa}$ and $10\,\mathrm{Pa}$ for over more than three orders of magnitude in frequency. This clearly shows that there is a remarkable difference between the viscoelastic properties on the micro- and the macroscale. Of course, it is known that the microrheolical properties of mucus depend on the particle size even well below 1 $\mu m$, but our optical detection method did not allow to explore this regime. In any case, as one expects to find an even lower viscosity for smaller particles, the difference in Fig \ref{HECMucusFreqSweep}b will be even more pronounced.\\
In Fig. \ref{MucusCSEM} the CSEM images of a HEC gel and a mucus sample are shown for two different spatial resolutions. The polymeric network of the HEC shows a typical homogeneous mesh for a gel. The mucus shows a more heterogeneous distribution of polymeric material and especially in the large magnification a heterogeneous porous structure is visible. This scaffold of pore walls is made out of much thicker polymeric material than the polymeric network of the HEC gel.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{HECMucusCSEM.png}
\caption{Cryo-SEM images of (a, b) HEC gel and (c, d) native mucus. For the HEC gel the mesh sizes are rather homogenous. For the mucus the mesh sizes range from tens of nanometers up to several micrometers and the structure resembles more a porous foam like network.}
\label{MucusCSEM}
\end{figure*}
\section{Discussion}
When comparing the microrheologic shear modulus of HEC and mucus (Fig. \ref{HECMucusOT}) we find similar viscoelastic properties. Both the elastic as well as the viscous modulus show a comparable response spectrum. It should be noted, though, that the local properties in mucus vary more significantly which is due to the heterogeneity of the material that could be observed in CSEM images. At frequencies above $10^2\,\mathrm{Hz}$, $G'$ roughly stays constant at $15\,\mathrm{Pa}$, a value that is significantly below the value of $6\cdot 10^{3}\,\mathrm{Pa}$ that is found in the macrorheology at a frequency of $1\,\mathrm{Hz}$. The LAOS measurements also revealed significant differences between the HEC gel and the mucus. The latter showed a nonlinear response behavior already at strain amplitudes of $\gamma =20\,\%$. A similar behavior was found by Ewoldt et al.\cite{Ewoldt2007} in LAOS experiments with gastropod pedal mucus. From the CSEM images we know that the mucus has a porous structure with a thick scaffold that builds the pore walls. While the rheometer probes the whole bulk of the fluid, the microrheology accesses mostly the contents of the pores which is formed by an aqueous solution of dissolved biopolymers. This structure is very similar to that of a foam. Foams in general consist of a porous material which is filled with another material of much lower stiffness. This foam like structure can be modeled only if we assume significant simplifications. A suitable approach is the Mori-Tanaka model\cite{Mori1973} which considers a foam-like material with elastic walls. In this case, the material is composed of two phases, one of which is the wall material and the other one of which is the material filling the pores. Due to the very large difference in elastic properties we will fully neglect the contribution of the aqueous solution in the pores and then the total macroscopic shear modulus of mucus is linked to the shear modulus of the material of the pore walls by
\begin{align}
G_{total}&=G_{walls}\left(1-\frac{c_p}{1-\beta(1-c_p)}\right)\,,
\end{align}
where $c_p$ is the volume fraction of the filling material and $\beta$ is a dimensionless number. Under the assumption that the wall material is isotropic and homogeneous, it is given by
\begin{align}
\beta&=\frac{2(4-5\nu)}{15(1-\nu)}
\end{align}
with Poisson's ratio $\nu$. Under the assumption of a volume fraction of the pores of $c_p=47\,\%$, which we determined from the CSEM images, while assuming incompressibility of the pore walls ($\nu=0.5$) the actual shear modulus of the wall or scaffold material lies above the values measured by the rheometer by a factor of $2.5$. This means that the gap between macro- and microrheology increases even further when taking material porosity into account. Given that the liquid inside the pores is rheologically comparable to an aqueous solution, the diffusion in mucus can be as fast as in water for small particles \cite{Lai2009b}. For larger particles, size exclusion effects occur. Particles above a certain cut-off size, which is determined by the pore size, can be trapped inside the mucus. However, also smaller particle can be retained in the mucus due to interactions with mucus components \cite{Lieleg2011}.
Our optical tweezers measurements showed the comparable microrheology of HEC gel and respiratory horse mucus. Thus HEC gel might be an appropriate model to study if diffusion of particles through mucus is impeded by size exclusion effects, given the mesh sizes are similar to mucus pore sizes. However, it needs to be considered that retention of particles due to interaction with mucus components cannot be evaluated by using HEC gel.
\section{Conclusion}
Rheological characteristics on the micro- and on the macroscale of native equine respiratory mucus were compared to a synthetic hydroxyethylcellulose (HEC) hydrogel for reference. Our measurements revealed that mucus has peculiar rheological properties that may be best explained by its foam-like microstrucure. This foam like structure is an unique property of the mucus and has to be considered if transport properties of drugs have to be optimized. As the physiologial function of mucus is different at various organs (e.\,g. respiratory, digestive, or reproductive tract), it appears intriguing to investigate whether such differences are also reflected in different structures and rheological properties across various organs and also species.\\
{\indent}Obviously, the entrapment and clearance by mucus as well as the penetration of micro- and nanoparticles by and through mucus, respectively, will strongly depend on the interaction with mucus and the particular path taken by such objects. Besides the chemistry of the interacting object the mucus behavior due to its structure is essential. Knowledge of the basic structure and the understanding of the impact of those structural and functional features of mucus will have important bearings for the design of pulmonary drug delivery systems. Of course, in any realistic situation of physiological relevance, the local ion strength, pH, temperature and local mechanical (shear) stresses will affect the mechanical properties of the mucus. These parameters might not only induce quantitative changes, but future studies have also to reveal if, e.g., under certain circumstances a collapse of the scaffold structure might occur.
\section{Acknowledgement}
Michael Hellwig and Andreas Schaper (Philipps-University Marburg) are acknowledged for assistance in CSEM measurements. We thank the German Research Association (DFG - LE 1053/16-1 and GRK 1276) for financial support. We thank Julian Kirch for the part of the recording of the cryogenic scanning-electron-micrographs. Afre Torge thanks the FiDel-project (“Cystic Fibrosis Delivery”, grant N° 13N12530) for financial support by the German Federal Ministry of Education and Research (BMBF).
\\
\\
|
\section*{References}
|
\section*{Introduction}
\renewcommand{\thetheorem}{\arabic{theorem}}
\subsection*{Context}
In \cite{gompf}, R. Gompf defined a homotopy invariant $\theta_G$ of oriented 2-plane fields in 3-manifolds. This invariant is defined for oriented 2-plane fields $\xi$ in a closed oriented 3-manifold $M$ when the first Chern class $c_1(\xi)$ is a torsion element of $H^2(M;\b Z)$. This invariant appears, for instance, in the construction of an absolute grading for the Heegaard-Floer homology groups, see \cite{GH}.
Since the positive unit normal of an oriented 2-plane field of a Riemannian 3-manifold $M$ is a section of its unit tangent bundle $UM$, homotopy classes of oriented 2-plane fields of $M$ are in one-to-one correspondence with homotopy classes of sections of $UM$. Thus, the invariant $\theta_G$ may be regarded as an invariant of homotopy classes of nowhere zero vector fields, also called \textit{combings}. In that setting, the Gompf invariant is defined for \textit{torsion combings} of closed oriented 3-manifolds $M$, \textit{ie}\ combings $X$ such that the Euler class $e_2(X^\perp)$ of the normal bundle $X^\perp$ is a torsion element of $H^2(M;\b Z)$. \\
In \cite{lescopcombing}, C. Lescop proposed an alternative definition of $\theta_G$ using a Pontrjagin construction from the combing viewpoint. Here, we use a similar approach to show how to define Pontrjagin numbers for torsion combings by using pseudo-parallelizations, which are a generalization of parallelizations. This enables us to define a relative extension of the Gompf invariant for torsion combings in all compact oriented 3-manifolds with boundary. We also study the iterated variations under Lagrangian-preserving surgeries of this extended invariant and prove that it is a degree two invariant with respect to a suitable finite type invariant theory. In such a study, pseudo-parallelizations reveal decisive since they are, in some sense, compatible with Lagrangian-preserving surgeries while genuine parallelizations are not.
\subsection*{Conventions}
In this article, compact oriented 3-manifolds may have boundary unless otherwise mentioned. All manifolds are implicitly equipped with Riemannian structures. The statements and the proofs are independent of the chosen Riemannian structures. \\
If $M$ is an oriented manifold and if $A$ is a submanifold of $M$, let $TM$, resp. $TA$, denote the tangent bundles to $M$, resp. $A$, and let $NA$ refer to the orthogonal bundle to $A$ in $M$, which is canonically isomorphic to the normal bundle to $A$ in $M$. The fibers of $NA$ are oriented so that $NA \oplus TA = TM$ fiberwise and the boundaries of all compact manifolds are oriented using the outward normal first convention. \\
If $A$ and $B$ are transverse submanifolds of an oriented manifold $M$, their intersection is oriented so that $N(A\cap B) = NA \oplus NB$, fiberwise. Moreover, if $A$ and $B$ have comple\-mentary dimensions, \textit{ie}\ if $\mbox{dim}(A)+\mbox{dim}(B)=\mbox{dim}(M)$, let $\varepsilon_{A\cap B}(x) = 1$ if $x \in A\cap B$ is such that $T_xA\oplus T_xB = T_xM$ and $\varepsilon_{A\cap B}(x) = -1$ otherwise. If $A$ and $B$ are compact transverse submani\-folds of an oriented manifold $M$ with complementary dimensions, the \textit{algebraic intersection of $A$ and $B$ in $M$} is
$$
\langle A, B \rangle_M = \sum_{x \in A\cap B} \varepsilon_{A\cap B}(x).
$$
Let $L_1$ and $L_2$ be two rational cycles of an oriented $n$-manifold $M$. Assume that $L_1$ and $L_2$ bound two rational chains $\Sigma_1$ and $\Sigma_2$, respectively. If $L_1$ is transverse to $\Sigma_2$, if $L_2$ is transverse to $\Sigma_1$ and if $\mbox{dim}(L_1)+\mbox{dim}(L_2) = n-1$, then the \textit{linking number of $L_1$ and $L_2$ in $M$} is
$$
lk_M(L_1,L_2) = \langle \Sigma_1 , L_2 \rangle_M = (-1)^{\tiny{n-\mbox{dim}(L_2)}}\langle L_1 , \Sigma_2 \rangle_M.
$$
\subsection*{Setting and statements}
A \textit{combing} $(X,\sigma)$ of a compact oriented 3-manifold $M$ is a section $X$ of the unit tangent bundle $UM$ together with a nonvanishing section $\sigma$ of the restriction $X^\perp_{|\partial M}$ of the normal bundle $X^\perp$ to $\partial M$. For simplicity's sake, the section $\sigma$ may be omitted in the notation of a combing. For any combing $(X,\sigma)$, note that $\rho(X)=(X_{|\partial M}, \sigma, X_{|\partial M} \wedge \sigma)$, where $\wedge$ denotes the cross product, is a trivialization of $TM_{|\partial M}$. So, a combing of a compact oriented 3-manifold $M$ may also be seen as a pair $(X,\rho)$ where $X$ is a section of $UM$ that is the first vector of a trivialization $\rho$ of $TM_{|\partial M}$ together with this trivialization. \\
Two combings $(X,\sigma_X)$ and $(Y,\sigma_Y)$ of a compact oriented 3-manifold $M$ are said to be \textit{transverse} when the graph $X(M)$ is transverse to $Y(M)$ and $-Y(M)$ in $UM$. The combings $(X,\sigma_X)$ and $(Y,\sigma_Y)$ are said to be \textit{$\partial$-compatible} when $X_{|\partial M}= Y_{|\partial M}$, $\sigma_X = \sigma_Y$, $X(\mathring M)$ is transverse to $Y(\mathring M)$ and $-Y(\mathring M)$ in $UM$, and
$$
\overline{X(\mathring M)\cap Y(\mathring M)} \cap UM_{|\partial M} =\emptyset.
$$
\iffalse
If $X(\mathring M)$ is transverse to $Y(\mathring M)$ and $-Y(\mathring M)$ in $UM$ and if
$$
\overline{X(\mathring M)\cap Y(\mathring M)} \cap UM_{|\partial M} =\emptyset
$$
then $(X,\sigma_X)$ and $(Y,\sigma_Y)$ are \textit{almost transverse}. Two combings $(X,\sigma_X)$ and $(Y,\sigma_Y)$ of $M$ are said to be \textit{coincide on $\partial M$} if $X_{|\partial M}= Y_{|\partial M}$ and $\sigma_X=\sigma_Y$. \fi
When $(X,\sigma_X)$ and $(Y,\sigma_Y)$ are $\partial$-compatible, define two links $L_{X=Y}$ and $L_{X=-Y}$ as follows. First, let $P_M$ denote the projection from $UM$ to $M$ and set
$$
L_{X=-Y} = P_M (X(M) \cap (-Y)(M)).
$$
Second, there exists a link $L_{X=Y}$ in $\mathring{M}$ such that
$$
P_M (X(M) \cap Y(M)) = \partial M \sqcup L_{X=Y}.
$$
If $(X,\sigma)$ is a combing of a compact oriented 3-manifold $M$, its \textit{relative Euler class} $e_2^M(X^{\perp}\hspace{-1mm}, \sigma)$ in $H^2(M,\partial M;\b Z)$ is an obstruction to extending the section $\sigma$ as a nonvanishing section of $X^\perp$. This obstruction is such that its Poincaré dual $P(e_2^M(X^\perp, \sigma))$ is represented by the zero set of a generic section of $X^\perp$ extending $\sigma$. This zero set is oriented by its coorientation induced by the orientation of $X^\perp$. When $M$ is closed, the \textit{Euler class} $e_2(X^\perp)$ of $X$ is just this obstruction to finding a nonvanishing section of $X^\perp$. \\
A combing $(X,\sigma)$ of a compact oriented 3-manifold $M$ is a \textit{torsion combing} if $e_2^M(X^\perp,\sigma)$ is a torsion element of $H^2(M,\partial M ; \b Z)$, \textit{ie}\ $\left[e_2^M(X^\perp,\sigma)\right]=0$ in $H^2(M,\partial M ; \b Q)$. \\
Let $M_1$ and $M_2$ be two compact oriented 3-manifolds. The manifolds $M_1$ and $M_2$ are said to have \textit{identified boundaries} if a collar of $\partial M_1$ in $M_1$ and a collar of $\partial M_2$ in $M_2$ are identified. In this case, ${TM_1}_{|\partial M_1}= \b R n_1 \oplus T \partial M_1$ is naturally identified with ${TM_2}_{|\partial M_2}= \b R n_2 \oplus T \partial M_2$ by an identification that maps the outward normal vector field $n_1$ to $M_1$ to the outward normal vector field $n_2$ to $M_2$. \\
If $\tau_1$ and $\tau_2$ are parallelizations of two compact oriented 3-manifolds $M_1$ and $M_2$ with identified boundaries such that $\tau_1$ and $\tau_2$ coincide on $\partial M_1 \simeq \partial M_2$, then the \textit{first relative Pontrjagin number of $\tau_1$ and $\tau_2$} is an element $p_1(\tau_1,\tau_2)$ of $\b Z$ which corresponds to the Pontrjagin obstruction to extending a specific trivialization $\tau(\tau_1,\tau_2)$ of $TW \otimes \b C$ defined on the boundary of a cobordism $W$ from $M_1$ to $M_2$ with signature zero (see Subsection~\ref{ssec_defpara} or \cite[Subsection 4.1]{lescopcombing}). In the case of a parallelization $\tau$ of a closed oriented 3-manifold $M$, we get an absolute version. The \textit{Pontrjagin number $p_1(\tau)$ of $\tau$} is the relative Pontrjagin number $p_1(\tau_\emptyset,\tau)$ where $\tau_\emptyset$ is the parallelization of the empty set. Hence, for two parallelizations $\tau_1$ and $\tau_2$ of some closed oriented 3-manifolds,
$$
p_1(\tau_1,\tau_2) = p_1(\tau_2) - p_1(\tau_1).
$$
In \cite{lescopcombing}, using an interpretation of the variation of Pontrjagin numbers of parallelizations as an intersection of chains, C. Lescop showed that such a variation can be computed using only the first vectors of the parallelizations. This led her to the following theorem, which contains a definition of the \textit{Pontrjagin numbers} for torsion combings of closed oriented 3-manifolds.
\begin{theorem}[{\cite[Theorem 1.2 \& Subsection 4.3]{lescopcombing}}] \label{thm_defp1X}
Let $M$ be a closed oriented 3-manifold. There exists a unique map
$$
p_1 \ : \ \lbrace \mbox{homotopy classes of torsion combings of } M \rbrace \longrightarrow \b Q
$$
such that :
\begin{enumerate}[(i)]
\item for any combing $X$ on $M$ such that $X$ extends to a parallelization $\tau$ of $M$ :
$$
p_1([X])=p_1(\tau),
$$
\item if $X$ and $Y$ are two transverse torsion combings of $M$, then
$$
p_1([Y])-p_1([X])= 4 \cdot lk(L_{X=Y},L_{X=-Y}).
$$
\end{enumerate}
Furthermore, $p_1$ coincides with the Gompf invariant : for any torsion combing $X$,
$$
p_1([X])=\theta_G(X^\perp).
$$
\end{theorem}
In this article, we study the variations of the Pontrjagin numbers of torsion combings of compact oriented 3-manifolds with respect to specific surgeries, called \textit{Lagrangian-preserving surgeries}, which are defined as follows. \\
A \textit{rational homology handlebody of genus $g\in \b N$}, or $\b Q$HH for short, is a compact oriented 3-manifold with the same homology with coefficients in $\b Q$ as the standard genus $g$ handlebody. Note that the boundary of a genus $g$ rational homology handlebody is homeomorphic to the standard closed connected oriented surface of genus $g$. The \textit{Lagrangian} of a $\b Q$HH $A$ is
$$
\go L_A := \mbox{ker}\left(i^A_* : H_1(\partial A; \b Q) \longrightarrow H_1(A;\b Q)\right)
$$
where $i^A$ is the inclusion of $\partial A$ into $A$. An \textit{LP$_\b Q$-surgery datum} in a compact oriented 3-manifold $M$ is a triple $(A,B,h)$, or $(\sfrac{B}{A})$ for short, where $A \subset M$, where $B$ and $A$ are rational homology handlebodies and where $h : \partial A \rightarrow \partial B$ is an identification homeomorphism, called \textit{LP$_\b Q$-identification}, such that $h_*(\go L_A)=\go L_B$. Performing the \textit{LP$_\b Q$-surgery} associated with the datum $(A,B,h)$ in $M$ consists in constructing the manifold :
$$
M \left( \sfrac{B}{A} \right) = \left( M \setminus \mathring A \right) \ \bigcup_h \ B.
$$
If $(M,X)$ is a compact oriented 3-manifold equipped with a combing, if $(A,B,h)$ is an LP$_\b Q$-surgery datum in $M$, and if $X_{B}$ is a combing of $B$ that coincides with $X$ on $\partial A \simeq \partial B$, then $(A,B,h,X_{B})$, or $(\sfrac{B}{A},X_B)$ for short, is an \textit{LP$_\b Q$-surgery datum in $(M,X)$}. Performing the \textit{LP$_\b Q$-surgery} associated with the datum $(A,B,h,X_{B})$ in $(M,X)$ consists in constructing the manifold $M \left( \sfrac{B}{A} \right)$ equipped with the combing :
$$
X(\sfrac{B}{A}) = \left\lbrace
\begin{aligned}
& X & &\mbox{on $M\setminus \mathring A$}, \\
& X_{B} & &\mbox{on $B$.}
\end{aligned}
\right.
$$
The main result of this article is a variation formula for Pontrjagin numbers -- see Theorem~\ref{thm_D2nd} below -- which reads as follows in the special case of compact oriented 3-manifolds without boundary.
\begin{theorem} \label{thm_D2nd0}
Let $(M,X)$ be a closed oriented 3-manifold equipped with a combing and let $\lbrace (\sfrac{B_i}{A_i},X_{B_i}) \rbrace_{i \in \lbrace 1,2 \rbrace}$ be two disjoint LP$_\b Q$-surgeries in $(M,X)$ (\textit{ie}\ $A_1$ and $A_2$ are disjoint). For all $I \subset \lbrace 1,2 \rbrace$, let $(M_I,X^I)$ be the combed manifold obtained by performing the surgeries associated to the data $ \lbrace (\sfrac{B_i}{A_i},X_{B_i})\rbrace_{i \in I}$. If $\lbrace X^I \rbrace_{I \subset \lbrace 1,2 \rbrace}$ is a family of torsion combings of the $\lbrace M_I \rbrace_{I\subset \lbrace 1,2 \rbrace}$, then
$$
\sum_{I\subset\lbrace 1 , 2 \rbrace} (-1)^{|I|} p_1([X^I]) = - 2 \cdot lk_M \left(L_{\lbrace X^I \rbrace}(\sfrac{B_1}{A_1}), L_{\lbrace X^I \rbrace}(\sfrac{B_2}{A_2})\right),
$$
where the right-hand side of the equality is defined as follows. For all $i \in \lbrace 1 , 2 \rbrace$, let
$$
H_1(A_i;\b Q) \stackrel{i^{A_i}_*}{\longleftarrow} \frac{H_1(\partial A_i;\b Q)}{\go L_{A_i}} \stackrel{h_*}{=} \frac{H_1(\partial B_i;\b Q)}{\go L_{B_i}} \stackrel{i^{B_i}_*}{\longrightarrow} H_1(B_i;\b Q)
$$
be the sequence of isomorphisms induced by the inclusions $i^{A_i}$ and $i^{B_i}$. There exists a unique homology class $L_{\lbrace X^I\rbrace}(\sfrac{B_i}{A_i})$ in $H_1(A_i; \b Q)$ such that for any nonvanishing section $\sigma_i$ of $X^\perp_{|\partial A_i}$~:
$$
L_{\lbrace X^I \rbrace}(\sfrac{B_i}{A_i}) = i^{A_i}_* \circ (i^{B_i}_*)^{-1}\left(P\big(e_2^{B_i}(X_{B_i}^\perp, \sigma_i)\big)\right) - P\big(e_2^{A_i}(X^\perp_{|A_i}, \sigma_i)\big),
$$
where $P$ stands for Poincaré duality isomorphisms from $H^2(A_i,\partial A_i;\b Q)$ to $H_1(A_i;\b Q)$ or from $H^2(B_i,\partial B_i;\b Q)$ to $H_1(B_i;\b Q)$. Furthermore, the homology classes $L_{\lbrace X^I \rbrace}(\sfrac{B_1}{A_1})$ and $L_{\lbrace X^I \rbrace}(\sfrac{B_2}{A_2})$ are mapped to zero in $H_1(M;\b Q)$ and the map
$$
lk_M : \mbox{\textup{ker}}\big(H_1(A_1;\b Q) \rightarrow H_1(M; \b Q)\big) \times \mbox{\textup{ker}}\big(H_1(A_2;\b Q) \rightarrow H_1(M; \b Q)\big) \longrightarrow \b Q
$$
is well-defined.
\end{theorem}
\begin{exampleemempty}
Consider $\b S^3$ equipped with a parallelization $\tau : \b S^3 \times \b R^3 \rightarrow T\b S^3$ which extends the standard parallelization of the unit ball. In this ball, consider a positive Hopf link and let $A_1 \sqcup A_2$ be a tubular neighborhood of this link. Let $X$ be the combing $\tau(e_1)=\tau(.,e_1)$, where $e_1=(1,0,0)\in \b S^2$, and let $B_1 = A_1$ and $B_2=A_2$. Identify $A_1$ and $A_2$ with $\b D^2 \times \b S^1$ and consider a smooth map $g : \b D^2 \rightarrow \b S^2$ such that $g(\partial \b D^2) = e_1$, and such that $-e_1$ is a degree 1 regular value of $g$ with a single preimage $\omega$. Finally, for $i \in \lbrace 1,2 \rbrace$, let $X_{B_i}$ be the combing :
$$
X_{B_i} :
\left\lbrace
\begin{aligned}
\b D^2 \times \b S^1 &\longrightarrow UM \\
(z,u) &\longmapsto \tau((z,u),g(u)).
\end{aligned}
\right.
$$
In this case, $L_{X(\lbrace \sfrac{B_i}{A_i} \rbrace_{i \in \lbrace 1,2 \rbrace})=-X} = L_{X_{B_1}=-X_{|A_1}} \cup L_{X_{B_2}=-X_{|A_2}}$, and, for $i \in \lbrace 1,2 \rbrace$, using the identification of $A_i$ with $\b D^2 \times \b S^1$, the link $L_{X_{B_i}=-X_{|A_i}}$ reads $\lbrace \omega \rbrace \times \b S^1$. As we will see in Proposition~\ref{prop_linksinhomologyI}, for $i \in \lbrace 1,2 \rbrace$, $L_{\lbrace X^I \rbrace}(\sfrac{B_i}{A_i}) = 2 [L_{X_{B_i}=-X_{|A_i}}]$. Eventually,
$$
\sum_{I\subset\lbrace 1 , 2 \rbrace} (-1)^{|I|} p_1([X^I]) = -8.
$$
\end{exampleemempty}
In general, for an LP$_\b Q$-surgery datum $(\sfrac{B}{A})$ in a compact oriented 3-manifold $M$, a trivialization of $TM_{|(M\setminus \mathring A)}$ cannot be extended as a parallelization of $M(\sfrac{B}{A})$. It follows that LP$_\b Q$-surgeries cannot be expressed as local moves on parallelized compact oriented 3-manifolds. This makes computing the variation of Pontrjagin numbers of torsion combings under LP$_\b Q$-surgeries tricky since Pontrjagin numbers of torsion combings are defined with respect to Pontrjagin numbers of parallelizations. \\
However, if $M$ is a compact oriented 3-manifold and if $\rho$ is a trivialization of $TM_{|\partial M}$, then the obstruction to finding a parallelization of $M$ which coincides with $\rho$ on $\partial M$ is an element of $H^2(M,\partial M; \sfrac{\b Z}{2\b Z})$ -- hence, its Poincaré dual is an element $[\gamma]$ of $H_1(M;\sfrac{\b Z}{2\b Z})$ -- and it is possible to get around such an obstruction thanks to the notion of \textit{pseudo-parallelization} developed by C.~Lescop. Let us postpone the formal definition to Subsection~\ref{ssec_defppara} (see also \cite{lescopcube}) and, for the time being, let us just mention that a pseudo-parallelization $\bar\tau$ of a compact oriented 3-manifold $M$ is a triple $(N(\gamma); \tau_e, \tau_d)$ where $N(\gamma)$ is a framed tubular neighborhood of a link $\gamma$ in $\mathring M$, $\tau_e$ is a parallelization of $M\setminus N(\gamma)$ and $\tau_d : N(\gamma)\times \b R^3 \rightarrow TN(\gamma)$ is a parallelization of $N(\gamma)$ such that there exists a section $E_1^d$ of $UM$ :
$$
E_1^d :
\left\lbrace
\begin{aligned}
m \in M\setminus \mathring N(\gamma) &\longmapsto \tau_e(m,e_1) \\
m \in N(\gamma) &\longmapsto \tau_d(m,e_1).
\end{aligned}
\right.
$$
Let us finally mention that $\bar\tau$ also determines a section $E_1^g$ of $UM$ which coincides with $E_1^d$ on $M\setminus \mathring N(\gamma)$. The sections $E_1^d$ and $E_1^g$ are the \textit{Siamese sections} of $\bar\tau$ and the link $\gamma$ is the \textit{link of the pseudo-parallelization $\bar\tau$}. \\
To a pseudo-parallelization, C. Lescop showed that it is possible to associate a \textit{complex trivialization} up to homotopy, see Definition~\ref{def_complextriv}. This leads to a natural extension of the notion of first relative Pontrjagin numbers of parallelizations to pseudo-parallelizations. Furthermore, as in the case of parallelizations, a pseudo-parallelization $\bar\tau$ of a compact oriented 3-manifold $M$ admits \textit{pseudo-sections} $\bar\tau(M\times\lbrace v\rbrace)$ which are 3-chains of $UM$, for all $v \in \b S^2$. In the special case $v =e_1$ the pseudo-section $\bar\tau(M\times \lbrace e_1 \rbrace)$ of $\bar\tau$ can be written as :
$$
\bar\tau(M\times \lbrace e_1 \rbrace) = \frac{E_1^d(M) + E_1^g(M)}{2}.
$$
A combing $(X,\sigma)$ of $M$ is said to be \textit{compatible with $\bar\tau$} if $(X,\sigma)$ is $\partial$-compatible with $(E_1^d,{E_2^e}_{|\partial M})$ and $(E_1^g,{E_2^e}_{|\partial M})$, where $E_2^e$ is the second vector of $\tau_e$, and if
$$
L_{E_1^d=X} \cap L_{E_1^g=-X}=\emptyset \mbox{ \ and \ } L_{E_1^g=X} \cap L_{E_1^d=-X}=\emptyset.
$$
If $(X,\sigma)$ and $\bar\tau$ are compatible, then $\rho(X)=\bar\tau_{|\partial M}$ and we get two disjoint rational combinations of oriented links in $\mathring M$ :
$$
L_{\bar\tau = X} = \frac{L_{E_1^d= X} + L_{E_1^g= X}}{2} \mbox{ \ and \ } L_{\bar\tau = - X} = \frac{L_{E_1^d=- X} + L_{E_1^g=- X}}{2}.
$$
Pseudo-parallelizations allow us to revisit the definition of Pontrjagin numbers and to generalize it to torsion combings of compact oriented 3-manifolds with non empty boundary as follows. Let $P_{\b S^2}$ denote the standard projection from $W\times \b S^2$ to $\b S^2$, for any manifold $W$.
\begin{lemma} \label{lem1}
Let $(X,\sigma)$ be a torsion combing of a compact oriented 3-manifold $M$, let $\bar\tau$ be a pseudo-parallelization of $M$, and let $E_1^d$ and $E_1^g$ be the Siamese sections of $\bar\tau$. If $\bar\tau$ and $(X,\sigma)$ are compatible, then the expression
$$
4\cdot lk_M(L_{\bar\tau=X} , L_{\bar\tau=-X}) - lk_{\b S^2} \left( e_1-(-e_1) \ , \ P_{\b S^2} \circ \tau_d^{-1} \circ X(L_{{E_1}^d={-E_1}^g}) \right)
$$
depends only on the homotopy class of $(X,\sigma)$. It will be denoted $p_1(\bar\tau,[X])$ and its opposite will be written $p_1([X],\bar\tau)$.
\end{lemma}
\begin{theorem} \label{thm_defp1Xb}
Let $(X_1,\sigma_{X_1})$ and $(X_2,\sigma_{X_2})$ be torsion combings of two compact oriented 3-manifolds $M_1$ and $M_2$ with identified boundaries such that $(X_1,\sigma_{X_1})$ and $(X_2,\sigma_{X_2})$ coincide on the boundary. For $i\in\lbrace 1,2\rbrace$, let $\bar\tau_i$ be a pseudo-parallelization of $M_i$ such that $\bar\tau_i$ and $(X_i,\sigma_{X_i})$ are compatible. The expression
$$
p_1([X_1],[X_2])= p_1( [X_1],\bar\tau_1) + p_1(\bar\tau_1, \bar\tau_2) + p_1(\bar\tau_2, [X_2])
$$
depends only on the homotopy classes of $(X_1,\sigma_{X_1})$ and $(X_2,\sigma_{X_2})$, and it defines \textup{the first relative Pontrjagin number of $(X_1,\sigma_{X_1})$ and $(X_2,\sigma_{X_2})$}. Moreover, if $M_1$ and $M_2$ are closed, then
$$
p_1([X_1],[X_2])=p_1([X_2])-p_1([X_1]).
$$
\end{theorem}
Under the assumptions of Theorem \ref{thm_D2nd0}, we see that it would be impossible to naively define $p_1([X_{|A_1}],[{X_{\lbrace 1 \rbrace}}_{|B_1}])$ as $p_1([X])-p_1([{X_{\lbrace 1 \rbrace}}])$, where $X$ extends $X_{|A_1}$ to the closed manifold $M$, and ${X_{\lbrace 1 \rbrace}}$ extends ${X_{\lbrace 1 \rbrace}}_{|B_1}$ in the same way to $M(\sfrac{B_1}{A_1})$. Indeed Theorem \ref{thm_D2nd0} and the example that follows it show that the expression $\left(p_1([X])-p_1([{X_{\lbrace 1 \rbrace}}])\right)$ depends on the combed manifold $(M,X)$ into which $(A_1,X_{|A_1})$ has been embedded. It even depends on the combing $X$ that extends the combing $X_{|A_1}$ of $A_1$ to $M$ for the fixed manifold $M$ of this example, since
$$
\left( p_1([X]) - p_1([{X_{\lbrace 1 \rbrace}}]) \right) - \left( p_1([{X_{\lbrace 2 \rbrace}}]) - p_1([{X_{\lbrace 1,2 \rbrace}}]) \right) =-8
$$
there. \\
Theorem \ref{thm_defp1Xb} translates as follows in the closed case and it bridges a gap between the two dissimilar generalizations of the Pontrjagin numbers of parallelizations for pseudo-parallelizations and for torsion combings in closed oriented 3-manifolds.
\begin{corollary} \label{cor_p1Xppara}
Let $X$ be a torsion combing of a closed oriented 3-manifold $M$ and let \linebreak $\bar\tau (N(\gamma);\tau_e,\tau_d)$ be a pseudo-parallelization of $M$. Let $E_1^d$ and $E_1^g$ denote the Siamese sections of $\bar\tau$. If $X$ and $\bar \tau$ are compatible, then
$$
\begin{aligned}
p_1([X])
= p_1(\bar \tau) &+ 4\cdot lk_M(L_{\bar\tau=X} , L_{\bar\tau=-X}) \\
&- lk_{\b S^2} \left( e_1-(-e_1) \ , \ P_{\b S^2} \circ \tau_d^{-1} \circ X(L_{{E_1}^d={-E_1}^g}) \right).
\end{aligned}
$$
\end{corollary}
Another special case is when genuine parallelizations can be used. The closed case with genuine parallelizations is nothing but C. Lescop's definition of the Pontrjagin number of torsion combings in closed oriented 3-manifolds stated above.
\begin{corollary} \label{cor_compactpara}
Let $(X_1,\sigma_1)$ and $(X_2,\sigma_2)$ be torsion combings of two compact oriented 3-manifolds $M_1$ and $M_2$ with identified boundaries such that $(X_1,\sigma_1)$ and $(X_2,\sigma_2)$ coincide on the boundary. If, for $i \in \lbrace 1 , 2 \rbrace$, $\tau_i \hspace{-1mm}=\hspace{-1mm} (E_1^i,E_2^i,E_3^i)$ is a parallelization of $M_i$ such that $(X_i,\sigma_i)$ and $(E_1^i,{E_2^i}_{|\partial M_i})$ are $\partial$-compatible, then
$$
p_1([X_1],[X_2]) = p_1(\tau_1 , \tau_2) + 4 \cdot lk_{M_2}(L_{E_1^{2}=X_2} \ , \ L_{E_1^{2}=-X_2}) - 4 \cdot lk_{M_1}(L_{E_1^{1}=X_1} \ , \ L_{E_1^{1}=-X_1}).
$$
\end{corollary}
Finally, for torsion combings defined on a fixed compact oriented 3-manifold (which may have boundary), we have the following simple variation formula, as in the closed case.
\begin{theorem} \label{formuleplus}
If $(X, \sigma)$ and $(Y,\sigma)$ are $\partial$-compatible torsion combings of a compact oriented 3-manifold $M$, then
$$
p_1([X],[Y]) = 4 \cdot lk_M(L_{X=Y}, L_{X=-Y}).
$$
\end{theorem}
\def\mbox{spin$^c$}{\mbox{spin$^c$}}
Let $M$ be a compact connected oriented 3-manifold. For all section $\sigma$ of $TM_{|\partial M}$, let $\mbox{spin$^c$}(M,\sigma)$ denote the \textit{set of $\mbox{spin$^c$}$-structures on $M$ relative to $\sigma$}, \textit{ie}\ the set of homotopy classes on $M \setminus \lbrace \omega \rbrace$ of combings $(X,\sigma)$ of $M$, where $\omega$ is any point in $\mathring M$ (see \cite{gmdeloup}, for a detailed presentation of $\mbox{spin$^c$}$-structures). Thanks to Theorem~\ref{formuleplus}, it is possible to classify the torsion combings of a fixed $\mbox{spin$^c$}$-structure up to homotopy, thus generalizing a property of the Gompf invariant in the closed case. I thank Gw\'{e}na\"{e}l Massuyeau for suggesting this statement.
\begin{theorem} \label{GM}
Let $(X,\sigma)$ and $(Y,\sigma)$ be $\partial$-compatible torsion combings of a compact connected oriented 3-manifold $M$ which represent the same $\mbox{spin$^c$}$-structure. The combings $(X,\sigma)$ and $(Y,\sigma)$ are homotopic relatively to the boundary if and only if $p_1([X],[Y]) =0$.
\end{theorem}
The key tool in the proof of Theorem~\ref{thm_defp1Xb} is the following generalization of the interpretation of the variation of the Pontrjagin numbers of parallelizations as an algebraic intersection of three chains.
\begin{theorem} \label{prop_varasint}
Let $\tau$ and $\bar \tau$ be two pseudo-parallelizations of a compact oriented 3-manifold $M$ that coincide on $\partial M$ and whose links are disjoint. For any $v \in \b S^2$, there exists a 4-chain $C_4(\tau,\bar\tau ;v)$ of $[0,1]\times UM$ transverse to the boundary of $[0,1] \times UM$ such that
$$
\partial C_4(\tau,\bar\tau ;v) = \lbrace 1 \rbrace \times \bar\tau(M\times\lbrace v \rbrace) - \lbrace 0 \rbrace \times \tau(M\times \lbrace v \rbrace) - [0,1]\times \tau(\partial M \times \lbrace v \rbrace)
$$
and for any $x,y$ and $z$ in $\b S^2$ with pairwise different distances to $e_1$ :
$$
p_1(\tau,\bar\tau)= 4 \cdot \langle C_4(\tau,\bar\tau ;x),C_4(\tau,\bar\tau ;y),C_4(\tau,\bar\tau ;z) \rangle_{[0,1]\times UM}
$$
for any triple of pairwise transverse $C_4(\tau,\bar\tau,x)$, $C_4(\tau,\bar\tau,y)$ and $C_4(\tau,\bar\tau,z)$ that satisfy the hypotheses above.
\end{theorem}
Our general variation formula for Pontrjagin numbers of torsion combings reads as follows for all compact oriented 3-manifolds.
\begin{theorem}\label{thm_D2nd}
Let $(M,X)$ be a compact oriented 3-manifold equipped with a combing, \linebreak let $\lbrace (\sfrac{B_i}{A_i},X_{B_i}) \rbrace_{i \in \lbrace 1,2 \rbrace}$ be two disjoint LP$_\b Q$-surgeries in $(M,X)$, and, for all $I \subset \lbrace 1 , 2 \rbrace$, \linebreak let $X^I = X(\lbrace \sfrac{B_i}{A_i} \rbrace_{i \in I})$. If $\lbrace X^I \rbrace_{I \subset \lbrace 1,2 \rbrace}$ is a family of torsion combings of the manifolds \linebreak $M_I=M(\lbrace \sfrac{B_i}{A_i} \rbrace_{i \in I})$, then
$$
p_1([X^{\lbrace 2 \rbrace}],[X^{\lbrace 1, 2 \rbrace}])-p_1([X],[X^{\lbrace 1 \rbrace}]) = - 2 \cdot lk_M \left(L_{\lbrace X^I \rbrace}(\sfrac{B_1}{A_1}), L_{\lbrace X^I \rbrace}(\sfrac{B_2}{A_2})\right),
$$
where the right-hand side is defined as in Theorem~\ref{thm_D2nd0}.
\end{theorem}
A direct consequence of this variation formula is that the extended Gompf invariant for torsion combings of compact oriented 3-manifolds is a degree two finite type invariant with respect to LP$_\b Q$-surgeries.
\begin{corollary} \label{cor_FTcombings}
Let $(M,X)$ be a compact oriented 3-manifold equipped with a combing, let $\lbrace ( \sfrac{B_i}{A_i} , X_{B_i} ) \rbrace_{i\in \lbrace 1, \ldots, k\rbrace}$ be a family of disjoint LP$_\b Q$-surgeries in $(M,X)$, and, for all $I \subset \lbrace 1 , \ldots, k \rbrace$, let $(M_I,X^I)$ be the combed manifold obtained by performing the surgeries associated to the data $\lbrace (\sfrac{B_i}{A_i}, X_{B_i}) \rbrace_{i \in I}$. If $k\geqslant 3$, and if $\lbrace X^I \rbrace_{I \subset \lbrace 1, \ldots, k \rbrace}$ is a family of torsion combings of the $\lbrace M_I \rbrace_{I \subset \lbrace 1, \ldots, k \rbrace}$, then
$$
\sum_{I \subset \lbrace 2 , \ldots , k \rbrace} (-1)^{\card(I)} \ p_1 \left( [X^I] , [X^{I\cup\lbrace 1 \rbrace}] \right)=0.
$$
If $\partial M=\emptyset$, this reads
$$
\sum_{I \subset \lbrace 1 , \ldots , k \rbrace} (-1)^{\card(I)} \ p_1 \left( [X^I] \right)=0.
$$
\end{corollary}
In the first section of this article, we give details on Lagrangian-preserving surgeries, combings and pseudo-parallelizations. Then, we review the definitions of Pontrjagin numbers of parallelizations and pseudo-parallelizations. The second section ends with a proof of Theorem~\ref{prop_varasint}. The third section is devoted to the proof of Theorem~\ref{thm_defp1Xb} and Theorem~\ref{GM}. Finally, we study the variations of Pontrjagin numbers with respect to Lagrangian-preserving surgeries, and finish the last section by proving Theorem~\ref{thm_D2nd}.\\
\begin{large}\textbf{Acknowledgments.}\end{large} First, let me thank C. Lescop and J.-B. Meilhan for their thorough guidance and support. I also thank M. Eisermann and G. Massuyeau for their careful reading and their useful remarks.
\renewcommand{\thetheorem}{\arabic{section}.\arabic{theorem}}
\section{More about ...}
\subsection{Lagrangian-preserving surgeries} \label{ssec_LPsurgeries}
Let us first note three easy lemmas, the proofs of which are left to the reader.
\begin{lemma} \label{prop_redefLPs}
Let $(\sfrac{B}{A})$ be an LP$_\b Q$-surgery datum in a compact oriented 3-manifold $M$ and let $L_1$ and $L_2$ be links in $M \minusens \mathring A$. If $L_1$ and $L_2$ are rationally null-homologous in $M$, then they are null-homologous in $M(\sfrac{B}{A})$ and
$$
lk_{M(\sfrac{B}{A})}(L_1,L_2) = lk_M(L_1,L_2).
$$
\end{lemma}
\iffalse
\begin{proof}
For $i \in \lbrace 1,2 \rbrace$, let $\Sigma_i$ be a 2-chain in $M$ such that $\partial \Sigma_i=L_i$ and $\Sigma_i$ is transverse to $\partial A$. Since $(\sfrac{B}{A})$ is an LP$_\b Q$-surgery, $[\Sigma_i\cap \partial A]= 0$ in $H_1(B; \b Q)$. Therefore, for all $i \in \lbrace 1,2 \rbrace$, there exists a 2-chain $\Sigma'_i$ in $M(\sfrac{B}{A})$ such that
$$
\partial \Sigma'_i = L_i \mbox{ \ and \ } \Sigma'_i \cap (M(\sfrac{B}{A})\minusens \mathring B) = \Sigma_i \cap (M \minusens \mathring A).
$$
As a consequence $L_1$ and $L_2$ are null-homologous in $M(\sfrac{B}{A})$ and, since $L_2 \subset M \minusens \mathring A$, it follows that
$$
lk_{M(\sfrac{B}{A})}(L_1,L_2) = \langle \Sigma'_1 , L_2 \rangle_{M(\sfrac{B}{A})} = \langle \Sigma_1 , L_2 \rangle_M = lk_M(L_1,L_2).
$$
\end{proof}
\fi
A \textit{rational homology 3-sphere}, or a $\b Q$HS for short, is a closed oriented 3-manifold with the same homology with rational coefficients as $\b S^3$.
\begin{lemma} \label{prop-LP2}
Let $(\sfrac{B}{A})$ be an LP$_\b Q$-surgery in a compact oriented 3-manifold $M$. If $M$ is a $\b Q$HS, then $M(\sfrac{B}{A})$ is a $\b Q$HS.
\end{lemma}
\iffalse
\begin{proof}
Let $M$ be a $\b Q$HS and let $A$ be a $\b Q$HH of genus $g\in \b N$. Using the Mayer-Vietoris sequence associated to $M=A\cup (M\minusens\mathring A)$ shows that $M\minusens \mathring A$ is a $\b Q$HH of genus $g$ and that the inclusions of $\partial A$ into $A$ and into $M\setminus \mathring A$ induce an isomorphism
$$
H_1(\partial A ; \b Q) \simeq H_1(A; \b Q)\oplus H_1(M\minusens\mathring A ; \b Q).
$$
For details, see \cite[Sublemma 4.6]{moussardFTIQHS}. It follows that $M(\sfrac{B}{A})\minusens \mathring B$ is also a genus $g$ $\b Q$HH and, since $(\sfrac{B}{A})$ is an LP$_\b Q$-surgery, the inclusions of $\partial B$ into $B$ and into $M(\sfrac{B}{A})\setminus \mathring B$ induce an isomorphism
$$
H_1(\partial B ; \b Q) \simeq H_1(B; \b Q)\oplus H_1(M(\sfrac{B}{A})\minusens\mathring B ; \b Q).
$$
Using this isomorphism in the Mayer-Vietoris sequence associated to the splitting \linebreak $M(\sfrac{B}{A})=B\cup (M(\sfrac{B}{A})\minusens\mathring B)$ shows that $M(\sfrac{B}{A})$ is a $\b Q$HS.
\end{proof}
\fi
\begin{lemma} \label{phom}
If $A$ is a compact connected orientable 3-manifold with connected boundary and if the map $i^A_* : H_1(\partial A ; \b Q) \rightarrow H_1(A; \b Q)$ induced by the inclusion of $\partial A$ into $A$ is surjective, then $A$ is a rational homology handlebody.
\end{lemma}
\iffalse
\begin{proof}
First, for such a manifold $A$, we have $H_0(A;\b Q)\simeq \b Q$ and $H_3(A;\b Q)\simeq 0$. Second, using the hypothesis on $i^A_*$ in the exact sequence associated to $(A, \partial A)$, we get that $H_1(A,\partial A;\b Q)= 0$. Using Poincaré duality and the universal coefficient theorem, it follows that
$$
H_2(A;\b Q) \simeq H^1(A,\partial A;\b Q) \simeq \mbox{Hom}(H_1(A,\partial A;\b Q),\b Q) = 0.
$$
Moreover, we get the following exact sequence from the exact sequence associated to $(A,\partial A)$~:
$$
0\rightarrow H_2(A,\partial A; \b Q) \rightarrow H_1(\partial A;\b Q) \rightarrow H_1(A;\b Q) \rightarrow 0.
$$
It follows that $\mbox{dim}(H_2(A,\partial A;\b Q))+\mbox{dim}(H_1(A;\b Q))=\mbox{dim}(H_1(\partial A;\b Q)) = 2g$, where $g$ denotes the genus of $\partial A$. However,
$$
H_2(A,\partial A;\b Q) \simeq H^1(A;\b Q) \simeq \mbox{Hom}(H_1(A;\b Q),\b Q),
$$
hence $\mbox{dim}(H_1(A;\b Q))=g$.
\end{proof}
\fi
\begin{proposition}
Let $A$ be a compact submanifold with connected boundary of a $\b Q$HS $M$, let $B$ be a compact oriented 3-manifold and let $h:\partial A \rightarrow \partial B$ be a homeomorphism. If the surgered manifold $M(\sfrac{B}{A})$ is a $\b Q$HS and if
$$ lk_{M(\sfrac{B}{A})}(L_1,L_2) = lk_M(L_1,L_2)
$$
for all disjoint links $L_1$ and $L_2$ in $M \minusens \mathring A$, then $(\sfrac{B}{A})$ is an LP$_\b Q$-surgery.
\end{proposition}
\begin{proof}
Using the Mayer-Vietoris exact sequences associated to $M=A\cup (M\minusens\mathring A)$ and \linebreak $M(\sfrac{B}{A})=B\cup (M(\sfrac{B}{A})\minusens\mathring B)$, we get that the maps $i_*^A : H_1(\partial A ; \b Q) \longrightarrow H_1(A; \b Q)$ and \linebreak $i_*^B : H_1(\partial B ; \b Q) \longrightarrow H_1(B; \b Q)$ induced by the inclusions of $\partial A$ and $\partial B$ into $A$ and $B$ are surjective. Using Lemma~\ref{phom}, it follows that $A$ and $B$ are rational homology handlebodies. Moreover, $A$ and $B$ have the same genus since $h : \partial A \rightarrow \partial B$ is a homeomorphism. \\
Let $P_{\go L_A}$ and $P_{\go L_B}$ denote the projections from $H_1(\partial A;\b Q)$ onto $\go L_A$ and $\go L_B$, respectively, with kernel $\go L_{M\setminus \mathring A}$. Consider a collar $[0,1]\times \partial A$ of $\partial A$ such that $\lbrace 0 \rbrace \times \partial A \simeq \partial A$ and note that for all 1-cycles $x$ and $y$ of $\partial A$ :
$$
\langle P_{\go L_A}(y), x \rangle_{\partial A}
= lk_M(\lbrace 1 \rbrace \times y, \lbrace 0 \rbrace \times x )
= lk_{M(\sfrac{B}{A})}(\lbrace 1 \rbrace \times y, \lbrace 0 \rbrace \times x)
=\langle P_{\go L_B}(y), x \rangle_{\partial B},
$$
so that $P_{\go L_B}=P_{\go L_A}$ and $h_*(\go L_A)=\go L_B$.
\end{proof}
\subsection{Combings}
\begin{proposition} \label{prop_linksandsigns}
If $X$ and $Y$ are $\partial$-compatible combings of a compact oriented 3-manifold $M$, then
$$
L_{X=Y}=L_{Y=X} \mbox{ \ and \ } L_{X=Y}= - L_{-X=-Y}.
$$
\end{proposition}
\begin{proof}
First, by definition, the link $L_{X=Y}$ is the projection of the intersection of the sections $X(\mathring M)$ and $Y(\mathring M)$. This intersection is oriented so that
$$
NX(\mathring M) \oplus NY(\mathring M) \oplus T(X(\mathring M)\cap Y(\mathring M))
$$
orients $UM$, fiberwise. Since the normal bundles $NX(\mathring M)$ and $NY(\mathring M)$ have dimension 2, the isomorphism permuting them is orientation-preserving so that $L_{X=Y}=L_{Y=X}$. Second, $(-X)(\mathring M)\cap(-Y)(\mathring M)$ is the image of $X(\mathring M)\cap Y(\mathring M)$ under the map $\iota$ from $UM$ to itself which acts on each fiber as the antipodal map. This map reverses the orientation of $UM$ as well as the coorientations of $X(\mathring M)$ and $Y(\mathring M)$, \textit{ie}\
$$
\begin{aligned}
&N(-X)(M)=-\iota(NX(M)), \ N(-Y)(M)=-\iota(NY(M)). \\
\end{aligned}
$$
Since $N(-X)(M) \oplus N(-Y)(M) \oplus T((-X)(M)\cap (-Y)(M))$ has the orientation of $UM$
$$
T((-X)(M)\cap (-Y)(M)) = -\iota(T(X(M)\cap Y(M))).
$$
Hence, $L_{X=Y}= - L_{-X=-Y}$.
\end{proof}
\begin{definition}
Let $M$ be a compact oriented 3-manifold and let $L$ be a link in $\mathring M$. Define the \textit{blow up of $M$ along $L$} as the 3-manifold $Bl(M,L)$ constructed from $M$ in which $L$ is replaced by its unit normal bundle in $M$. The 3-manifold $Bl(M,L)$ inherits a canonical differential structure. See \cite[Definition 3.5]{lescopcombing} for a detailed description.
\end{definition}
\begin{lemma} \label{lem_phomotopy}
Let $X$ and $Y$ be $\partial$-compatible combings of a compact oriented 3-manifold $M$. There exists a 4-chain $\bar F(X,Y)$ of $UM$ with boundary~:
$$
\partial \bar F(X,Y) = Y(M) - X(M) + UM_{| L_{X=-Y}}.
$$
\end{lemma}
\begin{proof}
To construct the desired 4-chain, start with the partial homotopy from $X$ to $Y$
$$
\tilde F(X,Y) :
\left\lbrace
\begin{aligned}
\ [0,1] \times (M\setminus L_{X=-Y}) &\longrightarrow UM\\
(s,m) & \longmapsto \left( m , \l H_X^Y(s,m) \right)
\end{aligned}
\right.
$$
where $\l H_X^Y(s,m)$ is the unique point of the shortest geodesic arc from $X(m)$ to $Y(m)$ such that
$$
d_{\b S^2}(X(m),\l H_X^Y(s,m)) = s \cdot d_{\b S^2}(X(m),Y(m))
$$
where $d_{\b S^2}$ denotes the usual distance on $\b S^2$. Next, extend the map
$$
(s,m) \longmapsto \l H_X^Y(s,m)
$$
on the blow up of $M$ along $L_{X=-Y}$. The section $X$ induces a map
$$
X : NL_{X=-Y} \longrightarrow -Y^\perp(L_{X=-Y})
$$
which is a diffeomorphism on a neighborhood of $\lbrace 0 \rbrace \times L_{X=-Y}$ since $X$ and $Y$ are $\partial$-compatible combings. Furthermore, this diffeomorphism is orientation-preserving by definition of the orientation on $L_{X=-Y}$. So, for $n \in UN_mL_{X=-Y}$, $\l H_X^Y(s,n)$ can be defined as the unique point at distance $s\pi$ from $X(m)$ on the unique half great circle from $X(m)$ to $Y(m)$ through $T_m X(n)$. Thanks to transversality again, the set $\lbrace \l H_X^Y(s,n) \ | \ s \in [0,1], \ n \in UN_mL_{X=-Y} \rbrace$ is a whole sphere $\b S^2$ for any fixed $m \in L_{X=-Y}$, so that
$$
\partial \tilde F(X,Y)([0,1]\times Bl(M,L_{X=-Y})) = Y(M) - X(M) + \partial_{int}
$$
where $\partial_{int} \simeq L_{X=-Y} \times \b S^2 $ (see \cite[Proof of Proposition 3.6]{lescopcombing} for the orientation of $\partial_{int}$). Finally, let $\bar F(X,Y)=\tilde F(X,Y)([0,1]\times Bl(M,L_{X=-Y}))$.
\end{proof}
If $X$ and $Y$ are $\partial$-compatible combings of a compact oriented 3-manifold $M$ and if $\sigma$ is a nonvanishing section of $X^\perp_{|\partial M}$, let $\l H^{-Y}_{X,\sigma}$ denote the map from $[0,1] \times (M\minusens L_{X=Y})$ to $UM$ such that, for all $(s,m)$ in $[0,1] \times \partial M$, $\l H^{-Y}_{X,\sigma}(s,m)$ is the unique point at distance $s\pi$ from $X(m)$ on the unique geodesic arc starting from $X(m)$ in the direction of $\sigma(m)$ to $-X(m)=-Y(m)$ and, for all $(s,m)$ in $[0,1] \times (\mathring M \setminus L_{X=Y})$, $\l H^{-Y}_{X,\sigma}(s,m)$ is the unique point on the shortest geodesic arc from $X(m)$ to $-Y(m)$ such that
$$
d_{\b S^2}(X(m),\l H^{-Y}_{X,\sigma}(s,m)) = s \cdot d_{\b S^2}(X(m),-Y(m)).
$$
As in the previous proof, $\l H^{-Y}_{X,\sigma}$ may be extended as a map from $[0,1]\times Bl(M,L_{X=Y})$ to $UM$. In the case of $X=Y$, for all section $\sigma$ of $X^\perp$, nonvanishing on $\partial M$, let $L_{\sigma=0}$ denote the oriented link $\lbrace m \in M \ | \ \sigma(m)= 0 \rbrace$ and define a map $\l H_{X,\sigma}^{-X}$ as the map from $[0,1] \times (M\minusens L_{\sigma=0})$ to $UM$ such that, for all $(s,m)$ in $[0,1] \times (M\minusens L _{\sigma=0})$, $\l H_{X,\sigma}^{-X}(s,m)$ is the unique point at distance $s\pi$ from $X(m)$ on the unique geodesic arc starting from $X(m)$ in the direction of $\sigma(m)$ to $-X(m)$. Note that $L_{\sigma=0}\cap\partial M = \emptyset$, and $[L_{\sigma=0}]=P(e^M_2(X,\sigma_{|\partial M}))$. Here again, $\l H_{X,\sigma}^{-X}$ may be extended as a map from $[0,1]\times Bl(M,L_{\sigma=0})$ to $UM$. \\
In order to simplify notations, if $A$ is a submanifold of a compact oriented 3-manifold $M$, we may implicitly use a parallelization of $M$ to write $UM_{|A}$ as $A\times \b S^2$.
\begin{proposition} \label{prop_links}
If $(X,\sigma)$ and $(Y,\sigma)$ are $\partial$-compatible combings of a compact oriented 3-manifold $M$, then, in $H_3(UM;\b Z)$,
$$
\begin{aligned}
\ [L_{X=-Y} \times \b S^2 ] &= [X(M) - Y(M)] \\
\ [L_{X=Y} \times \b S^2 ] &= [X(M) - (-Y)(M) + \l H^{-Y}_{X,\sigma} ([0,1]\times \partial M)].
\end{aligned}
$$
\end{proposition}
\begin{proof}
The first identity is a direct consequence of Lemma~\ref{lem_phomotopy}. The second one can be obtained using a similar construction. Namely, construct a 4-chain $\bar F(X,-Y)$ using the partial homotopy from $X$ to $-Y$ :
$$
\tilde F(X,-Y) : \left\lbrace
\begin{aligned}
\ [0,1] \times (M\setminus L_{X=Y}) &\longrightarrow UM\\
(s,m) & \longmapsto \left( m , \l H^{-Y}_{X,\sigma}(s,m) \right).
\end{aligned}
\right.
$$
As in the proof of Lemma~\ref{lem_phomotopy}, $\tilde F(X,-Y)$ can be extended to $[0,1] \times Bl(M,L_{X=Y})$. Finally, we get a 4-chain $\bar F(X,-Y)$ of $UM$ with boundary :
$$
\partial \bar F(X,-Y) = (-Y)(M) - X(M) - \l H^{-Y}_{X,\sigma} ([0,1]\times \partial M) + UM_{|L_{X=Y}}.
$$
\end{proof}
\begin{proposition} \label{prop_euler}
Let $X$ be a combing of a compact oriented 3-manifold $M$ and let \linebreak $P : H^2(M, \partial M;\b Z) \rightarrow H_1(M;\b Z)$ be the Poincaré duality isomorphism. If $M$ is closed, then, in $H_3(UM;\b Z)$,
$$
[P(e_2(X^\perp)) \times \b S^2 ] = [X(M) - (-X)(M)],
$$
where $[P(.)\times S^2]$ abusively denotes the homology class of the preimage of a representative of $P(.)$ under the bundle projection $UM \rightarrow M$. In general, if $\sigma$ is a section of $X^\perp$ such that $L_{\sigma=0}\cap\partial M = \emptyset$ then, in $H_3(UM;\b Z)$,
$$
[ P(e_2^M(X^\perp,\sigma_{|\partial M})) \times \b S^2 ] = [X(M) - (-X)(M) +\l H_{X,\sigma}^{-X}([0,1]\times \partial M)].
$$
\end{proposition}
\begin{proof}
Recall that $P(e_2^M(X^\perp,\sigma_{|\partial M}))=[L_{\sigma=0}]$. Perturbing $X$ by using $\sigma$, construct a section $Y$ homotopic to $X$ that coincides with $X$ on $\partial M$ and such that $[L_{X=Y}] = P(e_2^M(X^\perp,\sigma_{|\partial M}))$. Using Proposition~\ref{prop_links},
$$
[L_{X=Y} \times \b S^2 ] = [X(M) - (-Y)(M) + \l H^{-Y}_{X,\sigma_{|\partial M}} ([0,1]\times \partial M)],
$$
so that
$$
[ P(e_2^M(X^\perp,\sigma_{|\partial M})) \times \b S^2 ] = [X(M) - (-X)(M) +\l H_{X,\sigma}^{-X}([0,1]\times \partial M)].
$$
\end{proof}
\begin{proposition} \label{prop_linksinhomologyI}
If $(X,\sigma)$ and $(Y,\sigma)$ are $\partial$-compatible combings of a compact oriented 3-manifold $M$, then, in $H_1(M;\b Z)$,
$$
\begin{aligned}
2 \cdot [L_{X=-Y}] &= P(e_2^M(X^\perp,\sigma)) - P(e_2^M(Y^\perp,\sigma)), \\
2 \cdot [L_{X=Y}] &= P(e_2^M(X^\perp,\sigma)) + P(e_2^M(Y^\perp,\sigma)).
\end{aligned}
$$
\end{proposition}
\begin{proof}
Extend $\sigma$ as a section $\bar\sigma$ of $X^\perp$. Using Propositions~\ref{prop_linksandsigns},~\ref{prop_links}~and~\ref{prop_euler}, we get, in $H_3(UM;\b Z)$,
$$
\begin{aligned}
2 \ \cdot \ [L_{X=-Y}\times \b S^2] &= [L_{X=-Y} \times \b S^2] - [L_{-X=Y} \times \b S^2] \\
&= [X(M)-Y(M)]-[(-X)(M)-(-Y)(M)] \\
&= [X(M)-Y(M)-(-X)(M)+(-Y)(M) \\
& \hspace{5mm}+\l H_{X,\bar\sigma}^{-X}([0,1]\times \partial M) -\l H_{X,\bar\sigma}^{-X} ([0,1]\times \partial M) ] \\
&= [X(M)-(-X)(M)+\l H_{X,\bar\sigma}^{-X}([0,1]\times \partial M)] \\
&- [Y(M)-(-Y)(M)+\l H_{X,\bar\sigma}^{-X}([0,1]\times \partial M) ] \\
&= [P(e_2^M(X^\perp,\sigma)) \times \b S^2] - [P(e_2^M(Y^\perp,\sigma)) \times \b S^2 ],
\end{aligned}
$$
$$
\begin{aligned}
2 \cdot [L_{X=Y}\times \b S^2] &= [L_{X=Y} \times \b S^2] - [L_{-X=-Y} \times \b S^2] \\
&= [X(M) - (-Y)(M)+\l H_{X,\bar\sigma_{|\partial M}}^{-Y} ([0,1]\times \partial M)]\\
&- [(-X)(M) - Y(M) +\l H_{-X,\bar\sigma_{|\partial M}}^{Y} ([0,1]\times \partial M)] \\
&= [ P(e_2^M(X^\perp,\sigma)) \times \b S^2 ] - [ P(e_2^M((-Y)^\perp,\sigma)) \times \b S^2 ] \\
&= [ P(e_2^M(X^\perp,\sigma)) \times \b S^2 ] + [ P(e_2^M(Y^\perp,\sigma)) \times \b S^2 ].
\end{aligned}
$$
\iffalse &= [X(M) - (-Y)(M)+\l H_{X,\sigma}^{-X} ([0,1]\times \partial M)]\\
&- [(-X)(M) - Y(M) + \l H_{-Y,\sigma}^{Y} ([0,1]\times \partial M)] \\
&= [X(M)-(-X)(M) + \l H_{X,\sigma}^{-X} ([0,1]\times \partial M)] \\
&- [(-Y)(M)-Y(M)+ \l H_{-Y,\sigma}^{Y} ([0,1]\times \partial M)] \\ \fi
\end{proof}
\begin{remark}
If $M$ is a compact oriented 3-manifold and if $\sigma$ is a trivialization of $TM_{|\partial M}$, then the set $\mbox{spin$^c$}(M,\sigma)$ is a $H^2(M,\partial M; \b Z)$-affine space and the map
$$
c : \left\lbrace
\begin{aligned}
\mbox{spin$^c$}(M,\sigma) &\longrightarrow H^2(M,\partial M; \b Z) \\
[X]^c &\longmapsto e_2^M(X^\perp, \sigma)
\end{aligned}
\right.
$$
is affine over the multiplication by 2. Moreover, $[X]^c-[Y]^c \in H^2(M,\partial M; \b Z) \simeq H_1(M; \b Z)$ is represented by $L_{X=-Y}$, hence $2 \cdot [L_{X=-Y}] = P(e_2^M(X^\perp,\sigma)) - P(e_2^M(Y^\perp,\sigma))$. See \cite[Section 1.3.4]{gmdeloup} for a detailed presentation using this point of view. Both Proposition \ref{prop_linksinhomologyI} and Corollary \ref{corrplus} below are already-known results. For instance, Corollary \ref{corrplus} is also present in \cite{lescopcombing} (Lemma 2.16).
\end{remark}
\begin{corollary} \label{corrplus}
If $X$ and $Y$ are transverse combings of a closed oriented 3-manifold $M$, then, in $H_1(M;\b Z)$,
$$
\begin{aligned}
2 \cdot [L_{X=-Y}] &= P(e_2(X^\perp)) - P(e_2(Y^\perp)), \\
2 \cdot [L_{X=Y}] &= P(e_2(X^\perp)) + P(e_2(Y^\perp)).
\end{aligned}
$$
\end{corollary}
\subsection{Pseudo-parallelizations} \label{ssec_defppara}
\hspace{-3mm} A \textit{pseudo-parallelization} $\bar\tau \hspace{-1mm} = \hspace{-1mm} (N(\gamma); \tau_e, \tau_d)$ of a compact oriented 3-manifold $M$ is a triple~where
\begin{enumerate}[\textbullet]
\setlength{\itemsep}{0pt}
\setlength{\parskip}{5pt}
\item $\gamma$ is a link in $\mathring{M}$,
\item $N(\gamma)$ is a tubular neighborhood of $\gamma$ with a given product struture :
$$
N(\gamma)\simeq [a,b] \times \gamma \times [-1,1],
$$
\item $\tau_e$ is a genuine parallelization of $\smash{M\setminus\mathring{N(\gamma)}}$,
\item $\tau_d$ is a genuine parallelization of $N(\gamma)$ such that
$$
\tau_d = \left\lbrace
\begin{aligned}
& \tau_e &\mbox{ on } \partial (\left[ a , b \right] \times \gamma \times \left[ -1 , 1 \right]) \setminus \lbrace b \rbrace \times \gamma \times \left[ -1 , 1 \right] \\
& \tau_e \circ \l T_\gamma &\mbox{ on } \lbrace b \rbrace \times \gamma \times \left[ -1 , 1 \right]
\end{aligned}
\right.
$$
where $\l T_\gamma$ is
$$
\l T_\gamma :
\left\lbrace
\begin{aligned}
( [a,b] \times \gamma \times \left[ -1 , 1 \right] )\times \b R^3 & \longrightarrow ([a,b] \times \gamma \times \left[ -1 , 1 \right]) \times \b R^3 \\
((t,c,u),v) &\longmapsto ((t,c,u),R_{e_1, \pi+\theta(u)}(v)).
\end{aligned}
\right.
$$
where $R_{e_1, \pi+\theta(u)}$ is the rotation of axis $e_1$ and angle $\pi+\theta(u)$, and where $\theta : [-1,1] \rightarrow [-\pi,\pi]$ is a smooth increasing map constant equal to $\pi$ on the interval $[-1,-1+\varepsilon]$ ($\varepsilon \in ]0,\sfrac{1}{2}[$), and such that $\theta(-x)=-\theta(x)$.
\end{enumerate}
Note that a pseudo-parallelization whose link is empty is a parallelization.
\begin{lemma} \label{lem_extendpparallelization}
If $M$ is a compact oriented 3-manifold with boundary and if $\rho$ is a trivialization of $TM_{|\partial M}$, there exists a pseudo-parallelization $\bar\tau$ of $M$ that coincides with $\rho$ on $\partial M$.
\end{lemma}
\begin{proof}
The obstruction to extending the trivialization $\rho$ as a parallelization of $M$ can be represented by an element $[\gamma] \in H_1(M;\pi_1(SO(3)))$ where $\gamma$ is a link in $M$. It follows that $\rho$ can be extended on $M\setminus \mathring {N(\gamma)}$ where $N(\gamma)$ is a tubular neighborhood of $\gamma$. Finally, according to \cite[Lemma 10.2]{lescopcube}, it is possible to extend $\rho$ as a pseudo-parallelization on each torus of~$N(\gamma)$.
\end{proof}
Thanks to Lemma \ref{lem_extendpparallelization}, an LP$_\b Q$-surgery in a rational homology 3-sphere equipped with a pseudo-parallelization can be seen as a local move. This is not the case for an LP$_\b Q$-surgery in a rational homology 3-sphere equipped with a genuine parallelization. \\
\iffalse
Unlike the case of a parallelized rational homology 3-sphere, in the case of a rational homology 3-sphere equipped with a pseudo-parallelization, an LP$_\b Q$-surgery can be seen as a local move thanks to Lemma \ref{lem_extendpparallelization} above.
\fi
Before we move on to the definition of pseudo-sections \textit{ie}\ the counterpart of sections of parallelizations for pseudo-parallelizations, we need the following.
\begin{definition} \label{def_addinner}
Let $\bar \tau = (N(\gamma); \tau_e, \tau_d)$ be a pseudo-parallelization of a compact oriented 3-manifold. An \textit{additional inner parallelization} is a map $\tau_g$ such that
$$
\tau_g :
\left\lbrace
\begin{aligned}
\ [a,b]\times \gamma \times [-1,1] \times \b R^3 &\longrightarrow TN(\gamma) \\
((t,c,u),v) & \longmapsto \tau_d \left( \l T_\gamma^{-1} ((t,c,u),\l F(t,u)(v))\right)
\end{aligned}
\right.
$$
where, choosing $\varepsilon \in \ ]0,\sfrac{1}{2}[$, $\l F$ is a map such that
$$
\l F :
\left\lbrace
\begin{aligned}
\ [a,b]\times[-1,1] & \longrightarrow SO(3) \\
(t,u) & \longmapsto \left\lbrace
\begin{aligned}
& \mbox{Id}_{SO(3)} & \mbox{for $|u|>1-\varepsilon$} \\
& R_{e_1,\pi + \theta(u)} & \mbox{for $t < a + \varepsilon$} \\
& R_{e_1,-\pi - \theta(u)} & \mbox{for $t > b - \varepsilon$}
\end{aligned} \right.
\end{aligned}
\right.
$$
which exists since $\pi_1(SO(3))\hspace{-1mm}=\sfrac{\b Z}{2 \b Z}$ and which is well-defined up to homotopy since $\pi_2(SO(3))\hspace{-1mm}=~\hspace{-2mm}0$.
\end{definition}
From now on, we will always consider pseudo-parallelizations together with an additional inner parallelization. Finally, note that if $\bar \tau=(N(\gamma); \tau_e, \tau_d, \tau_g)$ is a pseudo-parallelization of a compact oriented 3-manifold together with an additional inner parallelization, then :
\begin{enumerate}[\textbullet]
\setlength{\itemsep}{0pt}
\setlength{\parskip}{5pt}
\item the parallelizations $\tau_e$, $\tau_d$ and $\tau_g$ agree on $\partial N(\gamma) \setminus \lbrace b \rbrace \times \gamma \times [-1,1]$,
\item $\tau_g = \tau_e \circ \l T_\gamma^{-1}$ on $\lbrace b \rbrace \times \gamma \times [-1,1]$.
\end{enumerate}
\begin{definition}
A \textit{pseudo-section} of a pseudo-parallelization of a compact oriented 3-manifold $M$ together with an additional inner parallelization, $\bar \tau=(N(\gamma); \tau_e, \tau_d, \tau_g)$, is a 3-cycle of \linebreak $(UM,UM_{|\partial M})$ of the following form :
$$
\begin{aligned}
\bar \tau (M\times \lbrace v \rbrace ) &= \tau_e((M\setminus \mathring N(\gamma))\times \lbrace v \rbrace ) \\
&+ \frac{ \tau_d(N(\gamma)\times \lbrace v \rbrace) + \tau_g(N(\gamma)\times \lbrace v \rbrace) + \tau_e( \lbrace b \rbrace \times \gamma \times C_2(v)) }{2}
\end{aligned}
$$
where $v\in \b S^2$ and $C_2(v)$ is the 2-chain of $\left[ -1 , 1 \right] \times \b S^1(v)$ of Figure~\ref{fig_C2v}, where $\b S^1(v)$ stands for the circle of $\b S^2$ that lies on the plane orthogonal to $e_1$ and passes through $v$. Note that :
$$
\begin{aligned}
\partial C_2(v) &= \lbrace (u, R_{e_1,\pi+\theta(u)} (v)) \ | u \in \left[ -1 , 1 \right] \rbrace \\
&+ \lbrace (u, R_{e_1,-\pi-\theta(u)} (v)) \ | u \in \left[ -1 , 1 \right] \rbrace - 2 \cdot \left[ -1, 1 \right] \times \lbrace v \rbrace.
\end{aligned}
$$
\end{definition}
\begin{center}
\definecolor{zzttqq}{rgb}{0.6,0.2,0}
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm]
\clip(-2.25,-1) rectangle (13,2.25);
\fill[color=zzttqq,fill=zzttqq,fill opacity=0.1] (0,0) -- (4,2) -- (0,2) -- cycle;
\fill[color=zzttqq,fill=zzttqq,fill opacity=0.1] (7.5,2) -- (7.5,0) -- (11.5,0) -- cycle;
\draw (0,2)-- (4,2);
\draw (4,0)-- (4,2);
\draw (4,0)-- (0,0);
\draw (0,0)-- (0,2);
\draw (7.5,0)-- (7.5,2);
\draw (7.5,0)-- (11.5,0);
\draw (11.5,0)-- (11.5,2);
\draw (7.5,2)-- (11.5,2);
\draw (0,0)-- (4,2);
\draw (0,2)-- (4,0);
\draw (7.5,2)-- (11.5,0);
\draw (11.5,2)-- (7.5,0);
\draw (-2,1.25) node[anchor=north west] {$C_2(v)=$};
\draw (-0.5,-0.25) node[anchor=north west] {$-1$};
\draw (3.75,-0.25) node[anchor=north west] {$1$};
\draw (7,-0.25) node[anchor=north west] {$-1$};
\draw (11.25,-0.25) node[anchor=north west] {$1$};
\draw [->] (0,0) -- (4,0);
\draw [->] (7.5,0) -- (11.5,0);
\draw [->] (11.5,0) -- (11.5,2);
\draw [->] (4,0) -- (4,2);
\draw (4,2) node[anchor=north west] {$\b S^1(v)$};
\draw (11.5,2) node[anchor=north west] {$\b S^1(v)$};
\draw [color=zzttqq] (0,0)-- (4,2);
\draw [color=zzttqq] (4,2)-- (0,2);
\draw [color=zzttqq] (0,2)-- (0,0);
\draw [color=zzttqq] (7.5,2)-- (7.5,0);
\draw [color=zzttqq] (7.5,0)-- (11.5,0);
\draw [color=zzttqq] (11.5,0)-- (7.5,2);
\draw (5.58,1.13) node[anchor=north west] {$-$};
\end{tikzpicture}
\captionof{figure}{The 2-chain $C_2(v)$ where we cut the annulus $\left[ -1,1 \right] \times \b S^1(v)$ along $\left[ -1, 1 \right] \times \lbrace v \rbrace$.} \label{fig_C2v}
\end{center}
\begin{definition}
If $\bar\tau = (N(\gamma);\tau_e,\tau_d, \tau_g)$ is a pseudo-parallelization of a compact oriented 3-manifold $M$, let the \textit{Siamese sections} of $\bar\tau$ denote the following sections of $UM$:
$$
E_1^{d}:
\left\lbrace
\begin{aligned}
m \in M\setminus \mathring N(\gamma) & \longmapsto \tau_e(m,e_1) \\
m \in N(\gamma) & \longmapsto \tau_d(m,e_1)
\end{aligned}
\right.
\hspace{3mm}\mbox{ and }\hspace{3mm}
E_1^{g} :
\left\lbrace
\begin{aligned}
m \in M\setminus \mathring N(\gamma) & \longmapsto \tau_e(m,e_1) \\
m \in N(\gamma) & \longmapsto \tau_g(m,e_1).
\end{aligned}
\right.
$$
\end{definition}
As already mentioned in the introduction, note that when $\bar\tau = (N(\gamma);\tau_e,\tau_d, \tau_g)$ is a pseudo-parallelization of a compact oriented 3-manifold $M$, its pseudo-section at $e_1$ reads
$$
\bar\tau(M\times \lbrace e_1 \rbrace) = \frac{E_1^d(M)+E_1^g(M)}{2}
$$
where $E_1^d$ and $E_1^g$ are the Siamese sections of $\bar\tau$.
\section{From parallelizations to pseudo-parallelizations}
\subsection{Pontrjagin numbers of parallelizations} \label{ssec_defpara}
In this subsection we review the definition of first relative Pontrjagin numbers for parallelizations of compact connected oriented 3-manifolds. For a detailed presentation of these objects we refer to \cite[Section 5]{lescopEFTI} and \cite[Subsection~4.1]{lescopcombing}. \\
Let $C_1$ and $C_2$ be compact connected oriented 3-manifolds with identified boundaries. Recall that a \textit{cobordism from $C_1$ to $C_2$} is a compact oriented 4-manifold $W$ whose boundary reads
$$
\partial W = - C_1 \bigcup_{\partial C_1 \simeq \lbrace 0 \rbrace \times C_1 } -[0,1] \times \partial C_1 \bigcup_{\partial C_2 \simeq \lbrace 1 \rbrace \times C_1} C_2.
$$
Moreover, we require $W$ to be identified with $[0,1[\times C_1$ or $]0,1]\times C_2$ on collars of $\partial W$. \\
Recall that any compact oriented 3-manifold bounds a compact oriented 4-manifold, so that a cobordism from $C_1$ to $C_2$ always exists. Also recall that the signature of a 4-manifold is the signature of the intersection form on its second homology group with real coefficients and that any 4-manifold can be turned into a 4-manifold with signature zero by performing connected sums with copies of $\pm \b C P^2$. So let us fix a connected cobordism $W$ from $C_1$ to $C_2$ with signature zero. \\
Now consider a parallelization $\tau_1$, resp. $\tau_2$, of $C_1$, resp. $C_2$. Define the vector field $\vec n$ on a collar of $\partial W$ as follows. Let $\vec n$ be the unit tangent vector to $[0,1]\times \lbrace x \rbrace$ where $x \in C_1$ or $C_2$. Define $\tau(\tau_1, \tau_2)$ as the trivialization of $TW \otimes \b C$ over $\partial W$ obtained by stabilizing $\tau_1$ or $\tau_2$ into $\vec n \oplus \tau_1$ or $\vec n \oplus \tau_2$ and tensoring with $\b C$. In general, this trivialization does not extend as a parallelization of $W$. This leads to a Pontrjagin obstruction class $p_1(W;\tau(\tau_1, \tau_2))$ in $H^4(W, \partial W, \pi_3(SU(4)))$. Since $\pi_3(SU(4)) \simeq \b Z$, there exists $p_1(\tau_1, \tau_2)\in \b Z$ such that $p_1(W;\tau(\tau_1, \tau_2))=p_1(\tau_1, \tau_2)[W,\partial W]$. Let us call $p_1(\tau_1, \tau_2)$ the \textit{first relative Pontrjagin number of $\tau_1$ and $\tau_2$}. \\
Similarly, define the \textit{Pontrjagin number $p_1(\tau)$ of a parallelization $\tau$} of a closed connected oriented 3-manifold $M$, by taking a connected oriented 4-manifold $W$ with boundary $M$, a collar of $\partial W$ identified with $]0,1] \times M$ and $\vec n$ as the outward normal vector field over $\partial W$. \\
We have not actually defined the sign of the Pontrjagin numbers. We will not give details here on how to define it, instead we refer to \cite[\S 15]{MS} or \cite[p.44]{lescopEFTI}. Let us only mention that $p_1$ is the opposite of the second Chern class $c_2$ of the complexified tangent bundle.
\subsection{Pontrjagin numbers for pseudo-parallelizations}
\begin{definition} \label{def_complextriv}
Let $\bar \tau=(N(\gamma); \tau_e, \tau_d, \tau_g)$ be a pseudo-parallelization of a compact oriented 3-manifold $M$, a \textit{complex trivialization} $\bar \tau _{\b C}$ associated to $\bar\tau$ is a trivialization of $TM \otimes \b C$ such that~:
\begin{enumerate}[\textbullet]
\setlength{\itemsep}{0pt}
\setlength{\parskip}{5pt}
\item $\bar \tau_\b C$ is \textit{special} (\textit{ie}\ its determinant is one everywhere) with respect to the trivialization of the determinant bundle induced by the orientation of $M$,
\item on $M\setminus \mathring{N(\gamma)}$, $\bar \tau_\b C = \tau_e \otimes 1_\b C$,
\item for $m=(t,c,u)\in [a,b] \times \gamma \times [-1,1]$, $\bar \tau_\b C (m,.)= (m,\tau_d (t,c,u)(\l G(t,u)(.)))$,
\end{enumerate}
where $\l G$ is a map so that :
$$
\l G : \left\lbrace
\begin{aligned}
\ [a,b] \times [-1,1] &\longrightarrow \ SU(3) \\
(t,u) &\longmapsto \left\lbrace
\begin{aligned}
& \mbox{Id}_{SU(3)} &\mbox{for $|u| > 1 - \varepsilon$} \\
& \mbox{Id}_{SU(3)} &\mbox{for $t < a+\varepsilon$} \\
& R_{e_1,-\pi-\theta(u)} &\mbox{for $t>b -\varepsilon$}.
\end{aligned}
\right.
\end{aligned}
\right.
$$
Note that such a smooth map $\l G$ on $[a,b]\times[-1,1]$ exists since $\pi_1(SU(3))=\lbrace 1 \rbrace$. Moreover, $\l G$ is well-defined up to homotopy since $\pi_2(SU(3))=\lbrace 0 \rbrace$.
\end{definition}
Pseudo-parallelizations, or pseudo-trivializations, have been first used in \cite[Section 4.3]{lescopKT}, but they have been first defined in \cite[Section 10]{lescopcube}. Note that our conventions are slightly different.
\begin{definition}
Let $\bar\tau_1$ and $\bar\tau_2$ be pseudo-parallelizations of two compact oriented 3-manifolds $M_1$ and $M_2$ with identified boundaries and let $W$ be a cobordism from $M_1$ to $M_2$ with signature zero. As in the case of genuine parallelizations, define a trivialization $\tau(\bar\tau_1, \bar\tau_2)$ of $TW \otimes \b C$ over $\partial W$ using the special complex trivializations $\bar\tau_{1,\b C}$ and $\bar\tau_{2,\b C}$ associated to $\bar\tau_1$ and $\bar\tau_2$, respectively. The \textit{first relative Pontrjagin number of $\bar\tau_1$ and $\bar\tau_2$} is the Pontrjagin obstruction $p_1(\bar\tau_1,\bar\tau_2)$ to extending the trivialization $\tau(\bar\tau_1, \bar\tau_2)$ as a trivialization of $TW\otimes \b C$ over $W$.
\end{definition}
Finally, if $\bar\tau$ is a pseudo-parallelization of a closed oriented 3-manifold $M$, then define the \textit{Pontrjagin number $p_1(\bar\tau)$ of the pseudo-parallelization $\bar\tau$} as $p_1(\tau_\emptyset, \bar \tau)$ as before.
\subsection[Variation of $p_1$ as an intersection of three 4-chains]{Variation of $p_1$ as an intersection of three 4-chains}
In this subsection, we give a proof of Theorem~\ref{prop_varasint}, which expresses the relative Pon\-trjagin numbers (resp. the variation of Pontrjagin numbers) of pseudo-parallelizations in compact (resp. closed) oriented 3-manifolds as an algebraic intersection of three 4-chains.
\begin{lemma} \label{simppara}
If $\bar \tau=(N(\gamma);\tau_e,\tau_d,\tau_g)$ is a pseudo-parallelization of a compact oriented 3-manifold $M$, if $E_1^d$ and $E_1^g$ are its Siamese sections and if $E_2^e$ denotes the second vector of $\tau_e$, then $P(e_2^M({E_1^d}^\perp,{E_2^e}_{|\partial M})) = -P(e_2^M({E_1^g}^\perp,{E_2^e}_{|\partial M})) = [\gamma]$ in $H_1(M;\b Z)$.
\end{lemma}
\begin{proof}
Since $E_1^d$ and $E_1^e$ coincide on $M\setminus \mathring N(\gamma)$, the obstruction to extending $E_2^e$ as a section of ${E_1^d}^\perp$ is the obstruction to extending $E_2^e$ as a section of ${E_1^d}^\perp_{|N(\gamma)}$. However, parallelizing $N(\gamma)$ with $\tau_d$ and using that
$$
\tau_d = \left\lbrace
\begin{aligned}
& \tau_e &\mbox{ on } \partial (\left[ a , b \right] \times \gamma \times \left[ -1 , 1 \right]) \setminus \lbrace b \rbrace \times \gamma \times \left[ -1 , 1 \right] \\
& \tau_e \circ \l T_\gamma &\mbox{ on } \lbrace b \rbrace \times \gamma \times \left[ -1 , 1 \right]
\end{aligned}
\right.
$$
we get that $E_2^e$ induces a degree +1 map ${E_2^e}_{|\alpha} : \alpha \rightarrow \b S^1$ on any meridian $\alpha$ of $N(\gamma)$. It follows that
$$
P(e_2^M({E_1^d}^\perp,{E_2^e}_{|\partial M})) = + [\gamma].
$$
Similarly, parallelizing $N(\gamma)$ with $\tau_g$, $E_2^e$ induces a degree -1 map ${E_2^e}_{|\alpha} : \alpha \rightarrow \b S^1$ on any meridian $\alpha$ of $N(\gamma)$, so that
$$
P(e_2^M({E_1^g}^\perp,{E_2^e}_{|\partial M})) = - [\gamma].
$$
\end{proof}
Recall that for a combing $X$ of a compact oriented 3-manifold $M$ and a pseudo-parallelization $\bar\tau$ of $M$, if $X$ and $\bar\tau$ are compatible, then
$$
L_{\bar\tau=X}=\frac{L_{E_1^d=X}+L_{E_1^g=X}}{2} \mbox{ \ \ and \ \ } L_{\bar\tau=-X}=\frac{L_{E_1^d=-X}+L_{E_1^g=-X}}{2}
$$
where $E_1^d$ and $E_1^g$ denote the Siamese sections of $\bar\tau$.
\begin{lemma} \label{nulexcep}
Let $\bar \tau=(N(\gamma);\tau_e,\tau_d,\tau_g)$ be a pseudo-parallelization of a compact oriented 3-manifold $M$. If $X$ is a torsion combing of $M$ compatible with $\bar \tau$, then $L_{\bar\tau=X}$ and $L_{\bar\tau=-X}$ are rationally null-homologous in $M$.
\end{lemma}
\begin{proof}
Let $E_1^d$ and $E_1^g$ be the Siamese sections of $\bar\tau$. Using Proposition~\ref{prop_linksinhomologyI} and the fact that $X$ is a torsion combing, we get, in $H_1(M;\b Q)$,
$$
\begin{aligned}
\ 2 \cdot [L_{X=-E_1^d}+L_{X=-E_1^g} ] &= [-P(e_2^M({E_1^d}^\perp,{E_2^e}_{|\partial M})) -P(e_2^M({E_1^g}^\perp,{E_2^e}_{|\partial M})) ] \\
\ 2 \cdot [L_{X=E_1^d}+L_{X=E_1^g} ] &= [P(e_2^M({E_1^d}^\perp,{E_2^e}_{|\partial M})) + P(e_2^M({E_1^g}^\perp,{E_2^e}_{|\partial M})) ]
\end{aligned}
$$
where $E_2^e$ is the second vector of $\tau_e$. Conclude with Lemma~\ref{simppara}.
\end{proof}
\begin{definition} \label{def_Ft}
Let $X$ and $Y$ be $\partial$-compatible combings of a compact oriented 3-manifold $M$. For all $t \in [0,1]$, let $\bar F_t(X,Y)$ denote the 4-chain of $[0,1]\times UM$~:
$$
\bar F_t(X,Y) = [0,t] \times X(M) + \lbrace t \rbrace \times \bar F(X,Y) + [t,1] \times Y(M),
$$
where $\bar F(X,Y)$ is a 4-chain of $UM$ as in Lemma~\ref{lem_phomotopy}. Note that :
$$
\partial \bar F_t(X,Y) = \lbrace 1 \rbrace \times Y(M) - \lbrace 0 \rbrace \times X(M) - [0,1] \times X(\partial M) + \lbrace t \rbrace \times UM_{L_{X=-Y}}.
$$
\end{definition}
\begin{lemma} \label{lem_transfert}
Let $\bar \tau=(N(\gamma);\tau_e,\tau_d,\tau_g)$ be a pseudo-parallelization of a compact oriented 3-manifold $M$. If $X$ is a torsion combing of $M$ compatible with $\bar \tau$, then there exist 4-chains of $[0,1]\times UM$, $C_4^\pm(\bar\tau,X)$ and $C_4^\pm(X,\bar\tau)$, with boundaries :
$$
\begin{aligned}
\partial C_4^\pm(\bar\tau,X) &= \lbrace 1 \rbrace \times (\pm X)(M) - \lbrace 0 \rbrace \times \bar\tau(M\times \lbrace \pm e_1 \rbrace ) - [0,1] \times (\pm X)(\partial M) \\
\partial C_4^\pm(X,\bar\tau) &= \lbrace 1 \rbrace \times \bar\tau(M\times \lbrace \pm e_1 \rbrace ) - \lbrace 0 \rbrace \times (\pm X)(M) - [0,1] \times (\pm X)(\partial M) \\
\end{aligned}
$$
\end{lemma}
\begin{proof}
Let $E_1^d$ and $E_1^g$ be the Siamese sections of $\bar\tau$ and just set
$$
\begin{aligned}
C_4^\pm(\bar\tau,X) &= \sfrac{1}{2} \cdot \left( \bar F_t( \pm E_1^d, \pm X) + \bar F_t( \pm E_1^g, \pm X) - \lbrace t \rbrace \times UM_{|\Sigma(\pm e_1)} \right) \\
C_4^\pm(X,\bar\tau) &= \sfrac{1}{2} \cdot \left( \bar F_t( \pm X, \pm E_1^d) + \bar F_t( \pm X, \pm E_1^g) - \lbrace t \rbrace \times UM_{|-\Sigma(\pm e_1)} \right)
\end{aligned}
$$
where the 4-chains $\bar F_t$ are as in Definition~\ref{def_Ft} and where $\Sigma(\pm e_1)$ are rational 2-chains of $M$ bounded by $\pm(L_{E_1^d=-X}+L_{E_1^g=-X})$, which are rationally null-homologous according to Lemma~\ref{nulexcep}.
\end{proof}
\begin{remark} \label{rmkpcase}
Recall that a genuine parallelization $\tau$ of a compact oriented 3-manifold is a pseudo-parallelization whose link is empty. In such a case, $E_1^d$ and $E_1^g$ are the first vector $E_1^\tau$ of the parallelization $\tau$ and the chains $C_4^{\pm}$ can be simply defined as
$$
\begin{aligned}
C_4^\pm(\tau,X) &= \bar F_t( \pm E_1^\tau, \pm X) - \lbrace t \rbrace \times UM_{|\Sigma(\pm e_1)} \\
C_4^\pm(X,\tau) &= \bar F_t( \pm X, \pm E_1^\tau) - \lbrace t \rbrace \times UM_{|-\Sigma(\pm e_1)}
\end{aligned}
$$
where the 4-chains $\bar F_t$ are as in Definition~\ref{def_Ft} and where $\Sigma(\pm e_1)$ are rational 2-chains of $M$ bounded by $\pm L_{E_1^\tau=-X}$.
\end{remark}
\begin{lemma} \label{lem_C4pme1}
Let $\tau$ and $\bar \tau$ be two pseudo-parallelizations of a compact oriented 3-manifold $M$ that coincide on $\partial M$ and whose links are disjoint. For all $v\in \b S^2$, there exists a 4-chain $C_4(M,\tau,\bar\tau ;v)$ of $[0,1] \times UM$ such that
$$
\partial C_4(M,\tau,\bar\tau ;v) = \lbrace 1 \rbrace \times \bar \tau (M \times \lbrace v \rbrace)- \lbrace 0 \rbrace \times \tau (M \times \lbrace v \rbrace) - [0,1] \times \tau (\partial M\times \lbrace v \rbrace).
$$
\end{lemma}
\begin{proof}
Let us write $C_4(\tau,\bar\tau ;v)$ instead of $C_4(M,\tau,\bar\tau ;v)$ when there is no ambiguity. Since the 3-chains $\partial C_4(\tau,\bar\tau ;v)$, where $v \in \b S^2$, are homologous, it is enough to prove the existence of $C_4(\tau,\bar\tau ;e_1)$. First, let $X$ be a combing of $M$ such that $X$ is compatible with $\tau$ and $\bar \tau$. In general, this combing is not a torsion combing. Second, let $E_1^d$ and $E_1^g$ (resp. $\bar E_1^d$ and $\bar E_1^g$) denote the Siamese section of $\tau$, (resp. $\bar\tau$) and set
$$
\begin{aligned}
\bar F(\tau,X) &= \sfrac{1}{2} \cdot \left( \bar F( E_1^d,X) + \bar F( E_1^g,X) \right) \\
\bar F(X,\bar\tau) &= \sfrac{1}{2} \cdot \left( \bar F(X, \bar E_1^d) + \bar F(X, \bar E_1^g)\right).
\end{aligned}
$$
These chains have boundaries :
$$
\begin{aligned}
\partial \bar F(\tau,X) &= X(M) - \tau(M\times \lbrace e_1 \rbrace) + UM_{|L_{\tau=-X}} \\
\partial \bar F(X,\bar\tau) &= \bar\tau(M\times \lbrace e_1 \rbrace) - X(M) + UM_{|-L_{\bar\tau=-X}}.
\end{aligned}
$$
Hence, for all $t \in [0,1]$, the 4-chain of $[0,1]\times UM$
$$
\bar F_t(\tau, \bar\tau; e_1) = [0,t] \times \tau(M\times \lbrace e_1 \rbrace) + \lbrace t \rbrace \times \left(\bar F(\tau,X) + \bar F(X,\bar\tau)\right) + [t,1] \times \bar\tau(M\times \lbrace e_1 \rbrace)
$$
has boundary :
$$
\begin{aligned}
\partial \bar F_t(\tau, \bar\tau; e_1) &= \lbrace 1 \rbrace \times \bar\tau(M\times \lbrace e_1 \rbrace) - \lbrace 0 \rbrace \times \tau(M\times \lbrace e_1 \rbrace) \\
&- [0,1] \times \tau(\partial M\times \lbrace e_1 \rbrace) + \lbrace t \rbrace \times UM_{|L_{\tau=-X}\cup - L_{\bar\tau=-X}}.
\end{aligned}
$$
Thanks to Proposition~\ref{prop_linksinhomologyI} and Lemma~\ref{simppara}, in $H_1(M;\b Q)$ :
$$
\begin{aligned}
4 \cdot [ L_{\tau=-X} ] &= 2\cdot [L_{E_1^d=-X} + L_{E_1^g=-X}] \\
&= [-2 \cdot P(e_2^M(X^\perp,{E_2^e}_{|\partial M}))\hspace{-1mm}+\hspace{-1mm}P(e_2^M({E_1^d}^\perp,{E_2^e}_{|\partial M}))\hspace{-1mm}+\hspace{-1mm}P(e_2^M({E_1^g}^\perp,{E_2^e}_{|\partial M}))] \\
&= -2 \cdot[P(e_2^M(X^\perp,{E_2^e}_{|\partial M}))]
\end{aligned}
$$
where $E_2^e$ is the second vector of $\tau_e$. Similarly, $2\cdot [L_{\bar\tau=-X}]=-[P(e_2^M(X^\perp,{E_2^e}_{|\partial M}))]$ in $H_1(M;\b Q)$. So, the link $L_{\tau=-X}\cup - L_{\bar\tau=-X}$ is rationally null-homologous in $M$, \textit{ie}\ there exists a rational 2-chain $\Sigma(\tau,\bar\tau)$ such that $\partial \Sigma(\tau,\bar\tau) = L_{\tau=-X}\cup - L_{\bar\tau=-X}$. Hence, we get a 4-chain $C_4(\tau,\bar\tau;e_1)$ as desired by setting
$$
C_4(\tau,\bar\tau;e_1) = \bar F_t(\tau, \bar\tau; e_1) - \lbrace t \rbrace \times UM_{|\Sigma(\tau,\bar\tau)}.
$$
\end{proof}
\begin{lemma} \label{lem_welldefined}
Let $\tau$ and $\bar \tau$ be two pseudo-parallelizations of a compact oriented 3-manifold $M$ that coincide on $\partial M$. If $x$, $y$ and $z$ are points in $\b S^2$ with pairwise different distances to $e_1$, then there exist pairwise transverse 4-chains $C_4(\tau,\bar\tau ;x)$, $C_4(\tau,\bar\tau ;y)$ and $C_4(\tau,\bar\tau ;z)$ as in Lemma~\ref{lem_C4pme1} and the algebraic intersection $\langle C_4(\tau,\bar\tau ;x), C_4(\tau,\bar\tau ;y), C_4(\tau,\bar\tau ;z) \rangle_{[0,1]\times UM}$ only depends on $\tau$ and~$\bar\tau$.
\end{lemma}
\begin{proof}
Pick any $x$, $y$ and $z$ in $\b S^2$ with pairwise different distances to $e_1$ and consider some 4-chains $C_4(\tau,\bar\tau ;x)$, $C_4(\tau,\bar\tau ;y)$ and $C_4(\tau,\bar\tau ;z)$ such that, for $v \in \lbrace x , y , z \rbrace$,
$$
\partial C_4(\tau,\bar\tau ;v) = \lbrace 1 \rbrace \times \bar\tau(M\times\lbrace v \rbrace) - \lbrace 0 \rbrace \times \tau (M\times \lbrace v \rbrace ) - [0,1] \times \tau (\partial M \times \lbrace v \rbrace ).
$$
The intersection of $C_4(\tau,\bar\tau;x)$, $C_4(\tau,\bar\tau;y)$ and $C_4(\tau,\bar\tau;z)$ is in the interior of $[0,1]\times UM$. The algebraic triple intersection of these three 4-chains only depends on the fixed boundaries and on the homology classes of the 4-chains. The space $H_4([0,1]\times UM ; \b Q)$ is generated by the classes of 4-chains $\Sigma \times \b S^2$ where $\Sigma$ is a surface in $M$. If $\Sigma \times \b S^2$ is such a 4-chain, then
$$
\begin{aligned}
& \langle C_4(\tau,\bar\tau ;x) + \Sigma\times \b S^2, C_4(\tau,\bar\tau ;y), C_4(\tau,\bar\tau ;z) \rangle - \langle C_4(\tau,\bar\tau ;x) , C_4(\tau,\bar\tau ;y), C_4(\tau,\bar\tau ;z) \rangle \\
&= \langle \Sigma\times \b S^2, C_4(\tau,\bar\tau ;y), C_4(\tau,\bar\tau ;z) \rangle \\
&= \left\lbrace
\begin{aligned}
& \langle \Sigma\times \b S^2, [0,1] \times \tau(M \times \lbrace y \rbrace ) , [0,1] \times \tau(M \times \lbrace z \rbrace) \rangle \mbox{ , pushing $\Sigma \times \b S^2$ near $0$,} \\
& \langle \Sigma\times \b S^2, [0,1] \times \bar\tau(M \times \lbrace y \rbrace ) , [0,1] \times \bar \tau(M \times \lbrace z \rbrace) \rangle \mbox{ , pushing $\Sigma \times \b S^2$ near $1$.} \\
\end{aligned}
\right.
\end{aligned}
$$
Hence, $\langle \Sigma\times \b S^2, C_4(\tau,\bar\tau ;y), C_4(\tau,\bar\tau ;z) \rangle $ is independent of $\tau$ and $\bar\tau$. So, use Lemma~\ref{lem_extendpparallelization} to extend a trivialization of $TM_{|\Sigma}$ as a pseudo-parallelization $\tau'$ that coincides with $\tau$ and $\bar\tau$ on $\partial M$. Considering this pseudo-parallelization we get
$$
\langle \Sigma\times \b S^2, C_4(\tau,\bar\tau ;y), C_4(\tau,\bar\tau ;z) \rangle \hspace{-1mm}=\hspace{-1mm} \langle \Sigma\times \b S^2, [0,1] \times \tau'(M \times \lbrace y \rbrace ) , [0,1] \times \tau'(M \times \lbrace z \rbrace) \rangle \hspace{-1mm}=\hspace{-1mm} 0,
$$
so that the algebraic triple intersection of the three chains $C_4(\tau,\bar\tau;x)$, $C_4(\tau,\bar\tau;y)$ and $C_4(\tau,\bar\tau;z)$ only depends on their fixed boundaries.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{prop_varasint}]
Let $\tau$ and $\bar \tau$ be two pseudo-parallelizations of a compact oriented 3-manifold $M$ that coincide on $\partial M$ and whose links are disjoint. To conclude the proof of Theorem~\ref{prop_varasint}, we have to prove that for any $x$, $y$ and $z$ in $\b S^2$ with pairwise different distances to $e_1$ :
$$
p_1(\tau, \bar\tau)= 4 \cdot \langle C_4(M,\tau,\bar\tau ;x),C_4(M,\tau,\bar\tau ;y),C_4(M,\tau,\bar\tau ;z) \rangle_{[0,1]\times UM}.
$$
First, we know from \cite[Lemma 10.9]{lescopcube} that this is true if $M$ is a $\b Q$HH of genus 1. Notice that it is also true if $M$ embeds in such a manifold. Indeed, if $\b H$ is a $\b Q$HH of genus 1 and if $M$ embeds in $\b H$ then, using Lemma~\ref{lem_extendpparallelization} and using that $\tau$ and $\bar \tau$ coincide on $\partial M$, there exists a pseudo-parallelization $\check \tau$ of $\b H \setminus \mathring M$ such that
$$
\bar \tau_{\b H} : \left\lbrace
\begin{aligned}
m \in M &\mapsto \bar\tau(m,.) \\
m \in \b H\setminus \mathring M & \mapsto \check \tau(m,.)
\end{aligned} \right.
\mbox{ \ and \ }
\tau_{\b H} : \left\lbrace
\begin{aligned}
m \in M &\mapsto \tau(m,.) \\
m \in \b H\setminus \mathring M & \mapsto \check \tau(m,.)
\end{aligned} \right.
$$
are pseudo-parallelizations of $\b H$. Furthermore, for any $v \in \b S^2$, let $C_4(\b H,v)$ be the 4-chain of $[0,1]\times U\b H$ :
$$
C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;v) = C_4(M,\tau,\bar\tau ;v)\cup [0,1] \times \check \tau( (\b H \setminus \mathring M) \times \lbrace v \rbrace )
$$
where $C_4(M,\tau,\bar\tau ;v)$ is as in Lemma~\ref{lem_C4pme1}. The boundary of $C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;v)$ is~:
$$
\partial C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;v) = \lbrace 1 \rbrace \times \bar\tau_{\b H}(M\times \lbrace v \rbrace ) - \lbrace 0 \rbrace \times \tau_{\b H}(M\times \lbrace v \rbrace ) - [0,1] \times \tau_{\b H}(\partial M\times \lbrace v \rbrace ).
$$
Using the definition of Pontrjagin numbers of pseudo-parallelizations and the hypothesis on $\b H$, it follows that if $x,y$ and $z$ are points in $\b S^2$ with pairwise different distances to $e_1$ :
$$
\begin{aligned}
p_1(\tau,\bar\tau) = p_1(\tau_{\b H},\bar\tau_{\b H})
= 4 \cdot \langle C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;x) , C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;y) , C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;z) \rangle.
\end{aligned}
$$
Now note that :
$$
\langle [0,1] \times \check \tau (\b H\setminus \mathring M \times \lbrace x \rbrace) , [0,1] \times \check \tau (\b H\setminus \mathring M \times \lbrace y \rbrace ), [0,1] \times \check \tau (\b H\setminus \mathring M \times \lbrace z \rbrace) \rangle = 0.
$$
Indeed, if $\check \tau = (N(\check \gamma); \check \tau_e,\check \tau_d,\check \tau_g)$, for all $v\in \b S^2$ the pseudo-section of $\check\tau$ reads
$$
\begin{aligned}
\check \tau (M\times \lbrace v \rbrace ) &= \check\tau_e((M\setminus \mathring N(\check\gamma))\times \lbrace v \rbrace ) \\
&+ \frac{\check \tau_d(N(\check\gamma)\times \lbrace v \rbrace) + \check\tau_g(N(\check\gamma)\times \lbrace v \rbrace) + \check\tau_e( \lbrace b \rbrace \times \check\gamma \times C_2(v)) }{2}.
\end{aligned}
$$
The 3-chains $\check\tau_e((M\setminus \mathring N(\check\gamma))\times \lbrace v \rbrace )$, for $v\in \lbrace x,y,z \rbrace$, are pairwise disjoint since $\check\tau_e$ is a genuine parallelization and since $x,y$ and $z$ are pairwise distinct points in $\b S^2$. Moreover, the 3-chains $\check\tau_e( \lbrace b \rbrace \times \check\gamma \times C_2(v))$, for $v\in \lbrace x,y,z\rbrace$, are also pairwise disjoint since they are subsets of the $\check\tau_e(\lbrace b \rbrace \times \check\gamma \times \b S^1(v))$ , $v \in \lbrace x,y,z \rbrace$, which are pairwise disjoint since $x,y$ and $z$ have pairwise different distances to $e_1$. Finally, we have~:
$$
\begin{aligned}
\langle \check\tau_d(N(\check \gamma) \times \lbrace x\rbrace)+\check\tau_g(N(\check \gamma) \times \lbrace x\rbrace) , \check\tau_d(N(\check \gamma) \times \lbrace y\rbrace)+\check\tau_g(N(\check \gamma) \times \lbrace y\rbrace) , \\ \check\tau_d(N(\check \gamma) \times \lbrace z\rbrace)+\check\tau_g(N(\check \gamma) \times \lbrace z\rbrace) \rangle
&= 0
\end{aligned}
$$
since a triple intersection between the 3-chains
$$
\lbrace \check\tau_d(N(\check \gamma) \times \lbrace v\rbrace)+\check\tau_g(N(\check \gamma) \times \lbrace v\rbrace)\rbrace_{v\in \lbrace x,y,z\rbrace}
$$ would be contained in an intersection between two of the $\lbrace \check\tau_d(N(\check \gamma)\times \lbrace v\rbrace)\rbrace_{v\in \lbrace x,y,z \rbrace}$ or between two of the $\lbrace \check\tau_g(N(\check \gamma)\times \lbrace v\rbrace)\rbrace_{v\in \lbrace x,y,z \rbrace}$ which must be empty since $\check\tau_d$ and $\check\tau_g$ are genuine parallelizations. It follows that
$$
\begin{aligned}
p_1(\tau,\bar\tau) &= 4 \cdot\langle C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;x) , C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;y) , C_4(\b H,\tau_{\b H},\bar\tau_{\b H} ;z) \rangle \\
&= 4 \cdot\langle C_4(M,\tau,\bar\tau ;x) , C_4(M,\tau,\bar\tau ;y) , C_4(M,\tau,\bar\tau ;z) \rangle.
\end{aligned}
$$
Using the same construction, note also that it is enough to prove the statement when $M$ is a closed oriented 3-manifold since any oriented 3-manifold embeds into a closed one. \\
Let us finally prove Theorem~\ref{prop_varasint} when $M$ is a closed oriented 3-manifold. Consider a Heegaard splitting $M = H_1 \cup_\Sigma H_2$ such that there is a collar $\Sigma\times[0,1] \subset H_2$ of $\Sigma$ verifying
$$
N(\bar \gamma) \cap \left(\Sigma\times[0,1] \right) = \emptyset \mbox{ \ \ and \ \ } N( \gamma) \cap \left(\Sigma\times[0,1] \right) = \emptyset
$$
where $\gamma$ and $\bar\gamma$ are the links of $\tau$ and $\bar\tau$, respectively, and such that $\Sigma=\Sigma\times \lbrace 0 \rbrace$. Such a splitting can be obtained by considering a triangulation of $M$ containing $\gamma$ and $\bar\gamma$ in its 1-skeleton, and then defining $H_1$ as a tubular neighborhood of this 1-skeleton.
Using Lemma~\ref{lem_extendpparallelization}, we can construct a pseudo-parallelization $\tau^c$ of $\Sigma\times [0,1]$ such that $\tau^c$ coincides with $\bar\tau$ on $\Sigma \times \lbrace 1 \rbrace$ and with $\tau$ on $\Sigma \times \lbrace 0 \rbrace$. Then, write $H'_1 = H_1 \cup (\Sigma\times[0,1])$ and $H'_2=H_2\setminus( \Sigma\times[0,1[)$ -- see Figure~\ref{figHH} -- and set
$$
\check \tau : \left\lbrace
\begin{aligned}
(m,v) \in UH_1 &\longmapsto \tau(m,v) \\
(m,v) \in U(\Sigma\times[0,1]) &\longmapsto \tau^c(m,v) \\
(m,v) \in UH'_2 &\longmapsto \bar\tau(m,v).
\end{aligned}
\right.
$$
\begin{center}
\definecolor{zzttqq}{rgb}{0.6,0.2,0}
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=0.75cm]
\clip(-3.5,-3.5) rectangle (5.5,3.5);
\fill[line width=0pt,color=zzttqq,fill=zzttqq,fill opacity=0.15] (-3,2) -- (0,2) -- (0,-2) -- (-3,-2) -- cycle;
\fill[line width=0pt,color=zzttqq,fill=zzttqq,fill opacity=0.05] (0,2) -- (2,2) -- (2,-2) -- (0,-2) -- cycle;
\fill[line width=0pt,color=zzttqq,fill=zzttqq,fill opacity=0.15] (2,2) -- (5,2) -- (5,-2) -- (2,-2) -- cycle;
\draw (0,2)-- (0,-2);
\draw (2,2)-- (2,-2);
\draw (0,2.4)-- (0,2.6);
\draw (0,2.6)-- (5,2.6);
\draw (5,2.4)-- (5,2.6);
\draw (2,-2.4)-- (2,-2.6);
\draw (2,-2.6)-- (-3,-2.6);
\draw (-3,-2.6)-- (-3,-2.4);
\begin{scriptsize}
\draw (-1.75,0.25) node[anchor=north west] {$H_1$};
\draw (-1.75,-0.5) node[anchor=north west] {$\tau$};
\draw (3.25,0.25) node[anchor=north west] {$H'_2$};
\draw (3.25,-0.5) node[anchor=north west] {$\bar\tau$};
\draw (0.75,-0.5) node[anchor=north west] {$\tau^c$};
\draw (-0.75,-2.75) node[anchor=north west] {$H'_1$};
\draw (2.44,3.11) node[anchor=north west] {$H_2$};
\draw (-0.5,-2) node[anchor=north west] {$\Sigma\times\lbrace 0\rbrace$};
\draw (1.5,2.5) node[anchor=north west] {$\Sigma\times \lbrace 1 \rbrace$};
\end{scriptsize}
\end{tikzpicture}
\captionof{figure}
{}\label{figHH}
\end{center}
For $v \in \b S^2$, consider some 4-chains $C_4(H_1,\tau,\check\tau ;v)$, $C_4(H_2,\tau,\check\tau ;v)$, $C_4(H'_1,\check\tau,\bar\tau ;v)$ and $C_4(H'_2,\check\tau,\bar\tau ;v)$ of $[0,1]\times UH_1$, $[0,1]\times UH_2$, $[0,1]\times UH'_1$ and $[0,1]\times UH'_2$, respectively, such that :
$$
\begin{aligned}
\partial C_4(H_1,\tau,\check\tau ;v) &\hspace{-1mm}=\hspace{-1mm} \lbrace 1 \rbrace \times \check \tau (H_1 \times \lbrace v \rbrace) - \lbrace 0 \rbrace \times \tau (H_1\times \lbrace v \rbrace ) - [0,1] \times \tau (\partial H_1 \times \lbrace v \rbrace) \\
\partial C_4(H_2,\tau,\check\tau ;v) &\hspace{-1mm}=\hspace{-1mm} \lbrace 1 \rbrace \times \check \tau (H_2 \times \lbrace v \rbrace) - \lbrace 0 \rbrace \times \tau (H_2\times \lbrace v \rbrace ) - [0,1] \times \tau (\partial H_2 \times \lbrace v \rbrace) \\
\end{aligned}
$$
and
$$
\begin{aligned}
\partial C_4(H'_1,\check\tau,\bar\tau ;v) &\hspace{-1mm}=\hspace{-1mm} \lbrace 1 \rbrace \times \bar \tau (H'_1 \times \lbrace v \rbrace) - \lbrace 0 \rbrace \times \check \tau (H'_1\times \lbrace v \rbrace ) - [0,1] \times \check \tau (\partial H'_1 \times \lbrace v \rbrace) \\
\partial C_4(H'_2,\check\tau,\bar\tau ;v) &\hspace{-1mm}=\hspace{-1mm}\lbrace 1 \rbrace \times \bar \tau (H'_2 \times \lbrace v \rbrace) - \lbrace 0 \rbrace \times \check \tau (H'_2\times \lbrace v \rbrace ) - [0,1] \times \check \tau (\partial H'_2 \times \lbrace v \rbrace). \\
\end{aligned}
$$
Since $H_1$ and $H_2$ embed in rational homology balls, for any $x,y$ and $z$ in $\b S^2$ with pairwise different distances to $e_1$
$$
\begin{aligned}
p_1(\tau_{|H_1},\check\tau_{|H_1}) &= 4 \cdot\langle C_4(H_1,\tau,\check\tau ;x) , C_4(H_1,\tau,\check\tau ;y) , C_4(H_1,\tau,\check\tau ;z) \rangle_{[0,1]\times UH_1} \\
p_1(\tau_{|H_2},\check\tau_{|H_2}) &= 4 \cdot\langle C_4(H_2,\tau,\check\tau ;x) , C_4(H_2,\tau,\check\tau ;y) , C_4(H_2,\tau,\check\tau ;z) \rangle_{[0,1]\times UH_2} \\
\end{aligned}
$$
so that, using $C_4(M,\tau,\check\tau;v) = C_4(H_1,\tau,\check\tau;v)+C_4(H_2,\tau,\check\tau;v)$ for $v\in \b S^2$,
$$
p_1(\tau,\check\tau) = 4 \cdot\langle C_4(M,\tau,\check\tau ;x) , \ C_4(M,\tau,\check\tau ;y) , \ C_4(M,\tau,\check\tau ;z) \rangle_{[0,1]\times UM}.
$$
Similarly, since $H'_1$ and $H'_2$ embed in rational homology balls, for any $x,y$ and $z$ in $\b S^2$ with pairwise different distances to $e_1$
$$
\begin{aligned}
p_1(\check\tau_{|H'_1},\bar\tau_{|H'_1}) &= 4 \cdot\langle C_4(H'_1,\check\tau,\bar\tau ;x) , C_4(H'_1,\check\tau,\bar\tau ;y) , C_4(H'_1,\check\tau,\bar\tau ;z) \rangle _{[0,1]\times UH'_1} \\
p_1(\check\tau_{|H'_2},\bar\tau_{|H'_2}) &= 4 \cdot\langle C_4(H'_2,\check\tau,\bar\tau ;x) , C_4(H'_2,\check\tau,\bar\tau ;y) , C_4(H'_2,\check\tau,\bar\tau ;z)\rangle_{[0,1]\times UH'_2} \\
\end{aligned}
$$
so that, using $C_4(M,\check\tau,\bar\tau;v) = C_4(H'_1,\check\tau,\bar\tau;v)+C_4(H'_2,\check\tau,\bar\tau;v)$ for $v\in \b S^2$,
$$
p_1(\check\tau,\bar\tau) = 4 \cdot\langle C_4(M,\check\tau,\bar\tau;x), C_4(M,\check\tau,\bar\tau;y), C_4(M,\check\tau,\bar\tau;z) \rangle_{[0,1]\times UM}.
$$
Eventually, reparameterizing and stacking $C_4(M,\tau,\check\tau;v)$ and $C_4(M,\check\tau,\bar\tau;v) $, for all $v \in \b S^2$ we get a 4-chain $C_4(M,\tau,\bar\tau; v)$ of $[0,1]\times UM$ such that
$$
\partial C_4(M,\tau,\bar\tau;v) = \lbrace 1 \rbrace \times \bar \tau (M \times \lbrace v \rbrace) - \lbrace 0 \rbrace \times \tau(M\times \lbrace v \rbrace) - [0,1]\times \tau (\partial M \times \lbrace v \rbrace)
$$
and such that for any $x,y$ and $z$ in $\b S^2$ with pairwise different distances to $e_1$
$$
p_1(\tau, \bar \tau)= 4 \cdot\langle C_4(M,\tau,\bar\tau;x), C_4(M,\tau,\bar\tau;y), C_4(M,\tau,\bar\tau;z) \rangle_{[0,1]\times UM}.
$$
\end{proof}
\section{From pseudo-parallelizations to torsion combings}
\subsection{Variation of $p_1$ as an intersection of two 4-chains}
\begin{definition}
Let $M$ be a compact oriented 3-manifold. A trivialization $\rho$ of $TM_{|\partial M}$ is \textit{admissible} if there exists a section $X$ of $UM$ such that $(X,\rho)$ is a torsion combing of $M$.
\end{definition}
\begin{lemma} \label{lem_HtMb}
Let $M$ be a compact oriented 3-manifold, let $\rho$ be an admissible trivialization of $TM_{|\partial M}$ and let $S_1, S_2 , \ldots , S_{\beta_1(M)}$ be surfaces in $M$ comprising a basis of $H_2(M;\b Q)$. The subspace $H_T^\rho(M)$ of $H_2(UM;\b Q)$ generated by $\lbrace [X(S_1)], \ldots, [X(S_{\beta_1(M)})] \rbrace$ where $X$ is a section of $UM$ such that $(X,\rho)$ is a torsion combing of $M$, only depends on $\rho$.
\end{lemma}
\begin{proof}
Let $(Y,\rho)$ be another choice of torsion combing of $M$. Assume, without loss of generality, that $(X,\rho)$ and $(Y,\rho)$ are $\partial$-compatible, and let $C(X,Y)$ be the 4-chain of $UM$
$$
C(X,Y) = \bar F(X,Y) - UM_{|\Sigma_{X=-Y}}
$$
constructed using Lemma~\ref{lem_phomotopy} and Proposition~\ref{prop_linksinhomologyI}, which provide $\bar F(X,Y)$ and a 2-chain $\Sigma_{X=-Y}$ of $M$ bounded by $L_{X=-Y}$, respectively. For $i \in \lbrace 1,2,\ldots,\beta_1(M)\rbrace$,
$$
Y(S_i) - X(S_i) = \partial (C(X,Y) \cap UM_{|S_i}).
$$
\end{proof}
\begin{lemma} \label{lem_evaluationII}
Let $M$ be a compact oriented 3-manifold, let $\rho$ be an admissible trivialization of $TM_{|\partial M}$ and let $(X,\rho)$ and $(Y,\rho)$ be $\partial$-compatible torsion combings of $M$. There exists a 4-chain $C_4(X,Y)$ of $[0,1]\times UM$ such that
$$
\partial C_4(X,Y) = \lbrace 1 \rbrace \times Y(M) - \lbrace 0 \rbrace \times X(M) - [0,1] \times X(\partial M).
$$
For any such chain $C_4(X,Y)$, if $C$ is a 2-cycle of $[0,1]\times UM$ then,
$$
[C] = \langle C , C_4(X,Y) \rangle_{[0,1] \times UM} [S] \mbox{ \ in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$,}
$$
where $[S]$ is the homology class of the fiber of $UM$ in $H_2([0,1]\times UM ; \b Q)$.
\end{lemma}
\begin{proof}
Observe that $H_2(UM;\b Q)$ is generated by the family $\lbrace [Z(S_1)], \ldots, [Z(S_{\beta_1(M)})], [S] \rbrace$ where $S_1, \ldots , S_{\beta_1(M)}$ are surfaces in $M$ comprising a basis of $H_2(M;\b Q)$ and where $Z$ is a torsion combing of $M$ that coincides with $X$ and $Y$ on $\partial M$. Let $C_4(X,Y)$ be the 4-chain
$$
C_4(X,Y) = \bar F_t(X,Y) - \lbrace t \rbrace \times UM_{|\Sigma_{X=-Y}}
$$
where $\bar F_t(X,Y)$ is a 4-chain as in Definition~\ref{def_Ft} and $\Sigma_{X=-Y}$ is a 2-chain of $M$ bounded by $L_{X=-Y}$ provided by Proposition~\ref{prop_linksinhomologyI}. The chain $C_4(X,Y)$ has the desired boundary. Note that $\langle [S], C_4(X,Y) \rangle = 1$. Moreover, $\langle [Z(\Sigma)], C_4(X,Y) \rangle = 0$ for any surface $\Sigma$ in $M$. Indeed, notice that
$$
\langle [Z(\Sigma)], C_4(X,Y) \rangle = \left\lbrace
\begin{aligned}
&\langle [Z(\Sigma)], [0,1]\times X(M) \rangle \mbox{ , pushing $Z(\Sigma)$ before $t$,}\\
&\langle [Z(\Sigma)], [0,1]\times Y(M) \rangle \mbox{ , pushing $Z(\Sigma)$ after $t$.}
\end{aligned}
\right.
$$
As a consequence, $\langle [Z(\Sigma)], C_4(X,Y) \rangle$ is independent of $X$ and $Y$. Let us prove that it is possible to construct a torsion combing $Z'$ that coincides with $X$ and $Y$ on $\partial M$ and such that
$$
\langle [Z(\Sigma)], [0,1]\times Z'(M) \rangle=0.
$$
Using the parallelization $\rho=(E_1^\rho,E_2^\rho,E_3^\rho)$ of $\partial M$ induced by $X$, define a homotopy
$$
\l Z : [0,1] \times \partial M \rightarrow [0,1]\times UM_{|\partial M}
$$
from $-Z_{|\partial M}$ to $Z_{|\partial M}$ along the unique geodesic arc passing through $E_3^\rho$. Since $\Sigma$ sits in $\mathring{M}$, we can get a collar $\l C\simeq [0,1]\times \partial M$ of $\partial M$ such that $\l C \cap \Sigma = \emptyset$ and $\lbrace 1 \rbrace \times \partial M = \partial M$. Finally set $Z'$ to coincide with $-Z$ on $M\setminus \mathring{\l C}$ and with the homotopy $\l Z$ on the collar. The combing $(Z',\rho)$ is a torsion combing. Indeed, $E_2^\rho$ can be extended as a nonvanishing section of ${Z'}^\perp_{|\l C}$ so that
$$
e_2^M(Z'^\perp,E_2^\rho)=e_2^{M\setminus \mathring{\l C}}(Z'^\perp,E_2^\rho)=e_2^M(-Z^\perp,E_2^\rho)=-e_2^M(Z^\perp,E_2^\rho).
$$
Finally, using the torsion combing $Z'$, we get $\langle [Z(\Sigma)], [0,1]\times Z'(M) \rangle = 0$. \\
To conclude the proof, assume that $C'_4(X,Y)$ is a 4-chain with same boundary as the chain $C_4(X,Y)$ we constructed, and let $C$ be a 2-cycle of $[0,1]\times UM$. The 2-cycle $C$ is homologous to a 2-cycle in $\lbrace 1 \rbrace \times UM$. Similarly, $(C'_4(X,Y) - C_4(X,Y))$ is homologous to a 4-cycle in $\lbrace 0 \rbrace \times UM$. Hence, $\langle C , \ C'_4(X,Y) - C_4(X,Y) \rangle = 0$.
\end{proof}
\begin{lemma} \label{bord}
Let $\tau$ and $\bar \tau$ be two pseudo-parallelizations of a compact oriented 3-manifold $M$ that coincide on $\partial M$. Let $C_4(\tau,\bar\tau;\pm e_1)$ denote 4-chains of $[0,1]\times UM$ as in Theorem~\ref{prop_varasint} for $v=\pm e_1$. If the 4-chains $C_4(\tau,\bar\tau;\pm e_1)$ are transverse to each other, then
$$
\begin{aligned}
\partial (C_4(\tau,\bar\tau;e_1)\cap C_4(\tau,\bar\tau;-e_1)) &= \sfrac{1}{4} \cdot \lbrace 1 \rbrace \times \left( \bar E_1^d(L_{\bar E_1^d=-\bar E_1^g}) - (-\bar E_1^d)(L_{\bar E_1^d=-\bar E_1^g}) \right) \\
&- \sfrac{1}{4} \cdot \lbrace 0 \rbrace \times \left( E_1^d(L_{E_1^d=-E_1^g}) - (-E_1^d)(L_{E_1^d=-E_1^g}) \right)
\end{aligned}
$$
where $E_1^d$ and $E_1^g$, resp. $\bar E_1^d$ and $\bar E_1^g$, are the Siamese sections of $\tau$, resp. $\bar\tau$.
\end{lemma}
\begin{proof}
Since $\tau$ and $\bar \tau$ coincide with a trivialization of $TM_{|\partial M}$ on $\partial M$, we have
$$
\begin{aligned}
\partial (C_4(\tau,\bar\tau;e_1)\cap C_4(\tau,\bar\tau;-e_1))
&= \lbrace 1 \rbrace \times \left( \bar\tau(M\times \lbrace e_1 \rbrace)\cap \bar\tau(M\times \lbrace -e_1 \rbrace) \right) \\
&- \lbrace 0 \rbrace \times \left( \tau(M\times \lbrace e_1 \rbrace)\cap \tau(M\times \lbrace -e_1 \rbrace) \right) \\
&=\sfrac{1}{4} \cdot \lbrace 1 \rbrace \times \left( \bar E_1^d(L_{\bar E_1^d=-\bar E_1^g}) + \bar E_1^g(L_{\bar E_1^g=-\bar E_1^d}) \right) \\
&- \sfrac{1}{4} \cdot \lbrace 0 \rbrace \times \left( E_1^d(L_{E_1^d=-E_1^g}) + E_1^g(L_{E_1^g=-E_1^d}) \right) .
\end{aligned}
$$
\iffalse
&= ( \lbrace 1 \rbrace \times UM ) \bigcap \left( C_4(\tau,\bar\tau;e_1)\cap C_4(\tau,\bar\tau;-e_1) \right) \\
&- (\lbrace 0 \rbrace \times UM ) \bigcap \left( C_4(\tau,\bar\tau;e_1)\cap C_4(\tau,\bar\tau;-e_1) \right) \\
&=\sfrac{1}{4} \cdot \lbrace 1 \rbrace \times \left( \bar E_1^d(L_{\bar E_1^d=-\bar E_1^g}) - (-\bar E_1^d)(L_{\bar E_1^d=-\bar E_1^g}) \right) \\
&- \sfrac{1}{4} \cdot \lbrace 0 \rbrace \times \left( E_1^d(L_{E_1^d=-E_1^g}) - (-E_1^d)(L_{E_1^d=-E_1^g}) \right).
\fi
\end{proof}
\begin{definition} \label{omega}
Let $\bar\tau=(N(\gamma); \tau_e, \tau_d, \tau_g)$ be a pseudo-parallelization of a compact oriented 3-manifold $M$, and let $E_1^d$ and $E_1^g$ denote its Siamese sections. Recall from Definition \ref{def_addinner} that the map
$$
\tau_d^{-1} \circ \tau_g : [a,b] \times \gamma \times [-1,1] \times \b R^3 \rightarrow [a,b] \times \gamma \times [-1,1] \times \b R^3
$$
is such that
$$
\begin{aligned}
&\forall t \in [a,b], \ u \in [-1,1], \ c \in \gamma, \ v \in \b R^3 : \\
&\tau_d^{-1} \circ \tau_g ((t,c,u),v) = \l T_\gamma^{-1} ((t,c,u),\l F(t,u)(v)),
\end{aligned}
$$
Hence, $L_{E_1^d=-E_1^g}$ consists in parallels of $\gamma$ of the form $\lbrace t \rbrace \times \gamma \times \lbrace u \rbrace$. For all component $L$ of $L_{E_1^d=-E_1^g}$, there exists a point $e_2^L$ in $\b S^1(e_2)$ such that $L \times \lbrace e_2^L \rbrace =\tau_d^{-1} \circ \tau_g (L \times \lbrace e_2 \rbrace)$. Choose a point $e_2^\Omega$ in $\b S^1(e_2)$ distinct from $e_2$ and from the points $e_2^L$. Finally, set
$$
\Omega(\bar\tau)=-\tau_d(L_{E_1^d=-E_1^g}\times[-e_1,e_1]_{e_2^\Omega})
$$
where $[-e_1,e_1]_{e_2^\Omega}$ is the geodesic arc from $-e_1$ to $e_1$ passing through $e_2^\Omega$. The 2-chain $\Omega(\bar\tau)$ can be seen as the projection of a homotopy from $-E_1^d$ to $E_1^d$ over $L_{E_1^d=-E_1^g}$. Note that
$$
\partial \Omega(\bar\tau)= (-E_1^d)(L_{E_1^d=-E_1^g})- E_1^d(L_{E_1^d=-E_1^g}).
$$
The choice of $e_2^\Omega$ ensures that $\Omega(\bar\tau) \cap \bar\tau(M\times \lbrace e_2 \rbrace)=\emptyset$. Note that $\Omega(\tau)=\emptyset$ when $\tau$ is a genuine parallelization.
\end{definition}
\begin{definition} \label{def_pgo}
Let $M$ be a compact oriented 3-manifold and let $\rho$ be an admissible trivialization of $TM_{|\partial M}$. Let $\tau$ and $\bar\tau$ be pseudo-parallelizations of a compact oriented 3-manifold $M$ which coincide with $\rho$ on $TM_{|\partial M}$ and let $C_4(\tau,\bar\tau;\pm e_1)$ denote 4-chains of $[0,1]\times UM$ as in Theorem~\ref{prop_varasint}. Set
$$
\go P(\tau,\bar\tau) = \lbrace 0 \rbrace \times \Omega(\tau) + 4 \cdot (C_4(\tau,\bar\tau;e_1) \cap C_4(\tau,\bar\tau;-e_1)) - \lbrace 1 \rbrace \times \Omega(\bar\tau).
$$
When $(X,\rho)$ is a torsion combing of $M$, let $C_4^+(X,\bar\tau)$ and $C_4^-(X,\bar\tau)$ be 4-chains of $[0,1]\times UM$ as in Lemma~\ref{lem_transfert} and set
$$
\begin{aligned}
\go P(X,\bar\tau) = 4\cdot(C_4^+(X,\bar\tau)\cap C_4^-(X,\bar\tau)) - \lbrace 1 \rbrace \times \Omega(\bar\tau), \\
\go P(\bar\tau,X) = \lbrace 0 \rbrace \times \Omega(\bar\tau) + 4\cdot(C_4^+(\bar\tau,X)\cap C_4^-(\bar\tau,X)).
\end{aligned}
$$
\end{definition}
According to Lemma~\ref{bord} and Definition~\ref{omega}, the 4-chains $\go P(\lambda,\mu)$ of Definition~\ref{def_pgo} above are cycles. In the remaining of this section, we prove that their classes read $p_1(\lambda,\mu)[S]$ in $H_2([0,1]\times UM;\b Q)/H_T^\rho(M)$.
\begin{proposition} \label{prop_ppara}
Let $M$ be a compact oriented 3-manifold, let $\rho$ be an admissible trivialization of $TM_{|\partial M}$ and let $\tau$ and $\bar \tau$ be two pseudo-parallelizations of $M$ that coincide with $\rho$ on $\partial M$. Under the assumptions of Definition~\ref{def_pgo}, the class of $\go P(\tau,\bar\tau)$ in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$ equals $p_1(\tau,\bar\tau)[S]$ where $[S]$ is the homology class of the fiber of $UM$ in $H_2([0,1]\times UM ; \b Q)$.
\end{proposition}
\begin{proof}
The class of $\go P(\tau,\bar\tau)$ in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$ is
$$
\left[ \go P(\tau,\bar\tau) \right] = \langle \go P(\tau,\bar\tau) , C_4(X,Y) \rangle \cdot [S]
$$
where $(X,\rho)$ and $(Y,\rho)$ are $\partial$-compatible torsion combings of $M$ and where $C_4(X,Y)$ is any 4-chain of $[0,1]\times UM$ as in Lemma~\ref{lem_evaluationII}. Let us construct a specific $C_4(X,Y)$ as follows. Let $C_4(\tau,\bar\tau;e_2)$ be as in Theorem~\ref{prop_varasint} where $e_2=(0,1,0)$. Since, $\partial C_4(\tau,\bar\tau;e_1)$ and $\partial C_4(\tau,\bar\tau;e_2)$ are homologous, it is possible to reparameterize and to stack the 4-chains $C_4^+(X,\tau)$, $C_4(\tau,\bar\tau;e_2)$ and $C_4^+(\bar\tau,Y)$ where the chains $C_4^+(X,\tau)$ and $C_4^+(\bar\tau,Y)$ are as in Lemma~\ref{lem_transfert}. It follows that, in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$,
$$
\begin{aligned}
\left[ \go P(\tau,\bar\tau) \right] &= \langle \go P(\tau,\bar\tau) , C_4(X,Y) \rangle [S] \\
&= \langle \go P(\tau,\bar\tau) , C_4(\tau,\bar\tau;e_2) \rangle [S] \\
&= 4 \cdot \langle C_4(\tau,\bar\tau;e_1) \cap C_4(\tau,\bar\tau;-e_1) ,C_4(\tau,\bar\tau;e_2) \rangle [S] \\
&+ \langle \lbrace 0 \rbrace \times \Omega(\tau)- \lbrace 1 \rbrace \times \Omega(\bar\tau),C_4(\tau,\bar\tau;e_2) \rangle [S].
\end{aligned}
$$
Now, note that $ \langle \lbrace 0 \rbrace \times \Omega(\tau)- \lbrace 1 \rbrace \times \Omega(\bar\tau),C_4(\tau,\bar\tau;e_2) \rangle=0$ since
$$
\Omega(\tau) \cap \tau(M\times \lbrace e_2 \rbrace) = \emptyset \mbox{ \ and \ } \Omega(\bar \tau) \cap \bar \tau(M\times \lbrace e_2 \rbrace) = \emptyset.
$$
Hence,
$$
\left[ \go P(\tau,\bar\tau) \right]= 4 \cdot \langle C_4(\tau,\bar\tau;e_1),C_4(\tau,\bar\tau;-e_1),C_4(\tau,\bar\tau;e_2) \rangle [S] = p_1(\tau,\bar\tau) [S].
$$
\end{proof}
\subsection[Pontrjagin numbers for combings of compact 3-manifolds]{Pontrjagin numbers for combings of compact 3-manifolds \\ Proof of Theorem~\ref{thm_defp1Xb}}
\begin{lemma} \label{cor_reformcombingsb}
Let $(X,\rho)$ be a torsion combing of a compact oriented 3-manifold $M$. Let $\bar\tau$ be a pseudo-parallelization of $M$ compatible with $X$. Let $\go P(\bar\tau,X)$ be as in Definition~\ref{def_pgo}. The class $[\go P(\bar\tau,X)]$ in $H_2([0,1]\times UM ; \b Q) / H_T^\rho(M)$ only depends on $\bar\tau$ and on the homotopy class of $X$. It will be denoted by $\tilde p_1(\bar\tau,[X])[S]$.
\end{lemma}
\begin{proof}
Let $\tau$ be another pseudo-parallelization of $M$ which is compatible with $X$. Let $C_4^+(X,\tau)$ and $C_4^-(X,\tau)$ be fixed choices of 4-chains of $[0,1]\times UM$ as in Lemma~\ref{lem_transfert}. Using these 4-chains, construct the cycle $\go P(X,\tau)$ as in Definition~\ref{def_pgo}. Then, in the space $H_2([0,1]\times UM ; \b Q) / H_T^\rho(M)$, we have
$$
\begin{aligned}
[\go P(\bar\tau,X)] + [\go P(X,\tau)] &= [\go P(\bar\tau,X) + \go P(X,\tau)] \\
&= [ \lbrace 0 \rbrace \times \Omega(\bar\tau) + 4\cdot(C_4^+(\bar\tau,X)\cap C_4^-(\bar\tau,X)) \\
& \hspace{7mm} + 4\cdot(C_4^+(X,\tau)\cap C_4^-(X,\tau)) - \lbrace 1 \rbrace \times \Omega(\tau) ].
\end{aligned}
$$
By reparameterizing and stacking $C_4^+(\bar\tau,X)$ and $C_4^+(X,\tau)$, resp. $C_4^-(\bar\tau,X)$ and $C_4^-(X,\tau)$, we get a 4-chain $C_4(\bar\tau,\tau,e_1)$, resp. $C_4(\bar\tau,\tau,-e_1)$, as in Lemma~\ref{lem_C4pme1}. It follows that
$$
\begin{aligned}
[\go P(\bar\tau,X)] + [\go P(X,\tau)] & \hspace{-1mm}=\hspace{-1mm} [ \lbrace 0 \rbrace \times \Omega(\bar\tau) + 4\cdot(C_4(\bar\tau,\tau,e_1) \cap C_4(\bar\tau,\tau,e_1)) - \lbrace 1 \rbrace \times \Omega(\tau) ] \\
&\hspace{-1mm}=\hspace{-1mm} [\go P(\bar\tau, \tau)]
\end{aligned}
$$
or, equivalently, $[\go P(\bar\tau,X)] = [\go P(\bar\tau, \tau)] - [\go P(X,\tau)]$. This proves the statement since $\go P(X,\tau)$ is independent of the choices made for $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$, and since, according to Proposition~\ref{prop_ppara}, the class $[\go P(\bar\tau, \tau)]$ is independent of the choices for $C_4(\bar\tau,\tau,e_1)$ and $C_4(\bar\tau,\tau,-e_1)$.
\end{proof}
\begin{proposition} \label{cor_reformcombings}
If $\bar\tau$ is a pseudo-parallelization of a closed oriented 3-manifold $M$ and if $X$ is a torsion combing of $M$ compatible with $\bar\tau$, then
$$
\tilde p_1(\bar\tau,[X]) = p_1([X])-p_1(\bar\tau).
$$
\end{proposition}
\begin{proof}
According to Lemma~\ref{cor_reformcombingsb}, $\tilde p_1(\bar\tau,[X])$ is independent of the choices for $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$. Let us construct convenient 4-chains $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$. Let $\tau$ be a ge\-nuine parallelization of $M$. Thanks to Theorem~\ref{prop_varasint}, there exist two 4-chains of $[0,1]\times UM$, $C_4(\bar\tau,\tau;e_1)$ and $C_4(\bar\tau,\tau;-e_1)$, such that
$$
\begin{aligned}
\partial C_4(\bar\tau,\tau;e_1) &= \lbrace 1 \rbrace \times \tau(M\times \lbrace e_1 \rbrace) - \lbrace 0 \rbrace \times \bar \tau(M\times \lbrace e_1 \rbrace), \\
\partial C_4(\bar\tau,\tau;-e_1) &= \lbrace 1 \rbrace \times \tau(M\times \lbrace -e_1 \rbrace) - \lbrace 0 \rbrace \times \bar \tau(M\times \lbrace -e_1 \rbrace).
\end{aligned}
$$
Furthermore, as in Remark~\ref{rmkpcase}, construct two 4-chains $C_4^+(\tau,X)$ and $C_4^-(\tau,X)$ as
$$
\begin{aligned}
C_4^+(\tau, X) &= \bar F_{t_1}(E_1^\tau, X) - \lbrace t_1 \rbrace \times UM_{|\Sigma_{E_1^\tau=-X}}\\
C_4^-(\tau, X) &= \bar F_{t_2}(-E_1^\tau, -X) - \lbrace t_2 \rbrace \times UM_{|\Sigma_{-E_1^\tau=X}}
\end{aligned}
$$
where $E_1^\tau$ stands for the first vector of the parallelization $\tau$, where $t_1, t_2\in \ ]0,1[$, and where $\Sigma_{E_1^\tau=-X}$ and $\Sigma_{-E_1^\tau=X}$ are 2-chains with boundaries $L_{E_1^\tau=-X}$ and $L_{-E_1^\tau=X}$, respectively. Eventually, define $C_4^+(\bar\tau,X)$, resp. $C_4^-(\bar\tau,X)$, by reparameterizing and stacking the chains $C_4(\bar\tau,\tau;e_1)$ and $ C_4^+(\tau, X)$, resp. $C_4(\bar\tau,\tau;-e_1)$ and $ C_4^-(\tau, X)$. \\
Let us finally compute $[\lbrace 0 \rbrace \times \Omega(\bar\tau)+4\cdot(C_4^+(\bar\tau,X) \cap C_4^-(\bar\tau,X))]$. By construction, we have :
$$
\begin{aligned}
&[\lbrace 0 \rbrace \times \Omega(\bar\tau)+4 \cdot (C_4^+(\bar\tau,X) \cap C_4^-(\bar\tau,X))] \\
&\hspace{-1mm}=\hspace{-1mm} [\lbrace 0 \rbrace \times \Omega(\bar\tau)+4 \cdot (C_4(\bar\tau,\tau;e_1) \cap C_4(\bar\tau,\tau;-e_1))] + 4 [C_4^+(\tau, X) \cap C_4^-(\tau, X)],
\end{aligned}
$$
so that, using Proposition~\ref{prop_ppara},
$$
\begin{aligned}
&[\lbrace 0 \rbrace \times \Omega(\bar\tau)+4 \cdot (C_4^+(\bar\tau,X) \cap C_4^-(\bar\tau,X))] \\
&= (p_1(\tau)-p_1(\bar\tau))[S] + 4 \cdot [C_4^+(\tau, X) \cap C_4^-(\tau, X)].
\end{aligned}
$$
Now, using Definition~\ref{def_Ft},
$$
\begin{aligned}
C_4^+(\tau, X) = \ & \left[ 0 , t_1 \right] \times E_1^\tau(M)
+ \lbrace t_1 \rbrace \times \bar F(E_1^\tau,X) \\
+ \ & \left[ t_1, 1 \right] \times X(M) - \lbrace t_1 \rbrace \times UM_{|\Sigma_{E_1^\tau=-X}}\\
C_4^-(\tau, X) = \ & \left[ 0 , t_2 \right] \times (-E_1^\tau)(M)
+ \lbrace t_2 \rbrace \times \bar F(-E_1^\tau,-X) \\
+ \ & \left[ t_2, 1 \right] \times (-X)(M) - \lbrace t_2 \rbrace \times UM_{|\Sigma_{-E_1^\tau=X}}
\end{aligned}
$$
so that, assuming $t_1 < t_2$ without loss of generality,
$$
\begin{aligned}
C_4^+(\tau, X) \cap C_4^-(\tau, X)
= - \ & \lbrace t_1 \rbrace \times (-E_1^\tau)\left( \Sigma_{E_1^\tau=-X} \right) \\
+ \ & [t_1 , t_2] \times (-E_1^\tau)\left( L_{-E_1^\tau = X} \right) \\
- \ & \lbrace t_2 \rbrace \times X(\Sigma_{-E_1^\tau=X}).
\end{aligned}
$$
It follows that, using Theorem~\ref{thm_defp1X} and Lemma~\ref{lem_evaluationII} with $C_4(E_1^\tau,E_1^\tau)=[0,1] \times E_1^\tau(M)$,
$$
\begin{aligned}
4 \cdot [C_4^+(\tau, X) \cap C_4^-(\tau, X)] & \hspace{-1mm}=\hspace{-1mm} 4 \cdot \langle C_4^+(\tau, X) , C_4^-(\tau, X) , [0,1] \times E_1^\tau(M) \rangle [S] \\
&\italicegal4 \cdot lk(L_{E_1^\tau = X},L_{E_1^\tau = - X}) [S] \\
&\hspace{-1mm}=\hspace{-1mm}(p_1([X]) - p_1(\tau))[S] \\
\end{aligned}
$$
in $H_2([0,1]\times UM; \b Q)/H_T(M)$, and, eventually,
$$
\begin{aligned}
[\lbrace 0 \rbrace \times \Omega(\bar\tau)&+4 \cdot (C_4^+(\bar\tau,X) \cap C_4^-(\bar\tau,X))] \\
&= (p_1(\tau)-p_1(\bar\tau))[S] + (p_1([X]) - p_1(\tau))[S] \\
&=(p_1([X]) - p_1(\bar\tau))[S].
\end{aligned}
$$
\end{proof}
\begin{lemma} \label{fond}
If $(X,\rho)$ is a torsion combing of a compact oriented 3-manifold $M$ and if \linebreak $\bar\tau=(N(\gamma); \tau_e,\tau_d,\tau_g)$ is a pseudo-parallelization of $M$ compatible with $X$, then
$$
\begin{aligned}
\tilde p_1(\bar\tau, [X]) &= lk_M\left(L_{E_1^d=X} + L_{E_1^g=X} \ , \ L_{E_1^d=-X} + L_{E_1^g=-X} \right) \\
&- lk_{\b S^2}\left(e_1-(-e_1), \ P_{\b S^2} \circ \tau_d^{-1} \circ X(L_{E_1^d=-E_1^g})\right)
\end{aligned}
$$
where $E_1^d$ and $E_1^g$ denote the Siamese sections of $\bar\tau$.
\end{lemma}
\begin{proof}
We just have to evaluate the class of the 4-cycle
$$
\go P(\bar\tau,X) = \lbrace 0 \rbrace \times \Omega(\bar\tau)+4\cdot(C_4^+(\bar\tau,X)\cap C_4^-(\bar\tau,X))
$$
in $H_2([0,1]\times UM ; \b Q) / H_T^\rho(M)$ for convenient 4-chains $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$ with the prescribed boundaries. Let $t_1$ and $t_2$ in $]0,1[$, with $t_1 > t_2$, and set
$$
\begin{aligned}
C_4^+(\bar\tau,X) &= \sfrac{1}{2} \cdot \left( \bar F_{t_1}(E_1^d,X)+\bar F_{t_1}(E_1^g,X) - \lbrace t_1 \rbrace \times UM_{|\Sigma(e_1)} \right)\\
C_4^-(\bar\tau,X) &= \sfrac{1}{2} \cdot \left( \bar F_{t_2}(-E_1^d,-X)+\bar F_{t_2}(-E_1^g,-X) - \lbrace t_2 \rbrace \times UM_{|\Sigma(-e_1)} \right)
\end{aligned}
$$
where the chains $\bar F_t$ are as in Definition~\ref{def_Ft} and where, using Lemma~\ref{nulexcep}, $\Sigma(e_1)$ and $\Sigma(-e_1)$ are 2-chains of $M$ so that
$$
\partial \Sigma(\pm e_1) = \pm(L_{E_1^d=-X} + L_{E_1^g=-X}).
$$
These 4-chains do have the expected boundaries. Let us now describe $C_4^+(\bar\tau,X) \cap C_4^-(\bar\tau,X)$ :
\begin{enumerate}[\textbullet]
\item on $[0,t_2[$ : The intersection between $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$ is
$$
\sfrac{1}{4} \cdot [0,t_2[ \ \times \ E_1^d(L_{E_1^d=-E_1^g}) + \sfrac{1}{4} \cdot [0,t_2[ \ \times \ E_1^g(L_{E_1^g=-E_1^d}).
$$
\item on $]t_2 , t_1[$ : The intersection between $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$ is
$$
\sfrac{1}{2} \ \cdot \ ]t_2,t_1[ \ \times \ (-X)(L_{E_1^d=-X}) + \sfrac{1}{2} \ \cdot \ ]t_2,t_1[ \ \times \ (-X)(L_{E_1^g=-X}).
$$
\item on $]t_1,1]$ : There is no intersection between $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$ since they consist in $]t_1,1]\times X(M)$ and $]t_1,1]\times (-X)(M)$.
\item at $t_2$ : The intersection between $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$ is
$$
\begin{aligned}
&\sfrac{1}{2} \cdot \lbrace t_2 \rbrace \times E_1^d(M) \cap \lbrace t_2 \rbrace \times \bar F(-E_1^g,-X) \\
+ &\sfrac{1}{2} \cdot \lbrace t_2 \rbrace \times E_1^g(M) \cap \lbrace t_2 \rbrace \times \bar F(-E_1^d,-X) \\
- & \sfrac{1}{4} \cdot \lbrace t_2 \rbrace \times E_1^d(\Sigma(-e_1)) - \sfrac{1}{4} \cdot \lbrace t_2 \rbrace \times E_1^g(\Sigma(-e_1))
\end{aligned}
$$
\item at $t_1$ : The intersection between $C_4^+(\bar\tau,X)$ and $C_4^-(\bar\tau,X)$ is
$$
\begin{aligned}
&\sfrac{1}{2}\cdot \lbrace t_1 \rbrace \times \bar F(E_1^d,X) \cap \lbrace t_1 \rbrace \times (-X)(M) \\
+ &\sfrac{1}{2}\cdot \lbrace t_1 \rbrace \times \bar F(E_1^g,X) \cap \lbrace t_1 \rbrace \times (-X)(M) \\
- & \sfrac{1}{2} \cdot \lbrace t_1 \rbrace \times (-X)(\Sigma(e_1)).
\end{aligned}
$$
\end{enumerate}
It follows that :
$$
\begin{aligned}
&\langle \sfrac{1}{4} \cdot \lbrace 0 \rbrace \times \Omega(\bar\tau)+ C_4^+(\bar\tau,X) \cap C_4^-(\bar\tau,X) , [0,1] \times X(M) \rangle_{[0,1]\times UM}\\
&= \sfrac{1}{4} \cdot \langle [0,t_2[ \ \times \ E_1^d(L_{E_1^d=-E_1^g}) + [0,t_2[ \ \times \ E_1^g(L_{E_1^g=-E_1^d}) , [0,1]\times X(M) \rangle_{[0,1]\times UM} \\
&+ \sfrac{1}{2} \cdot \langle \lbrace t_2 \rbrace \times E_1^d(M) \cap \lbrace t_2 \rbrace \times \bar F(-E_1^g,-X) , [0,1] \times X(M) \rangle_{[0,1]\times UM} \\
&+\sfrac{1}{2} \cdot \langle \lbrace t_2 \rbrace \times E_1^g(M) \cap \lbrace t_2 \rbrace \times \bar F(-E_1^d,-X) , [0,1] \times X(M) \rangle_{[0,1]\times UM} \\
&- \sfrac{1}{4} \cdot \langle \lbrace t_2 \rbrace \times E_1^d(\Sigma(-e_1)) + \lbrace t_2 \rbrace \times E_1^g(\Sigma(-e_1)) , [0,1] \times X(M) \rangle_{[0,1]\times UM} \\
&+ \sfrac{1}{4} \cdot \langle \Omega(\bar\tau) , X(M) \rangle_{UM}.
\end{aligned}
$$
Since $L_{E_1^g=X} \cap L_{E_1^d=-X}$ and $L_{E_1^g=-X} \cap L_{E_1^d=X}$ are empty :
$$
\langle [0,t_2[ \ \times \ E_1^d(L_{E_1^d=-E_1^g}) + [0,t_2[ \ \times \ E_1^g(L_{E_1^g=-E_1^d}) , [0,1]\times X(M) \rangle_{[0,1]\times UM} = 0.
$$
Furthermore, note that if $(m,v)$ is an intersection point of
$$E_1^d(M)\cap \bar F(-E_1^g,-X) \cap X(M)
$$
then, in particular, $v=E_1^d(m)=X(m)$ so that $-E_1^g(m)$ and $-X(m)$ are not antipodal since $L_{E_1^d=X} \cap L_{E_1^g=-X}=\emptyset$. It follows that $v=E_1^d(m)=X(m)$ should also sit on the shortest geodesic arc from $-E_1^g(m)$ to $-X(m)$. Since such a configuration is impossible, this triple intersection is empty, thus
$$
\langle \lbrace t_2 \rbrace \times E_1^d(M) \cap \lbrace t_2 \rbrace \times \bar F(-E_1^g,-X) , [0,1] \times X(M) \rangle_{[0,1]\times UM} = 0.
$$
Similarly,
$$
\langle \lbrace t_2 \rbrace \times E_1^g(M) \cap \lbrace t_2 \rbrace \times \bar F(-E_1^d,-X) , [0,1] \times X(M) \rangle_{[0,1]\times UM} = 0.
$$
Now, we have
$$
\begin{aligned}
&\langle \lbrace t_2 \rbrace \times E_1^d(\Sigma(-e_1)) + \lbrace t_2 \rbrace \times E_1^g(\Sigma(-e_1)) , [0,1] \times X(M) \rangle \\
&= - lk_M(L_{E_1^d=X} + L_{E_1^g=X}, \ L_{E_1^d=-X} + L_{E_1^g=-X}).
\end{aligned}
$$
Furthermore, recall Definition~\ref{omega}
$$
\begin{aligned}
\langle \Omega(\bar\tau) , X(M) \rangle_{UM} &= \langle \tau_d^{-1} (\Omega(\bar\tau)) , \tau_d^{-1} \circ X(L_{E_1^d=-E_1^g}) \rangle_{L_{E_1^d=-E_1^g} \times \b S^2} \\
&= - \langle L_{E_1^d=-E_1^g} \hspace{-0.5mm} \times \hspace{-0.5mm} [-e_1,e_1]_{e_2^\Omega} \ , \ \tau_d^{-1} \circ X(L_{E_1^d=-E_1^g}) \rangle_{L_{E_1^d=-E_1^g} \times \b S^2}
\end{aligned}
$$
where $[-e_1,e_1]_{e_2^\Omega}$ is the geodesic arc of $\b S^2$ from $-e_1$ to $e_1$ passing through $e_2^\Omega$. Now, $L_{E_1^d=-E_1^g} \times \b S^2$ is oriented and an intersection
$$
(m,v)\hspace{-0.5mm}\in\hspace{-0.5mm}L_{E_1^d=-E_1^g} \hspace{-0.5mm} \times \hspace{-0.5mm} [-e_1,e_1]_{e_2^\Omega} \cap \tau_d^{-1} \circ X(L_{E_1^d=-E_1^g})
$$
is positive when
$$
T_{(m,v)}(L_{E_1^d=-E_1^g} \times [-e_1,e_1]_{e_2^\Omega})\oplus T_{(m,v)}(\tau_d^{-1}\circ X(L_{E_1^d=-E_1^g}))=T_{(m,v)}(L_{E_1^d=-E_1^g}\times\b S^2)
$$
as an oriented sum, which is equivalent to
$$
T_{v}([-e_1,e_1]_{e_2^\Omega})\oplus T_{v}(P_{\b S^2}\circ\tau_d^{-1}\circ X(L_{E_1^d=-E_1^g}))=T_{v}(\b S^2)
$$
as an oriented sum, where $P_{\b S^2}$ is the standard projection from $M\times \b S^2$ to $\b S^2$. See Figure \ref{orientation}.
\begin{center}
\definecolor{ccqqtt}{rgb}{0.8,0,0.2}
\definecolor{zzttqq}{rgb}{0.6,0.2,0}
\definecolor{ccqqqq}{rgb}{0.8,0,0}
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=0.5cm,scale=0.9]
\clip(9,-2.5) rectangle (23,10);
\fill[line width=0pt,color=zzttqq,fill=zzttqq,fill opacity=0.1] (13,4) -- (14.5,4) -- (14.5,-1) -- (13,-1) -- cycle;
\fill[line width=0pt,dotted,color=zzttqq,fill=zzttqq,fill opacity=0.1] (16,4) -- (20,4) -- (20,-1) -- (16,-1) -- cycle;
\fill[line width=0pt,color=zzttqq,fill=zzttqq,fill opacity=0.1] (21.5,4) -- (23,4) -- (23,-1) -- (21.5,-1) -- cycle;
\draw [line width=1.4pt] (16,4)-- (20,4);
\draw [line width=1.4pt] (20,-1)-- (20,8);
\draw [line width=1.4pt] (20,8)-- (22,9);
\draw [->] (16,4) -- (20,4);
\draw [->,line width=1.4pt] (21.5,4) -- (23,4);
\draw [line width=1.4pt] (14.5,-1)-- (14.5,8);
\draw [line width=1.4pt] (14.5,8)-- (16.5,9);
\draw [->,line width=1.4pt,color=ccqqqq] (21.5,6) -- (23,6);
\draw [->,line width=1.4pt,color=ccqqqq] (13,6) -- (14.5,6);
\draw [->,line width=1.4pt] (13,4) -- (14.5,4);
\draw [->] (16,-1) -- (16,4);
\draw [->] (21.5,-1) -- (21.5,4);
\draw[line width=1.4pt,color=ccqqqq] (19.12,4.3) -- (19.11,4.24) -- (19.1,4.18) -- (19.1,4.12) -- (19.09,4.05) -- (19.08,3.99) -- (19.07,3.93) -- (19.07,3.87) -- (19.06,3.81) -- (19.05,3.74) -- (19.05,3.68) -- (19.04,3.62) -- (19.04,3.56) -- (19.03,3.49) -- (19.02,3.43) -- (19.02,3.37) -- (19.01,3.31) -- (19.01,3.25) -- (19,3.18) -- (18.99,3.12) -- (18.99,3.06) -- (18.98,3) -- (18.98,2.94) -- (18.97,2.88) -- (18.96,2.82) -- (18.96,2.76) -- (18.95,2.7) -- (18.95,2.65) -- (18.94,2.59) -- (18.93,2.53) -- (18.93,2.48) -- (18.92,2.42) -- (18.91,2.37) -- (18.9,2.31) -- (18.9,2.26) -- (18.89,2.21) -- (18.88,2.16) -- (18.87,2.11) -- (18.86,2.06) -- (18.85,2.01) -- (18.85,1.97) -- (18.84,1.92) -- (18.83,1.88) -- (18.82,1.83) -- (18.81,1.79) -- (18.79,1.75) -- (18.78,1.71) -- (18.77,1.67) -- (18.76,1.64) -- (18.75,1.6) -- (18.73,1.57) -- (18.72,1.54) -- (18.71,1.51) -- (18.69,1.48) -- (18.68,1.45) -- (18.66,1.42) -- (18.65,1.4) -- (18.63,1.38) -- (18.61,1.36) -- (18.59,1.34) -- (18.58,1.32) -- (18.56,1.31) -- (18.54,1.29) -- (18.52,1.28) -- (18.5,1.27) -- (18.48,1.27) -- (18.45,1.26) -- (18.43,1.26) -- (18.41,1.26) -- (18.38,1.26) -- (18.36,1.27) -- (18.33,1.27) -- (18.31,1.28) -- (18.28,1.29) -- (18.25,1.3) -- (18.22,1.32) -- (18.19,1.34) -- (18.16,1.36) -- (18.13,1.38) -- (18.1,1.41) -- (18.07,1.44) -- (18.03,1.47) -- (18.02,1.48) -- (18.01,1.49) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5)(20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (20,6) -- (19.99,6) -- (19.99,6) -- (19.98,6) -- (19.95,6) -- (19.92,5.99) -- (19.89,5.99) -- (19.86,5.98) -- (19.82,5.97) -- (19.79,5.96) -- (19.77,5.94) -- (19.74,5.93) -- (19.71,5.91) -- (19.68,5.89) -- (19.66,5.87) -- (19.63,5.84) -- (19.61,5.82) -- (19.58,5.79) -- (19.56,5.76) -- (19.54,5.73) -- (19.52,5.7) -- (19.49,5.67) -- (19.47,5.64) -- (19.45,5.6) -- (19.44,5.56) -- (19.42,5.52) -- (19.4,5.48) -- (19.38,5.44) -- (19.36,5.4) -- (19.35,5.36) -- (19.33,5.31) -- (19.32,5.26) -- (19.3,5.22) -- (19.29,5.17) -- (19.27,5.12) -- (19.26,5.07) -- (19.25,5.02) -- (19.24,4.97) -- (19.22,4.91) -- (19.21,4.86) -- (19.2,4.8) -- (19.19,4.75) -- (19.18,4.69) -- (19.17,4.63) -- (19.16,4.58) -- (19.15,4.52) -- (19.14,4.46) -- (19.13,4.4) -- (19.12,4.34) -- (19.12,4.3);
\draw[line width=1.4pt,dash pattern=on 2pt off 4pt,color=ccqqtt] (17.44,2.52) -- (17.44,2.54) -- (17.43,2.56) -- (17.42,2.58) -- (17.41,2.6) -- (17.4,2.63) -- (17.4,2.65) -- (17.39,2.67) -- (17.38,2.69) -- (17.37,2.72) -- (17.37,2.74) -- (17.36,2.76) -- (17.35,2.78) -- (17.34,2.81) -- (17.33,2.83) -- (17.33,2.85) -- (17.32,2.88) -- (17.31,2.9) -- (17.3,2.92) -- (17.3,2.95) -- (17.29,2.97) -- (17.28,3) -- (17.27,3.02) -- (17.26,3.05) -- (17.26,3.07) -- (17.25,3.1) -- (17.24,3.12) -- (17.23,3.15) -- (17.23,3.18) -- (17.22,3.2) -- (17.21,3.23) -- (17.2,3.25) -- (17.19,3.28) -- (17.19,3.31) -- (17.18,3.33) -- (17.17,3.36) -- (17.16,3.39) -- (17.15,3.42) -- (17.15,3.44) -- (17.14,3.47) -- (17.13,3.5) -- (17.12,3.53) -- (17.12,3.56) -- (17.11,3.59) -- (17.1,3.61) -- (17.09,3.64) -- (17.08,3.67) -- (17.08,3.7) -- (17.07,3.73) -- (17.06,3.76) -- (17.05,3.79) -- (17.05,3.82) -- (17.04,3.85) -- (17.03,3.88) -- (17.02,3.91) -- (17.01,3.94) -- (17.01,3.97) -- (17,3.99) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4)(17.49,2.04)(18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (18,1.5) -- (17.99,1.51) -- (17.99,1.51) -- (17.98,1.52) -- (17.97,1.53) -- (17.97,1.54) -- (17.96,1.54) -- (17.95,1.55) -- (17.94,1.56) -- (17.93,1.57) -- (17.93,1.58) -- (17.92,1.59) -- (17.91,1.6) -- (17.9,1.61) -- (17.9,1.62) -- (17.89,1.63) -- (17.88,1.64) -- (17.87,1.65) -- (17.86,1.66) -- (17.86,1.67) -- (17.85,1.69) -- (17.84,1.7) -- (17.83,1.71) -- (17.83,1.72) -- (17.82,1.73) -- (17.81,1.74) -- (17.8,1.76) -- (17.79,1.77) -- (17.79,1.78) -- (17.78,1.79) -- (17.77,1.81) -- (17.76,1.82) -- (17.76,1.83) -- (17.75,1.85) -- (17.74,1.86) -- (17.73,1.88) -- (17.72,1.89) -- (17.72,1.9) -- (17.71,1.92) -- (17.7,1.93) -- (17.69,1.95) -- (17.68,1.96) -- (17.68,1.98) -- (17.67,1.99) -- (17.66,2.01) -- (17.65,2.03) -- (17.65,2.04) -- (17.64,2.06) -- (17.63,2.08) -- (17.62,2.09) -- (17.61,2.11) -- (17.61,2.13) -- (17.6,2.14) -- (17.59,2.16) -- (17.58,2.18) -- (17.58,2.19) -- (17.57,2.21) -- (17.56,2.23) -- (17.55,2.25) -- (17.54,2.27) -- (17.54,2.29) -- (17.53,2.3) -- (17.52,2.32) -- (17.51,2.34) -- (17.51,2.36) -- (17.5,2.38) -- (17.49,2.4) -- (17.48,2.42) -- (17.47,2.44) -- (17.47,2.46) -- (17.46,2.48) -- (17.45,2.5) -- (17.44,2.52) -- (17.44,2.52);
\draw[line width=1.4pt,color=ccqqtt] (16.44,5.61) -- (16.44,5.62) -- (16.43,5.63) -- (16.42,5.65) -- (16.41,5.66) -- (16.4,5.67) -- (16.4,5.68) -- (16.39,5.7) -- (16.38,5.71) -- (16.37,5.72) -- (16.37,5.73) -- (16.36,5.74) -- (16.35,5.75) -- (16.34,5.77) -- (16.33,5.78) -- (16.33,5.79) -- (16.32,5.8) -- (16.31,5.81) -- (16.3,5.82) -- (16.3,5.83) -- (16.29,5.83) -- (16.28,5.84) -- (16.27,5.85) -- (16.26,5.86) -- (16.26,5.87) -- (16.25,5.88) -- (16.24,5.88) -- (16.23,5.89) -- (16.23,5.9) -- (16.22,5.91) -- (16.21,5.91) -- (16.2,5.92) -- (16.19,5.92) -- (16.19,5.93) -- (16.18,5.94) -- (16.17,5.94) -- (16.16,5.95) -- (16.15,5.95) -- (16.15,5.96) -- (16.14,5.96) -- (16.13,5.97) -- (16.12,5.97) -- (16.12,5.97) -- (16.11,5.98) -- (16.1,5.98) -- (16.09,5.98) -- (16.08,5.99) -- (16.08,5.99) -- (16.07,5.99) -- (16.06,5.99) -- (16.05,5.99) -- (16.05,6) -- (16.04,6) -- (16.03,6) -- (16.02,6) -- (16.01,6) -- (16.01,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6) -- (16,6)(17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4) -- (17,4.01) -- (17,4.01) -- (16.99,4.02) -- (16.99,4.04) -- (16.98,4.07) -- (16.97,4.1) -- (16.97,4.13) -- (16.96,4.16) -- (16.95,4.19) -- (16.94,4.22) -- (16.93,4.25) -- (16.93,4.28) -- (16.92,4.31) -- (16.91,4.34) -- (16.9,4.37) -- (16.9,4.4) -- (16.89,4.42) -- (16.88,4.45) -- (16.87,4.48) -- (16.86,4.51) -- (16.86,4.53) -- (16.85,4.56) -- (16.84,4.59) -- (16.83,4.61) -- (16.83,4.64) -- (16.82,4.66) -- (16.81,4.69) -- (16.8,4.71) -- (16.79,4.74) -- (16.79,4.76) -- (16.78,4.79) -- (16.77,4.81) -- (16.76,4.84) -- (16.76,4.86) -- (16.75,4.88) -- (16.74,4.91) -- (16.73,4.93) -- (16.72,4.95) -- (16.72,4.97) -- (16.71,5) -- (16.7,5.02) -- (16.69,5.04) -- (16.68,5.06) -- (16.68,5.08) -- (16.67,5.1) -- (16.66,5.13) -- (16.65,5.15) -- (16.65,5.17) -- (16.64,5.19) -- (16.63,5.21) -- (16.62,5.23) -- (16.61,5.24) -- (16.61,5.26) -- (16.6,5.28) -- (16.59,5.3) -- (16.58,5.32) -- (16.58,5.34) -- (16.57,5.36) -- (16.56,5.37) -- (16.55,5.39) -- (16.54,5.41) -- (16.54,5.42) -- (16.53,5.44) -- (16.52,5.46) -- (16.51,5.47) -- (16.51,5.49) -- (16.5,5.51) -- (16.49,5.52) -- (16.48,5.54) -- (16.47,5.55) -- (16.47,5.57) -- (16.46,5.58) -- (16.45,5.59) -- (16.44,5.61);
\draw (15,10) node[anchor=north west] {$\b S^2$};
\draw (20.5,10) node[anchor=north west] {$\b S^2$};
\draw (9.25,6.8) node[anchor=north west] {$\tau_d^{-1}\circ X(L_{E_1^d = -E_1^g})$};
\draw (14.8,-1.34) node[anchor=north west] {$(-e_1)$};
\draw (20.3,-1.34) node[anchor=north west] {$(-e_1)$};
\draw (15.3,4.4) node[anchor=north west] {$e_1$};
\draw (20.8,4.4) node[anchor=north west] {$e_1$};
\draw (11,4.64) node[anchor=north west] {$L_{E_1^d=-E_1^g}$};
\draw [color=black, line width=1.4pt] (18,1.5)-- ++(-4.5pt,0 pt) -- ++(9.0pt,0 pt) ++(-4.5pt,-4.5pt) -- ++(0 pt,9.0pt);
\draw[color=black] (17.5,0.5) node {$(m,v)$};
\end{tikzpicture}
\captionof{figure}{A positive intersection $(m,v)$ between the 2-chain $L_{E_1^d=-E_1^g}\times[-e_1,e_1]_{e_2^\Omega}$ and $\tau_d^{-1} \circ X (L_{E_1^d=-E_1^g})$ in $L_{E_1^d=-E_1^g}\times \b S^2$.\\}\label{orientation}
\end{center}
\noindent
It follows that
$$
\begin{aligned}
\langle \Omega(\bar\tau) , X(M) \rangle_{UM} &= - \langle [-e_1,e_1]_{e_2^\Omega}, \ P_{\b S^2} \circ \tau_d^{-1} \circ X(L_{E_1^d=-E_1^g}) \rangle_{\b S^2} \\
&= - lk_{\b S^2}\left(e_1-(-e_1), \ P_{\b S^2} \circ \tau_d^{-1} \circ X(L_{E_1^d=-E_1^g})\right),
\end{aligned}
$$
\iffalse
and, finally,
$$
\begin{aligned}
\langle \go P(\bar\tau,X), [0,1]\times X(M) \rangle_{[0,1]\times UM} &= lk_M\left(L_{E_1^d=X} + L_{E_1^g=X} \ , \ L_{E_1^d=-X} + L_{E_1^g=-X} \right) \\
&- lk_{\b S^2}\left(e_1-(-e_1), \ P_{\b S^2} \circ \tau_d^{-1} \circ X(L_{E_1^d=-E_1^g})\right).
\end{aligned}
$$
\fi
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem1}]
According to Lemmas~\ref{cor_reformcombingsb}~and~\ref{fond}, Lemma~\ref{lem1} is true for $p_1=\tilde p_1$.
\end{proof}
From now on, if $X$ is a torsion combing of a compact oriented 3-manifold $M$ and if $\bar\tau$ is a pseudo-parallelization of $M$ compatible with $X$, then set
$$
p_1([X],\bar\tau) = \tilde p_1([X],\bar\tau) \mbox{ \ and \ } p_1(\bar\tau,[X]) = \tilde p_1(\bar\tau,[X]).
$$
As an obvious consequence, we get the following lemma.
\begin{lemma} \label{lempluplus}
Under the assumptions of Lemma \ref{cor_reformcombingsb}, in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$, the class of $\go P(\bar\tau,X)$ is $p_1(\bar\tau, X)[S]$.
\end{lemma}
\begin{proof}[Proof of Theorem~\ref{thm_defp1Xb}]
Let $X_1$ and $X_2$ be torsion combings of two compact oriented 3-manifolds $M_1$ and $M_2$ with identified boundaries such that $X_1$ and $X_2$ coincide on the boundary. Let also $\bar\tau_1$ and $\bar\tau'_1$ be two pseudo-parallelizations of $M_1$ that extend the trivialization $\rho(X_1)$ and, similarly, let $\bar\tau_2$ be a pseudo-parallelization of $M_2$ that extends the trivialization $\rho(X_2)$. In such a context, let
$$
\begin{aligned}
p_1([X_1],[X_2])(\bar\tau_1,\bar\tau_2) &= p_1([X_1],\bar\tau_1) + p_1(\bar\tau_1,\bar\tau_2) + p_1(\bar\tau_2, [X_2]) \\
p_1([X_1],[X_2])(\bar\tau'_1,\bar\tau_2) &= p_1([X_1],\bar\tau'_1) + p_1(\bar\tau'_1,\bar\tau_2) + p_1(\bar\tau_2, [X_2]) \\
\end{aligned}
$$
and note that
$$
p_1([X_1],[X_2])(\bar\tau_1,\bar\tau_2) - p_1([X_1],[X_2])(\bar\tau'_1,\bar\tau_2) = p_1(\bar\tau_1,\bar\tau'_1) - \langle \go P(\bar\tau_1,\bar\tau'_1) , [0,1]\times X_1(M) \rangle.
$$
Using Proposition~\ref{prop_ppara}, we get $p_1([X_1],[X_2])(\bar\tau_1,\bar\tau_2) - p_1([X_1],[X_2])(\bar\tau'_1,\bar\tau_2)=0$. In other words $p_1([X_1],[X_2])(\bar\tau_1,\bar\tau_2)$ is independent of $\bar\tau_1$. Similarly, it is also independent of $\bar\tau_2$ so that we can drop the pseudo-parallelizations from the notation. Eventually, using Lemma~\ref{lem1}, we get the formula of the statement. \\
For the second part of the statement, if $M_1$ and $M_2$ are closed, conclude with Proposition~\ref{cor_reformcombings}, which ensures that $p_1(\bar\tau_i,[X_i]) = p_1([X_i])-p_1(\bar\tau_i)$ for $i\in \lbrace 1,2 \rbrace$.
\end{proof}
Let us now end this section by proving Theorem \ref{formuleplus} and Theorem \ref{GM}, starting with the following.
\begin{lemma} \label{lemmaplus1}
Let $(X,\rho)$ and $(Y,\rho)$ be $\partial$-compatible torsion combings of a compact oriented 3-manifold $M$. Let $C_4(X,Y)$ and $C_4(-X,-Y)$ be 4-chains of $[0,1]\times UM$ as in Lemma \ref{lem_evaluationII}. The class of $\go P(X,Y) \hspace{-1mm}=\hspace{-1mm} 4\big(C_4(X,Y) \cap C_4(-X,-Y)\big)$ in the space $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$ reads $p_1([X],[Y]) [S]$ where $[S]$ is the homology class of the fiber of $UM$ in $H_2([0,1]\times UM ; \b Q)$.
\end{lemma}
\begin{proof}
Let $\bar \tau$ be a pseudo-parallelization of $M$ compatible with $X$ and $Y$. By Theorem \ref{thm_defp1Xb}, in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$ :
$$
p_1([X],[Y])[S] = \big(p_1([X],\bar \tau) + p_1(\bar \tau,[Y]) \big)[S]
$$
Then, using Lemma \ref{lempluplus},
$$
\begin{aligned}
p_1([X],[Y])[S] &= [\go P(X,\bar \tau) + \go P(\bar \tau, Y)]\\
&= [4(C_4^+(X,\bar \tau)\cap C_4^-(X,\bar \tau)) + 4(C_4^+(\bar \tau, Y)\cap C_4^-(\bar \tau, Y)) ].
\end{aligned}
$$
Hence, reparameterizing and stacking $C_4^+(X,\bar\tau)$ and $C_4^+(\bar\tau,Y)$, resp. $C_4^-(X,\bar\tau)$ and $C_4^-(\bar\tau,Y)$, we get a 4-chain $C_4(X,Y)$, resp. $C_4(-X,-Y)$, as in Lemma \ref{lem_evaluationII} and
$$
p_1([X],[Y])[S]= 4 \cdot [C_4(X,Y)\cap C_4(-X,-Y)].
$$
To conclude the proof, see that if $C'_4(X,Y)$ is a 4-chain of $[0,1]\times UM$ with the same boundary as $C_4(X,Y)$, then $C'_4(X,Y)-C_4(X,Y)$ is homologous to a 4-cycle in $\lbrace 0 \rbrace \times UM$ so that
$$
\begin{aligned}
&[C'_4(X,Y) \cap C_4(-X,-Y)] - [C_4(X,Y) \cap C_4(-X,-Y)] \\
&= \left[\big(C'_4(X,Y)- C_4(X,Y) \big) \cap C_4(-X,-Y)\right]
\end{aligned}
$$
sits in $H_T^\rho(M)$. So, the class $[C_4(X,Y)\cap C_4(-X,-Y)]$ in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$ is independent of the choices for $C_4(X,Y)$. Similarly, it is independent of the choices for $C_4(-X,-Y)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{formuleplus}]
According to Lemma \ref{lemmaplus1}, it is enough to evaluate the class of the chain $4\big(C_4(X,Y) \cap C_4(-X,-Y)\big)$ in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$ where $\rho = \rho(X)$ and where $C_4(X,Y)$ and $C_4(-X,-Y)$ are 4-chains of $[0,1]\times UM$ as in Lemma~\ref{lem_evaluationII}. Let us consider the 4-chains
$$
\begin{aligned}
C_4(X,Y) &= \bar F_{t_1}(X,Y) - \lbrace t_1 \rbrace \times UM_{|\Sigma_{X=-Y}},\\
C_4(-X,-Y) &= \bar F_{t_2}(-X,-Y) - \lbrace t_2 \rbrace \times UM_{|\Sigma_{-X=Y}} ,
\end{aligned}
$$
where $0<t_1<t_2<1$, and where $\bar F_{t_1}(X,Y)$, resp. $\bar F_{t_2}(-X,-Y)$, is a 4-chain as in Definition~\ref{def_Ft} and $\Sigma_{X=-Y}$, resp. $\Sigma_{-X=Y}$, is a 2-chain of $M$ bounded by $L_{X=-Y}$, resp. $L_{-X=Y}$, provided by Proposition~\ref{prop_linksinhomologyI}. With these chains,
$$
\begin{aligned}
&C_4(X,Y) \cap C_4(-X,-Y) \\
&= -\lbrace t_1 \rbrace \times (-X)(\Sigma_{X=-Y}) + [t_1, t_2] \times (-X)(L_{-X=Y}) - \lbrace t_2 \rbrace \times Y(\Sigma_{-X=Y}).
\end{aligned}
$$
Hence, using Lemma \ref{lem_evaluationII} with $[0,1]\times X(M)$, in $H_2([0,1]\times UM ; \b Q)/H_T^\rho(M)$ :
$$
\begin{aligned}
[C_4(X,Y) \cap C_4(-X,-Y)] &= \langle C_4(X,Y)\cap C_4(-X,-Y) , [0,1] \times X(M) \rangle_{[0,1]\times UM} [S] \\
&= lk(L_{X=Y},L_{X=-Y}) [S].
\end{aligned}
$$
\iffalse
&= \langle - \lbrace t_1 \rbrace \times (-X)(\Sigma_{X=-Y}) + [t_1, t_2] \times (-X)(L_{-X=Y}) \\
& \hspace{6cm} - \lbrace t_2 \rbrace \times Y(\Sigma_{-X=Y}) , [0,1]\times X(M) \rangle_{[0,1]\times UM} [S] \\
&= - lk(L_{X=Y},L_{-X=Y}) [S] \\
\fi
\end{proof}
\iffalse
\begin{proof}
Let $(X,\sigma)$ and $(Y, \sigma)$ be $\partial$-compatible torsion combings of a compact oriented 3-manifold $M$ that represent the same $\mbox{spin$^c$}$-structure. By definition, $(X,\sigma)$ and $(Y, \sigma)$ are homotopic on $M\setminus \l B$ where $\l B$ is a 3-ball in $\mathring M$. \linebreak So, there exists a combing $Y'$ homotopic to $Y$ (on $M$) such that $L_{X=Y'} \cap \partial \l B = \emptyset$ and $L_{X=-Y'} \subset \l B$. Let $L^{\l B} = L_{X=Y'} \cap \l B$ and $L^{M\setminus \l B} = L_{X=Y'} \cap (M\setminus \l B)$. The link $L_{X=-Y'}$ bounds a compact oriented surface in $\l B$, hence
$$
\begin{aligned}
p_1([X], [Y]) = p_1([X], [Y']) &= 4\cdot lk(L_{X=Y'}, L_{X=-Y'})\\
&= 4\cdot lk(L^{M \setminus \l B} \sqcup L^{\l B}, L_{X=-Y'})\\
&= 4\cdot lk(L^{\l B}, L_{X=-Y'}).
\end{aligned}
$$
Finally, as in the closed case (see \cite[Subsection 2.3 and Corollary 2.22]{lescopcombing}), $X$ can be extended as a parallelization on $\l B$ so that $Y'$ induces a map from $\l B$ to $\b S^2$. Its class in $\pi_3(\b S^2)\simeq \b Z$ reads $lk(L^{\l B}, L_{X=-Y'})$. Hence, $p_1([X], [Y]) = 0$ if and only if $X$ and $Y$ are homotopic on $M$.
\end{proof}
\fi
\begin{proof}[Proof of Theorem \ref{GM}]
If $X$ and $Y$ are homotopic relatively to the boundary, then $p_1([X],[Y])=0$. \linebreak Conversely, consider two combings $X$ and $Y_0$ in the same $\mbox{spin$^c$}$-structure and assume that \linebreak $p_1([X],[Y_0])=0$. Since $Y_0$ is in the same $\mbox{spin$^c$}$-structure as $X$, there exists a homotopy from $Y_0$ to a combing $Y_1$ that coincides with $X$ outside a ball $\l B$ in $\mathring M$. \\
Let $\sigma$ be a unit section of $X^{\perp}_{|\l B}$, and let $(X, \sigma, X\wedge\sigma)$ denote the corresponding parallelization over $\l B$. Extend the unit section $\sigma$ as a generic section of $X^{\perp}$ such that $\sigma_{|\partial M}=\sigma(X)$, and deform $Y_1$ to $Y$ where
$$
Y(m)=\frac{Y_1(m) + \chi(m) \sigma(m)}{\parallel Y_1(m) + \chi(m)\sigma(m)\parallel}
$$
for a smooth map $\chi$ from $M$ to $[0,\varepsilon]$, such that $\chi^{-1}(0)=\partial M$ and $\chi$ maps the complement of a neighborhood of $\partial M$ to $\varepsilon$, where $\varepsilon$ is a small positive real number. The link $L_{X=Y}$ is the disjoint union of $L_{X=Y}\cap \l B$ and a link $L_2$ of $M \setminus \l B$, the link $L_{X=-Y}$ sits in $\l B$, and
$$
0=p_1(X,Y_0)=p_1(X,Y)= 4lk(L_{X=Y},L_{X=-Y})
$$
where $lk(L_{X=Y},L_{X=-Y})=lk(L_{X=Y}\cap \l B,L_{X=-Y})=0$.\\
The parallelization $(X, \sigma, X\wedge\sigma)$ turns the restriction $Y_{|\l B}$ into a map from the ball $\l B$ to $\b S^2 = \b S(\b R X \oplus \b R \sigma \oplus \b R X\wedge\sigma)$ constant on $\partial \l B$, thus into a map from $\l B / \partial \l B \simeq \b S^3$ to $\b S^2$, and it suffices to prove that this map is homotopic to the constant map to prove Theorem \ref{GM}. For this it suffices to prove that this map represents $0$ in $\pi_3(\b S^2) \simeq \b Z$. \\
There is a classical isomorphism from $\pi_3(\b S^2)$ to $\b Z$ that maps the class of a map $g$ from $\b S^3$ to $\b S^2$ to the linking number of the preimages of two regular points of $g$ under $g$ (see \cite{hopf} and \cite[Theorem 2]{pontrjagin}). It is easy to check that this map is well-defined, depends only on the homotopy class of $g$, and is a group morphism on $\pi_3(\b S^2)$ that maps the class of the Hopf fibration $\left((z_1,z_2) \in (\b S^3 \subset \b C^2) \mapsto (\sfrac{z_1}{z_2}) \in (\b C P^1=\b S^2) \right)$ to $\pm 1$. Therefore it is an isomorphism from $\pi_3(\b S^2)$ to $\b Z$. Since $Y$ is in the kernel of this isomorphism, it is homotopically trivial so that $Y$ is homotopic to a constant on $B$, relatively to the boundary of $B$, and $Y_0$ is homotopic to $X$ on $M$, relatively to the boundary of $M$.
\end{proof}
\section[Variation of Pontrjagin numbers under LP$_\b Q$-surgeries]{Variation of Pontrjagin numbers \\ under LP$_\b Q$-surgeries}
\subsection{For pseudo-parallelizations}
In this subsection we recall the variation formula and the finite type property of Pontrjagin numbers of pseudo-parallelizations, which are contained in \cite[Section 11]{lescopcube}.
\def\mbox{$\overline{\tau}$}{\mbox{$\overline{\tau}$}}
\begin{proposition} \label{prop_FTIppara}
\noindent
For $M$ a compact oriented 3-manifold and $(\sfrac{B}{A})$ an LP$_\b Q$-surgery in $M$, if ${\bar\tau}_{M}$ and ${\bar\tau}_{M(\sfrac{B}{A})}$ are pseudo-parallelizations of $M$ and $M(\sfrac{B}{A})$ which coincide on $M\setminus \mathring A$ and coincide with a genuine parallelization on $\partial A$, then
$$
p_1({\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})},{\mbox{$\overline{\tau}$}}_{M}) = p_1({{\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})}}_{|B},{{\mbox{$\overline{\tau}$}}_{M}}_{|A}).
$$
\end{proposition}
\begin{proof}
Let $W^-$ be a signature zero cobordism from $A$ to $B$. By definition, the obstruction $p_1({{\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})}}_{|B},{{\mbox{$\overline{\tau}$}}_{M}}_{|A})$ is the Pontrjagin obstruction to extending the complex trivialization \linebreak $\tau({{\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})}}_{|B},{{\mbox{$\overline{\tau}$}}_{M}}_{|A})$ of $TW^-_{\hspace{-1mm}|\partial W^-} \hspace{-1mm} \otimes \b C$ as a trivialization of $TW^- \hspace{-1mm}\otimes \b C$. Let $W^+ \hspace{-1.5mm}=\hspace{-1mm} [0,1]\times (M\setminus \mathring A)$ and let $V=-[0,1]\times\partial A$. As shown in \cite[Proof of Proposition 5.3 item 2]{lescopEFTI}, since $(\sfrac{B}{A})$ is an LP$_\b Q$-surgery in $M$, the manifold $W=W^+\cup_V W^-$ has signature zero. Furthermore, since $W$ has signature zero, $p_1({\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})},{\mbox{$\overline{\tau}$}}_{M})$ is the Pontrjagin obstruction to extending the triviali\-zation $\tau({\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})},{\mbox{$\overline{\tau}$}}_{M})$ of $TW_{|\partial W} \otimes \b C$ as a trivialization of $TW\otimes \b C$. Finally, it is clear that $\tau({\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})},{\mbox{$\overline{\tau}$}}_{M})$ extends as a trivialization of $TW_{|W^+} \otimes \b C$ so that
$$
p_1({\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})},{\mbox{$\overline{\tau}$}}_{M}) = p_1({{\mbox{$\overline{\tau}$}}_{M(\sfrac{B}{A})}}_{|B},{{\mbox{$\overline{\tau}$}}_{M}}_{|A}).
$$
\end{proof}
\begin{corollary} \label{cor_FTppara}
Let $M$ be a compact oriented 3-manifold and let $\lbrace \sfrac{B_i}{A_i} \rbrace_{i\in \lbrace 1, \ldots , k\rbrace}$ be a family of disjoint LP$_\b Q$-surgeries where $k\geqslant2$. For any family $\lbrace \bar \tau_I \rbrace_{I \subset \lbrace 1, \ldots , k \rbrace}$ of pseudo-parallelizations of the $\lbrace M(\lbrace \sfrac{B_i}{A_i}\rbrace_{i\in I}) \rbrace_{I \subset \lbrace 1, \ldots , k \rbrace}$ whose links sit in $M\setminus(\cup_{i=1}^k \partial A_i)$ and such that, for all subsets $I, J\subset \lbrace 1, \ldots , k \rbrace$, $\bar \tau^I$ and $\bar \tau^J$ coincide on $(M\setminus \cup_{I\cup J}A_i)\cup_{I\cap J} B_i$, the following identity holds :
$$
\sum_{I \subset \lbrace 2, \ldots , k \rbrace} (-1)^{\card(I)} p_1 (\bar \tau ^I,\bar \tau ^{I\cup\lbrace 1 \rbrace})=0.
$$
\end{corollary}
\subsection[Lemmas for the proof of Theorem~\ref{thm_D2nd}]{Lemmas for the proof of Theorem~\ref{thm_D2nd}}
\begin{lemma} \label{lem_Xdsection}
If $X$ is a combing of a compact oriented 3-manifold $M$ and if $\partial A$ is the connected boundary of a submanifold of $M$ of dimension 3, then the normal bundle $X^\perp_{|\partial A}$ admits a nonvanishing section.
\end{lemma}
\begin{proof}
Parallelize $M$ so that $X$ induces a map $X_{|\partial A} : \partial A \rightarrow \b S^2$. This map must have degree 0 since $X$ extends this map to $A$ (so that $(X_{|\partial A})_* : H_2(\partial A;\b Q)\rightarrow H_2(\b S^2;\b Q)$ factors through the inclusion $H_2(\partial A ; \b Q) \rightarrow H_2(A;\b Q)$, which is zero). It follows that $X_{|\partial A}$ is homotopic to the map $(m \in\partial A \mapsto e_1 \in \b S^2)$ whose normal bundle admits a nonvanishing section.
\end{proof}
\begin{lemma} \label{ind}
Let $(M,X)$ be a compact oriented 3-manifold equipped with a combing, let $(\sfrac{B}{A},X_B)$ be an LP$_\b Q$-surgery in $(M,X)$ and let $\sigma$ be a nonvanishing section of $X^\perp_{|\partial A}$. Let $P$ stand for Poincaré duality isomorphisms and recall the sequence of isomorphisms induced by the inclusions $i^{A_i}$ and $i^{B_i}$
$$
H_1(A;\b Q) \stackrel{i^{A}_*}{\longleftarrow} \frac{H_1(\partial A;\b Q)}{\go L_{A}} = \frac{H_1(\partial B;\b Q)}{\go L_{B}} \stackrel{i^{B}_*}{\longrightarrow} H_1(B;\b Q).
$$
The class $ \left(i^A_* \circ {(i^B_*)}^{-1}\left(P(e_2^B(X_B^\perp, \sigma))\right) - P(e_2^A(X_{|A}^\perp, \sigma))\right)$ in $H_1( A ; \b Q)$ is independent of the choice of the section $\sigma$.
\end{lemma}
\begin{proof}
Let us drop the inclusions $i^B_*$ and $i^A_*$ from the notation. According to Proposition~\ref{prop_euler}, the class $P(e_2^A(X_{|A}^\perp, \sigma))$ verifies
$$
[ X(A) - (-X)(A) +\l H_{X,\sigma}^{-X}(\partial A \times [0,1]) ] = [ P(e_2^A(X_{|A}^\perp, \sigma)) \times \b S^2 ] \mbox{ \ in \ } H_1(UA ;\b Q).
$$
It follows that, for another choice $\sigma'$ of section of $X^\perp_{|\partial A}$,
$$
\begin{aligned}
[P(e_2^B(X_B^\perp, \sigma)) \times \b S^2 & - P(e_2^B(X_B^\perp, \sigma')) \times \b S^2 ] \\
&= [\l H_{X,\sigma}^{-X}(\partial A\times [0,1])-\l H_{X,\sigma'}^{-X}(\partial A\times [0,1])] \\
&=[P(e_2^A(X_{|A}^\perp, \sigma)) \times \b S^2 - P(e_2^A(X_{|A}^\perp, \sigma')) \times \b S^2 ].
\end{aligned}
$$
\end{proof}
\begin{lemma} \label{zero}
Let $(M,X)$ be a compact oriented 3-manifold equipped with a combing and let $(\sfrac{B}{A},X_B)$ be an LP$_\b Q$-surgery in $(M,X)$. If $(X,\sigma)$ is a torsion combing then $(X(\sfrac{B}{A}),\sigma)$ is a torsion combing if and only if
$$
i_*^{A} \circ (i_*^B)^{-1} \big( P(e_2^B({X_{B}}^\perp, \zeta))\big) - P(e_2^A(X_{|A}^\perp, \zeta)) = 0 \mbox{ \ in $H_1(M; \b Q)$}
$$
for some nonvanishing section $\zeta$ of $X^\perp_{|\partial A}$.
\end{lemma}
\begin{proof}
By definition, we have
$$
\begin{aligned}
P(e_2(X^\perp,\sigma)) &= P(e_2^A(X_{|A}^\perp, \zeta)) + P(e_2^{M \setminus \mathring A}(X^\perp,\sigma \cup \zeta)) \\
P(e_2({X(\sfrac{B}{A})}^\perp,\sigma)) &= P(e_2^B(X_{B}^\perp, \zeta)) + P(e_2^{M \setminus \mathring A}(X'^\perp,\sigma \cup \zeta)) \\
\end{aligned}
$$
where $\zeta$ is any nonvanishing section of $X^\perp_{|\partial A}$. So, it follows that, using appropriate identifications,
$$
P(e_2(X(\sfrac{B}{A})^\perp,\sigma)) - P(e_2({X}^\perp,\sigma)) = P(e_2^B(X_{B}^\perp, \zeta)) - P(e_2^A(X_{|A}^\perp, \zeta)).
$$
If $X$ is a torsion combing, then $P(e_2({X}^\perp,\sigma))$ is rationally null-homologous in $M$. Hence, $X(\sfrac{B}{A})$ is a torsion combing if and only if
$$
i_*^{A} \circ (i_*^B)^{-1} \big( P(e_2^B({X_{B}}^\perp, \zeta))\big) - P(e_2^A(X_{|A}^\perp, \zeta)) = 0 \mbox{ \ in $H_1(M; \b Q)$}.
$$
\end{proof}
\begin{lemma} \label{norm}
Let $(M,X)$ be a compact oriented 3-manifold equipped with a combing. Let $\lbrace (\sfrac{B_i}{A_i} , X_{B_i}) \rbrace_{i \in \lbrace 1,\ldots,k\rbrace}$ be a family of disjoint LP$_\b Q$-surgeries in $(M,X)$, where $k \in \b N\setminus \lbrace 0 \rbrace$. For all $I\subset \lbrace 1, \ldots, k \rbrace$, let $M_I=M(\lbrace \sfrac{B_i}{A_i} \rbrace_{i \in I})$ and $X^I=X(\lbrace \sfrac{B_i}{A_i}\rbrace_{i \in I})$. There exists a family of pseudo-parallelizations $\lbrace\bar \tau^I\rbrace_{ I \subset \lbrace 1,\ldots,k\rbrace}$ of the $\lbrace M_I \rbrace_{I \subset \lbrace 1, \ldots , k \rbrace}$ such that :
\begin{enumerate}[(i)]
\item the third vector of $\bar\tau=\bar\tau^\emptyset$ coincides with $X$ on $\cup_{i=1}^k \partial A_i$,
\item for all $I \subset \lbrace 1 , \ldots , k \rbrace$, $\bar\tau^I$ is compatible with $X^I$,
\item for all $I \subset \lbrace 1,\ldots,k\rbrace$, if $\gamma^I$ denotes the link of $\bar \tau_I$, then $N(\gamma^I) \cap \left(\cup_{i=1}^k \partial A_i \right) = \emptyset$,
\item for all $I, J \subset \lbrace 1,\ldots,k\rbrace$, $\bar\tau^I$ and $\bar\tau^J$ coincide on $(M\setminus \cup_{i\in I\cup J} A_i)\cup_{i\in I\cap J}B_i$,
\item there exist links $L^\pm_{A_i}$ in $A_i$, $L^\pm_{B_i}$ in $B_i$ and $L^\pm_{ext}$ in $M \setminus \cup_{i=1}^k \mathring A_i$ such that, for all subset $I \subset \lbrace 1, \ldots , k \rbrace$~:
$$
\begin{aligned}
& 2 \cdot L_{\bar\tau^I=X^I} = L_{{E^d}^I=X^I} + L_{{E^g}^I=X^I} = L^+_{ext} + \sum_{i \in I} L^+_{B_i} + \sum_{i \in \lbrace 1,\ldots , k \rbrace \setminus I} L^+_{A_i} , \\
& 2 \cdot L_{\bar\tau^I=-X^I} = L_{{E^d}^I=-X^I} + L_{{E^g}^I=-X^I} = L^-_{ext} + \sum_{i \in I} L^-_{B_i} + \sum_{i \in \lbrace 1,\ldots , k \rbrace \setminus I} L^-_{A_i},
\end{aligned}
$$
where ${E^d}^I$ and ${E^g}^I$ are the Siamese sections of $\bar\tau^I$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\l C$ denote a collar of $\cup_{i=1}^k \partial A_i$. Using Lemma~\ref{lem_Xdsection}, construct a trivialization $\tau_e$ of $\cup_{i=1}^k TM_{|\l C}$ so that its third vector coincides with $X$ on $\l C$. Then use Lemma~\ref{lem_extendpparallelization} to extend $\tau_e$ as pseudo-parallelizations of the $\lbrace A_i \rbrace_{i\in \lbrace 1,\ldots , k \rbrace }$, of the $\lbrace B_i \rbrace_{i\in \lbrace 1,\ldots , k \rbrace}$ and of $M\setminus(\cup_{i=1}^k \mathring A_i)$. Finally, use these pseudo-parallelizations to construct the pseudo-parallelizations of the 3-manifolds $\lbrace M_I \rbrace_{I \subset \lbrace 1,\ldots , k \rbrace}$ as in the statement.
\end{proof}
\begin{lemma} \label{inter}
In the context of Lemma~\ref{norm}, using the sequence of isomorphisms induced by the inclusions $i^{A_i}$ and $i^{B_i}$
$$
H_1(A_i;\b Q) \stackrel{i^{A_i}_*}{\longleftarrow} \frac{H_1(\partial A_i;\b Q)}{\go L_{A_i}} = \frac{H_1(\partial B_i;\b Q)}{\go L_{B_i}} \stackrel{i^{B_i}_*}{\longrightarrow} H_1(B_i;\b Q),
$$
for all $i \in \lbrace 1 , \ldots , k \rbrace$, we have that, in $H_1(A_i ; \b Q)$,
$$
\left[i_{\ast}^{A_i} \circ \left(i_{\ast}^{B_i}\right)^{-1}\left( L_{B_i}^{\pm}\right)- L_{A_i}^\pm\right] = \pm \left( i^{A_i}_* \circ(i^{B_i}_*)^{-1}\big(P(e_2^{B_i}(X_{B_i}^\perp, \sigma_i))\big) - P(e_2^{A_i}(X_{|A_i}^\perp, \sigma_i)) \right)
$$
where $\sigma_i$ is any nonvanishing section of $X^\perp_{|\partial A_i}$.
\end{lemma}
\begin{proof}
Let us drop the inclusions $i^B_*$ and $i^A_*$ from the notation. Let $i \in \lbrace 1,\ldots,k \rbrace$. According to Lemma~\ref{zero}, it is enough to prove the statement for a particular non vanishing section $\sigma_i$ of $X^\perp_{|\partial A_i}$. Recall that $A_i$ is equipped with a combing $X_{|A_i}$ and a pseudo-parallelization \linebreak $\bar\tau_{|A_i}=(N(\gamma\cap A_i); {\tau_e}_{|A_i}, {\tau_d}_{|A_i}, {\tau_g}_{|A_i})$ such that $X_{|\partial A_i}$ coincides with ${E_3^e}_{|\partial A_i}$ where $\tau_e=(E_1^e,E_2^e,E_3^e)$. Furthermore,
$$
L^+_{A_i} = 2\cdot L_{\bar\tau_{|A_i}= X_{|A_i}} \mbox{ \ and \ } L^-_{A_i} = 2\cdot L_{\bar\tau_{|A_i}= -X_{|A_i}}.
$$
Construct a pseudo-parallelization $\check\tau = (N(\check\gamma) ; \check \tau_e ,\check \tau_d ,\check \tau_g)$ of $A_i$ by modifying $\bar\tau$ as follows so that $\check\tau$ and $X$ coincide on $\partial A_i$. Consider a collar $\l C=[0,1]\times \partial A_i$ of $\partial A_i$ such that $\lbrace 1 \rbrace \times \partial A_i = \partial A_i$ and $\l C \cap \gamma = \emptyset$. Without loss of generality, assume that $X_{|A_i}$ coincides with $E_3^e$ on the collar $\l C$. Let $\check\tau$ coincide with $\bar\tau_{|A_i}$ on $\overline{A_i \setminus \l C}$. End the construction of $\check\tau$ by requiring
$$
\forall (s,b) \in \l C = [0,1] \times \partial A_i, \ \forall v \in \b S^2 \ : \ \check\tau_e( (s,b),v) = \tau_e \left((s,b), R_{e_2, \frac{-\pi s}{2}}(v)\right).
$$
Note that $\check\tau$ and $X_{|A_i}$ are compatible and that
$$
L^+_{A_i} = 2\cdot L_{\check\tau= X_{|A_i}} \mbox{ \ and \ } L^-_{A_i} = 2\cdot L_{\check\tau= -X_{|A_i}}.
$$
Using $\check\tau$ and Proposition~\ref{prop_linksinhomologyI}, we get
$$
\begin{aligned}
[L^+_{A_i}] &= P(e_2^{A_i}(X^\perp_{|A_i}, {\check E^e}_{2|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({\check{E}}^{d \perp}_{1}, {\check E^e}_{2|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({\check{E}}^{g \perp}_{1}, {\check E^e}_{2|\partial A_i})) \\
[L^-_{A_i}] &= - P(e_2^{A_i}(X^\perp_{|A_i}, {\check E^e}_{2|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({\check{E}}^{d \perp}_{1}, {\check E^e}_{2|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({\check{E}}^{g \perp}_{1}, {\check E^e}_{2|\partial A_i}))
\end{aligned}
$$
where $\check\tau_e=(\check E_1^e,\check E_2^e,\check E_3^e)$ and where ${\check E_1}^d$ and ${\check E_1}^g$ are the Siamese sections of $\check\tau$. By construction, it follows that
$$
\begin{aligned}
[L^+_{A_i}] &= P(e_2^{A_i}(X^\perp_{|A_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({E_1^d}^\perp_{|A_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({E_1^g}^\perp_{|A_i}, {E_2^e}_{|\partial A_i})) \\
[L^-_{A_i}] &= -P(e_2^{A_i}(X^\perp_{|A_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({E_1^d}^\perp_{|A_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{A_i}({E_1^g}^\perp_{|A_i}, {E_2^e}_{|\partial A_i}))
\end{aligned}
$$
where $E_1^d$ and $E_1^g$ are the Siamese sections of $\bar\tau$. Using the same method, we also get that
$$
\begin{aligned}
[L^+_{B_i}] &= P(e_2^{B_i}(X^\perp_{B_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{B_i}({E_1^d}^\perp_{|B_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{B_i}({E_1^g}^\perp_{|B_i}, {E_2^e}_{|\partial A_i})) \\
[L^-_{B_i}] &= -P(e_2^{B_i}(X^\perp_{B_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{B_i}({E_1^d}^\perp_{|B_i}, {E_2^e}_{|\partial A_i})) + \frac{1}{2} P(e_2^{B_i}({E_1^g}^\perp_{|B_i}, {E_2^e}_{|\partial A_i})).
\end{aligned}
$$
Conclude with Lemma~\ref{simppara}.
\end{proof}
\subsection[Variation formula for torsion combings]{Variation formula for torsion combings}
\begin{proof}[Proof of Theorem~\ref{thm_D2nd}]
Let $(M,X)$ be a compact oriented 3-manifold equipped with a combing. Let $\lbrace (\sfrac{B_i}{A_i},X_{B_i}) \rbrace_{i\in \lbrace 1, 2\rbrace}$ be two disjoint LP$_\b Q$-surgeries in $(M,X)$ and assume that, for all subset $I\subset\lbrace 1,2 \rbrace$, $X^I=X(\lbrace\sfrac{B_i}{A_i}\rbrace_{i \in I})$ is a torsion combing of the 3-manifold $M_I = M(\lbrace \sfrac{B_i}{A_i} \rbrace_{i \in I})$. Note that, for all $I,J\subset \lbrace 1,2 \rbrace$, $X^I$ and $X^J$ coincide on $(M\setminus \cup_{i\in I\cup J} A_i)\cup_{i\in I\cap J}B_i$. Finally, let $\lbrace \bar\tau^I \rbrace_{I\subset \lbrace 1,2\rbrace}$ be a family of pseudo-parallelizations as in Lemma~\ref{norm}, let ${E_1^d}^I$ and ${E_1^g}^I$ denote the Siamese sections of $\bar\tau^I$ for all $I\subset \lbrace 1,2 \rbrace$ and let $L_I$ stand for $L_{{E_1^d}^I=-{E_1^g}^I}$. Using Corollary~\ref{cor_FTppara}, we have
$$
\begin{aligned}
p_1 \left([X^{\lbrace 2 \rbrace}],[X^{\lbrace 1,2 \rbrace}]\right) - p_1\left([X],[X^{\lbrace 1 \rbrace}]\right)
&= p_1 \left([X^{\lbrace 2 \rbrace}],[X^{\lbrace 1,2 \rbrace}]\right)- p_1 \left(\bar\tau^{\lbrace 2 \rbrace},\bar\tau^{\lbrace 1,2 \rbrace}\right) \\
& - p_1\left([X],[X^{\lbrace 1 \rbrace}]\right) + p_1\left(\bar\tau,\bar\tau^{\lbrace 1 \rbrace}\right) \\
\end{aligned}
$$
which, using Lemma~\ref{lem1} and Theorem~\ref{thm_defp1Xb}, reads
$$
\begin{aligned}
& 4 \cdot lk_{M_{\lbrace1,2\rbrace}}(L_{\bar\tau^{\lbrace 1,2 \rbrace}=X^{\lbrace 1,2 \rbrace}} \ , \ L_{\bar\tau^{\lbrace 1,2 \rbrace}=-X^{\lbrace 1,2 \rbrace}}) - 4 \cdot lk_{M_{\lbrace 2 \rbrace}}(L_{\bar\tau^{\lbrace 2 \rbrace}=X^{\lbrace 2 \rbrace}} \ , \ L_{\bar\tau^{\lbrace 2 \rbrace}=-X^{\lbrace 2 \rbrace}}) \\
& - lk_{\b S^2} \left(e_1-(-e_1) \ , \ P_{\b S^2} \circ (\tau^{\lbrace 1,2 \rbrace}_d)^{-1} \circ X^{\lbrace 1,2 \rbrace}(L_{\lbrace 1,2 \rbrace}) - P_{\b S^2} \circ (\tau^{\lbrace 2 \rbrace}_d)^{-1} \circ X^{\lbrace 2 \rbrace}(L_{\lbrace 2 \rbrace})\right)\\
&- 4 \cdot lk_{M_{\lbrace 1 \rbrace}}(L_{\bar\tau^{\lbrace 1 \rbrace}=X^{\lbrace 1 \rbrace}} \ , \ L_{\bar\tau^{\lbrace 1 \rbrace}=-X^{\lbrace 1 \rbrace}}) + 4 \cdot lk_{M}(L_{\bar\tau=X} \ , \ L_{\bar\tau=-X}) \\
& + lk_{\b S^2} \left(e_1-(-e_1) \ , \ P_{\b S^2} \circ (\tau^{\lbrace 1 \rbrace}_d)^{-1} \circ X^{\lbrace 1 \rbrace}(L_{\lbrace 1 \rbrace}) - P_{\b S^2} \circ (\tau_d)^{-1} \circ X(L) \right).
\end{aligned}
$$
This can further be reduced to the following by using Lemma~\ref{norm},
$$
\begin{aligned}
& lk_M \left( L^+_{ext} + L^+_{A_1} + L^+_{A_2}, \ L^-_{ext} + L^-_{A_1} + L^-_{A_2} \right) \\
&- lk_{M_{\lbrace 1 \rbrace}} \left( L^+_{ext} + L^+_{B_1} + L^+_{A_2}, \ L^-_{ext} + L^-_{B_1} + L^-_{A_2} \right) \\
&- lk_{M_{\lbrace 2 \rbrace}} \left( L^+_{ext} + L^+_{A_1} + L^+_{B_2}, \ L^-_{ext} + L^-_{A_1} + L^-_{B_2} \right) \\
&+ lk_{M_{\lbrace 1,2 \rbrace}} \left( L^+_{ext} + L^+_{B_1} + L^+_{B_2}, \ L^-_{ext} + L^-_{B_1} + L^-_{B_2} \right).
\end{aligned}
$$
In order to compute these linking numbers, let us construct specific 2-chains. Let us introduce a more convenient set of notations. For all $i \neq j \in \lbrace 1,2 \rbrace$, let
$$
L^\pm_{i}=L^\pm_{A_i}, \ L^\pm_{ij}= L^\pm_{A_i} + L^\pm_{A_j}, \ L^\pm_{eij}= L^\pm_{ext} + L^\pm_{A_i} + L^\pm_{A_j}.
$$
Set also similar notations with primed indices where a primed index $i'$, $i \in \lbrace 1,2 \rbrace$, indicates that $L^\pm_{A_i}$ should be replaced by $L^\pm_{B_i}$. For instance, $L^\pm_{i'}=L^\pm_{B_i}$, $L^\pm_{i,j'}= L^\pm_{A_i} + L^\pm_{B_j}$, \textit{etc}. Using these notations, $p_1 \left([X^{\lbrace 2 \rbrace}],[X^{\lbrace 1,2 \rbrace}]\right) - p_1\left([X],[X^{\lbrace 1 \rbrace}]\right)$ reads :
$$
\begin{aligned}
lk_M(L_{e12}^+,L_{e12}^-)- lk_{M_{\lbrace 1 \rbrace}}(L_{e1'2}^+,L_{e1'2}^-) - lk_{M_{\lbrace 2 \rbrace}}(L_{e12'}^+,L_{e12'}^-) + lk_{M_{\lbrace 1,2 \rbrace}} (L_{e1'2'}^+,L_{e1'2'}^-) .
\end{aligned}
$$
Recall from Lemma~\ref{nulexcep} that there exist rational 2-chains $\Sigma_{e12}^\pm$ of $M$ which are bounded by the links $L^\pm_{e12}$. Similarly, there exist rational two chains $\Sigma^\pm_{e1'2'}$ bounded by $L^\pm_{e1'2'}$ in $M_{\lbrace 1,2\rbrace}$. Note that, for all $i \in \lbrace 1,2 \rbrace$, the 2-chains $\Sigma_{e12}^\pm\cap A_i$ are cobordisms between the $L^\pm_i$ and 1-chains $\ell^\pm_i$ in $\partial A_i$. Similarly, for all $i \in \lbrace 1,2 \rbrace$, the 2-chains $\Sigma_{e1'2'}^\pm\cap B_i$ are cobordisms between the $L^\pm_{i'}$ and 1-chains $\ell^\pm_{i'}$ in $\partial B_i$. Furthermore, according to Lemma~\ref{inter}, for all $i\in \lbrace 1,2 \rbrace$ and for any nonvanishing section $\sigma_i$ of $X^\perp_{|\partial A_i}$, in $H_1(\partial A_i ; \b Q)/\go L_{A_i}$ :
$$
[\ell^\pm_{i'}-\ell^\pm_{i}]= \pm \big((i_*^{B_i})^{-1}(P(e_2^{B_i}(X_{B_i}^\perp, \sigma_i))) - (i_*^{A_i})^{-1}(P(e_2^{A_i}(X_{|A_i}^\perp, \sigma_i)))\big).
$$
So, according to Lemma~\ref{zero}, for all $I \subset \lbrace 1,2 \rbrace$, there exists a 2-chain $S_{\sfrac{B_i}{A_i}}^I$ in $M_I$ which is bounded by $\ell^+_{i'}-\ell^+_i$. Finally, since $(\sfrac{B_1}{A_1})$ and $(\sfrac{B_2}{A_2})$ are LP$_\b Q$-surgeries, we can construct these chains so that
$$
\begin{aligned}
S_{\sfrac{B_1}{A_1}}^{\lbrace 1\rbrace} \cap (M_{\lbrace 1 \rbrace}\minusens \mathring A_2) &= S_{\sfrac{B_1}{A_1}}^{\lbrace 1,2 \rbrace} \cap (M_{\lbrace 1,2 \rbrace}\minusens \mathring B_2), \\
S_{\sfrac{B_2}{A_2}}^{\lbrace 2\rbrace} \cap (M_{\lbrace 2 \rbrace}\minusens \mathring A_1) &= S_{\sfrac{B_2}{A_2}}^{\lbrace 1,2 \rbrace} \cap (M_{\lbrace 1,2 \rbrace})\minusens \mathring B_1).
\end{aligned}
$$
Let us now return to the computation of $p_1 \left([X^{\lbrace 2 \rbrace}],[X^{\lbrace 1,2 \rbrace}]\right) - p_1\left([X],[X^{\lbrace 1 \rbrace}]\right)$. Using the 2-chains we constructed, we have :
$$
\begin{aligned}
lk_M(L_{e12}^+,L_{e12}^-) &= \langle \Sigma^+_{e12}, L_{e12}^- \rangle \\
lk_{M_{\lbrace 1 \rbrace}}(L_{e1'2}^+,L_{e1'2}^-) &= \langle \Sigma^+_{e12} \cap (M \setminus \mathring A_1) + S_{\sfrac{B_1}{A_1}}^{\lbrace 1 \rbrace} + \Sigma^+_{e1'2'}\cap B_1, L_{e1'2}^- \rangle \\
lk_{M_{\lbrace 2 \rbrace}}(L_{e12'}^+,L_{e12'}^-) &= \langle \Sigma^+_{e12}\cap (M \setminus \mathring A_2) + S_{\sfrac{B_2}{A_2}}^{\lbrace 2 \rbrace} + \Sigma^+_{e1'2'}\cap B_2 , L_{e12'}^- \rangle \\
lk_{M_{\lbrace 1,2 \rbrace}} (L_{e1'2'}^+,L_{e1'2'}^-)&=\langle \Sigma^+_{e12}\cap(M\setminus(\mathring A_1\hspace{-1mm}\cup\hspace{-1mm}\mathring A_2)) +S_{\sfrac{B_1}{A_1}}^{\lbrace 1 , 2 \rbrace} \\
&\hspace{1cm}+S_{\sfrac{B_2}{A_2}}^{\lbrace 1 , 2 \rbrace} + \Sigma^+_{e1'2'}\cap(B_1\cup B_2) , L_{e1'2'}^- \rangle. \\
\end{aligned}
$$
So, the contribution of the intersections in $M \minusens (\mathring A_1\cup \mathring A_2)$ is zero since it reads :
$$
\begin{aligned}
& \langle \Sigma^+_{e12} \cap (M \minusens (\mathring A_1 \cup \mathring A_2)) , L_{e}^- \rangle_{M\setminus (\mathring A_1\cup \mathring A_2)} \\
& -\langle \Sigma^+_{e12} \cap (M \minusens (\mathring A_1 \cup \mathring A_2)) + S_{\sfrac{B_1}{A_1}}^{\lbrace 1 \rbrace}\cap (M \minusens (\mathring A_1 \cup \mathring A_2)) , L_{e}^- \rangle_{M \setminus (\mathring A_1\cup \mathring A_2)} \\
& -\langle \Sigma^+_{e12} \cap (M \minusens (\mathring A_1 \cup \mathring A_2)) + S_{\sfrac{B_2}{A_2}}^{\lbrace 2 \rbrace}\cap (M \minusens (\mathring A_1 \cup \mathring A_2)), L_{e}^- \rangle_{M\setminus (\mathring A_1\cup \mathring A_2)} \\
& +\langle \Sigma^+_{e12} \cap (M \minusens (\mathring A_1 \cup \mathring A_2)) + S_{\sfrac{B_1}{A_1}}^{\lbrace 1 , 2 \rbrace}\cap (M \minusens (\mathring A_1 \cup \mathring A_2)) \\
&\hspace{5cm}+ S_{\sfrac{B_2}{A_2}}^{\lbrace 1 , 2 \rbrace}\cap (M \minusens (\mathring A_1 \cup \mathring A_2)) , L_{e}^- \rangle_{M\setminus (\mathring A_1\cup \mathring A_2)} .
\end{aligned}
$$
The contribution in $A_1$ is
$$
\begin{aligned}
\langle \Sigma^+_{e12} \cap A_1 , L_{1}^- \rangle_{A_1} - \langle\Sigma^+_{e12} \cap A_1 + S_{\sfrac{B_2}{A_2}}^{\lbrace 2 \rbrace}\cap A_1 , L_{1}^- \rangle_{A_1}
= - \langle S_{\sfrac{B_2}{A_2}}^{\lbrace 2 \rbrace}\cap A_1, L_{1}^- \rangle_{A_1}.
\end{aligned}
$$
In $A_2$, we similarly get $- \langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 \rbrace}\cap A_2, L_{2}^- \rangle_{A_2}$. The contribution in $B_1$ is
$$
\begin{aligned}
&-\hspace{-1mm}\langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 \rbrace}\cap B_1 \hspace{-1mm}+\hspace{-1mm} \Sigma^+_{e1'2'}\cap B_1 , L_{1'}^- \rangle_{B_1} \hspace{-1mm}+\hspace{-1mm} \langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 , 2 \rbrace}\cap B_1 \hspace{-1mm} +\hspace{-1mm} S_{\sfrac{B_2}{A_2}}^{\lbrace 1 , 2 \rbrace}\cap B_1 \hspace{-1mm} +\hspace{-1mm} \Sigma^+_{e1'2'}\cap B_1 , L_{1'}^- \rangle_{B_1} \\
&= \langle S_{\sfrac{B_2}{A_2}}^{\lbrace 1 , 2 \rbrace}\cap B_1 , L_{1'}^- \rangle_{B_1}
\end{aligned}
$$
and, in $B_2$, we get $\langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 , 2 \rbrace}\cap B_2 , L_{2'}^- \rangle_{B_2}$. Eventually, $L_{\lbrace X^I \rbrace}(\sfrac{B_i}{A_i}) = i_*^{A_i}([\ell_{i'}^+ - \ell_{i}^+])$ for $i \in \lbrace 1,2\rbrace$. Moreover, recall that $[\ell^+_{i'}-\ell^+_{i}]=-[\ell^-_{i'}-\ell^-_{i}]$ in $H_1(\partial A_i ; \b Q)/\go L_{A_i}$, and complete the computations~:
$$
\begin{aligned}
p_1 & \left([X^{\lbrace 2 \rbrace}],[X^{\lbrace 1,2 \rbrace}]\right) - p_1\left([X],[X^{\lbrace 1 \rbrace}]\right) \\
&= \langle S_{\sfrac{B_2}{A_2}}^{\lbrace 1 , 2 \rbrace}\cap B_1 , L_{1'}^- \rangle_{B_1} + \langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 , 2 \rbrace}\cap B_2 , L_{2'}^- \rangle_{B_2} \\
&- \langle S_{\sfrac{B_2}{A_2}}^{\lbrace 2 \rbrace}\cap A_1, L_{1}^- \rangle_{A_1} - \langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 \rbrace} \cap A_2, L_{2}^- \rangle_{A_2} \\
&= \langle S_{\sfrac{B_2}{A_2}}^{\lbrace 1 , 2 \rbrace} , \ell_{1'}^- \rangle_{M_{\lbrace 1,2 \rbrace}} + \langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 , 2 \rbrace} , \ell_{2'}^- \rangle_{M_{\lbrace 1,2 \rbrace}} \\
&- \langle S_{\sfrac{B_2}{A_2}}^{\lbrace 2 \rbrace}, \ell_{1}^- \rangle_{M_{\lbrace 2 \rbrace}} - \langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 \rbrace} , \ell_{2}^- \rangle_{M_{\lbrace 1 \rbrace}} \\
&= \langle S_{\sfrac{B_2}{A_2}}^{\lbrace 1 , 2 \rbrace} , \ell_{1'}^- - \ell_{1}^- \rangle_{M_{\lbrace 1,2 \rbrace}} + \langle S_{\sfrac{B_1}{A_1}}^{\lbrace 1 , 2 \rbrace} , \ell_{2'}^- - \ell_{2}^- \rangle_{M_{\lbrace 1,2 \rbrace}} \\
&=- 2 \cdot lk_M \left(L_{\lbrace X^I \rbrace}(\sfrac{B_1}{A_1}) \ , \ L_{\lbrace X^I \rbrace}(\sfrac{B_2}{A_2}) \right).
\end{aligned}
$$
\iffalse
&= lk_{M_{\lbrace 1,2 \rbrace}} \left(L_{\lbrace X^I \rbrace}(\sfrac{B_2}{A_2}) \ , \ - L_{\lbrace X^I \rbrace}(\sfrac{B_1}{A_1})\right) \\
&+ \ lk_{M_{\lbrace 1,2 \rbrace}}\left(L_{\lbrace X^I \rbrace}(\sfrac{B_1}{A_1}) \ , \ - L_{\lbrace X^I \rbrace}(\sfrac{B_2}{A_2})\right) \\
&=- 2 \cdot lk_{M_{\lbrace 1,2 \rbrace}} \left(L_{\lbrace X^I \rbrace}(\sfrac{B_1}{A_1}) \ , \ L_{\lbrace X^I \rbrace}(\sfrac{B_2}{A_2}) \right) \\
\fi
\end{proof}
\newpage
\nocite{hirzebruch,KM,pontrjagin,turaev,lickorish,rolfsen,MR1030042,MR1189008,MR1712769}
\bibliographystyle{alpha}
|
\section*{Highlights}
\begin{itemize}
\item Remarkable new data on connectivity and activity raise the promise and raise the bar for linking structure and dynamics in neural networks.
\item Recent theories aim at a statistical approach, in which the enormous complexity of wiring diagrams is reduced to key features of that connectivity that drive coherent, network-wide activity.
\item We provide a unified view of three branches of this work, tied to a broadly useful ``neural response'' formula that explicitly relates connectivity to spike train statistics.
\item This isolates a surprisingly systematic role for the local structure and spatial scale of connectivity in determining spike correlations, and shows how the coevolution of structured connectivity and spiking statistics through synaptic plasticity can be predicted self-consistently.
\end{itemize}
\section*{Abstract}
An essential step toward understanding neural circuits is linking their structure and their dynamics. In general, this relationship can be almost arbitrarily complex. Recent theoretical work has, however, begun to identify some broad principles underlying collective spiking activity in neural circuits. The first is that local features of network connectivity can be surprisingly effective in predicting global statistics of activity across a network. The second is that, for the important case of large networks with excitatory-inhibitory balance, correlated spiking persists or vanishes depending on the {\it spatial scales} of recurrent and feedforward connectivity. We close by showing how these ideas, together with plasticity rules, can help to close the loop between network structure and activity statistics.
\section{Introduction}
Here, we focus on relating network connectivity to collective activity at the level of spike times, or {\it correlations} in the spike counts of cells on the timescale of typical synapses or membranes (See {\bf Box 1}). Such correlations are known to have complex but potentially strong relations with coding in single neurons \cite{dettner_temporal_2016} and neural populations~\cite{hu_sign_2014,moreno-bote_information-limiting_2014,zylberberg_direction-selective_2016, franke_structures_2016},
and can modulate the drive to a downstream population \cite{kumar_spiking_2010}.
Moreover, such correlated activity can modulate the evolution of synaptic strengths through spike timing dependent plasticity (STDP)(~\cite{markram_history_2011, ocker_self-organization_2015, tannenbaum_shaping_2016}, but see~\cite{graupner_natural_2016}).
Collective spiking arises from two mechanisms: connections among neurons within a population, and external inputs or modulations affecting the entire population~\cite{cohen_measuring_2011, mcginley_waking_2015, doiron_mechanics_2016}. Experiments suggest that both are important. Patterns of correlations in cortical micro-circuits have been related to connection probabilities and strengths~\cite{cossell_functional_2015}. At the same time, latent variable models of dynamics applied to cortical data have revealed a strong impact of global inputs to the population~\cite{ecker_state_2014, rosenbaum_spatial_2017, goris_partitioning_2014, ecker_structure_2016}.
At first, the path to understanding these mechanisms seems extremely complicated. Electron microscopy (EM) and allied reconstruction methods promise connectomes among thousands of nearby cells, tabulating an enormous amount of data ~\cite{lee_anatomy_2016, kasthuri_saturated_2015, bock_network_2011, kleinfeld_large-scale_2011, briggman_wiring_2011, helmstaedter_connectomic_2013, mishchenko_ultrastructural_2010}. This begs the question of what \emph{statistics} of connectivity matter most -- and least -- in driving the important activity patterns of neural populations. The answer would give us a set of meaningful ``features'' of a connectome that link to basic statistical features of the dynamics that such a network produces. Our aim here is to highlight recent theoretical advances toward this goal.
\section{Mechanisms and definitions: sources and descriptions of (co)variability in spike trains}
\begin{figure}[h!]
\begin{mdframed}
\center{ \includegraphics[width=4 in]{spike_fig2.png}}
\caption*{ \small
{\bf Box 1. Spike train statistics.} The spike train of neuron $i$ is defined as a sum of delta functions,
$y_i(t) = \sum_\mathrm{k} \delta \left(t - t_i^k\right).$ Spike train statistics can be obtained from samples of the spike trains of each neuron in a population. A joint moment density of $n$ spike trains is defined as a trial-average of products of those spike trains:
\begin{align} \label{E:moments}
\mathbf{m}_{i, j, \ldots, n}(t_i, \ldots, t_n) = \frac{1}{N} \sum_{\mathrm{trials}} \prod_{i=1}^n y_i(t_i),
\end{align}
where $N$ is the number of trials. In practice, time is discretized into increments of size $\Delta t$, and spike trains are binned. Equation~\eqref{E:moments} is recovered from its discrete counterpart in the limit $\Delta t \rightarrow 0$.
The first spike train moment is the instantaneous firing rate of a neuron $i$. The second spike train moment, $\mathbf{m}_{ij}(t_i, t_j),$ is the correlation function of the two spike trains; if $i=j$ it is an autocorrelation, otherwise a cross-correlation. It is frequently assumed that the spike trains are stationary, so that their statistics do not depend on time. We can then replace averages over trials with averages over time. Additionally, the correlation in this case only depends on the {\it time lag} in between spikes. This yields the spike train cross-correlation as,
\begin{align}
\mathbf{m}_{i, j, \ldots, n}(s_j, \ldots, s_n) = \frac{1}{T}\int_0^T dt_i \; \prod_{j=i+1}^n y_j(t_i + s_j),
\end{align}
where $s_j = t_j - t_i$ for $j=i+1, \ldots, n$ and $s_i$=0. The correlation function measures the frequency of spike pairs. Two uncorrelated Poisson processes with rates $r_i$ and $r_j$ have $\mathbf{m}_{ij}(s) = r_i r_j$, independent of the time lag $s$. The statistics of any linear functional of the spike trains (such as output spike counts, or synaptic outputs or inputs) can be derived from these spike train statistics \cite{rosenbaum_spatial_2017, ocker_linking_2016}.
Moments mix interactions of different orders. To account for lower-order contributions, we can define \emph{cumulants} of the spike trains. The first cumulant and the first moment both equal the instantaneous firing rate. The second cumulant is the covariance function of the spike train: $\mathbf{C}_{ij}(s) = \mathbf{m}_{ij}(s) - r_ir_j$. The third spike train cumulant similarly measures the frequency of triplets of spikes, above what could be expected by composing those triplets of individual spikes and pairwise covariances. Higher order cumulants have similar interpretations~\cite{novak_three_2012}.
}
\end{mdframed}
\end{figure}
Neurons often appear to admit spikes stochastically. Such variability can be due to noise from, e.g., synaptic release~\cite{faisal_noise_2008}, and can be internally generated via a chaotic ``balanced'' state ~\cite{van_vreeswijk_chaos_1996,van_vreeswijk_chaotic_1998,renart_asynchronous_2010}.
As a consequence, the structure of spike trains is best described statistically. The most commonly used statistics are the instantaneous firing rate of each neuron, the autocorrelation function of the spike train (the probability of observing pairs of spikes in a given cell separated by a time lag $s$),
and the cross-correlation function (likewise, for spikes generated by two different cells). As shown in {\bf Box 1}, even weak correlations yield coherent, population-wide fluctuations in spiking activity that can have a significant impact on cells downstream \cite{kumar_spiking_2010}.
Similarly, higher-order correlations are
related to the probability of observing triplets, quadruplets or more spikes in a group of neurons, separated by a given collection of time lags. All these quantities correspond to {\it moment densities} of the spiking processes ({\bf Box 1}).
\section{Spike train covariability from recurrent connectivity and external input}
\label{sec:linearResponse}
In recent years, neuroscientists have advanced a very general framework for predicting how spike train correlations (more specifically, cumulants; {\bf Box 1}) depend on the structure of recurrent connectivity and external input.
This framework is based on {\it linearizing the response} of a neuron around a baseline state of irregular firing. For simplicity we present the result as a matrix of spike train \emph{auto-} and \emph{cross-spectra}, $\mathbf{C}(\omega)$; this is the matrix of the Fourier transforms of the familiar auto- and cross-covariance functions. Under the linear response approximation, this is
\begin{align} \label{eq:TreeCov}
\mathbf{C}(\omega)
= \underbrace{\mathbf{\Delta}(\omega) \mathbf{C}^0(\omega)\mathbf{\Delta}^*(\omega)}_\mathrm{Internally \; generated} + \underbrace{\mathbf{\Delta}(\omega)\left(\mathbf{A}(\omega)\mathbf{C}^\mathrm{ext}(\omega)\mathbf{A}^*(\omega) \right)\mathbf{\Delta}^*(\omega)}_\mathrm{Externally \; applied}.
\end{align}
The auto- (cross-) spectra correspond to diagonal (off-diagonal) terms of $\mathbf{C}(\omega)$. In particular, evaluated at $\omega = 0$, these terms give the variance and co-variance of spike counts over any time window large enough to contain the underlying auto and cross-correlation functions~\cite{rocha07,bair01}. The matrix $\mathbf{C}^0(\omega)$ is diagonal, and $\mathbf{C}^0_{kk}(\omega)$ is the power spectrum of neuron $k$'s spike train in the baseline state, without taking into account the effect of the recurrence on its spiking variance. $\mathbf{\Delta}^*$ denotes the conjugate transpose of $\mathbf{\Delta}$.
Here, the $ij^{th}$ entry of the matrix $\mathbf{\Delta}(\omega)$ is called a \emph{propagator}, and reflects how a spike in neuron $j$ propagates through the network to affect activity in neuron $i$. When the activity can be linearized around a stable ``mean-field'' state, the matrix of propagators obeys:
\begin{align} \label{eq:propagator}
\mathbf{\Delta}(\omega) = \big(\mathbf{I} -\mathbf{K}(\omega) \big)^{-1},
\end{align}
where $\mathbf{K}(\omega) = \mathbf{A}(\omega)\mathbf{W}$ is the interaction matrix, which, importantly, encodes weights of synaptic connections $\mathbf{W}_{ij} $ between neurons $j$ and $i$. In general, the connection matrix can have a time dependence corresponding to the filtering and delay of synaptic interactions, and be written as $\mathbf{W}(\omega)$, but we suppress this $\omega$-dependence for ease of notation.
$\mathbf{A}$ is a diagonal matrix; $\mathbf{A}_{ii}(\omega)$ is the linear response (Fourier transform of its impulse response, or PSTH~\cite{gabbiani_mathematics_2010}) of neuron $i$ to a perturbation in its synaptic input.
The first term of Eq.~\eqref{eq:TreeCov} thus measures how the variability in the spiking of each neuron propagates through the network to give rise to co-fluctuations in pairs of neurons downstream. These correlations thus generated from within the population; the second term in Eq. \eqref{eq:TreeCov}
captures correlations generated by external inputs.
The external inputs are described by their power spectrum, $\mathbf{C}^\mathrm{ext}_{ij}(\omega)$. A global modulation in the activity of many neurons due to shifts in attention, vigilance state, and/or motor activity, would result in low-rank matrix $\mathbf{C}^\mathrm{ext}$. In this case the second, external term of Eq. \eqref{eq:TreeCov} will itself be low rank, since the rank of a matrix product $AB$ is bounded above by the ranks of $A$ and $B$. Experimentally obtained spike covariance matrices can be decomposed into a low-rank and ``residual'' terms~\cite{yatsenko_improved_2015,ecker14} that correspond to the two terms in the matrix decomposition Eq.~\eqref{eq:TreeCov}.
In the simplest case of an uncoupled pair of neurons $i$ and $j$ receiving common inputs, Eq.~\eqref{eq:TreeCov} reduces to $
\mathbf{C}_{ij}(\omega) = \mathbf{A}_i(\omega) \mathbf{C}^\mathrm{ext}_{ij}(\omega) \mathbf{A}_j(-\omega)$ \cite{de_la_rocha_correlation_2007,shea-brown_correlation_2008}. The covariance of the two spike trains is thus given by the input covariance, multiplied by the gain with which each neuron transfers those common inputs to its output.
Eq.~\eqref{eq:propagator} can be expanded in powers of the interaction matrix $\mathbf{K}(\omega)$ as
\begin{align}
\mathbf{\Delta}(\omega) = \sum_{m=0}^\infty \mathbf{K}^m(\omega).
\end{align}
This expansion has a simple interpretation: $\mathbf{K}^m_{ij}(\omega)$ represents paths from a neuron $i$ to neuron $j$ that are exactly $m$ synapses long (with synaptic weights $\mathbf{W}$ weighted by the postsynaptic response gain $\mathbf{A}(\omega)$)~\cite{pernice_how_2011,trousdale_impact_2012}. Using this expansion in Eq.~\eqref{eq:TreeCov}, without external input, also provides an intuitive description of the spike train cross-spectra in terms of paths through the network,
\begin{align} \label{eq:expandK}
\mathbf{C}_{ij}(\omega) \approx \sum_{k} \sum_{m=0}^X\sum_{n=0}^Y \underbrace{\Big(\mathbf{K}^m(\omega)\Big)_{ik} \Big(\mathbf{K}^n (-\omega) \Big)_{jk}}_{\text{path terms}} \mathbf{C}^0_{kk}(\omega).
\end{align}
This expression explicitly captures contribution to the cross-spectrum, $\mathbf{C}_{ij}(\omega),$ of all paths of up to $X$ synapses ending at neuron $i$, and all paths of up to $Y$ synapses ending at neuron $j$. The index $k$ runs over all neurons. A first step toward the path expansion was taken by Ostojic, Brunel \& Hakim, who explored the first-order truncation of Eq.~\eqref{eq:TreeCov} for two cells, thus capturing contributions from direct connections and common inputs ~\cite{ostojic_how_2009}.
The formulation of Eq.~\eqref{eq:TreeCov} has a rich history. Sejnowski used a similar expression in describing collective fluctuations in firing rate networks \cite{sejnowski_stochastic_1976}. Linder, Doiron, Longtin and colleagues derived this expression in the case of stochastically driven integrate and fire neurons \cite{doiron_oscillatory_2004,lindner_theory_2005}, an approach generalized by Trousdale et al.~\cite{trousdale_impact_2012}. Hawkes derived an equivalent expression for the case of linearly interacting point processes (now called a multivariate Hawkes process), as pointed out by Pernice et al., who applied and directly related them to neural models \cite{hawkes_spectra_1971,pernice_how_2011,pernice_recurrent_2012}. The correspondence of the rate dynamics of the Hawkes model, networks of integrate-and-fire neurons, and binary neuron models, was discussed in detail in~\cite{grytskyy_unified_2013} and the direct approximation of integrate-and-fire neurons by linear-nonlinear-Poisson models in~\cite{ostojic_spiking_2011}.
Moreover, Buice and colleagues \cite{buice_systematic_2010} developed a field theoretical method that encompasses the above approach and extends the formulation to correlations of arbitrary order. Importantly, this also allows for nonlinear interactions. An expansion can be derived via this method that describes the coupling of higher order correlations to lower moments: in particular, pairwise correlations can impact the activity predicted from mean field theory~\cite{ocker_linking_2016}. The field theoretic approach has also been applied to models of coupled oscillators related to neural networks, specifically the Kuramoto model \cite{hildebrand_kinetic_2007, buice_correlations_2007} and networks of ``theta" neurons \cite{buice_dynamic_2013}. Similarly, Rangan developed a motif expansion of the operator governing the stationary dynamics of an integrate-and-fire network \cite{rangan09a}.
\section{Network motifs shape collective spiking across populations}
The relationship of spike train cross-spectra to pathways through the network provides a powerful tool for understanding how network connectivity shapes the power spectrum of the population-wide network activity ({\bf Box 1}).
The population power spectrum is given by the average over the cross-spectral matrix: $C(\omega) = \left \langle \mathbf{C}_{ij}(\omega) \right \rangle_{i,j}$ (the angle brackets denote averaging over pairs of neurons within a given network).
Therefore $C(\omega)$ is the average of the left-hand side of Eq.~\eqref{eq:expandK}. We show that the right-hand side of Eq.~\eqref{eq:expandK} can in turn be linked to the \emph{motif moments} which describe the mean strength of different weighted microcircuits in the network.
Assuming cellular response properties are homogeneous, the interaction matrix can be written as $\mathbf{K}(\omega) = A(\omega) \mathbf{W}$, where the scalar kernel $A(\omega)$ is the same response kernel for all cells.
If the baseline auto-correlations, hence $\mathbf{C}^0_k$, are also equal across the network, the ``path terms'' appearing in Eq.~\eqref{eq:expandK} are directly proportional to the {\it motif moments} of the connectivity matrix $\mathbf{W},$ defined as:
\begin{align}
\mu_{m,n} = \left \langle \mathbf{W}^m \left(\mathbf{W}^T\right)^n \right \rangle_{i,j} / N^{m+n+1}.
\label{eq:def_motif_moment}
\end{align}
This measures the average strength of a $(m,n)$--motif composed of two paths of synapses emanating from a neuron $k$ with one path of $m$ synapses ending at neuron $i$, the other path of $n$ synapses ending at neuron $j$.
Examples of a $(1,1)$-, and $(1,2)$-motif are shown in Fig.~\ref{fig:motif}A. For networks where $\mathbf{W}_{ij}= 0$ or $1$, $\mu_{m,n}$ is also the frequency of observing a motif in the network. Eq.~\eqref{eq:expandK} thus provides a way to approximate how average correlations depend on the frequency of motifs in the network.
\begin{figure}
\centering \includegraphics[width=6 in]{motifs_schematic4}
\caption{(A) Various motifs are identified throughout the neural network. Their frequency can be measured by counting their occurrence. (B) The probability of a motif (motif moments $\mu_{m,n}$, see text) can be decomposed into cumulants of smaller motifs. (C) Comparing the average correlation between excitatory neurons calculated using motif statistics of all orders ($X,Y\rightarrow \infty$ in Eq.~\eqref{eq:expandK}) and approximations using up to second order ($X+Y\leq 2$) for 512 networks (different dots) with various motif structures. Each network is composed of 80 excitatory and 20 inhibitory neurons. The deviation from the dashed line $y=x$ shows that motif moments beyond second order are needed to accurately describe correlations. (D) Same as (C) for a motif cumulant expansion (Eq.~(44) in \cite{hu_motif_2013}, a generalization of the motif cumulant theory for networks with multiple neuron populations) truncated after second order motif terms; both panels adapted from~\cite{hu_motif_2013}.} \label{fig:motif}
\end{figure}
An accurate approximation of the population power spectrum typically requires keeping many terms in the series (Fig.~\ref{fig:motif}C). This is a challenge for applications to real neural networks, where the statistics of motifs involving many neurons are much harder to determine. Importantly, however, the contribution of these higher order motifs can often be decomposed into contributions of smaller, component motifs by introducing~\emph{motif cumulants} (Fig.~\ref{fig:motif} B) \cite{hu_motif_2013,hu_local_2014}. This approach allows us to remove redundancies in motif statistics, and isolate the impact solely due to higher order motif structure.
This motif cumulant expansion allows Eq.~\eqref{eq:expandK} to be re-arranged in order to only truncate higher-order motif cumulants, rather than moments, providing a much-improved estimation of the spike train covariances (Fig. \ref{fig:motif}C vs D, \cite{hu_motif_2013}) that is still based only on very small motifs.
In sum, population-averaged correlation can be efficiently linked to \emph{local connectivity structures} described by motif cumulants. Significant motif cumulant structure exists in local networks of both cortex \cite{song_highly_2005,perin_synaptic_2011} and area CA3 of the hippocampus, where they play a crucial role in pattern completion \cite{guzman_synaptic_2016}. Moreover, as we see in Sec.~\ref{sec:plasticity}, correlations in turn can shape a network's motif structure through synaptic plasticity.
\paragraph*{Higher-order correlations and network structure}
While we have discussed how network structure gives rise to correlations in pairs of spike trains, joint activity in larger groups of neurons, described by higher-order correlations, can significantly affect population activity \cite
ohiorhenuan_sparse_2010, shimazaki_state-space_2012,
tkacik_searching_2014}. Analogous results to Eq.~\eqref{eq:TreeCov} exist for higher-order correlations in stochastically spiking models \cite{buice_systematic_2010, buice_beyond_2013, jovanovic_cumulants_2015, ocker_linking_2016}. Network structure has been linked to the strength of third-order correlations in networks with narrow degree distributions \cite{jovanovic_interplay_2016}, and the allied motif cumulant theory developed~\cite{HuCosyne15}, advancing the aim of understanding how local connectivity structure impacts higher-order correlations across networks. Finally, a range of work has characterized the statistics of avalanches of neuronal activity: population bursts with power law size distributions \cite{plenz_organizing_2007}, which potentially suggest a network operating near an instability \cite{BuiceC07}.
\paragraph*{Motifs and stability}
The results discussed so far rely on an expansion of activity around a baseline state where neurons fire asynchronously. How such states arise is a question addressed in part by the mean-field theory of spiking networks \cite{ginzburg_theory_1994, van_vreeswijk_chaos_1996, brunel_dynamics_2000, renart_asynchronous_2010, tetzlaff_decorrelation_2012, helias_correlation_2014}. The existence of a stable stationary state depends on the structure of connectivity between neurons. In particular, when connectivity is strong and neurons have heterogeneous in-degrees, the existence of a stable mean-field solution can be lost \cite{pyle_highly_2016, landau_impact_2016}. One way to rescue a stable activity regime is to introduce correlations between neurons' in and out-degrees \cite{pyle_highly_2016}; these correspond to chain motifs ($\kappa_{2}$ in Fig. 1B). Therefore the motifs that control correlated variability also affect the stability of asynchronous balanced states.
Motif structure also affects oscillatory population activity. Roxin showed that in a rate model generating oscillations of the population activity, the variance of in-degrees (related to the strength of convergent motifs) controls the onset of oscillations \cite{roxin_role_2011}. Zhao et al. took a complementary approach of examining the stability of completely asynchronous and completely synchronous states, showing that two-synapse chains and convergent pairs of inputs regulate the stability of completely synchronous activity \cite{zhao_synchronization_2011}.
\section{Spatial scale of connectivity and inputs determines correlations in large networks}
Cortical neurons receive strong external and recurrent excitatory projections that would, if left unchecked, drive neuronal activity to saturated levels. Fortunately, strong recurrent inhibition {\it balances} excitation, acting to stabilize cortical activity and allow moderate firing. These large and balanced inhibitory and excitatory inputs are a major source of synaptic fluctuations, ultimately generating output spiking activity with Poisson-like variability \cite{doiron_balanced_2014,deneve2016}.
A central feature of balanced networks is that they produce {\it asynchronous}, uncorrelated spiking activity (in the limit of large networks). Original treatments of balanced networks by van Vreeswijk and Sompolinsky \cite{van_vreeswijk_chaos_1996,van_vreeswijk_chaotic_1998} and Amit and Brunel \cite{amit97} explained asynchronous activity by assuming sparse wiring within the network, so that shared inputs between neurons were negligible. While the connection probability between excitatory neurons is small~\cite{oswald09,Holmgren03,lefort2009}, there is abundant evidence that connections between excitatory and inhibitory neurons can be quite dense \cite{fino_dense_2011,oswald09}. Renart, de la Rocha, Harris and colleagues showed that homogeneous balanced networks admit an asynchronous solution despite dense wiring \cite{renart10} (for large networks). This result suggests a much deeper relationship between balance and asynchronous activity than previously realized. Building on this work, Rosenbaum, Doiron and colleagues extended the theory of balanced networks to include spatially dependent connectivity \cite{rosenbaum_balanced_2014,rosenbaum2017}. We review below how the spatial spread of connectivity provides new routes to correlated activity in balanced networks.
Consider a two-layer network, with the second layer receiving both feedforward ($F$) and recurrent ($R$) inputs (Fig. \ref{Fig_balanced_space}A). For simplicity we assume that the feedforward and recurrent projections have Gaussian profiles with widths $\sigma_{\textrm{F}}$ and $\sigma_{\textrm{R}}$, respectively. Each neuron receives the combined input $I=F+R$, decomposing the input covariance $C_{II}(d)$ to a representative pair of layer two neurons separated by a distance $d$ as:
\begin{equation}\label{E:balanced_asych}
\begin{aligned}
C_{II}(d)=C_{FF}(d)+C_{RR}(d)+2C_{RF}(d).
\end{aligned}
\end{equation}
Here $C_{FF}$ and $C_{RR}$ are the direct covariance contributions from feedforward and recurrent pathways, respectively, while $C_{RF}$ is the indirect contribution to covariance from the recurrent pathway tracking the feedforward pathway.
If the network coupling is dense and layer one neurons are uncorrelated with one another then $C_{FF}$, $C_{RR}$ and $C_{RF}$ are all $\mathcal{O}(1)$. This means that feedforward and recurrent projections are potential sources of correlations within the network. The asynchronous state requires that $C_{II}$ $\sim$ $\mathcal{O}(1/N)$. This can only be true if the feedforward and recurrent correlations are {\it balanced} so that the recurrent pathways tracks and cancels the correlations due to the feedfoward pathway. If we take $N \to \infty$ then in the asynchronous state, $C(d) \to 0$ implies:
\begin{equation}\label{E:C_RF}
\begin{aligned}
C_{RF}(d) = -\frac{1}{2} \left (C_{FF}(d)+C_{RR}(d) \right ).
\end{aligned}
\end{equation}
This must be true for {\it every distance $d$}, and from Eq. \eqref{E:balanced_asych} we derive \cite{rosenbaum2017} that the various spatial scales must satisfy:
\begin{equation}\label{E:balanced_sigmas}
\begin{aligned}
\sigma^2_{\textrm{F}}=\sigma^2_{\textrm{R}}+\sigma^2_{\textrm{\textrm{rate}}}.
\end{aligned}
\end{equation}
Here $\sigma^2_{\textrm{\textrm{rate}}}$ is the spatial scale of correlated firing within the network. The intuition here is that for cancellation at every $d$ the spatial scale of feedforward and recurrent correlations must match one another. While the spatial scale of $C_{FF}(d)$ is determined only by $\sigma_{F}$, the scale of recurrent correlations is calculated from the correlated spiking activity convolved with the recurrent coupling (hence the sum $\sigma^2_{\textrm{R}}+\sigma^2_{\textrm{\textrm{rate}}}$). While $\sigma_{F}$ and $\sigma_{R}$ are architectural parameters of the circuit (and hence fixed), $\sigma_{\textrm{rate}}$ is a model output that must be determined. For any solution to make sense we require that $\sigma_{\textrm{rate}} >0$. This gives a compact asynchrony condition: $\sigma_{F} > \sigma_{R}$. In other words for feedforward and recurrent correlations to cancel, the spatial spread of feedforward projections must be larger than the spatial spread of recurrent projections.
\begin{figure}[h!]
\vspace{-1mm}
\centering
\includegraphics[width=6.5in]{balanced_space.pdf}
\caption{Correlated activity in balanced networks with spatially dependent connections. (A) Schematic of the two layer network. The blue/red zones denotes the spatial scale of feedforward/recurrent connectivity. (B1) Broad feedforward and narrow recurrent connectivity (left) produce an asynchronous state (middle). The asynchrony requires a cancellation of $C_{RR}(d)$ and $C_{FF}(d)$ by $C_{RF}(d)$ at all distances (right). (B2) Narrow feedforward and broad recurrent connectivity (left) produce spatially structured correlations (middle) because $C_{RF}(d)$ does not cancel $C_{RR}(d)$ and $C_{FF}(d)$ (right). (C) When a one dimensional latent variable is extracted and removed from the primate V1 array data, the model predictions (left) are validated (right). Panels are from \cite{rosenbaum2017}.}
\label{Fig_balanced_space}
\end{figure}
To illustrate how the spatial scales of connectivity control the asynchronous solution, we analyze the activity of a balanced spiking network when $\sigma_{F} > \sigma_{R}$ is satisfied (Fig. \ref{Fig_balanced_space}B1, left). As expected, the spiking activity is roughly asynchronous with spike count correlations near zero (Fig. \ref{Fig_balanced_space}B1, middle). When we examine the contributions of the feedforward and recurrent pathways, we see that the relation in Eq. \eqref{E:C_RF} is satisfied (Fig. \ref{Fig_balanced_space}B1, right). We contrast this to the case when $\sigma_{F} < \sigma_{R}$, violating the asynchrony condition (Fig. \ref{Fig_balanced_space}B2, left). Here, a clear signature of correlations is found: neuron that are nearby one another are positively correlated while more distant neuron pairs are negatively correlated. Indeed, the cancellation condition Eq. \eqref{E:C_RF} is violated at almost all $d$ (Fig. \ref{Fig_balanced_space}B2, right).
Layer 2/3 of macaque visual area V1 is expected to have $\sigma_{F} < \sigma_{R}$, with long range projections within 2/3 being broader than L4 projections to L2/3 \cite{bosking1997,lund_anatomical_2003}. Smith and Kohn collected population activity over large distances in layer 2/3 of macaque V1 \cite{smith08}. When a one dimensional source of correlations is removed from the data then the model prediction is supported (Fig. \ref{Fig_balanced_space}C).
While these arguments give conditions for when the asynchronous solution will exist, it cannot give an estimate of correlated activity. Rather, when either $N < \infty$ or $\sigma_{F} < \sigma_{R}$, we require the linear response formulation of Eq.~\eqref{eq:TreeCov} to give a prediction of how network correlations scale with distance (red dashed in Fig. \ref{Fig_balanced_space}B1 and B2, middle). It is interesting to note that a majority of the fluctuations are internally generated within the balanced circuit and subsequently have a rich spectrum of timescales. For spikes counted over long windows, the linear response formalism of Eq.~\eqref{eq:TreeCov} once again predicts the underlying correlations (``theory" curves).
\section{Joint activity drives plasticity of recurrent connectivity}
\label{sec:plasticity}
The structure of neuronal networks is plastic, with synapses potentiating or depressing in a way that depends on pre- and postsynaptic spiking activity~\cite{feldman_spike-timing_2012}. When synaptic plasticity is slow compared to the timescales of spiking dynamics, changes in synaptic weights are linked to the {\it statistics} of the spike trains \cite{kempter_hebbian_1999}: specifically, the joint moment densities of the pre- and postsynaptic spike trains~\cite{gerstner_mathematical_2002} ({\bf Box 1}). For plasticity rules based on pairs of pre- and postsynaptic spikes, this results in a joint evolution of the weight matrix $\mathbf{W},$ the firing rates, $\vec{r}$, and the cross-covariances, $\mathbf{C}(s)$ (Fig. \ref{fig:plasticity}A-C).
As we saw above, spike train cumulants depend on the network structure. In the presence of plasticity mechanisms, the structure of neuronal networks thus controls its own evolution -- both directly by generating correlations \cite{ocker_self-organization_2015,tannenbaum_shaping_2016} and indirectly, by filtering the correlations inherited from external sources \cite{gilson_emergence_2009-1, ocker_training_2016}. Recent work has leveraged this connection to determine how particular structural motifs shape spike train correlations to drive plasticity~\cite{tannenbaum_shaping_2016}. A further step is to close the loop on motifs, leveraging approximations of the true spike-train correlations in order to predict the plasticity dynamics of motif cumulants \cite{ocker_self-organization_2015, ocker_training_2016}:
\begin{align}
\frac{d}{dt} \vec{\kappa} = F\left(\mathbf{m}\left(\vec{\kappa}, \mathbf{C}^\mathrm{ext}\right)\right),
\end{align}
where $\vec{\kappa}$ represents a chosen set of motif cumulants and the form of the function $F$ depends on the plasticity model used. Such analysis reveals that under an additive, pair-based plasticity rule where pre-post pairs cause potentiation and post-pre pairs cause depression, an unstructured weight matrix (with zero motif cumulants) is unstable: motifs will spontaneously potentiate or depress, creating structure in the synaptic weights (Fig. \ref{fig:plasticity}D),~\cite{ocker_self-organization_2015}. So far, such studies have focused on plasticity driven by spike pairs, relying on the linear response theories of \cite{Hawkes71a, pernice_how_2011, trousdale_impact_2012}. More biologically realistic plasticity models rely on multi-spike interactions and variables measuring postsynaptic voltage or calcium concentrations \cite{markram_history_2011}.
Theories describing higher-order spike-train and spike-voltage or spike-calcium correlations provide a new window through which to examine networks endowed with these richer plasticity rules.
\begin{figure}[h!]
\vspace{-1mm}
\centering
\includegraphics[width=6.5in]{plasticity_fig.pdf}
\caption{Spike timing-dependent plasticity gives rise to joint evolution of synaptic weights, spike train covariances, and motif statistics. Panels from Ocker et al., 2015. A) Diagram of network structure. B) Evolution of the synapses highlighted in panel A. B) Evolution of the spike train covariances between the pre- and postsynaptic cells of each synapse, predicted using the first-order truncation of Eq. \eqref{eq:expandK}. Shading corresponds to the time points marked with arrows below the time axis of panel B. D) Projection of the joint dynamics of two-synapse motifs into the (divergent, chain) plane under a pair-based, additive Hebbian plasticity rule.}
\vspace{-4mm}
\label{fig:plasticity}
\end{figure}
\section{Conclusion}
The next few years could be pivotal for the study of how network structure drives neural dynamics. Spectacular experimental methods are producing vast datasets that unite connectivity and activity data in new ways~\cite{ cossell_functional_2015, bock_network_2011, lee_anatomy_2016, briggman_wiring_2011}. Does our field have a theory equal to the data? We've reviewed how mathematical tools can separate the two main mechanisms giving rise to collective spiking activity -- recurrent connectivity and common inputs -- and how this connects with decomposition methods in data analysis ~\cite{ecker_state_2014,goris_partitioning_2014,rosenbaum_spatial_2017,yatsenko_improved_2015}. Moreover, we are beginning to understand how local and spatial structures scale up to global activity patterns, and how they drive network plasticity. Daunting challenges remain, from the role of myriad cell types to the impact of nonlinear dynamics. As the bar ratchets up ever faster, the field will be watching to see what theories manage to clear it.
\section*{Acknowledgements }
We acknowledge the support of the NSF through grants DMS-1514743 and DMS-1056125 (ESB), DMS-1313225 and DMS-1517082 (B.D.), DMS-1517629 (KJ), DMS-1517828 (R.R.) and NSF/NIGMS-R01GM104974 (KJ), as well as the Simons Fellowships in Mathematics (ESB and KJ). GO, MB, and ESB wish to thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their vision, encouragement and support.
\newpage
\section{Highlighted References}
\begin{itemize}
\item ** Pernice et al., PLOS Comp. Biol., 2011. Calculates the full matrix of pairwise correlation functions either in networks of interacting Poisson (Hawkes process) neurons. Shows how network-wide correlation is related to connectivity paths and motifs.
\item *Trousdale et al., PLOS Comp. Biol, 2012. Calculates the linear response predictions for the full matrix of pairwise correlation functions, and their relationship to connectivity paths, in networks of integrate-and-fire neurons.
\item ** Jovanovic \& Rotter, PLOS CB 2017. Building on their PRE 2015 paper, calculates the linear response predictions for the $n$-th order cumulant densities of multivariate Hawkes processes (i.e. linear-Poisson neural networks), and shows how these are related to connectivity in regular networks (those with narrow degree distributions).
\item ** Ocker et al, ArXiV 2016. Extends existing theories for spike correlations to allow for nonlinear interactions among inputs and firing rates, deriving series of ``diagrams'' that show how connectivity influences activity statistics for quadratic and higher orders firing rate responses.
\item ** Hu et al., J Stat Mech 2013. Decomposes the effects of higher order motifs on spike correlations into smaller ones, and show how this {\it motif cumulant} approach enables one to predict global, population-wide correlations from the statistics of local, two-synapse motifs.
\item * Hu et al., PRE 2014 . Formulates network motif decompositions via a combinatorial relationship between moments and cumulants, extends the theory to multi-branch motifs related to higher order correlations, and to the impact of heterogeneous connectivity.
\item ** Renart, de la Rocha et al. Science 2010. Shows that for balanced networks with dense and strong connectivity, an asynchronous state emerges in the large-N limit in which excitatory-inhibitory interactions dynamically cancel excitatory-excitatory and inhibitory-inhibitory correlations. This leads to spike correlations that vanish on average. Also in the large N limit, this arises from population responses that instantaneously and linearly track external inputs.
\item ** Tetzlaff, Helias et al., PLoS CB 2012. A highly complementary study to the more well-known work of Renart et al. Using linear response theory for finite-size integrate-and-fire networks, Tetzlaff, Helias et al. expose negative feedback loops in the dynamics of both purely inhibitory and excitatory-inhibitory networks, which give rise to the dynamical cancellation of correlations in finite-size networks.
\item * Helias et al., PLoS CB 2014. Building on the study of Renart, de la Rocha et al. and on their previous work, the authors disentangle correlation cancellation by excitatory-inhibitory interactions (reflected in the suppression of fluctuations in the population activity) from the tracking of external inputs.
\item ** Rosenbaum et al., Nature Neuroscience 2016. Building on their 2014 work in PRX, the authors extends the theory of correlations in balanced networks to systems with spatially dependent connectivity, showing that mismatches in the spatial scales of feedforward and recurrent inputs can give rise to stable average firing rates but significant spike correlations.
\item *Rangan, PRL 2009. Develops a diagrammatic expansion for the statistics of stationary integrate-and-fire networks with delta synapses in terms of subnetworks (i.e. motifs) via a functional representation of the full network dynamics. (See also the accompanying PRE article)
\item ** Zhao et al, Frontiers in Comp. Neuroscience, 2011. A pioneering study of how network motifs impact network-wide synchrony, including simulations, analytical results, and methods of generating useful networks to test theories linking structure to activity.
\item ** Gilson et al., Biol. Cyb. 2009iv. Building on the authors' previous work on STDP in networks of Poisson neurons (Gilson et al., Biol. Cyb. 2009i-iii), uses a combination of the Hawkes theory and a mean-field calculation of synaptic weights to show that externally-generated correlations can give rise to selective connectivity in recurrent networks of Poisson neurons through spike timing-dependent plasticity.
\item ** Ocker et al., PLoS CB 2015. Uses the linear response theory for integrate-and-fire neurons to construct self-consistent predictions for internally-generated correlations and synaptic plasticity, and then derive a reduced dynamics for the spike timing-dependent plasticity of 2-synapse motif cumulants. This shows that an initially unstructured (Erd\H{o}s-R\`enyi) network connectivity is unstable under additive, Hebbian STDP unless all synapses potentiate or depress together.
\end{itemize}
|
\section{Introduction}
\section{Proposed method}
\subsection{Pencil beam based treatment}
Treatment plans usually involve more the one proton field in order to guarantee
dose homogeneity over the target volume and to increase the stability of the plan.
The active scanning technique \cite{Pedroni1995} allows, in particular, to control the dose
deposition of each focused pencil beam within the patient volume; this implies that a field is conceived
as the collection of all the single dose spots delivered for a specific position of the patient with respect to the nozzle.
Thank to this peculiarity, the active scanning enables the construction of complex-shaped dose distributions.
Once the number and direction of the proton fields are selected for a treatment, there are two different
planning strategies implemented for active scanning based gantries: \emph{single field uniform dose (SFUD)} \cite{Scheib1993}, implying
a superposition of the individually optimized dose distributions; \emph{intensity modulated proton therapy
(IMPT)} \cite{Lomax1999}, consisting in a simultaneous optimization of all the fields. The latter modality can often provide better
tradeoffs between the dose coverage of the target and the sparing of the organs at risk.
\subsection{IMPT optimization}
The IMPT optimization consists in the following problem \cite{Albertini2011}:
\begin{equation}
\boldsymbol{\omega} = \argmin_{\boldsymbol{\omega}} F(\boldsymbol{\omega})
\end{equation}
\begin{equation}
F(\boldsymbol{\omega}) = \sum\limits_{i_{t}}^{}\left(P_{i_{t}}-D_{i_{t}}(\boldsymbol{\omega})\right)^{2} +
\sum\limits_{i_{o}}^{}g_{i_{o}}^{2}\:\left(P_{i_{o}}-D_{i_{o}}(\boldsymbol{\omega})\right)^{2}
\label{impt-functional}
\end{equation}
where the first sum runs on the voxel indices of the target $\{i_{t}\}$ and the second one on the voxel indices of the OARs
$\{i_{o}\}$, $P_{j}$ and $D_{j}$ are respectively the prescribed and the computed dose for the $j$-th voxel, $\{g_{i_{o}}\}$'s are
weights chosen by the planner and $\boldsymbol{\omega} = \{\omega_{1},\cdots,\omega_{N}\}$ corresponds to the fluences
of the pencil beams. The iterative minimizer of (\ref{impt-functional}) is with the OAR dose constraints,
applied through the weights $\{g_{i_{o}}\}$ and with $\boldsymbol{\omega}^{(0)}$, such that each field is characterized by a flat
spread-out bragg peak (SOBP) \cite{Bortfeld1996} within the target volume. Since
the number and direction of the fields have been already fixed and the pre-optimized fluences $\boldsymbol{\omega}^{(0)}$
are automatically computed on the basis of the target volume, the dose constraints are the only leverage left at disposal of the
planner to modify the resulting dose distribution.
\subsection{Opacity concept}
The \emph{opacity} is defined as a quality of the OARs in relation to the pencil beams and has to be considered as a simple weight
ranging in [0,1]. The opacity level (OL) of an OAR is thought to modify the pre-optimized fluences $\boldsymbol{\omega}^{(0)}$
of the pencil beams, whose trajectory
(the straight line connecting the Bragg peak position to the leaving point at the nozzle) intersects voxels of such structure.
We indicate with $\{k_{n}\} \subseteq \{1,\cdots,N\}$ the list of indices referring to the subset of pencil beams crossing the $n$-th
OAR.
If the OL = 0, the OAR is considered to be ``transparent'' and the pencil beams intersecting the structure are left unaltered,
$\{\tilde{\omega}_{k_{n}}\} = \{\omega_{k_{n}}\}$;
if the OL = 1, the OAR is completely ``opaque'' and $\{\tilde{\omega}^{(0)}_{k_{n}}\} = 0$; when OL $ = \alpha \in (0,1)$
is ``partially opaque'' and the pencil beam fluences are penalized in the following way:
\begin{equation}
\tilde{\omega}^{(0)}_{\bar{k}_{n}} = \alpha \cdot l_{\bar{k}_{n}} \cdot \omega^{(0)}_{\bar{k}_{n}}
\hspace*{0.6cm} \forall \bar{k}_{n} \in \{k_{n}\} \hspace*{0.3cm},
\label{smooth-penalization}
\end{equation}
where $l_{\bar{k}_{n}}$ corresponds to the path length travelled by the $\bar{k}_{n}$-th pencil beam
inside the $n$-th OAR. Clearly, in this latter case, the longer the path length inside the OAR, the more the initial fluence
of the pencil beam will be decreased. The three cases are shown in Fig.\ref{opacity-levels}.
\newline
The OL's represent essentially an additional degree of freedom at disposal of the planner to change the starting conditions
and, therefore, to steer the outcome of the IMPT optimization. The dose constraints modify $\boldsymbol{\omega}^{(0)}$ according to the dose
released by the pencil beam to the OAR, whereas the OL's acts on the basis of the pencil beam trajectory in relation
to the OAR.
\newline
The OL's penalization requires the knowledge of all the voxel indices that are crossed by every pencil beam.
This computation is efficiently performed by an optimized version of the Siddon algorithm \cite{Jacobs1998}, commonly
adopted for the fast calculation of radiological paths.
\begin{figure*}[!h]
\centering
\includegraphics[width=6.0in]{fig_01.pdf}
\caption{Graphical explanation of the opacity penalization, assuming the pencil beams all characterized by
the same initial fluence $\omega_{0}$; on the left, the case of the OAR being ``transparent'';
in the middle, the case of the OAR being partially ``opaque'' and penalizing more the pencil beams
travelling more inside the OAR itself; on the right, case of the OAR completely ``opaque''.}
\label{opacity-levels}
\end{figure*}
\section{Materials}
IMPT plans of different indications and tumour sites have been re-optimized using the OL penalization
and the resulting plans were compared to the delivered ones. In the following, we offer a description of
the clinical cases that were selected for this simulation and of the analysis used to benchmark the robustness
of the dose distributions.
\subsection{Patient with metallic cage}
The first clinical case concerns a patient with an external metallic cage. Patients with neck tumours may undego surgery
before any radiotherapic treatment with protons and a cage is, therefore, required, to stabilize the head to the neck.
Since no calibration curve of the cage material is available for dose calculation, no pencil beam has to cross the metallic
structure to prevent any wrong placement of Bragg peak inside the patient volume.
\newline
The strategy followed in the nominal treatment
(PLAN-NOM) was to draw safety volumes around the metallic rods and to apply hard dose constraint to such volumes of interest (VOIs),
as shown in Fig.\ref{plan-metallic-cage}, such that the pencil beams of the four fields (F1, F2, F3, F4) involved
in the treatment were safe from hitting the rods.
\newline
The plan has been re-calculated by replacing the safety VOIs and related dose constraints with the effective contours of the rods,
set with OL = 1. In a first attempt, the direction of the fields were kept the same (PLAN-OL-1) and , then, an other plan was
generated by considering a wider angular spacing between F1-F2 and F3-F4 (PLAN-OL-2). The angular increase of the new fields
with respect to the nominal ones was of 10\textdegree.
\subsection{Fields crossing nasal cavities}
In some clinical cases, the optimal proton fields planned for a treatment cross
anatomical cavities to deliver dose to the target volume. Cavities, like the nose or the bowel, may undergo relevant
density changes between the day of the treatment and the acquisition of the CT image (used to calculate the
dose distributions for all the treatment fractions). Since the pencil beams crossing the cavity encounter a different
density object, they may be affected by not negligible range uncertanties; in particular, they will undershoot, in case
of a density increase, and overshoot, in the opposite situation.
\newline
The case of a treatment with fields crossing the nasal cavity has been taken into consideration. The field directions
of the nominal plan are shown in Fig.\ref{field-nasal-cavity-plan}; the frontal fields F1 and F2 encounter
the nasal cavity to reach the target volume.
To evaluate the stability of the IMPT plan, first, the original CT image was modified to simulate the extreme scenarios,
of the cavity being completely empty (HU = 0) and filled with mucus (HU $\simeq$ 30), as shown in Fig.\ref{nasal-cavities};
the original plan (PLAN-NOM) and the plan with OL penalization (PLAN-OL), where an OL $\in (0,1)$ was selected for the nose VOI,
were re-computed for both extreme scenarios, to evaluate the potential variation range with respect to the planning CT.
We named PLAN-NOM-H and PLAN-NOM-L, respectively, the difference between the nominal plan recomputed on the CT
with low and high density nasal cavity and the plan computed on the original CT
and we did similarly for the extreme scenario plans with OL penalization (PLAN-OL-H and PLAN-OL-L).
\subsection{Head and neck tumour}
One common indication for proton therapy centers is represented by tumours extending both in the head and neck.
As the tumour extends over a relatively big area, the treatment can result rather toxic for many organs at risk.
These cases are usually treated with four fields, as shown in Fig.\ref{head-neck-fields-1}: two coming from the front,
aimed at covering the target volume at the level of the shoulders, two coming from behind, supposed to deliver the dose
in the head part of the tumour. The drawback of this geometry is that, despite the dose constraints,
the front fields irradiate OARs inside the head and, in the same way, the back fields release high dose to the shoulders, without being
crucial for target coverage in that point.
\newline
The OL penalization is, here, exploited to switch off the fields in selected areas and to test a new field geometry.
In PLAN-OL-1, two artificial VOIs were drawn at the level if the shoulders and were assigned with OL = 1,
to switch off, respectively, the posterior fields irradiating that area.
In PLAN-OL-2, two addional artificial VOIs are created at the level of the head and set with OL = 1, to switch off the
anteriori fields in the part of the patient volume and a fifth intra-cranial field is added to compensate for the target coverage,
as shown in Fig.\ref{head-neck-intracranial-field}.
\begin{figure*}[!h]
\centering
\includegraphics[width=3.5in]{fig_02.png}
\caption{Patient with metallic cage: nominal IMPT dose distribution; inside the patient volume, the green
contour corresponds to the PTV and the red ones to the OARs; otuside the patient volume,
the safety VOIs and the effective countours of the 4 metallic rods are visible;
the fields F1, F2, F3, F4 used for PLAN-NOM and PLAN-OL-1 are also shown.}
\label{plan-metallic-cage}
\end{figure*}
\begin{figure*}[!h]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.25]{fig_03.pdf}
\caption{Axial CT slice}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.25]{fig_04.pdf}
\caption{Azimuthal CT slice}
\end{subfigure}
\caption{Fields crossing the nasal cavities: axial and azimuthal CT slice of the nasal cavity plan showing the red contour
of the PTV; the fields F1, F2 and F3 involved in the treatment are also shown.}
\label{field-nasal-cavity-plan}
\end{figure*}
\begin{figure*}[!h]
\centering
\hspace*{-1cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_05.pdf}
\caption{Empty nasal cavity}
\end{subfigure}
\hspace*{1cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_06.pdf}
\caption{Nasal cavity filled with mucus}
\end{subfigure}
\caption{Fields crossing the nasal cavities: modified CT slices simulating th extreme scenarios, when the nasal cavity
is completely empty (HU=0) and filled with mucus (HU=30).}
\label{nasal-cavities}
\end{figure*}
\begin{figure*}[!h]
\centering
\hspace*{-1cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_07.png}
\caption{Empty nasal cavity}
\end{subfigure}
\hspace*{1cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_08.png}
\caption{Nasal cavity filled with mucus}
\end{subfigure}
\caption{Fields crossing the nasal cavities: axial CT slices with the green contour of the CTV and red contours of some OARs;
the four fields F1, F2, F3, F4 are also indicated.}
\label{head-neck-fields-1}
\end{figure*}
\begin{figure*}[!h]
\centering
\includegraphics[width=3.5in]{fig_09.pdf}
\caption{Head and neck tumour: azimuthal CT slice showing the fifth intracranial field added for PLAN-OL-2.}
\label{head-neck-intracranial-field}
\end{figure*}
\section{Results}
The results of the simulations performed with the OL penalization are here shown and compared
to the nominal plans, described in the previous section.
\subsection{Patient with metallic cage}
Both plans with OL penalization have been re-calculated by dropping the dose constraints on the
safety margins drawn around the metallic rods and by setting the VOIs of the actual rod contours
with OL = 1. In this way, all pencil beams crossing even a voxel of such strutures are completely
switched off and they cannot be activated in the subsequent optimization.
\newline
In Fig.\ref{metallic-cage-plan-1}, the dose distributions of PLAN-NOM and PLAN-OL-1 are shown, but it
is from tables \ref{table-metallic-cage-1} that we get an insight regarding the main differences of the
plans. In PLAN-OL-1, the dose coverage of the clinical target volume (CTV) improves significantly leading
to an increase of +5.4\% of the volume reached by 100\% of the prescribed dose ($\mathbf{V_{100}}$).
Moreover, left and right parotis, the myelon and the brainstem are characterized by a substantial decrease in both
the maximum and the mean dose ($\mathbf{D_{max}}$, $\mathbf{D_{mean}}$) in PLAN-OL-1. The improved sparing of the parotis is
well represented by the comparison of the cumulative DVHs in Fig.\ref{metallic-cage-parotis-dvh-1}, where the red line refers
to PLAN-NOM and the blue one to PLAN-OL-1.
\newline
PLAN-OL-2 was conceived with more angularly spaced fields than the original F1, F2, F3, F4; this configuration would
have been not feasible without the OL penalization, since the tough constraints of the rod safety margins would have
prevented a homogeneous coverage of the target volume, leading necessarily to underdosed areas in the periphery of the CTV
and high dose peaks in the middle. Fig.\ref{metallic-cage-plan-2} shows the dose distributions of PLAN-NOM and PLAN-OL-2,
whereas results are summarized in tables \ref{table-metallic-cage-2}. Despite the fact the increase of $\mathbf{V_{100}}$
is less accentuated in PLAN-OL-2 than in PLAN-OL-1, the improved target coverage is visible in the comparison of the
cumulative DVHs of the CTV in Fig.\ref{metallic-cage-parotis-ctv-2}, whereas no difference was noticeable in the cumulative
DVH of PLAN-OL-1 compared to PLAN-NOM. This new field configuration provides a remarkable sparing of the left and right parotis
and the brainstem at the cost of decreasing the improvement concerning the myelon.
\begin{figure*}[!h]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_10.png}
\caption{PLAN-NOM}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_11.png}
\caption{PLAN-OL-1}
\end{subfigure}
\caption{Patient with metallic cage: dose distributions for the nominal plan and the one re-computed with OL penalization, keeping
the same field configuration.}
\label{metallic-cage-plan-1}
\end{figure*}
\begin{figure*}[!h]
\centering
\hspace*{-3.0cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.35]{fig_12.png}
\caption{Cumulative DVHs of the left parotis}
\end{subfigure}
\hspace*{3.0cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.35]{fig_13.png}
\caption{Cumulative DVHs of the right parotis}
\end{subfigure}
\caption{Patient with metallic cage: comparison between the cumulative DVHs of PLAN-NOM (red line) and PLAN-OL-1 (blue line) for
the left and right parotis.}
\label{metallic-cage-parotis-dvh-1}
\end{figure*}
\begin{figure*}[!h]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_14.png}
\caption{PLAN-NOM}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_15.png}
\caption{PLAN-OL-1}
\end{subfigure}
\caption{Patient with metallic cage: dose distributions for the nominal plan and the one re-computed with OL penalization, choosing
new fields more angularly spaced (+10\textdegree) with respect to those used for the nominal plan.}
\label{metallic-cage-plan-2}
\end{figure*}
\begin{figure*}[!h]
\centering
\includegraphics[width=3.5in]{fig_16.png}
\caption{Patient with metallic cage: comparison of the cumulative DVHs of the CTV for PLAN-NOM and PLAN-OL-2.}
\label{metallic-cage-parotis-ctv-2}
\end{figure*}
\begin{figure*}[!h]
\centering
\hspace*{-3.0cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.35]{fig_17.png}
\caption{Cumulative DVHs of the left parotis}
\end{subfigure}
\hspace*{3.0cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.35]{fig_18.png}
\caption{Cumulative DVHs of the right parotis}
\end{subfigure}
\caption{Patient with metallic cage: comparison between the cumulative DVHs of PLAN-NOM (red line) and PLAN-OL-2 (blue line) for
the left and right parotis.}
\label{metallic-cage-parotis-dvh-2}
\end{figure*}
\begin{table*}[!b]\centering
\ra{1.3}
\hspace*{-1cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & CTV & \phantom{abc} & PTV \\ \midrule
$\mathbf{D_{mean}}$ && +0.7\% && +1.2\% \\ \midrule
$\mathbf{V_{100}}$ && +5.4\% && +7.2\% \\
\bottomrule
\end{tabular}
\hspace*{1.0cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & $\mathbf{D_{max}}$ & \phantom{abc} & $\mathbf{D_{mean}}$ \\ \midrule
Right parotis && -1.7\% && -1.1\% \\ \midrule
Left parotis && -3.3\% && -2.7\% \\ \midrule
Myelon center && -2.5\% && -0.4\% \\ \midrule
Brainstem center && -1.8\% && -1.7\% \\
\bottomrule
\end{tabular}\vskip 5 pt
\caption{Patient with metallica cage: tables summarizing the differences between PLAN-OL-1 and PLAN-NOM concerning the dose coverage
of the target (left table) and OARs (right table).}
\label{table-metallic-cage-1}
\end{table*}
\begin{table*}[!t]\centering
\ra{1.3}
\hspace*{-1cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & CTV & \phantom{abc} & PTV \\ \midrule
$\mathbf{D_{mean}}$ && +0.9\% && +1.2\% \\ \midrule
$\mathbf{V_{100}}$ && +1.5\% && +3.4\% \\
\bottomrule
\end{tabular}
\hspace*{1.0cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & $\mathbf{D_{max}}$ & \phantom{abc} & $\mathbf{D_{mean}}$ \\ \midrule
Right parotis && -3.4\% && -2.4\% \\ \midrule
Left parotis && -6.0\% && -1.4\% \\ \midrule
Myelon center && -1.3\% && -0.5\% \\ \midrule
Brainstem center && -2.1\% && -1.8\% \\
\bottomrule
\end{tabular}\vskip 5 pt
\caption{Patient with metallic cage: tables summarizing the differences between PLAN-OL-2 and PLAN-NOM concerning the dose coverage
of the target (left table) and OARs (right table).}
\label{table-metallic-cage-2}
\end{table*}
\subsection{Fields crossing the nasal cavities}
To re-calculate the plan on the original CT with OL penalization, the VOI of the nasal cavity was drawn and set
with an OL = 0.5, meaning that pencil beams were decreased in fluence proportionally to their path length inside the VOI,
according to formula (\ref{smooth-penalization}). In the nominal plan, no dose constraints or other countermeasures were
taken into consideration to deal with the potential density changes of the nasal cavities, since they lie close to the target
volume and any dose constraint would severely affect the target coverage.
\newline
Fig.\ref{nasal-cavity-plan} shows the dose distribution of PLAN-NOM and PLAN-OL, that were computed on the original CT.
The plans result to be identical up to differences of 0.2\%,
as far as concerns both the target coverage and the sparing of the OARs.
\newline
When PLAN-NOM and PLAN-OL are re-computed on the extreme scenario CTs of Fig.\ref{nasal-cavities} and the difference
between each of these extreme-case plan and the corresponding reference ones are considered,
interesting results emerge. Fig.\ref{nasal-cavity-plan} shows PLAN-NOM-H and PLAN-OL-H, whereas tables \ref{table-nasal-cavity}
highlights the fact that in both scenarios the target coverage is more remarkably more robust for PLAN-OL than for PLAN-NOM,
since $\mathbf{V_{100}}$ improves by 3.1\% in the high-density nose cavities case and by +5.3\% in the other one.
\begin{figure*}[!h]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_19.png}
\caption{PLAN-NOM}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_20.png}
\caption{PLAN-OL}
\end{subfigure}
\caption{Fields crossing the nasal cavities: dose distributions for the nominal plan and the one re-computed with OL penalization
for the nasal cavity case.}
\label{nasal-cavity-plan}
\end{figure*}
\begin{figure*}[!h]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_21.png}
\caption{PLAN-NOM-H}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_22.png}
\caption{PLAN-OL-H}
\end{subfigure}
\caption{Fields crossing the nasal cavities: difference between the dose distribution re-calculated on the
CT with nasal cavities filled with mucus and the one computed on the original CT,
for the nominal plan and the one re-computed with OL penalization.}
\label{nasal-cavity-plan-high-diff}
\end{figure*}
\begin{table*}[!t]\centering
\ra{1.3}
\hspace*{-1cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & CTV & \phantom{abc} \\ \midrule
$\mathbf{D_{mean}}$ && +0.4\% \\ \midrule
$\mathbf{V_{100}}$ && +3.1\% \\
\bottomrule
\end{tabular}
\hspace*{1.0cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & CTV & \phantom{abc} \\ \midrule
$\mathbf{D_{mean}}$ && +0.5\% \\ \midrule
$\mathbf{V_{100}}$ && +5.3\% \\
\bottomrule
\end{tabular}
\caption{Fields crossing the nasal cavities: tables summarizing the differences between PLAN-OL-H and PLAN-NOM-H (on the left)
and PLAN-OL-L and PLAN-NOM-L (on the right) concerning the dose coverage
of the target.}
\label{table-nasal-cavity}
\end{table*}
\subsection{Head and neck tumour}
In this clinical case, the OL penalization is aimed at optimizing the same fields of the nominal plan
(PLAN-OL-1) and at testing a new treatment configuration (PLAN-OL-2).
\newline
By comparing the dose distributions of PLAN-NOM and PLAN-OL-1 at the level of the shoulders in Fig.\ref{head-neck-plan-1}, it is
noticeable how the dose in PLAN-OL-1 results better confined inside the CTV. This fact determines a remarkable improvement
in the sparing of the esophagus and the spinal cord, as reported in Table \ref{table-head-neck-1} and in Fig.\ref{head-neck-dvh-1},
where the comparison between the cumulative DVHs of the two plans are shown. No relevant changes concern the coverage of the target
volume, since $\mathbf{D_{mean}}$ differs by 0.3\% from the nominal plan. These results clearly show that switching off the posterior
fields at the level of the shoulders allows to achieve a better sparing of some OARs, while keeping the same irradiation level for the
tumour.
\newline
In PLAN-OL-2, a different field geometry has been investigated: the posterior fields are switched off at the level of the shoulders,
the anterior ones at the level of the head and a fifth intracranial field is added (Fig.\ref{head-neck-intracranial-field}). The goal
was to achieve a better preservation of the OARs inside the head. The dose distributions (Fig.\ref{head-neck-plan-2}) share
a very similar coverage of the target volume, but they are characterized by substantial differences
in the dose delivered to the OARs, as shown by Table \ref{table-head-neck-2}. In fact, the huge improvements in the sparing of the chiasm,
right lens (Fig.\ref{head-neck-dvh-2}a), left lens (\ref{head-neck-dvh-2}b), left inner ear and the spinal cord (considering in first place the decrease of $\mathbf{D_{max}}$) are followed
by a significant increase of the peak dose inside the thyroid and the mean dose to the brainstem. In this case, PLAN-OL-2 is not necessarily
superior to PLAN-NOM, but it is undeniable that the OL penalization has allowed the planner to test a scenario, that the simple enforcement
of the dose constraints would not have provided.
\begin{figure*}[!t]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_23.png}
\caption{PLAN-NOM}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_24.png}
\caption{PLAN-OL-1}
\end{subfigure}
\caption{Head and neck tumour: dose distributions of PLAN-NOM and PLAN-OL-1 at the level of the shoulders.}
\label{head-neck-plan-1}
\end{figure*}
\begin{figure*}[!t]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_25.png}
\caption{Esophagus}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_26.png}
\caption{Spinal cord}
\end{subfigure}
\caption{Head and neck tumour: comparison between the cumulative DVHs of PLAN-NOM (red line) and PLAN-OL-1 (blue line)
for the esophagus and the spinal cord.}
\label{head-neck-dvh-1}
\end{figure*}
\begin{table*}[!t]\centering
\ra{1.3}
\hspace*{1.0cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & $\mathbf{D_{max}}$ & \phantom{abc} & $\mathbf{D_{mean}}$ \\ \midrule
Right inner ear && -0.3\% && -1.2\% \\ \midrule
Esophagus && / && -2.2\% \\ \midrule
Spinal cord && -0.2\% && -3.2\% \\ \midrule
Brainstem && -1.0\% && -0.6\% \\
\bottomrule
\end{tabular}\vskip 5 pt
\caption{Head and neck tumour: tables summarizing the differences between PLAN-OL-1 and PLAN-NOM concerning the
the sparing of the OARs.}
\label{table-head-neck-1}
\end{table*}
\begin{figure*}[!t]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_27.png}
\caption{Esophagus}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_28.png}
\caption{Spinal cord}
\end{subfigure}
\caption{Head and neck tumour: dose distributions of PLAN-NOM and PLAN-OL-2 at the level of the head.}
\label{head-neck-plan-2}
\end{figure*}
\begin{figure*}[!t]
\centering
\hspace*{-1.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_29.png}
\caption{Left lense}
\end{subfigure}
\hspace*{2cm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_30.png}
\caption{Right lense}
\end{subfigure}\\\vskip 5 pt
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{fig_31.png}
\caption{Right optical nerve}
\end{subfigure}
\caption{Head and neck tumour: comparison between the cumulative DVHs of PLAN-NOM (red line) and PLAN-OL-2 (blue line)
for the left and right lense and the right optical nerve.}
\label{head-neck-dvh-2}
\end{figure*}
\begin{table*}[!t]\centering
\ra{1.3}
\hspace*{1.0cm}
\begin{tabular}{@{}rrrrcrrrcrrr@{}}\toprule
& \phantom{abc} & $\mathbf{D_{max}}$ & \phantom{abc} & $\mathbf{D_{mean}}$ \\ \midrule
Right lens && -9.0\% && -7.1\% \\ \midrule
Left lens && -3.9\% && -4.1\% \\ \midrule
Left inner ear && -2.6\% && +0.3\% \\ \midrule
Left parotis gland && -0.2\% && -3.9\% \\ \midrule
Thyroid && +2.7\% && +4.9\% \\ \midrule
Spinal cord && -2.0\% && +2.0\% \\ \midrule
Brainstem && -1.9\% && +9.5\% \\ \midrule
Chiasm && -13.9\% && -4.3\% \\
\bottomrule
\end{tabular}\vskip 5 pt
\caption{Head and neck tumour: tables summarizing the differences between PLAN-OL-2 and PLAN-NOM concerning the
the sparing of the OARs.}
\label{table-head-neck-2}
\end{table*}
\clearpage\newpage
\section{Discussion}
In this work, we have introduced the concept of \emph{opacity} as a quality assigned to VOIs and aimed at steering the outcome of the
IMPT planning by changing the starting condition of the optimization. In particular, the opacity level (OL) of a VOI is a weight
ranging in [0,1] and is used to penalize the initial fluence of the pencil beams crossing the voxels of such structure:
when OL = 0, the VOI is transparent and no penalization is performed; when 0 $<$ OL $<$ 1, the pencil beams are decreased in fluence proportionally to
their path length inside the VOI; when OL = 1, they are totally switched off.
\newline
We have shown how this simple additional feature can lead to dose distributions of higher quality and robustness within the framework
the IMPT treatment planning. Three different clinical cases were taken into consideration to show the potential advantage of adopting the
OL penalization in the pre-optimization stage: a patient with a post-surgery metallic cage, a treatment where fields cross the nasal cavities and
the case of a complex head and neck tumour.
\newline
Concerning the first clinical case, the opacity tool has shown to perfectly tackle the presence of metallic structures:
the safety margins and hard dose constraint on the metal are replaced with the actual contour of the structure and set with OL = 1.
This approach has lead to a plan, where both the coverage of the target volume and the sparing of some OARs have improved. The strategy to handle
the metallic cage can be re-proposed for golden teeth or any other prothesis, whose composition and, therefore, calibration curve are not known.
\newline
As far as regards the second clinical case, the OL penalization has provided a more robust dose distribution in relation to the possible density
changes inside the nasal cavities. Since the the plan does not change when OL is set to a specific value in (0,1), we obtain a simple way to guarantee
the stability of the treatment. This approach may be applied for any other anatomical cavity, placed close to the tumour and that can undergo substantial
density variations within the treatment frametime.
\newline
For the case of the head and neck tumour, we have shown how the opacity allows to optimize the usage of each single field, especially when dealing
with a tumour mass of complex shape and considerable extension. We have also pointed out how the opacity can easily shape the fields in order to test
treatment scenario, that would not be possible to achieve only through dose constraints.
\newline
In conclusion, the opacity is a leverage that be used by the planner to mould the IMPT dose distributions, when seeking for the best tradeoff
between target coverage and sparing of the OARs. It is no meant to replace the dose constraints, but rather to be played as an additional card
in the IMPT treatment planning.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Load modelling and identification are of great importance to the security and stability of power systems. While the accurate models are available for generators, transmission lines and control devices, load modeling is still a challenging and open subject due to the fact that electric load at each substation is an aggregation of numerous individual loads with different behaviors \cite{Hill:1994}-\cite{Hiskens:2006}. In addition, the poor measurements, modeling, exchange information, as well as the uncertainties in customers behaviors/devices further result in load uncertainties \cite{Turistsyn:2016}. Indeed, load uncertainty is one of the main factors that affect the accuracy of the power dynamic models implemented by system operators over the world\cite{Hiskens:2006}.
Generally speaking, the load uncertainty comes from both model structure and parameter values. It has been shown in previous literature \cite{Najafabadi:2012}-\cite{Kao:1994} that the use of different load models leads to different and even contradictory results for dynamic stability studies. Even though the applied model structure is verified, different parameter values may also yield different damping performances in small signal stability \cite{Turistsyn:2016}\cite{Hiskens:1995}\cite{Hiskens:1995_1}. For instance, different time constants of loads may lead to either asymptotically stable system or systems experiencing oscillations (i.e., Hopf bifurcation occurs)\cite{Turistsyn:2016}. Both load modelling and parameter identifications are essential in studying the dynamic behaviors of power systems. This paper mainly focuses on parameter identification for a generic dynamic load model that is suitable for small signal stability analysis.
Different methods for dynamic load parameter identification have been proposed, which can be classified into two categories: component-based approach \cite{Price:1988} and measurement-based approach \cite{Chiang:2006}-\cite{Wen:2003}. The latter approach is more commonly applied because real-time load variations and dynamic characteristics can be taken into account\cite{Ajjarapu:2009}. Measurement-based model identification is typically solved through optimization methods that minimize the error between the measured output variables and the simulated ones. In particular, the nonlinear least-square curve fitting method has been implemented in \cite{Hill:1994}\cite{Chiang:2006}-\cite{Hill:2008}. Genetic algorithms, neural network-based methods and other heuristic techniques have been applied in \cite{Karlsson:1996}-\cite{Wen:2003}. However, the optimization-based methods are time consuming and thus can not be implemented online\cite{Najafabadi:2012}. More importantly, all those methods require measurement data from dynamic behaviors of system under big disturbances (e.g., during faults), which is not always available \cite{Han:2009}. Indeed, the variation of load parameters may be much faster than the occurrence rate of natural disturbances \cite{Hill:1994}.
In this paper, we propose a novel measurement-based method for dynamic load identification in ambient conditions, which does not require the existence of large disturbance. Particularly, the method combines the statistical properties extracted from PMU measurements and the inherent model knowledge, and is able to provide fairly accurate estimations for parameter values in near real-time. Note that a generic dynamic load model is implemented in this paper which is suitable for the purposes of small signal stability analysis and damping performance \cite{Turistsyn:2016}\cite{Hiskens:1995}\cite{Hiskens:1995_1}\cite{Turistsyn:2015}. The proposed method can be implemented in online security analysis to provide up-to-date dynamic load parameters accurately.
The rest of the paper is organized as follows. Section \ref{sectionmodel} introduces the power system stochastic dynamic model. Particularly, the generic dynamic load model used in small signal stability is presented. Section \ref{sectionmethod} elaborates the proposed method for estimating parameters of dynamic loads. Section \ref{sectioncasestudy} presents the validation of the proposed method through numerical simulations. The impact of measurement noise is also investigated. Conclusions and perspectives are given in Section \ref{sectionconclusion}.
\section{power system stochastic dynamic model}\label{sectionmodel}
Although we focus on load models, generator models are also incorporated to provide more realistic simulations. Specifically, the classical generator model which can reasonably represent the dynamics of generator in ambient conditions is implemented. The power system buses are numbered as follows: load buses: $k = 1,2,...,m$, and generators: $i=m+1,...,N$. Particularly, to include the effects of the loads, the structure preserving model \cite{Chiang:book}\cite{Pai:2012} is used:
\small{
\begin{eqnarray}
\dot{\delta}_i&=&{\omega_i}\label{swing-1}\\
M_i\dot{{\omega_i}}&=&{P_{mi}}-{P_{Gi}}(\delta_i,\theta_i,V_i)-{D_i}{\omega_i}\label{swing-2}\\
P_{Gi}(\delta_i,\theta_i,V_i)&=&\sum_{k=1}^N|V_i||V_k|(G_{ik}\cos\theta_{ik}+B_{ik}\sin\theta_{ik})\\
Q_{Gi}(\delta_i,\theta_i,V_i)&=&\sum_{k=1}^N|V_i||V_k|(G_{ik}\sin\theta_{ik}-B_{ik}\cos\theta_{ik})
\end{eqnarray}}
\normalsize
where
\begin{table}[!ht]\normalsize
\begin{tabular}{ll}
$\delta_i$& generator rotor angle\\
$\omega_i$& generator angular frequency\\
$M_i$& inertial constant \\
$P_{mi}$ & mechanical power input \\
$P_{Gi}(\delta_i,\theta_i,V_i)$ & real power injection \\
$Q_{Gi}(\delta_i,\theta_i,V_i)$ & reactive power injection \\
$D_i$ & damping coefficient \\
$N$& total number of buses\\
$\theta_{ij}$& voltage angle difference between bus $i$ and $j$\\
$|V_i|$ & voltage magnitude\\
$G_{ij}$& line conductance between bus $i$ and $j$\\
$B_{ij}$& line susceptance between bus $i$ and $j$
\end{tabular}
\end{table}
\noindent The detailed expressions of $P_{Gi}(\delta_i,\theta_i,V_i)$ and $Q_{Gi}(\delta_i,\theta_i,V_i)$ are neglected here for simplicity and can be found in many books (e.g., \cite{Chiang:book}).
\normalsize
Regarding dynamic loads, we use the following first-order load model proposed in \cite{Turistsyn:2015} that can represent the common types of loads (e.g., induction motors, thermostatically controlled loads) in ambient conditions:
\small{
\begin{eqnarray}
\dot{g}_k&=&-\frac{1}{\tau_{g_k}}(P_k-P_{k}^s)\label{loaddynamicp}\\%(1+\sigma_k^p\xi_k^p)]\\
\dot{b}_k&=&-\frac{1}{\tau_{b_k}}(Q_k-Q_{k}^s)\label{loaddynamicq}\\%(1+\sigma_k^q\xi_k^q)] \\
P_k=g_kV_k^2&=&\sum_{j=1}^N|V_k||V_j|(-G_{kj}\cos\theta_{kj}-B_{kj}\sin\theta_{kj})\\
Q_k=b_kV_k^2&=&\sum_{j=1}^N|V_k||V_j|(-G_{kj}\sin\theta_{kj}+B_{kj}\cos\theta_{kj})
\end{eqnarray}}
\normalsize
where
\begin{table}[!ht]\normalsize
\begin{tabular}{ll}
$g_k$& effective conductance of the load\\
$b_k$& effective susceptance of the load \\
$\tau_{gk}$ & active power time constant of the load \\
$\tau_{bk}$ & reactive power time constant of the load \\
$P_k$ & real power demand of the load\\
$Q_k$ & reactive power demand of the load\\
$P_k^s$ & steady-state real power demand of the load \\
$Q_k^s$ & steady-state reactive power demand of the load \\
\end{tabular}
\end{table}
\noindent The values $P_k^s$ and $Q_k^s$ describe the static (steady-state) power characteristics of the loads achieved in equilibrium. The instant real power and reactive power consumption can be characterized by the effective conductance $P_k=g_kV_k^2$ and susceptance $Q_k=b_kV_k^2$ at any time. The time constants $\tau_{gk}$ and $\tau_{bk}$ that typically depend on voltage and frequency represent the instant relaxation rate of the load.
To incorporate load variation, we apply a similar approach used in \cite{Nwankpa:2000}\cite{Nwankpa:1992} and modify the set of load equations (\ref{loaddynamicp})-(\ref{loaddynamicq}) as follows:
\begin{eqnarray}
\dot{g}_k&=&-\frac{1}{\tau_{g_k}}[P_k-P_{k}^s(1+\sigma_k^p\xi_k^p)]\label{stoloaddynamicp}\\
\dot{b}_k&=&-\frac{1}{\tau_{b_k}}[Q_k-Q_{k}^s(1+\sigma_k^q\xi_k^q)]\label{stoloaddynamicq}
\end{eqnarray}
where the steady-state real and reactive load demands are perturbed with independent Gaussian noise from their initial values. Specifically, $\xi_k^p$ and $\xi_k^q$ are standard Gaussian noise, and $\sigma_k^p$ and $\sigma_k^q$ represent the noise intensities for static real and reactive power, respectively.
As discussed in \cite{Turistsyn:2016}\cite{Turistsyn:2015}, this dynamic load model can naturally represent the most common types of loads in ambient conditions such as thermostatic load, induction motor, power electronic converter, aggregate effects of distribution load tap changer (LTC) transformers, etc.
However, the range of time constants is considerably large ranging from cycles to several minutes, and even hours for different types of loads. For industrial plants, such as aluminum smelters, the time constants are in the range of $0.1$s to $0.5$s; for tap changers and other control devices, they are in the range of minutes; for heating load, they may range up to hours \cite{Hiskens:1995_1}. As a result, the uncertainty of composition of different types of loads can be aggregated in time constants $\tau_{g}$ and $\tau_{b}$\cite{Turistsyn:2016}. This is reasonable in the situations when the network characteristics are known, generator models are validated and static load characteristics are understood better than their dynamic response which is the case in practical situations. In addition to a wide range of time constants, the variation of $\tau_g$ and $\tau_b$ can also be fast. For example, $\tau_b$ may change from $0.1$s to $24.1$s in one day (see Table I, II in \cite{Hill:1994}).
Because of wide range and fast variation of time constants $\tau_g$ and $\tau_b$, they need to be updated frequently to ensure the accuracy of dynamic load models used in online security and stability analysis. Conventionally, $\tau_g$ and $\tau_b$ are estimated from dynamic data by perturbing the system, for example, through changing the transformer tap\cite{Cutsem:book}. However it is impractical to perturb the system frequently for estimating parameter values of loads. In this paper, we propose a novel method to estimate $\tau_g$ and $\tau_b$ for the loads of interests from ambient PMU measurements in daily operation. In particular, the estimation process does not require the existence of disturbance to the system.
\section{Methodology}\label{sectionmethod}
In ambient conditions, the stochastic dynamic load equations (\ref{stoloaddynamicp})-(\ref{stoloaddynamicq}) can be linearized as below:
\begin{eqnarray}
\left[\begin{array}{c}\dot{\bm{g}}\\\dot{\bm{b}}\end{array}\right]&=&\left[\begin{array}{cc}-{T_g}^{-1}\frac{\partial{\bm{P}}}{\partial{\bm{g}}}&\bm{0}\\
\bm{0}&-{T_b}^{-1}\frac{\partial{\bm{Q}}}{\partial{\bm{b}}}\end{array}\right]
\left[\begin{array}{c}{\bm{g}}\\{\bm{b}}\end{array}\right]\nonumber\\
&+&\left[\begin{array}{cc}{T_g}^{-1}P^s\Sigma^p&\bm{0}\\
\bm{0}&{T_b}^{-1}Q^s\Sigma^q\end{array}\right]\nonumber
\left[\begin{array}{c}{\bm{\xi^p}}\\{\bm{\xi^q}}\end{array}\right]\\
&=&A\left[\begin{array}{c}{\bm{g}}\\{\bm{b}}\end{array}\right]
+B\left[\begin{array}{c}{\bm{\xi^p}}\\{\bm{\xi^q}}\end{array}\right]
\end{eqnarray}
where
\begin{table}[!ht]\normalsize
\begin{tabular}{ll}
$\bm{g}=[g_1,...,g_m]^T$, & $\bm{b}=[b_1,...,b_m]^T$, \\
$T_g=\mbox{diag}[\tau_{g1},...,\tau_{gm}]$, &$T_b=\mbox{diag}[\tau_{b1},...,\tau_{bm}]$,\\ $\bm{P}=[P_1,...,P_m]^T$,& $\bm{Q}=[Q_1,...,Q_m]^T$,\\
$P^s=\mbox{diag}[P_1^s,...,P_m^s]$,& $Q^s=\mbox{diag}[Q_1^s,...,Q_m^s]$, \\ $\Sigma^p=\mbox{diag}[\sigma^p_1,...,\sigma^p_m]$,& $\Sigma^q=\mbox{diag}[\sigma^q_1,...,\sigma^q_m]$, \\ $\bm{\xi^p}=[\xi^p_1,...,\xi^p_m]^T$,& $\bm{\xi^q}=[\xi^q_1,...,\xi^q_m]^T$.
\end{tabular}
\end{table}
It is observed that $[\bm{g},\bm{b}]^T$ is a vector Ornstein-Uhlenbeck process that is stationary, Gaussian and Markovian \cite{Gardiner:2009}\cite{Wangxz:2015}. Particularly, if the state matrix $A$ is stable, the stationary covariance matrix $C_{\bm{xx}}=\left[\begin{array}{cc}C_{\bm{gg}}&C_{\bm{gb}}\\C_{\bm{bg}}&C_{\bm{bb}}\end{array}\right]$ can be shown to satisfy
the following Lyapunov equation\cite{Gardiner:2009}\cite{Hines:2015}:
\begin{equation}
AC_{\bm{xx}}+C_{\bm{xx}}A^T=-BB^T \label{lyapunov}
\end{equation}
which nicely combines the model knowledge and the statistical properties of state variables.
Since $P_k=g_kV_k^2$, we have:
\begin{equation}
\frac{\partial{P_k}}{\partial{g_j}}=\left\{\begin{array}{cc}V_k^2+2g_jV_k\frac{\partial{V_k}}{\partial{g_j}}\approx{V_k^2}&\mbox{if }j=k\\
2g_jV_k\frac{\partial{V_k}}{\partial{g_j}}\approx 0&\mbox{if }j\not=k\end{array}\right.
\end{equation}
under the assumption that $\triangle V_k\approx 0$ in ambient conditions. Similar relation can be obtained for $\frac{\partial{Q_k}}{\partial{b_j}}$. As a result, the Jacobian matrix $A$ satisfies:
\begin{eqnarray}
A\approx \left[\begin{array}{cc}-{T_g}^{-1}V^2&\bm{0}\\
\bm{0}&-{T_b}^{-1}V^2\end{array}\right]\label{approxA}
\end{eqnarray}
where $V=\mbox{diag}[V_1,...,V_m]$. Substituting (\ref{approxA}) and the detailed expression of $C_{\bm{xx}}$ and $B$ into (\ref{lyapunov}),
and performing algebraic simplification, we have:
\begin{eqnarray}
&&C_{\bm{gg}}=\frac{1}{2}T_g^{-1}(P^s)^2(\Sigma^p)^2V^{-2}\label{Cgg}\\
&&C_{\bm{bb}}=\frac{1}{2}T_b^{-1}(Q^s)^2(\Sigma^q)^2V^{-2}\label{Cbb}\\
&&C_{\bm{gb}}=C_{\bm{bg}}=0
\end{eqnarray}
Particularly, we utilize the relations (\ref{Cgg})-(\ref{Cbb}) that link the measurements of stochastic load variation to the physical model, and provide an ingenious way to estimate the dynamic parameters $T_g$ and $T_b$ from measurements.
In practical applications, $V$, $C_{\bm{gg}}$, and $C_{\bm{bb}}$ need to be acquired or estimated from limited PMU measurements. A window size of $1000$s is used in the examples of this paper where time constants are up to several seconds. Note that the larger the time constants, the longer the sample window is needed to ensure accuracy. First, the sample mean $\bar{V}$ can be used as an estimation of $V$, then $\bm{g}$ and $\bm{b}$ can be estimated from PMU measurements (i.e., phasors $V_k$ and $I_k$) as follows:
\begin{eqnarray}
g_k&=&\mbox{Re}\{\frac{I_k}{V_k}\}\label{g}\\
b_k&=&\mbox{Im}\{\frac{I_k}{V_k}\}\label{b}
\end{eqnarray}
Regarding the covariance matrix $C_{\bm{g}\bm{g}}=\E[(\bm{g}-\E[\bm{g}])(\bm{g}-\E[\bm{g}])^T]$ and $C_{\bm{b}\bm{b}}=\E[(\bm{b}-\E[\bm{b}])(\bm{b}-E[\bm{b}])^T]$, we use their unbiased estimators---sample covariance matrixes $Q_{\bm{g}\bm{g}}$ and $Q_{\bm{b}\bm{b}}$ in practice, each entry of which is calculated as below:
\begin{eqnarray}
Q_{g_kg_j}&=&\frac{1}{n-1}\sum_{i=1}^n(g_{k}(i)-\bar{g}_k)(g_{j}(i)-\bar{g}_j)\label{qgg}\\
Q_{b_kb_j}&=&\frac{1}{n-1}\sum_{i=1}^n(b_{k}(i)-\bar{b}_k)(b_{j}(i)-\bar{b}_j)\label{qbb}
\end{eqnarray}
where $\bar{{g}}_k$ and $\bar{{b}}_k$ denote the sample mean of ${g_k}$ and ${b_k}$, respectively, and $n$ is the sample size.
Therefore, the proposed algorithm can be summarized as follows. We assume that PMUs are installed at the substations that the (aggregated) loads of interests are connected to. We also assume that the static characteristics of loads are well understood such that $P^s$, $Q^s$, $\Sigma^p$ and $\Sigma^q$ are prior known, which is reasonable as shown in\cite{Nwankpa:2000}\cite{Nwankpa:1992}. Then the following algorithm provides an estimation of $T_g$ and $T_b$ for the dynamic loads from ambient PMU measurements:
\begin{description}[\IEEEusemathlabelsep\IEEEsetlabelwidth{Step 1.}]
\item [\textbf{{Step 1.}}] Compute the sample mean $\bar{V}$ and estimate $\bm{g}$ and $\bm{b}$ from PMU measurements by (\ref{g})-(\ref{b}).
\item [\textbf{{Step 2.}}] Calculate the sample covariance matrix $Q_{\bm{g}\bm{g}}$ and $Q_{\bm{b}\bm{b}}$ by (\ref{qgg})-(\ref{qbb}).
\item [\textbf{{Step 3.}}] Approximate $T_g$ and $T_b$ as blow:
\begin{eqnarray}
T_g&=&\frac{1}{2}(P^s)^2(\Sigma^p)^2\bar{V}^{-2}Q_{\bm{gg}}^{-1}\label{Tg}\\
T_b&=&\frac{1}{2}(Q^s)^2(\Sigma^q)^2\bar{V}^{-2}Q_{\bm{bb}}^{-1}\label{Tb}
\end{eqnarray}
\end{description}
Note that (\ref{Tg})-(\ref{Tb}) are acquired by a simple algebraic manipulation of (\ref{Cgg})-(\ref{Cbb}).
\section{case studies}\label{sectioncasestudy}
In this section, the proposed algorithm to estimate time constants of dynamic loads are validated through numerical simulations. Furthermore, the robustness of the proposed method to measurement noise is also demonstrated via simulation. All case studies were done in PSAT-2.1.9 \cite{Milano:PSAT}.
\subsection{Validation of the Method}\label{sectionvalidation}
We consider the standard WSCC 3-generator, 9-bus system model (see, e.g. \cite{Chiang:book}). The classical generator models (\ref{swing-1})-(\ref{swing-2}) and the stochastic dynamic load models (\ref{stoloaddynamicp})-(\ref{stoloaddynamicq}) are implemented in the structure preserving framework. The system parameters are available online: https://github.com/xiaozhew/PES-load-parameter-estimation. Particularly, there are three dynamic loads at buses 1, 2 and 3, the time constants of which are $\tau_g=1,3,0.2\mbox{s}$ and $\tau_b=5,7,0.8\mbox{s}$, respectively. The trajectories of some state variables and algebraic variables are shown in Fig. \ref{9bus}, from which we see that the state variables are fluctuating around their nominal values in ambient conditions, yet larger time constants lead to slower variations as expected (e.g., the variations of $g_2$ and $b_2$ are slower than $g_3$ and $b_3$).
\begin{figure}[!ht]
\centering
\begin{subfigure}[t]{0.52\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{g2.eps}
\caption{Trajectory of $g_2$ on [0s,1000s]}\label{g2}
\end{subfigure
\begin{subfigure}[t]{0.48\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{b2.eps}
\caption{Trajectory of $b_2$ on [0s,1000s]}\label{b2}
\end{subfigure}
\begin{subfigure}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{g3.eps}
\caption{Trajectory of $g_3$ on [0s,1000s]}\label{g3}
\end{subfigure
\begin{subfigure}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{b3.eps}
\caption{Trajectory of $b_3$ on [0s,1000s]}\label{b3}
\end{subfigure}
\begin{subfigure}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{V2.eps}
\caption{Trajectory of $V_2$ on [0s,1000s]}\label{g3}
\end{subfigure
\begin{subfigure}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{V3.eps}
\caption{Trajectory of $V_3$ on [0s,1000s]}\label{b3}
\end{subfigure}
\caption{Trajectories of some dynamic conductances, susceptances, and voltages in the 9-bus system.}\label{9bus}
\end{figure}
By the proposed algorithm, we firstly compute the sample mean $\bar{V}=\mbox{diag}[0.9952, 1.0126,1.0155]$. Then we estimate the dynamic conductance, susceptances and their corresponding sample covariance matrixes:
\begin{eqnarray}
Q_{\bm{gg}}=\left[\begin{array}{ccc}
1.41\times10^{-3} & 9.84\times10^{-5} & 2.24\times10^{-4}\\
9.84\times10^{-5} & 4.16\times10^{-4} & 6.57\times10^{-5}\\
2.24\times10^{-4} & 6.57\times10^{-5} & 5.75\times10^{-3}
\end{array}\right]
\end{eqnarray}
\begin{eqnarray}
Q_{\bm{bb}}=\left[\begin{array}{ccc}
2.63\times10^{-4} & -1.47\times10^{-6} & 6.46\times10^{-6}\\
-1.47\times10^{-6} & 1.75\times10^{-4} & -7.57\times10^{-6}\\
6.46\times10^{-6} & -7.57\times10^{-6} & 1.62\times10^{-3}
\end{array}\right]
\end{eqnarray}
It is expected that both $Q_{\bm{gg}}$ and $Q_{\bm{bb}}$ are nearly diagonal as the stochastic perturbations are independent.
Since each entry of $(P^s)^2(\Sigma^p)^2$ and $(Q^s)^2(\Sigma^q)^2$ is set to be $0.0025$, $T_g$ and $T_b$ can be readily estimated from (\ref{Tg})-(\ref{Tb}). A comparison between the estimated $\tau_g$, $\tau_b$ and their actual values are shown in Table. \ref{taugb-9bus}. It's observed that the proposed algorithm provides fairly accurate estimation for time constants of each load.
\begin{table}[!ht]
\centering
\caption{A comparison between the actual and the estimated time constants in the 9-bus system}\label{taugb-9bus}
\begin{tabular}{|c|c|c|c|}
\hline
&actual value (s)&estimated value (s) &error\\
\hline
$\tau_{g1}$&1.0000&0.9145&8.55\%\\
\hline
$\tau_{g2}$&3.0000&2.9867&0.44\%\\
\hline
$\tau_{g3}$&0.2000&0.2122&6.1\%\\
\hline
$\tau_{b1}$&5.0000&4.7974&4.05\%\\
\hline
$\tau_{b2}$&7.0000&6.9777&0.32\%\\
\hline
$\tau_{b3}$&0.8000&0.7462&6.72\%\\
\hline
\end{tabular}
\end{table}
\subsection{Impact of Measurement Noise}
Like other measurement-based methods, the performance of the proposed algorithm may be affected by PMU measurement noise. In order to investigate the potential influence, measurement noises with standard deviation of $10^{-3}$ have been added to $\bm{g}$, $\bm{b}$ and ${V}$ in the 9-bus example shown in Section \ref{sectionvalidation} according to the IEEE Standards \cite{IEEEstandard}\cite{IEEEStandard_amd}. A comparison between the actual and the estimated time constants are presented in Table. \ref{taugb-9bus-noise}. It is observed that the proposed method provides similar accuracy to the case without the measurement noises, which indicates that the method is relatively robust under measurement noise.
\begin{table}[!ht]
\centering
\caption{A comparison between the actual and the estimated time constants in the 9-bus system with the measurement noises}\label{taugb-9bus-noise}
\begin{tabular}{|c|c|c|c|}
\hline
&actual value (s)&estimated value (s) &error\\
\hline
$\tau_{g1}$&1.0000&0.9144&8.56\%\\
\hline
$\tau_{g2}$&3.0000&2.9819&0.60\%\\
\hline
$\tau_{g3}$&0.2000&0.2121&6.06\%\\
\hline
$\tau_{b1}$&5.0000&4.7752&4.50\%\\
\hline
$\tau_{b2}$&7.0000&6.9426&0.82\%\\
\hline
$\tau_{b3}$&0.8000&0.7443&6.97\%\\
\hline
\end{tabular}
\end{table}
\subsection{Further Validation}
For further validation, we apply the method to a larger system---the IEEE 39-bus 10-generator test
system, the parameters of which are available online: https://github.com/xiaozhew/PES-load-parameter-estimation. In particular, 10 dynamic loads have been added to buses 1-10, and their corresponding time constants range from $0.1$s to $5$s. A comparison between the actual and the estimated time constants are presented in Table. \ref{taugb-39bus}. The simulation results further demonstrate that the proposed method is able to provide good estimations for time constants of the dynamic loads.
\begin{table}[!ht]
\centering
\caption{A comparison between the actual and the estimated time constants in the 39-bus system}\label{taugb-39bus}
\begin{tabular}{|c|c|c|c|}
\hline
&actual value (s)&estimated value (s) &error\\
\hline
$\tau_{g1}$&0.1000& 0.1225&22.55\%\\
\hline
$\tau_{g2}$&0.6000&0.5860&2.33\%\\
\hline
$\tau_{g3}$&1.1000&1.1086&0.78\%\\
\hline
$\tau_{g4}$&1.6000&1.6924&5.78\%\\
\hline
$\tau_{g5}$&2.1000&2.1323&1.54\%\\
\hline
$\tau_{g6}$&2.6000&2.6130&0.50\%\\
\hline
$\tau_{g7}$&3.1000&3.0715&0.92\%\\
\hline
$\tau_{g8}$&3.6000&3.3004&8.32\%\\
\hline
$\tau_{g9}$&4.1000& 4.5080&9.95\%\\
\hline
$\tau_{g10}$&4.6000&4.6636&1.38\%\\
\hline
$\tau_{b1}$&0.5000&0.5277&5.53\%\\
\hline
$\tau_{b2}$&1.0000&0.9831&1.69\%\\
\hline
$\tau_{b3}$&1.5000&1.5325&2.17\%\\
\hline
$\tau_{b4}$&2.0000&2.0559&2.79\%\\
\hline
$\tau_{b5}$&2.5000&2.5784&3.14\%\\
\hline
$\tau_{b6}$&3.0000&3.3522&11.74\%\\
\hline
$\tau_{b7}$&3.5000&3.1273&10.65\%\\
\hline
$\tau_{b8}$&4.0000&4.1105&2.76\%\\
\hline
$\tau_{b9}$&4.5000&4.3084&4.26\%\\
\hline
$\tau_{b10}$&5.0000&5.1723&3.45\%\\
\hline
\end{tabular}
\end{table}
\section{conclusions and perspectives}\label{sectionconclusion}
In this paper, we have proposed a novel method to estimate parameter values of dynamic load from ambient PMU measurements. The accuracy and robustness of the method have been demonstrated through numerical studies. Unlike conventional methods, the proposed technique does not require the existence of large disturbance to systems, and thus can be implemented continuously in daily operation to provide up-to-date dynamic load parameter values.
In the future, we plan to further validate the method by using real PMU data and extend the method to estimate dynamic load parameters without knowing their static characteristics.
|
\section{Introduction}
Throughout, let $\mathbbm{k}$ be an algebraically closed field of characteristic zero and let $\mathbbm{k}^*$ denote the multiplicative group $\mathbbm{k}-\{0\}.$ Unless otherwise specified, all algebraic structures and linear operations are over $\mathbbm{k}.$ Our main aim is to provide explicit and unified formulas for the cocycles on the normalized bar resolutions (normalized cocycles) of finite abelian groups. Some applications to braided linear Gr-categories and Dijkgraaf-Witten Invariants (DW invariant) are also considered.
The cohomology groups of finite abelian groups are computable thanks to the well known Lyndon-Hochschild-Serre spectral sequence \cite{hs,lyndon}. However, the explicit formulas of normalized cocycles are not clear in literatures. Such explicit formulas of normalized cocycles, instead of the cohomology groups, are necessary in many respects of mathematics and physics. Besides the connections to braided linear Gr-categories and DW invariants involved in the present paper, normalized 2-cocycles are necessary in projective representation theory of finite groups \cite{frucht, kar}; normalized 3-cocycles are indispensable in the classification program of pointed finite tensor categories and quasi-quantum groups \cite{eg1, eg2, g, qha6, bgrc1, qha3}; normalized cocycles of all degrees are very important in the theory of symmetry protected topological orders \cite{cw, cw2, spt}.
Our approach of formulating the normalized cocycles is straightforward and elementary. First we construct a Koszul-like resolution of a finite abelian group $G$ by tensoring the minimal resolutions of cyclic factors of $G$ and give a complete set of representatives of cocycles for this resolution. Then we construct a chain map from the normalized bar resolution to this Koszul-like resolution. Finally we get the desired explicit and unified formulas of normalized cocycles on $G$ by pulling back those on the Koszul-like resolution along the chain map. We remark that, in principle, the method of Lyndon-Hochschild-Serre spectral sequence may also help one formulate explicit forms of normalized cocycles with nearly as much effort as we need here.
Here is a brief description of the content. In Section 2, we provide formulas of normalized cocycles of all degrees on any finite abelian groups. In Section 3, we use the formula of normalized 3-cocycles to determine the braided monoidal structures on linear Gr-categories.
In Section 4, we give a formula for the Dijkgraaf-Witten invariant of the $n$-torus for all $n$ and obtain the dimension formula for irreducible projective representations of an arbitrary finite abelian group.
\section{Explicit formulas of normalized cocycles on finite abelian groups}
In this section, we use freely the concepts and notations about group cohomology in the book \cite{w} of Weibel. Let $G$ be a group and $(B_{\bullet},\partial_{\bullet})$ be its normalized bar resolution. Applying $\operatorname{Hom}_{\mathbb{Z}G}(-,\mathbbm{k}^{\ast})$ one gets a complex $(B^{\ast}_{\bullet},\partial^{\ast}_{\bullet}).$ Denote the group of normalized $n$-cocycles by $Z^n(G,\mathbbm{k}^*),$ which is $\operatorname{Ker} \partial^*_n.$ In general, it is hard to determine $Z^n(G,\mathbbm{k}^*)$ directly as the normalized bar resolution is far too large. Our strategy of overcoming this is to get first a simpler resolution of $G$ whose cocycles are easy to compute and then construct a chain map from the normalized bar resolution to it which will help to determine $Z^n(G,\mathbbm{k}^*)$ eventually.
\subsection{A Koszul-like resolution}
From now on let $G$ be a finite abelian group. Write $G=\mathbb{Z}_{m_{1}}\times\cdots \times\mathbb{Z}_{m_{n}}$ where $m_i|m_{i+1}$ for $1\le i\le n-1$ and for every $\mathbb{Z}_{m_{i}}$ fix a generator $g_{i}$ for $1\leq i\leq n.$ It is well known that the following periodic sequence is a free resolution of the
trivial $\mathbb{Z}_{m_{i}}$-module $\mathbb{Z}:$
\begin{equation}\label{resolution}
\cdots\longrightarrow \mathbb{Z}\mathbb{Z}_{m_{i}}\stackrel{T_{i}}\longrightarrow
\mathbb{Z}\mathbb{Z}_{m_{i}}\stackrel{N_{i}}\longrightarrow\mathbb{Z}\mathbb{Z}_{m_{i}}\stackrel{T_{i}}\longrightarrow
\mathbb{Z}\mathbb{Z}_{m_{i}}\stackrel{N_{i}}\longrightarrow
\mathbb{Z}\longrightarrow 0,\end{equation}
where $T_{i}=g_{i}-1$ and $N_{i}=\sum\limits_{j=0}^{m_{i}-1}g_{i}^{j}$.
Consider the tensor product of the above periodic resolutions of the cyclic factors of $G.$ The resulting complex, denoted by
$(K_{\bullet},d_{\bullet}),$ is as follows. For each sequence $a_{1},\ldots,a_{n}$ of nonnegative integers, let $\Phi(a_{1},\ldots,a_{n})$ be a free
generator in degree $a_{1}+\cdots+a_{n}$. Thus
$$K_{m}:=\bigoplus_{a_{1}+\cdots+a_{n}=m} (\mathbb{Z}G)\Phi(a_{1},\ldots,a_{n}).$$ For all $1\leq i\leq n,$ define
$$d_{i}(\Phi(a_{1},\ldots,a_{n}))=\left \{
\begin{array}{lll} 0, &\;\;\;\;a_{i}=0;
\\ (-1)^{\sum_{l<i}a_{l}}N_{i}\Phi(a_{1},\ldots,a_{i}-1,\ldots,a_{n}), & \;\;\;\;0\neq a_{i}\;\textrm{even;}
\\(-1)^{\sum_{l<i}a_{l}}T_{i}\Phi(a_{1},\ldots,a_{i}-1,\ldots,a_{n}), &
\;\;\;\;0\neq a_{i}\;\textrm{odd.}
\end{array} \right.$$
The differential $d$ is set to be $d_{1}+\cdots +d_{n}$. Then
$(K_{\bullet},d_{\bullet})$ is a free resolution of the trivial $\mathbb{Z}G$-module $\mathbb{Z}$. The main goal of this subsection is to determine the explicit cocycles of this Koszul-like resolution.
For the convenience of the exposition, we fix some notations before moving on. For any $1\le r_1<\cdots<r_l\le n$, define $\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}:=\Phi(0,\dots,\lambda_1,\dots,\lambda_l,\dots,0)$ where $\lambda_i\ge1$ lies in the $r_i$-th position. If $\lambda_i=1$ for some $1\le i\le l$, sometimes we drop it for brevity.
It is clear that any cochain $f\in\operatorname{Hom}_{\mathbb{Z}G}(K_k,\mathbbm{k}^*)$ is uniquely determined by its values on $\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}$. Write $f_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=f(\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}})$.
\begin{theorem}\label{complete}
The following \begin{equation} \label{ck} \left\{ f\in \operatorname{Hom}_{\mathbb{Z}G}(K_k,\mathbbm{k}^*)
\left|
\begin{array}{llll}
f_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=1\emph{ if }\lambda_1\emph{ is even}, \\
f_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=\zeta_{m_{r_1}}^{a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}}\emph{ if }\lambda_1\emph{ is odd}\\
\emph{and} \ 0\leq a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}<m_{r_1} \ \emph{for} \ 1\leq r_1<\cdots<r_l\leq n\\
\emph{where}\ \lambda_1+\cdots+\lambda_l=k,\ \lambda_i\ge1\ \emph{for}\ 1\le i\le l\\
\end{array}
\right. \right\} \end{equation}
makes a complete set of representatives of $k$-cocycles of the complex $(K_{\bullet}^*,d_{\bullet}^*)$.
\end{theorem}
\begin{proof}
Suppose $f\in \operatorname{Hom}_{\mathbb{Z}G}(K_k,\mathbbm{k}^*)$ is a $k$-cocycle. We will show that $f$ is cohomologous to one in \eqref{ck}. Let $g\in \operatorname{Hom}_{\mathbb{Z}G}(K_{k-1},\mathbbm{k}^*)$ be a $(k-1)$-cochain given by $g_{r_1^{\mu_1}\cdots r_l^{\mu_l}}=1$ if $\mu_1$ is even and
$g_{r_1^{\mu_1}\cdots r_l^{\mu_l}}=(f_{r_1^{\mu_1+1}\cdots r_l^{\mu_l}})^{\frac{1}{m_{r_1}}}$ if $\mu_1$ is odd. Consider $f'=f-d^*g.$ Then clearly $f'_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=1$ if $\lambda_1$ is even.
If $\lambda_1$ is odd, then by the cocycle condition for $f'$ we have
$$(f'_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}})^{m_{r_1}}\prod_{\begin{array}{cc}2\le i\le l\\\lambda_i\text{ even}\end{array}}(f'_{r_1^{\lambda_1+1}\cdots r_i^{\lambda_i-1}\cdots r_l^{\lambda_l}})^{(-1)^{\sum\limits_{j<i}\lambda_j+1}m_{r_i}}=1.$$
Hence $(f'_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}})^{m_{r_1}}=1$, so $f'_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=\zeta_{m_{r_1}}^{a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}}$ for some $0\le a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}<m_{r_1}$.
Suppose that $f_1$ and $f_2$ are two cocycles in \eqref{ck} satisfying $f_1-f_2=d^*h$ for some $(k-1)$-cochain $h\in \operatorname{Hom}_{\mathbb{Z}G}(K_{k-1},\mathbbm{k}^*).$ Similarly as above, after subtracting a $(k-1)$-coboundary from $h$, we can assume that $h_{r_1^{\mu_1}\cdots r_l^{\mu_l}}=1$ if $\mu_1$ is even.
If $\lambda_1$ is even, then
\begin{eqnarray*}
(f'_1-f'_2)_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}&=&(h_{r_1^{\lambda_1-1}\cdots r_l^{\lambda_l}})^{m_{r_1}}\prod_{\begin{array}{cc}2\le i\le l\\\lambda_i\text{ even}\end{array}}(h_{r_1^{\lambda_1}\cdots r_i^{\lambda_i-1}\cdots r_l^{\lambda_l}})^{(-1)^{\sum\limits_{j<i}\lambda_j}m_{r_i}}\\
&=&(h_{r_1^{\lambda_1-1}\cdots r_l^{\lambda_l}})^{m_{r_1}}=1.
\end{eqnarray*}
If $\lambda_1$ is odd, then by the preceding equation and the condition $m_i|m_{i+1}$ for $1\le i\le n-1$ we have
$$(f_1-f_2)_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=\prod_{\begin{array}{cc}2\le i\le l\\\lambda_i\text{ even}\end{array}}(h_{r_1^{\lambda_1}\cdots r_i^{\lambda_i-1}\cdots r_l^{\lambda_l}})^{(-1)^{\sum\limits_{j<i}\lambda_j}m_{r_i}}=1.$$
Hence $f_1=f_2$.
\end{proof}
\begin{corollary}
If $G=\mathbb{Z}_{m_{1}}\times\cdots \times\mathbb{Z}_{m_{n}}$ where $m_i|m_{i+1}$ for $1\le i\le n-1$, then
$$\operatorname{H}^k(G,\mathbbm{k}^*)=\prod_{r=1}^n\mathbb{Z}_{m_r}^{\sum\limits_{j=1}^k(-1)^{k+j}\binom{n-r+j-1}{j-1}}.$$
\end{corollary}
\begin{proof}
By Theorem \ref{complete}, $\operatorname{H}^k(G,\mathbbm{k}^*)=\prod_{r=1}^n\mathbb{Z}_{m_r}^{N_{k,r}}$ where
\begin{eqnarray*}
N_{k,r}&=&\#\{(r_2,\dots,r_l,\lambda_1,\dots,\lambda_l)\in\mathbb{N}^{2l-1}|r<r_2<\cdots<r_l\le n,\lambda_1+\cdots+\lambda_l=k,\lambda_1\text{ odd}\}\\
&=&\sum_{l=1}^k\binom{n-r}{l-1}\#\{(\lambda_1,\dots,\lambda_l)\in\mathbb{N}^l|\lambda_1+\cdots+\lambda_l=k,\lambda_1\text{ odd}\}.
\end{eqnarray*}
Denote $s_{k,l}=\#\{(\lambda_1,\dots,\lambda_l)\in\mathbb{N}^l|\lambda_1+\cdots+\lambda_l=k,\lambda_1\text{ odd}\}$ and $t_{k,l}=\#\{(\lambda_1,\dots,\lambda_l)\in\mathbb{N}^l|\lambda_1+\cdots+\lambda_l=k,\lambda_1\text{ even}\}.$ Then $s_{k,l}=t_{k+1,l}$ and $s_{k,l}+t_{k,l}=\binom{k-1}{l-1}$.
Hence $$N_{k,r}+N_{k-1,r}=\sum_{l=1}^k\binom{n-r}{l-1}\binom{k-1}{l-1}=\binom{n-r+k-1}{k-1}.$$
Therefore, $N_{k,r}=\sum\limits_{j=1}^k(-1)^{k+j}\binom{n-r+j-1}{j-1}$.
\end{proof}
\subsection{A chain map from $(B_{\bullet},\partial_{\bullet})$ to $(K_{\bullet},d_{\bullet})$}
We need some more notations to present our chain map. For any positive integers $s$ and $r,$ let $[\frac{s}{r}]$ denote the integer part of $\frac{s}{r}$ and let $s_r'$ denote the remainder of the division of $s$ by $r.$ When there is no risk of confusion, we omit the subscript in $s_r'.$ It is easy to observe that
\begin{equation} \label{f}
[\frac{s+t_r'}{r}]=[\frac{s+t-[\frac{t}{r}]r}{r}]=[\frac{s+t}{r}]-[\frac{t}{r}]
\end{equation}
for any three natural numbers $s, \ t $ and $r.$ We need the following technical lemma for later discussions.
\begin{lemma}\label{tl}
Let $r$ be a positive integer. For any $2l+1$ natural numbers $a_1, \ a_2, \dots, a_{2l+1},$ we have the following equation
\begin{eqnarray*}
&&\sum_{i=1}^l [\frac{a_{2l+1}+a_{2l}}{r}] \cdots [\frac{a_{2i+3}+a_{2i+2}}{r}] \left( [\frac{a_{2i+1}+(a_{2i}+a_{2i-1})'}{r}] - [\frac{(a_{2i+1}+a_{2i})'+a_{2i-1}}{r}] \right) \cdot \\
&& \qquad [\frac{a_{2i-2}+a_{2i-3}}{r}] \cdots [\frac{a_{2}+a_{1}}{r}] \\
&=& [\frac{a_{2l+1}+a_{2l}}{r}] \cdots [\frac{a_{3}+a_{2}}{r}] - [\frac{a_{2l}+a_{2l-1}}{r}] \cdots [\frac{a_{2}+a_{1}}{r}] .
\end{eqnarray*}
\end{lemma}
\begin{proof}
By \eqref{f}, we have \[ [\frac{a_{2i+1}+(a_{2i}+a_{2i-1})'}{r}] - [\frac{(a_{2i+1}+a_{2i})'+a_{2i-1}}{r}] = [\frac{a_{2i+1}+a_{2i}}{r}] - [\frac{a_{2i}+a_{2i-1}}{r}]. \]
Then the lemma follows by an obvious elimination of consecutive terms.
\end{proof}
Now we are ready to give a chain map from the normalized bar resolution $(B_{\bullet},\partial_{\bullet})$ to the Koszul-like resolution $(K_{\bullet},d_{\bullet}).$ Recall that $B_{m}$ is the free $\mathbb{Z}G$-module on the set of all symbols
$[h_{1},\ldots,h_{m}]$ with $h_{i}\in G$ and $m\geq 1$. In case $m=0$, the symbol $[\; ]$ denote $1\in \mathbb{Z}G$ and the map $\partial_{0}=\epsilon:\;B_{0}\to \mathbb{Z}$ sends
$[\; ]$ to $1$. For a more concise formulation, denote $(g_i)_r:=\sum\limits_{j=0}^{r-1}g_i^j$ for $1\leq i\leq n$ in the following.
The first four terms of the chain map, which will be used for later applications, are as follows:
\begin{eqnarray*}
F_{1}: &&B_{1}\longrightarrow K_{1}\\
&&[g_{1}^{i_{1}}\cdots g_{n}^{i_{n}}]\mapsto
\sum_{s=1}^{n}g_{1}^{i_{1}}\cdots g_{s-1}^{i_{s-1}}(g_{s})_{i_{s}}\Phi_{s};\\
F_{2}: &&B_{2}\longrightarrow K_{2}\\
&&[g_{1}^{i_{1}}\cdots g_{n}^{i_{n}},g_{1}^{j_{1}}\cdots g_{n}^{j_{n}}]\mapsto
\sum_{s=1}^{n}g_{1}^{i_{1}+j_{1}}\cdots g_{s-1}^{i_{s-1}+j_{s-1}}[\frac{i_{s}+j_{s}}{m_{s}}]\Phi_{s^2}\\
&&-\sum_{1\leq s<t\leq n}g_{1}^{i_{1}}\cdots g_{t-1}^{i_{t-1}}g_{1}^{j_{1}}\cdots g_{s-1}^{j_{s-1}}(g_{s})_{j_{s}}(g_{t})_{i_{t}}\Phi_{st};\\
F_{3}: &&B_{3}\longrightarrow K_{3}\\
&&[g_{1}^{i_{1}}\cdots g_{n}^{i_{n}},g_{1}^{j_{1}}\cdots g_{n}^{j_{n}},g_{1}^{k_{1}}\cdots g_{n}^{k_{n}}]\mapsto\\
&& \sum_{r=1}^{n}[\frac{j_{r}+k_{r}}{m_{r}}]g_{1}^{j_{1}+k_{1}}\cdots g_{r-1}^{j_{r-1}+k_{r-1}}g_{1}^{i_{1}}\cdots g_{r-1}^{i_{r-1}}(g_{r})_{i_{r}}\Phi_{r^3}+\\
&&\sum_{1\leq r<t\leq n}[\frac{j_{r}+k_{r}}{m_{r}}]g_{1}^{j_{1}+k_{1}}\cdots g_{r-1}^{j_{r-1}+k_{r-1}}g_{1}^{i_{1}}\cdots g_{t-1}^{i_{t-1}}(g_{t})_{i_{t}}\Phi_{r^2t}+\\
&&\sum_{1\leq r<t\leq n}[\frac{i_{t}+j_{t}}{m_{t}}]g_{1}^{i_{1}+j_{1}}\cdots g_{t-1}^{i_{t-1}+j_{t-1}}g_{1}^{k_{1}}\cdots g_{r-1}^{k_{r-1}}(g_{r})_{k_{r}}\Phi_{rt^2}-\\
&&\sum_{1\leq r<s<t\leq n}g_{1}^{i_{1}}\cdots g_{t-1}^{i_{t-1}}(g_{t})_{i_{t}}g_{1}^{j_{1}}\cdots g_{s-1}^{j_{s-1}}(g_{s})_{j_{s}}g_{1}^{k_{1}}\cdots g_{r-1}^{k_{r-1}}(g_{r})_{k_{r}}\Phi_{rst};\\
F_{4}: &&B_{4}\longrightarrow K_{4}\\
&&[g_{1}^{i_{1}}\cdots g_{n}^{i_{n}},g_{1}^{j_{1}}\cdots g_{n}^{j_{n}},g_{1}^{k_{1}}\cdots g_{n}^{k_{n}},g_{1}^{l_{1}}\cdots g_{n}^{l_{n}}]\mapsto\\
&& \sum_{r=1}^{n}[\frac{k_{r}+l_{r}}{m_{r}}][\frac{i_{r}+j_{r}}{m_{r}}]g_{1}^{i_{1}+j_{1}+k_{1}+l_{1}}\cdots g_{r-1}^{i_{r-1}+j_{r-1}+k_{r-1}+l_{r-1}}\Phi_{r^4}+\\
&&\sum_{1\leq r<s\leq n}[\frac{k_{r}+l_{r}}{m_{r}}]g_{1}^{k_{1}+l_{1}}\cdots g_{r-1}^{k_{r-1}+l_{r-1}}[\frac{i_{s}+j_{s}}{m_{s}}]g_{1}^{i_{1}+j_{1}}\cdots g_{s-1}^{i_{s-1}+j_{s-1}}\Phi_{r^2s^2}-\\
&&\sum_{1\leq r<s\leq n}[\frac{j_{s}+k_{s}}{m_{s}}]g_{1}^{j_{1}+k_{1}}\cdots g_{s-1}^{j_{s-1}+k_{s-1}}g_{1}^{l_{1}}\cdots g_{r-1}^{l_{r-1}}(g_{r})_{l_{r}}g_{1}^{i_{1}}\cdots g_{s-1}^{i_{s-1}}(g_{s})_{i_{s}}\Phi_{rs^3}-\\
&&\sum_{1\leq r<s\leq n}[\frac{k_{r}+l_{r}}{m_{r}}]g_{1}^{k_{1}+l_{1}}\cdots g_{r-1}^{k_{r-1}+l_{r-1}}g_{1}^{j_{1}}\cdots g_{r-1}^{j_{r-1}}(g_{r})_{j_{r}}g_{1}^{i_{1}}\cdots g_{s-1}^{i_{s-1}}(g_{s})_{i_{s}}\Phi_{r^3s}-\\
&&\sum_{1\leq r<s<t\leq n}[\frac{k_{r}+l_{r}}{m_{r}}]g_{1}^{k_{1}+l_{1}}\cdots g_{r-1}^{k_{r-1}+l_{r-1}}g_{1}^{j_{1}}\cdots g_{s-1}^{j_{s-1}}(g_{s})_{j_{s}}g_{1}^{i_{1}}\cdots g_{t-1}^{i_{t-1}}(g_{t})_{i_{t}}\Phi_{r^2st}-\\
&&\sum_{1\leq r<s<t\leq n}[\frac{j_{s}+k_{s}}{m_{s}}]g_{1}^{j_{1}+k_{1}}\cdots g_{s-1}^{j_{s-1}+k_{s-1}}g_{1}^{l_{1}}\cdots g_{r-1}^{l_{r-1}}(g_{r})_{l_{r}}g_{1}^{i_{1}}\cdots g_{t-1}^{i_{t-1}}(g_{t})_{i_{t}}\Phi_{rs^2t}-\\
&&\sum_{1\leq r<s<t\leq n}[\frac{i_{t}+j_{t}}{m_{t}}]g_{1}^{i_{1}+j_{1}}\cdots g_{t-1}^{i_{t-1}+j_{t-1}}g_{1}^{l_{1}}\cdots g_{r-1}^{l_{r-1}}(g_{r})_{l_{r}}g_{1}^{k_{1}}\cdots g_{s-1}^{k_{s-1}}(g_{s})_{k_{s}}\Phi_{rst^2}+\\
&&\sum_{1\leq r<s<t<u\leq n}g_{1}^{l_{1}}\cdots g_{r-1}^{l_{r-1}}(g_{r})_{l_{r}}g_{1}^{k_{1}}\cdots g_{s-1}^{k_{s-1}}(g_{s})_{k_{s}}g_{1}^{j_{1}}\cdots g_{t-1}^{j_{t-1}}(g_{t})_{j_{t}}g_{1}^{i_{1}}\cdots g_{u-1}^{i_{u-1}}(g_{u})_{i_{u}}\Phi_{rstu}
\end{eqnarray*}
for $0\leq i_{r},j_{r},k_{r},l_{r}< m_{r}$ and $1\leq r\leq n$.
In general, let $\alpha:=(\alpha_{11},\dots,\alpha_{1n},\dots,\alpha_{k1},\dots,\alpha_{kn})$ where each $\alpha_{ij} \in [0, m_j)$ and is viewed as an integer modulo $m_j$ for all $1\leq i\leq k.$ We also write $\alpha=(\alpha_1,\dots,\alpha_k)$ where $\alpha_u=(\alpha_{u1},\dots,\alpha_{un})$ for $1\leq u\leq k.$
For brevity, in the following we denote the group element $g_{1}^{\alpha_{i1}}\cdots g_{n}^{\alpha_{in}}$ by $g^{\alpha_i}.$ Given any $1\leq r\leq n$, $[a,b]\subseteq[1,k]$, $a,b\in\mathbb{N}$ and $\alpha$, denote
$$\xi_{r,[a,b]}^{\alpha}:=\left\{
\begin{array}{ll}[\frac{\alpha_{br}+\alpha_{b-1,r}}{m_{r}}]\cdots[\frac{\alpha_{a+1,r}+\alpha_{ar}}{m_{r}}]g_{1}^{\alpha_{b1}+\cdots+\alpha_{a1}}\cdots g_{r-1}^{\alpha_{b,r-1}+\cdots+\alpha_{a,r-1}},&\;\;\;a-b\text{ odd;}\\
[\frac{\alpha_{br}+\alpha_{b-1,r}}{m_{r}}]\cdots[\frac{\alpha_{a+2,r}+\alpha_{a+1,r}}{m_{r}}]g_{1}^{\alpha_{b1}+\cdots+\alpha_{a,1}}\cdots g_{r-1}^{\alpha_{b,r-1}+\cdots+\alpha_{a,r-1}}(g_{r})_{\alpha_{ar}},&\;\;\;a-b\text{ even.}
\end{array}\right.$$
Define
\begin{eqnarray}\label{Fkabelian}
&& F_{k} \colon B_{k}\longrightarrow K_{k} \\
&&[g^{\alpha_1},\dots,g^{\alpha_{k}}]\mapsto \sum_{l=1}^{k}\sum_{\begin{array}{ccc}1\leq r_{1}<\cdots<r_{l}\leq n\\\lambda_1+\cdots+\lambda_l=k\\\lambda_i\ge1\text{ for }1\le i\le l\end{array}}(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_{i}\lambda_{j}}\xi_{r_{1},[a_{1},b_{1}]}^{\alpha}\cdots\xi_{r_{l},[a_{l},b_{l}]}^{\alpha}\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}} \notag
\end{eqnarray}
where $a_u = {\sum\limits_{i=u+1}^l \lambda_i}+1$ and $b_u = {\sum\limits_{i=u}^l \lambda_i}$ for $1 \le u \le l.$ Clearly, the interval $[1,k]$ is the disjoint union of the $[a_i,b_i]$'s.
\begin{proposition} \label{chainmap}
The following diagram is commutative.
\begin{figure}[hbt]
\begin{picture}(150,50)(50,-40)
\put(0,0){\makebox(0,0){$ \cdots$}}\put(10,0){\vector(1,0){20}}\put(40,0){\makebox(0,0){$B_{3}$}}
\put(50,0){\vector(1,0){20}}\put(80,0){\makebox(0,0){$B_{2}$}}
\put(90,0){\vector(1,0){20}}\put(120,0){\makebox(0,0){$B_{1}$}}
\put(130,0){\vector(1,0){20}}\put(160,0){\makebox(0,0){$B_{0}$}}
\put(170,0){\vector(1,0){20}}\put(200,0){\makebox(0,0){$\mathbb{Z}$}}
\put(210,0){\vector(1,0){20}}\put(240,0){\makebox(0,0){$0$}}
\put(0,-40){\makebox(0,0){$ \cdots$}}\put(10,-40){\vector(1,0){20}}\put(40,-40){\makebox(0,0){$K_{3}$}}
\put(50,-40){\vector(1,0){20}}\put(80,-40){\makebox(0,0){$K_{2}$}}
\put(90,-40){\vector(1,0){20}}\put(120,-40){\makebox(0,0){$K_{1}$}}
\put(130,-40){\vector(1,0){20}}\put(160,-40){\makebox(0,0){$K_{0}$}}
\put(170,-40){\vector(1,0){20}}\put(200,-40){\makebox(0,0){$\mathbb{Z}$}}
\put(210,-40){\vector(1,0){20}}\put(240,-40){\makebox(0,0){$0$}}
\put(40,-10){\vector(0,-1){20}}
\put(80,-10){\vector(0,-1){20}}
\put(120,-10){\vector(0,-1){20}}
\put(158,-10){\line(0,-1){20}}\put(160,-10){\line(0,-1){20}}
\put(198,-10){\line(0,-1){20}}\put(200,-10){\line(0,-1){20}}
\put(60,5){\makebox(0,0){$\partial_{3}$}}
\put(100,5){\makebox(0,0){$\partial_{2}$}}
\put(140,5){\makebox(0,0){$\partial_{1}$}}
\put(60,-35){\makebox(0,0){$d$}}
\put(100,-35){\makebox(0,0){$d$}}
\put(140,-35){\makebox(0,0){$d$}}
\put(50,-20){\makebox(0,0){$F_{3}$}}
\put(90,-20){\makebox(0,0){$F_{2}$}}
\put(130,-20){\makebox(0,0){$F_{1}$}}
\end{picture}
\end{figure}
\end{proposition}
\begin{proof}
We start with some conventions. Denote
$$\mathcal{E}_{r^{\lambda}}:=\left\{
\begin{array}{ll}
N_r,&\;\;\;\lambda\text{ even;}\\
T_r,&\;\;\;\lambda\text{ odd.}
\end{array}\right.$$
Then
\begin{eqnarray*}
d\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}&=&\mathcal{E}_{r_1^{\lambda_1}}\Phi_{r_1^{\lambda_1-1}\cdots r_l^{\lambda_l}}+
(-1)^{\lambda_1}\mathcal{E}_{r_2^{\lambda_2}}\Phi_{r_1^{\lambda_1}r_2^{\lambda_2-1}\cdots r_l^{\lambda_l}} +\cdots+
(-1)^{\lambda_1+\cdots+\lambda_{l-1}}\mathcal{E}_{r_l^{\lambda_l}}\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l-1}} \ .
\end{eqnarray*}
For any given $\alpha=(\alpha_{11},\dots,\alpha_{1n},\dots,\alpha_{k1},\dots,\alpha_{kn}),$ let $\alpha'=(\alpha_2,\dots,\alpha_k), \ \alpha''=(\alpha_1,\dots,\alpha_{k-1})$ and $\alpha'_u=(\alpha_1,\dots,\alpha_{u-1},\alpha_u+\alpha_{u+1},\alpha_{u+2},\dots,\alpha_k), \quad \forall 1\leq u\leq k-1.$
With the above notations, $\partial_k([g^{\alpha_1}, \dots, g^{\alpha_k}])$ becomes
\[
g^{\alpha_1}[g^{\alpha_2}, \dots, g^{\alpha_k}]+\sum_{u=1}^{k-1}(-1)^u [g^{\alpha_1},\dots,
g^{\alpha_{u-1}}, g^{\alpha_u+\alpha_{u+1}}, g^{\alpha_{u+2}}, \dots, g^{\alpha_k}]+(-1)^k[g^{\alpha_1}, \dots, g^{\alpha_{k-1}}].
\]
Then the coefficient of $\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}$ in $F_{k-1}\partial_k([g^{\alpha_1}, \dots, g^{\alpha_k}])$ is
\begin{equation} \label{c1}
(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j} \left( g^{\alpha_1}\xi_{r_1,[a_1,b_1]}^{\alpha'}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha'}+\sum_{u=1}^{k-1}(-1)^u\xi_{r_1,[a_1,b_1]}^{\alpha'_u}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha'_u}
+(-1)^k\xi_{r_1,[a_1,b_1]}^{\alpha''}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha''} \right)
\end{equation}
where $a_u = {\sum_{i=u+1}^l \lambda_i}+1$ and $b_u = {\sum_{i=u}^l \lambda_i}.$
For $1\leq r\leq n$, $[a,b]\subseteq[1,k-1]$, $a,b\in\mathbb{N}$ and $\alpha$, we have
$$ \xi_{r,[a,b]}^{\alpha'}=\xi_{r,[a+1,b+1]}^{\alpha}, \ \ \xi_{r,[a,b]}^{\alpha''}=\xi_{r,[a,b]}^{\alpha} \quad \text{and} \quad
\xi_{r,[a,b]}^{\alpha'_u}=\left\{
\begin{array}{ll}
\xi_{r,[a+1,b+1]}^{\alpha}, \ \ \ &\text{ if } u<a;\\
\xi_{r,[a,b]}^{\alpha}, \ \ \ &\text{ if } u>b.
\end{array}\right.
$$
Hence
\begin{eqnarray*}
&&\sum_{u=1}^{k-1}(-1)^u\xi_{r_1,[a_1,b_1]}^{\alpha'_u}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha'_u}\\
&=&\sum_{i=1}^l \sum_{u=a_i}^{b_i}(-1)^u \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots\xi_{r_{i-1},[a_{i-1}+1,b_{i-1}+1]}^{\alpha} \xi_{r_i,[a_i,b_i]}^{\alpha'_u}\xi_{r_{i+1},[a_{i+1},b_{i+1}]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha} \\
&=&\sum_{i=1}^l \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots\xi_{r_{i-1},[a_{i-1}+1,b_{i-1}+1]}^{\alpha} \xi_{r_{i+1},[a_{i+1},b_{i+1}]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha} \sum_{u=a_i}^{b_i}(-1)^u \xi_{r_i,[a_i,b_i]}^{\alpha'_u}.
\end{eqnarray*}
Therefore we can rewrite \eqref{c1} as
\begin{eqnarray} \label{c1'}
&& (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j} \left( g^{\alpha_1}\xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots\xi_{r_l,[a_l+1,b_l+1]}^{\alpha} +(-1)^k\xi_{r_1,[a_1,b_1]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha} \right. \\
&&\qquad + \sum_{i=1}^l \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots\xi_{r_{i-1},[a_{i-1}+1,b_{i-1}+1]}^{\alpha} \xi_{r_{i+1},[a_{i+1},b_{i+1}]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha} \sum_{u=a_i}^{b_i}(-1)^u \left. \xi_{r_i,[a_i,b_i]}^{\alpha'_u} \right). \notag
\end{eqnarray}
It remains to compute $\sum\limits_{u=a_i}^{b_i}(-1)^u \xi_{r_i,[a_i,b_i]}^{\alpha'_u}.$ This is split into two cases according to the parity of $b_i-a_i.$
If $b_i-a_i$ is odd, then
\begin{eqnarray} \label{odd}
&&\sum_{u=a_i}^{b_i}(-1)^u\xi_{r_i,[a_i,b_i]}^{\alpha'_u}\\
&=&\left( (-1)^{a_i}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+2,r_i}+(\alpha_{a_i,r_i}+\alpha_{a_i+1,r_i})'}{m_{r_i}}] \right.\notag\\
&&\quad +(-1)^{a_i+1}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{(\alpha_{a_i+1,r_i}+\alpha_{a_i+2,r_i})'+\alpha_{a_i,r_i}}{m_{r_i}}] \notag\\
&&\quad +\cdots \left. +(-1)^{b_i}[\frac{(\alpha_{b_i,r_i}+\alpha_{b_i+1,r_i})'+\alpha_{b_i-1,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+1,r_i}+\alpha_{a_i,r_i}}{m_{r_i}}] \right) \cdot \notag \\
&& \qquad g_1^{\alpha_{b_i+1,1}+\cdots+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}} \notag \\
&\overset{Lemma \ref{tl}}{=}&\left( (-1)^{a_i}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+2,r_i}+\alpha_{a_i+1,r_i}}{m_{r_i}}] + (-1)^{b_i}[\frac{\alpha_{b_i,r_i}+\alpha_{b_i-1,r_i}}{m_{r_i}}]\cdots [\frac{\alpha_{a_i+1,r_i}+\alpha_{a_i,r_i}}{m_{r_i}}] \right) \cdot \notag\\
&& \qquad g_1^{\alpha_{b_i+1,1}+\cdots+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}} \notag\\
&=&(-1)^{b_i}\xi_{r_i,[a_i,b_i]}^{\alpha}g_1^{\alpha_{b_i+1,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}}+(-1)^{a_i}\xi_{r_i,[a_i+1,b_i+1]}^{\alpha}g_1^{\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{a_i,r_i-1}}. \notag
\end{eqnarray}
If $b_i-a_i$ is even, then similarly we have
\begin{eqnarray} \label{even}
&& \qquad \sum_{u=a_i}^{b_i}(-1)^u\xi_{r_i,[a_i,b_i]}^{\alpha'_u}\\
&=&(-1)^{a_i}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+3,r_i}+\alpha_{a_i+2,r_i}}{m_{r_i}}]g_1^{\alpha_{b_i+1,1}+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}} (g_{r_i})_{(\alpha_{a_i,r_i}+\alpha_{a_i+1,r_i})'} \notag\\
&&+\left( (-1)^{a_i+1}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]
\cdots[\frac{\alpha_{a_i+3,r_i}+(\alpha_{a_i+1,r_i}+\alpha_{a_i+2,r_i})'}{m_{r_i}}] \right. \notag\\
&&\qquad +(-1)^{a_i+2}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{(\alpha_{a_i+2,r_i}+\alpha_{a_i+3,r_i})'+\alpha_{a_i+1,r_i}}{m_{r_i}}] \notag\\
&&\qquad +\cdots \notag \\
&&\qquad \left. +(-1)^{b_i}[\frac{(\alpha_{b_i,r_i}+\alpha_{b_i+1,r_i})'+\alpha_{b_i-1,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+2,r_i}+\alpha_{a_i+1,r_i}}{m_{r_i}}] \right)\cdot \notag \\
&&\qquad \qquad g_1^{\alpha_{b_i+1,1}+\cdots+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}}(g_{r_i})_{\alpha_{a_i,r_i}}\notag\\
&=&(-1)^{a_i}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+3,r_i}+\alpha_{a_i+2,r_i}}{m_{r_i}}]g_1^{\alpha_{b_i+1,1}+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}}\cdot \notag\\
&&\qquad \left( (g_{r_i})_{\alpha_{a_i,r_i}+\alpha_{a_i+1,r_i}}-[\frac{\alpha_{a_i,r_i}+\alpha_{a_i+1,r_i}}{m_{r_i}}]N_{r_i} \right) \notag\\
&&+\left( (-1)^{a_i+1}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+3,r_i}+\alpha_{a_i+2,r_i}}{m_{r_i}}] + (-1)^{b_i}[\frac{\alpha_{b_i,r_i}+\alpha_{b_i-1,r_i}}{m_{r_i}}]\cdots [\frac{\alpha_{a_i+2,r_i}+\alpha_{a_i+1,r_i}}{m_{r_i}}] \right) \cdot \notag\\
&&\qquad g_1^{\alpha_{b_i+1,1}+\cdots+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}}(g_{r_i})_{\alpha_{a_i,r_i}}\notag\\
&=&(-1)^{a_i}[\frac{\alpha_{b_i+1,r_i}+\alpha_{b_i,r_i}}{m_{r_i}}]\cdots[\frac{\alpha_{a_i+3,r_i}+\alpha_{a_i+2,r_i}}{m_{r_i}}]g_1^{\alpha_{b_i+1,1}+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}}\cdot \notag\\
&&\qquad \left( (g_{r_i})_{\alpha_{a_i,r_i}+\alpha_{a_i+1,r_i}}-(g_{r_i})_{\alpha_{a_i,r_i}}-[\frac{\alpha_{a_i,r_i}+\alpha_{a_i+1,r_i}}{m_{r_i}}]N_{r_i} \right)\notag\\
&&+ (-1)^{b_i}[\frac{\alpha_{b_i,r_i}+\alpha_{b_i-1,r_i}}{m_{r_i}}]\cdots [\frac{\alpha_{a_i+2,r_i}+\alpha_{a_i+1,r_i}}{m_{r_i}}]
g_1^{\alpha_{b_i+1,1}+\cdots+\alpha_{a_i,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}+\cdots+\alpha_{a_i,r_i-1}}(g_{r_i})_{\alpha_{a_i,r_i}}\notag\\
&=&(-1)^{a_i}\xi_{r_i,[a_i+1,b_i+1]}^{\alpha}g_1^{\alpha_{a_i,1}}\cdots g_{r_i}^{\alpha_{a_i,r_i}} + (-1)^{a_i+1}\xi_{r_i,[a_i,b_i+1]}^{\alpha} \mathcal{E}_{{r_i}^{b_i-a_i}} + (-1)^{b_i}\xi_{r_i,[a_i,b_i]}^{\alpha}g_1^{\alpha_{b_i+1,1}}\cdots g_{r_i-1}^{\alpha_{b_i+1,r_i-1}}. \notag
\end{eqnarray}
On the other hand, the term $\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}$ in $dF_k([g^{\alpha_1}, \dots, g^{\alpha_k}])$ comes from the differential of the terms
\[ \Phi_{r_1^{\lambda_1+1} r_2^{\lambda_2} \cdots r_l^{\lambda_l}}, \ \cdots, \ \Phi_{r_1^{\lambda_1} \cdots r_{l-1}^{\lambda_{l-1}} r_l^{\lambda_l+1}}, \ \ \Phi_{s r_1^{\lambda_1}\cdots r_l^{\lambda_l}},\ \ \Phi_{r_1^{\lambda_1} s r_2^{\lambda_2} \cdots r_l^{\lambda_l}}, \ \cdots, \ \Phi_{r_1^{\lambda_1} \cdots r_{l-1}^{\lambda_{l-1}} s r_l^{\lambda_l}}, \ \ \Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}s} \]
in $F_k([g^{\alpha_1}, \dots, g^{\alpha_k}]).$ Therefore, its coefficient is
\begin{eqnarray} \label{c2}
&&\sum_{1\leq u \leq l}(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{j \ne u} \lambda_j} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha} \cdots \xi_{r_{u-1},[a_{u-1}+1,b_{u-1}+1]}^{\alpha} \xi_{r_u,[a_u,b_u+1]}^{\alpha} \cdot \\
&&\qquad \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l, [a_l, b_l]}^\alpha (-1)^{\lambda_1+\cdots+\lambda_{u-1}} \mathcal{E}_{r_u^{\lambda_u+1}} \notag \\
&+&\sum_{s=1}^{r_1-1}(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l \lambda_i} \xi_{r_1,[a_1,b_1]}^{\alpha} \cdots \xi_{r_l,[a_l, b_l]}^{\alpha}\xi_{s,[k,k]}^{\alpha}T_s\notag\\%
&+&\sum_{u=1}^{l-1} \sum_{s=r_u+1}^{r_{u+1}-1} (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} \cdot \notag \\
&& \qquad \xi_{s, [a_u, a_u]} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha}
(-1)^{\lambda_1 + \cdots + \lambda_u} T_s \notag\\
&+&\sum_{s=r_l+1}^n(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha} \cdots \xi_{r_l,[a_l+1,b_l+1]}^{\alpha} \xi_{s,[1,1]}^{\alpha}(-1)^{\sum\limits_{i=1}^l\lambda_i}T_s. \notag
\end{eqnarray}
Noting that $\xi_{s,[a,a]}^{\alpha}T_s=g_1^{\alpha_{a1}}\cdots g_{s-1}^{\alpha_{a,s-1}}(g_s^{\alpha_{a,s}}-1)$, then one has the following equations:
\begin{eqnarray*}
&&\sum_{s=1}^{r_1-1}(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l \lambda_i} \xi_{r_1,[a_1,b_1]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha}\xi_{s,[k,k]}^{\alpha}T_s\\
&=&(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l \lambda_i} \xi_{r_1,[a_1,b_1]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha} \sum_{s=1}^{r_1-1} \xi_{s,[k,k]}^{\alpha}T_s\\
&=&(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l\lambda_i}\xi_{r_1,[a_1,b_1]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha}(g_1^{\alpha_{k1}}\cdots g_{r_1-1}^{\alpha_{k,r_1-1}}-1),\\ &&\\
&&\sum_{u=1}^{l-1} \sum_{s=r_u+1}^{r_{u+1}-1} (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} \cdot \notag \\
&& \qquad \xi_{s, [a_u, a_u]} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha}
(-1)^{\lambda_1 + \cdots + \lambda_u} T_s \notag\\
&=&\sum_{u=1}^{l-1} (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} \cdot \notag \\
&& \qquad \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha}
\sum_{s=r_u+1}^{r_{u+1}-1} \xi_{s, [a_u, a_u]} T_s \notag\\%
&=&\sum_{u=1}^{l-1} (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} \cdot \notag \\
&& \qquad \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha}
(g_1^{\alpha_{a_u 1}} \cdots g_{r_{u+1}-1} ^{\alpha_{a_u, r_{u+1}-1}} - g_1^{\alpha_{a_u 1}} \cdots g_{r_u}^{\alpha_{a_u, r_u}}), \notag\\%
&&\\
&&\sum_{s=r_l+1}^n(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l \lambda_i} \xi_{r_1,[a_1+1, b_1+1]}^{\alpha} \cdots \xi_{r_l,[a_l+1,b_l+1]}^{\alpha} \xi_{s,[1,1]}^{\alpha}(-1)^{\sum_{i=1}^l\lambda_i}T_s\\
&=&(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j} \xi_{r_1,[a_1+1, b_1+1]}^{\alpha} \cdots \xi_{r_l,[a_l+1,b_l+1]}^{\alpha} (g_1^{\alpha_{11}}\cdots g_n^{\alpha_{1n}}-g_1^{\alpha_{11}}\cdots g_{r_l}^{\alpha_{1,r_l}}).
\end{eqnarray*}
With these, \eqref{c2} becomes
\begin{eqnarray} \label{c3}
&&\sum_{1\leq u \leq l}(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{j=u+1}^l \lambda_j} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha} \cdots \xi_{r_{u-1},[a_{u-1}+1,b_{u-1}+1]}^{\alpha} \cdot \\%1
&&\qquad \xi_{r_u,[a_u,b_u+1]}^{\alpha} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l, [a_l, b_l]}^\alpha \mathcal{E}_{r_u^{\lambda_u+1}} \notag \\
&+& (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=1}^l\lambda_i}\xi_{r_1,[a_1,b_1]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha}(g_1^{\alpha_{k1}}\cdots g_{r_1-1}^{\alpha_{k,r_1-1}}-1) \notag \\%2
&+& \sum_{u=1}^{l-1} (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j+\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} \cdot \notag \\
&& \qquad \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha}
(g_1^{\alpha_{a_u 1}} \cdots g_{r_{u+1}-1} ^{\alpha_{a_u, r_{u+1}-1}} - g_1^{\alpha_{a_u 1}} \cdots g_{r_u}^{\alpha_{a_u, r_u}}) \notag \\%3
&+& (-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_i\lambda_j} \xi_{r_1,[a_1+1, b_1+1]}^{\alpha} \cdots \xi_{r_l,[a_l+1,b_l+1]}^{\alpha} (g_1^{\alpha_{11}}\cdots g_n^{\alpha_{1n}}-g_1^{\alpha_{11}}\cdots g_{r_l}^{\alpha_{1,r_l}}). \notag
\end{eqnarray}
We need to prove that the two formulas \eqref{c1'} and \eqref{c3} are equal.
By cancelling their obvious common terms, namely the first two terms of \eqref{c1'}, it suffices to prove
\begin{eqnarray*}
&&\sum_{i=1}^l \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots\xi_{r_{i-1},[a_{i-1}+1,b_{i-1}+1]}^{\alpha} \xi_{r_{i+1},[a_{i+1},b_{i+1}]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha} \sum_{u=a_i}^{b_i}(-1)^u \xi_{r_i,[a_i,b_i]}^{\alpha'_u} \\
&=&\sum_{u=1}^l (-1)^{\sum\limits_{i=u+1}^l\lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha} \cdots \xi_{r_{u-1},[a_{u-1}+1,b_{u-1}+1]}^{\alpha} \xi_{r_u,[a_u,b_u+1]}^{\alpha} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l, [a_l, b_l]}^\alpha \mathcal{E}_{r_u^{\lambda_u+1}} \notag \\
&&+ \sum_{u=1}^{l-1} (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha} \cdot \notag \\
&& \qquad (g_1^{\alpha_{a_u 1}} \cdots g_{r_{u+1}-1} ^{\alpha_{a_u, r_{u+1}-1}} - g_1^{\alpha_{a_u 1}} \cdots g_{r_u}^{\alpha_{a_u, r_u}}) \notag \\%3
&&+ (-1)^{\sum\limits_{i=1}^l\lambda_i}\xi_{r_1,[a_1,b_1]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha} g_1^{\alpha_{k1}}\cdots g_{r_1-1}^{\alpha_{k,r_1-1}} - \xi_{r_1,[a_1+1, b_1+1]}^{\alpha} \cdots \xi_{r_l,[a_l+1,b_l+1]}^{\alpha} g_1^{\alpha_{11}}\cdots g_{r_l}^{\alpha_{1,r_l}}. \notag
\end{eqnarray*}
Note that the latter is equal to
\begin{eqnarray*}
&&\sum_{u=1}^l (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha} \cdots \xi_{r_{u-1},[a_{u-1}+1,b_{u-1}+1]}^{\alpha} \xi_{r_u,[a_u,b_u+1]}^{\alpha} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l, [a_l, b_l]}^\alpha \mathcal{E}_{r_u^{\lambda_u+1}}\\
&&+ \sum_{u=1}^{l} (-1)^{\sum\limits_{i=u}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_{u-1},[a_{u-1}+1,b_{u-1}+1]}^{\alpha} \xi_{r_{u},[a_{u},b_{u}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha} g_1^{\alpha_{a_{u-1} 1}} \cdots g_{r_{u}-1} ^{\alpha_{a_{u-1}, r_{u}-1}} \notag \\
&&- \sum_{u=1}^{l} (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha} \cdots \xi_{r_l,[a_l,b_l]}^{\alpha} g_1^{\alpha_{a_u 1}} \cdots g_{r_u}^{\alpha_{a_u, r_u}} \notag \\
&=&\sum_{u=1}^l \xi_{r_1,[a_1+1,b_1+1]}^{\alpha}\cdots\xi_{r_{u-1},[a_{u-1}+1,b_{u-1}+1]}^{\alpha} \xi_{r_{u+1},[a_{u+1},b_{u+1}]}^{\alpha}\cdots\xi_{r_l,[a_l,b_l]}^{\alpha} \cdot \left( (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_u,[a_u,b_u+1]}^{\alpha} \mathcal{E}_{r_u^{\lambda_u+1}}
\right. \\
&&\left. \qquad + (-1)^{\sum\limits_{i=u}^l \lambda_i} \xi_{r_{u},[a_{u},b_{u}]}^{\alpha} g_1^{\alpha_{a_{u-1} 1}} \cdots g_{r_{u}-1} ^{\alpha_{a_{u-1}, r_{u}-1}}
- (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} g_1^{\alpha_{a_u 1}} \cdots g_{r_u}^{\alpha_{a_u, r_u}} \right).
\end{eqnarray*}
Now it is enough to verify that
\begin{eqnarray} \label{me}
&&\sum_{u=a_i}^{b_i}(-1)^u \xi_{r_i,[a_i,b_i]}^{\alpha'_u}=(-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_u,[a_u,b_u+1]}^{\alpha} \mathcal{E}_{r_u^{\lambda_u+1}} + (-1)^{\sum\limits_{i=u}^l \lambda_i} \xi_{r_{u},[a_{u},b_{u}]}^{\alpha} g_1^{\alpha_{a_{u-1} 1}} \cdots g_{r_{u}-1} ^{\alpha_{a_{u-1}, r_{u}-1}} \\
&&\qquad \qquad \qquad \qquad \qquad - (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} g_1^{\alpha_{a_u 1}} \cdots g_{r_u}^{\alpha_{a_u, r_u}}. \notag
\end{eqnarray}
The verification is split into two cases. If $b_i-a_i$ is even, then the equality is immediate simply by noting that
\[a_u = {\sum_{i=u+1}^l \lambda_i}+1, \quad b_u = {\sum_{i=u}^l \lambda_i} . \] If $b_i-a_i$ is odd, noting that
\begin{eqnarray*}
(-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_u,[a_u,b_u+1]}^{\alpha} \mathcal{E}_{r_u^{\lambda_u+1}} &=& (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} g_1^{\alpha_{a_u,1}} \cdots g_{r_u-1}^{\alpha_{a_u,r_u-1}}(g_{r_u})_{\alpha_{a_u,r_u}} (g_{r_u}-1) \\
&=& (-1)^{\sum\limits_{i=u+1}^l \lambda_i} \xi_{r_u,[a_u+1,b_u+1]}^{\alpha} g_1^{\alpha_{a_u,1}} \cdots g_{r_u-1}^{\alpha_{a_u,r_u-1}} (g_{r_u}^{\alpha_{a_u,r_u}} -1),
\end{eqnarray*} then the equality \eqref{me} follows.
The proof is completed.
\end{proof}
\subsection{Normalized cocycles}
Denote
$$\eta_{r,[a,b]}^{\alpha}:=\left\{
\begin{array}{ll}[\frac{\alpha_{br}+\alpha_{b-1,r}}{m_{r}}]\cdots[\frac{\alpha_{a+1,r}+\alpha_{ar}}{m_{r}}],&\;\;\;b-a\text{ odd;}\\
[\frac{\alpha_{br}+\alpha_{b-1,r}}{m_{r}}]\cdots[\frac{\alpha_{a+2,r}+\alpha_{a+1,r}}{m_{r}}]\alpha_{ar},&\;\;\;b-a\text{ even.}
\end{array}\right.$$
\begin{corollary}
The following $k$-cochains $\omega\in \operatorname{Hom}_{\mathbb{Z}G}(B_k,\mathbbm{k}^*)$ given by
\begin{equation}\label{barcocycleabelian}
\omega([g^{\alpha_1}, \dots, g^{\alpha_{k}}])=\prod_{l=1}^{k}\prod_{\begin{array}{ccc}1\leq r_{1}<\cdots<r_{l}\leq n\\\lambda_1+\cdots+\lambda_l=k,\lambda_1\emph{ odd}\\\lambda_i\ge1\emph{ for }1\le i\le l\end{array}}\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_{i}\lambda_{j}}\eta_{r_{1},[a_{1},b_{1}]}^{\alpha}\cdots\eta_{r_{l},[a_{l},b_{l}]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}}
\end{equation}
where $a_u = {\sum\limits_{i=u+1}^l \lambda_i}+1, \quad b_u = {\sum\limits_{i=u}^l \lambda_i}$ and $0\leq a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}<m_{r_1}$ for $1\leq r_1<\cdots<r_l\leq n$
form a complete set of representatives of $k$-cocycles of the complex $(B_{\bullet}^*,\partial_{\bullet}^*)$.
\end{corollary}
\begin{proof}
It follows from the chain map \eqref{Fkabelian} and Theorem \ref{complete}.
\end{proof}
\subsection{A chain map from $(K_{\bullet},d_{\bullet})$ to $(B_{\bullet},\partial_{\bullet})$}
For completeness, we also include a chain map from the Koszul-like resolution $(K_{\bullet},d_{\bullet})$ to the normalized bar resolution $(B_{\bullet},\partial_{\bullet}).$ This chain map is very useful for comparing cohomology classes of normalized cocycles and for studying the whole cohomology ring structure, etc.
Denote an ordered set of $\lambda$ elements as
$$\Lambda_{r^{\lambda}}:=\left\{\begin{array}{ll}(N_r,g_r,N_r,g_r,\dots,N_r,g_r),&\lambda\text{ even;}\\(g_r,N_r,g_r,N_r,g_r\dots,N_r,g_r),&\lambda\text{ odd.}\end{array}\right.$$
Given a set of positive integers $\lambda_1, \ \lambda_2, \dots, \lambda_l$ with $\lambda_1+\cdots+\lambda_l=k$,
let $\operatorname{Shuffle}(\lambda_1,\dots,\lambda_l)$ be the subset of the permutation group $S_k$ such that the elements of it preserve the order of elements of each block of the partition $(\lambda_1,\dots,\lambda_l)$. For each $k,$ define a map
\begin{eqnarray*}
G_{k} \colon K_{k} &\longrightarrow& B_{k} \\
\Phi_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}} &\mapsto& \sum_{\sigma\in \operatorname{Shuffle}(\lambda_1,\dots,\lambda_l)}(-1)^{\sigma}[\sigma(\Lambda_{r_1^{\lambda_1}},\dots,\Lambda_{r_l^{\lambda_l}})].
\end{eqnarray*}
\begin{proposition}
We have the following commutative diagram
\begin{figure}[hbt]
\begin{picture}(150,50)(50,-40)
\put(0,0){\makebox(0,0){$ \cdots$}}\put(10,0){\vector(1,0){20}}\put(40,0){\makebox(0,0){$K_{3}$}}
\put(50,0){\vector(1,0){20}}\put(80,0){\makebox(0,0){$K_{2}$}}
\put(90,0){\vector(1,0){20}}\put(120,0){\makebox(0,0){$K_{1}$}}
\put(130,0){\vector(1,0){20}}\put(160,0){\makebox(0,0){$K_{0}$}}
\put(170,0){\vector(1,0){20}}\put(200,0){\makebox(0,0){$\mathbb{Z}$}}
\put(210,0){\vector(1,0){20}}\put(240,0){\makebox(0,0){$0$}}
\put(0,-40){\makebox(0,0){$ \cdots$}}\put(10,-40){\vector(1,0){20}}\put(40,-40){\makebox(0,0){$B_{3}$}}
\put(50,-40){\vector(1,0){20}}\put(80,-40){\makebox(0,0){$B_{2}$}}
\put(90,-40){\vector(1,0){20}}\put(120,-40){\makebox(0,0){$B_{1}$}}
\put(130,-40){\vector(1,0){20}}\put(160,-40){\makebox(0,0){$B_{0}$}}
\put(170,-40){\vector(1,0){20}}\put(200,-40){\makebox(0,0){$\mathbb{Z}$}}
\put(210,-40){\vector(1,0){20}}\put(240,-40){\makebox(0,0){$0$}}
\put(40,-10){\vector(0,-1){20}}
\put(80,-10){\vector(0,-1){20}}
\put(120,-10){\vector(0,-1){20}}
\put(158,-10){\line(0,-1){20}}\put(160,-10){\line(0,-1){20}}
\put(198,-10){\line(0,-1){20}}\put(200,-10){\line(0,-1){20}}
\put(60,5){\makebox(0,0){$d$}}
\put(100,5){\makebox(0,0){$d$}}
\put(140,5){\makebox(0,0){$d$}}
\put(60,-35){\makebox(0,0){$\partial_3$}}
\put(100,-35){\makebox(0,0){$\partial_2$}}
\put(140,-35){\makebox(0,0){$\partial_1$}}
\put(50,-20){\makebox(0,0){$G_{3}$}}
\put(90,-20){\makebox(0,0){$G_{2}$}}
\put(130,-20){\makebox(0,0){$G_{1}$}}
\end{picture}
\end{figure}
\end{proposition}
\begin{proof}
By direct verification similarly as the proof of Proposition \ref{chainmap}. The detail is omitted.
\end{proof}
\subsection{A translation to quantum field theory} \label{subspt}
Now we follow the notations in \cite{spt} and translate our result into quantum field theory language. Let $G=\mathbb{Z}_{N_1}\times\cdots\times\mathbb{Z}_{N_n}$ where $N_i|N_{i+1}$ for $1\le i\le n-1$.
Let $k=d+1$ be the spacetime dimension. For $1\le l\le d+1$, $1\le r_1<\cdots<r_l\le n$, $\lambda_i\ge1$ for $1\le i\le l$,
define $$\phi_{{r_i}^{\lambda_i}}=\left\{\begin{array}{ll}A_{r_i}dA_{r_i}\cdots dA_{r_i}, \quad &\text{ if }\lambda_i\text{ odd;}\\
dA_{r_i}\cdots dA_{r_i}, \quad &\text{ if }\lambda_i\text{ even.}\end{array}\right.$$
We generalize the correspondence between the partition functions of fields and cocycles given in \cite{spt}.
The generalized correspondence connects the part $$\zeta_{N_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_{i}\lambda_{j}}\eta_{r_{1},[a_{1},b_{1}]}^{\alpha}\cdots\eta_{r_{l},[a_{l},b_{l}]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}}$$
of $(d+1)$-cocycle $\omega_{d+1}$
and the partition function $$\zeta_{N_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_{i}\lambda_{j}}a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}\frac{N_{r_1}^{\lambda_1-2[\frac{\lambda_1}{2}]}\cdots N_{r_l}^{\lambda_l-2[\frac{\lambda_l}{2}]}}{(2\pi)^{[\frac{\lambda_1+1}{2}]+\cdots+[\frac{\lambda_l+1}{2}]}}\int\phi_{r_l^{\lambda_l}}\cdots\phi_{r_1^{\lambda_1}}}$$
where the corresponding terms of $A_u$ and $dA_u$ are given in \cite{spt} and the order of $A_u$ and $dA_u$ is so arranged that their positions indicate which component of $\alpha$ they correspond to. Note that $\alpha=(\alpha_1,\dots,\alpha_{d+1})$, $\lambda_1+\cdots+\lambda_l=d+1$, and $\lambda_1$ is odd. Our result reveals the fact that we don't need higher form fields $B,C$, etc, to get a complete set of representatives of cocycles.
Now we explain how we get these partition functions.
First, any 1-form field is the linear combination of the wedge products of some $A_u$ and $dA_u$ where each $A_u$ appears at most once, i.e. the linear combination of $\phi_{r_l^{\lambda_l}}\cdots\phi_{r_1^{\lambda_1}}$ for some $1\le l\le d+1$, $1\le r_1<\cdots<r_l\le n$, $\lambda_i\ge1$ for $1\le i\le l$.
After integration by part on $\int\phi_{r_l^{\lambda_l}}\cdots\phi_{r_1^{\lambda_1}}$, we need only consider those terms with $\lambda_1$ odd.
Due to a discrete $\mathbb{Z}_N$ gauge symmetry, and the gauge transformation must be identified by $2\pi$, we have the following general rules:
$$\oint A_u=\frac{2\pi n_u}{N_u}\mod 2\pi, \quad \quad \quad \oint\delta A_u=0\mod 2\pi.$$
We consider a spacetime with a volume size $L^{d+1}$ where $L$ is the length of one dimension, for example $T^{d+1}$ torus.
The allowed large gauge transformation implies that locally $A$ can be:
$$A_{u,\mu}=\frac{2\pi n_udx_{\mu}}{N_uL}, \quad \,\,\,\delta A_u=\frac{2\pi m_udx_{\mu}}{L}.$$
Now we consider the partition function $\exp(ik_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}\int\phi_{r_l^{\lambda_l}}\cdots\phi_{r_1^{\lambda_1}})$ with $\lambda_1$ odd.
Note that $\delta(dA_u)=0.$ Thus for the large gauge transformation, we have $k_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}\int\delta(\phi_{r_l^{\lambda_l}}\cdots\phi_{r_1^{\lambda_1}})=0\mod 2\pi$.
This implies $$k_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=p_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}\frac{N_{r_2}^{\lambda_2-2[\frac{\lambda_2}{2}]}\cdots N_{r_l}^{\lambda_l-2[\frac{\lambda_l}{2}]}}{(2\pi)^{[\frac{\lambda_1+1}{2}]+\cdots+[\frac{\lambda_l+1}{2}]-1}}$$
where $p_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}\in\mathbb{Z}$.
For the flux identification, we have $$k_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}\int\phi_{r_l^{\lambda_l}}\cdots\phi_{r_1^{\lambda_1}}=\frac{(2\pi)^{[\frac{\lambda_1+1}{2}]+\cdots+[\frac{\lambda_l+1}{2}]}n_{r_1}^{[\frac{\lambda_1+1}{2}]}\cdots n_{r_l}^{[\frac{\lambda_l+1}{2}]}}{N_{r_1}^{\lambda_1-2[\frac{\lambda_1}{2}]}\cdots N_{r_l}^{\lambda_l-2[\frac{\lambda_l}{2}]}}.$$
Hence $$(2\pi)^{[\frac{\lambda_1+1}{2}]+\cdots+[\frac{\lambda_l+1}{2}]-1}k_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}\simeq(2\pi)^{[\frac{\lambda_1+1}{2}]+\cdots+[\frac{\lambda_l+1}{2}]-1}k_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}+N_{r_1}^{\lambda_1-2[\frac{\lambda_1}{2}]}\cdots N_{r_l}^{\lambda_l-2[\frac{\lambda_l}{2}]}.$$
Here $\simeq$ means the level identification.
Therefore, the cyclic period of $p_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}$ is $N_{r_1}$.
Finally let $(-1)^{\sum\limits_{1\leq i<j\leq l}\lambda_{i}\lambda_{j}}p_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}=a_{r_1^{\lambda_1}\cdots r_l^{\lambda_l}}$.
Then we get the partition functions in correspondence with cocycles.
\section{On Braided linear Gr-categories}
The monoidal category of finite dimensional vector spaces graded by a group $G,$ with the usual tensor product and associativity constraint given by a 3-cocycle $\omega$ is denoted by $\operatorname{Vec}^\omega_G.$ Such a monoidal category is called a linear Gr-category. The terminology goes back to Ho\`{a}ng Xu\^{a}n S\'{i}nh \cite{grc}, a student of Grothendieck. The aim of this section is give a complete description to all braided linear Gr-categories with a help of the explicit formulas of normalized $3$-cocycles. This extends the related partial results obtained in \cite{bct, hly, js, ms} to full generality.
\subsection{Monoidal structures}
Recall that the category $\operatorname{Vec}_G$ of finite-dimensional $G$-graded vector spaces has simple objects $\{S_g|g \in G\}$ where $(S_g)_h=\delta_{g,h}\mathbbm{k}, \ \forall h \in G.$ The tensor product is given by $S_g \otimes S_h=S_{gh},$ and $S_1$ ($1$ is the identity of $G$) is the unit object. Without loss of generality we may assume that the left and right unit constraints are identities. If $a$ is an associativity constraint on $\operatorname{Vec}_G,$ then it is given by $a_{S_f,S_g,S_h}=\omega(f,g,h)\operatorname{id},$ where $\omega:G \times G \times G \rightarrow \mathbbm{k}^*$ is a function. The pentagon axiom and the triangle axiom give
\begin{gather*}
\omega(ef,g,h)\omega(e,f,gh)=\omega(e,f,g)\omega(e,fg,h)\omega(f,g,h), \\ \omega(f,1,g)=1,
\end{gather*}
which say exactly that $\omega$ is a normalized 3-cocycle on $G.$ Note that cohomologous cocycles define equivalent monoidal structures, therefore the classification of monoidal structures on $\operatorname{Vec}_G$ is equivalent to determining a complete set of representatives of normalized 3-cocycles on $G.$
\subsection{Normalized 3-cocycles}
In the special case $k=3$, if we abbreviate $a_{r^3}$ by $a_r$, $a_{rs^2}$ by $a_{rs}$, then \eqref{barcocycleabelian} becomes
\begin{eqnarray}\label{3cocycle}
&& \omega:\;B_{3}\longrightarrow \mathbbm{k}^{\ast} \\ \notag
&&[g_{1}^{i_{1}}\cdots g_{n}^{i_{n}},g_{1}^{j_{1}}\cdots g_{n}^{j_{n}},g_{1}^{k_{1}}\cdots g_{n}^{k_{n}}] \mapsto \prod_{r=1}^{n}\zeta_{m_r}^{a_{r}i_{r}[\frac{j_{r}+k_{r}}{m_{r}}]}
\prod_{1\leq r<s\leq n}\zeta_{m_r}^{a_{rs}k_{r}[\frac{i_{s}+j_{s}}{m_{s}}]}
\prod_{1\leq r<s<t\leq n}\zeta_{m_r}^{-a_{rst}k_{r}j_{s}i_{t}}
\end{eqnarray}
where $0\le a_r,a_{rs},a_{rst}<m_r$.
\begin{remark}
The present complete set of representatives of normalized $3$-cocycles is slightly different from that in \cite{bgrc1, hly}. Of course they are equivalent up to cohomology.
\end{remark}
\subsection{Braided structures} Now we consider the braided structures on a linear Gr-category $\operatorname{Vec}_G^\omega.$ Recall that a braiding in $\operatorname{Vec}_G^\omega$ is a commutativity constraint $c: \otimes \rightarrow \otimes^{\operatorname{op}}$ satisfying the hexagon axiom. Note that $c$ is given by $c_{S_x,S_y}=\mathcal{R}(x,y)\operatorname{id},$ where $\mathcal{R}: G \times G \rightarrow \mathbbm{k}^*$ is a function, and the hexagon axiom of $c$ says that
\begin{equation}\label{EM3cocycle}
\frac{\mathcal{R}(xy,z)}{\mathcal{R}(x,z)\mathcal{R}(y,z)}\frac{\omega(x,z,y)}{\omega(x,y,z)\omega(z,x,y)}=1=\frac{\mathcal{R}(x,yz)}{\mathcal{R}(x,y)\mathcal{R}(x,z)}\frac{\omega(x,y,z)\omega(y,z,x)}{\omega(y,x,z)}
\end{equation}
for all $x,y,z\in G$.
In other words, $\mathcal{R}$ is a quasi-bicharacter of $G$ with respect to $\omega.$ Therefore, the classification of braidings in $\operatorname{Vec}_G^\omega$ is equivalent to determining all the quasi-bicharacters of $G$ with respect to $\omega.$ It is interesting to remark that the braided monoidal structures $(\omega,\mathcal{R})$ on $\operatorname{Vec}_G$ appeared already in the 1950s in terms of the so-called abelian cohomology of Eilenberg and Mac Lane \cite{EMa, EMb}.
\subsection{Quasi-bicharacters} Clearly, any quasi-bicharacter $\mathcal{R}$ is uniquely determined by the following values:
$$r_{ij}:=\mathcal{R}(g_{i},g_{j}),\;\;\;\;\;\;\;\;\textrm{for all}\;\;1\leq i, \ j\leq n.$$
\begin{proposition} \label{condition}
Let $r_{ij}\in \mathbbm{k}^{\ast}$ for
$1\leq i, \ j\leq n$. Then there is a quasi-bicharacter $\mathcal{R}$ with respect to $\omega$ satisfying $\mathcal{R}(g_{i},g_{j})=r_{ij}$
if and only if the following equations are satisfied:
\begin{eqnarray*}
r_{ii}^{m_{i}}=\zeta_{m_{i}}^{a_{i}}=\zeta_{m_{i}}^{-a_{i}},&&\emph{for}\;\;1\leq i\leq n,\\
r_{ij}^{m_{i}}=r_{ji}^{m_i}=1,\ \ a_{ij}=0,&&\emph{for}\;\;1\le i<j\le n,\\
a_{rst}=0,&&\emph{for}\;\;1\leq r<s<t\leq n.
\end{eqnarray*}
\end{proposition}
\begin{proof} ``$\Rightarrow$".
For the case $r<s<t,$ consider $\mathcal{R}(g_{t}g_{s},g_{r})$ and $\mathcal{R}(g_{s}g_{t},g_{r})$ which obviously are equal. By \eqref{EM3cocycle}, we have
\begin{eqnarray*}\mathcal{R}(g_{t}g_{s}, g_{r})&=&\mathcal{R}(g_{t} , g_{r})\mathcal{R}(g_{s}, g_{r})\frac{\omega(g_{r},g_{t},g_{s})\omega(g_{t},g_{s},g_{r})}{\omega(g_{t},g_{r},g_{s})}\\
&=&\mathcal{R}(g_{t} , g_{r})\mathcal{R}(g_{s}, g_{r})\zeta_{m_r}^{-a_{rst}},\end{eqnarray*}
\begin{eqnarray*}\mathcal{R}(g_{s}g_{t}, g_{r})&=&\mathcal{R}(g_{s} , g_{r})\mathcal{R}(g_{t}, g_{r})\frac{\omega(g_{r},g_{s},g_{t})\omega(g_{s},g_{t},g_{r})}{\omega(g_{s},g_{r},g_{t})}\\
&=&\mathcal{R}(g_{s} , g_{r})\mathcal{R}(g_{t}, g_{r}).\end{eqnarray*}
Therefore, $\zeta_{m_r}^{-a_{rst}}=1$. Since $0\le a_{rst}<m_r$, we arrive at $a_{rst}=0.$
For any $1\leq i\leq n$, applying \eqref{EM3cocycle} iteratively, we have $\mathcal{R}(g_{i},g_{i}^{s})=\mathcal{R}(g_{i},g_{i})^{s}$ and $\mathcal{R}(g_{i}^{s},g_{i})=\mathcal{R}(g_{i},g_{i})^{s}$
for $1\leq s\leq m_{i}-1$. Then
$$1=\mathcal{R}(g_{i},g_{i}^{m_{i}})=\mathcal{R}(g_{i},g_{i})\mathcal{R}(g_{i},g_{i}^{m_{i}-1})
\frac{\omega(g_{i},g_{i},g_{i}^{m_{i}-1})}{\omega(g_i,g_i,g_i^{m_i-1})\omega(g_i,g_i^{m_i-1},g_i)}=\mathcal{R}(g_{i},g_{i})^{m_{i}}\zeta_{m_{i}}^{-a_{i}},$$
$$1=\mathcal{R}(g_{i}^{m_{i}},g_{i})=\mathcal{R}(g_{i}^{m_{i}-1},g_{i})\mathcal{R}(g_{i},g_{i})\frac{\omega(g_i^{m_i-1},g_i,g_i)\omega(g_{i},g_{i}^{m_{i}-1},g_{i})}{\omega(g_i^{m_i-1},g_i,g_i)}=\mathcal{R}(g_{i},g_{i})^{m_{i}}{\zeta_{m_{i}}^{a_{i}}}.$$
Thus $r_{ii}^{m_{i}}=\zeta_{m_{i}}^{a_{i}}=\zeta_{m_{i}}^{-a_{i}}$.
Assume $i<j.$ Applying \eqref{EM3cocycle} iteratively, one has $\mathcal{R}(g_{i}^{k},g_{j})=\mathcal{R}(g_{i},g_{j})^{k}$ for $1\leq k\leq m_{i}-1.$ Therefore,
$$1=\mathcal{R}(g_{i}^{m_{i}},g_{j})=\mathcal{R}(g_{i}^{m_{i}-1},g_{j})\mathcal{R}(g_{i},g_{j})\frac{\omega(g_i^{m_i-1},g_i,g_j)\omega(g_j,g_i^{m_i-1},g_i)}{\omega(g_i^{m_i-1},g_j,g_i)}
=\mathcal{R}(g_{i},g_{j})^{m_{i}}.$$
This implies that $r_{ij}^{m_{i}}=1$. Applying \eqref{EM3cocycle} iteratively, one has $\mathcal{R}(g_{i},g_{j}^{k})=\mathcal{R}(g_{i},g_{j})^{k}$ for $1\leq k\leq m_{j}-1$. Therefore,
$$1=\mathcal{R}(g_{i},g_{j}^{m_{j}})=\mathcal{R}(g_{i},g_{j})\mathcal{R}(g_i,g_j^{m_j-1})\frac{\omega(g_j,g_i,g_j^{m_j-1})}{\omega(g_i,g_j,g_j^{m_j-1})\omega(g_j,g_j^{m_j-1},g_i)}=r_{ij}^{m_{j}}\zeta_{m_i}^{-a_{ij}}.$$ This implies that $r_{ij}^{m_{j}}=\zeta_{m_{i}}^{a_{ij}}$.
Since $m_i|m_j$, we have $\zeta_{m_i}^{a_{ij}}=1$. Since $0\le a_{ij}<m_i$, we arrive at $a_{ij}=0$.
Assume $i>j.$ Applying \eqref{EM3cocycle} iteratively, one has $\mathcal{R}(g_{i}^{k},g_{j})=\mathcal{R}(g_{i},g_{j})^{k}$ for $1\leq k\leq m_{i}-1.$ Therefore,
$$1=\mathcal{R}(g_{i}^{m_{i}},g_{j})=\mathcal{R}(g_{i}^{m_{i}-1},g_{j})\mathcal{R}(g_{i},g_{j})\frac{\omega(g_i^{m_i-1},g_i,g_j)\omega(g_j,g_i^{m_i-1},g_i)}{\omega(g_i^{m_i-1},g_j,g_i)}
=\mathcal{R}(g_{i},g_{j})^{m_{i}}\zeta_{m_j}^{a_{ij}}.$$
This implies that $r_{ij}^{m_{i}}=\zeta_{m_j}^{-a_{ij}}=1$. Applying \eqref{EM3cocycle} iteratively, one has $\mathcal{R}(g_{i},g_{j}^{k})=\mathcal{R}(g_{i},g_{j})^{k}$ for $1\leq k\leq m_{j}-1$. Therefore,
$$1=\mathcal{R}(g_{i},g_{j}^{m_{j}})=\mathcal{R}(g_{i},g_{j})\mathcal{R}(g_i,g_j^{m_j-1})\frac{\omega(g_j,g_i,g_j^{m_j-1})}{\omega(g_i,g_j,g_j^{m_j-1})\omega(g_j,g_j^{m_j-1},g_i)}=r_{ij}^{m_{j}}.$$ This implies that $r_{ij}^{m_{j}}=1$.
The necessity is proved.
``$\Leftarrow$". Conversely, define a map $\mathcal{R}: G\times G \longrightarrow k^{\ast}$ by setting $$\mathcal{R}(g_{1}^{i_{1}}\cdots g_{n}^{i_{n}},g_{1}^{j_{1}}\cdots g_{n}^{j_{n}}):=\prod_{s=1}^n r_{ss}^{i_{s}j_{s}}.$$ We verify that $\mathcal{R}$ is a quasi-bicharacter of $G$ with respect to $\omega.$
Let $x=g_{1}^{i_{1}}\cdots g_{n}^{i_{n}}, \ y=g_{1}^{j_{1}}\cdots g_{n}^{j_{n}}, \ z=g_{1}^{k_{1}}\cdots g_{n}^{k_{n}},$ then
$$\mathcal{R}(g_{1}^{i_{1}}\cdots g_{n}^{i_{n}}\cdot g_{1}^{j_{1}}\cdots g_{n}^{j_{n}},g_{1}^{k_{1}}\cdots g_{n}^{k_{n}})=\prod_{s=1}^n r_{ss}^{(i_{s}+j_{s})'k_{s}},$$ where $(i_{s}+j_{s})'$ denotes the remainder of division of $i_{s}+j_{s}$ by $m_{s}.$
Consider $\mathcal{R}(x , z)\mathcal{R}(y ,z)\frac{\omega(z,x,y)\omega(x,y,z)}{\omega(x,z,y)}.$ By direct calculation, one has
$$\frac{\omega(z,x,y)\omega(x,y,z)}{\omega(x,z,y)}=\prod_{l=1}^{n}\zeta_{m_{l}}^{a_{l}k_{l}
[\frac{i_{l}+j_{l}}{m_l}]}.$$
Therefore,
\begin{eqnarray*}
&&\mathcal{R}(x,z)\mathcal{R}(y,z)\frac{\omega(z,x,y)\omega(x,y,z)}{\omega(x,z,y)}\\
&=&\prod_{s=1}^n r_{ss}^{(i_{s}+j_{s})k_{s}}\prod_{l=1}^{n}\zeta_{m_{l}}^{a_{l}k_{l}
[\frac{i_{l}+j_{l}}{m_{l}}]}\\
&=&\prod_{s=1}^n r_{ss}^{((i_{s}+j_{s})'+[\frac{i_{s}+j_{s}}{m_{s}}]m_{s})k_{s}}\prod_{l=1}^{n}\zeta_{m_{l}}^{a_{l}k_{l}
[\frac{i_{l}+j_{l}}{m_{l}}]}\\
&=&\prod_{s=1}^n r_{ss}^{(i_{s}+j_{s})'k_{s}}\prod_{l=1}^{n} c_{ll}^{[\frac{i_{l}+j_{l}}{m_{l}}]m_{l}k_{l}}\prod_{l=1}^{n}\zeta_{m_{l}}^{a_{l}k_{l}
[\frac{i_{l}+j_{l}}{m_{l}}]}\\
&=&\prod_{s=1}^n r_{ss}^{(i_{s}+j_{s})'k_{s}}\prod_{l=1}^{n}\zeta_{m_{l}}^{-a_{l}k_{l}
[\frac{i_{l}+j_{l}}{m_{l}}]}\prod_{l=1}^{n}\zeta_{m_{l}}^{a_{l}k_{l}
[\frac{i_{l}+j_{l}}{m_{l}}]}.
\end{eqnarray*}
This implies that $$\mathcal{R}(x,z)\mathcal{R}(y,z)\frac{\omega(z,x,y)\omega(x,y,z)}{\omega(x,z,y)}=\prod_{s=1}^n r_{ss}^{(i_{s}+j_{s})'k_{s}}=\mathcal{R}(xy,z).$$
The sufficiency is proved.
\end{proof}
\section{The Dijkgraaf-Witten invariant of the $n$-torus}
In this section, we give a formula of the Dijkgraaf-Witten invariant for an arbitrary $n$-torus $T^n$ associated to finite abelian groups. In the special case of $n=2,$ we recover and improve some results obtained in \cite{turaev}. This is due to the fact that as we have an explicit formula of $2$-cocycles, we are able to derive dimension formulas for irreducible projective representations of finite abelian groups. This is of independent interest.
\subsection{Definition of DW invariants}
Just as its name implies, such an invariant of 3-manifolds was introduced by Dijkgraaf and Witten in \cite{dw}. Then it was generalized to arbitrary dimension in \cite{freed} by Freed.
Now we recall briefly the definition of DW invariants. The reader is referred to \cite{dw, freed, turaev} for more details.
Let $G$ be a finite group and let $[\omega]\in \operatorname{H}^n(BG;\mathbbm{k}^*)$. For a closed oriented $n$-manifold $M$, the Dijkgraaf-Witten invariant of $M$ is defined as
$$Z^{[\omega]}(M)=\frac{1}{|G|}\sum_{\phi:\pi_1(M)\to G}\langle f_{\phi}^*[\omega],[M]\rangle,$$
where $f_{\phi}:M\to BG$ is a map inducing $\phi$ on the fundamental group which is determined by $\phi$ up to homotopy, $[M]$ is the fundamental class of $M$, and $\langle,\rangle$ is the pairing $\operatorname{H}^n(M;\mathbbm{k}^*) \times \operatorname{H}_n(M;\mathbb{Z})\to \mathbbm{k}^*$.
\subsection{The DW invariant of the $n$-torus}
The DW invariant of the $n$-torus is known to be the ground state degeneracy, which is the dimension of a Hilbert space, hence an integer. Some special cases were computed in \cite{ww, wen}.
Let $\mathbb{Z}_d$ denote the quotient ring $\mathbb{Z}/d\mathbb{Z}$ and $M_n(\mathbb{Z}_d)$ the ring of $d \times d$ matrices with entries in $\mathbb{Z}_d.$ Fix a $d$-th primitive root $\xi$ of 1 and define $$N_n(d):=\frac{\sum_{A\in M_n(\mathbb{Z}_d)}\xi^{\det A}}{d^n}.$$
\begin{lemma}\label{lemma}
The function $N_n(d)$ is integer-valued and is multiplicative on $d$, that is, if $d=d_1d_2$ with $(d_1,d_2)=1$, then
$N_n(d)=N_n(d_1)N_n(d_2)$. Moreover, for a prime $p,$
$$N_n(p^m)=\sum_{i=1}^mp^{m(n-2)}p^{(m-i)(n-2)(n-1)}N_{n-1}(p^i)(p^{ni}-p^{n(i-1)})+p^{m(n-2)n}.$$
\end{lemma}
\begin{proof}
Take $A=(a_{ij})\in M_n(\mathbb{Z}_d).$ Then
$\det A=a_{11}A_{11}+\cdots+a_{1n}A_{1n}$ where $A_{ij}$ is the algebraic cofactor of $a_{ij}$ and thus
\begin{eqnarray*}
\sum_{A\in M_n(\mathbb{Z}_d)}\xi^{\det A}&=&\sum_{a_{11}=0}^{d-1}\xi^{a_{11}A_{11}}\cdots\sum_{a_{1n}=0}^{d-1}\xi^{a_{1n}A_{1n}}\\
&=&d^n\#\{B\in M_{(n-1)\times n}(\mathbb{Z}_d)|\text{all }(n-1)\text{-minors of }B\text{ are }0\}.
\end{eqnarray*}
Hence $N_n(d)=\#\{B\in M_{(n-1)\times n}(\mathbb{Z}_d)|\text{all }(n-1)\text{-minors of }B\text{ are }0\}.$
Assume $B=(b_{ij})\in M_{(n-1)\times n}(\mathbb{Z}_d)$ is such a matrix all of whose $(n-1)$-minors are $0.$ Define $\text{ord}(b_{11},\dots,b_{1n})$ to be the smallest integer $r$ such that $d|rb_{11}, \ \dots, \ d|rb_{1n}.$ Clearly, $\text{ord}(b_{11},\dots,b_{1n})|d$. Now suppose $d=p^m$ where $p$ is prime.
If $\text{ord}(b_{11},\dots,b_{1n})=p^i$, then $p^ib_{11}=p^m\widetilde{b_{11}}, \ \dots, \ p^ib_{1n}=p^m\widetilde{b_{1n}}$. For $i\geq1$, if $p|\widetilde{b_{11}}, \ \dots, \ p|\widetilde{b_{1n}}$, then $\text{ord}(b_{11},\dots,b_{1n})\leq p^{i-1}$, contradiction.
So we may assume, without loss of generality, that $p \nmid \widetilde{b_{11}}.$ In this case, the matrix $P:=\left(\begin{array}{cccc}\overline{\widetilde{b_{11}}}&\overline{\widetilde{b_{12}}}&\ldots&\overline{\widetilde{b_{1n}}}\\
\overline{0}&\overline{1}&\ldots&\overline{0}\\
\vdots&\vdots&\ddots&\vdots\\
\overline{0}&\overline{0}&\ldots&\overline{1}\end{array}\right)$ is invertible in $M_n(\mathbb{Z}_{p^m}).$ Thus obviously,
$(\overline{b_{11}},\dots,\overline{b_{1n}})P^{-1}=(\overline{p^{m-i}},\overline{0},\dots,\overline{0})$. Assume $(\overline{b_{i1}},\dots,\overline{b_{in}})P^{-1}=(\overline{b_{i1}'},\dots,\overline{b_{in}'})$ for $i=2,\dots,n-1$.
Since all $(n-1)$-minors of $B$ are 0, all $(n-1)$-minors of $\left(\begin{array}{cccc}\overline{p^{m-i}}&\overline{0}&\ldots&\overline{0}\\
\overline{b_{21}'}&\overline{b_{22}'}&\ldots&\overline{b_{2n}'}\\
\vdots&\vdots&\ddots&\vdots\\
\overline{b_{n-1,1}'}&\overline{b_{n-1,2}'}&\ldots&\overline{b_{n-1,n}'}\end{array}\right)$ are 0. Hence all $(n-2)$-minors of $\left(\begin{array}{ccc}\overline{b_{22}'}&\ldots&\overline{b_{2n}'}\\
\vdots&\ddots&\vdots\\
\overline{b_{n-1,2}'}&\ldots&\overline{b_{n-1,n}'}\end{array}\right)$ are 0 modulo $p^i$.
So we have
$$N_n(p^m)=\sum_{i=1}^mp^{m(n-2)}p^{(m-i)(n-2)(n-1)}N_{n-1}(p^i)(p^{ni}-p^{n(i-1)})+p^{m(n-2)n}.$$
Denote $S_n(d)=\{B\in M_{(n-1)\times n}(\mathbb{Z}_d)|\text{all }(n-1)\text{-minors of }B\text{ are }0\}$.
Then $B\mapsto(B\mod d_1,B\mod d_2)$ defines a map from $S_n(d_1d_2)$ to $S_n(d_1)\times S_n(d_2)$.
If $(d_1,d_2)=1$, then this map is clearly injective and surjective by the Chinese Remainder Theorem.
Hence $N_n(d_1d_2)=N_n(d_1)N_n(d_2)$.
So if $d=p_1^{m_1}\cdots p_r^{m_r}$ where $p_1,\dots,p_r$ are distinct primes, then $N_n(d)=N_n(p_1^{m_1})\cdots N_n(p_r^{m_r})$.
\end{proof}
\begin{theorem}\label{mt}
The Dijkgraaf-Witten invariant of the $n$-torus $T^n$ for a finite abelian group $G$ is
\begin{equation}\label{dwtorus}
Z^{[\omega]}(T^n)=\frac{1}{|G|}\sum_{f_1,\dots,f_n\in G}\frac{\prod_{\sigma\in A_n}\omega(f_{\sigma(1)},\dots,f_{\sigma(n)})}{\prod_{\sigma\in S_n\setminus A_n}\omega(f_{\sigma(1)},\dots,f_{\sigma(n)})}.
\end{equation}
Let $G=\mathbb{Z}_{m_1}\times\cdots\times\mathbb{Z}_{m_l}$ where $m_i|m_{i+1}$ for $1\le i\le l-1$.
If $l<n$, then $Z^{[\omega]}(T^n)=|G|^{n-1}$.
If $l=n$, then $Z^{[\omega]}(T^n)=\frac{|G|^{n-1}}{d^{n(n-1)}}N_n(d)$ where $d=\frac{m_1}{(m_1,a_{1\cdots n})}$.
If $l>n$, then
$$Z^{[\omega]}(T^n)=\frac{1}{|G|}\sum_{A}\prod_{1\leq r_1<\cdots<r_n\leq l}\zeta_{m_{r_1}}^{a_{r_1\cdots r_n}\det A\left(\begin{array}{ccc}1&\ldots&n\\r_1&\ldots&r_n\end{array}\right)}$$ where
$A=(\alpha_{ij})_{n \times l}$ and $0\leq \alpha_{ij}<m_j$ for $1\leq i\leq n$.
\end{theorem}
\begin{proof}
The $n$-torus $T^n$ is obtained by gluing parallel edges of an $n$-dimensional cube. The cube can be subdivided into $n!$ $n$-simplexes such that each $n$-simplex has $n$ successive edges in common with the cube.
If we label the remaining $n$ edges of the cube after gluing by $f_1,\dots,f_n\in G$, then each $n$-simplex is uniquely determined by an permutation $(f_1',\dots,f_n')$ of $(f_1,\dots,f_n)$.
The fundamental class $[T^n]$ is represented by an $n$-chain $\sigma:\Delta^n\to T^n$ where $\sigma$ is the sum of those $n$-simplexes with the sign of which is positive if the permutation is even and negative otherwise.
By \cite[p89]{hatcher} we may identify $\operatorname{H}^n(BG;\mathbbm{k}^*)$ and $\operatorname{H}^n(G;\mathbbm{k}^*).$ Then when $\phi$ runs over all group homomorphisms from $\pi_1(T^n)$ to $G$, we have
$$\omega((f_{\phi})_*([T^n]))=\frac{\prod_{\sigma\in A_n}\omega(f_{\sigma(1)},\dots,f_{\sigma(n)})}{\prod_{\sigma\in S_n\setminus A_n}\omega(f_{\sigma(1)},\dots,f_{\sigma(n)})}$$
where $f_1,\dots,f_n$ run over $G$.
Hence \eqref{dwtorus} holds.
Now let $f_i=(\alpha_{i1},\dots,\alpha_{il})$ where $0\le \alpha_{ij}<m_j$ for $1\le i\le n$.
Recall that
$$\omega(f_1,\dots,f_n)=\prod_{k=1}^{n}\prod_{\begin{array}{ccc}1\leq r_{1}<\cdots<r_{k}\leq l\\\lambda_1+\cdots+\lambda_k=n,\lambda_1\text{ odd}\\\lambda_i\ge1\text{ for }1\le i\le k\end{array}}\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq k}\lambda_{i}\lambda_{j}}\eta_{r_{1},[a_{1},b_{1}]}^{\alpha}\cdots\eta_{r_{k},[a_{k},b_{k}]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_k^{\lambda_k}}}$$
where $a_{k}=1, \ b_{k}=\lambda_{k}, \ \dots, \ a_{1}=\lambda_{2}+\cdots+\lambda_{k}+1, \ b_{1}=\lambda_{1}+\cdots+\lambda_{k}=n$.
\begin{eqnarray*}
Z^{[\omega]}(T^n)&=&\frac{1}{|G|}\sum_{f_1,\dots,f_n\in G}\frac{\prod_{\sigma\in A_n}\omega(f_{\sigma(1)},\dots,f_{\sigma(n)})}{\prod_{\sigma\in S_n\setminus A_n}\omega(f_{\sigma(1)},\dots,f_{\sigma(n)})}\\
&=&\frac{1}{|G|}\sum_{f_1,\dots,f_n\in G}\prod_{k=1}^{n}\prod_{\begin{array}{ccc}1\leq r_{1}<\cdots<r_{k}\leq l\\\lambda_1+\cdots+\lambda_k=n,\lambda_1\text{ odd}\\\lambda_i\ge1\text{ for }1\le i\le k\end{array}}\\
&&\quad \frac{\prod_{\sigma\in A_n}\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq k}\lambda_{i}\lambda_{j}}\eta_{r_{1},[\sigma(a_{1}),\sigma(b_{1})]}^{\alpha}\cdots\eta_{r_{k},[\sigma(a_{k}),\sigma(b_{k})]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_k^{\lambda_k}}}}{\prod_{\sigma\in S_n\setminus A_n}\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq k}\lambda_{i}\lambda_{j}}\eta_{r_{1},[\sigma(a_{1}),\sigma(b_{1})]}^{\alpha}\cdots\eta_{r_{k},[\sigma(a_{k}),\sigma(b_{k})]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_k^{\lambda_k}}}}
\end{eqnarray*}
If $(\lambda_1,\dots,\lambda_k)$ is a partition of $n$ and $\lambda_i>1$ for some $i\in\{1,\dots,k\}$, then $b_i>a_i$ and $\eta_{r_i,[a_i,b_i]}^{\alpha}$ contains the term $[\frac{\alpha_{b_i,r_i}+\alpha_{b_i-1,r_i}}{m_{r_i}}]$.
Let $\tau$ be the transposition $(b_i,b_i-1)$, then
$$\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq k}\lambda_{i}\lambda_{j}}\eta_{r_{1},[a_{1},b_{1}]}^{\alpha}\cdots\eta_{r_{k},[a_{k},b_{k}]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_k^{\lambda_k}}}
=\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq k}\lambda_{i}\lambda_{j}}\eta_{r_{1},[\tau(a_{1}),\tau(b_{1})]}^{\alpha}\cdots\eta_{r_{k},[\tau(a_{k}),\tau(b_{k})]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_k^{\lambda_k}}}.$$
Hence
$$\frac{\prod_{\sigma\in A_n}\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq k}\lambda_{i}\lambda_{j}}\eta_{r_{1},[\sigma(a_{1}),\sigma(b_{1})]}^{\alpha}\cdots\eta_{r_{k},[\sigma(a_{k}),\sigma(b_{k})]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_k^{\lambda_k}}}}{\prod_{\sigma\in S_n\setminus A_n}\zeta_{m_{r_1}}^{(-1)^{\sum\limits_{1\leq i<j\leq k}\lambda_{i}\lambda_{j}}\eta_{r_{1},[\sigma(a_{1}),\sigma(b_{1})]}^{\alpha}\cdots\eta_{r_{k},[\sigma(a_{k}),\sigma(b_{k})]}^{\alpha}a_{r_1^{\lambda_1}\cdots r_k^{\lambda_k}}}}=1.$$
Therefore,
$$Z^{[\omega]}(T^n)=\frac{1}{|G|}\sum_{f_1,\dots,f_n\in G}\prod_{1\leq r_{1}<\cdots<r_{n}\leq l}\frac{\prod_{\sigma\in A_n}\zeta_{m_{r_1}}^{(-1)^{\frac{n(n-1)}{2}}\alpha_{\sigma(n),r_1}\cdots\alpha_{\sigma(1),r_n}a_{r_1\cdots r_n}}}{\prod_{\sigma\in S_n\setminus A_n}\zeta_{m_{r_1}}^{(-1)^{\frac{n(n-1)}{2}}\alpha_{\sigma(n),r_1}\cdots\alpha_{\sigma(1),r_n}a_{r_1\cdots r_n}}}.
$$
Hence if $l<n$, then each summand of \eqref{dwtorus} is 1 and $Z^{[\omega]}(T^n)=|G|^{n-1}$.
If $l=n$, then equation \eqref{dwtorus} becomes
\begin{eqnarray*}
Z^{[\omega]}(T^n)&=&\frac{1}{|G|}\sum_{A}(\zeta_{m_1}^{a_{1\cdots n}})^{\sum_{\sigma\in S_n}(-1)^{\text{sign of }\sigma}(-1)^{\frac{n(n-1)}{2}}\alpha_{\sigma(n)1}\cdots\alpha_{\sigma(1)n}}\\
&=&\frac{1}{|G|}\sum_{A}(\zeta_{m_1}^{a_{1\cdots n}})^{\sum_{\sigma\in S_n}(-1)^{\text{sign of }\sigma}\alpha_{\sigma(1)1}\cdots\alpha_{\sigma(n)n}}\\
&=&\frac{1}{|G|}\sum_{A}(\zeta_{m_1}^{a_{1\cdots n}})^{\det A}
\end{eqnarray*}
where $A=(\alpha_{ij})_{n \times l}$ and $0\leq \alpha_{ij}<m_j$ for $1\leq i\leq n$.
Denote $\xi=\zeta_{m_1}^{a_{1\cdots n}}$, then $\xi$ is a $d$-th primitive root of 1 where $d=\frac{m_1}{(m_1,a_{1\cdots n})}$. In this situation,
\begin{eqnarray*}
Z^{[\omega]}(T^n)&=&\frac{1}{|G|}\sum_{A}\xi^{\det A}\\
&=&\frac{1}{|G|}(\frac{m_1}{d}\cdots\frac{m_n}{d})^n\sum_{A\in M_n(\mathbb{Z}_d)}\xi^{\det A}\\
&=&\frac{|G|^{n-1}}{d^{n^2}}\sum_{A\in M_n(\mathbb{Z}_d)}\xi^{\det A}.
\end{eqnarray*}
If $l>n$, the formula is similarly derived as the case $l=n$.
Hence Lemma \ref{lemma} completes the proof.
\end{proof}
\subsection{The DW invariant of $T^2$ and projective representations}
In \cite{turaev} Turaev observed the connection between DW invariants of surfaces and projective representations of finite groups. In case of $T^2,$ our Theorem \ref{mt}
recovers some partial results of Turaev. Moreover, with a help of our explicit formula of $2$-cocycles, we are able to derive a formula for the dimension of an arbitrary projective representation of finite abelian groups. This is of independent interest on the one hand, and helps to improve some formulas in \cite{turaev} on the other hand.
Now let $G=\mathbb{Z}_{m_1}\times\cdots\times\mathbb{Z}_{m_l}$ with $m_1|m_2|\cdots|m_l$ and let $\omega$ be a $2$-cocycle on $G$.
In case $k=2$, \eqref{barcocycleabelian} becomes
$$\omega(g_1^{i_1}\cdots g_l^{i_l},g_1^{j_1}\cdots g_l^{j_l})=\prod_{1\leq r<s\leq l}\zeta_{m_r}^{-a_{rs}i_sj_r}.$$
Let $G_0$ be the set of all $\omega$-regular elements in $G$, i.e.,
$$G_0=\{x\in G|\omega(x,y)=\omega(y,x)\text{ for all }y\in G\}.$$ It is well known that the number of irreducible $\omega$-representations of $G$ is $|G_0|$ and all irreducible $\omega$-representations of $G$ share a common dimension $d=\sqrt{\frac{|G|}{|G_0|}},$ see \cite{frucht, kar}.
In the following we derive the formula of $|G_0|,$ hence of $d$ as well, in terms of the data $(a_{rs})_{1\leq r<s\leq l}$ of the given $2$-cocycle $\omega.$ Consider the following antisymmetric $l \times l$-matrix
$$B=\left(\begin{array}{cccc}0&b_{12}&\ldots&b_{1l}\\-b_{12}&0&\ldots&b_{2l}\\\vdots&\vdots&\ddots&\vdots\\-b_{1l}&-b_{2l}&\ldots&0\end{array}\right)$$
where $b_{ij}=\frac{m_l}{m_i}a_{ij}$. Assume that the invariant factors of $B$ are $\lambda_1, \ \dots, \ \lambda_k$ with $\lambda_1|\lambda_2|\cdots|\lambda_k.$
\begin{proposition}
Keep the above notations. Then we have \[ |G_0|=\frac{|G|}{\frac{m_l}{(m_l,\lambda_1)}\cdots\frac{m_l}{(m_l,\lambda_k)}}, \quad \quad d=\sqrt{\frac{m_l}{(m_l,\lambda_1)}\cdots\frac{m_l}{(m_l,\lambda_k)}}. \]
\end{proposition}
\begin{proof}
By direct computations, we have
\begin{eqnarray*}
|G_0|&=&\#\{(i_1,\dots,i_l)|0\le i_r<m_r\text{ for }1\le r\le l,\prod_{1\leq r<s\leq l}\zeta_{m_r}^{-a_{rs}i_sj_r}=\prod_{1\leq r<s\leq l}\zeta_{m_r}^{-a_{rs}i_rj_s}\\
&&\text{ for any }(j_1,\dots,j_l)\text{ where }0\le j_r<m_r\text{ for }1\le r\le l\}\\
&=&\frac{1}{\frac{m_l}{m_1}\cdots\frac{m_l}{m_{l-1}}}\#\{(i_1,\dots,i_l)|0\le i_r<m_l\text{ for }1\le r\le l,\left(\begin{array}{ccc}i_1&\ldots&i_l\end{array}\right)B\left(\begin{array}{ccc}j_1\\\vdots\\j_l\end{array}\right)\\
&&\equiv0\mod m_l\text{ for any }(j_1,\dots,j_l)\text{ where }0\le j_r<m_l\text{ for }1\le r\le l\}\\
&=&\frac{1}{\frac{m_l}{m_1}\cdots\frac{m_l}{m_{l-1}}}\#\{(i_1,\dots,i_l)|0\le i_r<m_l\text{ for }1\le r\le l,\lambda_ri_r\equiv0\mod m_l\text{ for }1\le r\le k\}\\
&=&\frac{1}{\frac{m_l}{m_1}\cdots\frac{m_l}{m_{l-1}}}(m_l,\lambda_1)\cdots(m_l,\lambda_k)m_l^{l-k}\\
&=&\frac{|G|}{\frac{m_l}{(m_l,\lambda_1)}\cdots\frac{m_l}{(m_l,\lambda_k)}}
\end{eqnarray*} and \[ d=\sqrt{\frac{|G|}{|G_0|}}=\sqrt{\frac{m_l}{(m_l,\lambda_1)}\cdots\frac{m_l}{(m_l,\lambda_k)}}. \]
\end{proof}
We recover a result of Turaev \cite{turaev} in the following
\begin{corollary}
Keep the previous assumptions and notations. We have $$Z^{[\omega]}(T^2)=|G_0|=\frac{|G|}{\frac{m_l}{(m_l,\lambda_1)}\cdots\frac{m_l}{(m_l,\lambda_k)}}.$$
\end{corollary}
\begin{proof}
If $l\ge2$, then
\begin{eqnarray*}
Z^{[\omega]}(T^2)&=&\sum_A\prod_{1\leq r_1<r_2\leq l}\zeta_{m_{r_1}}^{a_{r_1r_2}\det A\left(\begin{array}{cc}1&2\\r_1&r_2\end{array}\right)}\\
&=&\frac{1}{(\frac{m_l}{m_1})^2\cdots(\frac{m_l}{m_{l-1}})^2}\sum_{0\leq\alpha_{ij}<m_l}\zeta_{m_l}^{\left(\begin{array}{ccc}\alpha_{11}&\ldots&\alpha_{1l}\end{array}\right)B\left(\begin{array}{ccc}\alpha_{21}\\\vdots\\\alpha_{2l}\end{array}\right)}
\end{eqnarray*}
where $A=\left(\begin{array}{ccc}\alpha_{11}&\ldots&\alpha_{1l}\\\alpha_{21}&\ldots&\alpha_{2l}\end{array}\right)$ and $0\leq\alpha_{ij}<m_j$.
By assumption, there exist two invertible integral matrices $P, \ Q \in GL_l(\mathbb{Z})$ such that
$$B=P\left(\begin{array}{cccccc}\lambda_1&&&&&\\&\ddots&&&&\\&&\lambda_k&&&\\&&&0&&\\&&&&\ddots&\\&&&&&0\end{array}\right)Q.$$
Note that the images of $P$ and $Q$ in $M_l(\mathbb{Z}_{m_l})$ are also invertible. Hence
\begin{eqnarray*}
&&\sum_{0\leq\alpha_{ij}<m_l}\zeta_{m_l}^{\left(\begin{array}{ccc}\alpha_{11}&\ldots&\alpha_{1l}\end{array}\right)B\left(\begin{array}{ccc}\alpha_{21}\\\vdots\\\alpha_{2l}\end{array}\right)}\\
&=&\sum_{0\leq\alpha_{ij}<m_l}\zeta_{m_l}^{\left(\begin{array}{ccc}\alpha_{11}&\ldots&\alpha_{1l}\end{array}\right)\left(\begin{array}{cccccc}\lambda_1&&&&&\\&\ddots&&&&\\&&\lambda_k&&&\\&&&0&&\\&&&&\ddots&\\&&&&&0\end{array}\right)\left(\begin{array}{ccc}\alpha_{21}\\\vdots\\\alpha_{2l}\end{array}\right)}\\
&=&m_l^{2(l-k)}\sum_{\alpha_{11}=0}^{m_l-1}\sum_{\alpha_{21}=0}^{m_l-1}\zeta_{m_l}^{\alpha_{11}\lambda_1\alpha_{21}}\cdots\sum_{\alpha_{1k}=0}^{m_l-1}\sum_{\alpha_{2k}=0}^{m_l-1}\zeta_{m_l}^{\alpha_{1r}\lambda_r\alpha_{2r}}
\end{eqnarray*}
It is well known that if $\xi^m=1$, then
$$\sum_{i=0}^{m-1}\xi^{id}=\left\{\begin{array}{ll}m, &\text{if }\xi^d=1;\\0,&\text{if }\xi^d\ne1.\end{array}\right.$$
Thus we have
\begin{eqnarray*}
Z^{[\omega]}(T^2)&=&\frac{1}{|G|}\frac{1}{(\frac{m_l}{m_1})^2\cdots(\frac{m_l}{m_{l-1}})^2}m_l^{2(l-k)}m_l^k(m_l,\lambda_1)\cdots(m_l,\lambda_k) =\frac{|G|}{\frac{m_l}{(m_l,\lambda_1)}\cdots\frac{m_l}{(m_l,\lambda_k)}}.
\end{eqnarray*}
If $l=1$, $Z^{[\omega]}(T^2)=|G|=|G_0|$. The conclusion also holds.
\end{proof}
\section*{Acknowledgment}
The authors are grateful to J. C. Wang for bringing the works \cite{spt, ww, wen} to their attention, in connection with Subsection 2.5 and Subsection 4.2.
|
\section{Introduction}
\subsection{Measuring Argument Strength}
\label{SEC:IntroMeasure}
Probabilistic models of argumentation became popular in cognitive science and its subfields including psychology, philosophy, and computer science in recent years \cite<see, e.g.,>{hahn06a,haenni09,zenker13}. Like logic-based nonmonotonic approaches for defeasible argumentation \cite<see, e.g.>{prakken02}, probabilistic approaches allow for dealing with exceptions and retracting conclusions in the light of new evidence. However, in contrast to qualitative logical approaches, probability allows for managing \emph{degrees of belief} in the sentences involved in common sense argumentation. Moreover, degrees of belief can be used to model the strength of arguments \cite{hahn06a,oaksford07c,pfeifer06d,pfeifer07,pfeifer13a}.
The concept ``argument'' is ambiguous. In logic, it denotes a triple consisting of a (possibly empty) premise set, a conclusion indicator, and a conclusion set. Consider, for example, the following argument, which is an instance of modus ponens:
\begin{tabular}{ll}
(P1)&If I take the train at five ($T$), I'll be home at six ($H$).\\
(P2)&I take the train at five ($T$).\\
(C) &Therefore, I'll be home at six ($H$).\\
\end{tabular}
Here, (P1) and (P2) are the premises, ``Therefore'' the conclusion indicator and the sentence ``I'll be home at six'' is the conclusion. In argumentative contexts, ``argument'' may also denote a premise which speaks for or against a conclusion. For example ``The train conductors are on strike'', can serve as an argument for concluding that it is better to take the bus. In what follows, however, we will focus on arguments in the logical sense only.
How can we measure the strength of an argument? There are at least two formal approaches to study (probabilistic) argument strength. In the first approach argument strength is based on \emph{uncertain consequence} relations, i.e., by presupposing that the conclusion follows to some degree from the premises. Usually, this is modeled by a conditional probability of ``the conclusion \emph{given} (some combination of) the premises'' of the argument \cite<see, e.g.>{hahn06a,oaksford07c}. As pointed out by \citeA{osherson90}, measures of confirmation can serve as models for argument strength \cite<for an overview of measures of confirmation see>{crupi07}. Measures of confirmation and previous attempts to model argument strength by uncertain consequence relations are problematic when arguments involve conditionals, like the modus ponens above (see premise (P1)): it is far from clear to give a precise meaning of conditionalizing on a combination of premises, when the premise set contains conditional events. There is ample formal and experimental evidence that uncertain conditionals are best modeled by conditional probabilities \cite<see, e.g.,>{evans04,oaksford07,OverCruz17,pfeifer13,pfeifer13b}. Therefore, conditionals should be modeled by conditional probabilities. However, this requirement would imply to measure the uncertainty of a \emph{conclusion \underline{given} (some combination of) the premises}. Unfortunately, satisfactory semantics of expressions like $\overbrace{C}^{\text{conclusion}} | \overbrace{(A \text{ and } (C|A))}^{\text {premises}}$ do not exist yet. Such semantics would, however, be necessary to capture the underlying logical structure of the modus ponens \cite<for an approach where conditionals are interpreted as conditional random quantities which allows for dealing with nested conditionals, see>{2016:SMPS1,GOPSsubm}. Modus ponens is just a relatively simple example here: there are, of course, many other argument forms involving conditionals. The inability to deal with conditionals seems to us to be one of the main reasons, why currently no formally satisfactory measure of argument exists within the first approach: measures based on uncertain consequence relations do not seem to be able to deal with the logical form of the argument.
In this paper, we advocate the second approach to argument strength. It satisfies the requirement of doing justice to the logical form of arguments involving conditionals \cite{pfeifer07,pfeifer13a}. Specifically, we define argument strength based on the following ideas: (i) keep the consequence relation deductive, (ii) assign probabilities to the premises, and then (iii) define the measure of argument strength based on the propagated coherent lower and upper probability bounds on the conclusion \cite{pfeifer07,pfeifer13a}.
Probability propagation from the premises to the conclusion is governed by \emph{coherence based probability logic} \cite<see, e.g.>{coletti02,pfeifer09b,gilio16}. The \emph{coherence approach} to probability was originated by Bruno de Finetti \cite{definetti74}. It conceives probabilities as subjective degrees of belief. Conditional probabilities ($p(C|A)$) are primitive. This allows for zero probabilities of the conditioning event ($A$). Note that in standard approaches to probability, $p(C|A)$ is undefined if $p(A)=0$, which is problematic in many argument forms \cite<see, e.g.>{pfeifer13,gilio16}.
Moreover, coherence allows for managing \emph{imprecise probabilities} (set-valued probabilities involving lower and upper probability bounds), which is relevant for formalising arguments under incomplete probabilistic knowledge. The above mentioned modus ponens, for example, is formalised as follows:\\
\begin{tabular}{lp{7cm}}
(P1') & $p(H|T)=x$\\
(P2') &$p(T)=y$\\
(C') & Therefore, $z' \leq p(H)\leq z''$, where $z'=xy$ and $z''=xy+1-y$ are the best possible coherent probability bounds on the conclusion.\\
\end{tabular}\\
Following \citeA{pfeifer13a}, we define the measure of argument strength {$\mathfrak{s}$ } on an argument $\mathcal{A}$ as follows:
\begin{quote}
Let $z'$ and $z''$ denote the coherent lower and upper probability bounds, respectively, on the conclusion of argument $\mathcal{A}$. Then,
\begin{equation}\label{EQN:ArgStr}
\mathfrak{s}(\mathcal{A})=_{\text{def.}}\overbrace{(1-(z''-z'))}^{\text{precision}} \times \overbrace{\frac{z'+z''}{2}}^{\text{location}}\, .
\end{equation}
\end{quote}
Intuitively, measure {$\mathfrak{s}$ } combines the \emph{precision} and the \emph{location} of the coherent conclusion probability interval. Specifically, strong arguments are arguments with \emph{low imprecision} of the conclusion probability (measured by the one-complement of the distance between the upper and the lower probability bounds, $1-(z''-z')$) and with conclusion probabilities \emph{close to one} (measured by the mean of the lower and upper probability bound, $(z'+z'')/2$). Of course, precision and location could be modeled differently (e.g., by using the geometric or the harmonic mean instead of the arithmetic mean). Moreover, in contexts where the location is more important than the precision of the conclusion probability interval (or \emph{vice versa}), adding suitable weights to formula~(\ref{EQN:ArgStr}) can adjust the measure for such cases. However, for the purpose of our paper it is sufficient to keep the measure as simple as possible.
Measure {$\mathfrak{s}$ } has a number of plausible consequences:
it ranges always from zero to one (i.e., $0 \leq \mathfrak{s}\leq 1$, since $z'$ and $z''$ are probability values, which are also in the unit interval, $[0,1]$). The extreme ``0'' denotes weak arguments and ``1'' denotes strong arguments. Arguments with conclusion probability 1, are strong arguments, since $\mathfrak{s} = 1$ if $z'=z''=1$. Arguments with conclusion probability 0 (i.e., $z'=z''=0$) are weak arguments, since $\mathfrak{s} = 0$. Likewise, probabilistically non-informative arguments (i.e., $z'=0$ and $z''=1$) are weak arguments, since $\mathfrak{s} = 0$.
Interestingly, measure {$\mathfrak{s}$ } also provides a new solution to the Ellsberg paradox \cite{ellsberg61},\footnote{We thank Kevin T. Kelly for pointing us to the Ellsberg paradox.} which we describe in the next section.
\subsection{Modeling the Ellsberg Paradox by Measure {$\mathfrak{s}$ }}
\label{SEC:IntroEllsberg}
Ellsberg described the following situation \cite{ellsberg61}:
\begin{quote}
An urn contains 90 balls, of which 30 are red ($R$) and 60 are black or yellow ($B \vee Y$, where ``$\vee$'' denotes \emph{disjunction} (``or'') as defined in classical logic). The ratio of the black and yellow balls is unknown---there might be anything between 0 to 60 black (or yellow) balls. One ball is drawn from the urn and you are asked to choose a bet between two bets. If you take \textbf{Bet 1}, you will win \$100, if the ball drawn from the urn is red. If you take \textbf{Bet 2}, you will win \$100, if the ball drawn from the urn is black.
\end{quote}
Ellsberg predicted that most people choose {\bf Bet 1} when asked to decide which of the two bets they prefer. Then, considering again the same urn, Ellsberg predicted that people will choose {\bf Bet 4}, when they are asked to decide between the following two alternative bets:
\begin{quote}
If you take {\textbf{Bet 3}}, you will win \$100, if the ball drawn from the urn is red or yellow. If you take {\textbf{Bet 4}}, you will win \$100, if the ball drawn from the urn is black or yellow.
\end{quote}
Ellsberg's predictions create a well-known paradox as they violate the independence axiom of rational choice \cite<see, e.g.,>{sep-rationality-normative-utility}. Moreover, Ellsberg's predictions were experimentally confirmed in many studies \cite<see, e.g.,>{becker64,slovic74,maccrimmon79}.
We propose to frame the Ellsberg paradox in terms of probability logical arguments. Specifically, the \emph{premises} represent the probabilistic information given in the description of the urn, and the \emph{conclusions} represent the respective bets involved in the Ellsberg paradox. Thus, we obtain four arguments. Each argument speaks for choosing the corresponding bet. The associated argument to \textbf{Bet 2}, for example, is argument $\mathcal{A}_2$:
\begin{quote}
$p(R)=.33$\\ $p(B \vee Y)=.67$\\
Therefore, $0\leq p(B)\leq.67$ is coherent.
\end{quote}
The strength of this argument is denoted by $\mathfrak{s}(\mathcal{A}_2)$ and by applying equation~(\ref{EQN:ArgStr}) equal to .11 (i.e., $\mathfrak{s}(\mathcal{A}_2)=.11$).
Table~\ref{TAB:ArgEllsb} lists the conclusions and the argument strengths {$\mathfrak{s}$ } for each argument for the corresponding four bets involved in the Ellsberg paradox.
\begin{table}[!ht]
\begin{center}
\caption{Conclusions and normative strengths ($\mathfrak{s}$) of Arguments $\mathcal{A}_1, \ldots, \mathcal{A}_4$ associated with the four bets involved in the Ellsberg paradox. The premises are always $p(R)=.33$ and $p(B \vee Y)=.67$.}
\label{TAB:ArgEllsb}
\vskip 0.12in
\begin{tabular}{ccc}\hline
& Conclusion& Argument strength \\\hline
{\bf Bet 1}& $p(R)=.33$&$\mathfrak{s}(\mathcal{A}_1)=.33$ \\
{\bf Bet 2} & $0\leq p(B)\leq.67$&$\mathfrak{s}(\mathcal{A}_2)=.11$\\
{\bf Bet 3}& $.33 \leq p(R\vee Y) \leq 1$&$\mathfrak{s}(\mathcal{A}_3)=.22$\\
{\bf Bet 4}&{$p(B\vee Y)=.67$}&$\mathfrak{s}(\mathcal{A}_4)=.67$\\\hline
\end{tabular}
\end{center}
\end{table}
The four argument strength values in Table~\ref{TAB:ArgEllsb} induce the following preference orders in the classical Ellsberg task: {\bf Bet 1} $\succ$ {\bf Bet 2}, since $\mathfrak{s}(\mathcal{A}_1)=.33>\mathfrak{s}(\mathcal{A}_2)=.11$, and {\bf Bet 4} $\succ$ {\bf Bet 3}, since $\mathfrak{s}(\mathcal{A}_4)=.67>\mathfrak{s}(\mathcal{A}_3)=.22$ (where $ X \succ Y$ denotes \emph{$X$ is preferred over $Y$}). This preference order corresponds to Ellsberg's predictions and matches the data \cite<see, e.g.,>{becker64,slovic74,maccrimmon79}.
The functions of the four arguments can be understood in an \emph{epistemic} and in a \emph{persuasive} sense. The epistemic function of the arguments is to \emph{gain knowledge} about which bet should be preferred. The persuasive function of the arguments is to \emph{convince} someone which bet should be preferred.
In the following section we further investigate the psychological plausibility of {$\mathfrak{s}$ } by an experiment.
\section{Method}
\subsection{Participants}
In this experiment 60 university students (mean age 25.9 years ($SD=5.6$), 48 females, 12 males) participated for a compensation of 15\EUR. All of the participants were Finnish native speakers and none of them had studied psychology, mathematics, statistics or philosophy as their major.
\subsection{Design and Materials}
We used three target task types: argument ranking tasks, argument rating tasks, and the (original) Ellsberg tasks.
The \emph{argument ranking tasks} first instructed the participants to \emph{rank} the strength of arguments $\mathcal{A}_1$ and $\mathcal{A}_2$ (see Table~\ref{TAB:ArgEllsb}). Second, the participants were instructed to rank the strength of arguments $\mathcal{A}_3$ and $\mathcal{A}_4$. The \emph{argument rating tasks} instructed the participants to \emph{rate} the strength of each of the four arguments. In the original version of the Ellsberg task, participants had to rank which bets they preferred as described in the Introduction. We investigated the following questions which relate argument strength to the Ellsberg problem:
\begin{itemize}
\item Do the results of the argument strength rating tasks predict the responses in the Ellsberg tasks?
\item Do the results of the argument strength rating tasks predict the responses in the argument strength ranking tasks?
\end{itemize}
Moreover, we explored empirically, whether argument strength formulated in epistemic or in persuasive terms impacts participants' reasoning. Finally, we systematically manipulated the information conveyed in the argument rating and in the argument ranking tasks by the following independent variables: (i) only the uncertainty of the conclusion was presented, (ii) only the uncertainties of the premises were presented, and (iii) uncertainties of the premises and the conclusion were presented. The instructions introduced the following symbol for marking not conveyed information in the respective conditions which correspond the variables (i) and (ii): \includegraphics[width=.021\textwidth]{mute.eps}. By using a $2\times 3$ between-participant design we fully crossed epistemic versus persuasive formulations and the manipulated information conveyed in the arguments.
In the epistemic booklets we used knowledge-oriented phrasings like ``Which argument is stronger to know which bet to choose?'', whereas in the persuasive booklets we used according phrasings like ``Which argument convinces stronger which bet to choose?''. The experimental conditions are explained in Table~\ref{TAB:design}.
\begin{table}[!ht]
\begin{center}
\caption{Experimental conditions (Cd 1--Cd 6; $N=60$).}
\label{TAB:design}
\vskip 0.12in
\begin{tabular}{lll}
\hline
Presented probabilities &Epistemic & Persuasive \\\hline
Premise \& conclusion & Cd 1 ($n_1=10$) & Cd 2 ($n_2=10$) \\
Conclusion only& Cd 3 ($n_3=10$) & Cd 4 ($n_4=10$) \\
Premise only& Cd 5 ($n_5=10$) & Cd 6 ($n_6=10$) \\\hline
\end{tabular}
\end{center}
\end{table}
\paragraph{Argument ranking tasks}
In these tasks, the participants were instructed to imagine two friends arguing about which bet the participant should choose. Then, argument $\mathcal{A}_1$ for {\bf Bet~1}, and argument $\mathcal{A}_2$ for {\bf Bet~2} were presented to the participant, e.g.:
\begin{quote}
\noindent {\bf Argument 2} for {\bf Bet~2}\\
\fbox{
\begin{minipage}[c][7em][c]{0.43\textwidth}
I am \includegraphics[width=.035\textwidth]{mute.eps} \% sure that the ball drawn from the urn is red.\\
I am \includegraphics[width=.035\textwidth]{mute.eps} \% sure that the ball drawn from the urn is black or yellow.\\
{\underline{Therefore}}, I am at least 0 \% and at most 67 \% sure that the ball drawn from the urn is black.
\end{minipage}}
\end{quote}
The participants were then presented with the question ``Which argument is stronger to know which bet to choose?'' (\emph{Kumpi argumentti on vahvempi sen tiet\"amiseen, kumpi veto kannattaisi valita?}) in the epistemic condition. In the persuasive condition, they were asked \emph{``Which argument convinces you stronger which bet to choose?''} (\emph{Kumpi ar\-gu\-ment\-ti va\-kuut\-taa sinut vah\-vem\-min siit\"a, kumpi veto kan\-nat\-tai\-si valita?}). Then, the participants were instructed to indicate their choice by ticking the respective box for Argument~1 (i.e., $\mathcal{A}_1$) or Argument~2 (i.e., $\mathcal{A}_2$). Finally, the participants ranked Argument~3 (i.e., $\mathcal{A}_3$) and Argument~4 (i.e., $\mathcal{A}_4$).
\paragraph{Argument rating tasks}
In these tasks participants were presented with the same four arguments as in the argument ranking tasks. They were asked to carefully reconsider each. Instead of using forced choice response formats, each argument was followed by a question, e.g., {``How strong is {\bf Argument~2} for choosing {\bf Bet~2}?''} (\emph{Kuinka vahva \textbf{Argumentti 2} on \textbf{Vedon~2} valitsemiseksi?}; original epistemic formulation) or {``How strong is {\bf Argument~2} for convincing to choose {\bf Bet~2}?''} (\emph{Kuinka vahva \textbf{Argumentti 2} on vakuuttamaan \textbf{Vedon~2} va\-lit\-se\-mi\-ses\-ta?}; original persuasive formulation).
The participants were asked to mark their responses on a scale (see Figure~\ref{argscale}).
\begin{figure}[ht]
\begin{center}
\includegraphics[width=.5\textwidth]{AnswerScale_3.eps}
\end{center}\vspace{-1cm}
\caption{Answer scale used in the argument rating tasks.}
\label{argscale}
\end{figure}
\paragraph{Ellsberg tasks} Here, as explained in the introduction, the participants had to choose which rankings among bets they preferred ({\bf Bet~1} or {\bf Bet~2} and {\bf Bet~3} or {\bf Bet~4}). All participants were presented with the same Ellsberg tasks.
\subsection{Procedures}
Participants completed the booklets individually in a quiet room. At the beginning of the testing, participants were informed to take as much time as needed for completing the tasks. Furthermore, they were instructed not to look back on their previous responses.
After reading the introduction the participants worked on tasks which differed from the Ellsberg problem (and which are not in the scope of the present paper). After that, the target tasks were presented in the following order: (i) argument ranking tasks, (ii) argument rating tasks, and (iii) the Ellsberg tasks. Finally, the participants filled in demographic data and answered questions about the difficulty and clearness of the tasks. Each session concluded by an interview to further explore argument strength from a qualitative point of view: we asked how the participants solved the tasks and what they thought determined the strength of an argument. Participants used on the average 9.6 minutes ($SD=2.8$) to work on the target tasks and to fill in the final questions in the booklet.
\section{Results and Discussion}
We performed Fisher's exact tests to compare the impact of the different booklets on the response frequencies in the argument ranking tasks and in the Ellsberg tasks. Moreover, we tested influences of the different conditions in the argument rating tasks by analyses of variance. After performing Holm-Bonferroni corrections we did not observe any significant differences. We therefore pooled the data for further analysis ($N=60$).
\paragraph{Ellsberg's predictions}
The majority of responses in all three types of tasks (i.e., argument ranking, argument rating and Ellsberg task) are consistent with Ellsberg's predictions.
Our findings also replicate empirical findings reported in the literature \cite<see, e.g.,>{becker64,slovic74,maccrimmon79}. Moreover, our data suggest that classical findings in Ellsberg tasks carry over to (isomorphic) problems formulated in terms of argument strength.
Table~\ref{ellsbergfrequence} shows how the participants ranked the arguments in the argument ranking tasks and how they ranked the bets in the Ellsberg tasks. Bet~1 (resp., argument $\mathcal{A}_1$ supporting Bet~1) is more frequently chosen than Bet~2 (resp., $\mathcal{A}_2$ supporting Bet~2). Likewise, Bet~4 (resp., argument $\mathcal{A}_4$ supporting Bet~4) is more frequently chosen than Bet~3 (resp., $\mathcal{A}_3$ supporting Bet~3).
\begin{table}
\begin{center}
\caption{Percentages of argument preferences in the argument ranking tasks (rnk$(\mathcal{A})$) and in the Ellsberg tasks ($N=60$).}
\label{ellsbergfrequence}
\vskip 0.12in
\begin{tabular}{ccccccc}
\cline{1-3} \cline{5-7}
\% &rnk$(\mathcal{A})$ &Ellsberg & &\% &rnk$(\mathcal{A})$ &Ellsberg \\ \cline{1-3} \cline{5-7}
Bet1 &73,3 &93,3 & &Bet3 &25,0 &23,3 \\
Bet2 &26,7 &~6,7 & &Bet4 &75,0 &76,7 \\ \cline{1-3} \cline{5-7}
\end{tabular}
\end{center}
\end{table}
Moreover, we constructed the underlying preference orders of the argument strengths and the bets from the participants' responses in all the three task types. This allows one to see which choice strategies were most commonly used. In all tasks, strategies consistent with the independence axioms of rational choice
were less frequently preferred, as can be seen in Table~\ref{strategy}. For constructing the preference orders based on the responses in the argument strength ratings tasks, we made the following assumption: if the strength of an argument $\mathcal{A}_x$ was rated higher than the strength of an argument $\mathcal{A}_y$, then the corresponding Bet~$x$ is preferred over Bet~$y$. Again, our findings replicate the predictions of Ellsberg and the previous empirical findings \cite<see, e.g.,>{becker64,slovic74,maccrimmon79}.
\begin{table}[!ht]
\begin{center}
\caption{Percentages of responses consistent with Ellsberg's predictions ($E$), the independence axiom of rational choice ($I$). The preference order $R$ can be interpreted as a reversed version of $E$. ``$(x , y) \succ (u , v)$'' means ``arguments (resp. bets) $x$ and $y$ are preferred over arguments (resp. bets) $u$ and $v$''. Preference order responses consistent with {$\mathfrak{s}$ } are in \textbf{bold}.}
\label{strategy}
\vskip 0.12in
\begin{tabular}{cccc}
\hline
Preference&\multicolumn{3}{c}{Tasks ($N=60$)}\\\cline{2-4}
Order&$\mathcal{A}$ Ranking& Ellsberg & $\mathcal{A}$ Rating\\\hline
$(1 , 4) \succ (2 , 3)^E$\!\!\!\!& \textbf{56.67}& \textbf{71.67}& \textbf{56.10}\\
$(2 , 3) \succ (1 , 4)^R$\!\!\!\!&8.33 & 1.67& 4.88 \\
$ (1 , 3) \succ (2 , 4)^I$\!\!\!\!& 16.67& 21.67& 21.95\\
$ (2 , 4) \succ (1 , 3)^I$\!\!\!\!& 18.33& 5.00& 17.07\\\hline
\end{tabular}
\end{center}
\end{table}
Table~\ref{ellsbergratings} shows the mean argument strength rating responses. As predicted by measure $\mathfrak{s}$, the mean argument strength ratings reflect the Ellsberg predictions, i.e., $\text{mean rating}(\mathcal{A}_1)>\text{mean rating}(\mathcal{A}_2)$ and $\text{mean rating}(\mathcal{A}_4)>\text{mean rating}(\mathcal{A}_3)$.
\begin{table}[!ht]
\begin{center}
\caption{Means and standard deviations ($SD$) of the argument strength ratings on a scale from 0 (``extremely weak'') to 10 (``extremely strong''; $N=60$).}
\label{ellsbergratings}
\vskip 0.12in
\begin{tabular}{ccccc}
\hline
&$\mathcal{A}_1$ & $\mathcal{A}_2$ &$\mathcal{A}_3$ &$\mathcal{A}_4$ \\\hline
Mean & 5,20 & 3,98 & 5,77 & 6,95 \\
$SD$& 2,64 & 2,58 &1,74 &1,87 \\\hline
\end{tabular}
\end{center}
\end{table}
\paragraph{Consistency among the data}
Based on the argument strength ratings, we predicted the participants' choices in the ranking and in the Ellsberg tasks. The data support our predictions: the argument strength rating responses predict the ranking responses in the Ellsberg tasks. The rating responses also predict the responses in the argument strength ranking tasks (see Table~\ref{prediction1} and Table~\ref{prediction2}).
As some participants had rated the arguments for the bets equally strong, no predictions could be derived in these cases. When taking into account only those cases, in which making predictions was possible, the responses of roughly 3/4 of the participants were consistent with their responses in the ranking tasks. In the argument strength ranking tasks, 77.3 \% of the participants chose as predicted between the first two bets and 75.0 \% chose as predicted between the second two bets. For the Ellsberg tasks, we observed similarly high percentages (i.e., 75.0 \% and 70.8 \% of the participants, for the first and the second bet rankings, respectively). This is again strong experimental support for the psychological plausibility of measure $\mathfrak{s}$.
\begin{table}[!ht]
\begin{center}
\caption{Predictions of bet rankings in Ellsberg tasks based on responses in the argument strength rating tasks ($N=60$).}
\label{prediction1}
\vskip 0.12in
\begin{tabular}{lcc}
\hline
& \multicolumn{2}{c}{Ranking}\\ \cline{2-3}
\% & Bet 1 vs. Bet 2\!&Bet 3 vs. Bet 4 \\\hline
Chose as predicted & 55.00 & 56.67 \\
Did not choose as predicted\!\!\!\!\!\!\!\!\!\!\!& 18.33 & 23.33 \\
No prediction made & 26.67 & 20.00 \\\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!ht]
\begin{center}
\caption{Predictions of argument strength rankings based on the responses in argument strength rating tasks ($N=60$).}
\label{prediction2}
\vskip 0.12in
\begin{tabular}{lcc}
\hline
& \multicolumn{2}{c}{Ranking}\\ \cline{2-3}
\% & $\mathcal{A}_1$ vs. $\mathcal{A}_2$ &$\mathcal{A}_3$ vs. $\mathcal{A}_4$ \\\hline
Chose as predicted & 56.67 & 60.00 \\
Did not choose as predicted& 16.67 & 20.00 \\
No prediction made & 26.67 & 20.00 \\\hline
\end{tabular}
\end{center}
\end{table}
Finally, we discuss qualitative data taken from structured interviews on folk psychological conceptions on what argument strength means.
\paragraph{Interview results} After the participants completed the paper and pencil tasks, we collected folk psychological conceptions on what ``argument strength'' (\emph{argumentin vahvuus}) means by structured interviews.
We asked the participants how they would define argument strength in their own words. Participants who had received the \emph{persuasive} booklets, we hypothesized, mentioned persuasive aspects (like how \emph{convincing} arguments are) more frequently than those of the epistemic condition. Moreover, participants who had received the \emph{epistemic} booklets focused more on epistemic aspects (like truth and knowledge) than those of the persuasive condition. However, the interview responses do not confirm these hypotheses.
The responses to the interview question concerning the meaning of ``argument strength'' reflected features of our measure $\mathfrak{s}$.
Specifically, the \emph{location} of the coherent conclusion probability interval was referred to by almost all of the participants.
For many participants the location seemed to be more important than the \emph{precision} of the coherent conclusion probability interval. They had, for example, focused solely on the lower probability bound of the interval and ignored the upper bound or responded based on the mean value of the interval.
However, a few participants also referred to the \emph{precision} of the coherent conclusion probability interval by sentences like:
\begin{quote}
``The size of this gap between 33 [\%] and 100 [\%] is so big that it increases the uncertainty.'' \emph{(Ep\"avarmuutta lis\"a\"a se, ett\"a v\"ali 33:n ja 100:n v\"alill\"a on niin suuri)}
\end{quote}
Some participants also talked about the truth or correctness of the probability bounds of the conclusion. For them, the arguments were strong, when the probabilities in the conclusions were \emph{correct}, almost regardless of the values in them.
Finally, we note that the interview responses provide folk psychological evidence for using location and precision of conclusion probability intervals for evaluating the strength of uncertain arguments. Location and precision are the key ingredients of our measure of argument strength~$\mathfrak{s}$.
\section{Concluding Remarks}
We proposed a formal measure of argument strength and showed how it predicts responses in Ellsberg tasks. Specifically, we framed choices among bets in terms of probability logical argument forms. We confirmed experimentally that Ellsberg's predictions can be justified by argument strength rankings and argument strength ratings.
Since the proposed measure exploits tools available in coherence-based probability logic and since it is based on a deductive consequence relation, it allows for dealing with arguments involving conditionals. The proposed measure has many plausible consequences, which calls for future formal-normative and experimental research for modeling also other argument types, like the conditional syllogisms.
Understanding argument strength is important for theories about reasoning and argumentation in general. Our paper sheds formal and experimental light on what argument strength means.
\section{Acknowledgments}
This research was supported by the DFG project
PF~740/2-2 (awarded to Niki Pfeifer) as part of the Priority Program ``New Frameworks of Rationality'' (SPP1516).
|
\section{Introduction}\label{sec:intro}
Compressed sensing (CS) \cite{Donoho2006CompressedSensing,Candes2006CompressiveSampling} states that most of the signals of scientific interest can be approximated very accurately using a smaller number of measurements, compared to the dimension of the signal. For that, the signal needs to be sparse or have a sparse representation in terms of proper sparsifying bases. This observation has a huge impact in signal processing, machine learning, and statistics. Mathematically speaking, the goal of the CS problem is to recover the signal $\boldsymbol{x} = [x_1,x_2,\dots, x_N]^T$ with length $N$ from its undersampled random projections, also referred to as measurements. $M$ random projections are generated using a measurement matrix $\boldsymbol{\Phi} \in \mathbb{R}^{M\times N}$ from the linear measurement process,
$
\boldsymbol{y} = \boldsymbol{\Phi} \boldsymbol{x} + \boldsymbol{n},
$ where $\boldsymbol{y} = [y_1,y_2,\dots, y_M]^T$ represents the measurement vector and $\boldsymbol{n}$ denotes the corrupting noise. \par
Signal $\boldsymbol{x}$ is said to be $K$-sparse if it has at most $K$ non-zero entries in a proper basis. The sparsity of $\boldsymbol{x}$ can be exploited to find a unique solution of the underdetermined system equation with high probability from $\mathcal{O}(K\log(\frac{N}{K}))$ measurements \cite{Candes2006CompressiveSampling}. \par
In this paper, we consider the problem of reconstructing a correlated time series of such compressible vectors from their noisy undersampled measurement. Particularly, we are interested in approximating the time series $\{ \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \dots \}$ from the measurement time series $\{ \boldsymbol{y}^{(1)}, \boldsymbol{y}^{(2)}, \dots \}$. In many real-world applications, the signal of interest has a substantial correlation in time. The main idea is to incorporate the knowledge from the previous estimates of the signal to achieve a more accurate estimation of the signal at the current time step. \par
Moreover, in many applications, different parts of the signal have different recovery requirements. Thus, different coefficients of the signal have different \emph{importance levels}. For instance, in video processing, it is desired to recover the salient area more accurately. Moreover, if the signal is sparse in canonical basis, we are interested in reconstructing the large coefficients with less error. \emph{Non-uniform} acquisition and recovery of signal is desirable in many applications such as image processing \cite{shahrasbi2016model}, camera sensor networks \cite{Uddin2011PhotoNet:Awareness,Rahimpour2016DistributedNetworks}, wireless sensor networks \cite{Leinonen2015SequentialNetworks}, collaborative vector estimation \cite{Sani2016DistributedNetworks}, component analysis \cite{Rahmani2016ASampling,Hosseini2016Cloud-basedPrediction}, and internet of things \cite{Fragkiadakis2014AdaptiveApplications}.
In this work, we propose an adaptive framework to design a non-uniform measurement matrix, which contrast with \emph{dynamic CS algorithms} \cite{Ziniel2013DynamicPassing,Wijewardhana2016AMeasurements,Shahrasbi2011TC-CSBP:Propagation} focusing only on the recovery step. Our method is also distinct from the \emph{adaptive CS} \cite{Malloy2014Near-OptimalSensing,Braun2015Info-GreedySensing,Haupt2012SequentiallySensing} methods that are concerned with reconstructing signals, which are static over time. Here, similar to adaptive CS, the main idea is to concentrate the sensing energy on the more important coefficients, by designing a proper measurement matrix. However, due to dynamic nature of the problem, the algorithm should not make firm decisions about the location of more important coefficients. Hence, soft importance level information is advantageous. To infer the importance level of each coefficient at each time step, a generative model is imposed on the coefficients and the parameters of the model are updated in an online fashion. \par
\begin{figure}
\centering
\includegraphics[width=\columnwidth,angle=0]{Figures/model_ANCS.png}\vspace{-1.5mm}
\caption{\small Overall block diagram of the proposed framework.Reconstructed signal, at each time step, is utilized to generate the measurement matrix. }
\label{fig:block_diagram}
\vspace{-5mm}
\end{figure}
Figure \ref{fig:block_diagram} shows the overall architecture of the proposed method. At each time step, after reconstructing the signal, by using a conventional CS recovery algorithm, the importance levels of the coefficients are inferred mathematically. The importance levels are further employed to design the measurement matrix for the next time step. \par
The rest of this paper is organized as follows. In Section \ref{sec:system_model}, the system model is presented. Then, the generative model of the proposed Bayesian framework is introduced in Section \ref{sec:Bayesian}. In Section \ref{sec:matrix_design}, the inferred importance level information are used to design the measurement matrix for sensing. Finally, Section \ref{sec:results} presents the simulation results and Section \ref{sec:conclusions} draws conclusions. \par
\section{System Model}\label{sec:system_model}
We consider recovery of a vector-valued time series $\{ \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \dots \}$ from the linear measurements given by
\begin{equation}
\label{eq:meas_model}
\boldsymbol{y}^{(t)} = \boldsymbol{\Phi}^{(t)}\boldsymbol{x}^{(t)} + \boldsymbol{n}^{(t)}, \qquad t = 1,2,\dots
\end{equation}
where $\boldsymbol{n}^{(t)} \in \mathbb{R}^M$ represents the noise and is modeled as an additive white Gaussian noise (AWGN) with $\boldsymbol{n}^{(t)} \sim \mathcal{N}(\boldsymbol{0},\sigma_{n}^2\boldsymbol{I}_{M})$. \par
It is assumed that the signal of interest $\boldsymbol{x}^{(t)}$ is compressible and contains coefficients with different importance levels, which are not known a priori. In many scenarios, it is desirable to have non-uniform recovery performance on different parts of signal. More important coefficients may correspond to support of a sparse vector or the salient area in a video frame. \par
We also assume that, at each time step, using the estimation of the signal $\hat{\boldsymbol{x}}^{(t)}$, more important coefficients are tagged using a possibly erroneous algorithm. The variable $\boldsymbol{\alpha}^{(t)}$ marks the detected \emph{region of interest} (ROI) in the signal at time $t$. Specifically, $\alpha_n^{(t)} = 1$, if the $n^{\text{th}}$ coefficient of the signal is detected to be in the ROI, and $\alpha_n^{(t)} = 0$ otherwise. However, due to sensing failure, error in recovery of $\hat{\boldsymbol{x}}^{(t)}$, and/or misdetection of the ROI, $\boldsymbol{\alpha}^{(t)}$ may contain erroneous elements. \par
As mentioned earlier, the signal of interest often exhibits substantial temporal correlation. Here, we assume that ROI, and therefore the support of non-zero entries in $\boldsymbol{\alpha}^{(t)}$, changes slowly in time. Our goal is to employ the temporal correlation to infer reliable importance level information and employ the importance levels to design a non-uniform measurement matrix.
\section{Bayesian Inference of Importance Levels} \label{sec:Bayesian}
To extract reliable information from possibly faulty ROI data $\boldsymbol{\alpha}^{(t)}$ , we propose to employ Bayesian inference. In Bayesian framework, the goal is to infer the probability distribution of hidden variables given the observations. The hidden variables are often the parameters that are desired to be estimated. Specifically, in our model, the following hidden variables are introduced:
\begin{enumerate}
\item Coefficient-specific reliability $u_{n} \in \{ 0,1 \}$, which is either $0$ or $1$ and describes the reliability of ROI data of the $n^{\text{th}}$ coefficient.
\item Overall reliability $r \in [0,1]$, denoting the overall trustworthiness of the ROI detection algorithm. For small values of $r$, the algorithm is more prone to reporting faulty data. A generally reliable algorithm will report trustworthy measurements on most of the coefficients.
\item Importance level for each coefficient $c_{n} \in [0,1]$, describing the probability that coefficient $n$ is in ROI.
\end{enumerate}
As mentioned earlier, in the proposed generative model, $\alpha_n$ is the observed variable. If $\alpha_n = 1$, the $n^{\text{th}}$ coefficient is detected to be in ROI, and $\alpha_n = 0$ otherwise. In this model, coefficient-specific reliability and overall reliability model the faulty data. Without them, all the observations would be assumed to be trustworthy, which is not the case in real-world scenarios. \par
Figure \ref{fig:graph_model} illustrates the graphical representation of the proposed generative model. The arrows in the graph represent the dependency among the variables. Hence, the observed ROI data depends on the actual importance level of the coefficients and the reliability of algorithm in detecting the ROI coefficients.
The goal of the inference algorithm is to obtain the probability distribution of the overall reliability, coefficient-specific reliability, and the importance levels, given the ROI data. At each time step $t$, the proposed model can be formulated as follows, for $n=1, \dots, N$:
\begin{equation}
\begin{aligned}
r &\sim ~\operatorname{Beta}(b^{1},b^{0}) \\
u_{n} &\sim ~\operatorname{Bernoulli}(r) \\
c_{n} &\sim ~\operatorname{Beta}(\beta^{1}_{n},\beta^{0}_{n}) \\
\alpha_{n}^{(t)} &\sim ~ u_{n} \operatorname{Bernoulli}(c_{n}) + (1- u_{n}) \operatorname{Bernoulli}(1 - c_{n})
\label{eq:gen_model}
\end{aligned}
\end{equation}
\begin{figure}
\centering
\scalebox{1}{
\begin{tikzpicture}
\tikzstyle{main}=[circle, minimum size = 7mm, thick, draw =black!80, node distance = 5mm,inner sep=1pt]
\tikzstyle{connect}=[-latex, thick]
\tikzstyle{box}=[rectangle, draw=black!100]
\node[main, fill = black!10] (alpha1) [] {$\alpha_{1}$ };
\node[main, fill = black!10] (alpha2) [below=of alpha1] {$\alpha_{2}$ };
\node[main] (alphadots) [below=of alpha2,draw=none] {$\vdots$ };
\node[main, fill = black!10] (alphaN) [below=of alphadots] {$\alpha_{N}$ };
\node[main] (u1) [left= of alpha1] {$u_{1}$ };
\node[main] (u2) [below=of u1] {$u_{2}$ };
\node[main] (udots) [below=of u2,draw=none] {$\vdots$ };
\node[main] (uN) [below=of udots] {$u_{N}$ };
\node[main] (c1) [right= of alpha1] {$c_{1}$ };
\node[main] (c2) [below=of c1] {$c_{2}$ };
\node[main] (cdots) [below=of c2,draw=none] {$\vdots$ };
\node[main] (cN) [below=of cdots] {$c_{N}$ };
\node[main] (r) [ below left=of u2] {$r$ };
\path (u1) edge [connect] (alpha1)
(u2) edge [connect] (alpha2)
(uN) edge [connect] (alphaN)
(c1) edge [connect] (alpha1)
(c2) edge [connect] (alpha2)
(cN) edge [connect] (alphaN)
(r) edge [connect] (u1)
(r) edge [connect] (u2)
(r) edge [connect] (uN)
;
\end{tikzpicture}
}
\caption{\small Graphical representation of the generative model. }
\label{fig:graph_model}
\vspace{-4mm}
\end{figure}
The observed variable $\alpha_n^{(t)}$ is modeled with summation of two Bernoulli distributions. This means that if the ROI data for $n^{\text{th}}$ coefficient is reliable, i.e. $u_n = 1$, $\alpha_n$ will be sampled from a Bernoulli distribution with true parameter for importance level, i.e. $c_n$. Otherwise, it will be sampled from $\operatorname{Bernoulli}( 1 - c_{n})$ and will be more probable to report faulty data. Since $c_n$ is used as the parameter of a Bernoulli distribution, it is the natural choice to model it with a Beta distribution. This is due to the fact that the conjugate prior for Bernoulli distribution is Beta distribution. \par
Similarly, the variable representing the overall reliability, i.e. $r$, is modeled with a Beta distribution. This is because the coefficient-specific reliability variables are sampled from $\operatorname{Bernoulli}(r)$. This means that if the ROI detection is reliable in general, ROI data on most of the coefficients will be reliable. This prior links the performance of the algorithm on different coefficients and reduces the chance of overfitting the coefficient-specific reliability. \par
As mentioned earlier, the goal of the inference algorithm is to obtain the distribution of hidden variables, given the observations, i.e. $\mathbb{P} \{\boldsymbol{c},\boldsymbol{u}, r| \mathcal{A} \}$. For compactness of notation, we set $\boldsymbol{c} = \{c_1,c_2,\dots, c_N\}$, $\boldsymbol{u} = \{u_1,u_2,\dots, u_N\}$, and $\mathcal{A} = \{ \boldsymbol{\alpha}^{(1)}, \boldsymbol{\alpha}^{(2)},\dots \}$. At each time step, after receiving the ROI data, the distribution of hidden variables are inferred by exploiting the data and the prior belief, represented by the prior distribution $\mathbb{P} \{\boldsymbol{c},\boldsymbol{u}, r\}$.
For that, we need to specify the joint distribution of the observation and the hidden variables. Specifically, using the model formulated in (\ref{eq:gen_model}), we have:
\begin{equation}
\label{eq:joint_dist}
\begin{aligned}
\mathbb{P}\{\mathcal{A} ,\boldsymbol{u},\boldsymbol{c},r \} =
\prod_{t = 1}^{\infty}\prod_{n = 1}^{N}
&\mathbb{P}\{\alpha_{n}^{(t)}|u_{n},c_{n}\}
\mathbb{P}\{u_{n}|r\}\\
&\mathbb{P}\{c_{n}|\beta^1_n,\beta^0_n\}\mathbb{P}\{r|b^1,b^0\}
\end{aligned}
\end{equation}
However, due to obvious practical reasons and to limit the history of the inference, the inference is performed using a few of recent observations. For that, a sliding window of length $W$ is utilized and the parameters of the posterior distributions are inferred using only the last $W$ observations. \par
To infer the importance level of the coefficients as well as the reliability of the ROI data, we need to find the posterior distribution given the ROI data, i.e., $\mathbb{P}\{\boldsymbol{u},\boldsymbol{c},r | \mathcal{A}\}$. However, directly obtaining the posterior distributions is not computationally feasible and results in explosive number of probability factors growing exponentially with number of coefficients. To handle the intractable integrals of the inference procedure, \emph{variational inference} is often employed \cite{Jordan2008graphical,jordan1999variational,Babagholami-Mohamadabadi2014AEstimation}. \par
In variational inference, the posterior distribution is assumed to be fully factorized over all the hidden variables. In other words, the posterior distribution is being approximated by a family of distributions, for which the inference procedure is tractable. For our model, the fully factorized approximation of the posterior distribution, also referred to as the variational distribution, is defined as:
\begin{equation}
\label{eq:variational_dist}
\mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\} =
\prod_{n}\mathbb{Q}\{c_{n}|\hat{\beta}^1_{n},\hat{\beta}^0_{n}\}
\mathbb{Q}\{u_{n}|\tau_{n}\}\\
\mathbb{Q}\{r_{n}|\hat{b}^1,\hat{b}^0\}.
\end{equation}
where $\hat{b}^1$, $\hat{b}^0$, $\hat{\beta}^1_{n}$, $\hat{\beta}^0_{n}$, and $\tau_{n}$ are the parameters of the factorized distributions. By introducing the variable $\tau_{n}$, we are seeking the best approximate of $\mathbb{P}\{\boldsymbol{u},\boldsymbol{c},r | \mathcal{A}\}$ among all the distributions $\mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\}$, by factorizing the distribution over \emph{disjoint} groups of hidden variables. $\boldsymbol{u}$, $\boldsymbol{c}$, and $r$.
It is worthwhile to mention that we make no further assumption about the distributions and their functional forms.\par
Specifically, we aim to find the best set of distributions and parameters that maximizes the lower bound of log likelihood of the observations \cite{jordan1999variational,bishop2006pattern}. The lower bound of log-likelihood of the observations can be written as \cite[Chapter 10]{bishop2006pattern}:
\begin{equation}
\label{eq:lower_bound}
\begin{aligned}
\ln ( \mathbb{P} \{ \mathcal{A}\} ) \geq
& \int \mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\} \ln (\mathbb{P}\{\mathcal{A},\boldsymbol{c},\boldsymbol{u},r\}) - \\
& \quad \int \mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\} \ln (\mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\})\\
& = \mathbb{E}\{ \ln (\mathbb{P}\{ \mathcal{A}, \boldsymbol{c},\boldsymbol{u},r\}) \} \\
& - \mathbb{E}\{ \ln (\mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\}) \} \\
& \triangleq \mathcal{L}(\mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\}),
\end{aligned}
\end{equation}
where the expected value is with respect to variational distribution. Hence, the problem boils down to maximizing $\mathcal{L}(\mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\})$ to find the best variational distributions. Since the lower bound is concave with respect to each of the factorized distributions, i.e., $\mathbb{Q}\{c_{n}|\hat{\beta}^1_{n},\hat{\beta}^0_{n}\}$,$
\mathbb{Q}\{u_{n}|\tau_{n}\}$,and $\mathbb{Q}\{r_{n}|\hat{b}^1,\hat{b}^0\}$, we can determine the best approximate distributions by maximizing $\mathcal{L}(\mathbb{Q}\{\boldsymbol{c},\boldsymbol{u},r\})$ with respect to one factor at a time \cite{bishop2006pattern}. Thus, at each step, the lower bound is maximized over one factor, keeping all the other distributions. This procedure is repeated until convergence. \par
For simplicity of notation, let us denote the whole set of hidden variables with $\boldsymbol{Z} = \{\{ c_{n} \},\{ u_{n} \},r\}$. In (\ref{eq:variational_dist}), $\boldsymbol{Z}$ is divided into disjoint groups $Z_i,i = 1,\dots$, where each $Z_i$ is representing one of the hidden variables in $\boldsymbol{Z}$. By maximizing the lower bound $\mathcal{L}(\mathbb{Q}\{\boldsymbol{Z} \})$, the variational distribution of each partition $\mathbb{Q}\{\boldsymbol{Z}_i\}$ is given by \cite[Chapter 10]{bishop2006pattern}:
\begin{equation}
\ln(\mathbb{Q}\{\boldsymbol{Z}_i\}) = \mathbb{E}_{j\neq i}\{ \ln( \mathbb{P} \{ \mathcal{A},\boldsymbol{Z}\} )\} + const,
\label{eq:general_update_rule}
\end{equation}
where $\mathbb{E}_{j\neq i}\{. \}$ is the expectation with respect to distributions $\mathbb{Q}\{\boldsymbol{Z}_j\}, j \neq i$. Then by plugging in $\mathbb{P} \{ \mathcal{A},\boldsymbol{Z}\} = \mathbb{P} \{ \mathcal{A},\boldsymbol{c},\boldsymbol{u},r\}$ from (\ref{eq:joint_dist}) and employing the exponential form of the distributions, the variational distributions can be obtained. The constant value is determined by normalizing the distribution.
Using (\ref{eq:general_update_rule}), we can derive closed form expressions for parameters of the variational distributions. At each time step, after receiving the new observation vector, $\boldsymbol{\alpha}^{(t)}$, the distribution of the hidden variables are updated using the derived update rules. Then, the updated distributions are used to concentrate the sensing energy on the more important coefficients of the signal. \par
\section{Measurement Matrix Design}
\label{sec:matrix_design}
In this section, the distributions of the importance levels are exploited to design the measurement matrix at each time step $\boldsymbol{\Phi}^{(t)}$. The idea is to employ the information extracted from the previous measurements and focus the sensing energy on the ROI coefficients. \par
In conventional compressive sensing methods, the sensing energy is distributed uniformly among the coefficients of the signal. In many standard methods, it is assumed that the column of the measurement matrix are scaled to be of unit norm. Thus, the total amount of sensing energy is $\| \Phi \|_F^2 = N$. In this work, we also assume that the available sensing energy is $N$. A constraint on the available sensing energy is necessary for any practical implementation. Also, without the constraint, the issue of noise would be irrelevant. \par
In adaptive sensing procedures \cite{Malloy2014Near-OptimalSensing,Braun2015Info-GreedySensing,Haupt2012SequentiallySensing,Iwen2012AdaptiveEnvironments}, no energy is allocated to the coefficients that are not likely to be in support of the signal, i.e., ROI. However, in our problem, since we are dealing with time-varying signals, such hard decisions should be avoided. \par
The key aspect of the proposed method is the allocation of sensing energy across the coefficients of the signal. In Section \ref{sec:Bayesian}, a Bayesian framework is introduced to obtain the distribution of the importance of each coefficient. Specifically, the norm of the $n^{\text{th}}$ column of the measurement matrix $\boldsymbol{\Phi^{(t)}}$ is given as:
\begin{equation}
\label{eq:col_norm}
\gamma_n^{(t)} = \sqrt{N}\frac{\bar{c}_{n}}{\eta}
\end{equation}
where $\bar{c}_{n}$ is the expected value of the importance level of the $n^{\text{th}}$ coefficient of the signal, i.e. $\bar{c}_{n} = \mathbb{E}_{\mathbb{Q}\{ c_{n}\}}\{ c_{n}\}$. and $\eta$ is a constant to ensure that the energy constraint is met. Specifically, for $\eta = \sqrt{\sum_n \bar{c}_{n}^2}$, we will have $\| \Phi \|_F^2 = N$. \par
Thus, at each time step the estimate of the signal is used to update the distribution of the hidden variables. Then, the inferred importance levels are exploited to tune the energy allocated to each coefficient of the signal.
\section{Numerical Experiments}\label{sec:results}
In this section, a series of numerical experiments are presented to highlight the performance gain of the ANCS. The primary performance metric used in our studies is time averaged normalized MSE (TNMSE), which is defined as
$
\frac{1}{T}\sum_{t = 1}^{T}\frac{\| \boldsymbol{x}^{(t)} - \hat{\boldsymbol{x}}^{(t)} \|_2^2}{\| \boldsymbol{x}^{(t)} \|_2^2}.
$
where $T$ is the number of time slots of the signal, $\| . \|_2$ is the $\ell_2$-norm of a vector, and $\hat{\boldsymbol{x}}^{(t)}$ is the estimate of $\boldsymbol{x}^{(t)}$ at time $t$. \par
The parameters of the algorithm are set as follows. Since, no prior information is assumed on the importance levels of the coefficients, the parameters are initialized as $\beta^{1}_{n} = 1 = \beta^{0}_{n} = 1, ~\forall n$. This choice of parameters results in a uniform distribution for the importance levels. To initialize $b^1$ and $b^0$, it is reasonable to assume that at least half of the measurements are reliable. In our numerical experiments, we initialized $b^1 = 3$ and $b^0 = 1$, which means on average $75\%$ of the measurements are trustworthy. The maximum number of iterations for the inference algorithm is set to $40$, with possibility of early termination if $\frac{\sum_n(\bar{c}_n^{(k)} - \bar{c}_n^{(k-1)} )^2}{\sum_n(\bar{c}_n^{(k-1)})^2} \leq 10^{-6}$ at $k^{\text{th}}$ iteration. Moreover, a window length of $W = 5$ is used. \par
In all the simulations to construct the measurement matrices, elements of the matrix were drawn from an i.i.d zero mean Gaussian distribution. For uniform sampling, the columns of the matrices are scaled to have unit norm. On the other hand, for ANCS, (\ref{eq:col_norm}) is used to realize the non-uniform distribution of energy among the columns. The total sensing energy of all the methods is assumed to be the same, i.e. $\| \boldsymbol{\Phi}^{(t)} \|_F^2 = N, \ \forall t$.\par
As a performance benchmark and to quantify the performance improvement obtained by the ANCS, we exploit the proposed method as the sampling step of an $\ell_1$ minimization recovery algorithm. Specifically, the estimate of the signal is obtained by solving an $\ell_1$ minimization problem, given by:
$$
\hat{\boldsymbol{x}}^{(t)} = \arg \min \| \boldsymbol{x} \|_1 , \text{ s.t. } \| \boldsymbol{y}^{(t)} - \boldsymbol{\Phi}^{(t)} \boldsymbol{x} \| _2 \leq c,
$$
where $\|\boldsymbol{x} \|_1=\sum_n|x_n|$ and $c$ is set to be equal to $\sigma_n\sqrt{M}$. To solve the problem, CVX \cite{Grant2008GraphPrograms,Grant2014CVX:2.1}, which is a toolbox for specifying and solving convex problems, is used.
\subsection{Performance evaluation for sparse signals in canonical basis}
\label{subsec:results_canonical}
For the first experiment, the performance gain of ANCS is quantified for the signals that are sparse in canonical basis. To model the temporal correlation, both in amplitude and support of the signal, the signal is assumed to be outcome of two random processes. Specifically, a binary vector $\boldsymbol{s}^{(t)} = [s_1^{(t)},\dots,s_N^{(t)}]^T$ describes the support of the signal at time $t$. $s_n = 1$ indicates the coefficients in the support and $s_n^{(t)} = 0$ denotes the zero coefficients. Coefficients of $\boldsymbol{s}^{(t)}$ are assumed to be independent and a Markov chain process is defined for each of the coefficients. The Markov chain processes are described by $p_{01} = \mathbb{P} \{s_n^{(t)} = 1 | s_n^{(t-1)} = 0 \}$ and $\lambda = \mathbb{P} \{s_n^{(t)} = 1\}, \ \forall n,t $. Thus, $\lambda$ is related to the sparsity level of signal. \par
Furthermore, a second process models the amplitude of the large coefficients. We employ an independent Gauss-Markov process for each of the coefficients of the signal. Amplitude of the $n^{\text{th}}$ coefficient evolves over time as:
$a_n^{(t)} = (1 - \rho)a_n^{(t-1)} +
\rho \nu_n^{(t)}.
$
Here, $\rho$ is a constant between $0$ and $1$ and controls the degree of correlation. For $\rho = 1$, the amplitude would be an uncorrelated Gaussian random process. $\nu_n^{(t)}$ is the amount of variation among two consecutive time steps and is modeled with $\mathcal{N}(0,\sigma_{L}^2)$. Thus the mean of the process is assumed to be $0$. At each time step, the coefficients of the signal are constructed as $x_n^{(t)} = a_n^{(t)}s_n^{(t)}. $\par
\begin{figure}
\centering
\subfigure[ ]{
\includegraphics[width=1.5in,angle=0]{Figures/support30.eps}\vspace{-1.5mm}
\label{subfig:supportvstime}
}
\subfigure[ ]{
\includegraphics[width=1.7in,angle=0]{Figures/importancevstime30.eps}\vspace{-1.5mm}
\label{subfig:importancevstime}
}
\caption{ \small \subref{subfig:supportvstime} Support of the signal and \subref{subfig:importancevstime} the expected value of the inferred importance levels, i.e., $\bar{c}_{n}$, for the first $30$ coefficients of the signal. $M = 60$, $N = 200$, SNR $= 20$ dB, and $W = 5$. }
\label{fig:importanceovertime}
\vspace{-5mm}
\end{figure}
The simulation parameters are set as follows, unless otherwise is stated. We assume that the signal of interest is of length $N = 200$ with sparsity level of $\lambda = \frac{K}{N} = 0.1$. The variance of noise, i.e., $\sigma_n^2$, is set to have a signal-to-noise ratio (SNR) of $20$ dB. Other model parameters are set as $\rho = 0.2$, $p_{01} = 0.02$, $\sigma_L = 10$, and $T = 30$. \par
To detect the ROI, i.e., support of the signal, after determining the estimate of the signal $\hat{\boldsymbol{x}}^{(t)}$, a simple thresholding is performed. Specifically, $\alpha_n^{(t)}$ is set to $1$, if $\hat{x}_n^{(t)} \geq 1$.
Figure \ref{fig:importanceovertime} shows the evolution of the inferred importance levels, i.e., $\bar{c}_n$, over time for $n = 1,2, \dots,30$. In other words, this figure illustrates how the sensing energy is distributed among the coefficients at each time step. As it is clear, at the first time step, $\bar{c}_n = 0.5, \forall n$, indicating unbiased estimate of importance levels when no further information is available. However, as more measurements are received, uncertainty decreases and the support of the signal is revealed. It is also worthwhile to point out that an error in the ROI detection procedure can potentially impact up to $W = 5$ time slots. Error propagation, as well as computational complexity, are the main reasons that choosing large values for $W$ should be avoided. \par
\begin{figure}
\centering
\includegraphics[width=2.6in,angle=0]{Figures/TNMSE_vsP_M60_SNR20_05.eps}\vspace{-1.5mm}
\caption{\small Performance of ANCS for different values of $p_{01}$. The total sensing energy is the same for all the methods. $N = 200$, SNR $= 20$ dB, and $T = 30$, and $M = 60$. }
\label{fig:TNMSEvsP}
\vspace{-5mm}
\end{figure}
To study the performance of ANCS for different levels of temporal correlation, Figure \ref{fig:TNMSEvsP} illustrates the TNMSE of ANCS for different values of $p_{01}$. The results are averaged over substantially large number of Monte-Carlo Trials. Here, ANCS is employed also as the sampling step of Support-aware MMSE, as well as the $\ell_1$ minimization recovery method. SA-MMSE calculates the minimum mean square error estimate of the signal when the support of the sparse signal is known. The actual support of the signal, $\sigma_L^2$, $\sigma_n^2$, and $\rho$ are provided as the inputs of the SA-MMSE algorithm. The performance of SA-MMSE is an indicator of lowest MSE achievable by a recovery algorithm.\par
For small values of $p_{01}$, the signal is nearly static over time. Thus, the method is able to detect the support accurately and the TNMSE is decreased significantly. Furthermore, since the signal is sparse in canonical basis and the support of the signal is set to be the ROI, overall recovery error is the same as the recovery error of the ROI coefficients. It is due to the fact that whole energy of the signal is concentrated in the ROI. As it can be noticed in the figure, for $p_{01} = 0$, ANCS can enhance the performance of the $\ell_1$ minimization algorithm substantially.
As $p_{01}$ increases, the support of the signal changes over time and the observations of previous time steps become less informative about the signal and the performance gain of ANCS decreases. However, for values of $p_{01} < 0.3$, nonuniform recovery of the signal is still achieved. \par
\begin{figure}
\centering
\includegraphics[width=2.6in,angle=0]{Figures/TNMSE_vsM_P02_SNR20.eps}\vspace{-1.5mm}
\caption{ \small TNMSE (in dB) for
of different recovery algorithm with and without ANCS as the sampling step for $N = 200$, SNR $= 20$ dB, and $T = 30$. }
\label{fig:ErrorvsM}
\vspace{-5mm}
\end{figure}
Figure \ref{fig:ErrorvsM} compares the performance of different recovery algorithms with uniform sampling and ANCS as the sampling step for different number of measurements. As it is clear, ANCS can decrease the TNMSE up to $7$ dB, compared to $\ell_1$ minimization recovery, and can reduce the required number of measurements. As an example, to achieve a TNMSE of $-15$ dB, ANCS employs about $50\%$ of the measurements required by uniform sampling, highlighting one of the major benefits of ANCS: for a sparse signal in canonical basis, ANCS is able to reduce the recovery error and number of required measurements substantially.\par
To highlight the performance gain achieved by ANCS in low SNR regimes, Figure \ref{fig:ErrorvsSNR} depicts TNMSE of different methods versus SNR. It is easy to notice that the performance of ANCS is very close to SA-MMSE with uniform sampling, which is MSE-optimal. This is one of the main benefits of adaptive CS. As mentioned in Section \ref{sec:intro}, it is known that adaptive CS provides the opportunity to detect and estimate signals at lower SNRs. Furthermore, performance of SA-MMSE algorithm in Figure \ref{fig:ErrorvsM} and Figure \ref{fig:ErrorvsSNR} illustrates that ANCS is able reduce the lower bound of recovery error by up to $6$ dB. \par
\begin{figure}
\centering
\includegraphics[width=2.6in,angle=0]{Figures/TNMSE_vsSNR_M60_P02.eps}\vspace{-1.5mm}
\caption{ \small Performance of $\ell_1$ minimization and SA-MMSE with and without ANCS as the sampling step in terms of TNMSE (in dB). of different methods for $N = 200$, $p_{01} = 0.02$, and $M = 60$. }
\label{fig:ErrorvsSNR}
\vspace{-5mm}
\end{figure}
\subsection{Performance evaluation for sparse signals in the DCT domain}
\label{subsec:DCT}
In this series of experiments, the performance gain achieved by ANCS is evaluated for signals that are not sparse in canonical basis, but has a sparse representation in some proper domain. In our numerical experiments, we employed DCT domain as the sparsifying basis. \par
To generate the sparse signal in DCT domain, the same procedure explained in Section \ref{subsec:results_canonical} is exploited. Specifically, let $\boldsymbol{u}^{(t)} = \boldsymbol{\Psi}\boldsymbol{x}^{(t)}$ represent the sparse representation of the signal of interest, $\boldsymbol{x}^{(t)}$, in DCT domain. $\boldsymbol{\Psi}$ denotes the DCT transform matrix. To generate a time correlated signal, elements of $\boldsymbol{u}^{(t)}$ are constructed as $u_n^{(t)} = s_n^{(t)} a_n^{(t)}$, where $s_n^{(t)}$ and $a_n^{(t)}$ are outcome of two random processes described in Section \ref{subsec:results_canonical}. To reconstruct the signal, we use $\hat{\boldsymbol{x}}^{(t)} = \boldsymbol{\Psi}^T\hat{\boldsymbol{u}}^{(t)}$, where
$$
\hat{\boldsymbol{u}}^{(t)} = \arg \min \| \boldsymbol{u} \|_1 , \text{ s.t. } \| \boldsymbol{y}^{(t)} - \boldsymbol{\Xi}^{(t)} \boldsymbol{u} \| _2 < c,
$$
and $\boldsymbol{\Xi}^{(t)} = \boldsymbol{\Phi}^{(t)}\boldsymbol{\Psi}^T$.\par
Furthermore, to model the variation of ROI over time, a new set of binary Markov processes is employed. This means that the probability of a coefficient being in the ROI is independent from its location and its value. To describe this Markov process, for simplicity, we use the same set of parameters as the random process corresponding to the support of the signal, i.e., $\lambda$ and $p_{01}$. Hence, the rates of change for support of $\boldsymbol{u}^{(t)}$ and the ROI in $\boldsymbol{x}^{(t)}$ are assumed to be the same. It is also assumed that the ROI detection algorithm may report erroneous observations to ANCS. \par
In Figure \ref{fig:DCTvs_M}, we evaluate the performance of ANCS versus the number of measurements $M$ for fault rate of $10\%$. This experiment also shows that the proposed ANCS is able to decrease the error of ROI coefficients up to $3$-$4$ dB for different number of measurements. This benefit comes at the cost of losing performance on total recovery error. Interestingly, for smaller values of $M$, this benefit comes at almost no cost and without losing any performance for non-ROI entries. \par
\begin{figure}
\centering
\includegraphics[width=2.6in,angle=0]{Figures/DCT_TNMSE_vsM_P02_SNR20.eps}\vspace{-1.5mm}
\caption{\small TNMSE (in dB) versus $M$ of ANCS. $N = 200$, SNR $= 20$ dB, and $T = 30$, and $M = 60$. }
\label{fig:DCTvs_M}
\vspace{-5mm}
\end{figure}
Finally, as it was expected, Figure \ref{fig:DCT_TNMSEvsMICmodel} illustrates that as ANCS receives more faulty data from the ROI detection algorithm, its performance becomes more similar to conventional CS with uniform sampling. This is because the faulty data prevents the inference algorithm from gaining certainty on the location of ROI coefficients. However, even for fault rates of as much as $50\%$, non-uniform recovery of the signal is achieved. \par
\begin{figure}
\centering
\includegraphics[width=2.6in,angle=0]{Figures/DCT_TNMSE_vsFault_P02_M60_SNR20.eps}\vspace{-1.5mm}
\caption{\small Plot of TNMSE (in dB) of ANCS for different values of fault rate.
$N = 200$, SNR $= 20$ dB, and $T = 30$, and $M = 60$. }
\label{fig:DCT_TNMSEvsMICmodel}
\vspace{-3mm}
\end{figure}
\section{conclusions}\label{sec:conclusions}
This work presented \emph{adaptive non-uniform compressive sampling} (ANCS) for time-varying sparse signals.
The main idea is to employ the observations of previous time slots to infer the region of interest (ROI) in the signal and concentrate the sensing energy on the corresponding coefficients.
For that, we presented a Bayesian framework, by modeling the overall and coefficient-specific reliability of the ROI detection algorithm. \par
The results show that the proposed framework is able to achieve the desired non-uniform recovery and can decrease the error in ROI significantly for signals that are sparse or have a sparse representation in a proper basis. The results also illustrated that the proposed method is particularly advantageous for signals that are sparse in canonical basis. For such signals, ANCS results in substantial improvement in accuracy of estimation. \par
\section{Acknowledgments}
This material is based upon work supported by the National Science Foundation under Grant No. ECCS-1418710.
\footnotesize{
\balance
\bibliography {Mendeley}
\bibliographystyle{ieeetr}
}
\end{document}
|
\section{Introduction}
Spontaneous symmetry breaking is a ubiquitous phenomenon in physics.
Its wide range of applications includes,
for instance, the Higgs mechanism in particle physics,
which allows to give mass to the particles of the Standard Model,
or the phase transitions of ferromagnets in solid state physics.
Here we consider this phenomenon in gravity.
In particular, we consider an Ellis wormhole
in General Relativity
\cite{Ellis:1973yv,Ellis:1979bh,Bronnikov:1973fh,Morris:1988cz,Morris:1988tu,ArmendarizPicon:2002km,Sushkov:2005kj,Lobo:2005us}.
Such a wormhole
connects two asymptotically flat spacetimes by a throat.
In order to allow for the non-trivial topology in General Relativity
a phantom field is included, i.e., a real scalar field with a
reversed sign in front of its kinetic term in the action.
When symmetric boundary conditions are specified,
the Ellis wormhole is reflection symmetric with respect to its throat.
In suitable coordinates, the wormhole metric in $D=4$ spacetime dimensions reads
\begin{equation}
ds^2 = - dt^2 + d \eta^2 + \left(\eta^2 + \eta_0^2 \right) d \Omega^2 \ .
\label{ellis_sym}
\end{equation}
Then the wormhole throat is located at $\eta=0$,
and the throat parameter $\eta_0$ characterizes the size of the throat.
Of course, by choosing asymmetric boundary conditions also
asymmetric wormhole solutions can be created,
where the reflection symmetry $\eta \to - \eta$ no longer holds.
However, here we are not interested in such an enforced symmetric breaking.
Instead we would like to demonstrate that asymmetric wormholes can appear
also for symmetric boundary conditions, when matter is included.
Thus the symmetry breaking happens spontaneously, and energy considerations
will tell us, that the asymmetric solutions are energetically preferred.
Wormholes immersed in matter have been studied before with
various types of matter. Examples include nuclear matter
\cite{Dzhunushaliev:2011xx,Dzhunushaliev:2012ke,Charalampidis:2013ixa}
and bosonic matter \cite{Dzhunushaliev:2014bya}.
However, most solutions studied so far were symmetric,
and the asymmetric solutions were asymmetric by construction,
because of an asymmetric choice of boundary conditions.
Thus - to our knowledge - for wormhole solutions immersed in matter
spontaneous symmetry breaking has not been observed before.
Recently, the static Ellis wormhole solution has
been generalized to spacetimes with higher dimensions
\cite{Torii:2013xba,Dzhunushaliev:2013jja}.
Moreover, rotating generalizations of the Ellis wormhole
in four and five dimensions
have been found \cite{Dzhunushaliev:2013jja,Kleihaus:2014dla,Chew:2016epf}.
We will not address the question of rotation here.
Instead we will consider non-rotating wormhole solutions in $D$ dimensions
immersed in bosonic matter, with a focus on five dimensions.
The bosonic matter consists of a complex scalar field with a
self-interaction such that it allows for localized solutions
already in a flat spacetime background, which correspond to
non-topological solitons \cite{Friedberg:1976me,Friedberg:1986tq} or $Q$-balls
\cite{Coleman:1985ki}.
When gravity is coupled to this bosonic matter boson stars arise
\cite{Friedberg:1986tq,Lee:1991ax}, which can get very close
to the black hole limit.
In \cite{Dzhunushaliev:2014bya} in addition a phantom field was included,
to obtain solutions with a non-trivial topology, i.e.,
$Q$-balls and boson stars harbouring wormholes at their core.
However, all these solutions were reflection symmetric.
Here we study the emergence of the asymmetric solutions.
Let us denote with ${\cal M}_+$ the part of the manifold
with positive radial coordinate $\eta$, and with ${\cal M}_-$
the part with negative $\eta$. Since we are dealing with
configurations with two asymptotically flat regions,
we obtain two values for the mass, $M_\pm$,
as measured asymptotically in ${\cal M}_\pm$.
These values can be read off the asymptotic behaviour
of the metric. For symmetric solutions both masses agree. However,
for asymmetric solutions the two masses can differ widely.
We note, that the asymmetric solutions always come in pairs,
since they are transformed into each other
by the reflection transformation $\eta \to -\eta$.
Consequently, their two masses are simply interchanged.
To obtain the particle number
in the case of $Q$-balls and boson stars,
volume integrals are performed.
For the solutions with
non-trivial topology one can proceed
analogously, when the solutions are symmetric. For the
asymmetric solutions, however, the calculation of the
particle number via such integrals can become ambiguous,
since the inner boundary is not provided by symmetry.
Therefore we here propose an unambiguous procedure
to obtain the particle number.
This is mandatory,
since we need the particle number in order to demonstrate that
the asymmetric solutions are energetically favourable.
The presence of the matter in the wormhole spacetime has further notable
consequences. In particular, the backreaction of the matter
on the geometry can cause a drastic change of the
geometry, giving rise to a transition
from configurations with a single throat to
configurations with a double throat and an equator in between.
Such a transition is known to occur for symmetric wormholes
with bosonic matter \cite{Dzhunushaliev:2014bya}
and for other types of matter as well
\cite{Hauser:2013jea,Dzhunushaliev:2014mza}.
This paper is organized as follows:
In section II we present the theoretical setting
for obtaining both symmetric and asymmetric
configurations of wormholes immersed in bosonic matter
in $D$ dimensions.
The numerical results are shown in section III,
starting with the probe limit,
where the spontaneous symmetry breaking is already observed.
We then discuss in detail the five-dimensional families
of symmetric and asymmetric solutions in the presence of gravity.
Finally, we include a brief discussion
of the asymmetric configurations in four dimensions,
and show that the spontaneous symmetry breaking is present as well.
We conclude in section V.
In appendix A we briefly address our method of extracting
the mass and the particle number for asymmetric solutions.
\section{Theoretical setting}
\subsection{Action}
We consider General Relativity with a minimally coupled complex scalar field $\Phi$
and phantom field $\Psi$ in $D$ spacetime dimensions.
Besides the Einstein-Hilbert action
with curvature scalar $\cal R$, coupling constant $\kappa$
and metric determinant $g$, the action
\begin{equation}
S=\int \left[ \frac{1}{2 \kappa}{\cal R} +
{\cal L}_{\rm ph} +{\cal L}_{\rm bs} \right] \sqrt{-g}\ d^Dx
\label{action}
\end{equation}
then contains the respective matter contributions,
the Lagrangian ${\cal L}_{\rm ph}$ of the phantom field $\Psi$,
and the Lagrangian ${\cal L}_{\rm bs}$ of the complex scalar field $\Phi$.
The kinetic term for the phantom field carries the reverse sign
\begin{equation}
{\cal L}_{\rm ph} = \frac{1}{2}\partial_\mu \Psi\partial^\mu \Psi \ ,
\label{lpsi}
\end{equation}
as compared to the kinetic term of the complex scalar field
\begin{equation}
{\cal L}_{\rm bs} =
-\frac{1}{2} g^{\mu\nu}\left( \partial_\mu\Phi^* \partial_\nu\Phi
+ \partial_\nu\Phi^* \partial_\mu\Phi
\right) - U( \left| \Phi \right|) \ .
\label{lphi}
\end{equation}
The asterisk denotes complex conjugation,
and $U$ denotes the potential with the mass term and the self-interaction
\begin{equation}
U(|\Phi|) = \lambda |\Phi|^2 \left( |\Phi|^4 -c|\Phi|^2 +b \right)
\ . \label{U}
\end{equation}
The global minimum of the potential resides at $\Phi =0$, where $U(0)=0$,
while a local minimum is found at some finite value of $|\Phi|$.
The mass of the bosons $m_b=\sqrt{\lambda b}$ is specified by the
quadratic term.
The potential is chosen such that it allows for non-topological soliton solutions
\cite{Friedberg:1976me,Friedberg:1986tq} or $Q$-balls
\cite{Coleman:1985ki}.
Variation of the action with respect to the metric
leads to the Einstein equations
\begin{equation}
G_{\mu\nu}= {\cal R}_{\mu\nu}-\frac{1}{2}g_{\mu\nu}{\cal R} = \kappa T_{\mu\nu}
\label{ee}
\end{equation}
with stress-energy tensor
\begin{equation}
T_{\mu\nu} = g_{\mu\nu}{{\cal L}}_M
-2 \frac{\partial {{\cal L}}_M}{\partial g^{\mu\nu}} \ ,
\label{tmunu}
\end{equation}
where we denoted by ${\cal L}_{\rm M} =
{\cal L}_{\rm ph}+{\cal L}_{\rm bs} $ the sum of the scalar field Lagrangians.
\subsection{Ans\"atze}
For the line element
of the spherically symmetric solutions
with a non-trivial topology we choose
\begin{equation}
ds^2 = -e^{(D-3)a} dt^2
+p e^{-a}\left[d\eta^2 + \left(\eta^2 + \eta_0^2\right)\Omega^2_{D-2}\right] \ .
\label{lineel}
\end{equation}
Here $d\Omega^2_{D-2}$
denotes the metric on the unit $(D-2)$-sphere, while $a$ and $p$ are functions of
the radial coordinate $\eta$,
which takes positive and negative
values, i.e. $-\infty< \eta < \infty$.
The two limits $\eta\to \pm\infty$
correspond to two distinct asymptotically flat regions,
associated with ${\cal M}_+$ and ${\cal M}_-$, respectively.
Note that in Eq.~(\ref{lineel}) we have introduced the parameter $\eta_0$,
which we will refer to as the throat parameter.
As for spherically symmetric $Q$-balls and boson stars,
we parametrize the complex scalar field $\Phi$ via
\begin{equation}
\Phi
= \phi (\eta) ~ e^{ i\omega_s t } \ , \label{ansatzp}
\end{equation}
where $\phi (\eta)$ is a real function,
and $\omega_s$ denotes the boson frequency.
The phantom field $\Psi$ depends only on the radial coordinate,
\begin{equation}
\Psi
= \psi (\eta) \ .
\label{ansatzph}
\end{equation}
\subsection{Einstein and Matter Field Equations}
Substituting the above Ans\"atze into the Einstein equations
$G_\mu^\nu=\kappa T_\mu^\nu$ leads to the following field equations
\begin{eqnarray}
-\frac{D-2}{2p}e^a (a''-p''/p)
+\frac{(D-2)(D-3)}{8p}e^a (a'^2-2 a' p'/p) & & \nonumber\\
+\frac{(D-2)(D-7)}{8p^3}e^a p'^2
-\frac{(D-2)^2}{2ph}e^a \eta ( a'-p'/p) & & \nonumber\\
-\frac{(D-2)(D-5)}{2 p h^2}e^a \eta_0^2
& = &
-\kappa \left(U(\phi) +\frac{e^a}{2p}(2\phi'^2-\psi'^2) +\omega_s^2 e^{-(D-3)a}\phi^2\right)
\label{eeq_00}
\end{eqnarray}
\begin{eqnarray}
-\frac{(D-2)(D-3)}{8p}e^a a'^2
+\frac{(D-2)(D-3)}{8p^3}e^a p'^2 & & \nonumber\\
+\frac{(D-2)(D-3)}{2p^2 h}e^a \eta p'
-\frac{(D-2)(D-3)}{2 p h^2}e^a \eta_0^2
& = &
-\kappa \left(U(\phi) -\frac{e^a}{2p}(2\phi'^2-\psi'^2) -\omega_s^2 e^{-(D-3)a}\phi^2\right)
\label{eeq_rr}
\end{eqnarray}
\begin{eqnarray}
\frac{D-3}{2p^2}e^a p''
+\frac{(D-2)(D-3)}{8p}e^a a'^2& & \nonumber\\
+\frac{(D-3)(D-8)}{8p^3}e^a p'^2
+\frac{(D-3)^2}{2p^2 h}e^a \eta p'& & \nonumber\\
-\frac{(D-3)(D-6)}{2 p h^2}e^a \eta_0^2
& = &
-\kappa \left(U(\phi) +\frac{e^a}{2p}(2\phi'^2-\psi'^2) -\omega_s^2 e^{-(D-3)a}\phi^2\right)
\label{eeq_tt}
\end{eqnarray}
which derive from the $tt$, $\eta\eta$ and $\theta\theta$ components, respectively.
For convenience we use the abbreviation $h=\eta^2+\eta_0^2$.
Variation of the action with respect to the
complex scalar field and to the phantom field
leads to the equations
\begin{eqnarray}
\phi'' +\left(\frac{D-3}{2}\frac{p'}{p}+(D-2)\frac{\eta}{h}\right) \phi'
& = &
\frac{1}{2} p e^{-a} \frac{dU}{d\phi}-\omega_s^2 p e^{-(D-2)a}\phi \ ,
\label{eqSk}\\
\left(\left(ph\right)^{\frac{D-2}{2}}\frac{1}{\sqrt{p}}\psi'\right)' & = & 0 \ ,
\label{eqPh}
\end{eqnarray}
where integration of the last equation leads to
\begin{equation}
\psi' = \sqrt{p}\left(ph\right)^{-\frac{D-2}{2}}{\cal D}\ .
\label{phip}
\end{equation}
Here the constant $\cal D$ denotes the scalar charge of the phantom field.
By inserting Eq.~(\ref{phip}) into Eq.~(\ref{eeq_rr})
the scalar charge $\cal D$
can be expressed via
\begin{eqnarray}
{\cal D}^2
= \left(ph\right)^{D-2}\left[
\frac{(D-2)(D-3)}{4p}
\left(a'^2 - p'^2/p^2-4\frac{\eta}{h} p'/p +4\frac{\eta_0^2}{h^2}\right)
-2\kappa e^{-a}\left(U(\phi) -\frac{e^a}{p}\phi'^2 -\omega_s^2 \phi^2 e^{-(D-3)a}\right)
\right] \ .
\label{eqD2}
\end{eqnarray}
We now eliminate the $\phi'^2$ term
and the $\psi'^2$ term in the Einstein equations
by adding Eq.~(\ref{eeq_rr}) to Eq.~(\ref{eeq_00})
and to Eq.~(\ref{eeq_tt}).
This provides us with
the final set of Einstein equations
\begin{eqnarray}
a''
+\frac{D-2}{h}a'
-\frac{D-3}{2p} \eta a' p'
& = & -4\kappa \frac{p e^{-a}}{(D-2)(D-3)}\left(U(\phi)-\omega_s^2(D-2) \phi^2 e^{-(D-3)a}\right)
\label{eqa}\\
p''+
\frac{(2 D-5)}{h}\eta p'
+\frac{D-5}{2p} p'^2
-2\frac{(D-4)\eta_0^2}{h^2}p
& = & -4\kappa \frac{p^2 e^{-a}}{D-3}\left(U(\phi)-\omega_s^2 \phi^2 e^{-(D-3)a}\right)
\label{eqp}
\end{eqnarray}
to be solved
together with the equation for the bosonic matter, Eq.~(\ref{eqSk}).
Clearly, the system of equations allows for reflection symmetric
solutions ${\cal S}$, i.e., solutions whose functions are either symmetric
or antisymmetric under $\eta \to - \eta$.
\subsection{Boundary Conditions}
We need to solve the above set of three coupled ordinary differential equations
of second order. Thus we have to impose six boundary conditions.
Here we would like to impose symmetric boundary conditions, which are the same
for $\eta \to \infty$ and $\eta \to -\infty$. Therefore any asymmetry in the
solutions is not enforced via boundary conditions but arises spontaneously.
The boundary conditions for the boson field function $\phi(\eta)$
are then given by
\begin{equation}
\phi(\eta\to \pm \infty) \to 0 \ .
\label{bcasym}
\end{equation}
These conditions ensure, that the configurations possess finite energy.
For the metric we require asymptotic flatness in both asymptotic regions.
By imposing on the metric function $a$ the conditions
\begin{equation}
a(\eta\to \pm \infty) \to 0 \
\label{bcbsym}
\end{equation}
we also set the time scale.
For the metric function $p$ asymptotic flatness in both asymptotic regions
implies
\begin{equation}
p(\eta\to \pm \infty) \to 1 \ .
\label{bccsym}
\end{equation}
\subsection{Single and multiple throats}
Here we discuss the throat properties in terms of the
circumferential function $R(\eta)= \sqrt{p h}e^{-a/2}$, which represents
the radius of a circle in the equatorial plane with constant coordinate
$\eta$.
Let us first consider symmetric solutions ${\cal S}$.
In symmetric solutions, the metric functions are symmetric under
$\eta \to - \eta$.
Consequently, $\eta=0$ plays a special role.
The metric functions $a$ and $p$ possess a vanishing derivative at $\eta=0$,
$a'(0)=p'(0)=0$, which implies that they assume extremal values.
In particular, from the circumferential radius $R$ we can conclude,
that if $R(0)$ is a minimum, a throat is located at $\eta=0$,
since $\eta=0$ then corresponds to a minimal surface.
If $R(0)$ is a maximum, it represents only a local maximum,
since $R(\eta) \to |\eta|$ asymptotically. Such a local maximum
corresponds to a maximal surface and thus an equator.
Clearly, between the equator and asymptotic infinity (at least) one
throat should be localized in each of the regions ${\cal M}_+$ and ${\cal M}_-$.
In that case double (or multiple) throat configurations are present.
Transitions between single and double throat configurations
in four spacetime dimensions were observed for symmetric solutions in
\cite{Dzhunushaliev:2014bya}.
These transitions occur, when the second and first derivative of $R$ vanish,
\begin{equation}
R''(0) = 0 \ \ \Longleftrightarrow \ \ \
\left[ (D-2)(D-3)e^a - 2 p \eta_0^2 \kappa U(\phi) \right]_{\eta=0} =0 \ .
\end{equation}
We note that for symmetric solutions the surface gravity vanishes at
the throat for configurations with a single
throat, since the surface gravity is given by
$\frac{D-3}{2}\left[ e^{(D-2)a/2} a'/\sqrt{p}\right]_{\eta_{\rm th}}$,
and $a'(0)=0$
\cite{Dzhunushaliev:2014bya}.
In contrast to symmetric solutions, where a single throat
must be located at $\eta=0$, a throat of asymmetric solutions
can be located anywhere. Moreover, the transition from
single to double throat solutions can happen, when the
circumferential radius develops a turning point
at some value of the radial coordinate $\eta_{0\, \rm cr}$, where
$R'(\eta_{0\, \rm cr})=R''(\eta_{0\, \rm cr})=0$.
Thus for asymmetric solutions the single throat does not degenerate
to form an equator and a double throat, but an equator together with
a second throat appear spontaneously somewhere else
in the manifold.
\subsection{Energy conditions}
When the null energy condition (NEC) is violated,
also the weak and the strong energy condition are violated.
It is therefore sufficient to address only the NEC,
which requires
\begin{equation}
\Xi = T_{\mu\nu} k^\mu k^\nu \ge 0 \
\end{equation}
for all (future-pointing) null vector fields $k^\mu$.
This condition can be expressed via the Einstein tensor
by making use of the Einstein equations.
For spherically symmetric solutions we then obtain
the new conditions
\begin{equation}
-G_t^t+G_\eta^\eta \geq 0 \ , \ \ \ {\rm and }
\ \ \ -G_t^t+G_\theta^\theta\geq 0 \ ,
\label{Nulleng}
\end{equation}
both of which must be obeyed everywhere in order to respect
the NEC.
For the present set of solution the NEC is always violated.
\subsection{Mass and scalar charge}
In the presence of gravity, the mass of a
stationary asymptotically flat solution in $D$ dimensions
can be obtained from the Komar integral \cite{Wald:1984}
\begin{equation}
M = - \frac{1}{16 \pi G_D} \frac{D-2}{D-3} \int_{S_{\infty}^{D-2}} \alpha \ ,
\label{komarM}
\end{equation}
with $\alpha_{\mu_1 \dots \mu_{D-2}} \equiv \epsilon_{\mu_1 \dots \mu_{D-2}
\rho \sigma} \nabla^\rho \xi^\sigma$ and $\xi \equiv \partial_t$.
By choosing the surface $S$ at spatial infinity in ${\cal M}_+$ (${\cal M}_-$)
we obtain the mass $M_+$ ($M_-$).
Both values of the mass are encoded in the metric function $g_{tt}$
and can easily be extracted. We obtain $M_+$ from
\begin{equation}
g_{tt} \underset{\eta \to \infty}{\longrightarrow} - 1 + \frac{\mu}{\eta^{D-3}} \ , \ \ \
M_+= \frac{(D-2) \Omega_{D-2}}{16 \pi G} \mu
\ , \label{komarM2}
\end{equation}
where $\Omega_{D-2}$ is the area of the unit $D-2$-sphere,
and $M_-$ analogously. When we use Stokes' theorem to convert the integral
into a volume integral, we obtain a boundary term at some $\eta_b$, e.g.,
\begin{equation}
{M_+} =
\frac{1}{{4\pi}} \int_{\Sigma}
R_{\mu\nu}n^\mu\xi^\nu dV
- \frac{1}{16 \pi G_D} \frac{D-2}{D-3} \int_{S_{\eta_b}^{D-2}} \alpha \
= \frac{1}{{4\pi}} \int_{\Sigma}
R_{\mu\nu}n^\mu\xi^\nu dV + M_{\eta_b}
\ . \label{komarM1}
\end{equation}
Here $\Sigma$ denotes an asymptotically flat spacelike hypersurface,
$n^\mu$ is normal to $\Sigma$ with $n_\mu n^\mu = -1$,
and $dV$ is the natural volume element on $\Sigma$
\cite{Wald:1984}.
Obviously, the volume integral only agrees
with the mass $M_+$ when the surface term $M_{\eta_b}$ vanishes.
Since the surface term is proportional to the product of the surface gravity and the
surface area, this is the case when the surface gravity vanishes.
In the symmetric case, the surface gravity always vanishes at $\eta=0$,
i.e., at the throat (or at the equator).
Then the volume integral may be evaluated to give the mass.
However, in the asymmetric case, the surface gravity is finite at the
throat (and at the equator) \cite{Dzhunushaliev:2014bya}.
In that case the surface term would have to be included to obtain
the proper value of the mass from the volume integral.
A simple evaluation of the volume integral from say $\eta=0$ to
$\eta=\infty$ alone would not yield the correct value for the mass
in the asymmetric case. These considerations are important
for evaluating the mass of the asymmetric configurations in the
probe limit. Therefore we present an appropriate procedure for extracting the
mass in the probe limit in Appendix A.
\subsection{Charge or particle number}
The Lagrange density is invariant under the global phase transformation
\begin{equation}
\displaystyle
\Phi \rightarrow \Phi e^{i\chi} \ .
\end{equation}
This leads to the conserved current
\begin{eqnarray}
j^{\mu} & = & - i \left( \Phi^* \partial^{\mu} \Phi
- \Phi \partial^{\mu}\Phi ^* \right) \ , \ \ \
j^{\mu} _{\ ; \, \mu} = 0 \ .
\label{current}
\end{eqnarray}
In globally regular topologically trivial spacetimes
the associated conserved charge $Q$ is then obtained
by integrating the time-component
of the current over the entire space
\begin{eqnarray}
Q &= &- \int j^t \left| g \right|^{1/2} d\eta d\Omega_{D-2}
\nonumber \\
&=& 2 \Omega_{D-2} \omega \int_0^{\infty} |g| ^{1/2} \frac{\phi^2}{A^2} \,d\eta \, \ .
\label{Qc}
\end{eqnarray}
The global charge $Q$ then corresponds to the particle number
of the complex boson field.
In wormhole spacetimes, we have to reconsider the above
definition of the charge or particle number.
For symmetric configurations with a single throat it is clear that we should integrate
from $\eta=0$ to $\pm \infty$, to obtain the particle numbers $Q_\pm$
for the regions ${\cal M}_\pm$. This should remain true, when
the throat at $\eta=0$ turns into an equator \cite{Dzhunushaliev:2014bya}.
However, in the asymmetric case it is no longer obvious, where the inner
integration boundary $\eta_b$ should reside.
If we were to retain $\eta_b$ at the throat, for instance,
an ambiguity would arise
at the very least when the second throat and the equator emerge.
To solve this question, we reconsider the case of the mass.
In the presence of gravity,
the mass is obtained unambiguously and in a simple way,
when we use surface integrals at plus and minus infinity.
If we were to use surface integrals for the charge or particle number as well,
we would also have a clean definition of the charges $Q_\pm$
to be associated with the masses $M_\pm$.
To achieve this feat we employ the following trick.
We minimally couple the complex scalar field to a
fictitious U(1) gauge field, which is not allowed to backreact
on the configuration.
Thus the configuration is not changed, while we can employ
the Gauss law to read off the charge $Q_\pm$ of the configurations
in the asymptotically flat regions
\begin{equation}
Q_\pm = \int_{S_{\pm \infty} } {^*F} .
\label{chargeNew}
\end{equation}
This definition then yields the same charge as the above volume integral in the symmetric case,
while it leads to meaningful and unambiguous values in the asymmetric case,
which can be compared with the respective masses
to extract the binding energies of these configurations.
\section{Wormholes immersed in bosonic matter}
When solving the field equations subject to the given set
of symmetric boundary conditions, we were in for a surprise,
since the numerical procedure led to asymmetric solutions
in addition to the expected symmetric solutions.
Our further studies then revealed the phenomenon of
spontaneous symmetry breaking in wormhole spacetimes
with bosonic matter. The demonstration of this
phenomenon is the main focus of this section.
While we have presented the formalism in the last section
for $D$ spacetime dimensions,
we have performed the numerical calculations
mainly in five dimensions,
although we have also performed a study in four dimensions to
convince ourselves that the analogous
features are present.
In this section we first address the probe limit,
where the boson field equation is solved in the
background of the Ellis wormhole.
Already here the spontaneous symmetry breaking
is observed.
Depending on the throat size,
the asymmetric solutions (${\cal A}_\pm$) can bifurcate
from the symmetric solutions (${\cal S}$)
at critical values of the boson frequency.
The asymmetric solutions are always energetically favoured.
Subsequently, we couple to gravity, and thus take
the backreaction of the boson field into account.
The phenomenon of spontaneous symmetry breaking
in then retained.
However, the structure of the families of solutions
becomes much richer. In particular, both
symmetric and asymmetric solutions exhibit transitions
to double throat wormhole configurations.
Finally, we address this phenomenon in four dimensions,
were contact to astrophysics can in principle be made.
Here the symmetric solutions were studied in detail
before, but the asymmetric ones had not been seen
because the calculations had been restricted to ${\cal M}_+$
\cite{Dzhunushaliev:2014bya}.
In the following we focus on the fundamental solutions.
We refer to solutions as fundamental
when they do not possess a radially excited boson field,
i.e., when their boson field function does not have nodes.
However, we have observed that the
radially excited solutions exhibit the analogous pattern
of symmetric and asymmetric solutions.
For the numerical calculations we have introduced
the compactified radial coordinate
\begin{equation}
x= {\rm atan}\left(\frac{\eta}{r_0}\right) \ ,
\end{equation}
where $r_0$ is some constant, for which we chose $r_0=3$.
We have then employed a collocation method for boundary-value ordinary
differential equations, equipped with an adaptive mesh selection procedure
\cite{COLSYS}.
We have used typical mesh sizes with $10^3-10^4$ points,
reaching a relative accuracy of $10^{-10}$ for the functions.
Estimates of the relative errors for the mass and the angular momentum
have been of order $10^{-6}$.
We have employed the condition ${\cal D}=\mbox{const}$, Eq.~(\ref{eqD2}),
to monitor the quality of the numerical solutions.
The variation of ${\cal D}$ has been typically less than $10^{-9}$.
For the self-interaction potential $U(\phi)$ we have chosen
the parameters $\lambda=1$, $c=2$ and $b=1.1$,
since this allowed us to compare with previous
calculations of bosonic configurations without a wormhole
\cite{Hartmann:2010pm}.
With these potential parameters fixed, the further parameters were
the parameter $\kappa$, which includes the
gravitational coupling strength,
the boson frequency $\omega_s$,
and the throat parameter $\eta_0$.
In most of the results shown, we have fixed the throat parameter $\eta_0=3$,
leaving only $\kappa$ and $\omega_s$ as free parameters.
\subsection{Probe Limit}
We here demonstrate, that the phenomenon of
spontaneous symmetry breaking occurs already
in the probe limit.
We begin by illustrating the symmetric and asymmetric solutions,
and then discuss their dependence on the two parameters,
the boson frequency $\omega_s$ and the
throat size $\eta_0$. We point out that depending on the throat size
bifurcations occur.
Subsequently, we consider the
masses and particle numbers of the configurations
and show that the asymmetric solutions
are more strongly bound.
\subsubsection{Parameters}
\begin{figure}[t!]
\begin{center}
\vspace{0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig1a.eps}
\label{fig1a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig1b.eps}
\label{fig1b}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig1c.eps}
\label{fig1c}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig1d.eps}
\label{fig1d}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig1e.eps}
\label{fig1e}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig1f.eps}
\label{fig1f}
}
}
\end{center}
\vspace{-0.7cm}
\caption{Probe limit:
(a) The boson field function $\phi$ versus the
compactified coordinate $x={\rm atan}\, (\eta/r_0)$
for a fixed throat parameter $\eta_0=3.0$
and decreasing values of the boson frequency
$\omega_s$.
(b) Same as (a) for asymmetric solutions.
(c) The boson field function $\phi$ versus the
compactified coordinate $x={\rm atan}\, (\eta/r_0)$
for asymmetric solutions at a fixed boson frequency
$\omega_s=0.6$ and decreasing values of the
throat parameter $\eta_0$.
(d) The value $\phi_0$ of the boson field functions $\phi$
at the throat $\eta=0$ versus the boson frequency
$\omega_s$ for decreasing values of the
throat parameter $\eta_0$.
(e) The mass $M$ (dotted) of the symmetric solutions
and the mass $M_+$ (solid) and $M_-$ (dashed)
of the asymmetric solutions versus the boson frequency
$\omega_s$ for decreasing values of the
throat parameter $\eta_0$.
(f) Same as (e) for the particle number.
\label{fig1}
}
\end{figure}
Usually the probe limit is obtained by simply taking the coupling $\kappa$
of gravity and matter to zero. However, since we want to
retain the non-trivial topology,
we must be careful here as to take only the coupling to the complex
boson field to zero, while retaining the coupling to the phantom field.
This can be achieved by an appropriate scaling of the phantom field,
$\Psi \to \Psi/\sqrt{\kappa}$.
The Einstein equations then yield the Ellis wormhole
\cite{Torii:2013xba,Dzhunushaliev:2013jja},
and only the boson field equation needs to be solved
in the background of the Ellis wormhole.
The solutions depend on
the boson frequency $\omega_s$ and on the throat parameter $\eta_0$.
In the limit $\eta_0 \to 0$, contact must be made with
topologically trivial solutions, since the two asymptotically flat
regions are then separated.
Indeed, for $\eta_0 \to 0$, the so called
non-topological soliton or $Q$-ball limit is reached
\cite{Friedberg:1976me,Coleman:1985ki}.
However, it can either be reached in both regions
${\cal M}_+$ and ${\cal M}_-$ at the same time (symmetric case),
or only in a single region (asymmetric case), as we show below.
The domain of existence of
non-rotating $Q$-balls is restricted to a certain frequency range,
$\omega_{\rm min} < \omega_s < \omega_{\rm max}$
\cite{Friedberg:1976me,Coleman:1985ki,Volkov:2002aj}.
The maximal frequency $\omega_{\rm max}$
\begin{equation}
\label{cond1}
\omega^2_{\rm max} =
\frac{1}{2} U''(0) = \lambda \, b = m_b^2 \ ,
\end{equation}
ensures that the solutions possess an exponential fall-off
at spatial infinity.
The minimal frequency $\omega_{\rm min}$ is based on
an argument to allow for localized solutions
\cite{Friedberg:1976me,Coleman:1985ki,Volkov:2002aj,Hartmann:2010pm}
and given by
\begin{equation}
\label{cond2}
\omega^2_{\rm min} =
\min_{\phi} \left[{U(\phi)}/{\phi^2} \right] \; = \;
\lambda \left(b- \frac{c^2}{4} \right) \ .
\end{equation}
These limits are retained for the topologically non-trivial solutions,
where the Minkowski background is replaced by an Ellis background
\cite{Dzhunushaliev:2014bya}.
\subsubsection{Bifurcations}
In Fig.~\ref{fig1} we show features of the symmetric solutions in the probe limit which are analogous to those
known in four dimensions \cite{Dzhunushaliev:2014bya}.
Here the boson function $\phi$ is symmetric with respect to
the reflection $\eta \to - \eta$, as seen in Fig.~\ref{fig1a},
where the throat parameter $\eta_0$ is fixed while
the boson frequency $\omega_s$ is varied.
Note, that as the boson frequency $\omega_s$ tends to $\omega_{\rm min}$,
the boson function $\phi$ tends to a constant in an increasingly
large region.
We show the boson function $\phi$ of the asymmetric solutions
for the same parameters in Fig.~\ref{fig1b}.
Here the reflection symmetry of the solutions is broken.
However, since the field equation is reflection symmetric,
all asymmetric solutions come in pairs.
To distinguish the two solutions of a pair, let us include an index $\pm$
associated with the two regions ${\cal M}_\pm$.
The index $+$ ($-$) then indicates that more matter is localized
in the region ${\cal M}_+$ (${\cal M}_-$).
The functions $\phi_-$ are obtained from
the functions $\phi_+$ by reflection ($\eta \to - \eta$)
\begin{equation}
\label{reflection} \phi_-(\eta) = \phi_+(-\eta) \ ,
\end{equation}
and vice versa.
Let us now vary the throat parameter $\eta_0$ while keeping
the frequency $\omega_s$ fixed.
In the symmetric case, when $\eta_0 \to 0$,
a {\sl double}
$Q$-ball solution is approached, where
a $Q$-ball is localized in each of the regions
${\cal M}_+$ and ${\cal M}_-$, which
simply represent two disjunct
Minkowski spacetimes in the limit.
In contrast, in the asymmetric case for fixed frequency
and $\eta_0 \to 0$,
the solutions approach a {\sl single}
$Q$-ball, that is
localized in one of the regions, either
${\cal M}_+$ for $\phi_+$ or ${\cal M}_-$ for $\phi_-$, while the
complementary region becomes completely empty in the limit.
This is demonstrated in Fig.~\ref{fig1c}.
To monitor the different solutions, it is useful to keep track of
the value of the boson function at the throat, $\phi_0$.
This parameter then maps out the domain of existence of
the symmetric and asymmetric solutions
concerning the dependence on the frequency $\omega_s$
and on the throat parameter $\eta_0$.
We note, that both asymmetric solutions possess the same
value of $\eta_0$.
To illustrate this dependence,
we exhibit in Fig.~\ref{fig1d} the value of $\phi_0$ versus the
boson frequency $\omega_s$ for four values of the
throat parameter, $\eta_0=3$, 2.8, 2 and 0.1.
Recall, that for $\eta_0=0$ the non-trivial topology is lost.
The figure shows that a bifurcation phenomenon must take place
at a critical value $\eta_{0\, \rm cr}$ of the throat parameter between 2.8 and 2,
that is associated with a critical value of the
boson frequency $\omega_{\rm cr}$.
For $\eta_0 \le \eta_{0\, \rm cr}$ symmetric and asymmetric
solutions exist in the full interval $[\omega_{\rm min}, \omega_{\rm max}]$.
At the critical value $\eta_{0\, \rm cr}$, symmetric and asymmetric solutions
precisely touch at the critical value $\omega_{\rm cr}$ of the boson frequency.
For $\eta_0 > \eta_{0\, \rm cr}$ a frequency gap
$\omega_{\rm crl} \le \omega_s \le \omega_{\rm cru}$
appears, where only symmetric solutions exist.
The pairs of asymmetric solutions then bifurcate
from the symmetric ones at the end points of the gap, i.e., at the upper
critical frequency $\omega_{\rm cru}$
and at the lower critical frequency $\omega_{\rm crl}$.
The asymmetric solutions
then persist for $\omega_{\rm min} < \omega_s <\omega_{\rm crl}$
and $\omega_{\rm cru} < \omega_s < \omega_{\rm max}$.
We note that the value of the boson field at the throat, $\phi_0$,
tends to a limiting value, $\phi_0(\omega_{\rm min})=1$ in the symmetric case,
when $\omega_s \to \omega_{\rm min}$.
Here the field equation is solved by $\phi(\eta)=1$.
At first glance it may be surprising, that for small $\eta_0$
($\eta_0=0.1$ in the figure) the value
of $\phi_0$ of the asymmetric solutions is much lower than
the corresponding value of the symmetric solutions,
while both approach a $Q$-ball solution in the limit
$\eta_0 \to 0$. The reason for this is, that in the asymmetric case,
in the limit $\eta_0 \to 0$
the $\phi_+$ field has to jump from its maximal value to zero
at $\eta=0$ (and likewise the $\phi_-$ field).
Therefore for sufficiently small values of
$\eta_0$ the asymmetric field assumes about half its maximal value
at $\eta=0$.
\subsubsection{Mass and particle number}
We now turn to the global charges of these solutions, beginning with the mass.
In Fig.~\ref{fig1e} we exhibit the mass
versus the boson frequency $\omega_s$
for several values of the throat parameter $\eta_0$,
including the critical value $\eta_{0\, \rm cr}$.
The mass $M$ of the symmetric solutions is shown by dotted curves,
the mass $M_+$ of the solutions with more matter localized
in ${\cal M}_+$ is represented by solid curves, and
the mass $M_-$ of the solutions with more matter localized
in ${\cal M}_-$ by dashed curves.
We recall that these masses
refer to the asymptotic behaviour in ${\cal M_+}$.
Because of the symmetry of the pair of asymmetric solutions,
the values of their masses would be interchanged
when read off in ${\cal M_-}$.
The figure nicely illustrates the bifurcation phenomenon seen already in
Fig.~\ref{fig1d}.
Below $\eta_{0\, \rm cr}$ there are three distinct curves
for $M$, $M_+$ and $M_-$ in the full frequency range.
At $\eta_{0\, \rm cr}$ the three curves touch at the critical
boson frequency $\omega_{\rm cr}$,
while beyond $\eta_{0\, \rm cr}$ there are two
bifurcation points of the boson frequency,
$\omega_{\rm crl}$ and $\omega_{\rm cru}$,
where the pairs of asymmetric solutions bifurcate from
the symmetric ones.
Away from such bifurcation points, the mass $M_-$ is much smaller
than the other masses.
The particle number $Q$ is exhibited in Fig.~\ref{fig1f}.
Employing the same style for the respective $Q$ curves as for the $M$
curves, we see, that the dependence of the particle number on the
throat size and on the boson frequency is completely analogous to the
dependence of the mass.
\begin{figure}[t!]
\begin{center}
\vspace{0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig2a.eps}
\label{fig2a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig2b.eps}
\label{fig2b}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig2c.eps}
\label{fig2c}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig2d.eps}
\label{fig2d}
}
}
\end{center}
\vspace{-0.7cm}
\caption{Probe limit:
(a) The mass $M$ (dotted) of the symmetric solutions
and the mass $M_+$ (solid) and $M_-$ (dashed)
of the asymmetric solutions versus the respective
particle number $Q$, $Q_+$, and $Q_-$
for throat parameter $\eta_0=2$ in ${\cal M}_+$.
(b) Same as (a) for the critical value of $\eta_0$.
(c) Same as (a) for $\eta_0=2.8$.
(d) Same as (a) for $\eta_0=3$.
The mass of $Q$ free bosons is also shown (thin-dotted).
\label{fig2}
}
\end{figure}
Le us now demonstrate that due to spontaneous symmetry breaking
the asymmetric solutions are energetically favourable.
To this end we address the question, in which of the solutions
the bosons are most strongly bound.
The mass of $Q$ free bosons is given by
\begin{equation}
M_{\bf free} = m_b \, Q \ .
\end{equation}
The binding energy of solutions with $Q$ ($Q_\pm$)
particles can then be extracted
from the difference of their mass $M$ ($M_\pm$)
and the mass of $Q$ ($Q_\pm$) free bosons.
When considering $M(Q)$, $M_+(Q_+)$ and $M_-(Q_-)$
one obtains cusp-like structures,
as illustrated in Fig.~\ref{fig2}, where
the masses $M$, $M_+$ and $M_-$ are shown
versus their respective particle numbers
for the same set of throat parameters as in Fig.~\ref{fig1e}.
Also shown in the figure is the mass of $Q$ free bosons.
The symmetric solutions always form a cusp at some minimal
value of the mass and the particle number.
From this cusp two branches emerge, where the lower branch
soon becomes bound, i.e., $M<M_{\bf free}$, whereas the
upper branch remains unbound.
The branch structure of the asymmetric solutions depends
on the throat parameter and the bifurcation phenomenon.
For $\eta_0 < \eta_{0\, \rm cr}$, the mass $M_+$ of the
asymmetric solutions possesses an analogous cusp structure
to the symmetric one. However beyond $\eta_{0\, \rm cr}$
the respective upper and lower branch are no longer
connected.
Then one clearly notices the bifurcations at $\omega_{\rm crl}$
and $\omega_{\rm cru}$, where in each case two asymmetric branches
bifurcate from a symmetric branch.
In the vicinity of each bifurcation point, both of the emerging asymmetric solutions
possess a lower mass for the same particle number than
the respective symmetric solutions. Thus they are energetically favoured,
and they remain energetically favoured,
also far from their respective bifurcation points.
As expected, the spontaneous symmetry breaking
leads to energetically more favourable solutions.
\subsection{Gravitating Solutions}
We now consider the backreaction of the boson field on the wormhole.
This means that the full set of coupled nonlinear Einstein-matter equations is solved.
Then in addition to the throat parameter $\eta_0$ and the
boson frequency $\omega_s$ the coupling constant to
gravity, which is contained in the parameter $\kappa$,
enters as another continuous parameter.
To reduce the resulting amount of data, we here fix the throat
parameter to $\eta_0=3$, a value above the critical value.
This will allow us to see, how the coupling to gravity
affects the symmetry breaking.
Again we will focus on solutions in five dimensions.
Fixing the values of $\kappa$ and $\eta_0$ (as well as the number
of dimensions), we then
obtain families of gravitating solutions,
formed again for both symmetric (${\cal S}$) and asymmetric (${\cal A}_\pm$)
solutions, which depend on the boson frequency $\omega_s$.
However, unlike the case of the probe limit, the dependence of these
families of solutions on $\omega_s$ can become very involved
in the presence of gravity. Moreover, the wormhole geometry changes
at certain frequencies from single throat to double throat configurations.
\subsubsection{Symmetry Breaking}
\begin{figure}[t!]
\begin{center}
\vspace{0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig3a.eps}
\label{fig3a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig3b.eps}
\label{fig3b}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig3c.eps}
\label{fig3c}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig3d.eps}
\label{fig3d}
}
}
\end{center}
\vspace{-0.7cm}
\caption{
Properties of gravitating solutions versus
the boson frequency $\omega_s$ ($\eta_0=3$).
(a) $\kappa=0.001$: The mass $M$ of the symmetric solutions (dotted green),
and the masses
$M_+$ (solid red) and $M_-$ (dashed lilac)
of the asymmetric solutions.
For double throat configurations the colour is changed to blue.
(b) Same as (a) for the particle number $Q$.
(c) Same as (a) for $\kappa=0.01$
(d) Same as (a) for $\kappa=0.1$.
Also indicated are the masses of the respective boson stars (solid black).
The mass and particle number in the probe limit
are shown in (a) and (b) for $\omega_s>0.6$ (dotted black).
The thin vertical lines indicate $\omega_{\rm max}$.
\label{fig3}
}
\end{figure}
Let us begin by considering the spontaneous symmetry breaking
in the presence of gravity.
In Fig.~\ref{fig3} we illustrate the dependence of the families of
gravitating solutions on the coupling constant $\kappa$,
keeping the throat parameter $\eta_0$ fixed.
We start with a rather small value of the coupling constant $\kappa$.
In particular, we exhibit in Fig.~\ref{fig3a}
the masses versus the boson frequency for
$\kappa=0.001$ and in Fig.~\ref{fig3b} the respective
particle numbers. We then increase the coupling to
$\kappa=0.01$ and $\kappa=0.1$, where the respective masses are
shown in Fig.~\ref{fig3c} and Fig.~\ref{fig3d}.
First of all we observe that the symmetry breaking persists
when gravity is coupled.
Thus besides the family of symmetric solutions
also the families of asymmetric solutions are retained.
The bifurcation frequencies
$\omega_{\rm crl}$ and $\omega_{\rm cru}$,
where the pairs of asymmetric solutions bifurcate from
the symmetric solutions,
vary only slightly with increasing $\kappa$.
As in the probe limit, in between these two bifurcation frequencies
only symmetric solutions exist.
Clearly, in a large range of boson frequencies,
the structure of the gravitating solutions
follows the structure of the solutions in the probe limit.
In fact, as seen in Fig.~\ref{fig3a} and \ref{fig3b},
the solutions possess very similar global charges there.
However, for very large frequencies and for small frequencies
significant deviations occur.
In particular,
a third bifurcation frequency $\omega_{\rm crg} < \omega_{\rm crl}$
arises, where the asymmetric solutions merge again with the
symmetric ones. This bifurcation is highlighted in the inset
in Fig.~\ref{fig3c}.
This new gravity induced bifurcation at boson frequency $\omega_{\rm crg}$
marks the endpoint of the asymmetric branches.
Let us now go into more detail.
For small $\kappa$ a very prominent new feature
is the spiral structure at large values of the mass
and particle number.
This structure is present for both symmetric and
asymmetric solutions, and resembles at first glance
very much the spiral structure of boson stars.
Indeed, the branches of solutions of topologically
non-trivial solutions follow closely the
branch of boson stars up into the spiral.
However, unlike boson stars, the present spirals unwind again
as seen in the inset in Fig.~\ref{fig3a}.
From a physical point of view,
we consider that the most relevant branches of solutions are
those that start from the minimal value of $M$ (${\cal S}$) or from
the bifurcation frequency $\omega_{\rm crl}$ (${\cal A}_\pm$)
and continue to smaller frequencies until
(for the ${\cal S}$ and ${\cal A}_+$ solutions)
the mass and the particle number reach a maximum.
At this maximum a change of the stability
of the solutions is expected to occur, in analogy to boson stars.
Namely the solutions should acquire an (additional) unstable mode.
By symmetry the ${\cal A}_-$ would of course also
change stability at the same frequency.
As the spirals unwind, the symmetric and asymmetric configurations
change their geometry and evolve a double throat structure,
as indicated by the change of colour (blue) in the figures.
This will be discussed in more detail below.
Here we note that the families of solutions then
bend backwards and form descending branches
with respect to their global charges, which follow
closely those of the physically more relevant ascending branches.
When the descent is slowed down, the solutions approach
the third bifurcation frequency $\omega_{\rm crg}$,
where the asymmetric solutions merge again with the
symmetric ones.
Only this symmetric branch then continues
to smaller frequencies $\omega_s$.
We remark, that for frequencies very close to the maximum frequency
also significant deviations from the probe limit arise,
not discerned in Fig.~\ref{fig3a}.
However, the inset of Fig.~\ref{fig3b}, where the particle number
is shown, reveals that some interesting further branch structure
is present close to $\omega_{\rm max}$ for these small values
of the coupling $\kappa$.
In general, the dependence of the particle number on the frequency
follows very closely the dependence of the mass.
We therefore exhibit the particle number only for a single coupling
constant $\kappa$.
The only noticeable exception to this behaviour occurs for
$\omega_s < \omega_{\rm crg}$ on the last part of the branch of
symmetric solutions (independent of $\kappa$).
Here the particle number increases,
whereas the mass remains almost constant.
To address the $\kappa$-dependence of the solutions,
we exhibit in Fig.~\ref{fig3} also the masses for
$\kappa=0.01$ (c) and $\kappa=0.1$ (d).
For $\kappa=0.01$ the spiral structure is still
present at a rudimentary level. Thus the general
structure of the solutions is analogous to the one
of the lower couplings.
When $\kappa$ is further increasing, however, the spiral
behaviour disappears, and the branch structure
of the families of solutions radically simplifies.
This also holds for frequencies close to the
maximal frequency.
The mass and the particle number of a given family
are then simple functions of the boson frequency $\omega_s$,
since the backbendings present for lower $\kappa$ disappear,
as illustrated for $\kappa=0.1$.
We have focussed our discussion here on the fundamental solutions.
But we would like to remark, that the families of
first excitations follow the pattern
of the fundamental solutions.
\begin{figure}[t!]
\begin{center}
\vspace{0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig4a.eps}
\label{fig4a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig4b.eps}
\label{fig4b}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig4c.eps}
\label{fig4c}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig4d.eps}
\label{fig4d}
}
}
\end{center}
\vspace{-0.7cm}
\caption{Gravitating solutions for throat parameter $\eta_0=3$:
(a) The mass $M$ (dotted) of the symmetric solutions
and the mass $M_+$ (solid) and $M_-$ (dashed)
of the asymmetric solutions versus the
the particle number $Q$, $Q_+$, and $Q_-$, respectively,
for coupling $\kappa=0.001$
in the vicinity of the bifurcations $\omega_{\rm crl}$
and $\omega_{\rm cru}$.
(b) Same as (a) for the full range of masses.
(c) Same as (a) for $\kappa=0.01$.
(d) Same as (a) for $\kappa=0.1$, $\kappa=0.01$ and $\kappa=0.001$.
Note the different colour coding here.
In all figures the mass of $Q$ free bosons is shown for comparison
(thin black lines).
The black dot in the inset in (b) denotes the
endpoint at $\omega_{\rm max}$.
\label{fig4}
}
\end{figure}
Let us now consider the $M(Q)$ dependence in order
to demonstrate that
the asymmetric solutions are energetically favoured
also in the presence of gravity.
To this end we exhibit in Fig.~\ref{fig4} the masses
$M$, $M_+$ and $M_-$
versus the respective particle numbers
$Q$, $Q_+$ and $Q_-$
for several values of the coupling constant $\kappa$.
Also shown is the mass of $Q$ free bosons for comparison.
Fig.~\ref{fig4a} presents the masses
versus the particle numbers
for the small coupling $\kappa = 0.001$,
zooming into the small mass region,
where the bifurcations at $\omega_{\rm crl}$ and
$\omega_{\rm cru}$ are clearly visible.
As expected,
for these small masses and particle numbers
there is very little difference to the probe limit,
except for an overall scaling parameter
(compare Fig.~\ref{fig2d}).
Thus the asymmetric solutions are clearly energetically favoured,
when they bifurcate from the symmetric ones.
But the inset in Fig.~\ref{fig4a} shows already a part
of the additional structure of the
symmetric solutions close to $\omega_{\rm max}$.
The full structure of the solutions is shown in
Fig.~\ref{fig4b}. Interestingly, the various branches of $M(Q)$
for the symmetric and asymmetric solutions are all very close
to each other on such a large scale. The large mass solutions
are all strongly bound as a comparison with the free case shows.
The spiral structure gives rise to a number of cusps at
large masses. However, since the masses and particle numbers
are all very close, we cannot discern these cusps in the
figure. (Here only a schematic plot would clearly exhibit the cusp
structure.)
The inset in Fig.~\ref{fig4a} shows already a part
of the additional structure of the
symmetric solutions close to $\omega_{\rm max}$.
The black dot in the inset in Fig.~\ref{fig4b}
denotes the endpoint reached by the configurations
at $\omega_{\rm max}$.
This endpoint is universal, i.e., independent of the coupling
$\kappa$.
It also represents a fourth bifurcation point, since
the asymmetric and symmetric solutions merge again
at $\omega_{\rm max}$.
In Fig.~\ref{fig4c} we demonstrate the
$M(Q)$ dependence for $\kappa=0.01$ in the small mass region.
The figure shows that
the asymmetric solutions remain energetically preferred,
when the coupling is increased.
We observe that the figures \ref{fig4}(a)-(c) look very much the same.
This is demonstrated in \ref{fig4d} for all values of $\kappa$ considered.
Indeed, there is only a slight dependence on the coupling $\kappa$
apart from the overall scaling factor.
Clearly, the asymmetric solutions remain preferred.
\subsubsection{Wormhole Geometries}
\begin{figure}[t!]
\begin{center}
\vspace{0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig5a.eps}
\label{fig5a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig5b.eps}
\label{fig5b}
}
}
\end{center}
\vspace{-0.7cm}
\caption{
Throat structure of non-rotating gravitating solutions ($\eta_0=3$).
(a) The circumferential radius function $R(\eta)$ ($R_+(\eta)$)
versus the compactified coordinate $x={\rm atan}\, \eta$
for a sequence of symmetric
and asymmetric solutions with various values
of the frequency $\omega_s$ and coupling constant $\kappa = 0.001$.
(b) Same as (a) for the boson function $\phi(\eta)$ ($\phi_+(\eta)$).
\label{fig5}
}
\end{figure}
\begin{figure}[t!]
\begin{center}
\vspace{0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig6a.eps}
\label{fig6a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig6b.eps}
\label{fig6b}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig6c.eps}
\label{fig6c}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig6d.eps}
\label{fig6d}
}
}
\end{center}
\vspace{-0.7cm}
\caption{Gravitating solutions for throat parameter $\eta_0=3$:
(a) Location of the single throat respectively
the double throat and the equator versus the frequency $\omega_s$
for coupling $\kappa = 0.001$.
(b) Zoom of (a) for frequencies close to $\omega_{\rm max}$.
(c) Same as (a) for $\kappa=0.01$.
(d) Same as (a) for $\kappa=0.1$.
\label{fig6}
}
\end{figure}
Next we address the geometry of the solutions.
In particular, we are interested in the transition from single throat
configurations to configurations featuring a double throat
with an equator in between the two throats.
To illustrate this transition, we exhibit
in Fig.~\ref{fig5a} the circumferential radius function $R(\eta)$
($R_+(\eta)$)
for a sequence of symmetric (asymmetric) solutions,
where the coupling constant has the value $\kappa=0.001$.
Here a minimum of $R(\eta)$ ($R_+(\eta)$)
corresponds to a throat, and
a maximum to an equator. The transition from a single
throat to a double throat configuration occurs at an
inflection point.
For the symmetric solutions we display in Fig.~\ref{fig5a} a
single throat solution for the frequency $\omega_s=0.2456$,
the transition from a single to a double throat
configuration happening at $\omega_{\rm tr}=0.2439$,
and a double throat solution with $\omega_s=0.2435$.
The asymmetric double throat solutions bifurcate at $\omega_{\rm crg}=0.4942$
from the symmetric ones.
For the asymmetric solutions the respective solutions have frequencies
$\omega_s=0.2724$ (double throat), $\omega_{\rm tr}=0.2518$ and $\omega_s=0.2494$
(single throat).
As seen in the figure, in the asymmetric case the equator and
the second throat emerge asymmetrically.
Thus
the single throat does not degenerate at the transition
frequency, where the equator and the second throat arise.
Instead, an inflection point arises in the other part
of the manifold, which splits into a maximum and a minimum as the
frequency $\omega_s$ is increased.
In addition we display in the figure
the double throat solution at the third
bifurcation frequency $\omega_{\rm crg}$,
where the symmetric and asymmetric solutions merge again.
To get an idea of the matter distributions associated with
these configurations and, in particular, with the transitions,
we exhibit in Fig.~\ref{fig5b} the boson function $\phi(\eta)$
($\phi_+(\eta)$) for the
same set of solutions.
We observe that the value of $\phi_0$
increases from the single to the double throat solutions.
In particular, we note that the inflection point arises
close to the peak of $\phi_+(\eta)$
in the part of the manifold, where most of the matter resides.
To see the evolution of the throats of the symmetric and
asymmetric configurations, we exhibit the
dependence of their locations in Fig.~\ref{fig6},
beginning in Fig.~\ref{fig6a} with the coupling $\kappa=0.001$.
Clearly,
for symmetric solutions their single throat (solid green)
is localized at $\eta_t=0$,
while beyond their transition frequency $\omega_{\rm tr}$
their equator resides at $\eta_e=0$ (dotted black).
At the transition frequency the double throats (solid and dashed
black) emerge and stride away.
For the larger frequencies the single throat
of the asymmetric solutions is located close to $\eta_t=0$.
However, for the solutions ${\cal A}_+$ it is then shifted into
the region ${\cal M}_-$ (and vice versa) (solid dark red),
because of the backreaction of the matter on the geometry.
At the transition frequency this throat continues to
stride away (solid dark blue), while a cusp arises
in the region ${\cal M}_+$, formed by the second throat
(dashed dark blue) and the equator (dotted dark blue).
Note that the light colours (red and blue) in the figure
represent the second asymmetric solution.
At the bifurcation $\omega_{\rm crg}$,
the asymmetric throats merge with the symmetric throats,
and the asymmetric equator with the symmetric one.
In Fig.~\ref{fig6b} we zoom into the region close to $\omega_{\rm max}$.
Here only single throat solutions exist. The inset demonstrates
the merging of the solutions at $\omega_{\rm max}$.
We note that in this region the asymmetric throats first stride far away
from $\eta=0$, before they return towards the limiting symmetric solution.
As $\kappa$ is increased, the frequency range of
the asymmetric double throat solutions decreases,
as seen in Fig.~\ref{fig6c} for $\kappa = 0.01$,
where the same colour coding is used.
For large $\kappa$ only symmetric double throat solutions
are left, as illustrated in Fig.~\ref{fig6d} for $\kappa=0.1$.
These symmetric double throat solutions persist
at small frequencies for any value of the coupling $\kappa$.
\begin{figure}[t!]
\begin{center}
\vspace{-0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig7a.eps}
\label{fig7a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig7b.eps}
\label{fig7b}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig7c.eps}
\label{fig7c}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig7d.eps}
\label{fig7d}
}
}
\end{center}
\vspace{-0.7cm}
\caption{
Throat geometry:
Three dimensional view of the isometric embedding.
(a) Symmetric solution with a single throat for
$\kappa=0.001$, $\omega=0.3$.
(b) Asymmetric solution with a single throat for
$\kappa=0.001$, $\omega=0.3$.
(c) Symmetric solution with a double throat for
$\kappa=0.001$, $\omega=0.3$.
(d) Asymmetric solution with a double throat for
$\kappa=0.001$, $\omega=0.3$.
\label{fig7}
}
\end{figure}
Let us now visualize the geometry of the wormholes by
considering embeddings of spatial hypersurfaces.
We display an isometric embedding of one of the equatorial planes
($\theta=0,\ \pi/2$)
for several wormhole spacetimes in Fig.~\ref{fig7}.
For the embedding we employ the parametric representation
\begin{equation}
\rho(\eta) = R(\eta) \ , \ \ \ \
z(\eta)= \int_0^\eta\sqrt{1-R'^2} \ d\eta' \ .
\label{embedd}
\end{equation}
Below the transition value $\omega_{\rm tr}$
the solutions possess a single throat, as seen in
Fig.~\ref{fig7} for a symmetric (a) and
an asymmetric (b) solution.
For the parameters chosen in the figure,
$\kappa=0.001$ and $\omega_s=0.3$,
the single throat of the asymmetric solution
has been pushed into ${\cal M}_-$.
The respective double throat solutions are shown
in Fig.~\ref{fig7}(c) and (d).
For the symmetric ones the equator resides at the centre (c),
while for the asymmetric ones it is located
off the centre (d).
We did not find solutions with more than two throats.
Finally, we would like to briefly comment on the limiting solution
for small boson frequencies.
As discussed above, for small $\omega_s$ only symmetric solutions
persist.
When $\omega_s$ is decreased here, the boson function becomes
steeper and steeper in the vicinity of the equator,
with its peak strongly increasing.
At the same time the circumferential function $R(\eta)$
peaks more and more strongly at the equator.
This behaviour indicates that a singular limiting solution is reached,
whose Kretschmann scalar diverges in the limit.
This agrees with the previous four-dimensional study
\cite{Dzhunushaliev:2014bya}.
\subsection{Spontaneous Symmetry Breaking in Four Spacetime Dimensions}
\begin{figure}[t!]
\begin{center}
\vspace{0.5cm}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig8a.eps}
\label{fig8a}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig8b.eps}
\label{fig8b}
}
}
\mbox{\hspace{-0.5cm}
\subfigure[][]{\hspace{-1.0cm}
\includegraphics[height=.25\textheight, angle =0]{Fig8c.eps}
\label{fig8c}
}
\subfigure[][]{\hspace{-0.5cm}
\includegraphics[height=.25\textheight, angle =0]{Fig8d.eps}
\label{fig8d}
}
}
\end{center}
\vspace{-0.7cm}
\caption{
Properties of gravitating solutions versus
the boson frequency $\omega$ ($\eta_0=3$) in four dimensions.
(a) $\kappa=0.01$: The mass $M$ of the symmetric solutions (dotted green),
and the masses
$M_+$ (solid red) and $M_-$ (dashed lilac)
of the asymmetric solutions.
For double throat configurations the colour is changed to blue.
(b) Same as (a) for the particle number $Q$.
(c) The mass $M$ (dotted) of the symmetric solutions
and the mass $M_+$ (solid) and $M_-$ (dashed)
of the asymmetric solutions versus the
the particle number $Q$, $Q_+$, and $Q_-$, respectively,
for coupling $\kappa=0.01$
in the vicinity of the bifurcations $\omega_{\rm crl}$
and $\omega_{\rm cru}$.
(d) The limit $\omega_s \to \omega_{\rm max}$ is demonstrated
for the boson field function $\phi$.
\label{fig8}
}
\end{figure}
After having discussed in detail the properties of the
wormholes within bosonic matter in five spacetime dimensions,
we now briefly demonstrate that the features observed
in five dimensions also hold in four dimensions.
We recall, that
in \cite{Dzhunushaliev:2014bya} only the symmetric solutions
have been studied, which we now complement with the asymmetric ones.
First of all we note, that the analogous bifurcation phenomenon
takes place in four dimensions.
In the probe limit,
below a critical value $\eta_{0\, \rm cr}$ of the throat parameter,
symmetric and asymmetric solutions exist in the full
frequency interval $[\omega_{\rm min}, \omega_{\rm max}]$.
Above this critical value a frequency gap
$\omega_{\rm crl} \le \omega_s \le \omega_{\rm cru}$
appears, where only symmetric solutions are found.
At the end points of the gap, i.e., at $\omega_{\rm cru}$
and $\omega_{\rm crl}$,
pairs of asymmetric solutions bifurcate from the symmetric ones,
and persist for $\omega_{\rm min} < \omega_s <\omega_{\rm crl}$
and $\omega_{\rm cru} < \omega_s < \omega_{\rm max}$.
As in five dimensions,
this bifurcation phenomenon is retained in the presence of gravity.
Depending on the coupling constant $\kappa$, the branch structure
of the solutions changes considerably, as already demonstrated in
\cite{Dzhunushaliev:2014bya} for the symmetric solutions
(and for a smaller throat parameter).
The asymmetric solutions follow this general pattern.
Thus with increasing $\kappa$ the branch structure simplifies,
and the spiral part disappears,
analogously to the five-dimensional solutions.
We here demonstrate the branch structure
for the value of $\kappa=0.01$.
Fig.~\ref{fig8a} exhibits the masses
$M$ (dotted green),
$M_+$ (solid red) and $M_-$ (dashed lilac),
where the bifurcations are clearly visible.
The asymmetric solutions emerge from the symmetric ones at
the bifurcations $\omega_{\rm crl}$ and $\omega_{\rm cru}$,
and merge again at
$\omega_{\rm crg}$ (as well as at $\omega_{\rm max}$).
Also, the transition from single to double throat solutions
is observed (blue)
analogously to the five-dimensional case.
The respective particle numbers are shown in Fig.~\ref{fig8b}.
Again, they follow the behaviour of the masses rather closely,
except for very small frequencies $\omega_s$, where only
the symmetric solutions are retained.
In Fig.~\ref{fig8c}, the masses $M$, $M_+$ and $M_-$
are shown versus the respective particle numbers
$Q$, $Q_+$ and $Q_-$ in the bifurcation region
(i.e., for the larger values of the frequency $\omega_s$),
together with the mass of $Q$ free bosons.
The inset in Fig.~\ref{fig8c} exhibits
the solutions in their full range of existence.
As expected, the asymmetric solutions are energetically
favoured over the symmetric ones. Thus the spontaneous
symmetry breaking leads to more strongly bound systems.
We conjecture, that the analogous phenomenon is present
also in more than five dimensions.
Let us remark, that
in four dimensions we are in principle entitled to also address
astrophysical aspects of these solutions.
Then one may view them as
boson stars with a nontrivial topology.
However, this aspect will be addressed elsewhere.
As a final point, we remark that in four dimensions
the limiting solution for $\omega_s \to \omega_{\rm max}$
becomes rather trivial in the sense, that the
boson field tends to zero and the solution has vanishing
mass. This is different for five dimensions,
where the limiting solution has a non-trivial boson field
and a finite mass.
We exhibit in Fig.~\ref{fig8d} a sequence of solutions
close to $\omega_{\rm max}$, which demonstrates
the way the boson field approaches its limit.
\section{Conclusions and Outlook}
In this paper we have considered Ellis wormholes in the presence of
a complex bosonic matter field and encountered the phenomenon
of spontaneous symmetry breaking of the solutions.
Starting with the probe limit, we have seen that besides
the symmetric configurations there are also asymmetric configurations
present.
The latter always come in pairs, and are related to each other by
a reflection with respect to the radial coordinate $\eta=0$.
For a small throat size, the symmetric and asymmetric solutions
are present in the full frequency interval
[$\omega_{\rm min}$,$\omega_{\rm max}$], whereas for a larger
throat size this is only true for the symmetric solutions.
There the asymmetric solutions branch off the symmetric ones
at critical values of the frequency,
$\omega_{\rm crl}$ and $\omega_{\rm cru}$.
At the critical value of the throat size $\eta_{0\, \rm cr}$
these critical frequency values coincide.
Both symmetric and asymmetric solutions satisfy the same
set of boundary conditions. In that sense
the asymmetric solutions appear spontaneously,
without any external trigger. The reason for their appearance
is that the asymmetric solutions are energetically
favourable, as we have shown.
To this end, we have analyzed the masses and particle numbers
of the solutions. For the symmetric solutions one finds the
same mass $M$ in both asymptotic regions, and likewise
the same particle number.
For the asymmetric solutions this is different. Here
the solution with most of its mass located in the region ${\cal M}_+$
possesses the mass $M_+$ in this region
and the mass $M_-$ in ${\cal M}_-$. For the second asymmetric
solution of the pair these masses are interchanged
because the two are related by reflection.
The same holds for the particle number.
When the mass of the three solutions
is then considered versus their respective particle number,
it becomes clear, that asymmetric solutions are
more strongly bound than the symmetric solutions.
Thus the spontaneous symmetry breaking leads to
energetically favoured configurations.
All of this remains valid when gravity is coupled.
However, gravity introduces further new features.
First of all,
the backreaction of the boson field on the
metric removes the lower frequency bound $\omega_{\rm min}$.
For small values of the coupling constant $\kappa$
a spiralling behaviour arises, that is known from compact stars.
However, unlike for compact stars, here the spirals
unwind again. We attribute this effect to the presence
of the negative energy density in the form of a
phantom field, since such unwinding has, for instance, also been observed
for bosonic configurations in Einstein-Gau\ss -Bonnet theory
\cite{Hartmann:2013tca}.
Second, the presence of gravity generates a further bifurcation
phenomenon at $\omega_{\rm crg}$. Here the pair of asymmetric
solutions merges again with the symmetric ones.
Only the symmetric solutions then persist to small boson frequencies.
Third, we observe a transition in the geometry of the
solutions. For small $\kappa$ this transition arises
in the vicinity of the spiral. Here the single throat solutions
develop an equator and a second throat. For the symmetric
solutions, the equator is then localized at the radial coordinate
$\eta=0$, surrounded symmetrically by both throats.
For the asymmetric solutions, in contrast, the equator and
the second throat originate far from the first throat
in the other part of the manifold.
Here we have performed most of the calculations in
five spacetime dimensions. However, the formalism
is general for $D$ dimensions. Pure Ellis wormholes
in $D$ dimensions have been obtained in \cite{Torii:2013xba}.
It should be straightforward to include bosonic
matter and obtain the analogues of the
configurations studied here also in $D>5$ dimensions.
In particular, we expect that the phenomenon of
spontaneous symmetry breaking will be present
independent of the dimension.
We have already shown that this phenomenon is also
present in four dimensions. In this case,
such configurations of bosonic matter
surrounding wormholes might also be of
potential astrophysical interest \cite{Dzhunushaliev:2014bya}.
In particular,
one can obtain solutions which differ widely in mass and size
by varying the potential for the complex scalar field.
For instance, there may also be solutions which
mimic compact astrophysical objects like neutron stars
or black holes.
These solutions could be studied in the context
of gravitational lensing
\cite{Abe:2010ap,Toki:2011zu,Takahashi:2013jqa},
with respect to their light curves \cite{Dzhunushaliev:2014mza},
their geodesics
\cite{Eilers:2013lla}, etc.
Let us conclude with some remarks on the stability of these
solutions.
The Ellis wormholes are known to be unstable
\cite{Shinkai:2002gv,Gonzalez:2008wd,Gonzalez:2008xk,Torii:2013xba}.
The unstable radial mode of the wormholes was shown to
persist in the presence of bosonic matter
for symmetric solutions in four dimensions
\cite{Dzhunushaliev:2014bya},
where the instability was weakened by the presence of matter.
Now, that we have seen, that spontaneous symmetry breaking occurs,
we conjecture that the symmetric solutions will acquire
another unstable mode from the matter side. In contrast,
the asymmetric solutions will only possess the single
unstable wormhole mode (on their fundamental branch).
While we defer a full analyis of the stability of the symmetric and
asymmetric solutions to a later time,
we note already, that the analysis of the solutions in the
probe limit precisely conforms to this expectation.
Here, at the bifurcation frequencies
$\omega_{\rm crl}$ and $\omega_{\rm cru}$
the symmetric solutions indeed exhibit a zero mode,
which then turns into a second unstable mode
in the parameter space where the asymmetric solutions exist.
For the asymmetric solutions
stability could possibly be achieved
by removing the phantom field
and modifying gravity instead
\cite{Hochberg:1990is,Fukutaka:1989zb,Ghoroku:1992tz,Furey:2004rq,Bronnikov:2009az,Kanti:2011jz,Kanti:2011yv,Lobo:2009ip,Harko:2013yb}.
\section*{Acknowledgement}
We gratefully acknowledge support by the German Research Foundation
within the framework of the DFG Research Training Group 1620
{\it Models of gravity}
as well as support by the Volkswagen Stiftung, and support from FP7, Marie Curie
Actions, People, International Research Staff Exchange
Scheme (IRSES-606096).
BK
gratefully acknowledges support from Fundamental Research in Natural Sciences
by the Ministry of Education and Science of Kazakhstan.
\section{Appendix}
We briefly explain, how we obtain the mass
and the particle number of the asymmetric solutions.
For these special care is needed, since there is some ambiguity
as to where to put the lower limit of the respective volume integrals.
We therefore extract these global charges from the
asymptotic behaviour of the solutions.
For the symmetric solutions there is no such ambiguity.
\subsection{Mass in the probe limit}
In the probe limit the backreaction of the matter on the
spacetime is not taken into account, since the matter equation
is solved in the background of the Ellis wormhole.
Then the mass cannot be extracted from the asymptotic form of the metric.
To obtain the mass anyway asymptotically, we resort to the
following construction.
We consider the Einstein equation
\begin{equation}
R^0_0\sqrt{-g} = \kappa \left(T^0_0 +\frac{1}{2-D}T_\mu^\mu\right)\sqrt{-g}
\end{equation}
and treat the metric function $a$ as order ${\cal O}(\kappa)$. Consequently,
only the background metric enters the right hand side.
Evaluation yields
\begin{equation}
-\frac{D-3}{2}\left(
\left[\left(ph\right)^{\frac{D-2}{2}}\frac{1}{\sqrt{p}}\right]a'\right)'
=
\kappa \left(T^0_0 +\frac{1}{2-D}T_\mu^\mu\right)
\left(ph\right)^{\frac{D-2}{2}}\sqrt{p} \ .
\end{equation}
Defining $\rho_{\rm m} = \left[\left(ph\right)^{\frac{D-2}{2}}\frac{1}{\sqrt{p}}\right]a'$
we find
\begin{equation}
\rho_{\rm m}(\eta) = -\frac{2}{D-3} \kappa
\int_{-\infty}^\eta { \left(T^0_0 +\frac{1}{2-D}T_\mu^\mu\right)
\left(ph\right)^{\frac{D-2}{2}}\sqrt{p} } d\eta'
+ \rho_{\rm m 0} \ ,
\end{equation}
where $\rho_{\rm m \ 0}$ is an integration constant. In the next step we integrate
$a' =\left[\left(ph\right)^{\frac{2-D}{2}}\sqrt{p}\right] \rho_{\rm m}$,
\begin{eqnarray}
a(\eta) & = & -\frac{2\kappa}{D-3}\int_{-\infty}^\eta {
\left[\left(ph\right)^{\frac{2-D}{2}}\sqrt{p}\right]
\left\{
\int_{-\infty}^{\eta'} { \left(T^0_0 +\frac{1}{2-D}T_\mu^\mu\right)
\left(ph\right)^{\frac{D-2}{2}}\sqrt{p} } d\eta''
\right\} } d\eta'
\nonumber\\
& &
+ \rho_{\rm m 0}\int_{-\infty}^\eta {\left[\left(ph\right)^{\frac{2-D}{2}}\sqrt{p}\right]} d\eta'
+a_0 \ ,
\end{eqnarray}
where $a_0$ is again an integration constant.
Next we find the integration constants $\rho_{\rm m 0}$ and $a_0$ from
the boundary conditions $a(\pm \infty)=0$,
\begin{equation}
\rho_{\rm m 0}=
\frac{2\kappa}{D-3}
\frac{
\int_{-\infty}^\infty {
\left[\left(ph\right)^{\frac{2-D}{2}}\sqrt{p}\right]
\left\{
\int_{-\infty}^{\eta} { \left(T^0_0 +\frac{1}{2-D}T_\mu^\mu\right)
\left(ph\right)^{\frac{D-2}{2}}\sqrt{p} } d\eta'
\right\} } d\eta
}
{\int_{-\infty}^\infty {\left[\left(ph\right)^{\frac{2-D}{2}}\sqrt{p}\right]} d\eta}
\ , \ \ a_0=0 \ .
\end{equation}
The masses $M_\pm$ are related to the asymptotic behaviour of the function $a$,
\begin{equation}
M_\pm = \pm \frac{D-3}{2\kappa}\left[\eta^{D-2} a'\right]_{\pm \infty}\Omega_{D-2}
= \pm \frac{D-3}{2\kappa}\rho_{\rm m}(\pm \infty)\Omega_{D-2} \ .
\end{equation}
Explicitly,
\begin{eqnarray}
M_+ & = & \left\{-\frac{D-2}{D-3}\int_{-\infty}^\infty{\left(T^0_0 +\frac{1}{2-D}T_\mu^\mu\right)
\left(ph\right)^{\frac{D-2}{2}}\sqrt{p} } d\eta
+\frac{D-2}{2\kappa} \rho_{\rm m 0}\right\}\Omega_{D-2} \ ,
\nonumber\\
M_- & = & -\frac{D-2}{2\kappa}\rho_{\rm m 0}\Omega_{D-2} \ .
\end{eqnarray}
We note that $M_+$ can be written as
\begin{equation}
M_+ = \frac{D-2}{D-3} \int_\Sigma {
\left(T_{\mu\nu} +\frac{1}{2-D} g_{\mu\nu}T_\lambda^\lambda\right)n^\mu \xi^\nu }dV
+\frac{D-2}{2\kappa} \rho_{\rm m 0}\Omega_{D-2} \ ,
\end{equation}
where $\Sigma$ denotes a spacelike hypersurface including both asymptotic regions
of ${\cal M}_+$ and ${\cal M}_-$.
\subsection{Particle number via electric charge}
We consider a fictitious electrostatic potential $\Phi_{\rm el}(\eta)$
sourced by the current $j^\mu$, Eq.~(\ref{current}),
\begin{equation}
\partial_\mu\left(
\sqrt{-g} g^{\mu \nu} g^{t t}\partial_\nu \Phi_{\rm el}
\right)
=
-j^t \sqrt{-g} \ .
\end{equation}
For the spherically symmetric Ansatz this yields
\begin{equation}
\left(e^{(D-3)a}\left(ph\right)^{\frac{D-2}{2}}\frac{1}{\sqrt{p}}
\Phi'_{\rm el}\right)'
= j^t e^{-a}\left(ph\right)^{\frac{D-2}{2}}\sqrt{p}
\ .
\label{Maxw}
\end{equation}
The charges $Q_{\pm}$ can then be obtained from
\begin{equation}
Q_{\pm} =
\mp \left[\
\eta^{D-2} \Phi'_{\rm el}
\right]_{\pm\infty} \Omega_{D-2} \ .
\end{equation}
Introducing the auxiliary quantity
$\rho_q= e^{(D-3)a}\left(ph\right)^{\frac{D-2}{2}}\frac{1}{\sqrt{p}}
\Phi'_{\rm el}$
we obtain from Eq.~(\ref{Maxw})
\begin{equation}
\rho_q(\eta) =
\int_{-\infty}^\eta {j^t e^{-a}\left(ph\right)^{\frac{D-2}{2}}\sqrt{p}} d\eta'
+\rho_{q 0}
\end{equation}
with integration constant $\rho_{q 0}$. On the other hand, expressing
$\Phi'_{\rm el}$ in terms of $\rho_q$ and integrating yields
\begin{equation}
\Phi_{\rm el}(\eta)=
\int_{-\infty}^\eta{
e^{-(D-3)a}\left(ph\right)^{-\frac{D-2}{2}}\sqrt{p}
\left[
\int_{-\infty}^{\eta'} {j^t e^{-a}\left(ph\right)^{\frac{D-2}{2}}\sqrt{p}} d\eta''
\right]}
d\eta'
+\rho_{q 0}\int_{-\infty}^\eta{e^{-(D-3)a}\left(ph\right)^{-\frac{D-2}{2}}\sqrt{p}}
d\eta' + \Phi_{\rm el 0} \ ,
\end{equation}
where $\Phi_{\rm el 0}$ is another integration constant.
In order to detemine the integration constants $\rho_{q 0}$ and $\Phi_{\rm el 0}$ we
impose the boundary conditions $\Phi_{\rm el}(\pm\infty)=0$.
This yields $\Phi_{\rm el 0}=0$ and
\begin{equation}
\rho_{q 0}=
-\frac{
\int_{-\infty}^\infty{
e^{-(D-3)a}\left(ph\right)^{-\frac{D-2}{2}}\sqrt{p}
\left[
\int_{-\infty}^\eta {j^t e^{-a}\left(ph\right)^{\frac{D-2}{2}}\sqrt{p}} d\eta'
\right]}
d\eta}{
\int_{-\infty}^\infty{e^{-(D-3)a}\left(ph\right)^{-\frac{D-2}{2}}\sqrt{p}}
d\eta} \ .
\end{equation}
We observe that $Q_\pm = \mp\rho_q(\pm \infty)\Omega_{D-2}$. Consequently,
\begin{equation}
Q_+ =
-\left(\int_{-\infty}^\infty {j^t e^{-a}\left(ph\right)^{\frac{D-2}{2}}\sqrt{p}} d\eta
+\rho_{q 0}\right)\Omega_{D-2} \ , \ \
Q_- = \rho_{q 0}\Omega_{D-2} \ .
\end{equation}
|
\section{Introduction} \label{sec1}
\rev2{Two- and three-dimensional lattices of connected beams can provide pentamode-like behavior in the static limit
\cite{Milton95,Norris14}. One reason for interest in such structures is that they exhibit one-wave behaviour characteristic of scalar or acoustic wave systems, while also displaying material anisotropy, so that anisotropic acoustic effects are possible. Such scalar dynamic effective properties are perhaps surprising since periodic lattice structures support
multiple wave types yielding complex dispersion properties} described by Bloch-Floquet spectra, particularly band gaps
\cite{Martinsson03d,Phani06} and anisotropic propagation \cite{Gonella08}. At the same time, the relatively simple geometry of two- and three-dimensional lattices allows for the possibility of mechanical simplifications that maintain the underlying structural dynamics of the continuous beam elements while leading to accurate predictions for the dispersion properties. This paper focuses on the latter aspect, as we develop a semi-analytical formalism that reduces the Bloch wave problem to an analytically simple form while retaining the crucial mechanics of the structure.
\reva{Two distinct approaches \cite{WITTRICK1971} to analysing waves in periodic structures can be distinguished based on the number of degrees of freedom: {\it finite models} and {\it infinite models}. The former naturally includes the
finite elements method (FEM) for which strategies have been developed that are specifically designed to treat lattice structures. Thus, \cite{Phani06} developed a FEM procedure for calculating dispersion curves of hexagonal, square and triangular lattice structures. FEM has been used by \cite{Gonella08} to consider waves in regular and re-entrant hexagonal lattices and by \cite{Spadoni09} to examine hexagonal chiral lattices as phononic crystals.
Infinite methods, as in this paper, retain some of the characteristics of the continuous nature of the structure at the smallest scale. A simple beam considered as a separate entity displays an infinite number of modes; it is therefore no surprise that a model based on such elements has an infinite number of Bloch-Floquet branches.
}
\reva{
The fundamental step in deriving the dispersion relation for waves in a periodic structure is the application of the Floquet condition on the unit cell of the lattice. Whatever approach is used, whether FEM or semi-analytical, this step reduces the problem to a generalized eigenvalue problem for the system (or stiffness) matrix.
The main distinction between finite and infinite models is that the former reduce to linear systems with eigenvalue equal to the square of the frequency, whereas infinite models necessarily involve finding roots of transcendental equations. There are, however, computational approaches adapted to this problem, such as that of Wittrick and Williams \cite{WITTRICK1971} based on Householder's algorithm.
}
Regarding other {\reva{infinite } methods for solving dynamic waves problems in lattices, we note that
an interesting alternative wave-based approach for determining the Bloch waves in 2D periodic structures was proposed by \cite{Leamy12}. The semi-analytical method considers the explicit waves propagating back and forth on each member, coupled by reflection and transmission matrices at joints. The present method is similar to that of \cite{Leamy12} in that both approaches yield exact dispersion relations within the context of the beam theories employed (Timoshenko beam theory was used in \cite{Leamy12}). However, the present approach is arguably simpler in that it does not require propagation and reflection/transmission matrices for the multiple wave types. Instead, the crucial ingredient in the present method is the dynamic stiffness matrix that relates forces at the two ends of a beam member to the displacements at either end.
An important limit of any dynamic model is the low frequency, quasistatic or homogenization limit. Although static homogenization theory for quite general lattice structures has been developed by several authors, e.g. \cite{Martinsson03a,Gonella08b}, these approaches do not derive the homogenized properties from the limit of a dynamic model. An exception is the paper by
\cite{Colquitt11} who showed for a triangular lattice that only by including the flexural wave effects is the effective mass properly modeled in the low frequency limit. Simpler beam models which ignore flexural waves, or bending, show quasistatic wave speeds with effective mass that is less than the total mass of the unit cell \cite{Colquitt11}. This suggests that models ignoring flexural effects do not properly account for the distributed mass on the wave-bearing segments of the structure, and cannot yield the correct quasistatic results.
The analytical approach used here represents the lattice members as uni-dimensional beams supporting longitudinal and flexural waves. A strategy for implementing this was outlined by \cite{Martinsson03d} who introduced the necessary stiffness matrix relating forces and displacements at the ends of a beam. By combining these matrices it is possible to represent any periodic lattice, in principle. The method of \cite{Martinsson03d} was used in \cite{Colquitt11} to consider lattices with triangular unit cell structure, and for square cell lattices in \cite{Colquitt13}.
In this paper we develop further the approach proposed by \cite{Martinsson03d} and \cite{Colquitt11}. We present, for the first time, analysis of a general hexagonal unit cell lattice, a structure of great interest in relation to graphene and other phenomena. Also, the method is extended into 3D to analyze the tetrahedral unit cell lattice.
\reva{The formulation is semi-analytical to the extent that all matrix elements are explicit, the dispersion relation for square and cubic lattices are derived analytically. Although one could obtain analytic dispersion relations for hexagonal and tetrahedral lattices using symbolic computation \cite{Williams1995}, direct numerical methods are employed at the final stage to perform computation}. The semi-analytical nature of the solution allows us to extract the low frequency asymptotics, and to find closed-form expressions for the quasistatic Christoffel matrix, as demonstrated for hexagonal and rectangular unit cell lattices in 2D.
\rev2{In this sense the present study is step in the continuation from low frequency (quasi-static) response governed by effective elastic stiffness and density to dynamic effective medium models. }
\rev2{ The present analysis does not include torsion in the individual members. The beams are assumed to have large length to thickness ratio, and hence a static applied macroscopic torsion is borne at the level of the unit cell by flexure of the members. Bending is the dominant effect for producing torsion in the lattice structures considered here. This can be seen {\it a posteriori} from the comparisons below with full elastodynamic simulations which do not display Bloch-Floquet branches with significant torsional effects at level of the lattice member. In other words, torsion in individual members is ignored because we are only including the dynamic counterparts of the micro-effects that lead to the static effective medium. Note that the present model allows for rigid body rotation at the unit cell level, which is consistent with static homogenization \cite{Norris14}.
}
The format of the paper is as follows. The solution method for hexagonal and tetrahedral lattices is summarized in \S \ref{sec2}, where the Bloch wave condition is explicitly used to derive the dispersion relation for Floquet modes. The detailed derivation of the system matrix for the hexagonal lattice is presented in \S \ref{sec3}. The low frequency asymptotics are examined in \S \ref{sec4} where the explicit form of the quasi-static Christoffel matrix is derived. The dynamic and quasi-static solutions are obtained for the rectangular lattice in \S \ref{sec5}, and for cubic lattice in \S \ref{sec6}. In addition, numerical examples in \S \ref{sec5} and \S \ref{sec6} compare results from the present theory with fully elastodynamic FEM computations for hexagonal, rectangular, tetrahedral and cubic lattices.
\section{\rev2{Dispersion relation}} \label{sec2}
\subsection{\rev2{Structures and structural parameters}}
We focus our attention on two example structures in 2D and 3D, hexagonal and tetrahedral lattices, respectively. Each may be defined by two points ${\bf a}_1$, ${\bf a}_2$ inside the unit cell $\mP$ spanned by vectors ${\bf e}_1$, ${\bf e}_2$ (and ${\bf e}_3$ in 3D), see Fig.\ \ref{fig1}. The unit cell $\mP$
is then periodically translated to cover the whole plane (space in 3D) and thereby make the infinitely extended lattice.
We assume that all material parameters are periodic such that the properties in any translated cell $\mP+n{\bf e}_1+m{\bf e}_2\ \ (+l{\bf e}_3\ \ {\text {in 3D}})$ coincide with those in $\mP$.
Every point ${\bf
a}_i$ in the lattice is connected to three (four in 3D) neighboring points ${\bf a}_j$ by rods
$[{\bf a}_i,{\bf a}_j]$ with length $l_{ij}=|{\bf a}_i-{\bf a}_j|$
and direction ${\bf e}_{ij}=l_{ij}^{-1}({\bf a}_i-{\bf a}_j)$.
There are masses $m_1$, $m_2$ with moments of inertia $I_1$, $I_2$ at the points ${\bf a}_1$, ${\bf a}_2$.
The rod $[{\bf a}_i,{\bf a}_j]$ has axial stiffness $\m_{ij}$, beam flexural coefficient $\l_{ij}$ and lineal density $\rho} \def{\omega_{ij}$ (these are related to the rod Young's modulus $E_{ij}$, cross-sectional area $A_{ij}$, radius of gyration $\kappa_{ij}$ and volumetric density $\rho_{ij}^V$ by
$\m_{ij} =E_{ij}A_{ij}$, $\l_{ij} =E_{ij}A_{ij}\kappa_{ij}^2$, $\rho} \def{\omega_{ij} = \rho_{ij}^VA_{ij}$).
\begin{figure}[ht]
\centering
\setlength{\unitlength}{1mm}
\begin{picture}(85,55
\thinlines
\multiput(5,5)(2,0){30}
{\line(1,0){1}}
\multiput(12,19)(2,0){30}
{ \line(1,0){1}}
\multiput(19,33)(2,0){30}
{ \line(1,0){1}}
\multiput(26,47)(2,0){30}
{ \line(1,0){1}}
\multiput(0,0)(0.9,1.8){24}
{\qbezier(5, 5)(5.225, 5.45)(5.45,5.9)}
\multiput(10,0)(0.9,1.8){24}
{\qbezier(15, 5)(15.225, 5.45)(15.45,5.9)}
\multiput(20,0)(0.9,1.8){24}
{\qbezier(25, 5)(25.225, 5.45)(25.45,5.9)}
\multiput(30,0)(0.9,1.8){24}
{\qbezier(35, 5)(35.225, 5.45)(35.45,5.9)}
\thicklines
\multiput(5,5)(20,0){3}
{\qbezier(12, 5)(8, 9)(8,9)}
\multiput(8.5,12)(20,0){3}
{\qbezier(15.5, 12)(11.5, 16)(11.5,16)}
\multiput(12,19)(20,0){3}
{\qbezier(19, 19)(15, 23)(15,23)}
\multiput(-5,5)(20,0){4}
{\qbezier(2, 5)(18, 9)(18,9)}
\multiput(-1.5,12)(20,0){4}
{\qbezier(5.5, 12)(21.5, 16)(21.5,16)}
\multiput(2,19)(20,0){4}
{\qbezier(9, 19)(25, 23)(25,23)}
\multiput(5,5)(20,0){3}
{\qbezier(12, 5)(1, -5)(1,-5)}
\multiput(5,5)(20,0){3}
{\qbezier(8, 9)(19, 19)(19,19)}
\multiput(8.5,12)(20,0){3}
{\qbezier(11.5, 16)(22.5, 26)(22.5,26)}
\multiput(12,19)(20,0){3}
{\qbezier(15, 23)(26, 33)(26,33)}
\put(42,27.5){$\mathbf{a}_2$}
\put(44,21){$\mathbf{a}_1$} %
\put(53,29){$\mP$}
\multiput(17,10)(20,0){3}{\circle*{1.7}}
\multiput(13,14)(20,0){3}{\circle*{1.7}}
\multiput(24,24)(20,0){3}{\circle*{1.7}}
\multiput(20,28)(20,0){3}{\circle*{1.7}}
\multiput(31,38)(20,0){3}{\circle*{1.7}}
\multiput(27,42)(20,0){3}{\circle*{1.7}}
\end{picture}
\caption{The hexagonal (honeycomb) lattice}
{\label{fig1}}
\end{figure}
\begin{figure}[ht]
\centering
\setlength{\unitlength}{0.254mm}
\begin{picture}(290,270)(160,-320)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(275,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-55)(295,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(250,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(280,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-110)(240,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(225,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-75)(210,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(165,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-130)(265,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(250,-185)
\put(210,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\put(285,-131){\shortstack{{\textbf a$_2$}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(225,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(225,-165)
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(165,-95)
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(275,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-55)(295,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(250,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(280,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-110)(240,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(225,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-75)(210,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(165,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-130)(265,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(250,-185)
\put(210,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(225,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(225,-165)
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(165,-95)
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(445,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(385,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-55)(405,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(415,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(360,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(390,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-110)(350,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(335,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-75)(320,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(275,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-130)(375,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(385,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(360,-185)
\put(320,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(335,-165)
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(275,-95)
\put(385,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(445,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(385,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-55)(405,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(415,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(360,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(390,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-110)(350,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(335,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-75)(320,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(275,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-130)(375,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(385,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(360,-185)
\put(320,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(335,-165)
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(275,-95)
\put(385,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(225,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-275)(165,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-165)(295,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(280,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-220)(240,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(250,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(225,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-185)(210,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-240)(265,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(250,-295)
\put(210,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(225,-275)
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(225,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-275)(165,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-165)(295,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(280,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-220)(240,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(250,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(225,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-185)(210,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-240)(265,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(250,-295)
\put(210,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(225,-275)
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(385,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-165)(405,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(415,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(390,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-220)(350,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(360,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(335,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-185)(320,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-240)(375,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(385,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(360,-295)
\put(320,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-275)(335,-275)
\put(445,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(385,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-165)(405,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(415,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(390,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-220)(350,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(360,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(335,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-185)(320,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-240)(375,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(385,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(360,-295)
\put(320,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-275)(335,-275)
\put(445,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(275,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-55)(295,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(250,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(280,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-110)(240,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(225,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-75)(210,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(165,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-130)(265,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(250,-185)
\put(210,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(225,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(225,-165)
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(165,-95)
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(275,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-55)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(165,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-55)(295,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(250,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-90)(280,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-110)(240,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-145)(225,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-75)(210,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(165,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-110)(195,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-130)(265,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(220,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-165)(250,-185)
\put(210,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-95)(225,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(225,-165)
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(165,-95)
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(445,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(385,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-55)(405,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(415,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(360,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(390,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-110)(350,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(335,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-75)(320,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(275,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-130)(375,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(385,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(360,-185)
\put(320,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(335,-165)
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(275,-95)
\put(385,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(445,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(385,-95)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-55)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-55)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-55)(405,-90)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(415,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(360,-75)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-90)(390,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-110)(350,-145)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-145)(335,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-75)(320,-110)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(275,-95)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-110)(305,-130)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-130)(375,-165)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(385,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(330,-150)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-165)(360,-185)
\put(320,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-145){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-130){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-110){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-75){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-90){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-95)(335,-55)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(335,-165)
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-55){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-95)(275,-95)
\put(385,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-95){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(225,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-275)(165,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-165)(295,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(280,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-220)(240,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(250,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(225,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-185)(210,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-240)(265,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(250,-295)
\put(210,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(225,-275)
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(275,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-315)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-165)(225,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(165,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(225,-275)(165,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(335,-165)(295,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(250,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(295,-200)(280,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(280,-220)(240,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(250,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(240,-255)(225,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(250,-185)(210,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(165,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(210,-220)(195,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(305,-240)(265,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(220,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(265,-275)(250,-295)
\put(210,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(220,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(240,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(280,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(250,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(295,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(165,-205)(225,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(225,-275)
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(225,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(165,-205)
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(165,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(265,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(385,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-165)(405,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(415,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(390,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-220)(350,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(360,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(335,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-185)(320,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-240)(375,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(385,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(360,-295)
\put(320,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-275)(335,-275)
\put(445,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(385,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(445,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(385,-205)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-165)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-315)(445,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-165)(335,-275)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(275,-315)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(335,-275)(275,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(445,-165)(405,-200)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(415,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(360,-185)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(405,-200)(390,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(390,-220)(350,-255)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(360,-295)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(350,-255)(335,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(360,-185)(320,-220)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(275,-205)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(320,-220)(305,-240)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(415,-240)(375,-275)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(385,-315)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(330,-260)
\special{color rgb 0 0 0}\allinethickness{0.508mm}\path(375,-275)(360,-295)
\put(320,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(305,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(330,-260){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-295){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(385,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(415,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(390,-220){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-185){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(405,-200){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(445,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-315){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(275,-205)(335,-165)
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(445,-275)(335,-275)
\put(445,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-165){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(385,-205)(275,-205)
\put(385,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(275,-205){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(375,-275){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\put(260,-156){\shortstack{{\textbf a$_1$}}}
\special{color rgb 0 0 0}
\end{picture}
\caption{The tetrahedral lattice}
\label{diamond}
\end{figure}
\subsection{Analytic dispersion relation}
Considering the rod ${\bf e}_{ij}$, let ${\bf u}_i$, ${\bf u}_j$ denote the displacement at the end points ${\bf a}_i$, ${\bf a}_j$, respectively. Let ${\bf f}_{ij}$ denote the force at point ${\bf a}_i$ from the rod ${\bf e}_{ij}$. The precise form of the displacement and force 3-vectors (6-vectors for 3D case) will be defined in Section \ref{sec3}, for the moment we do not need to know their specific nature, except to note that they include both longitudinal and flexural effects. The equilibrium equation at point ${\bf a}_i$ is then
\beq{8} \sum\limits_{j\in \mathcal{N}_i} {\bf f}_{ij} = -\omega^2
{\bf M}_i {\bf u}_i, \quad
{\bf M}_i = \mathrm{diag} ( m_i, m_i, I_i) \ \ \text {in 2D or } \mathrm{diag} ( m_i, m_i, m_i, I_i, I_i, I_i) \ \ \text {in 3D},
\relax
where ${\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_i$ is the set of points connected with ${\bf a}_i$. \reva{It is notable that this approach allows concentrated masses at the junctions which are included in the matrix ${\bf M}_i$.}
The force ${\bf f}_{ij}$ may be expressed in terms of the end point displacements
\beq{17} {\bf f}_{ij} =
{\bf P}_{ij}^{(2)} {\bf u}_j - {\bf P}_{ij}^{(1)} {\bf u}_i ,
\relax
where the frequency dependent stiffness matrices ${\bf P}_{ij}^{(1)} (\omega)$, ${\bf P}_{ij}^{(2)} (\omega)$ are derived in Section \ref{sec3}.
Applying the Floquet periodic conditions
\[\label{221}
\begin{aligned}
{\bf u}_j&=\exp(i{\bf k}\cdot{\bf g}_{j}){\bf u}_1,\ \ {\bf g}_j={\bf
a}_j-{\bf a}_1,\ \ j\in{\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_2 ,
\\
{\bf u}_j&=\exp(i{\bf k}\cdot{\bf g}_{j}){\bf u}_2,\ \ {\bf g}_j={\bf
a}_j-{\bf a}_2,\ \ j\in{\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_1
\end{aligned}
\]
and using eqs.\ \er{8} and \er{17} leads to
\beq{3-3}
\begin{aligned}
\sum\limits_{j\in
\mathcal{N}_1} \big(
{\bf P}_{1j}^{(2)} \exp(i{\bf k}\cdot{\bf g}_{j}) \, {\bf u}_2
- {\bf P}_{1j}^{(1)} {\bf u}_1 \big)
&= -\omega^2 {\bf M}_1 {\bf
u}_1 ,
\\
\sum\limits_{j\in \mathcal{N}_2} \big(
{\bf P}_{2j}^{(2)} \exp(i{\bf k}\cdot{\bf g}_{j}) \, {\bf u}_1
- {\bf P}_{2j}^{(1)} {\bf u}_2 \big)
&= -\omega^2 {\bf M}_2 {\bf
u}_2.
\end{aligned}
\relax
For each $j\in \mathcal{N}_2$ there is a unique $\bar j\in \mathcal{N}_1$ such that
\beq{012}
{\bf P}_{2j}^{(2)} e^{i {\bf k}\cdot{\bf g}_j}
=
\Big( {\bf P}_{1\bar j}^{(2)} e^{i {\bf k}\cdot{\bf g}_{\bar j}}
\Big)^{+},
\ \
{\bf P}_{2j}^{(1)} = {\bf P}_{1\bar j}^{(3)},
\ \
\relax
where $+$ denotes the Hermitian conjugation and the matrices $ {\bf P}_{1j}^{(3)}$ are defined in Section \ref{sec3}. Hence it is possible to express the second equation of \eqref{3-3} in terms of a sum over neighboring links of ${\bf a}_1$.
Introducing matrices
\[\label{223}
{\bf H}_1=\sum\limits_{j\in\mathcal{N}_1}
{\bf P}_{1j}^{(1)},\ \ {\bf H}_2=-\sum\limits_{j\in\mathcal{N}_1}
{\bf P}_{1j}^{(2)}\exp(i{\bf k}\cdot{\bf g}_{j})
,\ \ {\bf H}_3=\sum\limits_{j\in\mathcal{N}_1}{\bf P}_{1j}^{(3)}
\]
equations \er{3-3} can then be rewritten in the form
\[\label{224}
{\bf H}{\bf u}=\omega^2{\bf M}{\bf u}
\]
with
\[\label{225}
{\bf u}=\begin{pmatrix} {\bf u}_1 \\ {\bf u}_2 \end{pmatrix},\ \ {\bf M}=\mathrm{diag}({\bf M}_1,{\bf
M}_2),\ \ {\bf H}\equiv{\bf H}(\omega,{\bf k})=\begin{pmatrix} {\bf H}_1 & {\bf H}_2 \\ {\bf H}_2^+ & {\bf
H}_3 \end{pmatrix} \ \ \big( = {\bf H}^+\big).
\]
Then Floquet curves (dispersion curves) $\omega_n({\bf k})$ can be found
from the equation
\[\label{226}
\det({\bf H}(\omega,{\bf k})-\omega^2{\bf M})=0.
\]
Note that according to Section \ref{sec3} (see eqs.\ \eqref{223}, \eqref{10}, \eqref{003} and \eqref{3}), the matrices ${\bf H}_1$ and ${\bf H}_3$ are real symmetric, so that the matrix $ {\bf H}$
is Hermitian, in turn guaranteeing that the dispersion relation
\er{226} is real valued for real $\omega$, ${\bf k}$. We will return to
this equation in Section \ref{sec3} after we have described the
displacements and forces, and derived the stiffness matrices.
\section{Dynamic stiffness matrices} \label{sec3}
\subsection{Longitudinal wave equation} Consider the rod ${\bf
e}_{ij}$ with uniform Young's modulus $\mu_{ij}$ and density $\rho} \def{\omega_{ij}$.
Let $u_{ij}(x)$ denote the component of the displacement in the ${\bf e}_{ij}-$direction at any point $x$ (a one-dimensional linear coordinate parameter) of $[{\bf a}_i,{\bf a}_j]$. The displacement $u_{ij}$ satisfies the wave equation for longitudinal wave motion and its associated boundary
conditions (BCs)
\[\label{201}
\m_{ij}\frac{\partial^2 }{\partial x^2}u_{ij}=-\omega^2\rho} \def{\omega_{ij}u_{ij},\ \ u_{ij}(0)={\bf
e}_{ij}\cdot{\bf u}_{i},\ \ u_{ij}(l_{ij})={\bf e}_{ij}\cdot{\bf
u}_j.
\]
Solving \er{201},
\[\label{203}
u_{ij}(x)=\frac{
{\bf e}_{ij}\cdot{\bf u}_{i}\sin(s_{ij}\omega (l_{ij}-x))
+
{\bf e}_{ij}\cdot{\bf u}_{j} \sin(s_{ij}\omega x)}{\sin(s_{ij}\omega l_{ij})},
\ \ s_{ij}=\sqrt{\frac{\rho} \def{\omega_{ij}}{\m_{ij}}},
\]
implying that the longitudinal force ${\bf f}_{ij}$ acting on the point ${\bf a}_i$ is
\[\label{204}
{\bf f}_{ij\ wave} \equiv
\m_{ij}\frac{\partial u_{ij} }{\partial x}(0)\, {\bf e}_{ij}
=\frac{ \m_{ij}s_{ij}\omega}{\sin(s_{ij}\omega l_{ij})}\,
{\bf e}_{ij}{\bf e}_{ij}^{T}
\big( {\bf u}_{j}-{\bf
u}_{i}\cos(s_{ij}\omega l_{ij})\big)
.
\]
\subsection{Flexural wave equation}
\rev2{The kinematic BCs for flexural wave motion involve both the flexural displacement and the non-torsional rotation at the ends of the rod. In 2D, define the unit vector perpendicular to the plane of the lattice, ${\bf e}_b=
{\bf e}_1\wedge {\bf e}_2 /|{\bf e}_1\wedge {\bf e}_2|$.
The flexural displacement $v_{ij}(x)$ at any point $x$ on the rod ${\bf e}_{ij}$ is then defined as the component of the displacement in the ${\bf e}_{ij}^\perp-$ direction,
where
}
\beq{6}
{\bf e}_{ij}^\perp = { {\bf e}_{b} \wedge {\bf e}_{ij} }.
\relax
\rev2{The generalized 2D displacement vectors are therefore ``three-dimensional" with two components for the longitudinal motion and one for flexural.
The flexural wave equation and its BCs are, with $v ' = \partial v /\partial x$, }
\bal{4}
& - \l_{ij} \frac{\partial^4
v_{ij}}{\partial x^4} =- \omega^2 \rho} \def{\omega_{ij} v_{ij} , \quad
\\
v_{ij}(0)= {\bf e}_{ij}^\perp \cdot {\bf u}_i, \ \ &
v_{ij}(l_{ij})= {\bf e}_{ij}^\perp \cdot{\bf u}_j, \ \ v_{ij}'(0)=
{\bf e}_b \cdot{\bf u}_i, \ \ v_{ij}'(l_{ij})= {\bf e}_b\cdot {\bf
u}_j . \nonumber \eal
The generalized force (shear force and bending moment) at point ${\bf a}_i$ due to bending is \beq{7} {\bf f}_{ij\ bending} = -\lambda_{ij}
\frac{\partial^3 v_{ij}}{\partial x^3} {\bf e}_{ij}^\perp +
\lambda_{ij} \frac{\partial^2 v_{ij}}{\partial x^2} {\bf e}_b. \relax
\rev2{In 3D we extend the definition of the end point flexural displacement by defining two non-torsional rotation components in the directions ${\bf e}_b$ and ${\bf e}_b^\prime$, and the related shear force components along
${\bf e}_{ij}^\perp $ and ${\bf e}_{ij}^{\perp \prime}$, where }
\beq{new perp}
{\bf e}_{ij}^\perp = ({\bf r}, {\bf 0}_3),\ \ {\bf e}_b = ({\bf 0}_3, {\bf e}_{ij} \wedge {\bf r}),\ \ {\bf e}_{ij}^{\perp \prime} = ({\bf e}_{ij} \wedge {\bf r}, {\bf 0}_3),\ \ {\bf e}_b^\prime = -({\bf 0}_3, {\bf r}) .
\relax
Here ${\bf 0}_3$ is 3D zero-vector, ${\bf r} = {\bf e}_{\alpha} \wedge {\bf e}_{ij}/|{\bf e}_{\alpha} \wedge {\bf e}_{ij}|$ and ${\bf e}_\alpha$ is some (any) vector ${\bf e}_1$ or ${\bf e}_2$ or ${\bf e}_3$ whichever is not parallel to ${\bf e}_{ij}$.
\rev2{Let $w_{ij}(x)$ denote the displacement in direction ${\bf e}_{ij}^{\perp \prime}$, $w_{ij} ' = \partial w_{ij} /\partial x$,
then the flexural wave equation and its BCs for the extra dimension in 3D case can be written as eq.\ \eqref{4} combined with }
\bal{42}
& - \l_{ij} \frac{\partial^4
w_{ij}}{\partial x^4} =- \omega^2 \rho} \def{\omega_{ij} w_{ij} , \quad
\\
w_{ij}(0)= {\bf e}_{ij}^{\perp\prime} \cdot {\bf u}_i, \ \ &
w_{ij}(l_{ij})= {\bf e}_{ij}^{\perp\prime} \cdot{\bf u}_j, \ \ w_{ij}'(0)=
{\bf e}_b^\prime \cdot{\bf u}_i, \ \ w_{ij}'(l_{ij})= {\bf e}_b^\prime \cdot {\bf
u}_j . \nonumber \eal
The additional generalized force term at point ${\bf a}_i$ is
\beq{7_2} {\bf f}^{\prime}_{ij\ bending} = -\lambda_{ij}
\frac{\partial^3 w_{ij}}{\partial x^3} {\bf e}_{ij}^{\perp \prime} +
\lambda_{ij} \frac{\partial^2 w_{ij}}{\partial x^2} {\bf e}_b^\prime. \relax
\rev2{In summary, the components of $ {\bf u}_i$
in the ${\bf e}_{ij}^\perp$, ${\bf e}_{ij}^{\perp\prime}$ and ${\bf e}_b$, ${\bf e}_b^\prime$ directions are the transverse deflection and beam rotation angle, respectively.
The force components in direction ${\bf e}_{ij}^\perp$ and ${\bf e}_{ij}^{\perp\prime}$ are the resultant shear force while the
${\bf e}_b$ and ${\bf e}_b^\prime$ "force" components represent the bending moment. Note that in 3D, there are three displacement components and three rotation components.
In this way the coupled longitudinal and flexural dynamics of the 2D lattice are described in terms of "three-dimensional"vectors for displacement and forces in 2D, and "six-dimensional" vectors for 3D lattices.}
The generalized forces at the two ends of the rod are related to the
displacements there by the stiffness matrix ${\bf K}$, defined such that
\bal{003}
\begin{pmatrix} {\bf e}_{ij}^{\perp}\cdot {\bf f}_{ij} \\ {\bf e}_b\cdot{\bf f}_{ij}\\ {\bf e}_{ij}^{\perp}\cdot {\bf f}_{ji} \\ {\bf e}_b\cdot{\bf f}_{ji} \end{pmatrix}
= -\l_{ij}
{\bf K}(\omega)
\begin{pmatrix} {\bf e}_{ij}^{\perp}\cdot {\bf u}_i \\ {\bf e}_b\cdot{\bf u}_i \\ {\bf e}_{ij}^{\perp}\cdot {\bf u}_j \\ {\bf e}_b\cdot{\bf u}_j \end{pmatrix}
,
\quad {\bf K} =
\begin{pmatrix}
{\bf K}_1 & {\bf K}_2
\\
{\bf K}_2^T & {\bf K}_3
\end{pmatrix}.
\eal
The bending forces \eqref{7} and \eqref{7_2} at lattice site $i$ from rod $ij$ therefore becomes
\beq{7a}
\begin{aligned}
&{\bf f}_{ij\ bending} = -\lambda_{ij}\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big)\Big( {\bf
K}_1
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big)^{T}{\bf u}_i+ {\bf
K}_2
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big)^{T}{\bf u}_j\Big),
\\
&{\bf f}^{\prime}_{ij\ bending} = -\lambda_{ij}\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big)\Big( {\bf
K}_1
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big)^{T}{\bf u}_i+ {\bf
K}_2
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big)^{T}{\bf u}_j\Big).
\end{aligned}
\relax
We next derive the explicit form of the stiffness matrix.
\subsection{Solution of the flexural stiffness matrix}
With eqs.\ \er{4} and \er{42} in mind, consider the
solution to
\beq{1} \frac{\partial^4 w}{\partial x^4} - \gamma^4
w=0, \quad x\in[0,l], \relax in the form \bal{2} &w(x) = \frac 1{2(1-
c c_h)} \big\{ \big[ ( c -c_h )(\cos \gamma x-\cosh \gamma x) + (s+
s_h)(\sin \gamma x -\sinh \gamma x )\big] w(l)
\nonumber \\
& + \frac 1{\gamma}\big[ ( s_h-s )(\cos \gamma x-\cosh \gamma x) +
(c-c_h)(\sin \gamma x -\sinh \gamma x )\big] w'(l)
\nonumber \\
& + \big[ (1-c c_h +s s_h)\cos \gamma x +(1-c c_h -s s_h)\cosh
\gamma x + (c s_h+s c_h)(\sinh \gamma x -\sin \gamma x )\big] w(0)
\nonumber \\
& + \frac 1{\gamma} \big[ (s c_h-c s_h)(\cos \gamma x -\cosh \gamma
x )+ (1-c c_h -s s_h)\sin \gamma x +(1-c c_h +s s_h)\sinh \gamma x
\big] w'(0) \big\},
\ea
where $c=\cos \gamma l$, $s=\sin \gamma l$, $c_h=\cosh \gamma l$,
$s_h=\sinh \gamma l$. $v(x)$ and $w(x)$ have the same form of solution, so that the stiffness matrix is the same. According to its definition in \eqref{003} the stiffness matrix ${\bf K}$ satisfies
\bal{3=5}
\begin{pmatrix}
w'''(0)
\\
-w''(0)
\\
-w'''(l)
\\
w''(l)
\end{pmatrix}
&= {\bf K}(\omega)
\begin{pmatrix}
w(0)
\\
w'(0)
\\
w(l)
\\
w'(l)
\end{pmatrix} .
\eal
The explicit form of the stiffness matrix then follows from \eqref{2} as
\bal{3}
{\bf K}(\omega)
= \frac {\gamma^2}{1- c c_h}
\begin{pmatrix}
\gamma(cs_h+sc_h) & ss_h & -\gamma (s+s_h) &
c_h-c
\\
s s_h & \gamma^{-1} (sc_h - c s_h ) & c-c_h & \gamma^{-1}
(s_h -s)
\\
-\gamma (s+s_h) & c- c_h & \gamma(cs_h+sc_h) &
- ss_h
\\
c_h-c & \gamma^{-1} (s_h-s) & - s s_h & \gamma^{-1}
(sc_h-cs_h)
\end{pmatrix}.
\eal
\subsection{Total force and stiffness matrices }
The total force at point $i$ from rod ${\bf e}_{ij}$ now follows from
\eqref{204} and \eqref{7},
\beq{71} {\bf f}_{ij
} = {\bf f}_{ij\ wave}(0) + {\bf f}_{ij\ bending}(0) + {\bf f}^{\prime}_{ij\ bending}(0),
\relax
where ${\bf f}^{\prime}_{ij\ bending}(0)$ doesn't exist in 2D case. Set
\bal{23} &\tilde{\mu}_{ij} = {\mu}_{ij}/l_{ij}, \quad
\tilde{s}_{ij} (\omega) = \omega s_{ij} l_{ij}, \quad
\gamma_{ij} (\omega) =\big(\omega^2 \rho} \def{\omega_{ij}/\l_{ij} \big)^{1/4} ,\quad
{\bf A}_{ij} = {\bf e}_{ij}
{\bf e}_{ij}^T,
\eal
The dynamic stiffness matrices introduced in eqs.\ \er{17} and \er{012} then follow from
\er{204}, \er{7a} and \er{71} as
\bal{10} {\bf P}_{ij}^{(1)} & =
\tilde\mu_{ij} \tilde s_{ij}\cot \tilde s_{ij} {\bf
A}_{ij} +\lambda_{ij}
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big) {\bf K}_1
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big)^{T} +\lambda_{ij}
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big) {\bf K}_1
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big)^{T} ,
\nonumber \\
{\bf P}_{ij}^{(2)}&= \tilde\mu_{ij}\tilde s_{ij} \csc \tilde
s_{ij} {\bf A}_{ij} - \lambda_{ij}
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big) {\bf K}_2
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big)^{T} - \lambda_{ij}
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big) {\bf K}_2
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big)^{T} ,
\\
{\bf P}_{ij}^{(3)} & = \tilde\mu_{ij} \tilde s_{ij}\cot
\tilde s_{ij} {\bf A}_{ij} +\lambda_{ij}
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big) {\bf K}_3
\big( {\bf e}_{ij}^\perp , \, {\bf e}_b\big)^{T} +\lambda_{ij}
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big) {\bf K}_3
\big( {\bf e}_{ij}^{\perp \prime} , \, {\bf e}_b^\prime \big)^{T}. \nonumber \eal
where ${\bf K}$ is defined by eq.\ \eqref{3} with
\beq{-12}
\gamma = \gamma_{ij} , \ \
c=\cos \gamma_{ij}l_{ij},\ \ s=\sin \gamma_{ij}l_{ij},\ \ c_h=\cosh \gamma_{ij}l_{ij},\ \ s_h=\sinh \gamma_{ij}l_{ij} .
\relax
The identities \eqref{012} are a consequence of the relations
${\bf K}_2^T = {\bf J}{\bf K}_2{\bf J}$, ${\bf K}_3 = {\bf J}{\bf K}_1{\bf J}$
where ${\bf J} = \mathrm{diag}(1,-1)$.
The force ${\bf f}_{ij}$ at point ${\bf a}_i$ given by eq.\ \eqref{17}
then follows from \eqref{204}, \eqref{7a} and \eqref{71}.
\section{Effective wave speeds at low frequency} \label{sec4}
\subsection{Low frequency asymptotics}
The low-frequency asymptotic behavior of ${\bf K}$ defined in
\er{3} is, using ${\bf K}_3 = {\bf J}{\bf K}_1{\bf J}$,
\beq{45}
\begin{aligned}
{\bf K}_1 &= l^{-2}
\begin{pmatrix} {12}\,l^{-1} & 6
\\
6 & {4}\,{l}
\end{pmatrix} + \frac{\gamma^4 l^2}{35}
\begin{pmatrix}
-13\,l^{-1} & -\frac{11}6
\\
-\frac{11}6 & -\frac{l}3
\end{pmatrix} + \text{O}( \gamma^8) ,
\\
{\bf K}_2 &= l^{-2}\begin{pmatrix} -{12}\,l^{-1} & 6
\\
-6 & 2\,l
\end{pmatrix} + \frac{\gamma^4 l^2}{70}
\begin{pmatrix}
-9\,l^{-1} & \frac{13}6
\\
-\frac{13}6 & \frac{l}2
\end{pmatrix} + \text{O}( \gamma^8) ,
\end{aligned}
\relax
implying that ${\bf K}(0)$ is positive semi-definite
having eigenvalues $30$ and $2$ with non-normalized eigenvectors
$(2,1,-2,1)^T$ and $(0,1,0,-1)^T$, respectively. The null vectors of
${\bf K}(0)$, $(1,0,1,0)^T$ and $(-l,2,l,2)^T$, correspond to rigid
body displacement and rotation, respectively.
The low frequency expansions of the dynamic stiffness matrices of eq. \eqref{10} are
\beq{11}
\begin{aligned}
{\bf P}_{ij}^{\big(\stackrel{1}{3} \big)} (\omega )=& \tilde\mu_{ij}{\bf
A}_{ij} + 2
\lambda_{ij} l_{ij}^{-3} \Big(
6({\bf A}_{ij}^{\perp }+{\bf A}_{ij}^{\perp \prime})
\pm 3({\bf A}_{ij}^{b\perp }+{\bf A}_{ij}^{b\perp \prime}
+{\bf A}_{ij}^{\perp b }+{\bf A}_{ij}^{\perp b \prime})l_{ij}
+ 2 ({\bf A}_b+{\bf A}_b^\prime)l_{ij}^2
\Big)
\\
& - \frac 13 \omega^2 \rho} \def{\omega_{ij} l_{ij} \Big( {\bf A}_{ij} +
\frac{1}{70} \big(
78({\bf A}_{ij}^{\perp }+{\bf A}_{ij}^{\perp \prime})
\pm 11({\bf A}_{ij}^{\perp b }+{\bf A}_{ij}^{\perp b \prime}
+{\bf A}_{ij}^{b\perp }+{\bf A}_{ij}^{b\perp \prime})l_{ij}\\
& + 2
({\bf A}_b+{\bf A}_b^\prime)l_{ij}^2\big) \Big) +\text{O}(\omega^4),
\\
{\bf P}_{ij}^{(2)} (\omega )=& \tilde\mu_{ij}{\bf A}_{ij} + 2\lambda_{ij}l_{ij}^{-3}\Big(
6({\bf A}_{ij}^{\perp }+{\bf A}_{ij}^{\perp \prime})
+ 3({\bf A}_{ij}^{b\perp }+{\bf A}_{ij}^{b\perp \prime}
- {\bf A}_{ij}^{\perp b }-{\bf A}_{ij}^{\perp b \prime})l_{ij}
- ({\bf A}_b+{\bf A}_b^\prime)l_{ij}^2
\Big)
\\ & + \frac 16 \omega^2 \rho} \def{\omega_{ij} l_{ij} \Big( {\bf A}_{ij} +
\frac{1}{70} \big(
54({\bf A}_{ij}^{\perp }+{\bf A}_{ij}^{\perp \prime})
+ 13({\bf A}_{ij}^{b\perp }+{\bf A}_{ij}^{b\perp \prime}
-{\bf A}_{ij}^{\perp b }-{\bf A}_{ij}^{\perp b \prime} )l_{ij}
\\
& - 3
({\bf A}_b+{\bf A}_b^\prime)l_{ij}^2 \big) \Big) +\text{O}(\omega^4) ,
\end{aligned}
\relax
where
\bal{62}
&{\bf A}_{ij}^\perp = {\bf e}_{ij}^\perp {{\bf
e}_{ij}^\perp }^T, \ \ \quad {\bf A}_{ij}^{\perp b} = {\bf
e}_{ij}^\perp {\bf e}_b^T, \ \ \quad {\bf A}_{ij}^{b\perp } =
{\bf e}_b {{\bf e}_{ij}^\perp }^T, \ \ \ \ {\bf A}_b = {\bf e}_b
{\bf e}_b^T , \\
&{\bf A}_{ij}^{\perp \prime} = {\bf e}_{ij}^{\perp \prime} {{\bf
e}_{ij}^{\perp \prime} }^T, \quad {\bf A}_{ij}^{\perp b \prime} = {\bf
e}_{ij}^{\perp \prime} {{\bf e}_b^{\prime}}^T, \quad {\bf A}_{ij}^{b\perp \prime} =
{\bf e}_b^\prime {{\bf e}_{ij}^{\perp \prime} }^T, \quad {\bf A}_b^\prime = {\bf e}_b^\prime
{{\bf e}_b^\prime}^T . \label{62_2}
\eal
Note that the terms with the primes are not present for the 2D lattice, and hence eq.\ \eqref{62_2} applies only for the 3D case.
The zero frequency limit of the system matrix ${\bf H}$ defined in eq.\ \eqref{225} has the following form
\[\label{400}
{\bf H}^{(0)}\equiv {\bf H}(0,{\bf 0}) =\begin{pmatrix} {\bf H}_+^{(0)} & -{\bf H}_+^{(0)} \\ -{\bf H}_-^{(0)} & {\bf H}_-^{(0)} \end{pmatrix}
+\begin{pmatrix} {\bf 0} & {\bf R}_+ \\ {\bf R}_- & {\bf 0} \end{pmatrix}
\]
with
\beq{401}
\begin{aligned}
{\bf H}_\pm ^{(0)}&=\sum_{j\in{\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_1}\Big(
\tilde\mu_{1j}{\bf A}_{1j} + 2 \lambda_{1j}l_{1j}^{-3} \big(
6({\bf A}_{ij}^{\perp }+{\bf A}_{ij}^{\perp \prime})
\pm 3({\bf A}_{ij}^{b\perp }+{\bf A}_{ij}^{b\perp \prime}
+{\bf A}_{ij}^{\perp b }+{\bf A}_{ij}^{\perp b \prime})l_{ij}
+ 2 ({\bf A}_b+{\bf A}_b^\prime)l_{ij}^2
\big)
\Big),
\\
{\bf R}_\pm &=\sum_{j\in{\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_1}6\l_{1j}l_{1j}^{-3}\Big( ({\bf A}_b+{\bf A}_b^\prime )l_{ij}^2 \pm 2({\bf A}_{1j}^{\perp b}+{\bf A}_{1j}^{\perp b \prime})l_{ij} \Big).
\end{aligned}
\relax
The effective quasi-static
speeds are defined as
\[\label{408}
c(\pmb{\k})=\lim_{k\to0}\frac{\omega({\bf k})}{k},\ \ {\bf k}=k\pmb{\k},\ \ |\pmb{\k}|=1.
\]
We consider the following perturbation {\it ansatz} for small
${\bf k}$,
\[\label{410}
\omega^2({\bf k})= k\omega_1+
k^2\omega_2+O(k^3)
\]
with associated displacement
\[\label{411}
{\bf u}({\bf k})={\bf u}^0+k{\bf u}^{1}+k^2{\bf u}^2+{\bf O}(k^3) .
\]
The asymptotic behavior of ${\bf H}$ for small $\omega$ and $k$ is
\[\label{409}
{\bf H}(\omega,{\bf k})={\bf H}^{(0)}+k{\bf H}^{(1)}(\pmb{\k})+
k^2{\bf H}^{(2)}(\pmb{\k})+\omega^2{\bf H}^{(3)}+{\bf O}(\omega^4)+{\bf O}(k\omega^2)+{\bf O}(k^3).
\]
Substituting \er{411}-\er{409} into \er{224} and identifying terms with
the same power of $k$,
yields at O$(k)$
\[\label{412}\
{\bf H}^{(0)} {\bf u}^1 +{\bf H}^{(1)}{\bf u}^0
= \omega_1 ( {\bf M}-{\bf H}^{(3)}){\bf u}^0 .
\]
The matrix ${\bf H}^{(3)}$ follows from eqs.\ \eqref{223}, \eqref{225}, \eqref{11} and
\eqref{409}.
The subsequent general analysis applies only to the 2D lattice for which the vectors ${\bf u}^0$, ${\bf u}^1$, ${\bf u}^2$ are 6-dimensional. The analogous derivation for the 3D case, which involves 12-dimensional vectors, is not considered here, although we note that some explicit low frequency asymptotic results are given in \S\ref{sec6}.
\subsection{Effective speeds in 2D lattices
Consider the equation
\[\label{403}
{\bf H}^{(0)}\begin{pmatrix} {\bf u}_1 \\ {\bf u}_2 \end{pmatrix}={\bf 0}.
\]
Since ${\bf H}_\pm^{(0)}= ({\bf H}_\pm^{(0)})^+>0$ then it is not difficult to show that the
solution of \er{403} satisfies
\[\label{404}
{\bf u}_1={\bf u}_2,\ \ {\bf u}_1\perp{\bf e}_b.
\]
Based on eqs.\ \eqref{403} and \eqref{404}
we obtain the following result:
The dimension of $\mathop{\mathrm{Ker}}\nolimits{\bf H}^{(0)}$ \er{403} is equal to $2$ and the basis can
be chosen as
\[\label{406}
{\bf u}^{01}=2^{-\frac12}\begin{pmatrix} {\bf e}^{01} \\ {\bf e}^{01} \end{pmatrix},\ \
{\bf u}^{02}=2^{-\frac12}\begin{pmatrix} {\bf e}^{02} \\ {\bf e}^{02}\end{pmatrix},\ \ {\bf
e}^{01}=\ma1 \\ 0 \\ 0\end{pmatrix},\ \ {\bf
e}^{02}=\ma0 \\ 1 \\ 0\end{pmatrix}.
\]
In summary, ${\bf H}^{(0)}$ possesses an eigenvalue $\omega=0$ with multiplicity $2$. We next obtain the equation that determines the associated pair of wave speeds.
Using the properties derived previously for ${\bf H}^{(0)}$ of \eqref{400} it follows that the leading order displacement ${\bf u}^0$ is spanned by $\{{\bf u}^{01}, {\bf u}^{02}\}$, see eq.\ \eqref{406}.
Using the identity
${\bf A}_{ij} + {\bf A}_{ij}^{\perp} = \mathrm{diag} (1,1,0)$ for any pair $ij$, it follows that
\[ \label{4-4}
(({\bf M}-{\bf
H}^{(3)}){\bf u}^{0i}\cdot{\bf u}^{0j})_{i,j=1}^2 =
\frac m2 \mathrm{diag} (1,1)
\]
where
$m$ is the total mass per unit cell,
\beq{467}
m = m_1+m_2 + \sum\limits_{j\in\mathcal{N}_1}
\rho} \def{\omega_{1j} l_{1j} .
\relax
The appearance of the total mass is significant, bearing in mind that dynamic lattice models which do not include both flexural and longitudinal waves are known to produce quasistatic wave speeds with effective mass less than the total mass of the unit cell \cite{Colquitt11}.
Scalar multiplying \er{4-4} by ${\bf u}^0$
and using ${\bf H}^{(1)}{\bf u}^0\cdot{\bf u}^0=0$
implies
\[\label{4-3}
\omega_1 = 0,\ \
{\bf u}^1= - ({\bf H}^{(0)})^{-1} {\bf H}^{(1)}{\bf u}^0
\]
where $({\bf H}^{(0)})^{-1}$ is uniquely defined acting on the subspace orthogonal to $\spn \{{\bf u}^{01},{\bf u}^{02}\}$.
At O$(k^2)$ we have
\[\label{413}
\omega_2({\bf M}-{\bf H}^{(3)}){\bf u}^0\cdot{\bf u}^{0j} = {\bf H}^{(2)}{\bf u}^0\cdot{\bf
u}^{0j}+ {\bf H}^{(1)}{\bf u}^1\cdot{\bf u}^{0j} , \ \ j=1,2.
\]
Hence we
deduce that the squares of the effective speeds $c_{\rm eff}^2 =\omega_2$ are
eigenvalues of the following $2\ts2$ matrix
\beq{415}
{\bf C}_{\rm eff}^2= \frac 2m
\big\{
({\bf H}^{(2)}{\bf u}^{0i}\cdot{\bf u}^{0j})_{i,j=1}^2-
(({\bf H}^{(0)})^{-1}{\bf H}^{(1)}{\bf u}^{0i}\cdot{\bf H}^{(1)}{\bf
u}^{0j})_{i,j=1}^2 \big\}
\relax
where $m$ defined in \er{467} is the total mass per unit cell, ${\bf H}^{(0)}$ is given in \er{400} and
\beq{4-9}
\begin{aligned}
{\bf H}^{(1)} &= \begin{pmatrix} {\bf 0} & {\bf A} \\ {\bf A}^+ & {\bf 0} \end{pmatrix} ,
\ \
{\bf A} = -i
\sum\limits_{j\in\mathcal{N}_1}
{\bf P}_{1j}^{(2)} (0) \, ({\bf g}_{j}\cdot \boldsymbol\kappa),
\\
{\bf H}^{(2)} &= \begin{pmatrix} {\bf 0} & {\bf B} \\ {\bf B}^+ & {\bf 0} \end{pmatrix} ,
\ \
{\bf B} = \frac 12
\sum\limits_{j\in\mathcal{N}_1}
{\bf P}_{1j}^{(2)} (0) \, ({\bf g}_{j}\cdot \boldsymbol\kappa)^2 .
\end{aligned}
\relax
The expression \er{415} can be simplified as follows, with ${\bf I}_{2,3}=\begin{pmatrix} {\bf e}^{01} & {\bf e}^{02}
\end{pmatrix}$,
\[\label{effn1}
{\bf C}_{\rm eff}^2= \frac 1m{\bf I}_{2,3}^{T}({\bf B}+{\bf B}^+-2{\bf A}^+(2{\bf H}_+^{(0)}+{\bf R}_+-2{\bf H}^{(0)}_+{\bf A}_b)^{-1}{\bf A}){\bf
I}_{2,3} .
\]
Introducing the matrices
\beq{effn2}
\begin{aligned}
{\bf B}_1 &=\sum_{j\in{\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_1}\big(
\tilde\mu_{1j}{\bf A}_{1j} + 12 \lambda_{1j}l_{1j}^{-3}
{\bf A}_{1j}^{\perp }\big)({\bf g}_{j}\cdot \boldsymbol\kappa)^2,
\\
{\bf B}_2 & = \sum_{j\in{\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_1}\big(
\tilde\mu_{1j}{\bf A}_{1j} + 6 \lambda_{1j}l_{1j}^{-3} (
2{\bf A}_{1j}^{\perp }
+ {\bf A}_{1j}^{b\perp }) \big)({\bf g}_{j}\cdot \boldsymbol\kappa),
\\
{\bf B}_3 &=\sum_{j\in{\mathcal N}} \def\mP{{\mathscr P}} \def\C{{\mathbb C}_1}\big(
\tilde\mu_{1j}{\bf A}_{1j} + 3 \lambda_{1j}l_{1j}^{-3} (
4{\bf A}_{1j}^{\perp }
+ 2{\bf A}_{1j}^{\perp b } + 2{\bf A}_{1j}^{b\perp } +
{\bf A}_b ) \big),
\end{aligned}
\relax
we can rewrite \er{effn1} succinctly as
\[\label{effn5}
{\bf C}_{\rm eff}^2= \frac 1m{\bf I}^{T}_{2,3}({\bf B}_1
-{\bf B}_2^{T}{\bf B}_3^{-1}{\bf B}_2){\bf I}_{2,3}.
\]
\section{2D Examples} \label{sec5}
\subsection{Rectangular lattice}
\begin{figure}[h!]
\begin{center}
\setlength{\unitlength}{0.12mm}
\begin{picture}(229,236)(120,-298)
\put(127,-182){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\special{color rgb 1 0 0}\allinethickness{0.254mm}\path(135,-183)(230,-183)
\special{color rgb 1 0 0}\allinethickness{0.254mm}\path(240,-183)(335,-183)
\put(341,-183){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(232,-83)(232,-178)
\put(231,-75){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\put(232,-183){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\put(231,-290){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\special{color rgb 0 0 1}\allinethickness{0.254mm}\path(232,-190)(232,-285)
\special{color rgb 0 0 0}\put(250,-160){\makebox(0,0)[cc]{\shortstack{{\Large a$_0$}}}}
\special{color rgb 0 0 0}\put(336,-164){\makebox(0,0)[cc]{\shortstack{{\Large a$_1$}}}}
\special{color rgb 0 0 0}\put(203,-92){\makebox(0,15)[cc]{\shortstack{{\Large a$_2$}}}}
\special{color rgb 0 0 0}\put(122,-170){\makebox(0,15)[cc]{\shortstack{{\Large a$_3$}}}}
\special{color rgb 0 0 0}\put(197,-290){\makebox(0,15)[cc]{\shortstack{{\Large a$_4$}}}}
\special{color rgb 0 0 0}
\end{picture}
\end{center}
\caption{The rectangular lattice cell with mass at ${\bf a}_0$, showing the neighboring masses in the adjoining cells. }
{\label{fig2}}
\end{figure}
\subsubsection{Dispersion relation}
The unit cell for the rectangular lattice, shown in Fig.\ \ref{fig2}, possesses a mass at the central node. Enforcing the equilibrium condition at the single mass and the Bloch-Floquet condition, it may be shown that the equations of motion for this structure reduce to
\begin{equation} \label{589213}
\sum\limits_{j=1,2,3,4}\big( {\bf P}_{0j}^{(1)} - {\bf P}_{0j}^{(2)}{\bf e}^{i{\bf k}\cdot{\bf g}_j} \big){\bf u}_0 = \omega^2 {\bf M}_0 {\bf u}_0,
\ \ {\bf M}_0 = \mathrm{diag}(m_0,m_0,I_0).
\end{equation}
The derivation is entirely similar to that for the hexagonal lattice in Sections \ref{sec2} and \ref{sec3}, with the same notation employed.
We assume the members are of two types: $1$ for horizontal, and $2$ for vertical members,
with parameters denoted by $\rho_j, {\bf K}^{(j)}$, etc.\ $j=1,2$. Then it may be shown that eq.\
\eqref{589213} becomes
\beq{-5}
\begin{pmatrix}
\begin{matrix}
\tilde{\mu}_1 \tilde{s}_1(\cot\tilde{s}_1 -\csc\tilde{s}_1 \cos \tilde{k}_x)
\\
+
\l_2 (K^{(2)}_{11}+K^{(2)}_{13}\cos \tilde{k}_y)
\end{matrix}
& 0 & i\l_2K^{(2)}_{14}\sin \tilde{k}_y
\\ 0 &
\begin{matrix}
\tilde{\mu}_2 \tilde{s}_2(\cot\tilde{s}_2 -\csc\tilde{s}_2 \cos \tilde{k}_y )
\\
+\l_1 (K^{(1)}_{11}+K^{(1)}_{13}\cos \tilde{k}_x)
\end{matrix}
& -i\l_1 K^{(1)}_{14}\sin \tilde{k}_x
\\ -i\l_2 K^{(2)}_{14}\sin \tilde{k}_y & i\l_1 K^{(1)}_{14}\sin \tilde{k}_x &
\begin{matrix}
\l_1 (K^{(1)}_{22} + K^{(1)}_{24} \cos \tilde{k}_x )
\\
+\l_2( K^{(2)}_{22} + K^{(2)}_{24} \cos \tilde{k}_y )
\end{matrix}
\end{pmatrix}
{\bf u}_0
= \frac {\omega^2}2 {\bf M}_0{\bf u}_0
\relax
where ${\bf k}= (k_x,k_y)$ and
$
\tilde{k}_x = l_1k_x$, $\tilde{k}_y=l_2k_y$.
\subsubsection{Quasi-static effective speeds for the rectangular lattice}
Using ${\bf k}=k\boldsymbol\kappa$ the second order asymptotics of \er{-5} \rev2{are}
\[\label{eff1}
(k^2{\bf A}+\omega^2{\bf B}+k{\bf D}+{\bf E})({\bf u}_0+k{\bf u}_1+k^2{\bf
u}_2)={\bf 0}
\]
with matrices of the form
\[\label{eff2}
{\bf A}=\mathrm{diag}(A_j),\ \ {\bf B}=\mathrm{diag}(B_j),\ \ {\bf
E}=\mathrm{diag}(0,0,E),\ \ {\bf D}=\begin{pmatrix}
0 & 0 & d_1 \\
0 & 0 & d_2 \\
d_1^* & d_2^* & 0
\end{pmatrix}
\]
where (in the following calculations we do not need exact values of
$A_3$, $B_3$)
\[\label{eff2a}
\begin{aligned}
A_1&=\frac 12{\m_1 l_1 \k_x^2}+ {6\l_2l_2^{-1}\k_y^2},\ \
A_2=\frac 12{\m_2 l_2\k_y^2 }+ {6\l_1 l_1^{-1}\k_x^2},
\ \
E=6\l_1 l_1^{-1}+6\l_2 l_2^{-1} ,
\\
B_1&=B_2=- \frac 12(\rho} \def{\omega_1 l_1+\rho} \def{\omega_2 l_2 +m_0),
\ \ d_1={6i\l_2l_2^{-1} \k_y},\ \ d_2=-{6i\l_1 l_1^{-1}\k_x} .
\end{aligned}
\]
Substituting $\omega=c_jk$, ${\bf u}_i={\bf u}_{ji}$, $j=1,2$ (because
for $\omega,k=0$ we have two solutions) into \er{eff1} we obtain
\begin{eqnarray} \label{eff5}
k^0&:&{\bf E}{\bf u}_{j0}={\bf 0}, \label{eff3} \notag \\
k^1&:&{\bf D}{\bf u}_{j0}+{\bf E}{\bf u}_{j1}={\bf 0}, \label{eff4} \\
k^2&:&({\bf A}+c_j^2{\bf B}){\bf u}_{j0}+{\bf D}{\bf u}_{j1}+{\bf E}{\bf
u}_{j2}={\bf 0}.\notag
\end{eqnarray}
Scalar multiplying the O$(k^2)$ equation by ${\bf u}_{j0}$ and using
\er{eff2}-\er{eff4} with self-adjointness of all matrices we deduce
that
\[\label{eff6}
\begin{pmatrix}{\bf u}_{10} & {\bf u}_{20}\end{pmatrix}^{T}({\bf A}+c_j^2{\bf B}){\bf u}_{j0}-\frac1E\begin{pmatrix}{\bf u}_{10} & {\bf u}_{20}\end{pmatrix}^{T}
\begin{pmatrix} d_1 \\ d_2 \\ 0 \end{pmatrix}\begin{pmatrix} d_1^* & d_2^* & 0 \end{pmatrix}{\bf
u}_{j0}={\bf 0}.
\]
Using \er{eff2a} we can rewrite the effective
equations \er{eff6} as
\[\label{eff9}
\begin{pmatrix}
\m_1l_1\k_x^2+\frac{12\k_y^2}{ \l_1^{-1}l_1+\l_2^{-1}l_2}
& -\frac{12\k_x\k_y}{\l_1^{-1}l_1 +\l_2^{-1}l_2} \\
-\frac{12\k_x\k_y}{\l_1^{-1}l_1 +\l_2^{-1}l_2}
& \m_2l_2\k_y^2+\frac{12\k_x^2}{\l_1^{-1}l_1 +\l_2^{-1}l_2}
\end{pmatrix}{\bf v}_{j0}=mc_j^2{\bf v}_{j0}
\]
with, as expected \cite{Colquitt11}, the total mass per unit cell
\[\label{eff10}
m=m_0 + \rho} \def{\omega_1 l_1 +\rho} \def{\omega_2 l_2 .
\]
The equation \er{eff9} with constant matrix has two solutions:
effective speeds $c_j^2$ and corresponding constant displacements
${\bf v}_{j0}$, $j=1,2$, which are eigenvalues and eigenvectors of
the left matrix divided by $m$.
\subsubsection{Numerical example}
We consider wave propagation in the $x-$direction $(\tilde{k}_y=0)$
in which case the solutions of \eqref{-5} simplify as follows: (i) a
quasi-longitudinal solution ${\bf u}_0 = (1,0,0)^T$ with
$\tilde{k}_x$ given explicitly in terms of $\omega$ from
\beq{46-}
\cos \tilde{k}_x = \cos\tilde{s}_1 +
\Big(
{\l_2 }{}(K^{(2)}_{11}+K^{(2)}_{13}) - \frac 12 {m_0 \omega^2}
\Big)\, \frac{ \sin\tilde{s}_1}{ \tilde{\mu}_1\tilde{s}_1} .
\relax
Note that this mode couples longitudinal effects in the $x-$direction with flexural effects in the $y-$direction.
(ii) a quasi-flexural solution ${\bf u}_0 = (0,a,b)^T$ with dispersion relation in the form of a quadratic equation for $\cos \tilde{k}_x $
\bal{-45}
&\Big(\l_1 (K^{(1)}_{11}+K^{(1)}_{13}\cos \tilde{k}_x )
+\tilde{\mu}_2\tilde{s}_2 (\cot\tilde{s}_2-\csc\tilde{s}_2 ) - \frac{1}{2}m_0\omega^2\Big)
\notag \\
& \quad \times
\Big( \l_1 (K^{(1)}_{22} + K^{(1)}_{24} \cos \tilde{k}_x )
+\l_2( K^{(2)}_{22} + K^{(2)}_{24} )
-\frac{1}{2}I_0\omega^2\Big)
-\big(\l_1 K^{(1)}_{14}\sin \tilde{k}_x \big)^2=0.
\eal
Solutions of this dispersion relation couple the flexural wave in the $x-$direction with both the longitudinal and flexural waves in the $y-$direction.
We consider a lattice with square unit cell of size $L^2$, with all members the same and of thickness $t$ (and therefore radius of gyration $\k = t/\sqrt{12}$). The dimensions and properties used are given in Table \ref{tab1}, which corresponds to an example considered in \cite{Leamy12}.
\begin{table}[ht]
\caption{Parameters of the square lattice.}
\centering
\begin{tabular}{ |c | c | c | c| c |}
\hline
$E$ (GPa) & $\nu$ & $\rho^V$ (kg/m$^3$)& $L$ (mm)& $t$ (mm) \\
\hline \hline
70 & .33 & 2.7 $\cdot$ 10$^3$ & 10 & 1 \\
\hline
\end{tabular}
\label{tab1}
\end{table}
Results based on eq.\ \eqref{-5} are shown in Fig.\ \ref{fig1numeric} along with a comparison against results found using FEM (COMSOL).
The two types of wave solutions defined by eqs.\ \er{46-} and \er{-45} are distinguished in Fig.\ \ref{fig1numeric}.
Based on the comparison with the FEM calculations in Fig.\ \ref{fig1numeric} it is evident that the present theory provides an excellent match to the first six Floquet branches for waves propagating in the $x-$direction.
\begin{figure}[h!]
\begin{center}
\subfloat[]{
\includegraphics[width=2.5in]{figs/UsedExample_LeamySquare_Euler.eps}
}
\subfloat[]{
\includegraphics[width=2.5in]{figs/UsedExample_LeamySquare_COMSOL.eps}
}
\end{center}
\caption{Dispersion curves of the square lattice of Table \ref{tab1} for $k_y=0$. (a) The blue curves correspond to quasi-longitudinal motion described by eq.\ \eqref{46-}; the black and red curves correspond to the pair of quasi-transverse solutions described by
eq.\ \eqref{-45}. (b) Dispersion curves calculated using FEM (COMSOL).
}
\label{fig1numeric}
\end{figure}
\subsection{Hexagonal lattice}
\begin{figure}[h!] \label{hex_unit}
\centering
\subfloat[]{
\setlength{\unitlength}{0.15mm}
\begin{picture}(201,175)(160,-260)
\special{color rgb 0 0 0}\put(331,-179){\makebox(0,0)[cc]{\shortstack{{\Large a$_1$}}}}
\special{color rgb 0 0 0}\put(193,-117){\makebox(0,15)[cc]{\shortstack{{\Large a$_5$}}}}
\special{color rgb 0 0 0}\put(227,-175){\makebox(0,15)[cc]{\shortstack{{\Large a$_2$}}}}
\special{color rgb 0 0 0}\put(192,-260){\makebox(0,15)[cc]{\shortstack{{\Large a$_6$}}}}
\special{color rgb 0 0 0}\put(332,-260){\makebox(0,15)[cc]{\shortstack{{\Large a$_4$}}}}
\special{color rgb 0 0 0}\put(332,-115){\makebox(0,15)[cc]{\shortstack{{\Large a$_3$}}}}
\put(214,-181){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(215,-180)(305,-180)
\put(305,-180){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\put(353,-107){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\put(167,-107){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\put(352,-252){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(305,-180)(355,-105)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(305,-180)(355,-255)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(165,-105)(215,-180)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(165,-255)(215,-180)
\put(167,-252){\special{color rgb 0 0 0}{\ellipse*{15}{15}}}
\special{color rgb 0 0 0}
\end{picture}
}
\hspace{20pt}
\subfloat[]{
\includegraphics[width=0.3\textwidth]{figs/Hex_Brill_zone2.eps} \label{ibz_hex}
}
\caption{Hexagonal lattice. (a) The unit cell. (b) The irreducible Brillouin zone \cite{Maldovanbook09}.}
\end{figure}
\subsubsection{Quasi-static effective speeds for the hexagonal lattice}
Consider the special case in which the lattice is a regular hexagon with uniform properties $\mu$, $\lambda$ and $l$. It follows from eq.\ \eqref{effn5} that
\beq{-31}
{\bf C}_{\rm eff}^2 =
\frac{3l}{2m} \,\Big[ \Big(\frac 1\mu + \frac {l^2}{12\lambda}
\Big)^{-1} \mathrm{diag} (1,1) + \frac\mu 2 \boldsymbol\kappa\bk^T
\Big]
\relax
The eigenvectors of the matrix ${\bf C}_{\rm eff}^2$ are then purely longitudinal and transverse, i.e. parallel and perpendicular to $\boldsymbol\kappa$, with wave speeds $c_L$ and $c_T$, respectively, where
\beq{-32}
c_T^2 = \frac{3l}{2m} \, \Big(\frac 1\mu + \frac {l^2}{12\lambda}
\Big)^{-1},
\quad
c_L^2 = c_T^2 + \frac{3l}{4m} \mu .
\relax
\subsubsection{Numerical result}
We consider an example for which all members have the same uniform properties and are arranged in a regular hexagonal lattice.
The numeric computations are based on the properties in Table \ref{tab2} and the path of the wave vector taken is along the perimeter of the Brillouin zone shown in Fig.\ \ref{ibz_hex}.
\begin{table}[h!]
\caption{Hexagonal lattice parameters.}
\centering
\begin{tabular}{| c | c | c | c| c |}
\hline
$E$ (GPa) & $\nu$ & $\rho^V$ (kg/m$^3$)& $l$ (mm)& $t$ (mm) \\
\hline \hline
70 & .33 & 2.7 $\cdot$ 10$^3$ & 10 & 1 \\
\hline
\end{tabular}
\label{tab2}
\end{table}
The dispersion curves in Fig.\ \ref{figsec2hex2numeric}(a) \reva{were obtained from eq.\ \eqref{226} } using a combination of minimum value threshold and minimum peak finding methods for the $6\times 6$ determinant evaluated on a discretized grid of wave vector and frequency. This provides a fast solution technique, which can be refined by taking smaller grid steps. Figure \ref{figsec2hex2numeric} shows that the dispersion curves computed by the present simplified theory agree well with those found using FEM. A close comparison shows some small deviations from the FEM results (which can safely be considered as an accurate benchmark) but the overall agreement is remarkable considering the simplicity of the present approach. \rev2{The hexagonal system displays a strong one-wave effect between approximately 15 and 30 kHz. In this range the dispersion is weak, as indicated by the almost straight line branches. Furthermore, the hexagonal symmetry ensures isotropy in the long-wavelength limit, which is the original reason \cite{Norris11mw} for our interest in this particular structure.
}
\reva{Note that the roots obtained in Fig.\ \ref{figsec2hex2numeric}(a) were numerically checked using a symbolic algebra -generated expression for the determinant of eq.\ \eqref{226}. Although significant speedup in computing time was not observed, this was not the primary purpose and future work could use such very lengthy but precise expressions to better computational advantage.
}
\begin{figure}[h!]
\begin{center}
\subfloat[]{
\includegraphics[width=2.5in]{figs/UsedExample_Hex_Example.eps}
}
\subfloat[]{
\includegraphics[width=2.5in]{figs/UsedExample_Hex_Example_COMSOL.eps}
}
\end{center}
\caption{Dispersion curves of the regular hexagonal lattice with properties in Table \ref{tab2}. (a) The first six Floquet branches for wave-vector along the perimeter of the Brillouin zone. (b) Dispersion curves calculated using FEM (COMSOL).
\rev2{Note the almost non-dispersive one-wave behaviour between 15 and 30 kHz.}}
\label{figsec2hex2numeric}
\end{figure}
\section{3D Examples} \label{sec6}
\subsection{Cubic lattice}
\begin{figure}[H] \label{cubic_cell_lattice}
\centering
\setlength{\unitlength}{0.18mm}
\begin{picture}(161,173)(180,-260)
\special{color rgb 0 0 0}\put(245,-170){\makebox(0,0)[cc]{\shortstack{{\Large a$_0$}}}}
\special{color rgb 0 0 0}\put(341,-169){\makebox(0,0)[cc]{\shortstack{{\Large a$_1$}}}}
\special{color rgb 0 0 0}\put(243,-117){\makebox(0,15)[cc]{\shortstack{{\Large a$_5$}}}}
\special{color rgb 0 0 0}\put(182,-175){\makebox(0,15)[cc]{\shortstack{{\Large a$_2$}}}}
\special{color rgb 0 0 0}\put(242,-265){\makebox(0,15)[cc]{\shortstack{{\Large a$_6$}}}}
\special{color rgb 0 0 0}\put(197,-240){\makebox(0,15)[cc]{\shortstack{{\Large a$_4$}}}}
\special{color rgb 0 0 0}\put(312,-130){\makebox(0,15)[cc]{\shortstack{{\Large a$_3$}}}}
\put(185,-180){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(185,-180)(335,-180)
\put(260,-180){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(335,-180){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(260,-105){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(260,-255){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(260,-105)(260,-255)
\put(305,-135){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(215,-225){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(305,-135)(215,-225)
\special{color rgb 0 0 0}
\end{picture}
\caption{Cubic unit cell}
\end{figure}
\subsubsection{Dispersion relations}
Similar to the rectangular lattice, the equation of motion can be written as
\begin{equation} \label{cubic_eom}
\sum\limits_{j=1,2,3,4,5,6}\big( {\bf P}_{0j}^{(1)} - {\bf P}_{0j}^{(2)}{\bf e}^{i{\bf k}\cdot{\bf g}_j} \big){\bf u}_0 = \omega^2 {\bf M}_0 {\bf u}_0,
\ \ {\bf M}_0 = \mathrm{diag}(m_0,m_0,m_0,I_0,I_0,I_0).
\end{equation}
We assume the members are of three types: 1, 2, 3 for the $x$, $y$, and $z$-directions, respectively, with parameters denoted by $\rho_j$, ${\bf K}^{(j)}$, etc. $j=1,2,3$, then eq.\
\eqref{cubic_eom} becomes,
\begin{footnotesize}
\begin{align}\label{-4}
&\begin{pmatrix} \zeta_1 & 0&0&0 & i\l_3K^{(3)}_{14}\sin \tilde{k}_z & -i\l_2K^{(2)}_{14}\sin \tilde{k}_y\\ 0& \zeta_2 & 0 & -i\l_3K^{(3)}_{14}\sin \tilde{k}_z & 0 & i\l_1K^{(1)}_{14}\sin \tilde{k}_x\\ 0&0& \zeta_3& i\l_2K^{(2)}_{14}\sin \tilde{k}_y & -i\l_1K^{(1)}_{14}\sin \tilde{k}_x &0\\ 0& i\l_3K^{(3)}_{14}\sin \tilde{k}_z & -i\l_2K^{(2)}_{14}\sin \tilde{k}_y & \zeta_4&0&0\\ -i\l_3K^{(3)}_{14}\sin \tilde{k}_z & 0 & i\l_1K^{(1)}_{14}\sin \tilde{k}_x & 0 & \zeta_5 & 0\\i\l_2K^{(2)}_{14}\sin \tilde{k}_y & -i\l_1K^{(1)}_{14}\sin \tilde{k}_x & 0 & 0 & 0 & \zeta_6\end{pmatrix} 2 {\bf u}_0
\notag \\
& \qquad \qquad
=\omega^2{\bf M}_0{\bf u}_0 ,
\end{align}
\end{footnotesize}
where ${\bf k}= (k_x,k_y,k_z)$,
$( \tilde{k}_x , \tilde{k}_y, \tilde{k}_z) = (l_1k_x,l_2k_y,l_3k_z)$ and
\begin{equation}
\begin{aligned}
&\zeta_1=\l_2 (K^{(2)}_{11}+K^{(2)}_{13}\cos \tilde{k}_y)+\l_3 (K^{(3)}_{11}+K^{(3)}_{13}\cos \tilde{k}_z)
+\tilde{\mu}_1 \tilde{s}_1(\cot\tilde{s}_1 -\csc\tilde{s}_1 \cos \tilde{k}_x),
\\&\zeta_2=\l_3 (K^{(3)}_{11}+K^{(3)}_{13}\cos \tilde{k}_z)+\l_1 (K^{(1)}_{11}+K^{(1)}_{13}\cos \tilde{k}_x)
+\tilde{\mu}_2 \tilde{s}_2(\cot\tilde{s}_2 -\csc\tilde{s}_2 \cos \tilde{k}_y),
\\&\zeta_3=\l_1 (K^{(1)}_{11}+K^{(1)}_{13}\cos \tilde{k}_x)+\l_2(K^{(2)}_{11}+K^{(2)}_{13}\cos \tilde{k}_y )
+\tilde{\mu}_3 \tilde{s}_3(\cot\tilde{s}_3 -\csc\tilde{s}_3 \cos \tilde{k}_z ),
\\&\zeta_4=\l_2 (K^{(2)}_{22}+K^{(2)}_{24}\cos \tilde{k}_y)+\l_3 (K^{(3)}_{22}+K^{(3)}_{24}\cos \tilde{k}_z),
\\&\zeta_5=\l_3 (K^{(3)}_{22}+K^{(3)}_{24}\cos \tilde{k}_z)+\l_1 (K^{(1)}_{22}+K^{(1)}_{24}\cos \tilde{k}_x),
\\&\zeta_6= \l_1(K^{(1)}_{22}+K^{(1)}_{24}\cos \tilde{k}_x)+\l_2 (K^{(2)}_{22}+K^{(2)}_{24}\cos \tilde{k}_y).
\end{aligned}
\end{equation}
\subsubsection{Quasi-static effective elastic moduli for the cubic lattice}
Considering wave propagation in the $(100)$ and $(110)$ directions of the lattice with pure cubic symmetry ($l_1=l_2=l_3$ etc.) and taking the low frequency limit, we obtain
\beq{eff_moduli}
C_{11}= {EA}/{l^2},\ \ C_{66}={6EI}/{l^4},\ \ C_{12}=0,
\ \ \rho_{\text {eff}} = (3\rho Al+3m_0)/l^3 .
\relax
The moduli are in agreement with known results, e.g. \cite{Norris14}, and the effective mass density is, as expected, identical to the actual density.
$C_{12}=0$ indicates that Poisson's ratio $\nu_{12}=0$ which can be interpreted as applying a displacement in the $(100)$ direction does not cause deformation in the $(010)$ direction.
\subsubsection{Example: wave propagation in the $x-$direction}
We consider wave propagation along one axis of a lattice structure with uniform material and structural properties as given in Table 3 and with members of square cross-section.
\begin{table}[h!]
\caption{Cubic lattice parameters.}
\centering
\begin{tabular}{| c | c | c | c| c |}
\hline
$E$ (GPa) & $\nu$ & $\rho^V$ (kg/m$^3$)& $l$ (mm)& $t$ (mm) \\
\hline \hline
70 & .33 & 2.7 $\cdot$ 10$^3$ & 10 & 1 \\
\hline
\end{tabular}
\label{tab3}
\end{table}
Setting $\tilde{k}_y=\tilde{k}_z=0$, we find that the first pure-longitudinal solution ${\bf u}_0 = (1,0,0,0,0,0)^T$ of \eqref{-4}, has wavenumber $\tilde{k}_x$ in terms of $\omega$ as
\begin{equation}\label{426000}
\cos \tilde{k}_x = \cos\tilde{s} +
\Big(2
\l (K_{11}+K_{13}) - \frac 12 m_0 \omega^2
\Big) \frac{\ \sin\tilde{s}}{ \tilde{\mu} \tilde{s}} .
\end{equation}
The flexural solution ${\bf u}_0 = (0,1,l,0,\alpha,\beta)^T$ reduces the $6\times6$ equation of motion matrix to a $4\times4$ one.
\begin{equation}
\begin{pmatrix}B - m_0\omega^2&0&0&D\\0&B - m_0\omega^2&-D&0\\0&D&C - I_0\omega^2&0\\-D&0&0&C - I_0\omega^2 \end{pmatrix}\begin{pmatrix}1\\l\\ \alpha\\ \beta \end{pmatrix} = {\bf 0},
\end{equation}
where
\begin{equation}
\begin{aligned}
&B=2\tilde{\mu} \tilde{s}(\cot\tilde{s}-\csc\tilde{s})+2\l \big( 2K_{11}+K_{13}(\cos\tilde{k}_x+1)\big) ,
\\&C=2\l\big( 2K_{22}+K_{24}(\cos\tilde{k}_x+1)\big) ,
\\&D=i2\l K_{14}\sin\tilde{k}_x.
\end{aligned}
\end{equation}
Then calculate the determinant to obtain the flexural dispersion relation
\begin{equation}\label{430000}
\begin{aligned}
&\Big(\l\big(2K_{11}+K_{13}(\cos \tilde{k}_x+1)\big)+\tilde{\mu}\tilde{s}(\cot\tilde{s}-\csc\tilde{s}) - \frac{1}{2}m_0\omega^2\Big)
\\
& \quad \times
\Big( \l\big(2K_{22} + K_{24} (\cos \tilde{k}_x +1 )\big)
-\frac{1}{2}I_0\omega^2\Big)
-\big(\l K_{14}\sin \tilde{k}_x \big)^2=0.
\end{aligned}
\end{equation}
\rev2{
In addition to the propagating wave branches the model also displays pure resonances. These are modes that are independent of $k_x$ and hence non-propagating, i.e. with zero group velocity. They correspond to the generalized displacement
${\bf u}_0 = (0,0,0,1,0,0)^T$ which
represents flexural resonances (deflection in $x$-direction) of the beams oriented in the $y$- and $z$-directions. The mode is a solution of eq.\ \eqref{-4} at resonance frequencies
that satisfy
\beq{-56-}
2\l \big(K_{22}+K_{24}\big) - \omega^2 I_0 =0.
\relax
In the case considered with $I_0=0$, eq.\ \eqref{-56-} reduces to
\beq{-5=7}
\Big( \sin \frac{\gamma l}2 \cosh \frac{\gamma l}2 + \cos \frac{\gamma l}2 \sinh \frac{\gamma l}2 \Big)
\, \sin \frac{\gamma l}2 = 0
\relax
where $\gamma$ is the flexural wavenumber of Euler beam theory. The first two lowest solutions of eq.\ \eqref{-5=7} are $\gamma l = 1.5000 \pi$ and $2\pi$.
}
The dispersion curves for the cubic lattice are shown in Fig. \ref{fig:dispersion}. \rev2{The analytic results for the propagating wave branches eqs.\ \eqref{426000} and \eqref{430000} match with the FEM simulation. The first two resonance frequencies of eq.\ \eqref{-5=7} are at
$51.949$ kHz and $92.354$ kHz, and are shown as flat branches in Fig.\ \ref{fig:dispersion}(a).
The first/lowest solution corresponds to the flat branch in Fig.\ \ref{fig:dispersion}(b). The branch in Fig.\ \ref{fig:dispersion}(b) corresponding to the $92.354$ kHz resonance shows slight variation with wavenumber, but is well approximated by the flat branch in Fig.\ \ref{fig:dispersion}(a).
We can conclude from the comparison in Fig.\ \ref{fig:dispersion} that the analytical model predicts the first eight branches to a remarkable degree of approximation.
}
\begin{figure}[h!]
\begin{center}
\subfloat[]{
\includegraphics[width=2.5in]{figs/cubic_theory.eps}
}
\subfloat[]{
\includegraphics[width=2.5in]{figs/cubic_comsol.eps}
}
\end{center}
\caption{Dispersion curves of the cubic lattice of Table \ref{tab3}. (a) The green lines are dispersion curves of longitudinal waves, the black and red lines are dispersion curves of shear waves, and the blue curves are flexural resonances of the beams oriented in the $y$- and $z$-directions. (b) Dispersion curves calculated using FEM (COMSOL). }\label{fig:dispersion}
\end{figure}
\subsection{Tetrahedral lattice}
\begin{figure}[H]
\centering
\subfloat[]{
\setlength{\unitlength}{0.18mm}
\begin{picture}(228,132)(160,-245)
\put(230,-180){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(230,-180)(310,-180)
\put(310,-180){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(310,-180)(365,-125)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(310,-180)(360,-215)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(310,-180)(350,-240)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(175,-235)(230,-180)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(180,-150)(230,-180)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(195,-125)(230,-180)
\put(175,-235){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(180,-150){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(195,-125){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(360,-215){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(350,-240){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\put(365,-125){\special{color rgb 0 0 0}{\ellipse*{10}{10}}}
\special{color rgb 0 0 0}\put(230,-171){\shortstack{${a}_2$}}
\special{color rgb 0 0 0}\put(290,-171){\shortstack{${a}_1$}}
\special{color rgb 0 0 0}\put(335,-126){\shortstack{${a}_3$}}
\special{color rgb 0 0 0}\put(320,-241){\shortstack{${a}_4$}}
\special{color rgb 0 0 0}\put(350,-206){\shortstack{${a}_5$}}
\special{color rgb 0 0 0}\put(160,-226){\shortstack{${a}_6$}}
\special{color rgb 0 0 0}\put(205,-131){\shortstack{${a}_7$}}
\special{color rgb 0 0 0}\put(170,-171){\shortstack{${a}_8$}}
\special{color rgb 0 0 0}
\end{picture}
}
\hspace{20pt}
\subfloat[]{
\includegraphics[width=1in]{figs/diamond_IBZ.eps} \label{ibz_dia}
}
\caption{Tetrahedral lattice. (a) The unit cell. (b) The irreducible Brillouin zone \cite{Curtarolo2010}.}{\label{tetra_unit}}
\end{figure}
\subsubsection{Numerical result}
We consider an example for which all members are rods of radius $t$ and have the same uniform properties and are arranged in a regular tetrahedral lattice. The \rev2{numerical} computations are based on the properties in Table \ref{tab4} and the path of the wave vector taken is along $\Gamma - {\text L}$ of the Brillouin zone shown in Fig.\ \ref{ibz_dia}.
\begin{table}[ht]
\caption{Tetrahedral lattice parameters.}
\centering
\begin{tabular}{| c | c | c | c| c |}
\hline
$E$ (GPa) & $\nu$ & $\rho^V$ (kg/m$^3$)& $l$ (mm)& $t$ (mm) \\
\hline \hline
70 & .33 & 2.7 $\cdot$ 10$^3$ & 10 & .5 \\
\hline
\end{tabular}
\label{tab4}
\end{table}
The dispersion curves in Fig.\ \ref{diamond_disp}(a) were obtained by finding the smallest eigenvalue of a positive definite matrix, and plotting the corresponding wave number and frequency of the discretized grid where the smallest eigenvalue is smaller than $\epsilon$ (a small value). Figure \ref{diamond_disp} shows that the dispersion curves computed by the present simplified theory agree well with those found using FEM.
\rev2{As with the 2D hexagonal structure, the tetrahedral lattice displays a broad frequency range with one-wave behaviour: 5 to 20 kHz. The wave is almost non-dispersive, and isotropic in the long-wavelength regime on account of the symmetry of the lattice.}
\begin{figure}[h!]
\begin{center}
\subfloat[]{
\includegraphics[width=2.5in]{figs/Diamond_5mm_rod.eps}
}
\subfloat[]{
\includegraphics[width=2.5in]{figs/comsol_5mm.eps}
}
\end{center}
\caption{Dispersion curves of the regular tetrahedral lattice with properties in Table \ref{tab4}. (a) The first seven Floquet branches for wave-vector along the perimeter of the Brillouin zone. (b) Dispersion curves calculated using FEM (COMSOL).
\rev2{Note the clear one-wave behaviour between about 5 and 20 kHz.}
}
\label{diamond_disp}
\end{figure}
\section{Conclusions}\label{sec7}
Dynamic modeling of 2D and 3D lattices can be accurately modeled using a low order model with minimal degrees of freedom described by thin beam members. The dispersion relations for rectangular and cubic lattices have been derived analytically by imposing the Bloch-Floquet periodicity condition, yielding an Hermitian eigenvalue problem for the unknown frequencies. Numerical methods were used to compute the band-diagrams for hexagonal and tetrahedral lattices. The semi-analytical approach allowed us to extract the low frequency asymptotics. In particular, the closed-form explicit expressions for the Christoffel matrix in the quasistatic regime for rectangular, hexagonal and cubic lattices were presented. Numerical comparisons of wave dispersion diagrams with FEM simulations indicate that the beam model provides good accuracy for lower modes. The semi-analytical nature of the present model makes it the natural extension of purely static methods for periodic lattice structures, e.g. \cite{Norris14}.
\rev2{It accurately predicts the one-wave behaviour in the hexagonal and tetrahedral lattices. These particular structures are distinct in that they provide effective in the long-wavelength limit, and hence quasi-acoustic wave effects in the one-wave regions. By breaking the symmetry one can extend the scalar one-wave effect to display anisotropy, an important subject for future investigation with the semi-analytic model.}
In summary, our beam model provides a novel and fast approach to calculate the band-diagrams for 2D and 3D lattices. This semi-analytical method may prove useful in designing phononic crystals and pentamode structures.
\section*{Acknowledgments}
AK was partially supported by the RSF project N\textsuperscript{\underline{o}}15-11-30007 and TRR 181 project. X.S. acknowledges support under ONR MURI Grant No. N000141310631. A.N.N. acknowledges support from Institut de M\'{e}canique et d'Ing\'{e}nierie, Universit\'{e} de Bordeaux. The reviewers are thanked for providing suggestions that improved the paper.
|
\section{Introduction}
The very first successfully imaged exoplanet is Fomalhaut b \citep{Kalas2008}. The Fomalhaut system consists of a young main-sequence star of spectral type A4, Fomalhaut A, and there are two other physically bound stars in this system, Fomalhaut B, which is of spectral type K, and Fomalhaut C, which is of spectral type M. Both are located at very wide distances from the primary star, at ca.\ 100'' corresponding to ca.\ 750~AU. Fomalhaut A is located at a distance of 7.5~pc from the Sun. It is surrounded by a debris disk that has been imaged in the optical and in the infrared. Repeated observations performed with the coronograph onboard \textit{HST} at red optical wavelengths revealed an object that appeared to be moving on an orbit in the same plane as the debris disk with an orbital distance of ca.\ 100~AU; this object was called Fomalhaut b. An exoplanet at that distance to the host star is expected to be cold, and respective models yielded that the object likely has a mass roughly similar to that of Jupiter \citep{Kalas2008}.
However, observations at other wavelengths yielded surprising results. The planet was repeatedly detected at optical and near-infrared wavelengths of around 6000 and 8100\,\AA~\citep{Kalas2008, Currie2012}, but observations at even longer wavelengths, where the exoplanet is expected to be brighter, yielded strong upper limits that were incompatible with typical exoplanetary model spectra \citep{Janson2012}. Even more confusingly, Fomalhaut b was detected at blue optical wavelengths (ca.\ 4400\,\AA), where it should have been very dim \citep{Currie2012}. Recent detections at optical wavelengths also revealed that the best-fit orbit of this object is not, as previously thought, aligned with the debris disk; instead, the object seems to be on a disk-crossing orbit \citep{Kalas2013}. Several possible explanations have been brought forward to explain the surprising spectral shape of this object, including a planetary accretion disk around a Jupiter-mass planet, explaining the blue excess by reflection, but not explaining the out-of-disk orbit \citep{Kalas2008, Janson2012}; a small hot planet with a very large cloud of planetesimals, explaining the brightness and misaligned orbit without producing unobserved disk perturbations, but not explaining how it arrived there since no scattering objects of sufficient mass are available \citep{Janson2012}; and an icy dust cloud without a core, explaining the brightness, but being implausible in terms of age estimates \citep{Currie2012, Galicher2013}. None of these hypotheses is fully consistent with the object properties observed so far.
\begin{figure}
\includegraphics[width=0.48\textwidth]{hrcimage_zoomed.pdf}
\caption{The X-ray image collected with HRC-I shows that the positions of Fomalhaut and Fomalhaut~b (marked with small circles) do not display any significant photon excess over the background (large circle).}
\label{hrcimagezoomed}
\end{figure}
Given that the latest observations of relative motion observed between Fomalhaut A and Fomalhaut b do not support an in-disk orbit of the planet candidate, one should consider the possibility that Fomalhaut b might be a background object that just happens to pass by Fomalhaut~A at a close projected distance. This has recently been discussed by \cite{Neuhaeuser2015}.
The faintness of Fomalhaut b at optical wavelengths (ca.\ 25 mag) and the blue-leaning SED constrain what kind of object it can be. It cannot be an ordinary star, since any star with such a blue SED would be intrinsically bright. To match Fomalhaut b's faintness, the hypothetical star would have to be so far away that substantial reddening caused by the interstellar medium would occur, so that the colors would not fit any more. It also cannot be a white dwarf; white dwarfs typically have absolute visible magnitudes of 10--15 mag \citep{WoodOswalt1998}. Placing a white dwarf at a distance of 0.5--5 kpc to match the brightness would mean that the projected 2D spatial velocity of the white dwarf would have to be unrealistically large in order to fit the observed trajectory of Fomalhaut~b \citep{Neuhaeuser2015}.
This leaves the possibility that Fomalhaut b may be a neutron star. Neutron stars have absolute visual magnitudes of ca.\ 22--25 mag, depending on temperature \citep[for example]{Kaplan2011}. The apparent magnitude of Fomalhaut b is ca.\ 25 in the visual, so an old and cool neutron star at ca.\ 10 pc distance or a somewhat hotter one farther away would generally fit the observed apparent brightness of the object. Neutron stars start out with surface temperatures of several MK after formation and cool over time scales of megayears to temperatures of around 100\,000 K \citep{Page1992, Yakovlev2004review}. Several nearby thermally emitting neutron stars have been detected in X-rays with sufficient signal to noise to allow spectral fitting, and those neutron stars typically display surface temperatures of 500\,000 to 1\,000\,000~K; however, they are estimated to be rather young with $\lesssim 1$~Myr \citep{Kaplan2011}. An older, cooler neutron stars will still look blue at optical wavelengths, matching the few observational detections available so far for Fomalhaut~b. The emission from an old neutron star would peak in the far-UV, however the brightness of the central star Fomalhaut A at optical and UV wavelengths makes such observations difficult and requires sophisticated subtraction methods for the stellar point spread function (PSF). However, in soft X-rays, where the hard tail of the neutron star spectrum may be detectable, a star with the mass of Fomalhaut is expected to be X-ray dark, making X-ray observations a good opportunity to study the object Fomalhaut~b.
\section{Observations and data analysis}
The Fomalhaut system has been imaged in X-rays previously, albeit spatially unresolved, with the \textit{Einstein} and \textit{ROSAT} observatories. The deepest archival observation was performed with a \textit{ROSAT} PSPC pointing in 1992, and we present that data set shortly in the following section. We have performed a new, deeper X-ray observation which can spatially resolve Fomalhaut and Fomalhaut~b using \textit{Chandra}'s HRC-I instrument and present that data in detail in section~\ref{newobs}.
\subsection{The archival \textit{ROSAT} pointing}
\textit{ROSAT} was a space telescope that operated from 1990 to 1999, observing the soft X-ray sky \citep{Truemper1982, Aschenbach1988}. Its PSPC camera \citep{Pfeffermann1987} provided soft X-ray images with a spatial resolution of 25 arsec (full-width half maximum of the point spread function) in an energy range of 0.1--2.4~keV. The Fomalhaut system was observed in a PSPC pointing for 6.3~ks in 1992.
The spatial resolution of \textit{ROSAT}'s PSPC is not sufficient to resolve the Fomalhaut system since the spatial separation between Fomalhaut and Fomalhaut~b is 13.8 arcsec.
The Fomalhaut system is undetected in X-rays by \textit{ROSAT}.
\cite{Schmitt1997} give an upper limit on the X-ray count rate of
0.0066~cts/s for this observation. In the interest of a consistent
analysis of the \textit{ROSAT} and \textit{Chandra} data, we
re-analyzed the data set and derive a 3$\sigma$ upper limit on
the count rate of 0.008~cts/s. The flux that corresponds to this
upper limit depends on the assumed spectrum of the undetected source,
and we discuss this in more detail in section~\ref{results}. Assuming
a black-body spectrum with a temperature of 100,000~K, we calculate
an upper limit on the X-ray flux of $9.0\times 10^{-14}$~erg\,cm$^{-2}$\,s$^{-1}$.
We used the WebPIMMS
count rate simulator for this \citep{Mmukai1993}, which is accessible
at \url{http://heasarc.gsfc.nasa.gov/cgi-bin/Tools/w3pimms/w3pimms.pl}. We took into
account that the Boron filter was used during this \textit{ROSAT} observations,
which reduces the throughput at soft X-ray energies by ca.\ 90\%.
\subsection{The new \textit{Chandra} observation}\label{newobs}
We performed an X-ray pointing with \textit{Chandra}'s HRC-I \citep{Weisskopf2000} with a duration of 32.7~ks. Details of the observation are given in Table~\ref{xrayobs}. We used the standard data analysis tools of \textit{Chandra}'s \textit{CIAO} software, version 4.7 \citep{ciao}. Specifically, we extracted counts from a source detect cell placed on the nominal positions of Fomalhaut and Fomalhaut~b, each with a radius of $1^{\prime\prime}$. This was compared to a background signal which was extracted from a source-free region on the chip with a radius of $20^{\prime\prime}$. Since HRC-I provides no useful energy resolution for the collected X-ray photons (in contrast to other X-ray instruments onboard \textit{Chandra}), no energy filtering was performed. The extracted photons therefore can come from the full range of energies to which HRC-I is sensitive, i.e.\ 0.08--10~keV.
We show the extracted image collected with HRC-I in
Fig.~\ref{hrcimagezoomed}.
The indicated regions show the positions of Fomalhaut and Fomalhaut~b,
as well as our chosen background region. The background region contains
294 photon counts, while the source detect cells for Fomalhaut and Fomalhaut~b contain
6 and 7 photon counts, respectively. Given that the background extraction region has an area
100 larger than the individual source detect cells, we would expect a pure background signal of
2.94 in a source detect cell.
\begin{table}
\caption{X-ray observation parameters for Fomalhaut~b.}
\label{xrayobs}
\begin{tabular}{l l}
\hline \hline
Instrument & \textit{Chandra} HRC-I\\ \hline
Observation ID & 17896 \\
Exposure time & 32691.2 s\\
Start time & 2016-08-09 10:38:42 \\
End time & 2016-08-09 20:22:24\\
Energy range & 0.08--10 keV\\
\hline
\end{tabular}
\end{table}
\begin{table*}
\caption{Observed fluxes and upper limits for Fomalhaut~b in different wavelength bands.}
\label{allobs}
\begin{tabular}{l l l l l l l}
\hline \hline
Obs.\ year & Telescope/Instr. & Filter & Detection & flux & flux density & Reference \\
& & & & [mag] & [erg\,cm$^{-2}$\,s$^{-1}$\,\AA$^{-1}$] & \\
\hline
1992 & ROSAT PSPC & Boron filter & no & ... & $< 7.55\times10^{-16}$ & this work \\
2004 & HST ACS/HRC & F606W & yes & $24.92\pm0.10$ & $3.14\pm0.29\times10^{-19}$ & Currie et al.\ (2012) \\
2004 & HST ACS/HRC & F606W & yes & $24.42\pm0.09$ & $4.93\pm0.41\times10^{-19}$ & Kalas et al.\ (2008) \\
2004 & HST ACS/HRC & F606W & yes & $24.29\pm0.08$ & $5.61\pm0.41\times10^{-19}$ & Kalas et al.\ (2008) \\
2005 & KeckII NIRC2 & H & no & $\geq22.9$ & $\leq 8.28\times10^{-20}$ & Kalas et al.\ (2008) \\
2005 & KeckII NIRC2 & CH$_4$S & no & $\geq20.6$ & ... & Kalas et al.\ (2008) \\
2006 & HST ACS/HRC & F435W & no & $\geq24.7$ & $8.36\times10^{-19}$ & Kalas et al.\ (2008) \\
2006 & HST ACS/HRC & F435W & yes & $25.22\pm0.18$ & $5.18\pm0.86\times10^{-19}$ & Currie et al.\ (2012) \\
2006 & HST ACS/HRC & F606W & yes & $25.13\pm0.09$ & $2.59\pm0.21\times10^{-19}$ & Kalas et al.\ (2008) \\
2006 & HST ACS/HRC & F606W & yes & $24.97\pm0.09$ & $3.00\pm0.25\times10^{-19}$ & Currie et al.\ (2012) \\
2006 & HST ACS/HRC & F814W & yes & $24.55\pm0.13$ & $1.69\pm0.20\times10^{-19}$ & Kalas et al.\ (2008) \\
2006 & HST ACS/HRC & F814W & yes & $24.91\pm0.20$ & $1.21\pm0.22\times10^{-19}$ & Currie et al.\ (2012) \\
2008 & Gemini North NIRI & L$^\prime$ & no & $\geq 16.6$ & $\leq 1.22\times10^{-18}$ & Kalas et al.\ (2008) \\
2009 & Subaru IRCS & J & no & $\geq 22.22$ & $\leq 4.01\times10^{-19}$ & Currie et al.\ (2012) \\
2010/2011 & Spitzer IRAC & 4.5\,$\upmu$m & no & $\geq 16.7$ & $\leq 5.71\times10^{-19}$ & Janson et al.\ (2012) \\
2013 & Spitzer IRAC & 4.5\,$\upmu$m & no & $\geq 17.3$ & $\leq 3.30\times10^{-19} $ & Janson et al.\ (2015) \\
2016 & Chandra HRC-I & 0.08-10 keV & no & ... & $< 8.58\times10^{-17}$ & this work \\
\hline
\end{tabular}
\end{table*}
Using Poissonian counting statistics, a signal of
9~counts
in a detect cell would correspond to a $\geq 3\sigma$ excess,
i.e.\ a probability smaller than 0.03\% that this signal would
be detected as a random fluctuation of the background. However,
the signal in the detect cells is lower than that, meaning that
a significant count excess is not detected for either
Fomalhaut or Fomalhaut~b.
\footnote{In contrast, the source to the lower left of the background region in
Fig.~\ref{hrcimagezoomed} has 33 photon counts in a $1^{\prime\prime}$
detect cell and is significantly detected. No matches were found
for this source position in major catalogs at various wavelengths,
so that we can make no further
statements about the nature of this source.}
We therefore adopt a signal of
9~counts
counts and, given the exposure time of 32691.2~s, a count rate of
0.00028~cts/s
as the upper limit for Fomalhaut and Fomalhaut~b in the full HRC energy
band of 0.08--10~keV.
What we are specifically interested in is an upper limit on the X-ray flux and X-ray luminosity of Fomalhaut~b and Fomalhaut. Often this is done by spectral fitting to model the energy sensitivity of the detector and derive the actual source flux. However, since the \textit{{HRC-I}} instrument has virtually no intrinsic energy resolution, one needs to find an appropriate counts-to-energy conversion factor for the observation. This conversion factor will depend on the underlying X-ray spectrum one assumes for a source, because \textit{{HRC-I}} has different sensitivities to photons at the upper and lower end of the energy range it is sensitive to. To give the reader a rough idea, the effective area of \textit{{HRC-I}} starts at 0.08~keV with a few cm$^2$ and rises in several steps to ca.\ 200~cm$^2$ at 1~keV (with a narrow downward spike around 0.3~keV due to an absorption edge of carbon residuals on the mirror assembly), and decreases again at energies higher than 1~keV. Therefore a spectrum peaking at soft energies where the effective area is small will lead to a small number of counts being detected, whereas a spectrum with the same flux, but peaking at harder X-ray energies where \textit{{HRC-I}} has a larger effective area, will lead to more counts being detected.
\section{Results}\label{results}
\subsection{Fomalhaut b as a neutron star}
Neutron stars emit strongly in the far-UV and in soft X-rays. Following \citet{Neuhaeuser2015}, we will now explore the hypothesis of Fomalhaut~b being a neutron star in the near fore- or background of the Fomalhaut disk, given the new observations we performed with \textit{Chandra}.
\subsubsection{The X-ray flux upper limit}
To determine the appropriate upper limit to the X-ray flux of Fomalhaut~b from the upper limit on the X-ray count rate, one needs to assume some sort of underlying hypothetical source spectrum. Neutron stars are generally soft X-ray sources \citep{Truemper2005, Haberl2007}. Soon after their formation their surface cools down to a few MK, and reaches temperatures of a few $100,000$~K and lower after several million years \citep{Yakovlev2004review}.
We can derive an appropriate count-to-energy conversion factor for \textit{Chandra} HRC-I by using a blackbody spectral model in WebPIMMS. Using a hotter neutron star model with a surface temperature of 1~MK over an X-ray energy range of 0.08--10~keV yields a conversion factor of $8.18\times 10^{-12}$~ergs\,cm$^{-2}$\,s$^{-1}$\,count$^{-1}$ for HRC-I; using a colder neutron star model with $100\,000$~K yields a conversion factor of $4.79\times 10^{-11}$~ergs\,cm$^{-2}$\,s$^{-1}$\,count$^{-1}$, i.e.\ larger by a factor of about five. An upper limit on the X-ray flux will therefore depend on the assumed surface temperature of the neutron star.
Fortunately, this dependence and the actual fitting of possible spectral models including the optical data are `self-correcting': Say we assume a surface temperature that is too high. This will lead to a small conversion factor, meaning a very low, i.e.\ very restrictive, upper limit on the X-ray flux. This in turn will force us to fit the data with a rather cool neutron star model, because a hot neutron star would be too bright in X-rays and therefore violate the upper limit. The mismatch between our initial assumption for the surface temperature and the derived fit will indicate that the assumption was not correct. Similarly, assuming a neutron star that is too cold will lead to a large conversion factor, a non-restrictive upper limit on the X-ray flux, and therefore allow the data to be fit with a hotter neutron star. Finally, a roughly correct assumption about the surface temperature will lead to a matching temperature in the fit.
Having gone through this exercise, we assume a cool
surface temperature of $100\,000$~K for Fomalhaut~b as a
neutron star and use a counts-to-energy conversion factor
of $4.79\times 10^{-11}$~ergs\,cm$^{-2}$\,s$^{-1}$\,count$^{-1}$.
The upper limit on the count rate of
0.00028~cts/s
therefore corresponds to an upper limit on the X-ray flux of
$1.3\times 10^{-14}$~ergs\,cm$^{-2}$\,s$^{-1}$. We calculate the upper limit on the average flux density for
the wavelength range covered by \textit{HRC-I}, i.e.\ 0.08-10~keV, to be
$8.6\times 10^{-17}$~ergs\,cm$^{-2}$\,s$^{-1}$\,\AA$^{-1}$.
\begin{figure*}
\includegraphics[width=0.48\textwidth]{spectrum_4fits.pdf}
\includegraphics[width=0.48\textwidth]{spectrum_4fits_zoom.pdf}
\caption{The detections (black) and upper limits (red) for Fomalhaut b, together with an expected spectrum for a 400~K cold giant planet in light grey (from the pre-computed model grids available at \url{https://phoenix.ens-lyon.fr/Grids/}; these spectra are based on the PHOENIX code \citep{Hauschildt1997} with adaptations for sub-stellar objects \citep{Chabrier2000, Allard2001}) and blackbody spectra for a cool neutron star with temperatures of 200\,000~K, 140\,000~K, 91\,000~K, and 60\,000~K at a distance of 13.5~pc. The plot to the right shows a zoom-in onto the optical and infrared part of the spectrum. A neutron star with $T=91\,000$~K and a distance of 13.5~pc is the best-fitting neutron star model that fits the optical/near-IR detections and grazes the X-ray upper limit. }
\label{spectra}
\end{figure*}
\subsubsection{Optical and near-infrared data}
Fomalhaut~b has been detected by several groups in the optical and the near-infrared. Some detections in the optical have been reported to have significantly differing fluxes when taking the reported error bars at face value; however, given the sophisticated image analysis necessary to detect Fomalhaut~b's emission at all, we consider it is possible that the actual flux errors may be larger than initially reported. Upper limits at longer wavelengths have been derived from observations as well. \citet{Neuhaeuser2015} have given a summary on available pre-2015 detections and upper limits, and we give an updated overview of the available data in Table~\ref{allobs}, including a new data point in the infrared by \citet{Janson2015} and the X-ray flux upper limits derived in this work.
We can use those data sets to constrain an assumed neutron star spectrum for Fomalhaut~b. One might in principle assume a reddening effect on the spectrum occurs from dust absorption along the line of sight between Fomalhaut~b and the observer. However, interstellar absorption is negligible for nearby objects. Dust absorption from the Fomalhaut disk is also unlikely to produce significant reddening: \cite{Marsh2005} observed the Fomalhaut disk at 350~$\upmu$m and modelled a reddening of 0.005~mag at a wavelength of 24~$\upmu$m from those observations. From this we can extrapolate, using appropriate conversions from the 24~micron band to the $K$ and $V$ bands \citep{Indebetouw2005, Xue2016}, that the reddening in the $V$ band will be ca.\ 0.17~mag. This is smaller than the rms scatter of the Fomalhaut~b detections in the observed optical bands, which is of the order of 0.3~mag (see Table~\ref{allobs}), and we therefore regard a potential dust reddening due to the Fomalhaut disk as irrelevant to our further analysis.
\subsubsection{Possible neutron star parameters for Fomalhaut~b}
The emission from a neutron star in its simplest form can be described by a blackbody spectrum. It has been shown that often neutron stars are more accurately described by adding an additional harder X-ray component to the spectrum. Physically, this is likely due to hot polar caps covering a certain fraction of the neutron star surface (see \citet{Becker1997} and references therein) or the presence of a pulsar wind nebula \citep{Stappers2003}; Those harder X-ray spectral components are typically described by a power-law or a thermal spectrum. However, as the simplest possible model, we are assuming a single temperature blackbody component for the neutron star surface. Furthermore, we assume a typical radius of the hypothetical neutron star of 10~km \citep{Oezel2016review}. Any discrepancy to this assumed radius will only cause a slightly smaller or larger distance to the object in the spectral fit.
We produce a least-squares fit of a blackbody model to the available data, using the detections of Fomalhaut~b at optical wavelengths and treating the Chandra upper limit as a detection data point for the sake of the fit. The other upper limits in the optical and IR do not restrict the fit (see Figure~\ref{spectra}).
The best fit is found for a blackbody temperature of 91\,000~K and a distance of 13.5~pc. The constraint on the temperature is driven by the X-ray upper limit, while the detections in the optical and near-infrared pinpoint the distance for a given temperature. We indicate the effect of changing the temperature for a given distance (the latter being interchangeable with the radius of the object) in Figure~\ref{spectra}; a closer distance moves the spectrum upwards, while a cooler temperature moves the spectrum to the lower right according to Wien's displacement law. This means that one cannot assume a neutron star hotter than 91\,000~K and place it at a larger distance, because such a model would either lie below the optical data points while satisfying the X-ray upper limit, or be a reasonable fit to the optical data while violating the X-ray upper limit. Note that the best-fitting neutron star model is found for the case where the object is indeed detected at the X-ray upper limit, which it is not. Therefore, if Fomalhaut~b is a neutron star, it has to be cooler than 91\,000~K and closer than 13.5~pc.
Additionally, using the largest distance estimate of 13.5~pc, we can convert the upper limit to Fomalhaut~b's X-ray flux into an upper limit to its X-ray luminosity. We calculate this upper limit to be $L_\mathrm{X} < 2.7\times10^{26}$~erg/s.
\subsection{The A-type star Fomalhaut}
The central star Fomalhaut is an intermediate-mass star of spectral type A4V. Massive stars (spectral types O and early B) drive winds which produce X-rays through shocks, while low-mass stars (mid-F to late M) emit in X-rays due to their magnetically heated coronae. Stars in the intermediate mass regime, i.e. of spectral types mid-B to early F, typically do not show X-ray emission, with the exception of the chemically peculiar Ap/Bp stars. Among regular, i.e.\ non-chemically-peculiar A-type stars, only a handful have been detected in X-rays. Those rare examples display weak and very soft X-ray emission that can be described by a coronal plasma with dominant temperature components around 1~MK in the case of the A5V star beta Pictoris \citep{Guenther2012betaPic} and temperature components in the range from 1-4~MK in the case of the A7V star Altair \citep{Robrade2009Altair}. Other X-ray observations of A-type stars have not led to detections \citep{Pease2006, Ayres2008Vega, Drake2014}, with the exception of HR~8799, which is listed by SIMBAD as an A6 star, but was shown by \cite{Robrade2010HR8799} to be a peculiar type of early-F star with coronal emission, therefore not being representative for the intermediate-mass stellar regime.
In our observation, Fomalhaut is undetected in X-rays,
with an upper limit to the count rate of 0.00028~cts/s. Using a soft coronal
plasma model with a temperature of 1~MK to derive the appropriate
counts-to-energy conversion factor ($1.10\times10^{-11}$~erg/s/count), this upper
limit corresponds to an upper limit to the X-ray flux of $F_\mathrm{X} < 3.03\times10^{-15}$~erg\,s$^{-1}$\,cm$^{-2}$
and, using Fomalhaut's distance of 7.7~pc, an X-ray luminosity of $L_\mathrm{X} < 2.0\times10^{25}$~erg/s
over the energy range of 0.08--10~keV. Converting this to the more regularly used
energy range of 0.2--2~keV, we derive an upper limit of
$L_\mathrm{X,\, 0.2-2} < 6.3\times10^{24}$~erg/s.
The ratio of X-ray and bolometric luminosity, often used
as an activity indicator for stars that can display magnetic activity
\citep{Telleschi2007, Poppenhaeger2010, Wright2011},
is $L_\mathrm{X}/L_\mathrm{bol} < 3.2\times10^{-10}$ for the energy range of 0.08--10~keV
and $L_\mathrm{X,\, 0.2-2}/L_\mathrm{bol} < 9.9\times10^{-11}$ for the energy range of
0.2--2~keV; a bolometric luminosity of $16.63$ times the solar bolometric luminosity,
i.e.\ $6.36\times10^{34}$\,erg$\,$cm$^{-2}$$\,$s$^{-1}$\, \citep{Mamajek2012, Davis2005} was adopted for
Fomalhaut. Fomalhaut is therefore a star for which magnetic activity is
present at vanishingly low levels, if at all.
\begin{figure*}
\includegraphics[width=0.48\textwidth]{spectra_starlight.pdf}
\includegraphics[width=0.48\textwidth]{spectra_starlight_zoomin.pdf}
\caption{Left: The emission from the central A4V star Fomalhaut~A dominates at wavelengths below 100~\AA (blue dotted line). Right: Both a cool neutron star and reflected starlight, i.e.\ a scaled down spectrum of Fomalhaut~A, could explain the optical detections for Fomalhaut~b (black solid line and green dashed line). }
\label{Astarspectra}
\end{figure*}
To set this into context, this is a very restrictive upper limit, similar to the upper limit derived with \textit{Chandra}'s ACIS and HRC instruments for the A0V star Vega, which has been observed to be $L_\mathrm{X} \leq 2\times 10^{25}$~erg/s \citep{Pease2006, Ayres2008Vega}.
\section{Discussion}
Our non-detection of Fomalhaut~b in X-rays places strong constraints on the neutron star hypothesis. In order to fit the available data with a single-temperature blackbody model and a typical neutron star radius of 10~km, the hypothetical neutron star would need to be colder than ca.\ 90\,000~K, i.e.\ it would need to be very old with $\gtrsim 10$~Myr according to typical cooling models \citep{Yakovlev2004review}, and need to be located at a distance closer than 13.5~pc.\footnote{As already argued by \citet{Neuhaeuser2015}, any accompanying supernovae remnant from the formation of the neutron star would have diffused after ca.\ $100\,000$~years, consistent with an old age of the hypothetical neutron star.} This would make Fomalhaut~b the nearest known neutron star to the solar system, and furthermore an extremely old and cool one. If the neutron star radius was very small, for example 2~km, the temperature restriction would still hold, but the upper limit on its distance would only be ca.\ 0.5~pc. To set this into context, we can compare Fomalhaut~b to the extremely cool and old neutron star reported by \citet{Keane2013}. They derive an upper limit on temperature and black-body luminosity in the soft X-ray band of $<280\,000$~K and $<4.3\times10^{30}$~erg/s for the neutron star J1840--1419, which is located at a distance of 900~pc. Obviously, the close distance to the hypothetical neutron star for Fomalhaut~b is a main driver of the very low upper limit we have derived. However, it is still noteworthy that this would be the most near-by and coldest neutron star known, i.e.\ even with a neutron star explanation for Fomalhaut~b's surprising colours, we would be facing a very unusual scenario for \textit{detected} neutrino stars. It is worth pointing out that in general one can expect old and cool neutron stars to be quite common from supernova occurrence rates in the Milky Way (see for example \citealt{Camenzind2007book} p.269). However, since they are dim objects at all wavelengths, detecting those old and cool neutron stars is extremely challenging; if Fomalhaut~b truly is a neutron star, it would be the coldest one detected so far.
While a neutron star scenario is not completely ruled out, it is worthwhile to compare it to a scenario in which the detected emission of Fomalhaut~b is reflected light from the central star, and the reflecting object would be an object residing within the Fomalhaut disk (see for example \citet{Kalas2008}). As a rough guideline, we show a blackbody spectrum with appropriate parameters to approximate an A4V star like Fomalhaut ($T_\mathrm{eff} = 8270$~K \citep{Pecaut2013} and $L_\mathrm{bol} = 16.63 L_\mathrm{bol,\,\odot}$ \citep{Mamajek2012}, and from that calculated $R_\ast = 1.38\times 10^{11}$~cm) in Fig.~\ref{Astarspectra} (left-hand side). Scaling down this spectrum to account for a smaller reflecting surface at the position of Fomalhaut~b, we find a good match to the optical data when scaling down by a factor of $10^{-9}$, as shown in Fig.~\ref{Astarspectra} (right-hand side). Indeed, the reflected starlight hypothesis matches the optical data point in the blue ($435$~\AA) and in the near-infrared ($814$~\AA) better than the neutron star models which over-estimate the blue part and under-estimate the near-infrared part. We can perform a rough estimate of what size such a reflecting object would need to have to yield this observed flux. Assuming the reflector is in the Fomalhaut disk (i.e.\ at a distance of 115~AU to the central star) and has perfect reflectivity, we know that in order to reflect $10^{-9}$ of the starlight the reflector needs to have $10^{-9}$ times the area of a sphere with a radius of $115$~AU (as this sphere would encompass the complete flux of Fomalhaut). This corresponds to an area of ca.\ $4\times10^{12}$\,km$^2$. For comparison, the area enclosed by Saturn's outermost rings is ca.\ $6\times10^{10}$\,km$^2$. Much larger ring systems around planets may exist; the system 1SWASP J140747.93-394542.6 has been recently reportet to undergo extreme brightness changes which can be explained by the presence of a large ring system with a maximum radius of 0.6~AU \citet{Mamajek2012, Kenworthy2015}. That ring system would enclose an area of ca.\ $2.5\times10^{16}$\,km$^2$. For the Fomalhaut system, it is therefore conceivable that a smaller planet with a relatively large ring system with an area of $4\times10^{12}$\,km$^2$, i.e.\ a radius of ca.\ 600,000~km, could reflect enough starlight to match the observations.
To distinguish between these two hypotheses, imaging of Fomalhaut~b at far-UV or EUV wavelengths would be more definitive, however difficult since there are currently no telescopes in operation at those wavelengths. In the wavelength range of 200--400~\AA\ the neutron star and reflected light models begin to significantly differ from each other (compared to the optical); unfortunately, this is also the spectral range where the emission from the central star is strongest. In the EUV around 100~\AA\ a cool neutron star would display its strongest emission, while the central star is already relatively faint.
\section{Conclusion}
We have presented new X-ray observations of the Fomalhaut system that can spatially resolve the position of the central A4V star and the suspected planet Fomalhaut~b. We report non-detections of both the central star and the suspected planet in X-rays.
The non-detection of the central star places a strict upper limit on its X-ray flux and its fractional X-ray luminosity $L_\mathrm{X}/L_\mathrm{bol} < 10^{-10}$ for the energy range of 0.2--2~keV, comparable to the best X-ray upper limits which have been obtained for A-type stars, albeit in a small sample.
The hypothesis that Fomalhaut~b is a neutron star instead of a planet, located in the near fore- or background of the Fomalhaut disk, is strongly constrained by our non-detection. Possible neutron star models fitting the observational data have to be extremely cool ($\lesssim 90\,000$~K) and nearby ($\lesssim 13$~pc). An alternative explanation that the emission seen from Fomalhaut~b is reflected light from the central star is possible, and fits the optical detection in the blue better than the neutron star models which over-estimate the blue emission compared to the observations. Further observations in the UV and EUV would likely be able to clarify the mysterious nature of Fomalhaut~b.
\section{Acknowledgements}
The scientific results reported in this article are based on observations made by the Chandra X-ray Observatory. Support for this work was provided by the National Aeronautics and Space Administration through Chandra Award Number GO6-17056X issued by the Chandra X-ray Observatory Center, which is operated by the Smithsonian Astrophysical Observatory for and on behalf of the National Aeronautics Space Administration under contract NAS8-03060.
|
\section{Applications}
\label{sec:6}
In this section we explore {four} applications for illustrating the advantages and
flexibility
of the methods proposed in the previous sections. In the first application, we apply
Theorem~\ref{t:1} to the obstacle problem in PDE's in which dropping the extra forward
step decreases the computational cost per iteration because
the computation of an extra gradient step is numerically expensive.
In the second application, devoted to empirical risk minimization (ERM), we illustrate
the flexibility of using non self-adjoint linear operators. We derive different sequential
algorithms depending on the nature of the linear operator involved. { In the third
application, we develop a distributed operator-splitting scheme which allows for
time-varying communication graphs. Finally, the last application focuses in nonlinear
constrained optimization, in which monotone non-Lipschitz operators arise naturally. }
\subsection{Obstacle problem}
The obstacle problem is to find the equilibrium position of an elastic membrane on a
domain $\Omega$, whose boundary is fixed and is restricted to remain above the
some obstacle, given by the function $\varphi\colon\Omega{\tilde{o}}\mathbb{R}$. This problem can
be applied to fluid filtration in porous media, elasto-plasticity, optimal control among
other disciplines (see, e.g., \cite{Caf88} and the references therein).
Let $u\colon\Omega{\tilde{o}}\mathbb{R}$ be a function representing the vertical displacement of the
membrane and let $\psi\colon\Gamma{\tilde{o}}\mathbb{R}$ be the function representing the fixed
boundary, where $\Gamma$ is the smooth boundary of $\Omega$. Assume that
$\psi\in H^{1/2}(\Gamma)$ and $\varphi\in C^{1,1}(\Omega)$ satisfy
$\mathrm{T}\varphi\leq\psi$, and consider the
problem
\begin{align}
\min_{\mathrm{u}\in H^1(\Omega)}&\frac{1}{2}\int_{\Omega}|\nabla
\mathrm{u}|^2\mathrm{dx}\nonumber\label{e:obstprob}\\
\text{s.t. } \mathrm{T}\mathrm{u}&=\psi,\quad\text{a.e. on
}\Gamma; \\
\mathrm{u}&\geq \varphi,\quad\text{a.e. in }\Omega\nonumber,
\end{align}
where $\mathrm{T}\colon H^1(\Omega){\tilde{o}} H^{1/2}(\Gamma)$ is
the (linear) trace operator and $H^1(\Omega)$ is endowed with the
scalar product
$\scal{\cdot}{\cdot}\colon(\mathrm{u},\mathrm{v})\mapsto
\int_{\Omega}\mathrm{u}\mathrm{v}\,dx+\int_{\Omega}\nabla\mathrm{u}
\cdot\nabla\mathrm{v}\,dx$. There is a unique solution to this
obstacle problem \cite{Caff98}.
In order to set this problem in our context, let us define the
operator
\begin{equation}
\label{e:Q}
Q\colon H^{-1}(\Omega)\times H^{-1/2}(\Gamma){\tilde{o}} H^1(\Omega)
\end{equation}
which associates to each $(\mathrm{q},\mathrm{w})\in
H^{-1}(\Omega)\times H^{-1/2}(\Gamma)$ the unique weak solution
(in the sense of
distributions) to \cite[Section~25]{ZeidlerIIB}
\begin{equation}
\begin{cases}
-\Delta \mathrm{u}+\mathrm{u}=\mathrm{q},\quad&\text{in
}\Omega;\\
\frac{\partial \mathrm{u}}{\partial \nu}=\mathrm{w},&\text{on
}\Gamma,
\end{cases}
\end{equation}
where $\nu$ is outer unit vector normal to $\Gamma$. Hence, $Q$
satisfies
\begin{equation}
\label{e:Qweak}
(\forall \mathrm{v}\in
\mathrm{H})\qquad\scal{Q(\mathrm{q},\mathrm{w})}{\mathrm{v}}=
\scal{\mathrm{w}}{\mathrm{T}\mathrm{v}}_{{-1/2},{1/2}}
+\scal{\mathrm{q}}{\mathrm{v}}_{{-1},1},
\end{equation}
where $\scal{\cdot}{\cdot}_{{-1/2},{1/2}}$ and
$\scal{\cdot}{\cdot}_{{-1},{1}}$ stand for the dual pairs
$H^{-1/2}(\Gamma)-H^{1/2}(\Gamma)$ and
$H^{-1}(\Omega)-H^{1}(\Omega)$, respectively. Then, by defining
$\mathrm{H}=H^1(\Omega)$, $\mathrm{G}=H^{1/2}(\Gamma)$,
$\mathrm{f}\colon \mathrm{u}\mapsto
\frac{1}{2}\int_{\Omega}|\nabla
\mathrm{u}|^2\mathrm{dx}$, $\mathrm{g}=\iota_{\mathrm{C}}$,
where
$\mathrm{C}=\menge{\mathrm{u}\in\mathrm{H}}{\mathrm{u}\geq\varphi\;\text{
a.e.
in }\Omega}$, let $\mathrm{D}={\{\psi\}}$, and
let $\mathrm{L}=\mathrm{T}$, \eqref{e:obstprob} can be
written equivalently as
\begin{equation}
\label{e:probobstheo}
\min_{\mathrm{L}\mathrm{u}\in\mathrm{D}}\mathrm{f}(\mathrm{u})+
\mathrm{g}(\mathrm{u}).
\end{equation}
Moreover, it is easy to verify that $\mathrm{f}$ is convex and, by
using integration by parts and
\eqref{e:Qweak}, for every
$\mathrm{h}\in\mathrm{H}$ we have
\begin{align}
\mathrm{f}(\mathrm{u}+\mathrm{h})-\mathrm{f}(\mathrm{u})-
\Scal{\!Q\left(\!-\Delta\mathrm{u},\frac{\partial\mathrm{u}}{\partial\nu}\right)\!}{\!\mathrm{h}}
&=\frac{1}{2}\int_{\Omega}|\nabla\mathrm{h}|^2dx+\int_{\Omega}\!\!
\nabla\mathrm{u}\cdot\nabla\mathrm{h}\,dx+\scal{\Delta\mathrm{u}}{\mathrm{h}}_{-1,1}\nonumber\\
&\hspace{20pt}-\scal{\frac{\partial\mathrm{u}}{\partial\nu}}{\mathrm{T}\mathrm{h}}_{-1/2,1/2}\nonumber\\
&=\frac{1}{2}\int_{\Omega}|\nabla\mathrm{h}|^2dx,
\end{align}
which yields
\begin{equation}
\lim_{\|\mathrm{h}\|\to0}\frac{\left|\mathrm{f}(\mathrm{u}+\mathrm{h})-\mathrm{f}(\mathrm{u})-
\Scal{Q\left(-\Delta\mathrm{u},\frac{\partial\mathrm{u}}{\partial\nu}\right)}
{\mathrm{h}}\right|}{\|\mathrm{h}\|}=\frac{1}{2}\lim_{\|\mathrm{h}\|\to0}\frac{\|\nabla
\mathrm{h}\|_{L^2}^2}{\|\mathrm{h}\|}=0.
\end{equation}
Hence, $\mathrm{f}$ is Fr\'echet differentiable with a linear gradient given by
$\nabla\mathrm{f}\colon
\mathrm{u}\mapsto Q\left(\!-\Delta\mathrm{u},\frac{\partial\mathrm{u}}
{\partial\nu}\right)$. Moreover, from integration by parts we have
\begin{equation}
\Scal{Q\left(-\Delta\mathrm{u},\frac{\partial\mathrm{u}}{\partial\nu}\right)}{\mathrm{h}}=
\Scal{\frac{\partial\mathrm{u}}{\partial\nu}}{\mathrm{T}\mathrm{h}}_{-1/2,1/2}
-\scal{\Delta\mathrm{u}}{\mathrm{h}}_{-1,1}=\int_{\Omega}
\nabla\mathrm{u}\cdot\nabla\mathrm{h}\,dx\leq\|\mathrm{u}\|\|\mathrm{h}\|,
\end{equation}
which yields $\|\nabla\mathrm{f}(\mathrm{u})\|\leq\|\mathrm{u}\|$
and, hence, it is $1$-cocoercive \cite{baillon1977quelques}. In addition, the trace
operator is
linear and bounded \cite{Grisv86} and we have from \eqref{e:Qweak}
that
\begin{equation}
\label{e:Qweak2}
(\forall \mathrm{v}\in
\mathrm{H})(\forall \mathrm{w}\in
H^{1/2}(\Gamma))\qquad\scal{Q(0,\mathrm{w})}{\mathrm{v}}=
\scal{\mathrm{w}}{\mathrm{T}\mathrm{v}}_{{-1/2},{1/2}},
\end{equation}
which yields
$\mathrm{L}^*\colon\mathrm{w}\mapsto\mathrm{Q}(0,\mathrm{w})$ and
since $\mathrm{C}$ is non-empty closed convex, $\mathrm{g}$ is
convex, proper, lower semicontinuous and $\mathbf{prox}_{\gamma
\mathrm{g}}=P_{\mathrm{C}}$, for any $\gamma >0$.
Since first order conditions of \eqref{e:probobstheo} reduce to find
$(\mathrm{u},\mathrm{w})\in\mathrm{H}\times\mathrm{G}$ such that
$0\in N_{\mathrm{C}}(\mathrm{u})+\nabla\mathrm{f}(\mathrm{u})+
\mathrm{T}^*N_{\mathrm{D}}(\mathrm{T}\mathrm{u})$, which is a particular case
of Problem~\ref{prob:mi} and from Corollary~\ref{c:pd} when $\theta=1$ the method
\begin{equation}
\label{e:genVUobst}
\begin{array}{l}
\left\lfloor
\begin{array}{l}
\mathrm{v}^k
=P_{\mathrm{C}}
\left(\mathrm{u}^k-\sigma_0Q\left(-\Delta\mathrm{u}^k,
\frac{\partial\mathrm{u}^k}{\partial\nu}+\mathrm{w}^k\right)\right)\\[2mm]
\mathrm{t}^k = \mathrm{w}^k +
\sigma_1
\left(\mathrm{T}(2\mathrm{y}^k-\mathrm{x}^k)-\psi\right)\\[2mm]
\mathrm{u}^{k+1}=
\mathrm{u}^k+\frac{\lambda}{\sigma_0}\left(\mathrm{v}^k-\mathrm{u}^k+
\sigma_0
Q(0,\mathrm{w}^k-\mathrm{t}^k)\right)\\[2mm]
\mathrm{w}^{k+1}=\mathrm{w}^k+\frac{\lambda}{\sigma_1}
\left(\mathrm{t}^k-\mathrm{w}^k-
\sigma_1\mathrm{T}(\mathrm{v}^k-\mathrm{u}^k)\right)\\[2mm]
\end{array}
\right.\\[2mm]
\end{array}
\end{equation}
generates a weakly convergent sequence $(\mathrm{u}^k)_{k\in\ensuremath{\mathbb N}}$ to the
unique
solution to the obstacle problem provided, for instance (see
Remark~\ref{r:pd2}.\ref{r:pd21}), that
$\max\{\sigma_0,\sigma_1\}+2\sqrt{\sigma_0\sigma_1}\|\mathrm{T}\|<2$.
Note that $\nabla\mathrm{f}$ must be computed only once at each iteration, improving
the performance with respect to primal-dual methods following Tseng's approach, in
which $\nabla\mathrm{f}$ must be computed twice by iteration (see, e.g.,
\cite{briceno2011monotone+,tseng2000modified}). The method proposed in
{\cite{condat2013primal,vu2013splitting}} can also solve this problem but with
stronger
conditions on
constants $\sigma_0$ and $\sigma_1$ as studied in Remark~\ref{r:pd2}. Moreover,
our approach may include variable stepsizes together with different assymetric linear
operators which may improve the performance of the method.
On the other hand, the general version of our method in
Theorem~\ref{thm:asymmetric_metric} allows for an
additional projection onto a closed convex set. In this case this can be useful to impose
some of the constraints of the problem in order to guarantee that iterates at each
iteration satisfy such constraints. An additional projection step may accelerate the
method as it has been studied in \cite{BAKS16}.
Numerical comparisons
among these methods are part of further research.
\subsection{An Incremental Algorithm for Nonsmooth Empirical Risk
Minimization}
In machine learning~\cite{Shalev-Shwartz:2014:UML:2621980}, the Empirical Risk
Minimization (ERM) problem seeks to minimize a finite sample approximation of an
expected loss, under conditions on the feasible set and the loss function. If the solution
to the sample approximation converges to a minimizer of the expected loss when the
size of the sample increases, we say that the problem is learnable.
Suppose that we have a sample of size $m$, and, for every $i\in\{1,\ldots,m\}$,
the loss function associated to the sample $\mathrm{z}_i$ is given by $l(\cdot;
\mathrm{z}_i) \colon
\mathrm{x}\mapsto \mathrm{f}_i(\mathrm{a}_i^{\top} x)$, where each
$\mathrm{a}_i \in \mathbb{R}^{d}\backslash \{0\}$
and each $\mathrm{f}_i : \mathbb{R} \rightarrow (-\infty, \infty]$ is closed, proper, and
convex.
Then the ERM problem is to
\begin{align}\label{eq:ERM}
\Min_{\mathrm{x} \in \mathbb{R}^d} \frac{1}{m} \sum_{i=1}^{m}
\mathrm{f}_i(\mathrm{a}_i^{\top} \mathrm{x}).
\end{align}
This form features in support vector machines, logistic regression, linear regression, least-absolute deviations, and many other common models in machine learning.
The parameter $m$ indicates the size of the training set and is typically large.
Parallelizing a (sub)gradient computation of~\eqref{eq:ERM} is straightforward, but in
general, because training sets are large, we may not have enough processors to do so.
Thus, when only a few processors are available, incremental iterative algorithms, in
which one or a few training samples are used per iteration to update our solution
estimate, are a natural choice.
Several incremental algorithms are available for solving~\eqref{eq:ERM}, including
incremental (sub)gradient descent and incremental aggregated gradient
methods~\cite{schmidt2013minimizing,defazio2014finito,johnson2013accelerating,defazio2014saga,bertsekas2015incremental,wang2013incremental,bertsekasincrementalproximal,doi:10.1137/S1052623499362111,bianchi2015ergodic}.
The former class requires
diminishing stepsizes (e.g., of size
$O(k^{-1/2})$) and, hence, their convergence may be very slow, while the latter
class of algorithms is
usually restricted to the cases in which either $\mathrm{f}_i$ is smooth or the dual
problem
of~\eqref{eq:ERM} is smooth (in which case~\eqref{eq:ERM} is strongly convex). In
contrast, we now develop an
incremental proximal algorithm, which imposes no smoothness or
strong convexity assumptions. It has a Gauss-Seidel
structure and is obtained by an application of
Theorem~\ref{p:primdumi}. The involved
stepsizes may vary among iterations but they are set to be constants
for simplicity.
The method follows from the following first-order optimality
conditions obtained
assuming some qualification condition:
\begin{equation}
\mathrm{x}\quad\text{solves } \eqref{eq:ERM}\quad\Leftrightarrow\quad
0\in\sum_{i=1}^{m}\mathrm{a}_i\partial
\mathrm{f}_i(\mathrm{a}_i^{\top}\mathrm{x}),
\end{equation}
which is a particular case of Problem~\ref{prob:mi} when
$\mathrm{H}=\mathbb{R}^d$,
$\mathrm{A}\equiv\{0\}$, $\mathrm{C}_1=\mathrm{C}_2\equiv0$ and, for
every
$i\in\{1,\ldots,m\}$, $\mathrm{G}_i=\mathbb{R}$, $\mathrm{D}_i^{-1}=0$,
$\mathrm{L}_i=\mathrm{a}_i^{\top}$, and
$\mathrm{B}_i=\partial\mathrm{f}_i$. By using
Theorem~\ref{p:primdumi} in this case for matrices
$(\mathrm{P}_{ij})_{0\leq i<j\leq m}$ given by
\begin{equation}
(\forall 0\leq j<i\leq m)\quad \mathrm{P}_{ij}=
\begin{cases}
\frac{\ensuremath{\operatorname{Id}}\,}{\sigma_0},\quad&\text{if }i=j=0;\\
\frac{1}{\sigma_i},\quad&\text{if }i=j>0;\\
-\mathrm{a}_i^{\top},&\text{if }j=0;\\
\sigma_0 \mathrm{a}_i^{\top}\mathrm{a}_j,&\text{if }0<j<i,
\end{cases}
\end{equation}
we obtain
\begin{equation}
\label{e:algomicompERM3}
\begin{array}{l}
\left\lfloor
\begin{array}{l}
\mathrm{v}_1^k = \mathbf{prox}_{
\sigma_{1}\mathrm{f}_1^{*}}\left(\mathrm{u}_1^k +
\sigma_{1}\left(
\mathrm{a}_1^{\top}\mathrm{x}^k-\sigma_{0}\sum_{i=1}^m
\mathrm{a}_1^{\top}\mathrm{a}_i\mathrm{u}_i^k\right)\right)\\
\mathrm{v}_2^k = \mathbf{prox}_{
\sigma_{2}\mathrm{f}_2^{*}}\left(\mathrm{u}_2^k+
\sigma_{2}\left(
\mathrm{a}_2^{\top}\mathrm{x}^k-\sigma_{0}\left(\mathrm{a}_2^{\top}
\mathrm{a}_1\mathrm{v}_1^k+\sum_{i=2}^m
\mathrm{a}_2^{\top}\mathrm{a}_i\mathrm{u}_i^k\right)\right)\right)\\[1mm]
\hspace{0.5cm}\vdots\\
\mathrm{v}_m^k = \mathbf{prox}_{
\sigma_{m}\mathrm{f}_m^{*}}\left(\mathrm{u}_m^k +
\sigma_{m}\left(
\mathrm{a}_m^{\top}\mathrm{x}^k-\sigma_{0}\left(\sum_{i=1}^{m-1}
\mathrm{a}_m^{\top}\mathrm{a}_i\mathrm{v}_i^k
+\|\mathrm{a}_m\|^2\mathrm{u}_m^k\right)\right)\right)\\[1mm]
\mathrm{x}^{k+1}=
\mathrm{x}^k-{\lambda}\sum_{i=1}^m
\mathrm{a}_i\mathrm{v}_i^k\\
\mathrm{u}_1^{k+1}=\mathrm{u}_1^k+\frac{\lambda}{\sigma_1}\left(
\mathrm{v}_1^k-\mathrm{u}_1^k\right)\\
\hspace{0.8cm} \vdots\\
\mathrm{u}_m^{k+1}=\mathrm{u}_m^k+\frac{\lambda}{\sigma_m}\left(
\mathrm{v}_m^k-\mathrm{u}_m^k\right)+
\sigma_0\sum_{j=1}^{m-1}\mathrm{a}_m^{\top}\mathrm{a}_j(\mathrm{v}_j^k
-\mathrm{u}_j^k).
\end{array}
\right.\\[2mm]
\end{array}
\end{equation}
Since conditions \eqref{e:condiPii}-\eqref{e:metricconditionpd}
hold if
\begin{equation}
\label{e:condERM1}
\sqrt{\sum_{i=1}^m\|\mathrm{a}_i\|^2}+\sigma_0\sum_{i=1}^m
\|\mathrm{a}_i\|^2+\frac{\sigma_0}{2}
\left(\max_{i=1,\ldots,m}\|\mathrm{a}_i\|^2-
\min_{i=1,\ldots,m}\|\mathrm{a}_i\|^2\right)<
\frac{1}{\max\limits_{i=0,\ldots,m}\sigma_i},
\end{equation}
by choosing $(\sigma_i)_{0\leq i\leq m}$ satisfying
\eqref{e:condERM1}
the sequence $(\mathrm{x}^k)_{k\in\ensuremath{\mathbb N}}$ generated by
\eqref{e:algomicompERM3} converges to a solution provided that
$\lambda<M^{-1}$ where
$$M=\left(\min_{i=0,\ldots,m}\sigma_i\right)^{-1}+\frac{1}{2}
\sqrt{\sum_{i=1}^m\|\mathrm{a}_i\|^2}+\frac{\sigma_0}{2}\left(\sum_{i=1}^m
\|\mathrm{a}_i\|^2+\max_{i=1,\ldots,m}\|\mathrm{a}_i\|^2\right).$$
Note that, without loss of generality, we can assume, for every
$i\in\{1,\ldots,m\}$, $\|\mathrm{a}_i\|=1$, since
$\mathrm{f}_i(\mathrm{a}_i^{\top}\mathrm{x})=
\mathrm{g}_i((\mathrm{a}_i/\|\mathrm{a}_i\|)^{\top}\mathrm{x})$
with $\mathrm{g}_i\colon \mathrm{x}\mapsto
\mathrm{f}_i(\|\mathrm{a}_i\|\mathrm{x})$ and
$\mathbf{prox}_{\mathrm{g}_i}\colon\mathrm{x}\mapsto
\mathbf{prox}_{\|\mathrm{a}_i\|^2\mathrm{f}_i}
(\|\mathrm{a}_i\|\mathrm{x})/\|\mathrm{a}_i\|$. Therefore, condition
\eqref{e:condERM1} can be reduced to
$\sqrt{m}+m\sigma_0<(\max_{i=0,\ldots,m}\sigma_i)^{-1}$, which, in
the
case $\sigma_0=\cdots=\sigma_m$ reduces to
$\sigma_0<(\sqrt{5}-1)/(2\sqrt{m})$.
{
\subsection{A Distributed Operator Splitting Scheme with Time-Varying Networks}
In this section we develop an extension of the popular distributed operator splitting
scheme \emph{PG-Extra}~\cite{shi2015extra,shi2015proximal} to time-varying graphs. The
problem data are a collection of cost functions $f_1, \ldots, f_n$ on a Hilbert space ${\mathcal{H}}$
and a sequence of connected, undirected communication graphs $ G_t = (V_t, E_t)$ with
vertices $V_t = \{1, \ldots, n\}$ and edges $ E_t \subseteq \{1, \ldots, n \}^2$. Then the goal
of distributed optimization is to
\begin{align}\label{eq:logs}
\Min_{x \in {\mathcal{H}}} \; \sum_{i =1}^n f_i(x),
\end{align}
through an iterative algorithm that, at every time $t \in \ensuremath{\mathbb N}$, only allows communication between neighbors in $G_t$. For simplicity, we focus on the case wherein
$f_i : {\mathcal{H}} \rightarrow \ensuremath{\left]-\infty,+\infty\right]}$ is proper, lower semicontinuous and convex.
A well-known distributed operator splitting schemes is known as \emph{PG-Extra}. This method applies to fixed communicated graphs $G_t \equiv G$, and can be viewed as an instance of modern primal-dual algorithms, such as Condat-Vu~\cite{condat2013primal,vu2013splitting}. To the best of our knowledge there is no known extension of \emph{PG-Extra} to time-varying graphs that may also be applied to monotone inclusions. We will now develop such an extension.
For the graph $G_t$, let $A_t$ denote its adjacency matrix and let $D_t$ denote its degree matrix.\footnote{Briefly, $(A_t)_{ij} = 1$ if $(i,j) \in E$ and is zero otherwise, while $D$ is a diagonal matrix with diagonal entries $D_{ii} = \text{deg}(i)$.} The Laplacian matrix of $G_t$ is defined as the difference
$$
L_t := D_t - A_t.
$$
It is well-known that, for fully connected graphs, we have the identity $\ker(L_t) = \mathrm{span}(\mathbf{1}_n)$~\cite{chung1997spectral}. We may exploit this fact to develop an equivalent formulation of~\eqref{eq:logs}.
The Laplacian operator has a natural extension to the product space ${\mathcal{H}}^n$. It is then a straightforward exercise to show that the extension induces the following identity:
\begin{align*}
\left(\forall \mathbf{x} := (x_1, \ldots, x_n) \in {\mathcal{H}}^n\right) &&L_t\mathbf{x} = 0 \iff x_1 = x_2 = \cdots = x_n.
\end{align*}
Therefore, a family of equivalent formulations of~\eqref{eq:logs} is given by
\begin{align*}
\Min_{\mathbf{x} \in {\mathcal{H}}^n} &\; \sum_{i \in V} f_i(x_i) \\
\text{subject to:} & \; L_t \mathbf{x} = 0. \numberthis\label{eq:new_reform}
\end{align*}
The constraint $L_t\mathbf{x} = 0$ is equivalent to the constraint $x \in {\mathcal{U}} := \{x \in {\mathcal{H}}^n
\mid x_1 = \ldots = x_n\}$. Thus, one could apply a splitting method to derive a distributed
algorithm consisting of decoupled proximal steps on the $f_i$ followed by \emph{global
averaging} steps induced by the projection onto ${\mathcal{U}}$. However, in order to develop an
algorithm that respects the \emph{local communication structure} of the graphs $G_t$, we
must avoid computing such projections onto ${\mathcal{U}}$. For any fixed $t$, we may develop
such a method as a special case of modern primal-dual algorithms.
Indeed, a straightforward application of Condat-Vu~\cite{condat2013primal,vu2013splitting} yields the update rule
\begin{align*}
\text{For all $i \in V$} & \text{ in parallel} \\
x_i^{k+1} &= \mathbf{prox}_{\gamma f_i}(x_i^{k} -\gamma (L_t\mathbf{y}^k)_i)\\
\mathbf{y}^{k+1} &= \mathbf{y}^{k} + \tau L_t (2\mathbf{x}^{k+1} - \mathbf{x}^k), \numberthis\label{eq:condat_vu_decentralized}
\end{align*}
where $\gamma, \tau > 0$ are appropriately chosen stepsizes. This algorithm is fully decentralized because multiplications by $L_t$ only induce communication among neighbors in the graph $G_t$.
If we allow $t = k$, this Condat-Vu~\cite{condat2013primal,vu2013splitting} algorithm has, to the best of our knoweldge, no supporting convergence theory, although each of the optimization problems~\eqref{eq:new_reform} have the same set of solutions. The lack of convergence theory arises because Condat-Vu measures convergence in the product space $({\mathcal{H}}^n \times {\mathcal{H}}^n, \|\cdot\|_{P_t})$, where $P_t$ is a metric inducing linear transformation depending on $L_t$:
\begin{align*}
P_t &:= \begin{bmatrix}
\frac{1}{\gamma} \ensuremath{\operatorname{Id}}\, & -L_t \\
-L_t & \frac{1}{\tau} \ensuremath{\operatorname{Id}}\,
\end{bmatrix}.
\end{align*}
One may hope to apply standard variable metric operator-splitting
schemes~\cite{combettes2012variable,vu2013variableFBF}, but the compatibility condition
cannot hope to be satisfied. Thus, instead of Condat-Vu, we apply the variable metric
technique developed in this manuscript.
Mathematically, we let
\begin{align*}
{\mathcal{S}}_t : {\mathcal{H}}^n \times {\mathcal{H}}^n & \rightarrow {\mathcal{H}}^{n} \times {\mathcal{H}}^n \\
(\mathbf{x}, \mathbf{y}) &\mapsto ( (\mathbf{prox}_{\gamma f_i}(x_i^{k} - \gamma (L_ty^k)_i))_{i=1}^n , \mathbf{y}^{k} + \tau L_t (2\mathbf{x}^{k+1} - \mathbf{x}^k)).
\end{align*}
Given a proper choice of $\gamma$ and $\tau$, the results of~\cite{condat2013primal,vu2013splitting} show that ${\mathcal{S}}_t$ is of $\mathfrak{T}$-class in the space $({\mathcal{H}}^n \times {\mathcal{H}}^n, \|\cdot \|_{P_t})$ (indeed, ${\mathcal{S}}_t$ is a resolvent). Thus, for any $0 < \mu \leq \|P_t\|^{-1}$, Proposition~\ref{prop:classT} implies that
$$
{\mathcal{Q}}_t = \ensuremath{\operatorname{Id}}\, - \mu P_t(\ensuremath{\operatorname{Id}}\, - {\mathcal{S}}_t),
$$
is of $\mathfrak{T}$-class in the space $({\mathcal{H}}^n \times {\mathcal{H}}^n, \|\cdot \|)$ and $\ensuremath{\operatorname{Fix}}({\mathcal{Q}}_t) = \ensuremath{\operatorname{Fix}}({\mathcal{S}}_t)$. Like ${\mathcal{S}}_t$, the operator ${\mathcal{Q}}_t$ may be computed in a decentralized fashion, as communication between agents is only induced through multiplications by $L_t$.
The algorithm resulting from applying $Q_t$ is a time-varying distributed operator-splitting scheme:
\begin{align*}
(\mathbf{x}^{k+1}, \mathbf{y}^{k+1}) = {\mathcal{Q}}_k (\mathbf{x}^{k}, \mathbf{y}^{k}).
\end{align*}
The convergence of this iteration may be proved using an argument similar to Theorem~\ref{cor:asymmetricnoinversion} (which does not capture the case in which the operator at hand is varying). To prove convergence of this iteration, one must observe that the $\ensuremath{\operatorname{Fix}}({\mathcal{Q}}_k)$ is constant, that for all $(\mathbf{x}^\ast, \mathbf{y}^\ast) \in \ensuremath{\operatorname{Fix}}(Q_k)$ the sequence $\|((\mathbf{x}^{k}, \mathbf{y}^{k}) - (\mathbf{x}^\ast, \mathbf{y}^\ast)\|$ is nonincreasing, and that $\sum_{k=0}^\infty \| (\mathbf{x}^{k+1}, \mathbf{y}^{k+1}) - (\mathbf{x}^{k}, \mathbf{y}^{k})\|^2 < \infty $. A standard argument then shows that $(\mathbf{x}^{k}, \mathbf{y}^{k})$ converges to an element of $\ensuremath{\operatorname{Fix}}({\mathcal{Q}}_k) \equiv \ensuremath{\operatorname{Fix}}({\mathcal{Q}}_0)$.
\subsection{Nonlinear constrained optimization problems}
\label{sec:NLMP}
In this application we aim at solving the nonlinear constrained optimization problem
\begin{equation}
\label{e:nonlMP}
\Min_{x\in C}{f(x)+h(x)},
\end{equation}
where $C=\menge{x\in\ensuremath{{\mathcal H}}}{(\forall i\in\{1,\ldots,p\})\quad g_i(x)\le 0}$, $f\colon\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{\left]-\infty,+\infty\right]}$
is lower semicontinuous, convex and proper, for every $i\in\{1,\ldots,p\}$,
$g_i\colon\ensuremath{\operatorname{dom}}(g_i)\subset \ensuremath{{\mathcal H}}{\tilde{o}}\mathbb{R}$ and $h\colon\ensuremath{{\mathcal H}}{\tilde{o}}\mathbb{R}$ are $\mathcal{C}^1$
convex functions in $\ensuremath{\operatorname{int}}\ensuremath{\operatorname{dom}} g_i$ and $\ensuremath{{\mathcal H}}$, respectively, and $\nabla h$ is
$\beta^{-1}-$Lipschitz. A
solution of the optimization
problem \eqref{e:nonlMP}
can be found via the saddle points of the Lagrangian
\begin{equation}
L(x,u)=f(x)+h(x)+u^{\top}{g(x)}-\iota_{\mathbb{R}^p_+}(u),
\end{equation}
which, under standard qualification conditions can be found by solving the monotone
inclusion (see \cite{rockafellar1970saddle})
\begin{equation}
{\rm find}\quad x\in Y\quad \text{ such that }\quad (\ensuremath{\exists\,} u\in\mathbb{R}_+^p)\quad (0,0)\in
A(x,u)+B_1(x,u)+B_2(x,u),
\end{equation}
where $Y\subset\ensuremath{{\mathcal H}}$ is a nonempty closed convex set modeling apriori information on the
solution (eventually we can take $Y=\ensuremath{{\mathcal H}}$), $A\colon (x,u)\mapsto \partial f(x)\times
N_{\mathbb{R}^p_+}u$ is maximally
monotone, $B_1\colon (x,u)\mapsto (\nabla h(x),0)$ is $\beta-$cocoercive, and
$$B_2\colon (x,u)\mapsto \left(\sum_{i=1}^pu_i\nabla g_i(x),-g_1(x),\ldots,-g_p(x)\right)$$
is nonlinear,
monotone and continuous
\cite{rockafellar1970saddle}. If $Y\subset \ensuremath{\operatorname{dom}}\partial f\subset\cap_{i=1}^p\ensuremath{\operatorname{int}}\ensuremath{\operatorname{dom}} g_i$
we have that $X:=Y\times\mathbb{R}_+^p\subset \ensuremath{\operatorname{dom}} A=\ensuremath{\operatorname{dom}}\partial f\times \mathbb{R}_+^p\subset
\ensuremath{\operatorname{dom}} B_2=\cap_{i=1}^p\ensuremath{\operatorname{int}}\ensuremath{\operatorname{dom}} g_i\times\mathbb{R}^p$ and, from
\cite[Corollary~25.5]{bauschke2017convex},
we have that $A+B_2$ is maximally monotone.
The method proposed in Theorem~\ref{t:1} reduces to
\begin{equation}
\label{e:algonlc}
(\forall k\in\ensuremath{\mathbb N})\quad
\begin{array}{l}
\left\lfloor
\begin{array}{l}
y^k=\mathbf{prox}_{\gamma_k f}\left(x^k-\gamma_k(\nabla h(x^k)+\sum_{i=1}^pu_i^k\nabla
g_i(x^k))\right)\\[0.5mm]
\text{For every } i=1,\ldots,p\\[0.5mm]
\left\lfloor
\begin{array}{l}
\eta_i^k=\max\left\{0,u_i^k+\gamma_kg_i(x^k)\right\}\\[0.5mm]
u_i^{k+1}=\max\left\{0,\eta_i^k-\gamma_k(g_i(x^k)-g_i(y^k))\right\}\\[0.5mm]
\end{array}
\right.\\[4mm]
x^{k+1}=P_Y\left(y^k+\gamma_k\sum_{i=1}^p(u_i^k\nabla
g_i(x^k)-\eta_i^k\nabla
g_i(y^k))\right),
\end{array}
\right.
\end{array}
\end{equation}
where, for every $k\in\ensuremath{\mathbb N}$, $\gamma_k$ is found by the backtracking procedure
defined in \eqref{e:armijocond}. Note that, since $B_2$ is nonlinear, the approaches
proposed in
\cite{condat2013primal,vu2013splitting} cannot be applied to this instance.
In the particular instance when $f=\iota_{\Omega}$ for
some
nonempty closed convex set $\Omega$, we can choose, among other options,
$Y=\Omega$ since we know
that any solution must belong to $\Omega$.
Moroever, when, for every $i\in\{1,\ldots,p\}$, $g_i\colon x\mapsto d_i^{\top}x$, where
$d_i\in\mathbb{R}^N$, we have $B_2\colon (x,u)\mapsto (D^{\top}u,-Dx)$, where
$D=[d_1,\ldots,d_p]^{\top}$. This is a particular instance of problem
\eqref{eq:primal_before_PD} and $B_2$ is $\|D\|-$Lipschitz in this case, which allows us to
use constant stepsizes $\gamma_k=\gamma\in]0,\chi[$, where $\chi$ is defined in
\eqref{e:chi} and $L=\|D\|$.
Theorem~\ref{t:1} guarantees the convergence of the
iterates $\{x^k\}_{k\in\ensuremath{\mathbb N}}$ thus generated to a solution to \eqref{e:nonlMP} in any case.
In the next section, we explore some numerical results showing the good performance
of this method and the method with constant step-size when $g_i$ are affine linear.
}
\section{Conclusion}
In this paper, we systematically investigated a new extension of Tseng's
forward-backward-forward method and the forward-backward method. The three primary
contributions of this investigation are (1) a lower per-iteration complexity variant of
Tseng's method which activates the cocoercive operator only once; (2) the ability to
incorporate variable metrics in operator-splitting
schemes, which, unlike typical variable metric methods, do not enforce
compatibility conditions between metrics employed at successive time steps; and (3)
the ability to incorporate modified resolvents $J_{P^{-1} A}$ in iterative fixed-point
algorithms, which, unlike typical preconditioned fixed point iterations, can be formed
from non self-adjoint linear operators $P$, which lead to new
Gauss-Seidel style operator-splitting schemes.
\vspace{.3cm}
{\footnotesize {\bf Acknowledgments:} This work is partially supported by NSF GRFP grant
DGE-0707424,
by CONICYT grant FONDECYT 11140360, and by ``Programa de financiamiento
basal'' from CMM, Universidad de Chile. We want to thank the two anonymous reviewers,
whose comments and concerns allowed us to improve the quality of this manuscript.}
\bibliographystyle{siam}
\section{Numerical simulations}
\label{sec:7}
In this section we provide two instances of Section~\ref{sec:NLMP} and we compare
our proposed method with available algorithms in the literature.
\subsection{Optimization with linear inequalities}
In the context of problem~\eqref{e:nonlMP}, suppose that $\ensuremath{{\mathcal H}}=\mathbb{R}^N$, $h\colon
x\mapsto \|Ax-b\|^2/2$, $A$ is a $m\times N$ real matrix with $N=2m$ and $b\in\mathbb{R}^m$,
$f=\iota_{[0,1]^N}$, and
$$(\forall i\in\{1,\ldots,p\})\quad g_i(x)=d_i^{\top}x,$$
where $d_1,\ldots,d_p\in\mathbb{R}^N$. In this case, $B_1\colon (x,u)\mapsto (A^{\top}(Ax-b),0)$,
$B_2\colon (x,u)\mapsto (D^{\top}u,-Dx)$, where $D=[d_1,\ldots, d_p]^{\top}$,
$\beta=\|A\|^{-2}$ and $L=\|D\|$.
We compare the
method proposed in \eqref{e:algonlc} using the line search (FBHF-LS),
the version with constant stepsize (FBHF), the method proposed by
Condat and V\~u \cite{condat2013primal,vu2013splitting} (CV), the method proposed by
Tseng \cite{tseng2000modified} with line search (Tseng-LS) and with constant stepsize
(Tseng) for randomly generated matrices and vectors $A$, $D$ and $b$.
We choose the same starting point for each method
and the parameters for the line search for Tseng-LS and FBHF-LS are $\theta=0.316$,
$\varepsilon=0.88$ and $\sigma=0.9$. For the constant stepsizes versions of Tseng and
FBHF, we use $\gamma=\delta/(\beta^{-1}+L)$ and
$\gamma=\delta\beta/(1+\sqrt{1+16\beta^2L^2})$, respectively, and for $\bar{\sigma}>0$ we
select $\tau=1/(1/2\beta+\bar{\sigma} L^2)$
in order to satisfy
the convergence conditions on the parameters of each algorithm.
We choose several values of $\delta$ and $\bar{\sigma}$ for studying the behavior and
we use the stopping criterion
$\|(x_{k+1}-x_k,u_{k+1}-u_k)\|/\|(x_k,u_k)\|<10^{-7}$.
In Table~\ref{tab:0} we show the performance of the five algorithms
for random matrices $A$ and $D$ and a random vector $b$ with $N=2000$ and
$p=100$ and a selection of the parameters $\sigma,\delta$. We see that for Tseng and
FBHF
the performance improve for larger choices of $\delta$, while for CV it is not clear how to
choose $\bar{\sigma}$ in general. Even if the theoretical bound of FBHF does not permit
$\delta$ to go beyond $4$, for $\delta=4.4$ the convergence is also obtained for this case
with a better performance.
We suspect that the particular structure of this particular case can be exploited for
obtaining a
better bound. We also observe that the best performance in time is obtained for the lowest
number of iterations for each method. In addition, for this instance, algorithms
FBHF, CV and FBHF-LS are comparable in computational time, while the algorithms by
Tseng \cite{tseng2000modified} are considerably less efficient in time and in number of
iterations. In Table~\ref{tab:01} we compare the average time and iterations that the more
efficient methods in the first simulation take to achieve the stop criterion
($\epsilon=10^{-7}$) for $20$ random realizations of matrices $A$ and $D$ and a random
vector $b$, with $N=2000$ and
$p=100$. We use the parameters yielding the best performance of each method in the first
simulation. For FBHF we also explore the case when
$\delta=4.7$, which gives the best performance. We also observe that FBHF for
$\delta=3.999$ is comparable with CV in average time, while FBHF-LS is slower for this
instance.
\begin{table}[]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{|l||l|l|l||l|l|l||l|l|l|l||l||l|}
\hline
$\epsilon=10^{-7}$ & \multicolumn{3}{c||}{Tseng} &
\multicolumn{3}{c||}{FBHF}
& \multicolumn{4}{c||}{CV} & Tseng-LS & FBHF-LS \\ \hline
$\delta$,$\bar{\sigma}$ & 0.8 & 0.9 & 0.99 & 3.2 & 3.99 & 4.4 &
0.0125 &
0.0031 & 0.0008 & 0.0002 & LS & LS \\ \hline
$h(x^*)$ & 158.685 & 158.684 & 158.684 & 158.681 & 158.680 & 158.679 &
158.674 &
158.674 & 158.676 & 158.687 & 158.683 & 158.680 \\ \hline
iter. & 20564 & 18482 & 16791 & 11006 & 8915 & 8243 & 9384 &
9158 & 8516 & 13375 & 14442 & 10068 \\ \hline
time (s) & 41.55 & 37.18 & 33.76 & 13.11 & 10.48 & 9.76 & 10.70 & 10.61
& 9.67 & 15.27 & 94.86 & 12.40 \\ \hline
\end{tabular}%
}
\vspace{0.2cm}
\caption{Comparison of Tseng, FBHF and CV (with different values of $\delta$ and
$\bar{\sigma}$),
Tseng-LS and FBHF-LS for a stop criterion of $\epsilon=10^{-7}$.}
\label{tab:0}
\end{table}
\begin{table}[]
\centering
\begin{tabular}{|l|l|l|}
\hline
$\epsilon=10^{-7}$ & av. iter. & av. time (s) \\ \hline
FBHF-LS & 36225 & 43.96 \\ \hline
FBHF ($\delta=3.999$) & 32563 & 39.07 \\ \hline
FBHF ($\delta=4.7$) & 28364 & 34.14 \\ \hline
CV ($\sigma=0.0008$) & 33308 & 38.60 \\ \hline
\end{tabular}
\caption{Average performance of the more efficient methods for $20$ random
realizations of $A$, $D$ and $b$ with $N=2000$ and $p=100$.}
\label{tab:01}
\end{table}
\subsection{Entropy constrained optimization}
\label{sec:71}
In the context of problem~\eqref{e:nonlMP}, suppose that $\ensuremath{{\mathcal H}}=\mathbb{R}^N$, $h\colon
x\mapsto x^{\top}Qx-d^{\top}x+c$, $Q$ is a $N\times N$ semidefinite positive real
matrix, $b\in\mathbb{R}^N$, $c\in\mathbb{R}$, $f=\iota_{\Omega}$,
$\Omega$ is a closed convex subset of $\mathbb{R}^N$,
$p=1$, and $$g_1\colon\mathbb{R}^N_+{\tilde{o}}\mathbb{R}\colon x\mapsto
\sum_{i=1}^Nx_i\left(\ln\left(\frac{x_i}{a_i}\right)-1\right)-r,
$$
where $-\sum_{i=1}^Na_i<r<0$, $a\in\mathbb{R}^N_{++}$ and we use the convention $0\ln(0)=0$.
This problem appears
in robust least squares estimation when a relative entropy constraint is included
\cite{levy04}.
This constraint can be seen as a distance constraint with respect to the vector $a$, where
the distance is measured by the Kullback-Leibler divergence \cite{basseville13}.
In our numerical experience, we assume $Q=A^{\top}A$, $d=A^{\top}b$ and $c=\|b\|^2/2$,
where $A$ is a $m\times N$ real matrix with $N=2m$ and $b\in\mathbb{R}^m$, which yields
$h\colon x\mapsto \|Ax-b\|^2/2$, $\beta=\|A\|^{-2}$, $\Omega=[0.001,1]^N$, and
$a=(1,\ldots,1)^{\top}$.
In this context, $g_1$ achieves its minimum in $\bar{x}=(1,\ldots,1)^{\top}$ and
$g_1(\bar{x})=-N$
and we choose $r\in]-N,0[$.
Since the constraint is not linear, we cannot
use the methods proposed in \cite{vu2013splitting,condat2013primal}. We compare the
method proposed in \eqref{e:algonlc} with line search (FBHF-LS) with the Tseng's method
with linesearch
\cite{tseng2000modified} (Tseng-LS) and two routines in matlab: {{\tilde{t}} fmincon.interior-point}
(FIP) and
{{\tilde{t}} fmincon.sqp} (SQP). For $m=100, 200, 300$, we generate $20$ random matrices $A$
and random vectors $b$ and we compare the previous methods by changing
$r\in\{-0.2N,-0.4N,-0.6N,-0.8N\}$
in order to vary the feasible regions. We choose the same starting point for each method
and the parameters for the line search for Tseng-LS and FBHF-LS are $\theta=0.707$,
$\varepsilon=0.88$ and $\sigma=0.9$.
The stopping criterion is
$\|(x_{k+1}-x_k,u_{k+1}-u_k)\|/\|(x_k,u_k)\|<\epsilon$ with $\epsilon=10^{-11}$. In
Table~\ref{tab:1} we show,
for $m=300$,
the value of the objective function $h$, the nonlinear constraint $g_1$ and time for
achieving the stopping criterion for a fixed random matrix $A$ and vector $b$
by moving $r\in\{-0.2N,-0.4N,-0.6N,-0.8N\}$. We observe that all methods achieve
almost the same value of the objective function and satisfy the constraints, but
in time FBHF-LS obtains the best performance, even if the number of iterations are larger
than
that of FIP and SQP. Tseng-LS has also a better performance in time than FIP and SQP, with
a much larger number of iterations. We also observe that, the smaller the feasible
set is, the harder is for all the methods to approximate the solution and the only case
when the constraint is inactive is when $r=-0.2N$. On the other hand, even if in the cases
$r=-0.6N$ and $r=-0.8N$ we have $g_1(x^*)>0$, the value is $\approx 10^{-6}$ which is
very near to feasibility.
This behavior is confirmed in Table~\ref{tab:2}, in which we show,
for each $m\in\{100,200,300\}$, the average time and iterations
obtained from the $20$ random realizations by moving
$r\in\{-0.2N,-0.4N,-0.6N,-0.8N\}$. We observe that FBHF-LS takes considerably less time
than the other algorithms to reach the stopping criterion and the difference is more when
dimension is higher.
Since FIP and SQP are very slow for high dimensions, in Table~\ref{tab:3}
we compare the efficiency of Tseng-LS and FBHF-LS for $20$ random realizations
of $A$ and $b$ with $N\in\{1000,2000,3000\}$ for $r=-0.4N$ and $\epsilon=10^{-5}$.
The computational time of both methods are reasonable, but again FBHF-LS is faster.
FBHF-LS use less iterations than Tseng-LS for achieving the same criterion and, even if
we reduce $\epsilon$ from $10^{-5}$ to $10^{-10}$ and the number of iterations are more
than 3 times that of Tseng-LS for the weaker criterion, the computational time is similar.
We also observe that the percentage of relative improvement of an algorithm $A$ with
respect to Tseng-LS,
measured via $\% imp.(A)=100*(f(x_A)-f(x_T))/f(x_T)$, where $x_T$ and $x_A$ are the
approximative solutions obtained by Tseng-LS and $A$, is bigger for smaller dimensions.
For instance, in the case $500\times 1000$, FBHF-LS obtain an approximative solution
for which the objective function has a $12\%$ of relative improvement with respect to that
of
Tseng-LS for $\epsilon=10^{-5}$ and, if the criterion is strengthened to $10^{-10}$, the
improvement raises to $20\%$. For higher dimensions, this quantities are considerably
reduced.
\begin{table}[]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{|l||l|l|l|l||l|l|l|l||l|l|l|l||l|l|l|l|}
\hline
& \multicolumn{4}{c||}{$r=-0.2N$} &
\multicolumn{4}{c||}{$r=-0.4N$} & \multicolumn{4}{c||}{$r=-0.6N$} &
\multicolumn{4}{c|}{$r=-0.8N$} \\ \hline
$300\times600$ & $h(x^*)$ & $g_1(x^*)$ & time (s) &iter. & $h(x^*)$ & $g_1(x^*)$
&
time (s) & iter.& $h(x^*)$ & $g_1(x^*)$ & time (s) &iter.& $h(x^*)$ & $g_1(x^*)$ &
time (s)&iter. \\ \hline
FIP & 6.06E-10 & -105.038& 165.190 & 562 & 3.73E-09 & -7.78E-02 &
183.467 & 574 & 244.551 & -3.39E-08 & 218.413 & 794 & 4075.6824 &
-3.20E-09 & 378.269 & 1197 \\
SQP & 6.06E-10 & -105.038 & 372.210 & 357 & 3.73E-09 & -7.78E-02 &
598.258 & 341 & 244.551 & -3.39E-08 & 515.568 & 653 & 4075.6824 &
-3.20E-09 & 988.143 & 655 \\
Tseng-LS & 5.47E-15 & -119.365 & 13.682 & 13785 & 1.41E-14 & -9.81E-09 &
29.160 & 30110 & 244.551 & 1.16E-06 & 44.248 & 20717 & 4075.6822 &
2.67E-06 & 110.916 & 75254 \\
FBHF-LS & 2.15E-15 & -119.338 & 1.053 & 9680 & 5.56E-15 & -4.71E-09 &
2.220 & 21106 & 244.551 & 1.10E-06 & 10.381 & 19492 & 4075.6822 &
2.07E-06 & 17.464 & 60442\\
\hline
\end{tabular}%
}
\vspace{.2cm}
\caption{Comparison of objective function and constraints values, time and number of
iterations of FIP, SQP, Tseng-LS and FBHF-LS algorithms for solving the entropy
constrained
optimization when $N=600$, $m=300$ and $r\in
\{-0.2N,-0.4N,-0.6N,-0.8N\}$.}
\label{tab:1}
\end{table}
\begin{table}[]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{|l||l|l|l|l||l|l|l|l||l|l|l|l|}
\hline
Time (s) & \multicolumn{4}{c||}{$100\times 200$} &
\multicolumn{4}{c||}{$200\times
400$}
&
\multicolumn{4}{c|}{$300\times 600$} \\ \hline
constraint & $r=-0.2N$ & $r=-0.4N$ & $r=-0.6N$ & $r=-0.8N$ & $r=-0.2N$ &
$r=-0.4N$ & $r=-0.6N$
& $r=-0.8N$ & $r=-0.2N$ & $r=-0.4N$ & $r=-0.6N$ & $r=-0.8N$ \\ \hline
FIP & 8.24 & 9.50 & 11.92 & 11.22 & 52.95 & 57.92 & 75.02 & 76.29 &
142.51 & 183.22 & 253.80 & 324.42 \\
SQP & 8.60 & 11.18 & 14.28 & 18.51 & 70.88 & 98.70 & 122.03 & 209.73
& 313.52 & 489.95 & 569.34 & 1075.42 \\% \hline
Tseng-LS & 22.92 & 10.71 & 7.92 & 9.91 & 81.16 & 13.46 & 39.17 & 83.01
& 139.47 & 26.50 & 84.07 & 111.44 \\
FBHF-LS & 2.21 & 0.99 & 2.72 & 2.51 & 7.06 & 0.95 & 10.23 & 18.09
& 12.48 & 1.88 & 20.59 & 18.30 \\ \hline
\end{tabular}%
}
\vspace{.2cm}
\caption{Average time (s) to reach the stopping criterion of $20$ random realizations for
FIP,
SQP, Tseng-LS and FBHF-LS for a matrix $A$ with
dimension $100\times 200$, $200\times 400$ and $300\times 600$ and $r\in
\{-0.2N,-0.4N,-0.6N,-0.8N\}$.}
\label{tab:2}
\end{table}
\begin{table}[]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{|l|l||l|l|l||l|l|l||l|l|l|}
\hline
& & \multicolumn{3}{c||}{$500\times1000$} &
\multicolumn{3}{c||}{$1000\times2000$} & \multicolumn{3}{c|}{$1500\times3000$} \\
\hline
stop crit. & Algorithm & time (s) & iter. & \%
imp. & time (s) & iter. & \% imp. & time (s) & iter. & \% imp. \\ \hline
\multicolumn{1}{|c|}{\multirow{2}{*}{$10^{-5}$}} & Tseng-LS & 17.22 & 2704 &
0 & 52.80 & 4027 & 0 & 91.23 & 3349 & 0 \\
\multicolumn{1}{|c|}{} & FBHF-LS & 2.33 & 1993 &
12.2 & 6.26 & 3239 & 3.2 & 9.56 & 2474 & 0.1 \\ \hline
\multicolumn{1}{|c|}{$10^{-10}$} & FBHF-LS & 10.67 & 10092
&
20.1 & 71.51
&
33637 & 6.5 & 53.47 & 12481 & 0.2 \\ \hline
\end{tabular}%
}
\vspace{.2cm}
\caption{Comparison between Tseng-LS and FBHF-LS for higher dimensions. We compare
average time and average number of iterations for achieving stop criteria together with
percentage of relative improvement with respect to Tseng-LS approximate solution.}
\label{tab:3}
\end{table}
}
\section{Primal-dual composite
monotone inclusions with non self-adjoint linear operators}
\label{sec:5}
In this section, we apply our algorithm to composite primal-dual monotone
inclusions involving a cocoercive and a lipschitzian monotone
operator.
\begin{problem}
\label{prob:mi}
Let ${\mathrm H}$ be a real Hilbert space, let ${\mathrm
X}\subset {\mathrm H}$ be closed and convex, let $z\in{\mathrm
H}$, let ${\mathrm A}\colon {\mathrm H}
{\tilde{o}}
2^{\mathrm H}$ be maximally monotone, let
${\mathrm C}_1 \colon {\mathrm H} {\tilde{o}}{\mathrm H}$ be
$\mu$-cocoercive, for some $\mu\in\ensuremath{\left]0,+\infty\right[}$, and let ${\mathrm C}_2
\colon {\mathrm H} {\tilde{o}}{\mathrm H}$ be a monotone and
$\delta$-lipschitzian operator, for some $\delta\in\ensuremath{\left]0,+\infty\right[}$. Let $m\geq 1$ be an
integer,
and, for
every
$i\in\{1,\ldots,m\}$, let ${\mathrm G}_i$ be a real Hilbert
space, let $r_i\in{\mathrm G}_i$, let ${\mathrm B}_i\colon
{\mathrm G}_i {\tilde{o}} 2^{{\mathrm G}_i}$ be maximally monotone,
let ${\mathrm D}_i\colon {\mathrm G}_i {\tilde{o}} 2^{{\mathrm G}_i}$ be
maximally monotone and $\nu_i$-strongly monotone, for some
$\nu_i\in\ensuremath{\left]0,+\infty\right[}$, and suppose that ${\mathrm L}_i\colon {\mathrm
H} {\tilde{o}} {\mathrm G}_i$ is a nonzero linear bounded operator. The
problem is to solve the primal inclusion.
\begin{equation}
\label{e:miprimal}
\text{find }\quad {\mathrm x}\in {\mathrm X}\quad \text{such that
}\quad {\mathrm z}\in
{\mathrm A}{\mathrm x}+\sum_{i=1}^m{\mathrm
L}_i^*({\mathrm B}_i\ensuremath{\mbox{\small$\,\square\,$}} {\mathrm D}_i)(\mathrm
{L}_i\mathrm
{x}-\mathrm
{r}_i)+\mathrm{C}_1\mathrm{x}+\mathrm{C}_2\mathrm{x}
\end{equation}
together with the dual inclusion
\begin{align*}
&\text{find }\quad \mathrm{v}_1\in
\mathrm{G}_1,\ldots,\mathrm{v}_m\in
\mathrm{G}_m\quad \\
&\text{such that
}\quad
(\exists \mathrm{x\in
X})\:
\begin{cases}
\mathrm{z}-\sum_{i=1}^m\mathrm{L}_i^*\mathrm{v}_i\in
\mathrm{Ax+C}_1\mathrm{x}+\mathrm{C}_2\mathrm{x}\\
(\forall i\in\{1,\ldots,m\})\:\:\mathrm{v}_i\in({\mathrm
B}_i\ensuremath{\mbox{\small$\,\square\,$}} {\mathrm D}_i)(\mathrm
{L}_i\mathrm
{x}-\mathrm
{r}_i) \numberthis\label{e:midual}
\end{cases}
\end{align*}
under the assumption that a solution exists.
\end{problem}
In the case when ${\mathrm X}={\mathrm H}$ and ${\mathrm
C}_2=0$,
Problem~\ref{prob:mi} is studied in
\cite{vu2013splitting}\footnote{Note that in \cite{vu2013splitting},
weights
$(\omega_i)_{1\leq i\leq m}$ multiplying operators
$(\mathrm{B}_i\ensuremath{\mbox{\small$\,\square\,$}}\mathrm{D}_i)_{1\leq i\leq m}$ are considered.
They can be retrieved in \eqref{e:miprimal} by considering
$(\omega_i\mathrm{B}_i)_{1\leq i\leq m}$ and
$(\omega_i\mathrm{D}_i)_{1\leq i\leq m}$ instead of
$(\mathrm{B}_i)_{1\leq i\leq m}$ and
$(\mathrm{D}_i)_{1\leq i\leq m}$. Then both formulations are
equivalent.}
and
models a large class of problems including
optimization problems, variational inequalities, equilibrium
problems, among others (see
{\color{red}\cite{briceno2011monotone+,he2012convergence,vu2013splitting,condat2013primal}}
and
the
references therein). In \cite{vu2013splitting} the author
rewrite \eqref{e:miprimal} and \eqref{e:midual} in the case
${\mathrm X}={\mathrm H}$ as
\begin{equation}
\label{e:mipd}
\text{find}\quad {z}\in\ensuremath{{\mathcal H}}\quad\text{such that
}\quad
{0}\in
{M}{z}+{S}{z}+Q{z},
\end{equation}
where $\ensuremath{{\mathcal H}}=\mathrm{H\times G}_1\times\cdots\times\mathrm{G}_m$,
${M}\colon\ensuremath{{\mathcal H}}{\tilde{o}}
2^{\ensuremath{{\mathcal H}}}\colon (\mathrm{x},\mathrm{v}_1,\ldots,
\mathrm{v}_m)\mapsto (\mathrm{Ax-z})\times
(\mathrm{B}_1^{-1}\mathrm{v}_1+\mathrm{r}_1)\times\cdots\times
(\mathrm{B}_m^{-1}\mathrm{v}_m+\mathrm{r}_m)$ is maximally monotone,
${S}\colon
\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}\colon (\mathrm{x},\mathrm{v}_1,\ldots,
\mathrm{v}_m)\mapsto
(\sum_{i=1}^m\mathrm{L}_i^*\mathrm{v}_i,
-\mathrm{L}_1\mathrm{x},\ldots,-\mathrm{L}_m\mathrm{x})$ is skew linear,
and
$Q\colon
\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}\colon
(\mathrm{x},\mathrm{v}_1,\ldots,\mathrm{v}_m)\mapsto
(\mathrm{C}_1\mathrm{x},\mathrm{D}_1^{-1}\mathrm{v}_1,\ldots,
\mathrm{D}_m^{-1}\mathrm{v}_m)$ is cocoercive.
If $(\mathrm{x},\mathrm{v}_1,\ldots,\mathrm{v}_m)$ is a
solution in the primal-dual space $\ensuremath{{\mathcal H}}$ to \eqref{e:mipd}, then
$\mathrm{x}$ is a solution to \eqref{e:miprimal} and
$(\mathrm{v}_1,\ldots,\mathrm{v}_m)$ is a solution to \eqref{e:midual}. The author
provide an algorithm for solving
\eqref{e:miprimal}--\eqref{e:midual} in this particular instance,
which is an application of the {\em
forward-backward} splitting (FBS) applied to the inclusion
\begin{equation}
\label{e:mipdvm}
\text{find}\quad {z}\in\ensuremath{{\mathcal H}}\quad\text{such that
}\quad
{0}\in
V^{-1}({M}+{S}){z}+V^{-1}Q{z},
\end{equation}
where $V$ is a specific symmetric strongly monotone operator.
Under the metric $\scal{V\cdot}{\cdot}$, $V^{-1}({M}+{S})$ is
maximally monotone and $V^{-1}Q$ is cocoercive and, therefore,
the FBS converges weakly to a primal-dual solution.
In order to tackle the case $\mathrm{C}_2\neq 0$, we propose to
use the method in Theorem~\ref{cor:asymmetricnoinversion} for solving
$0\in Ax+B_1x+B_2x$ where $A=M$, $B_1=Q$, $B_2=S+C_2$,
and $C_2\colon(\mathrm{x},\mathrm{v}_1,\ldots,\mathrm{v}_m)\mapsto
(\mathrm{C}_2\mathrm{x},0,\ldots,
0)$
allowing, in that way, non self-adjoint linear operators which may vary among
iterations.
The following result provides the method thus obtained, where the dependence
of the non self-adjoint linear operators with respect to iterations has been avoided for
simplicity.
\begin{theorem}
\label{p:primdumi}
In Problem~\ref{prob:mi}, set $\mathrm{X}=\mathrm{H}$, set
$\mathrm{G}_0=\mathrm{H}$, for every $i\in\{0,1,\ldots,m\}$ and
$j\in\{0,\ldots, i\}$, let
$\mathrm{P}_{ij}\colon\mathrm{G}_j{\tilde{o}}\mathrm{G}_i$ be a linear
operator satisfying
\begin{equation}
\label{e:condiPii}
(\forall
\mathrm{x}_{i}\in\mathrm{G}_{i})\quad
\scal{\mathrm{P}_{ii}\mathrm{x}_{i}}{\mathrm{x}_{i}}
\geq\varrho_i\|\mathrm{x}_{i}\|^2
\end{equation}
for some $\varrho_i>0$. Define the $(m+1)\times(m+1)$
symmetric real
matrices $\Upsilon$, $\Sigma$, and $\Delta$ by
\begin{align*}
(\forall i\in\{0,\ldots,m\})(\forall j<i)\quad \Upsilon_{ij}&=
\begin{cases}
0,\quad&\text{if }i=j;\\
\|\mathrm{P}_{ij}\|/2,&\text{if }i>j,\\
\end{cases}
\quad \\
\Sigma_{ij}&=
\begin{cases}
\|\mathrm{P}_{ii}-\mathrm{P}_{ii}^*\|/2,\quad&\text{if }i=j;\\
\|\mathrm{L}_i+\mathrm{P}_{i0}/2\|,&\text{if }i\geq 1;j=0;\\
\|\mathrm{P}_{ij}\|/2,&\text{if }i>j>0,
\end{cases}\numberthis \label{e:defUpsilon}
\end{align*}
and $\Delta={\rm Diag}(\varrho_0,\ldots,\varrho_m)$.
Assume that $\Delta-\Upsilon$
is positive definite with smallest eigenvalue $\rho>0$ and
that
\begin{equation}
\label{e:metricconditionpd}
(\|\Sigma\|_2+\delta)^2<\rho\left(\rho-\frac{1}{2\beta}\right),
\end{equation}
where $\beta=\min\{\mu,\nu_1,\ldots,\nu_m\}$.
Let
$M=\max_{i=0,\ldots,m}\|\mathrm{P}_{ii}\|+\|\Upsilon\|_2$, let
$\lambda\in]0,M^{-1}[$,
let $(\mathrm{x}^0,\mathrm{u}_1^0,\ldots,\mathrm{u}_m^0)\in
\mathrm{H\times
G}_1\times\cdots\times
\mathrm{G}_m$, and let $\{\mathrm{x}^k\}_{k\in\ensuremath{\mathbb N}}$
and
$\{\mathrm{u}_i^k\}_{k\in\ensuremath{\mathbb N}, 1\leq i\leq m}$ the sequences
generated by the
following
routine:
for every $k\in\ensuremath{\mathbb N}$
\begin{equation}
\label{e:algomicomp1}
\hspace{-.2cm}
\begin{array}{l}
\left\lfloor
\begin{array}{l}
\!\!\mathrm{y}^k
\!=\!J_{
\mathrm{P}_{00}^{-1}\mathrm{A}}
\left(\mathrm{x}^k-\mathrm{P}_{00}^{-1}\bigg(\mathrm{C}_1\mathrm{x}^k+
\mathrm{C}_2\mathrm{x}^k+\sum_{i=1}^m
\mathrm{L}_i^*\mathrm{u}_i^k\bigg)\right)\\
\!\!\mathrm{v}_1^k \!=\! J_{\mathrm{P}_{11}^{-1}\mathrm{B}_1^{-1}}
\!\left(\mathrm{u}_1^k -
\mathrm{P}_{11}^{-1}\bigg(\mathrm{D}_1^{-1}\mathrm{u}_1^k-
\mathrm{L}_1\mathrm{x}^k-\mathrm{P}_{10}(\mathrm{x}^k-\mathrm{y}^k)\bigg)\right)\\
\!\!\mathrm{v}_2^k\!=\! J_{ \mathrm{P}_{22}^{-1}\mathrm{B}_2^{-1}}
\!\left(\mathrm{u}_2^k -
\mathrm{P}_{22}^{-1}\bigg(\mathrm{D}_2^{-1}\mathrm{u}_2^k-
\mathrm{L}_2\mathrm{x}^k-\mathrm{P}_{20}(\mathrm{x}^k-\mathrm{y}^k)
-\mathrm{P}_{21}(\mathrm{u}_1^k-\mathrm{v}_1^k)\bigg)\right)\\
\vdots\\
\!\!\mathrm{v}_m^k \!=\!J_{\mathrm{P}_{mm}^{-1}\mathrm{B}_m^{-1}}
\!\!\left(\!\mathrm{u}_m^k\! -\!
\mathrm{P}_{mm}^{-1}\bigg(\!\mathrm{D}_m^{-1}\mathrm{u}_m^k\!-\!
\mathrm{L}_m\mathrm{x}^k\!-\!\mathrm{P}_{m0}(\mathrm{x}^k\!-\!\mathrm{y}^k)
\!-\!\sum_{j=1}^{m-1}\!\mathrm{P}_{mj}(\mathrm{u}_j^k\!-\!\mathrm{v}_j^k)\bigg)\!\right)\\
\!\!\mathrm{x}^{k+1}=
\mathrm{x}^k+\lambda\left(\mathrm{P}_{00}(\mathrm{y}^k-\mathrm{x}^k)+
\left(\mathrm{C}_2\mathrm{x}^k-\mathrm{C}_2\mathrm{y}^k+\sum_{i=1}^m
\mathrm{L}_i^*(\mathrm{u}_i^k-\mathrm{v}_i^k)\right)\right)\\
\! \!\mathrm{u}_1^{k+1}=\mathrm{u}_1^k+\lambda\Big(\mathrm{P}_{10}
(\mathrm{y}^k-\mathrm{x}^k)+
\mathrm{P}_{11}(\mathrm{v}_1^k-\mathrm{u}_1^k)-
\mathrm{L}_1(\mathrm{x}^k-\mathrm{y}^k)\Big)\\
\vdots\\
\!\! \mathrm{u}_m^{k+1}
=\mathrm{u}_m^k+\lambda\left(\mathrm{P}_{m0}(\mathrm{y}^k-\mathrm{x}^k)+
\sum_{j=1}^m\mathrm{P}_{mj}(\mathrm{v}_j^k-\mathrm{u}_j^k)-
\mathrm{L}_m(\mathrm{x}^k-\mathrm{y}^k)\right).
\end{array}
\right.\\[2mm]
\end{array}
\end{equation}
Then there exists a primal-dual solution
$(\mathrm{x}^*,\mathrm{u}_1^*,\ldots,\mathrm{u}_m^*)\in
\mathrm{H\times
G}_1\times\cdots\times
\mathrm{G}_m$ to
Problem~\ref{prob:mi} such that $\mathrm{x}^k\rightharpoonup
\mathrm{x}^*$ and, for every $i\in\{1,\ldots,m\}$,
$\mathrm{u}_i^k\rightharpoonup \mathrm{u}_i^*$.
\end{theorem}
\begin{proof}
Consider the real Hilbert space $\ensuremath{{\mathcal H}}=\mathrm{H\oplus
G}_1\oplus\cdots\oplus
\mathrm{G}_m$, where its scalar product and norm are denoted by
$\pscal{\cdot}{\cdot}$ and $|||\cdot|||$, respectively, and
$x=(\mathrm{x}_0,\mathrm{x}_1,\ldots,\mathrm{x}_m)$ and
$y=(\mathrm{y}_0,\mathrm{y}_1,\ldots,\mathrm{y}_m)$ denote generic
elements of $\ensuremath{{\mathcal H}}$. Similarly as in
\cite{vu2013splitting}, note that the set of
primal-dual solutions
$x^*=(\mathrm{x}^*,\mathrm{u}_1^*,\ldots,\mathrm{u}_m^*)\in\ensuremath{{\mathcal H}}$ to
Problem~\ref{prob:mi} in the case $\mathrm{X}=\mathrm{H}$ coincides
with the set of
solutions to the monotone inclusion
\begin{equation}
\text{find}\quad x\in\ensuremath{{\mathcal H}}\quad\text{such that}\quad 0\in Ax+B_1x+B_2x,
\end{equation}
where the operators ${A}\colon\ensuremath{{\mathcal H}}{\tilde{o}}
2^{\ensuremath{{\mathcal H}}}$, $B_1\colon \ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}$, and
${B}_2\colon \ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}$ {($\ensuremath{\operatorname{dom}} B_2=\ensuremath{{\mathcal H}}$)} defined by
\begin{equation}
\begin{cases}
A&\colon (\mathrm{x},\mathrm{v}_1,\ldots,
\mathrm{v}_m)\mapsto (\mathrm{Ax-z})\times
(\mathrm{B}_1^{-1}\mathrm{v}_1+\mathrm{r}_1)\times\cdots\times
(\mathrm{B}_m^{-1}\mathrm{v}_m+\mathrm{r}_m)\\
B_1&\colon
(\mathrm{x},\mathrm{v}_1,\ldots,\mathrm{v}_m)\mapsto
(\mathrm{C}_1\mathrm{x},\mathrm{D}_1^{-1}\mathrm{v}_1,\ldots,
\mathrm{D}_m^{-1}\mathrm{v}_m)\\
B_2&\colon (\mathrm{x},\mathrm{v}_1,\ldots,
\mathrm{v}_m)\mapsto
(\mathrm{C}_2\mathrm{x}+\sum_{i=1}^m\mathrm{L}_i^*\mathrm{v}_i,
-\mathrm{L}_1\mathrm{x},\ldots,-\mathrm{L}_m\mathrm{x}),
\end{cases}
\end{equation}
are maximally monotone,
$\beta$-cocoercive, and monotone-Lipschitz, respectively
{ (see \cite[Proposition~20.22\,and\,20.23]{bauschke2017convex} and
\cite[Eq. (3.12)]{vu2013splitting}).}
Now let
$P\colon\ensuremath{{\mathcal H}}{\tilde{o}} \ensuremath{{\mathcal H}}$ defined by
\begin{equation}
\label{e:def_P}
P\colon x\mapsto \left(\mathrm{P}_{00}\mathrm{x}_0,
\mathrm{P}_{10}\mathrm{x}_0+\mathrm{P}_{11}\mathrm{x}_1,\ldots,\sum_{j=0}^m
\mathrm{P}_{mj}\mathrm{x}_j\right)=\left(\sum_{j=0}^i
\mathrm{P}_{ij}\mathrm{x}_j\right)_{i=0}^m.
\end{equation}
Then $P^*\colon x\mapsto(\sum_{j=i}^m
\mathrm{P}_{ji}^*\mathrm{x}_j)_{i=0}^m$ and
$U\colon\ensuremath{{\mathcal H}}{\tilde{o}} \ensuremath{{\mathcal H}}$ and $S\colon\ensuremath{{\mathcal H}}{\tilde{o}} \ensuremath{{\mathcal H}}$ defined by
\begin{align}
\label{e:def_U}
U&\colon x\mapsto \left(\frac{1}{2}\sum_{j=0}^{i-1}
\mathrm{P}_{ij}\mathrm{x}_j+
\left(\frac{\mathrm{P}_{ii}+\mathrm{P}_{ii}^*}{2}\right)\mathrm{x}_i+
\frac{1}{2}\sum_{j=i+1}^m
\mathrm{P}_{ji}^*\mathrm{x}_j\right)_{i=0}^m\\
\label{e:def_S}
S&\colon x\mapsto \left(\frac{1}{2}\sum_{j=0}^{i-1}
\mathrm{P}_{ij}\mathrm{x}_j+
\left(\frac{\mathrm{P}_{ii}-\mathrm{P}_{ii}^*}{2}\right)\mathrm{x}_i-
\frac{1}{2}\sum_{j=i+1}^m
\mathrm{P}_{ji}^*\mathrm{x}_j\right)_{i=0}^m
\end{align}
are the self-adjoint and skew components of $P$, respectively, satisfying $P=U+S$.
Moreover, for every
$x=(\mathrm{x}_0,\mathrm{x}_1,\ldots,\mathrm{x}_m)$
in $\ensuremath{{\mathcal H}}$, we have
\begin{align}
\label{e:Ustronglymon}
\pscal{Ux}{x}&=\sum_{i=0}^m\frac{1}{2}\sum_{j=0}^{i-1}
\scal{\mathrm{P}_{ij}\mathrm{x}_j}{\mathrm{x}_i}+
\scal{\mathrm{P}_{ii}\mathrm{x}_i}{\mathrm{x}_i}+
\frac{1}{2}\sum_{j=i+1}^m
\scal{\mathrm{P}_{ji}^*\mathrm{x}_j}{\mathrm{x}_i}\nonumber\\
&=\sum_{i=0}^m\scal{\mathrm{P}_{ii}\mathrm{x}_i}{\mathrm{x}_i}
+\sum_{i=1}^m\sum_{j=0}^{i-1}\scal{\mathrm{P}_{ij}\mathrm{x}_j}{\mathrm{x}_i}
\nonumber\\
&\geq\sum_{i=0}^m\varrho_i\|\mathrm{x}_i\|^2-
\sum_{i=1}^m\sum_{j=0}^{i-1}\|\mathrm{P}_{ij}\|\,\|\mathrm{x}_i\|\,\|
\mathrm{x}_j\|
\nonumber\\
&=\xi\cdot(\Delta-\Upsilon)\xi\geq\rho|\xi|^2=\rho\,|||x|||^2,
\end{align}
where $\xi:=(\|\mathrm{x}_i\|)_{i=0}^m\in\mathbb{R}^{m+1}$, $\Upsilon$ is
defined in \eqref{e:defUpsilon}, and $\rho$ is the smallest (strictly
positive)
eigenvalue of $\Delta-\Upsilon$. In addition, we can write
$B_2-S=C_2+R$,
where $C_2\colon x\mapsto
(\mathrm{C}_2\mathrm{x},0,\ldots,0)$ is monotone and
$\delta$-lipschitzian, and $R$ is a skew
linear operator satisfying, for every
$x=(\mathrm{x}_0,\mathrm{x}_1,\ldots,\mathrm{x}_m)\in\ensuremath{{\mathcal H}}$,
$Rx=(\sum_{j=0}^mR_{i,j}\mathrm{x}_j)_{0\leq i\leq m}$, where the
operators
$R_{i,j}\colon\mathrm{G}_j{\tilde{o}} \mathrm{G}_i$ are defined by
$R_{i,j}=-\mathrm{P}_{ij}/2$ if $i>j>0$,
$R_{i,j}=-(\mathrm{L}_{i}+\mathrm{P}_{i0})/2$ if $i>j=0$,
$R_{i,i}=(\mathrm{P}_{ii}^*-\mathrm{P}_{ii})/2$ and the other
components follow from the skew property of $R$.
Therefore,
\begin{align}
\label{e:desigS}
|||Rx|||^2\!=\!\sum_{i=0}^m\left\|\sum_{j=0}^mR_{i,j}\mathrm{x}_j\right\|^2
\!\!\!\leq\sum_{i=0}^m\left(\sum_{j=0}^m\|R_{i,j}\|\,\|\mathrm{x}_j\|\right)^{\!\!\!2}
\!\!=|\Sigma\xi|^2\!\leq\|\Sigma\|^2_2|\xi|^2\!=\|\Sigma\|^2_2|||x|||^2,
\end{align}
from which we obtain that
$B_2-S$ is $(\delta+\|\Sigma\|_2)$-lipschitzian.
Altogether, by noting that, for every $x\in\ensuremath{{\mathcal H}}$,
$\|Ux\|\leq M$, all the
hypotheses of Theorem~\ref{cor:asymmetricnoinversion} hold in this
instance and by developing \eqref{eq:FBF-asymmetric-no-U} for this
specific choices of $A$, $B_1$, $B_2$, $P$, $\gamma$, and setting,
for every $k\in\ensuremath{\mathbb N}$,
$z^k=(\mathrm{x}^k,\mathrm{u}_1^k,\ldots,\mathrm{u}_m^k)$ and
$x^k=(\mathrm{y}^k,\mathrm{v}_1^k,\ldots,\mathrm{v}_m^k)$, we obtain
\eqref{e:algomicomp1} after straighforward computations and using
\begin{equation}
{ x^k=J_{ P^{-1}A}(z^k-P^{-1}(B_1z^k+B_2z^k))\quad
\Leftrightarrow\quad P(z^k-x^k)-(B_1z^k+B_2z^k)\in Ax^k.}
\end{equation}
The result follows, hence, as a consequence of
Theorem~\ref{cor:asymmetricnoinversion}.
\end{proof}
\begin{remark}
\begin{enumerate}
\item As in Theorem~\ref{cor:asymmetricnoinversion}, the algorithm in
Theorem~\ref{p:primdumi} allows for linear operators
$(\mathrm{P}_{ij})_{0\leq i,j\leq m}$ depending on the iteration,
whenever \eqref{e:metricconditions2} holds for the corresponding
operators defined in
\eqref{e:def_P}--\eqref{e:def_S}. We omit this generalization in
Theorem~\ref{p:primdumi} for the sake of simplicity.
\item In the particular case when, for every $i\in\{1,\ldots,m\}$,
$\textrm{B}_i=\widetilde{\textrm{B}}_i\ensuremath{\mbox{\small$\,\square\,$}} \textrm{M}_i$, where
$\textrm{M}_i$ is such that $\textrm{M}_i^{-1}$ is monotone and
$\sigma_i$-Lipschitz, for some $\sigma_i>0$, Problem~\eqref{prob:mi}
can be solved
in a similar way if, instead of $B_2$ and $\delta$, we consider
$\widetilde{B}_2\colon
(\mathrm{x},\mathrm{v}_1,\ldots,
\mathrm{v}_m)\mapsto
(\mathrm{C}_2\mathrm{x}+\sum_{i=1}^m\mathrm{L}_i^*\mathrm{v}_i,
\mathrm{M}_1^{-1}\mathrm{v}_1-\mathrm{L}_1\mathrm{x},\ldots,
\mathrm{M}_m^{-1}\mathrm{v}_m-\mathrm{L}_m\mathrm{x})$ and
$\widetilde{\delta}=\max\{\delta,\sigma_1,\ldots,\sigma_m\}$. Again,
for the sake of
simplicity, this extension has not been considered in
Problem~\ref{prob:mi}.
\item If the inversion of the matrix $U$ is not difficult or no
variable metric is used and the projection onto
$\mathrm{X}\subset\mathrm{H}$ is computable, we can also use
Theorem~\ref{thm:asymmetric_metric} for solving Problem~\ref{prob:mi}
in the general case $\mathrm{X}\subset\mathrm{H}$.
\end{enumerate}
\end{remark}
\begin{corollary}
\label{c:pd}
In Problem~\ref{prob:mi}, let $\theta\in[-1,1]$, let
$\sigma_0,\ldots,\sigma_m$ be strictly positive real numbers and let
$\Omega$ the
$(m+1)\times(m+1)$ symmetric real matrix given by
\begin{equation}
\label{e:Omega}
(\forall i,j\in\{0,\ldots,m\})\quad \Omega_{ij}=
\begin{cases}
\frac{1}{\sigma_i},\quad&\text{if }i=j;\\
-(\frac{1+\theta}{2})\|\mathrm{L}_i\|,&\text{if }0=j<i;\\
0,&\text{if }0<j<i.
\end{cases}
\end{equation}
Assume that $\Omega$ is positive definite with $\rho>0$ its
smallest eigenvalue and that
\begin{equation}
\label{e:conditioncor}
\left(\delta+\left(\frac{1-\theta}{2}\right)
\sqrt{\sum_{i=1}^m\|\mathrm{L}_i\|^2}\right)^2
<\rho\left(\rho-\frac{1}{2\beta}\right),
\end{equation}
where $\beta=\min\{\mu,\nu_1,\ldots,\nu_m\}$.
Let
$M=(\min\{\sigma_0,\ldots,\sigma_m\})^{-1}+
(\frac{1+\theta}{2})\sqrt{\sum_{i=1}^m\|\mathrm{L}_{i}\|^2}$,
let
$\lambda\in]0,M^{-1}[$,
let $(\mathrm{x}^0,\mathrm{u}_1^0,\ldots,\mathrm{u}_m^0)\in
\mathrm{H\times
G}_1\times\cdots\times
\mathrm{G}_m$, and let $\{\mathrm{x}^k\}_{k\in\ensuremath{\mathbb N}}$
and
$\{\mathrm{u}_i^k\}_{k\in\ensuremath{\mathbb N}, 1\leq i\leq m}$ the sequences
generated by the
following
routine:
\begin{equation}
\label{e:genVU}
(\forall k\in\ensuremath{\mathbb N})\quad
\begin{array}{l}
\left\lfloor
\begin{array}{l}
\mathrm{y}^k
=J_{\sigma_0\mathrm{A}}
\left(\mathrm{x}^k-\sigma_0\left(\mathrm{C}_1\mathrm{x}^k+
\mathrm{C}_2\mathrm{x}^k+\sum_{i=1}^m
\mathrm{L}_i^*\mathrm{u}_i^k\right)\right)\\[2mm]
\text{For every } i=1,\ldots,m\\
\left\lfloor
\mathrm{v}_i^k = J_{
\sigma_i\mathrm{B}_i^{-1}}\left(\mathrm{u}_i^k -
\sigma_i\left(\mathrm{D}_i^{-1}\mathrm{u}_i^k-
\mathrm{L}_i(\mathrm{y}^k+\theta(\mathrm{y}^k
-\mathrm{x}^k))\right)\right)
\right.\\[2mm]
\mathrm{x}^{k+1}=
\mathrm{x}^k+\frac{\lambda}{\sigma_0}\left(\mathrm{y}^k-\mathrm{x}^k+
\sigma_0\left(\mathrm{C}_2\mathrm{x}^k-\mathrm{C}_2\mathrm{y}^k+\sum_{i=1}^m
\mathrm{L}_i^*(\mathrm{u}_i^k-\mathrm{v}_i^k)\right)\right)\\
\text{For every } i=1,\ldots,m\\
\left\lfloor
\mathrm{u}_i^{k+1}=\mathrm{u}_i^k+\frac{\lambda}{\sigma_i}
\left(\mathrm{v}_i^k-\mathrm{u}_i^k-
\sigma_i\theta\mathrm{L}_i(\mathrm{y}^k-\mathrm{x}^k)\right),
\right.\\[2mm]
\end{array}
\right.\\[2mm]
\end{array}
\end{equation}
Then there exists a primal-dual solution
$(\mathrm{x}^*,\mathrm{u}_1^*,\ldots,\mathrm{u}_m^*)\in
\mathrm{H\times
G}_1\times\cdots\times
\mathrm{G}_m$ to
Problem~\ref{prob:mi} such that $\mathrm{x}^k\rightharpoonup
\mathrm{x}^*$ and, for every $i\in\{1,\ldots,m\}$,
$\mathrm{u}_i^k\rightharpoonup \mathrm{u}_i^*$.
\end{corollary}
\begin{proof}
This result is a consequence of Theorem~\ref{p:primdumi} when, for
every
$i\in\{0,\ldots,m\}$, $\mathrm{P}_{ii}=\ensuremath{\operatorname{Id}}\,/\sigma_i$,
$\mathrm{P}_{i0}=-(1+\theta)\mathrm{L}_{i}$,
and, for every $0<j<i$, $\mathrm{P}_{ij}=0$. Indeed,
we have from \eqref{e:condiPii} that
$\varrho_i=1/\sigma_i$, and from \eqref{e:defUpsilon} we deduce that,
for every $x=(\xi_i)_{0\leq i\leq m}\in\mathbb{R}^{m+1}$,
\begin{equation}
\|\Sigma
x\|^2=\left(\frac{1-\theta}{2}\right)^2\left[\left(\sum_{i=0}^m\|\mathrm{L}_i\|\xi_i\right)^2+\xi_0^2
\sum_{i=1}^m\|\mathrm{L}_i\|^2\right]\leq
\left(\frac{1-\theta}{2}\right)^2\left(\sum_{i=0}^m\|\mathrm{L}_i\|^2\right)\|x\|^2,
\end{equation}
from which we obtain
$\|\Sigma\|_2\leq
(\frac{1-\theta}{2})\sqrt{\sum_{i=1}^m\|\mathrm{L}_i\|^2}$. Actually,
we have the equality by choosing $\bar{x}=(\bar{\xi}_i)_{0\leq i\leq
m}$ defined
by
$\bar{\xi}_i=\|\mathrm{L}_i\|/\sqrt{\sum_{j=1}^m\|\mathrm{L}_j\|^2}$
for
every $i\in\{1,\ldots,m\}$ and $\bar{\xi}_0=0$, which satisfies
$\|\bar{x}\|=1$
and $\|\Sigma
\bar{x}\|=(\frac{1-\theta}{2})\sqrt{\sum_{i=1}^m\|\mathrm{L}_i\|^2}$.
Therefore, condition
\eqref{e:metricconditionpd} reduces to \eqref{e:conditioncor}.
On the other hand, from \eqref{e:defUpsilon} we deduce that
$\Omega=\Delta-\Upsilon$ and
$\Upsilon=(\frac{1+\theta}{1-\theta})\Sigma$, which yields
$\|\Upsilon\|_2=(\frac{1+\theta}{2})\sqrt{\sum_{i=1}^m\|\mathrm{L}_i\|^2}$
and $\max_{i=0,\ldots,m}\|\mathrm{P}_{ii}\|
=(\min\{\sigma_0,\ldots,\sigma_m\})^{-1}$. Altogether, since
\eqref{e:genVU} is exactly \eqref{e:algomicomp1} for this choice of
matrices $(\mathrm{P}_{i,j})_{0\leq i,j,\leq m}$, the result is a
consequence of Theorem~\ref{p:primdumi}.
\end{proof}
\begin{remark}
\label{r:pd2}
\begin{enumerate}
\item\label{r:pd21} Note that, the condition $\rho>0$ where $\rho$ is the smallest
eigenvalue of $\Omega$ defined in \eqref{e:Omega}, is
guaranteed if
$\sigma_0(\frac{1+\theta}{2})^2\sum_{i=1}^m\sigma_i\|\mathrm{L}_i\|^2<1$.
Indeed, by repeating the procedure in \cite[(3.20)]{vu2013splitting}
in finite dimension we obtain, for every $x=(\xi_i)_{0\leq i\leq
m}\in\mathbb{R}^{m+1}$,
\begin{align}
\hspace{-.5cm}x\cdot\Omega x&=
\sum_{i=0}^m\frac{\xi_i^2}{\sigma_i}-\sum_{i=1}^m
2\left(\frac{1+\theta}{2}\right)\xi_0\|\mathrm{L}_i\|\xi_i\nonumber\\
&=\sum_{i=0}^m\frac{\xi_i^2}{\sigma_i}-\left(\frac{1+\theta}{2}\right)
\sum_{i=1}^m2
\frac{\sqrt{\sigma_i\|\mathrm{L}_i\|\xi_0}}{(\sigma_0\sum_{j=1}^m
\sigma_j\|\mathrm{L}_j\|^2)^{1/4}}
\frac{(\sigma_0\sum_{j=1}^m
\sigma_j\|\mathrm{L}_j\|^2)^{1/4}\xi_i}{\sqrt{\sigma_i}}\label{e:necess}\\
&\ge\sum_{i=0}^m\frac{\xi_i^2}{\sigma_i}-\left(\frac{1+\theta}{2}\right)
\left(\frac{\xi_0^2}{\sqrt{\sigma_0}}\sqrt{\sum_{j=1}^m\sigma_j\|\mathrm{L}_j\|^2}
+\sqrt{\sigma_0\sum_{j=1}^m\sigma_j\|\mathrm{L}_j\|^2}
\sum_{j=1}^m\frac{\xi_j^2}{\sigma_j}\right)\nonumber\\
&=\left(1-\left(\frac{1+\theta}{2}\right)\sqrt{\sigma_0\sum_{j=1}^m\sigma_j\|\mathrm{L}_j\|^2}\right)
\sum_{i=0}^m\frac{\xi_i^2}{\sigma_i}\nonumber\\
&\ge \rho_v\|x\|^2\label{e:ineqrho}
\end{align}
with
\begin{equation}
\label{e:rhov}
\rho_v=\max\{\sigma_0,\ldots,\sigma_m\}^{-1}\left(1-
\left(\frac{1+\theta}{2}\right)
\sqrt{\sigma_0\sum_{j=1}^m\sigma_j\|\mathrm{L}_j\|^2}\right).
\end{equation}
Note that $\rho_v$ coincides with the constant obtained in
\cite{vu2013splitting} in the case $\theta=1$ and we have $\rho\geq
\rho_v$. Moreover,
$\sigma_0(\frac{1+\theta}{2})^2\sum_{i=1}^m\sigma_i\|\mathrm{L}_i\|^2<1$
is also necessary for obtaining $\rho>0$, since in \eqref{e:necess}
we can choose a particular vector $x$ for
obtaining the equality. Of course, this choice
does not guarantee to also have equality in the last inequality in
\eqref{e:ineqrho} and, hence, $\rho\geq\rho_v$ in general.
\item\label{r:pd22} If we set $\theta=1$ and
$\mathrm{C}_2=0$ and, hence, $\delta=0$,
\eqref{e:conditioncor} reduces to $2\beta\rho>1$
and we obtain from \eqref{e:genVU} a variant of
\cite[Theorem~3.1]{vu2013splitting}
including an extra forward step involving only the operators
$(\mathrm{L}_i)_{1\leq i\leq m}$.
However, our condition is less restrictive, since $\rho\geq\rho_v$,
where $\rho_v$ is defined in \eqref{e:rhov} and it is obtained in
\cite{vu2013splitting} as we have seen in the last remark. Actually,
in the particular case when $m=1$,
$\mathrm{L}_1=\alpha\ensuremath{\operatorname{Id}}\,$, $\sigma_0=\eta^2\sigma_1=:\eta\sigma$ for
some $0<\eta<1$, constants $\rho_v$ and $\rho$ reduce to
$$\rho_v(\eta)=\frac{1-\eta\sigma\alpha}{\sigma}\quad\text{and}\quad
\rho(\eta)=\frac{1}{2\sigma}\left(\frac{\eta^2+1}{\eta^2}-\sqrt{\left(\frac{\eta^2-1}
{\eta^2}\right)^2+4\alpha^2\sigma^2}\right),$$
respectively. By straightforward computations we deduce that
$\rho(\eta)>\rho_v(\eta)$ for every
$0<\eta<(\alpha\sigma)^{-1}$, and hence our constant can strictly
improve the condition $2\beta\rho>1$, needed in both approaches.
Moreover, since Theorem~\ref{p:primdumi} allows for non self-adjoint
linear operators varying among iterations, we can permit variable
stepsizes $\sigma_{0}^k,\ldots,\sigma_{m}^k$ in
Theorem~\ref{p:primdumi}, which could not
be used in \cite{vu2013splitting} because of the variable metric
framework.
\item\label{r:pd23} In the particular case when $\mathrm{C}_1=0$ and
$\mathrm{C}_2=0$ we can take $\beta{\tilde{o}}+\infty$ and, hence, condition
\eqref{e:conditioncor} reduces to
\begin{equation}
\label{e:conditioncorem}
\left(\frac{1-\theta}{2}\right)
\sqrt{\sum_{i=1}^m\|\mathrm{L}_i\|^2}
<\rho,
\end{equation}
which is stronger than the condition in \cite{he2012convergence}
for the case $m=1$, in which it is only needed that $\rho>0$ for
achieving convergence. Indeed, in the case $m=1$,
\eqref{e:conditioncorem} reduces to
$2-2\theta\sigma_0\sigma_1\|\mathrm{L}_1\|^2>
(1-\theta)(\sigma_0+\sigma_1)\|\mathrm{L}_1\|$,
which coincides with the condition in \cite{he2012convergence} in the
case $\theta=1$, but they differ if $\theta\neq 1$
because of the extra forward step coming from the Tseng's splitting
framework. Actually, in the case $\theta=0$ it reduces to
$\sigma_0+\sigma_1<2/\|\mathrm{L}_1\|$ and in the case $\theta=-1$
we obtain the stronger condition
$\max\{\sigma_0,\sigma_1\}<1/\|\mathrm{L}_1\|$.
Anyway, in our context we can use constants
$\sigma_0^k,\ldots,\sigma_m^k$ varying among iterations and we have a
variant of the method in \cite{he2012convergence} and, in the case
when $\theta=1$, of Chambolle-Pock's splitting
\cite{chambolle2011first}.
\item\label{r:pd24} Since $\rho_v$ defined in \eqref{e:rhov} satisfies $\rho_v\leq
\rho$ in the case when $\mathrm{C}_1=\mathrm{C}_2=0$, a sufficient condition for
guaranteeing \eqref{e:conditioncorem} is
$(1-\theta)\sqrt{\sum_{i=1}^m\|\mathrm{L}_i\|^2}/2<\rho_v$, which implied
by the condition
\begin{equation}
\max\{\sigma_0,\ldots,\sigma_m\}\sqrt{\sum_{i=1}^m\|\mathrm{L}_i\|^2}<1.
\end{equation}
\item\label{r:pd25} Consider the case of composite optimization problems, i.e.,
when $\mathrm{A}=\partial \mathrm{f}$, $\mathrm{C}_1=\nabla \mathrm{h}$
for every $i=1,\ldots,m$, $\mathrm{B}_i=\partial
\mathrm{g}_i$ and $\mathrm{D}_i=\partial\mathrm{\ell}_i$, where, for
every
$i=1,\ldots,m$,
$\mathrm{f}\colon\mathrm{H}{\tilde{o}}\ensuremath{\left]-\infty,+\infty\right]}$ and
$\mathrm{g}_i\colon\mathrm{G}_i{\tilde{o}}\ensuremath{\left]-\infty,+\infty\right]}$ are proper lower
semicontinuous and convex functions and
$\mathrm{h}\colon\mathrm{H}{\tilde{o}}\mathbb{R}$ is differentiable, convex, with
$\beta^{-1}$-Lipschitz gradient. In this case, any solution to
Problem~\ref{prob:mi} when $\mathrm{C}_2=0$ is a solution to the
primal-dual optimization problems
\begin{equation}
\min_{\mathrm{x}\in\mathrm{H}}{\mathrm{f}(\mathrm{x})+\mathrm{h}(\mathrm{x})
+\sum_{i=1}^m(\mathrm{g}_i\ensuremath{\mbox{\small$\,\square\,$}}\mathrm{\ell}_i)(\mathrm{L}_i\mathrm{x})}
\end{equation}
and
\begin{equation}
\min_{\mathrm{u}_1\in\mathrm{G}_1,\ldots,\mathrm{u}_m\in\mathrm{G}_m}
{(\mathrm{f}^*\ensuremath{\mbox{\small$\,\square\,$}}\mathrm{h}^*)
\left(-\sum_{i=1}^m\mathrm{L}_i^*\mathrm{u}_i\right)+
\sum_{i=1}^m\mathrm{g}_i^*(\mathrm{u}_i)+\mathrm{\ell}_i^*(\mathrm{u}_i)},
\end{equation}
and the equivalence holds under some qualification condition. In this particular case,
\eqref{e:genVU} reduces to
\begin{equation}
\label{e:genVUopti}
\begin{array}{l}
\left\lfloor
\begin{array}{l}
\mathrm{y}^k
=\mathbf{prox}_{\sigma_0\mathrm{f}}
\left(\mathrm{x}^k-\sigma_0\left(\nabla\mathrm{h}(\mathrm{x}^k)+
\sum_{i=1}^m
\mathrm{L}_i^*\mathrm{u}_i^k\right)\right)\\[2mm]
\text{For every } i=1,\ldots,m\\
\left\lfloor
\mathrm{v}_i^k = \mathbf{prox}_{
\sigma_i\mathrm{g}_i^*}\left(\mathrm{u}_i^k -
\sigma_i\left(\nabla\mathrm{\ell}_i^*(\mathrm{u}_i^k)-
\mathrm{L}_i(\mathrm{y}^k+\theta(\mathrm{y}^k
-\mathrm{x}^k))\right)\right)
\right.\\[2mm]
\mathrm{x}^{k+1}=
\mathrm{x}^k+\frac{\lambda}{\sigma_0}\left(\mathrm{y}^k-\mathrm{x}^k+
\sigma_0\sum_{i=1}^m
\mathrm{L}_i^*(\mathrm{u}_i^k-\mathrm{v}_i^k)\right)\\
\text{For every } i=1,\ldots,m\\
\left\lfloor
\mathrm{u}_i^{k+1}=\mathrm{u}_i^k+\frac{\lambda}{\sigma_i}
\left(\mathrm{v}_i^k-\mathrm{u}_i^k-
\sigma_i\theta\mathrm{L}_i(\mathrm{y}^k-\mathrm{x}^k)\right),
\right.\\[2mm]
\end{array}
\right.\\[2mm]
\end{array}
\end{equation}
which, in the case $m=1$, is very similar to the method proposed in
\cite[Algorithm~3]{patrinos2016asym} (by taking $\mu=(1-\theta)^{-1}$ for
$\theta\in[-1,0]$), with a slightly different choice of the
parameters involved in the last two lines in \eqref{e:genVUopti}. {On the other
hand, in the case when $\ell=0$ and $\theta=1$, it differs from
\cite[Algorithm~5.1]{condat2013primal} in the last two steps, in which linear operators are
involved in our case. } An advantage of our
method, even in the case $m=1$, is that the stepsizes $\sigma_0$ and $\sigma_1$
may vary among iterations.
\end{enumerate}
\end{remark}
\section{Convergence theory}
\label{sec:2}
This section is devoted to study the conditions ensuring the convergence
of {the iterates generated recursively by} $z^{k+1}=T_{\gamma_k}z^k$ for any
starting point
$z^0\in\ensuremath{{\mathcal H}}$,
where, for every $\gamma>0$, $T_{\gamma}$ is defined in \eqref{eq:ouroperator}.
We first prove that $T_{\gamma}$ is quasi-nonexpansive for a suitable
choice of $\gamma$ and satisfies
$\ensuremath{\operatorname{Fix}}(T_{\gamma}) = \ensuremath{\operatorname{zer}}(A+B_1+B_2)\cap X$. Using these results we prove the weak
convergence of iterates $\{z^k\}_{k\in\ensuremath{\mathbb N}}$ to a solution to Problem~\ref{prob:main}.
\begin{proposition}[Properties of
$T_{\gamma}$]\label{prop:Tproperties}
\label{p:prop1}Let $\gamma>0$, assume that hypotheses of Problem~\ref{prob:main}
hold, {and set $S_{\gamma}:=(\ensuremath{\operatorname{Id}}\, - \gamma B_2) \circ J_{\gamma A} \circ (\ensuremath{\operatorname{Id}}\,-
\gamma (B_1+B_2)) + \gamma B_2$.}
Then,
\begin{enumerate}
\item \label{prop:Tproperties:item:fixed-points}
{We have $\ensuremath{\operatorname{zer}}(A+B_1+B_2)\subset\ensuremath{\operatorname{Fix}} S_{\gamma}$ and
$\ensuremath{\operatorname{zer}}(A+B_1+B_2)\cap X\subset\ensuremath{\operatorname{Fix}} T_{\gamma}$. Moreover, if $B_2$ is
$L$-Lipschitz in $\ensuremath{\operatorname{dom}} B_2$ for some $L>0$ and $\gamma<L^{-1}$ we have
$
\ensuremath{\operatorname{Fix}}(S_{\gamma}) = \ensuremath{\operatorname{zer}}(A+B_1+B_2)
$
and
$
\ensuremath{\operatorname{Fix}}(T_{\gamma}) = \ensuremath{\operatorname{zer}}(A+B_1+B_2) \cap X.
$
\item \label{prop:Tproperties:item:quasi0}
For all $z^\ast \in
\ensuremath{\operatorname{Fix}}(T_{\gamma})$ and $z\in\ensuremath{\operatorname{dom}} B_2$, by denoting $x:=J_{\gamma A} (z - \gamma
(B_1+B_2)z
)$ we have, for every $\varepsilon>0$,
\begin{align*}
\|T_\gamma z - z^\ast\|^2 &\leq\|z - z^\ast\|^2 - (1-\varepsilon) \|z - x\|^2 + \gamma^2
\|B_2z -
B_2x\|^2 \\
&\hspace{.35cm} -\! \frac{\gamma}{\varepsilon}\left(2\beta\varepsilon -\!
{\gamma}\right)\!\|B_1z - B_1 z^\ast\|^2
\!-\!\varepsilon\left\|z-x-\frac{\gamma}{\varepsilon}(B_1z-B_1z^*)\right\|^2.
\numberthis\label{eq:fejer0}
\end{align*}}
\item\label{prop:Tproperties:item:quasi} Suppose that
$B_2$ is $L$-Lipschitz {in $\ensuremath{\operatorname{dom}} A\cup X$ for some $L>0$. For all $z^\ast \in
\ensuremath{\operatorname{Fix}}(T_{\gamma})$ and $z\in\ensuremath{\operatorname{dom}} B_2$, by denoting $x:=J_{\gamma A} (z - \gamma
(B_1+B_2)z
)$ we have
\begin{align*}
\|T_\gamma z - z^\ast\|^2 &\leq \|z - z^\ast\|^2 - L^{2}(\chi^2-
\gamma^2)\|z - x\|^2-\frac{2\beta\gamma}{\chi}\left(\chi
- \gamma\right)\|B_1z - B_1 z^\ast\|^2\nonumber\\
&-\frac{\chi}{2\beta}\left\|z-x-\frac{2\beta\gamma}{\chi}(B_1z-B_1z^*)\right\|^2,
\numberthis\label{eq:fejer}
\end{align*}
}
where
\begin{equation}
\label{e:chi}
\chi:=\frac{4\beta}{1+\sqrt{1+16\beta^2 L^2}}\leq
\min\{2\beta,L^{-1}\}.
\end{equation}
\end{enumerate}
\end{proposition}
\begin{proof}
Part~\ref{prop:Tproperties:item:fixed-points}: Let $z^*\in\ensuremath{{\mathcal H}}$. We have
{
\begin{align}
\label{e:auxequiv}
z^*\in \ensuremath{\operatorname{zer}}(A+B_1+B_2)\quad&\Leftrightarrow\quad 0\in Az^*+B_1z^*+B_2z^*\nonumber\\
&\Leftrightarrow\quad -\gamma (B_1z^*+B_2z^*)\in \gamma Az^*\nonumber\\
&\Leftrightarrow\quad z^*=J_{\gamma A}\left(z^*-\gamma
(B_1z^*+B_2z^*)\right).
\end{align}
}
Then, since $B_2$ is single-valued in $\ensuremath{\operatorname{dom}} A$, if {$z^*\in \ensuremath{\operatorname{zer}}(A+B_1+B_2)$}
we have $B_2z^*=B_2J_{\gamma A}(z^*-\gamma
(B_1z^*+B_2z^*))$ and, hence, {$S_{\gamma}z^*=z^*$ }which yields
{$\ensuremath{\operatorname{zer}}(A+B_1+B_2)\subset \ensuremath{\operatorname{Fix}} S_{\gamma}$. Hence, if $z^\ast\in
\ensuremath{\operatorname{zer}}(A+B_1+B_2)\cap X$ then $z^\ast\in\ensuremath{\operatorname{Fix}} P_X$ and $z^\ast\in\ensuremath{\operatorname{Fix}} S_{\gamma}$,
which yields $z^\ast\in \ensuremath{\operatorname{Fix}} P_X\circ S_{\gamma}=\ensuremath{\operatorname{Fix}} T_{\gamma}$.} Conversely,
if { $B_2$ is $L$-Lipschitz in $\ensuremath{\operatorname{dom}} B_2$ and $z^*\in \ensuremath{\operatorname{Fix}} S_{\gamma}$ we
have }
\begin{equation*}
z^*-J_{\gamma A}(z^*-\gamma(B_1+B_2)z^*)=\gamma\left(B_2z^*-B_2J_{\gamma
A}(z^*-\gamma(B_1+B_2)z^*) \right),
\end{equation*}
which, from the Lipschitz continuity of $B_2$ yields
\begin{align*}
\|z^*-J_{\gamma A}(z^*-\gamma(B_1+B_2)z^*)\|&=\gamma\|B_2z^*-B_2J_{\gamma
A}(z^*-\gamma(B_1+B_2)z^*)\| \\
&\leq \gamma L\|z^*-J_{\gamma
A}(z^*-\gamma(B_1+B_2)z^*)\|.
\end{align*}
Therefore, if $\gamma<L^{-1}$ we deduce $z^*=J_{\gamma
A}(z^*-\gamma(B_1+B_2)z^*)$ and from \eqref{e:auxequiv}, we deduce
{$\ensuremath{\operatorname{zer}}(A+B_1+B_2)=\ensuremath{\operatorname{Fix}} S_{\gamma}$. Since $T_{\gamma}=P_XS_{\gamma}$ and
$P_X$ is
strictly quasi-nonexpansive, the result follows from
\cite[Proposition~4.49]{bauschke2017convex}.}
Part~\ref{prop:Tproperties:item:quasi0}:
Let $z^*\in \ensuremath{\operatorname{Fix}} T_{\gamma}$, {$z\in\ensuremath{\operatorname{dom}} B_2$} and define $B := B_1 + B_2$, $y:=
z - \gamma Bz$, $x:=
J_{\gamma A} y$, and $z^+ = T_{\gamma}z$. Note that $(x, y-x) \in
\gra(\gamma A)$ and, from Part~\ref{prop:Tproperties:item:fixed-points}, $(z^\ast,
-\gamma Bz^\ast)
\in \gra(\gamma A)$. Hence, by the monotonicity of
$A$ and $B_2$, we have $\dotp{x - z^\ast, x-y -\gamma
Bz^\ast} \leq 0$ and $\dotp{x - z^\ast, \gamma B_2 z^\ast - \gamma B_2x}\leq 0$. Thus,
\begin{align*}
\dotp{x - z^\ast, x-y - \gamma B_2x} &= \dotp{x - z^\ast, \gamma B_1z^\ast} + \dotp{x -
z^\ast, x-y -\gamma Bz^\ast} \\
&\hspace{2.8cm}+\dotp{x - z^\ast, \gamma B_2z^\ast - \gamma B_2x} \\
&\leq \dotp{x - z^\ast, \gamma B_1z^\ast}.
\end{align*}
Therefore, we have
\begin{align}
\label{e:desig1sec2}
2\gamma \dotp{x - z^\ast, B_2z - B_2x}
&= 2\dotp{ x - z^\ast,
\gamma B_2z + y - x} + 2 \dotp{ x- z^\ast, x - y - \gamma B_2x} \nonumber\\
&\leq 2 \dotp{ x - z^\ast, \gamma Bz + y - x} + 2\dotp{x - z^\ast, \gamma B_1z^\ast -
\gamma B_1z} \nonumber\\
&= 2 \dotp{ x - z^\ast, z - x} + 2\dotp{x - z^\ast, \gamma B_1z^\ast - \gamma B_1z}
\nonumber\\
&= \|z - z^\ast\|^2 \!- \!\|x - z^\ast\|^2 \!-\! \|z - x\|^2 \!+\! 2\dotp{x - z^\ast, \gamma
B_1z^\ast - \gamma B_1z} .
\end{align}
In addition, by cocoercivity of $B_1$, for all $\varepsilon > 0$, we have
\begin{align}
\label{e:desig2sec2}
2\dotp{x - z^\ast, \gamma B_1z^\ast - \gamma B_1z}
&= 2\dotp{z - z^\ast, \gamma B_1z^\ast - \gamma B_1z} + 2\dotp{x - z, \gamma B_1z^\ast -
\gamma B_1z}\nonumber \\
&\leq - 2\gamma \beta\|B_1z - B_1 z^\ast\|^2 + 2\dotp{x - z, \gamma B_1z^\ast -
\gamma B_1z}\nonumber \\
&= - 2\gamma \beta\|B_1z - B_1 z^\ast\|^2 + \varepsilon\|z-x\|^2 +
\frac{\gamma^2}{\varepsilon}\|B_1z - B_1
z^\ast\|^2\nonumber \\
&\hspace{4cm}-\varepsilon\left\|z-x-\frac{\gamma}{\varepsilon}(B_1z-B_1z^*)\right\|^2
\nonumber \\
&= \varepsilon\|z-x\|^2 - \gamma\left(2\beta -
\frac{\gamma}{\varepsilon}\right)\|B_1z - B_1
z^\ast\|^2\nonumber \\
&\hspace{4cm}-\varepsilon\left\|z-x-\frac{\gamma}{\varepsilon}(B_1z-B_1z^*)\right\|^2.
\end{align}
Hence, combining \eqref{e:desig1sec2} and \eqref{e:desig2sec2}, it follows from $z^*\in
X$, the nonexpansivity of $P_X$, and the Lipschitz property of $B_2$ in $\ensuremath{\operatorname{dom}} B_2\supset
X\cup\ensuremath{\operatorname{dom}} A$ that
{
\begin{align}
\label{e:auxprt3}
\|z^+ - z^\ast\|^2
&\leq \|x - z^\ast + \gamma B_2z - \gamma B_2x\|^2 \nonumber\\
&= \|x - z^\ast\|^2 + 2\gamma\dotp{x - z^\ast, B_2z - B_2x} + \gamma^2\| B_2z - B_2x\|^2
\nonumber\\
&\leq \|z - z^\ast\|^2 - \|z - x\|^2 + \gamma^2 \|B_2z - B_2x\|^2 \nonumber\\
&\hspace{.35cm}+ \varepsilon\|z\!- x\|^2\! -\! \gamma\left(2\beta -\!
\frac{\gamma}{\varepsilon}\right)\!\|B_1z - B_1 z^\ast\|^2
\!-\!\varepsilon\left\|z-x-\frac{\gamma}{\varepsilon}(B_1z-B_1z^*)\right\|^2,
\end{align}
and the result follows.
Part 3: It follows from \eqref{e:auxprt3} and the Lipschitz property on $B_2$ that
\begin{align*}
&\|z^+ - z^\ast\|^2\leq \|z - z^\ast\|^2 - L^2\left(\frac{1-
\varepsilon}{L^2}-\gamma^2\right)\|z - x\|^2 -
\frac{\gamma}{\varepsilon}\left(2\beta\varepsilon -
\gamma\right)\|B_1z - B_1 z^\ast\|^2 \\
&\hspace{.35cm}-\varepsilon\left\|z-x-\frac{\gamma}{\varepsilon}(B_1z-B_1z^*)\right\|^2.
\end{align*}}
In order to obtain the largest interval for $\gamma$ ensuring that
{the second and third terms on the right of} the above equation are
negative, we choose
the value $\varepsilon$ so that $\sqrt{1-\varepsilon}/L =
2\beta\varepsilon$, which yields
$\varepsilon=(-1+\sqrt{1+16\beta^2L^2})(8\beta^2L^2)^{-1}$. For this
choice of $\varepsilon$ we obtain $\chi=\sqrt{1-\varepsilon}/L =
2\beta\varepsilon$.
\end{proof}
{ In the case when $B_2$ is merely continuous, we need the following result,
which gives additional information to
\cite[Lemma~3.3]{tseng2000modified} and allows us to guarantee the convergence of
the algorithm under weaker assumptions than \cite[Theorem~3.4]{tseng2000modified}.
\begin{lemma}
\label{lem:Ts}
In the context of Problem~\ref{prob:main}, define, for every $z\in\ensuremath{\operatorname{dom}} B_2$ and $\gamma
>0$,
\begin{equation}
\label{e:defxphi}
x_{z}\colon \gamma\mapsto J_{\gamma
A}(z-\gamma(B_1+B_2)z)\quad \text{and \quad }\varphi_z\colon \gamma\mapsto
\frac{\|z-x_z(\gamma)\|}{\gamma}.
\end{equation}
Then, the following hold:
\begin{enumerate}
\item \label{lem:Tsi}$\varphi_z$ is nonincreasing and
\begin{equation*}
(\forall z\in\ensuremath{\operatorname{dom}} A)\quad \lim_{\gamma\downarrow
0^+}\varphi_z(\gamma)=
\|(A+B_1+B_2)^0(z)\|:=\inf_{w\in(A+B_1+B_2)z}\|w\|.
\end{equation*}
\item\label{lem:Tsii} For every $\theta\in]0,1[$ and $z\in\ensuremath{\operatorname{dom}} B_2$, there exists
$\gamma(z)>0$ such
that, for every $\gamma\in]0,\gamma(z)]$,
\begin{equation}
\label{e:armijocond}
\gamma\|B_2z-B_2x_z(\gamma))\|
\leq\theta\|z-x_z(\gamma)\|.
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
Part \ref{lem:Tsi}: Denote $B:=B_1+B_2$.
If $z\in\ensuremath{\operatorname{zer}}(A+B)$ then it follows from \eqref{e:auxequiv} that $\varphi_z\equiv 0$ and there
is nothing to prove. Hence,
assume $z\in\ensuremath{\operatorname{dom}} B_2\setminus\ensuremath{\operatorname{zer}}(A+B)$ which yields $\varphi_z(\gamma)>0$ for every
$\gamma>0$. From the definition of
$J_{\gamma A}$, we have $(z-x_z(\gamma))/\gamma-Bz\in A(x_z(\gamma))$ for every
$\gamma>0$ and,
from the monotonicity of $A$, we deduce that, for every strictly positive constants
$\gamma_1$ and $\gamma_2$ we have
\begin{align}
0&\le \Scal{\frac{z-x_z(\gamma_1)}{\gamma_1}-\frac{z-x_z(\gamma_2)}{\gamma_2}}
{x_z(\gamma_1)-x_z(\gamma_2)}\nonumber\\
&=-\frac{\|z-x_z(\gamma_1)\|^2}{\gamma_1}+\left(\frac{1}{\gamma_1}+\frac{1}{\gamma_2}\right)
\scal{z-x_z(\gamma_1)}{z-x_z(\gamma_2)}-\frac{\|z-x_z(\gamma_2)\|^2}{\gamma_2}.
\end{align}
Therefore
\begin{align}
\label{e:DOng}
\gamma_1\varphi_z(\gamma_1)^2+\gamma_2\varphi_z(\gamma_2)^2&\le
(\gamma_1+\gamma_2)
\Scal{\frac{z-x_z(\gamma_1)}{\gamma_1}}{\frac{z-x_z(\gamma_2)}{\gamma_2}}\nonumber\\
&\le
\frac{\gamma_1+\gamma_2}{2}(\varphi_z(\gamma_1)^2+\varphi_z(\gamma_2)^2),
\end{align}
which is equivalent to
$
(\gamma_1-\gamma_2)(\varphi_z(\gamma_1)^2-\varphi_z(\gamma_2)^2)\le 0,
$
and the monotonicity of $\varphi_z$ is obtained. The limit follows from
\cite[Lemma~3.3\&Eq~(3.5)]{tseng2000modified}.
Part \ref{lem:Tsii}: As before, if $z\in\ensuremath{\operatorname{zer}}(A+B)$ we have $z=x_z(\gamma)$ for every
$\gamma>0$ and,
hence, there is nothing to prove. From \ref{lem:Tsi} we have that, for every
$z\in\ensuremath{\operatorname{dom}}(A)\setminus
\ensuremath{\operatorname{zer}}(A+B)$,
\begin{equation*}
0<\|z-x_z(1)\|\le \lim_{\gamma\downarrow
0^+}\frac{\|z-x_z(\gamma)\|}{\gamma}= \|(A+B_1+B_2)^0(z)\|.
\end{equation*}
Therefore, $\lim_{\gamma\downarrow
0^+}x_z(\gamma)=z$ and from continuity of $B_2$, $\lim_{\gamma\downarrow
0^+}\|B_2z-B_2x_z(\gamma)\|=0$. This ensures the existence of $\gamma(z)>0$
such that, for every $\gamma\in]0,\gamma(z)]$, \eqref{e:armijocond} holds.
\end{proof}
\begin{remark}
Note that the previous lemma differs from \cite[Lemma~3.3]{tseng2000modified}
because we provide the additional information $\varphi_z$ nonincreasing. This
property is used in \cite{Nghia},
proved in \cite{Dong}, and will be crucial
for obtaining the convergence of the algorithm with line search to a solution to
Problem~\ref{prob:main} under weaker assumptions. We keep our proof for
the sake of completeness and because the inequality \eqref{e:DOng} is slightly stronger
than that obtained in \cite{Dong}.
\end{remark}
}
\begin{theorem}[Forward-backward-half forward algorithm]
\label{t:1}
{Under the assumptions of Problem~\ref{prob:main},
let $z^0 \in \ensuremath{\operatorname{dom}} A\cup X$, and consider the sequence $\{z_k\}_{k\in\ensuremath{\mathbb N}}$ recursively
defined
by $z^{k+1} := T_{\gamma_k} z^k$ or, equivalently,
\begin{equation}
\label{e:algomain1}
(\forall k\in\ensuremath{\mathbb N})\quad
\begin{array}{l}
\left\lfloor
\begin{array}{l}
x^k = J_{\gamma_k A}(z^k - \gamma_k (B_1 + B_2)z^k) \\[2mm]
z^{k+1} = P_X \big(x^k + \gamma_k B_2z^k - \gamma_k
B_2x^k\big),
\end{array}
\right.
\end{array}
\end{equation}
where $\{\gamma_k\}_{k \in \ensuremath{\mathbb N}}$ is a sequence of
stepsizes satisfying one of the following conditions:
\begin{enumerate}
\item Suppose that $B_2$ is $L$-Lipschitz in $\ensuremath{\operatorname{dom}} A\cup X$. Then, for every $k\in\ensuremath{\mathbb N}$,
$\gamma_k\in[\eta,\chi-\eta]$, where $\eta\in\left]0,\chi/2\right[$ and
$\chi$ is defined in~\eqref{e:chi}.
\item Suppose $X\subset\ensuremath{\operatorname{dom}} A$ and let $\varepsilon\in\left]0,1\right[$,
$\sigma\in]0,1[$,
and $\theta\in\left]0,\sqrt{1-\varepsilon}\right[$. Then, for every $k\in\ensuremath{\mathbb N}$, $\gamma_k$ is
the largest
$\gamma\in
\{2\beta\varepsilon\sigma,2\beta\varepsilon\sigma^2,\cdots\}$ satisfying
\eqref{e:armijocond} with $z=z^k$, and at least one of the following additional conditions
holds:
\begin{enumerate}
\item $\liminf_{k{\tilde{o}}\infty}\gamma_k=\delta>0$.
\item $B_2$ is uniformly continuous in any weakly compact subset of
$X$.
\end{enumerate}
\end{enumerate}
Then, $\{z_k\}_{k\in\ensuremath{\mathbb N}}$ converges weakly to a solution
to Problem~\ref{prob:main}.
}
\end{theorem}
\begin{proof}
{In the case when $B_2$ is $L$-Lipschitz in $\ensuremath{\operatorname{dom}} A\cup X$, it} follows from
Proposition~\ref{prop:Tproperties}(\ref{prop:Tproperties:item:quasi}) that the sequence
$\{z^{k}\}_{k \in
\ensuremath{\mathbb N}}$ is Fej{\'e}r monotone with
respect to $\ensuremath{\operatorname{zer}}(A+B_1+B_2) \cap X$. Thus, to show that $\{z^k\}_{k \in \ensuremath{\mathbb N}}$
converges weakly to a solution to Problem~\ref{prob:main}, we only need to prove that all
of
its weak subsequential limits lie in $\ensuremath{\operatorname{zer}}(A+B_1+B_2) \cap X$ {\cite[Theorem
5.33]{bauschke2017convex}.} Indeed,
it follows from Proposition~\ref{prop:Tproperties} and our hypotheses on the stepsizes
that, for every $z^*\in\ensuremath{\operatorname{Fix}}
T_{\gamma}$,
\begin{align}
\label{e:auxfejer}
\|z^{k}-z^*\|^2-\|z^{k+1}-z^*\|^2&\ge
L^2\eta^2\|z^k-x^k\|^2+\frac{2\beta\eta^2}{\chi}\|B_1z^k-B_1z^*\|^2\nonumber\\
&\hspace{20pt}+\frac{\chi}{2\beta}
\left\|z^k-x^k-\frac{2\beta\gamma_k}{\chi}(B_1z^k-B_1z^*)\right\|^2.
\end{align}
Therefore, we deduce from \cite[Lemma~3.1]{combettes2001quasi} that
\begin{equation}
\label{e:tozero}
z^k-x^k{\tilde{o}} 0
\end{equation}
when $L>0$ and
$0<\beta<\infty$\footnote{The case $B_1=0$ ($\beta=+\infty$)
has been studied by Tseng in \cite{tseng2000modified}. In the case when $B_2=0$
we
can also obtain convergence from
Proposition~\ref{prop:Tproperties}, since $L=0$ implies $\chi=2\beta$ and even
since the first term in the right hand side of \eqref{e:auxfejer} vanishes, the other two
terms yield $z^k-x^k\to0$.}. Now let $z\in\ensuremath{{\mathcal H}}$ be the weak limit point of some
subsequence of
$\{z^k\}_{k\in\ensuremath{\mathbb N}}$. Since $z^k\in X$ for every $k\ge 1$ and $X$ is weakly
sequentially closed {\cite[Theorem 3.34]{bauschke2017convex}}
we deduce $z\in X$. Moreover, {by denoting $B:=B_1+B_2$,} it follows from $x^k
= J_{\gamma_k
A}(z^k - \gamma_k Bz^k)$ that {$u^k := \gamma_k^{-1}(z^k - x^k) - Bz^k +
Bx^k\in
(A+B)x^k$. Then, \eqref{e:tozero}, $\gamma_k\ge \eta>0$ and the Lipschitz continuity of
$B$ yield $u^k
\rightarrow 0$.
Now, since $A+B_2$ is maximally monotone and $B_1$ is cocoercive with full
domain,
$A+B$ is maximally monotone and its graph is}
closed in the weak-strong topology in $\ensuremath{{\mathcal H}}\times\ensuremath{{\mathcal H}}$, which yields
{$0\in Az+Bz$} and the result follows.
{
In the second case, we deduce from
Proposition~\ref{prop:Tproperties}
(\ref{prop:Tproperties:item:fixed-points}\&\ref{prop:Tproperties:item:quasi0}) and
$\gamma_k\leq 2\beta\varepsilon\sigma$ that,
for every $z^*\in\ensuremath{\operatorname{zer}}(A+B_1+B_2)\cap X$ we have
\begin{align*}
\|z^{k} - z^\ast\|^2- \|z^{k+1} - z^\ast\|^2&\geq (1-\varepsilon) \|z^k - x^k\|^2
+\! \frac{\gamma_k}{\varepsilon}\left(2\beta\varepsilon -
{\gamma_k}\right)\|B_1z^k - B_1 z^\ast\|^2 \\
&\hspace{.5cm}+\varepsilon\left\|z^k-x^k-\frac{\gamma_k}{\varepsilon}(B_1z^k-B_1z^*)\right\|^2
-\gamma_k^2 \|B_2z^k - B_2x^k\|^2\\
&\geq (1-\varepsilon-\theta^2) \|z^k - x^k\|^2
+\! 2\beta\varepsilon(1-\sigma)\gamma_k\|B_1z^k - B_1 z^\ast\|^2,
\numberthis\label{eq:fejer1}
\end{align*}
where in the last inequality we use the conditions on $\{\gamma_k\}_{k\in\ensuremath{\mathbb N}}$,
whose existence is guaranteed by Lemma~\ref{lem:Ts}\eqref{lem:Tsii} because $z^k\in
X\subset \ensuremath{\operatorname{dom}} A$.
Then, we deduce from
\cite[Lemma~3.1]{combettes2001quasi} that
$z^k-x^k{\tilde{o}} 0$.
Now let $z$ be a weak limit point of a subsequence $\{z^{k}\}_{k\in K}$, with $K\subset
\ensuremath{\mathbb N}$.
If $\liminf_{k{\tilde{o}}\infty}\gamma_k=\delta>0$, from \eqref{e:armijocond} and $z^k-x^k{\tilde{o}} 0$
we have $B_2x^k-B_2z^k{\tilde{o}} 0$
and the proof is analogous to the previous case. Finally, for the last case, suppose that
there exists a subsequence of $\{\gamma_{k}\}_{k\in K}$ (called similarly) satisfying
$\lim_{k{\tilde{o}}\infty,k\in K}\gamma_k=0$. Our choice of $\gamma_k$ yields, for every $k\in K$,
\begin{equation}\label{e:palotro}
\theta\|z^k-J_{\tilde{\gamma}_k A}(z^k-\tilde{\gamma}_k B
z^k)\|/\tilde{\gamma}_k<\|B_2z^k-B_2J_{\tilde{\gamma}_k A}(z^k-\tilde{\gamma}_k B
z^k)\|,
\end{equation}
where $\tilde{\gamma}_k=\gamma_k/\sigma>\gamma_k$ and, from
Lemma~\ref{lem:Ts}\eqref{lem:Tsi} we
have
\begin{align}
\label{e:auxLS}
\sigma\|z^k-J_{\tilde{\gamma}_k A}(z^k-\tilde{\gamma}_k B
z^k)\|/{\gamma}_k&= \|z^k-J_{\tilde{\gamma}_k A}(z^k-\tilde{\gamma}_k B
z^k)\|/\tilde{\gamma}_k\nonumber\\
&\le \|z^k-J_{{\gamma}_k A}(z^k-{\gamma}_k B
z^k)\|/{\gamma}_k,
\end{align}
which, from $z^k-x^k{\tilde{o}} 0$, yields
$$\|z^k-J_{\tilde{\gamma}_k A}(z^k-\tilde{\gamma}_k B
z^k)\|\le \|z^k-x^k\|/\sigma{\tilde{o}} 0$$
as $k{\tilde{o}}\infty, k\in K$. Therefore, since $z_k\ensuremath{\:\rightharpoonup\:} z$, the sequence $\{\tilde{x}^k\}_{k\in
K}$ defined by
$$(\forall k\in K)\quad \tilde{x}^k:=J_{\tilde{\gamma}_k A}(z^k-\tilde{\gamma}_k B
z^k)$$ satisfies $\tilde{x}^k\ensuremath{\:\rightharpoonup\:} z$ as $k{\tilde{o}}+\infty,k\in K$ and
\begin{equation}
\tilde{w}^k:=\frac{z^k-\tilde{x}^k}{\tilde{\gamma}_k}+ B \tilde{x}^k- B
z^k\in (A+B_1+B_2)\tilde{x}^k.
\end{equation}
Hence, since $\{z\}\cup\bigcup_{k\in\ensuremath{\mathbb N}}[\tilde{x}^k,z^k]$ is a weakly compact
subset
of $X$ \cite[Lemma~3.2]{Salzo}, it follows from the uniform continuity of $B_2$ that the
right hand side of
\eqref{e:palotro} goes to $0$ and, hence, ${(z^k-\tilde{x}^k)}/{\tilde{\gamma}_k}{\tilde{o}} 0$ as
$k{\tilde{o}}\infty,k\in K$. Moreover, since $B_1$ is uniformly continuous, $B=B_1+B_2$
is also locally uniformly continuous and $B \tilde{x}^k- B
z^k{\tilde{o}} 0$, which yields $\tilde{w}^k{\tilde{o}} 0$ as $k{\tilde{o}}+\infty,k\in K$. The result is obtained as in
the first case since the graph of $A+B$ is weakly-strongly closed in the product topology.
}
\end{proof}
\begin{remark}
\begin{enumerate}
\item In \cite[Theorem~3.4]{tseng2000modified} the local boundedness of $z\mapsto
\min_{w\in(A+B)z}\|w\|$ is needed to guarantee the convergence of the method with
line
search. We drop this assumption by using the monotonicity of $\varphi_z$ in
Lemma~\ref{lem:Ts}\eqref{lem:Tsi}, which leads us to the inequality \eqref{e:auxLS}.
\item Since continuity on compact sets yields uniform continuity, in the finite
dimensional setting, the assumption on $B_2$ reduces to the mere continuity on $X$
(see \cite[Remark~3.1(v)]{Salzo}). In this case, we do not need to assume further
assumptions than those given in Problem~\ref{prob:main}.
\end{enumerate}
\end{remark}
\begin{remark}
The maximal monotonicity assumption on $A+B_2$ is satisfied, for
instance, if $\mathrm{cone}(\ensuremath{\operatorname{dom}} A-\ensuremath{\operatorname{dom}}
B_2)=\overline{\mathrm{span}}(\ensuremath{\operatorname{dom}} A-\ensuremath{\operatorname{dom}} B_2)$, where, for any set
$D\subset\ensuremath{{\mathcal H}}$,
$\mathrm{cone}(D)=\menge{\lambda d}{\lambda\in\mathbb{R}_+, d\in D}$
and $\overline{\mathrm{span}}(D)$ is the smallest closed linear
subspace of $\ensuremath{{\mathcal H}}$ containing $D$ \cite[Theorem~3.11.11]{Zalinescu}.
\end{remark}
\begin{remark}
\label{rem:chi}
{In the case when $B_2$ is $L$-Lipschitz in $\ensuremath{\operatorname{dom}} A\cup X$, the} stepsize
upper
bound $\chi=
\chi(\beta,L)$ defined in \eqref{e:chi}
depends on the
cocoercivity parameter $\beta$ of $B_1$ and the Lipschitz parameter
$L$ of $B_2$. In order to fully recover Tseng's splitting algorithm or the
forward-backward algorithm in the cases when $B_1$ or $B_2$ are zero,
respectively, we study the asymptotic behaviour of $\chi(\beta,L)$
when $L{\tilde{o}} 0$ and $\beta{\tilde{o}}+\infty$. It is easy to verify that
\begin{align*}
\lim_{L{\tilde{o}} 0}\chi(\beta,L)=2\beta\quad\text{and}\quad \lim_{\beta{\tilde{o}}
+\infty}\chi(\beta,L)=\frac{1}{L},
\end{align*}
which are exactly the bounds on the stepsizes of forward-backward and
Tseng's splittings.
{ On the other hand, when $B_2$ is continuous, if we choose
$\varepsilon\in\left]0,1\right[$ close to 1,
$\{\gamma_k\}_{k\in\ensuremath{\mathbb N}}$ could be larger since the line search starts from
$2\beta\varepsilon\sigma$. However, $\theta<\sqrt{1-\varepsilon}$
should be close to 0 in this case, and condition \eqref{e:armijocond} is more restrictive
and satisfied only for
small values of $\gamma_k$. Conversely, for small values of $\varepsilon$ we restrict
the sequence $\{\gamma_k\}_{k\in\ensuremath{\mathbb N}}$ in a small interval but \eqref{e:armijocond} is more
easily satisfied. The optimal choice of $\varepsilon$ in order to obtain an optimal sequence
$\{\gamma_k\}_{k\in\ensuremath{\mathbb N}}$ depends on the properties of the operators involved.
Note that, in the particular case when $B_2\equiv 0$, \eqref{e:armijocond} is satisfied for
$\theta=0$ and we can choose $\varepsilon=1$, recovering forward-backward splitting.
On the other hand, when $B_1\equiv0$, we can take $\varepsilon=0$ and
$\theta\in\left]0,1\right[$ recovering the Tseng's method with backtracking proposed in
\cite{tseng2000modified}.}
\end{remark}
\section{Introduction}
This paper is devoted to the numerical resolution of following problem.
\begin{problem}
\label{prob:main}
Let $X$ be a nonempty closed convex subset of a real Hilbert space ${\mathcal{H}}$, let $A
: {\mathcal{H}} \rightarrow 2^{\mathcal{H}}$ {and $B_2 : {\mathcal{H}}
\rightarrow 2^{{\mathcal{H}}}$ be maximally monotone operators, with $B_2$ single valued in $\ensuremath{\operatorname{dom}}
B_2\supset\ensuremath{\operatorname{dom}} A\cup X$, and let $B_1 : {\mathcal{H}}
\rightarrow {\mathcal{H}}$ be $\beta$-cocoercive\footnote{An operator $C : {\mathcal{H}} \rightarrow
{\mathcal{H}}$ is $\beta$-cocoercive for some $\beta > 0$ provided that $\dotp{Cx - Cy, x-y}
\geq \beta\|Cx - Cy\|^2$.}, for
some $\beta>0$. Moreover
assume that $B_2$ is continuous on $\ensuremath{\operatorname{dom}} A\cup
X$
and that $A+B_2$ is
maximally monotone. }The
problem is to
\begin{equation}
\label{e:main}
\text{find }\quad x\in X\quad \text{such that }\quad 0\in Ax+B_1x+B_2x,
\end{equation}
under the assumption that the set of solutions to \eqref{e:main} is nonempty.
\end{problem}
The wide variety of applications of Problem~\ref{prob:main} involving optimization
problems, variational inequalities,
partial differential equations, image processing, saddle point problems,
game theory, among others can be explored in
{\cite{bauschke2017convex,combetteswajs2005} }and the
references therein.
As an important application, consider the case of composite optimization problems of
the form
\begin{align}\label{eq:primal_before_PD}
\Min_{\mathrm{x} \in \mathrm{H}}\, \mathrm{f}(\mathrm{x}) +
\mathrm{g}(\mathrm{L}\mathrm{x}) + \mathrm{h}(\mathrm{x}),
\end{align}
where $\mathrm{H}$ and $\mathrm{G}$ are real Hilbert spaces, $\mathrm{L} :
\mathrm{H} \rightarrow \mathrm{G}$ is linear and bounded,
$\mathrm{f} :\mathrm{H} \rightarrow (-\infty, \infty]$ and $\mathrm{g} : \mathrm{G}
\rightarrow (-\infty, \infty]$ are
lower semicontinuous, convex, and proper, and $\mathrm{h} : \mathrm{H} \rightarrow
\mathbb{R}$ is
convex
differentiable with $\beta^{-1}$-Lipschitz gradient. Since $g$ may be non smooth,
primal algorithms in this context
need to evaluate $\mathbf{prox}_{\mathrm{g} \circ L}$ or invert $L$ which can be costly
numerically. In order to overcome this difficulty, fully split primal-dual algorithms are
proposed, e.g., { in
\cite{briceno2011monotone+,condat2013primal,vu2013splitting}}, in
which only
$\mathbf{prox}_{\mathrm{g}}$, $L$, and $L^\ast$ are computed. These algorithms follow
from the first order optimality conditions of \eqref{eq:primal_before_PD}, which,
under qualification conditions,
can be written as Problem~\ref{prob:main} with
\begin{align}\vspace{-10pt}
\label{e:primdualops}
X=\ensuremath{{\mathcal H}}=\mathrm{H}\times \mathrm{G},&&A = \partial \mathrm{f} \times
\partial \mathrm{g}^\ast, &&
B_1 = \nabla \mathrm{h} \times \{0\}, && B_2 =
\begin{bmatrix} 0 & \mathrm{L}^\ast \\ -\mathrm{L} & 0 \end{bmatrix},
\end{align}
{where we point out that $B_2$ is monotone and Lipschitz but {\em not
cocoercive},
because it is skew linear and, for every $x\in\ensuremath{{\mathcal H}}$, $\scal{x}{B_2x}=0$.}
We have that, for any solution $x=(\mathrm{x}_1^\ast, \mathrm{x}_2^\ast)\in
\ensuremath{\operatorname{zer}}(A+ B_1 + B_2)$, $\mathrm{x}_1^\ast$
solves~\eqref{eq:primal_before_PD}, where we denote $\ensuremath{\operatorname{zer}} T=\menge{x\in\ensuremath{{\mathcal H}}}{0\in Tx}$
for any set valued operator $T\colon\ensuremath{{\mathcal H}}{\tilde{o}} 2^{\ensuremath{{\mathcal H}}}$.
A method proposed in \cite{vu2013splitting} solves \eqref{eq:primal_before_PD} in a more
general context
by using forward-backward splitting (FB) in the product space with {the} metric
$\scal{\cdot}{\cdot}_V=\scal{V\cdot}{\cdot}$ for the operators
$V^{-1}(A+B_2)$ and $V^{-1}B_1$ with a specific choice of self-adjoint strongly
monotone linear operator $V$. We recall that the forward-backward splitting
\cite{combettes2004solving,bruck1975,lions1979splitting,goldstein1964}
finds a zero of the sum of a maximally monotone and a cocoercive operator, which
is a particular case of Problem~\ref{prob:main} when $X=\ensuremath{{\mathcal H}}$ and $B_2=0$. This
method provides a sequence obtained from the fixed point iteration of the
nonexpansive operator (for some
$\gamma\in]0,2\beta[$)
$$
\TFB := J_{\gamma A}\circ({\ensuremath{\operatorname{Id}}\,} - \gamma B_1),
$$
which converges weakly to a zero of $A+B_1$. Here ${\ensuremath{\operatorname{Id}}\,}$ stands for the
identity map in
$\ensuremath{{\mathcal H}}$ and{, for every set valued operator $M\colon\ensuremath{{\mathcal H}}{\tilde{o}} 2^{\ensuremath{{\mathcal H}}}$,
$J_{M}=({\ensuremath{\operatorname{Id}}\,}+M)^{-1}\colon\ensuremath{{\mathcal H}}{\tilde{o}} 2^{\ensuremath{{\mathcal H}}}$ is the
resolvent of $M$, which is single valued and
nonexpansive when $M$ is maximally monotone.}
In the context
of \eqref{e:primdualops}, the operators $V^{-1}(A+B_2)$ and $V^{-1}B_1$ are
maximally monotone and $\beta$-cocoercive in the metric
$\scal{\cdot}{\cdot}_V=\scal{V\cdot}{\cdot}$,
respectively, which ensures the convergence of the forward-backward splitting. The
choice of $V$
permits the explicit computation of $J_{V^{-1}(A+B_2)}$, which leads to a sequential
method that generalizes the algorithm proposed in \cite{chambolle2011first}. A variant
for solving
\eqref{eq:primal_before_PD} in the case when $h=0$ is proposed in
\cite{he2012convergence}. However, previous methods need the skew linear
structure of $B_2$ in order to obtain an implementable method.
{ An example in which a non-linear continuous operator $B_2$ arises naturally is
the convex constrained optimization problem
\begin{equation}
\min_{\substack{x\in C\\g(x)\le 0}}f(x),
\end{equation}
where $f\colon\ensuremath{{\mathcal H}}{\tilde{o}}\mathbb{R}$ is convex differentiable with $\beta^{-1}$-Lipschitz-gradient,
$C\subset\ensuremath{{\mathcal H}}$ is nonempty, closed and convex, and $g\colon\ensuremath{{\mathcal H}}{\tilde{o}}\mathbb{R}$ is a
$\mathcal{C}^1$ and convex function. The Lagrangian function in this case takes the form
\begin{equation}
L(x,\lambda)=\iota_C(x)+f(x)+\lambda g(x)-\iota_{\mathbb{R}_+}(\lambda),
\end{equation}
which, under standard qualification conditions can be found by solving the monotone
inclusion (see \cite{rockafellar1970saddle})
\begin{equation}
0\in A(x,\lambda)+B_1(x,\lambda)+B_2(x,\lambda),
\end{equation}
where $A\colon (x,\lambda)\mapsto N_Cx\times N_{\mathbb{R}_+}\lambda$ is maximally
monotone, $B_1\colon (x,\lambda)\mapsto (\nabla f(x),0)$ is cocoercive, and
$B_2\colon (x,\lambda)\mapsto (\lambda\nabla g(x),-g(x))$ is monotone and continuous
\cite{rockafellar1970saddle}. Of course, the problem can be easily extended to
consider finitely many
inequality and equality constraints and allow for more general lower semicontinuous
convex functions than $\iota_C$, but we prefer the simplified version for the ease of
presentation. Note that the non-linearity of $B_2$ does not allow to use previous
methods in this context.
}
In the {case when $B_2$ is $L$-Lipschitz for some $L>0$}, since $B:=B_1+B_2$
is monotone
and $(\beta^{-1}+L)$--Lipschitz
continuous, the forward-backward-forward splitting (FBF) proposed by Tseng in
\cite{tseng2000modified} solves Problem~\ref{prob:main}. This method
generates a sequence from the fixed point iteration of the operator
\begin{align*}
T_{\mathrm{FBF}} := P_X\circ \left[({\ensuremath{\operatorname{Id}}\,} - \gamma B) \circ J_{\gamma A} \circ
({\ensuremath{\operatorname{Id}}\,}-
\gamma B) + \gamma B\right],
\end{align*}
which converges weakly to a zero of $A+B$, provided that
$\gamma\in]0,(\beta^{-1}+L)^{-1}[$. However, this approach has two drawbacks:
\begin{enumerate}
\item FBF needs to evaluate $B=B_1+B_2$ twice per iteration, without taking into
advantage the cocoercivity property of $B_1$. In the particular case when $B_2=0$,
this method computes $B_1$ twice at each iteration, while the forward-backward
splitting needs only one computation of $B_1$ for finding a zero of $A+B_1$. Even if
we cannot ensure that FB is more efficient than FBF in this context, the cost
of each iteration of FB is lower than that of FBF, especially when the computation
cost of $B_1$ is high. This is usually the case, for instance, when $A$, $B_1$, and
$B_2$
are as in \eqref{e:primdualops} and we aim at solving \eqref{eq:primal_before_PD}
representing a
variational formulation of some partial differential equation (PDE). In this case,
the computation of $\nabla\mathrm{h}$ frequently amounts to solving a PDE, which
is computationally costly.
\item The step size $\gamma$ in FBF is bounded above by $(\beta^{-1}+L)^{-1}$,
which in the case when the influence of $B_2$ in the problem is low ($B_2\approx 0$)
leads to a method whose step size cannot go too far beyond $\beta$. In the case
$B_2=0$, the step size $\gamma $ in FB is bounded by $2\beta$. This can affect
the performance of the method, since very small stepsizes can lead to slow
algorithms.
\end{enumerate}
{In the general case when $B_2$ is monotone and continuous, we can also
apply a version of the method in \cite{tseng2000modified} which uses line search for
choosing the step-size at each iteration. However, this approach share the disadvantage of
computing
twice $B_1$ by iteration and, moreover, in the line search $B_1$ has to be computed several
times up to find a sufficiently small step-size, which can be computationally costly.}
In this paper we propose a splitting algorithm for solving Problem~\ref{prob:main}
which overcomes previous drawbacks. The method is derived from the
fixed point iteration of the operator $T_{\gamma}: {\mathcal{H}}
\rightarrow {\mathcal{H}}$, defined by
\begin{align}\label{eq:ouroperator}
T_\gamma := P_X\circ \left[({\ensuremath{\operatorname{Id}}\,}- \gamma B_2) \circ J_{\gamma A}
\circ ({\ensuremath{\operatorname{Id}}\,} -
\gamma (B_1+B_2)) + \gamma B_2\right],
\end{align}
for some $\gamma\in]0,\chi(\beta,L)[$, where $\chi(\beta,L)\leq\min\{2\beta,L^{-1}\}$ in the
case when $B_2$ is $L$-Lipschitz.
The algorithm thus obtained implements $B_1$ only once by iteration and it reduces to
FB or FBF when $X=\ensuremath{{\mathcal H}}$ and $B_2=0$, or $B_1=0$, respectively, and in these cases
we have $\chi(\beta,0)=2\beta$ and $\lim_{\beta{\tilde{o}}+\infty}\chi(\beta,L)=L^{-1}$.
{Moreover,
in the case when $B_2$ is merely continuous,
the step-size is found
by a line search in which $B_1$ is only computed once at each backtracking step.}
These results
can be found in Theorem~\ref{t:1} in Section~\ref{sec:2}.
Moreover,
a generalization of FB for finding a point in $X\cap\ensuremath{\operatorname{zer}}(A+B_1)$ can be derived
when $B_2=0$. This can be useful when the solution is known to belong to a closed
convex set $X$, which is the case, for example, in convex constrained minimization.
The additional projection onto $X$ can improve the performance of the method (see,
e.g., \cite{BAKS16}).
Another contribution of this paper is to include in our method non self-adjoint
linear operators in the computation of resolvents and other operators involved. More
precisely, in Theorem~\ref{thm:asymmetric_metric} in Section~\ref{sec:asymm}, for
an
invertible linear operator
$P$ (not necesarily self-adjoint) we justify
the computation of $P^{-1}(B_1+B_2)$ and
$J_{P^{-1}A}$, respectively.
In the case when $P$ is self-adjoint and strongly monotone,
the properties that $A$, $B_1$ and $B_2$ have with the standard metric
are preserved by $P^{-1}A$, $P^{-1}B_1$, and $P^{-1}B_2$ in the metric
$\scal{\cdot}{\cdot}_P=\scal{P\cdot}{\cdot}$. In this context, variable metric versions
of FB and FBF have been developed in
\cite{combettes2012variable,vu2013variableFBF}. Of course, a similar generalization
can be done for our algorithm, but we go beyond this self-adjoint case and we
implement $P^{-1}(B_1+B_2)$ and $J_{P^{-1}A}$, where the linear
operator $P$ is strongly monotone but non necesarily self-adjoint. The key for this
implementation is the decomposition $P=S+U$, where $U$ is self-adjoint and strongly
monotone and $S$ is skew linear. Our implementation follows after coupling $S$ with
the monotone and Lipschitz component $B_2$ and using some resolvent identities
valid for
the metric $\scal{\cdot}{\cdot}_U$. One of the important implications of this issue is
the justification of the convergence of some Gauss-Seidel type methods in product
spaces, which are deduced from our setting for block triangular linear operators $P$.
Additionally, we
provide a modification
of the previous method{ in Theorem~\ref{cor:asymmetricnoinversion}, in which
linear
operators
$P$ may vary among iterations. }In the case when, for every iteration $k\in\ensuremath{\mathbb N}$,
$P_k$ is self-adjoint, this feature has also been implemented for FB and FBF in
\cite{combettes2012variable,vu2013variableFBF}
but with a strong dependence between $P_{k+1}$ and $P_k$ coming from
the variable metric approach. Instead, in the general case, we modify our method for
allowing variable metrics and ensuring convergence under weaker
conditions. For instance, in the case when $B_2=0$ and $P_k$ is self-adjoint and
$\rho_k$-strongly monotone for some $\rho_k>0$,
our condition on our FB variable metric version reduces to
$(2\beta-\varepsilon)\rho_k>1$ for every
$k\in\ensuremath{\mathbb N}$. In the case when $P_k={\ensuremath{\operatorname{Id}}\,}/\gamma_k$ this
condition reduces to $\gamma_k<2\beta-\varepsilon$ which is a standard assumption
for FB with variable stepsizes. Hence, our condition on operators $(P_k)_{k\in\ensuremath{\mathbb N}}$
can be interpreted as ``step-size'' bounds.
Moreover, in Section~\ref{sec:5} we use our methods in composite primal-dual
inclusions, obtaining
generalizations and new versions of several primal-dual methods
\cite{chambolle2011first,vu2013variableFBF,patrinos2016asym,combettes2012primal}.
We provide comparisons among methods and new bounds on stepsizes which improve
several bounds in the literature. Finally, for illustrating the flexibility of the proposed
methods,
in Section~\ref{sec:6} we apply them to the obstacle problem in PDE's, to empirical
risk minimization, to distributed operator splitting schemes and to nonlinear constrained
optimization.
In the first example, we take advantage to dropping the extra forward step on $B_1$,
which amounts to reduce the computation of a PDE by iteration. In the
second example, we use non self-adjoint linear operators in order to obtain a
Gauss-Seidel
structure which can be preferable to parallel architectures {for high dimensions.
The third example illustrates how the variable metrics allowed by our proposed algorithm
can be used to develop distributed operator splitting schemes with time-varying
communication networks. The last example illustrates our backtracking line search
procedure for nonlinear constrained optimization wherein the underlying operator
$B_2$ is nonlinear and non Lipschitz. Finally, some numerical examples show the
performance of the proposed algorithms.
}
\section{Forward-backward-half forward splitting with
non self-adjoint linear operators}
\label{sec:asymm}
In this section, we introduce modified resolvents $J_{ P^{-1} A}$, which depend on an
invertible
linear mapping $P$. In some cases, it is preferable to compute the { modified
resolvent instead of the standard resolvent $J_{A} = ({\ensuremath{\operatorname{Id}}\,}+A)^{-1}$} because
the former may be easier to compute than the latter or, when $P$ is
triangular by blocks in a product space, the former may \textit{order the component
computation}
of the resolvent, replacing a parallel computation with a
Gauss-Seidel style sequential computation. However, ${P^{-1}A}$ may
not be maximally monotone. The following result allows us to use
some non self-adjoint linear operators in the computation of the resolvent
by using specific metrics. For simplicity, we assume from here
that $B_2$ is $L$-Lipschitz in $\ensuremath{\operatorname{dom}} A\cup X$, for some $L\geq 0$.
\begin{proposition}
\label{prop:new}
Let $A\colon\ensuremath{{\mathcal H}}{\tilde{o}} 2^{\ensuremath{{\mathcal H}}}$ be a maximally monotone operator, let $P\colon\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}$
be a linear bounded operator, and let $U := (P + P^\ast)/2$ and $S := (P -
P^\ast)/2$ be the
self-adjoint and skew symmetric components
of $P$, respectively. Assume that there exists $\rho>0$ such that
\begin{equation}
(\forall x\in\ensuremath{{\mathcal H}})\quad \rho\|x\|^2\le \scal{Ux}{x}=:\|x\|^2_U.
\end{equation}
Then, we have
\begin{equation}
J_{P^{-1}A}=J_{U^{-1}(A+S)}({\ensuremath{\operatorname{Id}}\,}+U^{-1}S).
\end{equation}
In particular, $J_{P^{-1}A}\colon\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}$ is single valued, everywhere defined and
satisfies
\begin{equation}
(\forall (x,y)\in\ensuremath{{\mathcal H}}^2)\quad \scal{J_{P^{-1}A}x-J_{P^{-1}A}y}{Px-Py}\ge
\|J_{P^{-1}A}x-J_{P^{-1}A}y\|^2_U
\end{equation}
and, hence, $U^{-1}P^*J_{P^{-1}A}$ is firmly nonexpansive in $(\ensuremath{{\mathcal H}},\scal{\cdot}{\cdot}_U)$,
where $\scal{\cdot}{\cdot}_U\colon (x,y)\mapsto\scal{Ux}{y}$.
\end{proposition}
\begin{proof}
Indeed, since $S$ is monotone and everywhere defined, $A+S$ is maximally monotone in
$\ensuremath{{\mathcal H}}$ \cite[Corollary~25.5]{bauschke2017convex} and, from
\cite[Lemma~3.7]{combettes2012variable}
we have that $U^{-1}(A+S)$ is maximally monotone in $\ensuremath{{\mathcal H}}$ with the metric
$\scal{\cdot}{\cdot}_U\colon(x,y)\mapsto\scal{x}{Uy}$. Hence, $J_{U^{-1}(A+S)}$ is single
valued (indeed firmly nonexpansive) and, for every $(x,z) \in {\mathcal{H}}^2$, we have
\begin{align*}
x = J_{ U^{-1}(A+S)}(z + U^{-1} Sz) &\quad \Leftrightarrow\quad z
+ U^{-1} S z - x \in
U^{-1}(A+S )x \\
&\quad\Leftrightarrow \quad (U+S)z- (U+S)x \in
Ax \\
&\quad\Leftrightarrow \quad x = J_{ P^{-1} A} z.
\end{align*}
Hence, for every
$(x,y)\in\ensuremath{{\mathcal H}}^2$, denoting by $p=J_{P^{-1}A}x=J_{U^{-1}(A+S)}(x+U^{-1} Sx)$
and $q=J_{P^{-1}A}y=J_{U^{-1}(A+S)}(y+U^{-1} Sy)$, the firm
nonexpansivity of
$J_{U^{-1}(A+S)}$ in $(\ensuremath{{\mathcal H}},\scal{\cdot}{\cdot}_U)$ yields
\begin{align*}
\scal{p-q}{Px-Py}&=
\scal{p-q}{U\left(x+U^{-1}
Sx-(y+U^{-1} Sy)\right)}\\
&=\scal{p-q}{x+U^{-1}
Sx-(y+U^{-1} Sy)}_U\\
&\ge \|p-q\|^2_U,
\end{align*}
and the result follows from $\scal{p-q}{Px-Py}=\scal{U^{-1}P^*(p-q)}{x-y}_U$.
\end{proof}
\begin{theorem}[New Metrics and $T_\gamma$]\label{thm:asymmetric_metric}
{Under the hypotheses of Problem~\ref{prob:main} and assuming additionally
that $B_2$ is $L$-Lipschitz in $\ensuremath{\operatorname{dom}} A\cup X$}, let $P : {\mathcal{H}} \rightarrow
{\mathcal{H}}$ be {a}
bounded linear
operator, let $U := (P + P^\ast)/2$ and $S := (P - P^\ast)/2$ be the the
self-adjoint and skew symmetric components
of $P$, respectively. Suppose that there exists $\rho>0$ such
that
\begin{align}
\label{e:metricconditions}
\left(\forall x \in {\mathcal{H}} \right) \qquad \rho\|x\|^2 \leq
\dotp{Ux, x}\quad\text{and}\quad
K^2<\rho\left(\rho-\frac{1}{2\beta}\right),
\end{align}
where $K\ge0$ is the Lipschitz constant of $B_2-S$.
Let $z^0 \in{\ensuremath{\operatorname{dom}} A\cup X}$ and let $\{z^k\}_{k\in\ensuremath{\mathbb N}}$ be the sequence defined
by
the
following iteration:
\begin{equation}
\label{eq:compressedalg}
(\forall k\in\ensuremath{\mathbb N})\quad
\begin{array}{l}
\left\lfloor
\begin{array}{l}
x^k = J_{ P^{-1} A}(z^k - P^{-1}(B_1 + B_2)z^k) \\[2mm]
z^{k+1} = P_X^U(x^k + U^{-1}(B_2z^k -
B_2x^k-S(z^k-x^k))),
\end{array}
\right.
\end{array}
\end{equation}
where $P_X^U$ is the projection operator of $X$ under the inner
product $\dotp{\cdot, \cdot}_U$.
Then $\{z^k \}_{k \in\ensuremath{\mathbb N}}$ converges weakly to a solution to Problem~\ref{prob:main}.
\end{theorem}
\begin{proof}
Note that, since $U$ is invertible from \eqref{e:metricconditions}, by adding and
subtracting the skew term $S$,
Problem~\ref{prob:main} is equivalent to
\begin{align}\label{eq:modifiedmono}
\text{find $x \in X$ such that} \quad 0 \in
U^{-1}(A+S)x + U^{-1}B_1{x} +
U^{-1}(B_2-S)x.
\end{align}
Because $S$ and $-S$ are both monotone and Lipschitz,
$\mathcal{A}:=U^{-1}(A+S)$ is monotone;
$\mathcal{B}_1:=U^{-1}B_1$ is $\rho \beta$-cocoercive~\cite[Proposition
1.5]{davis2014convergenceprimaldual};
and $\mathcal{B}_2:=U^{-1} (B_2 - S)$ is
monotone and $\rho^{-1}K$-Lipschitz {in $\ensuremath{\operatorname{dom}} A\cup X$} under the inner
product
$\dotp{\cdot,
\cdot}_U=\scal{U\cdot}{\cdot}$, where $K$
is the Lipschitz constant of $C:=B_2-S$.\footnote{Note that $K\leq
L+\|S\|$, but this constant is not precise when, for instance,
$B_2=S$. } For the last assertion
note that, for every $ x,y\, {\in \ensuremath{\operatorname{dom}} A\cup X}$, $$\|\mathcal{B}_2x -
\mathcal{B}_2y\|^2_{U} = \dotp{U^{-1}(Cx - Cy), Cx - Cy}
\leq \rho^{-1}K^2\|x-y\|^2\leq
\rho^{-2}K^2\|x-y\|^2_U.$$
Moreover, the stepsize condition reduces to
\begin{equation}
\label{e:auxiliary}
\gamma=1<\frac{4\beta\rho}{1+\sqrt{1+16\beta^2
K^2}}=\frac{-\rho+\sqrt{\rho^2+16\beta^2\rho^2
K^2}}{4\beta K^2}
\end{equation}
or, equivalently,
\begin{equation}
\label{e:aux221}
(4\beta
K^2+\rho)^2<{\rho^2+16\beta^2\rho^2
K^2}\quad
\Leftrightarrow\qquad
2\beta K^2+\rho<2\beta\rho^2,
\end{equation}
which yields the second condition in \eqref{e:metricconditions}.
Therefore, since $\mathcal{A}+\mathcal{B}_2=U^{-1}(A+B_2)$ is maximally monotone
in $({\mathcal{H}},\|\cdot\|_U)$, the inclusion~\eqref{eq:modifiedmono} meets the conditions of
Theorem~\ref{t:1} under this metric. {Therefore, by considering the sequence
generated
by
$z^{k+1}=T_1z^k$ for the
quasi-nonexpansive operator
\begin{align}
\label{e:defT1}
T_{1} = P_X^U\circ \left[({\ensuremath{\operatorname{Id}}\,} -
\mathcal{B}_2) \circ J_{ \mathcal{A}}
\circ \big({\ensuremath{\operatorname{Id}}\,} -
(\mathcal{B}_1+\mathcal{B}_2)\big) + \mathcal{B}_2\right],
\end{align}
which, from Proposition~\ref{prop:new} reduces to \eqref{eq:compressedalg},}
we obtain a sequence that weakly converges to a fixed point of
$T_{1}$, and hence, to a solution of $\ensuremath{\operatorname{zer}}(A+B_1+B_2)\cap
X$.
\end{proof}
\begin{remark}
\begin{enumerate}
\item Note that, in the particular case when $P=\ensuremath{\operatorname{Id}}\,/\gamma$, the
algorithm
\eqref{eq:compressedalg}
{reduces to \eqref{e:algomain1} when the stepsizes are constant}. Moreover,
$U=P$, $S=0$, $K=L$, $\rho=1/\gamma$ and the second condition in
\eqref{e:metricconditions} reduces to $\gamma<\chi$ with $\chi$
defined in \eqref{e:chi}. Hence, this assumption can be seen as a
kind of ``step size'' condition on $P$.
\item \label{rem:4}
As in Remark~\ref{rem:chi}, note that the second condition in
\eqref{e:metricconditions} depends
on the cocoercivity parameter $\beta$ and the Lipschitz constant
$L$. In the
case when $B_1$ is zero, we can take $\beta{\tilde{o}}+\infty$ and this
condition reduces to $K<\rho$. On the other hand, if $B_2$ is zero
we can take $L= 0$, then $K= \|S\|$ and, hence, the condition
reduces to $\|S\|^2<\rho(\rho-1/(2\beta))$.
In this way we obtain convergent versions
of Tseng's splitting
and forward-backward algorithm with non self-adjoint linear operators by
setting $B_1=0$ or $B_2=0$ in \eqref{eq:compressedalg}, respectively.
\item When $S=0$ {and $B_1=0$ or $B_2=0$, }
from Theorem~\ref{thm:asymmetric_metric} we
recover {the versions of Tseng's
forward-backward-forward splitting
\cite[Theorem~3.1]{vu2013variableFBF}
or forward-backward \cite[Theorem~4.1]{combettes2012variable},
respectively, when the step-sizes and the
non-standard metrics involved are constant. }Of course, when
$S=0$, $U=\ensuremath{\operatorname{Id}}\,/\gamma$, and $\rho=1/\gamma$, we recover
the classical bound for step-sizes in the standard metric case for each method.
\item For a particular choice
of operators and metric, the forward-backward method {
with non-standard metric} discused before
has been
used for solving
primal-dual composite inclusions and primal-dual optimization problems
{\cite{condat2013primal,vu2013splitting}}. This approach generalizes,
e.g., the method in \cite{chambolle2011first}.
In Section~\ref{sec:5} we compare the application of our method in the primal-dual
context
with \cite{vu2013splitting} and other methods in the literature.
\item In the particular instance when
$B_1=B_2=0$, we need $\|S\|<\rho$ and we obtain from
\eqref{eq:compressedalg} the
following version of the proximal point algorithm (we consider $X=\ensuremath{{\mathcal H}}$ for simplicity)
\begin{align*}
z^0\in\ensuremath{{\mathcal H}},\quad (\forall k\in\ensuremath{\mathbb N})\quad z^{k+1}&=J_{ P^{-1}
A}z^k+U^{-1}S(J_{ P^{-1}
A}z^k-z^k) \\
&=(\ensuremath{\operatorname{Id}}\,-U^{-1}P)z^k+U^{-1}PJ_{ P^{-1} A}z^k.\numberthis\label{e:ppavm}
\end{align*}
Moreover, in the case when $A=B_2=0$, since $U^{-1}\circ S\circ
P^{-1}=U^{-1}-P^{-1}$, we
recover from
\eqref{eq:compressedalg} the gradient-type method:
\begin{equation}
\label{e:gradvm}
z^0\in\ensuremath{{\mathcal H}},\quad (\forall k\in\ensuremath{\mathbb N})\quad z^{k+1}=z^k-
U^{-1}B_1z^k.
\end{equation}
\item In the particular case when $X=\ensuremath{{\mathcal H}}$ and $B_2$ is linear, in
\cite{patrinos2016asym} a method involving $B_2^*$ is proposed. In
the case when, $B_2$ is skew linear, i.e., $B_2^*=-B_2$
\eqref{e:metricconditions} reduces to this method in the case
$\alpha_n\equiv 1$ and $S=P$. The methods are different in general.
\end{enumerate}
\end{remark}
\section{Allowing variable $P$ and avoiding inversion of $U$}
\label{sec:4}
In Algorithm~\eqref{eq:compressedalg}, the linear operator $U$
must be inverted. In this section, for the special case ${\ensuremath{\operatorname{dom}} B_2=}X = {\mathcal{H}}$,
we show how to replace this sometimes costly inversion with a
single multiplication by the map $P$, which, in addition, may
vary at each iteration.
This new feature is a consequence of Proposition~\ref{prop:classT} below,
which allows us to obtain from an operator of the class
$\mathfrak{T}$ in $(\ensuremath{{\mathcal H}},\|\cdot\|_U)$, another operator of the same class
in $(\ensuremath{{\mathcal H}},\|\cdot\|)$ preserving the set of fixed points. This change
to the standard metric allows us to use different linear operators at each
iteration by avoiding classical restrictive additional assumptions of
the type $U_{n+1}\preccurlyeq U_n(1+\eta_n)$ with $(\eta_n)_{n\in\ensuremath{\mathbb N}}$
in $\ell^1_+$.
We recall
that an operator $\mathcal{S}\colon\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}$ belongs to the class
$\mathfrak{T}$
in $(\ensuremath{{\mathcal H}},\|\cdot\|)$ if and only if $\ensuremath{\operatorname{dom}} \mathcal{S}=\ensuremath{{\mathcal H}}$ and
$(\forall
y\in\ensuremath{\operatorname{Fix}} \mathcal{S})(\forall
x\in\ensuremath{{\mathcal H}})\quad
\|x-\mathcal{S}x\|^2\leq \scal{x-\mathcal{S}x}{x-y}$.
\begin{proposition}
\label{prop:classT}
Let $U\colon\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}$ {be} a self-adjoint bounded linear operator such
that, for every $x\in\ensuremath{{\mathcal H}}$, $\scal{Ux}{x}\geq\rho\|x\|^2$, for some
$\rho>0$, let $0<\mu\leq \|U\|^{-1}$, and let
$\mathcal{S}\colon\ensuremath{{\mathcal H}}{\tilde{o}}\ensuremath{{\mathcal H}}$
be
an
operator in the class $\mathfrak{T}$ in $(\ensuremath{{\mathcal H}},\|\cdot\|_U)$. Then, the
operator $\mathcal{Q}=\ensuremath{\operatorname{Id}}\,-\mu U(\ensuremath{\operatorname{Id}}\,-\mathcal{S})$ belongs to the
class
$\mathfrak{T}$
in $(\ensuremath{{\mathcal H}},\|\cdot\|)$ and $\ensuremath{\operatorname{Fix}} \mathcal{S}=\ensuremath{\operatorname{Fix}} \mathcal{Q}$.
\end{proposition}
\begin{proof}
First note that, under the assumptions on $U$ it is invertible and,
from \cite[Lemma~2.1]{combettes2013variable}, we deduce
\begin{equation}
\label{e:ineqU}
(\forall
x\in\ensuremath{{\mathcal H}})\quad \|x\|_U^2=\scal{Ux}{x}=\scal{Ux}{U^{-1}Ux}
\geq \|U\|^{-1}\|Ux\|^2,
\end{equation}
and $\ensuremath{\operatorname{Fix}} \mathcal{S}=\ensuremath{\operatorname{Fix}} \mathcal{Q}$ thus follows from the
definition of $\mathcal{Q}$.
Now let $y\in\ensuremath{\operatorname{Fix}}
\mathcal{S}$ and $x\in\ensuremath{{\mathcal H}}$. We have from
\eqref{e:ineqU} that
\begin{align}
\|x-\mathcal{S}x\|_U^2\leq \scal{x-\mathcal{S}x}{x-y}_U\:
&\Leftrightarrow\: \|x-\mathcal{S}x\|_U^2\leq
\scal{U(x-\mathcal{S}x)}{x-y}\nonumber\\
&\Rightarrow\: \|U\|^{-1}\|U(x-\mathcal{S}x)\|^2\leq
\scal{U(x-\mathcal{S}x)}{x-y}\nonumber\\
&\Leftrightarrow\: \frac{\|U\|^{-1}}{\mu}\|\mu
U(x-\mathcal{S}x)\|^2\leq
\scal{\mu U(x-\mathcal{S}x)}{x-y}\nonumber\\
&\Leftrightarrow\: \frac{\|U\|^{-1}}{\mu}\|x-\mathcal{Q}x\|^2\leq
\scal{x-\mathcal{Q}x}{x-y}
\end{align}
and, hence, if $\mu\in]0,\|U\|^{-1}]$ we deduce the result.
\end{proof}
\begin{theorem}\label{cor:asymmetricnoinversion}
{Under the hypotheses of Problem~\ref{prob:main}and assuming additionally
that $B_2$ is $L$-Lipschitz in $\ensuremath{\operatorname{dom}} B_2=\ensuremath{{\mathcal H}}$}, let $\{P_k\}_{k \in
\ensuremath{\mathbb N}}$ be a
sequence of bounded,
linear maps from ${\mathcal{H}}$ to ${\mathcal{H}}$. For each $k \in \ensuremath{\mathbb N}$, let
$U_k := (P_k + P_k^\ast)/2$ and ${S}_k := (P_k -
P_k^\ast)/2$
be the self-adjoint and skew symmetric components
of $P_k$, respectively. Suppose that
$M:=\sup_{k\in\ensuremath{\mathbb N}}\|U_k\|<\infty$ and that there
exist $\varepsilon\in]0,(2M)^{-1}[$, $\rho > 0$, and $\{\rho_k\}_{k
\in \ensuremath{\mathbb N}}
\subseteq [\rho,
\infty[$
such that, for every $k\in\ensuremath{\mathbb N}$,
\begin{align}
\label{e:metricconditions2}
\left(\forall x \in {\mathcal{H}} \right) \qquad \rho_k\|x\|^2 \leq
\dotp{U_kx, x} && \text{and} &&
K_k^2\leq\frac{\rho_k}{1+\varepsilon}\left(\frac{\rho_k}
{1+\varepsilon}-\frac{1}{2\beta}\right),
\end{align}
where $K_k\ge0$ is the Lipschitz constant of $B_2-S_k$.
Let $\{\lambda_k\}_{k \in \ensuremath{\mathbb N}}$ be a sequence in $[\varepsilon,
\|U_k\|^{-1} - \varepsilon]$, let $z^0 \in\ensuremath{{\mathcal H}}$, and let $\{z^k\}_{k
\in \ensuremath{\mathbb N}}$ be a sequence of points defined by the following iteration:
\begin{equation}
\label{eq:FBF-asymmetric-no-U}
(\forall k\in\ensuremath{\mathbb N})\quad
\begin{array}{l}
\left\lfloor
\begin{array}{l}
x^k = J_{P_k^{-1} A}(z^k - P_k^{-1}(B_1 + B_2)z^k) \\[2mm]
z^{k+1} = z^k + \lambda_k\left( P_k(x^k - z^k) + B_2z^k -
B_2x^k\right).
\end{array}
\right.
\end{array}
\end{equation}
Then $\{z^k \}_{k \in\ensuremath{\mathbb N}}$ converges weakly to a solution to Problem~\ref{prob:main}.
\end{theorem}
\begin{proof}
For every invertible and bounded linear map $P : {\mathcal{H}} \rightarrow
{\mathcal{H}}$, let us denote by
$\mathcal{T}_{ P} : {\mathcal{H}} \rightarrow {\mathcal{H}}$ the
forward-backward-forward
operator of Theorem~\ref{thm:asymmetric_metric} in the case
$X={\mathcal{H}}$, which associates, to every $z\in\ensuremath{{\mathcal H}}$,
\begin{align*}
\mathcal{T}_{P}z &= x_z + U^{-1}( B_2z - B_2x_z-S( z-x_z)),
\end{align*}
where $x_z = J_{P^{-1} A}(z -P^{-1}(B_1 + B_2)z)$. Recall that, from \eqref{eq:fejer}
and the proof of Theorem~\ref{thm:asymmetric_metric}, $\mathcal{T}_P$
is
a quasi-nonexpansive mapping in $\ensuremath{{\mathcal H}}$ endowed with the scalar product
$\scal{\cdot}{\cdot}_U$.
Observe that multiplying ${\ensuremath{\operatorname{Id}}\,} - \mathcal{T}_{P}$ by $U$ on the left
yields a $U^{-1}$-free
expression:
\begin{align*}
({\ensuremath{\operatorname{Id}}\,} - \mathcal{T}_{P})(z) &= (z - x_z) +U^{-1}S(z-x_z) -U^{-1}(B_2z -
B_2x_z)\\
\Leftrightarrow\qquad U({\ensuremath{\operatorname{Id}}\,} - \mathcal{T}_{P})(z) &= (U+S)(z - x_z) +
B_2x_z - B_2z\\
&= P(z - x_z) + B_2x_z -
B_2z\numberthis\label{eq:Snotation}.
\end{align*}
Note that, since $\mathcal{T}_P$ is quasi-nonexpansive in
$(\ensuremath{{\mathcal H}},\|\cdot\|_U)$,
it follows from
\cite[Proposition~2.2]{combettes2001quasi} that
$\mathcal{S}:=(\ensuremath{\operatorname{Id}}\,+\mathcal{T}_P)/2$
belongs to the class $\mathfrak{T}$ in $(\ensuremath{{\mathcal H}},\|\cdot\|_U)$ and, from
Proposition~\ref{prop:classT} and \eqref{eq:Snotation} we obtain that the operator
\begin{equation}
\label{e:QP}
\mathcal{Q}_{P}:={\ensuremath{\operatorname{Id}}\,}-\|U\|^{-1}U({\ensuremath{\operatorname{Id}}\,}-\mathcal{S})=
{\ensuremath{\operatorname{Id}}\,}-\frac{\|U\|^{-1}}{2}U({\ensuremath{\operatorname{Id}}\,}-\mathcal{T}_P)
\end{equation}
belongs to the class $\mathfrak{T}$ in
$(\ensuremath{{\mathcal H}},\|\cdot\|)$ and $\ensuremath{\operatorname{Fix}} \mathcal{S}=\ensuremath{\operatorname{Fix}}
\mathcal{Q}_{P}=\ensuremath{\operatorname{zer}}(U({\ensuremath{\operatorname{Id}}\,} -
\mathcal{T}_{P}))=\ensuremath{\operatorname{Fix}}(\mathcal{T}_{P})=\ensuremath{\operatorname{zer}}(A+B_1+B_2)$.
Hence, from \eqref{eq:Snotation} and \eqref{e:QP}, the algorithm
\eqref{eq:FBF-asymmetric-no-U} can be written equivalently as
\begin{align}
z^{k+1}&=z^k-\lambda_k(P_k(z^k - x_{z^k}) + B_2x_{z^k} -
B_2z^k)\nonumber\\
&=z^k+2\lambda_k\|U_k\|(\mathcal{Q}_{P_k}z^k-z^k).
\end{align}
Hence, since $0<\liminf\lambda_k\|U_k\|\le \limsup \lambda_k\|U_k\|<1$,
it follows from
\cite[Proposition~4.2 and
Theorem~4.3]{combettes2001quasi}
that
$(\|z^k-\mathcal{Q}_{P_k}z^k\|^2)_{k\in\ensuremath{\mathbb N}}$ is a summable sequence
and
$\{z^k\}_{k\in\ensuremath{\mathbb N}}$ converges weakly in $(\ensuremath{{\mathcal H}},\scal{\cdot}{\cdot})$ to
a solution to
$\cap_{k\in\ensuremath{\mathbb N}}\ensuremath{\operatorname{Fix}} \mathcal{T}_{P_k}=\ensuremath{\operatorname{zer}}(A+B_1+B_2)$ if and only if
every
weak limit of the sequence is a solution. Note that, since \eqref{e:metricconditions2} yields
$\|U_k^{-1}\|\leq \rho_k^{-1}$, we have
\begin{align}
\label{e:tozero1}
\|z^k-\mathcal{T}_{P_k}z^k\|_{U_k}^2&=\scal{U_{k}(z^k-\mathcal{T}_{P_k}z^k)}{z^k-
\mathcal{T}_{P_k}z^k}\nonumber\\
&\le \|U_{k}(z^k-\mathcal{T}_{P_k}z^k)\|\,\|z^k-\mathcal{T}_{P_k}z^k\|\nonumber\\
&=\|U_k^{-1}\|\|U_{k}(z^k-\mathcal{T}_{P_k}z^k)\|^2\nonumber\\
&\le 4\|U_k\|^2\rho_k^{-1}\,\|z^k-\mathcal{Q}_{P_k}z^k\|^2\nonumber\\
&\le 4M^2\rho^{-1}\|z^k-\mathcal{Q}_{P_k}z^k\|^2{\tilde{o}} 0.
\end{align}
Moreover, since $\mathcal{T}_{P_k}$ coincides
with $T_1$ defined in \eqref{e:defT1} involving the operators
$\mathcal{A}_k:=U_k^{-1}(A+S_k)$, $\mathcal{B}_{1,k}=U_k^{-1}B_1$,
and
$\mathcal{B}_{2,k}=U_k^{-1}(B_2-S_k)$ which are
monotone, $\rho_k\beta$-cocoercive, and monotone and
$\rho_k^{-1}K_k$-lipschitzian in $(\ensuremath{{\mathcal H}},\|\cdot\|_{U_k})$,
respectively,
we deduce from \eqref{eq:fejer}
that, for every
$z^{\ast}\in\ensuremath{\operatorname{zer}}(A+B_1+B_2)=\cap_{k\in\ensuremath{\mathbb N}}\ensuremath{\operatorname{zer}}(\mathcal{A}_k+
\mathcal{B}_{1,k}+\mathcal{B}_{2,k})$ we have
\begin{align}
\label{e:aux331}
&\rho_k^{-2}K_k^{2}(\chi_k^2-
1)\|z^k - J_{P_k^{-1} A}(z^k - P_k^{-1}(B_1+
B_2)z^k)\|_{U_k}^2 \nonumber\\
&+ \frac{2\beta\rho_k}{\chi_k}\left(\chi_k
- 1\right)\|U_k^{-1}(B_1z^k - B_1 z^\ast)\|_{U_k}^2\nonumber \\
&+\frac{\chi_k}{2\beta\rho_k}\left\|z^k-J_{P_k^{-1}A} (z^k - P_k^{-1}
(B_1z^k+ B_2z^k))-
\frac{2\beta\rho_k}{\chi_k}U_k^{-1}(B_1z^k-B_1z^*)\right\|_{U_k}^2\nonumber\\
&\leq \|z^k - z^\ast\|^2_{U_k} - \|\mathcal{T}_{P_k}z^k -
z^\ast\|^2_{U_k}\nonumber\\
&= -\|\mathcal{T}_{P_k}z^k - z^k\|_{U_k}^2 -
2\dotp{\mathcal{T}_{P_k}z^k - z^k,
z^\ast -
z^k}_{U_k}\nonumber\\
&\leq-\|\mathcal{T}_{P_k}z^k - z^k\|_{U_k}^2 +
2M\|\mathcal{T}_{P_k}z^k - z^k\|_{U_k}
\|z^\ast - z^k\|,
\end{align}
where
\begin{equation}
\label{e:chikbound}
\chi_k:=\frac{4\beta
\rho_k}{1+\sqrt{1+16\beta^2K_k^2}}\leq\rho_k\min\{2\beta,K_k^{-1}\}.
\end{equation}
By straightforward computations in the line
of \eqref{e:auxiliary} and \eqref{e:aux221} we deduce that
\eqref{e:metricconditions2}
implies, for all $k\in\ensuremath{\mathbb N}$, $\chi_k\geq1+\varepsilon$, $K_k\le \rho_k\le \|U_k\|\le M$ and,
hence,
we deduce from \eqref{e:aux331} and \eqref{e:metricconditions2} that
\begin{multline}
\label{e:tozero2}
\frac{\varepsilon\rho K_k^2}{M^2}\|z^k - J_{P_k^{-1} A}(z^k -
P_k^{-1}(B_1+
B_2)z^k)\|^2+{\varepsilon\rho}\|U_k^{-1}(B_1z^k - B_1
z^\ast)\|^2\\
+\frac{\rho}{2\beta M}\left\|z^k-J_{P_k^{-1}A} (z^k - P_k^{-1}
(B_1z^k+ B_2z^k))-
\frac{2\beta\rho_k}{\chi_k}U_k^{-1}(B_1z^k-B_1z^*)\right\|^2\\
\leq-\|\mathcal{T}_{P_k}z^k - z^k\|_{U_k}^2 +
2M\|\mathcal{T}_{P_k}z^k - z^k\|_{U_k}
\|z^\ast - z^k\|.
\end{multline}
Now, let $z$ be a weak limit of some subsequence of $(z^{k})_{k\in\ensuremath{\mathbb N}}$ called
similarly for simplicity.
We have that $(\|z^\ast -
z^k\|)_{k\in\ensuremath{\mathbb N}}$ is bounded and, since \eqref{e:tozero1} implies
$\|z^k-\mathcal{T}_{P_k}z^k\|_{U_k}^2\to0$ we deduce from
\eqref{e:tozero2}
that{, by denoting $x^k:=J_{P_k^{-1} A}(z^k - P_k^{-1}(B_1+
B_2)z^k)$, that $z^k-x^k\to0$. Hence, since, for every $x\in\ensuremath{{\mathcal H}}$,
\begin{equation}
\|S_kx\|\leq \|(S_k-B_2)x-(S_k-B_2)0\|+\|B_2x-B_20\|\leq
(K_k+L)\|x\|\leq\left(M+L\right)\|x\|,
\end{equation}
we have
\begin{align}
\label{e:goingto0}
\|P_k(z^k-x^k)\|&=\|(U_k+S_k)(z^k-x^k)\|\nonumber\\
&\leq \|U_k(z^k-x^k)\|+\|S_k(z^k-x^k)\|\nonumber\\
&\leq
(2M+L)\|z^k-x^k\|{\tilde{o}} 0.
\end{align}
Finally, denoting by $B:=B_1+B_2$
we have
\begin{equation}
u^k:=P_k(z^k-x^k)-(Bz^k-Bx^k)\in (A+B)x^k,
\end{equation}
and since $z^k-x^k\to0$ and $B$ is continuous, it follows from
\eqref{e:goingto0} that $u^k{\tilde{o}} 0$ and the result follows from the
weak-strong closedness of the maximally monotone operator $A+B$
and \cite[Theorem
5.33]{bauschke2017convex}. }
\end{proof}
\newpage7
\begin{remark}
\begin{enumerate}
\item Note that, in the particular case when $S_k\equiv 0$ and
$P_k=U_k=\gamma_k^{-1}V_k^{-1}$, we have from
\cite[Lemma~2.1]{combettes2013variable} that
$\rho_k=\gamma_k^{-1}\|V_k^{-1}\|$, the conditions on the constants
involved in Theorem~\ref{cor:asymmetricnoinversion} reduce to
\begin{equation}
\label{e:condit_sym_case}
\frac{\|V_k^{-1}\|}{M}\leq\gamma_k\leq
\frac{\|V_k^{-1}\|}{\rho},\quad L^2\leq
\frac{\gamma_k^{-1}\|V_k^{-1}\|}{1+\varepsilon}
\left(\frac{\gamma_k^{-1}\|V_k^{-1}\|}{1+\varepsilon}-\frac{1}{2\beta}\right),
\end{equation}
for some $0<\rho<M$, for every $k\in\ensuremath{\mathbb N}$, and \eqref{eq:FBF-asymmetric-no-U}
reduces to
\begin{equation}
\label{eq:FBF-symmetric-no-U}
(\forall k\in\ensuremath{\mathbb N})\quad
\begin{array}{l}
\left\lfloor
\begin{array}{l}
x^k = J_{\gamma_kV_k A}(z^k - \gamma_k V_k(B_1 + B_2)z^k) \\[2mm]
z^{k+1} = z^k + \frac{\lambda_k}{\gamma_k}\left( V_k^{-1}(x^k - z^k)
+ \gamma_kB_2z^k - \gamma_kB_2x^k\right).
\end{array}
\right.
\end{array}
\end{equation}
If in addition we assume that $B_2=0$ and, hence $L=0$,
\eqref{e:condit_sym_case} reduces to $\gamma_k\leq
\|V_k^{-1}\|2\beta/(1+\varepsilon)$ which is more general than the
condition in \cite{combettes2012variable} and, moreover,
we do not need any {compatibility assumption on} $(V_k)_{k\in\ensuremath{\mathbb N}}$ for
achieving
convergence. Similarly, if $B_1=0$, and hence, we can take
$\beta{\tilde{o}}\infty$, \eqref{e:condit_sym_case} reduces to $\gamma_k\leq
\|V_k^{-1}\|/(L(1+\varepsilon))$ which is more general than the condition in
\cite{vu2013variableFBF} and no additional assumption on
$(V_k)_{k\in\ensuremath{\mathbb N}}$ is needed. However, \eqref{eq:FBF-symmetric-no-U}
involves an additional computation of $V_k^{-1}$ in the last step of
each iteration $k\in\ensuremath{\mathbb N}$.
\item In the particular case when, for every $k\in\ensuremath{\mathbb N}$,
$P_k=U_k=\ensuremath{\operatorname{Id}}\,/\gamma_k$, where $(\gamma_k)_{k\in\ensuremath{\mathbb N}}$ is a real
sequence, we have $S_k\equiv0$, $K_k\equiv L$,
$\|U_k\|=\rho_k=1/\gamma_k$, and conditions
$\sup_{k\in\ensuremath{\mathbb N}}\|U_k\|<\infty$ and \eqref{e:metricconditions2} reduce
to
\begin{equation}
0<\inf_{k\in\ensuremath{\mathbb N}}\gamma_k\leq\sup_{k\in\ensuremath{\mathbb N}}\gamma_k<\chi,
\end{equation}
where $\chi$ is defined in \eqref{e:chi} and
\eqref{eq:FBF-asymmetric-no-U} reduces to
\begin{equation*}
(\forall k\in\ensuremath{\mathbb N})\quad
\begin{array}{l}
\left\lfloor
\begin{array}{l}
x^k =J_{\gamma_k A}(z^k - \gamma_k(B_1 + B_2)z^k) \\[2mm]
z^{k+1} = z^k + \eta_k\left( x^k +
\gamma_kB_2z^k - \gamma_kB_2x^k- z^k\right),
\end{array}
\right.
\end{array}
\end{equation*}
where $\eta_k\in[\varepsilon,1-\varepsilon]$, which is a relaxed version of
Theorem~\ref{t:1}.
\item As in Remark~\ref{rem:4}, by setting $B_1=0$ or $B_2=0$, we can
derive from \eqref{eq:FBF-asymmetric-no-U} versions of Tseng's
splitting and forward-backward algorithm with non self-adjoint linear operators
but without needing the inversion of $U$. In particular, the proximal
point algorithm in
\eqref{e:ppavm} reduces to
\begin{equation}
\label{e:ppavmwinv}
z^0\in\ensuremath{{\mathcal H}},\quad (\forall k\in\ensuremath{\mathbb N})\quad z^{k+1}=z^k+\lambda
P(J_{P^{-1} A}z^k-z^k)
\end{equation}
for $\lambda<\|U\|^{-1}$ and, in the case of \eqref{e:gradvm},
to avoid inversion is to
come back to the gradient-type method with the standard metric.
\end{enumerate}
\end{remark}
|
\section{Introduction}
The communication paradigm of cooperative relaying has recently received considerable attention due to its effectiveness in alleviating the effects of multipath fading, pathloss and shadowing, and its ability to deliver improved performance in cognitive radio systems and wireless sensor networks. Relay-assisted cellular networks are a promising solution for enhancing coverage and are included in standards, such as IEEE 802.16j/m and 3GPP Long Term Evolution-Advanced (LTE-A).
Several protocols for cooperative relaying were presented in \cite{LAN} where the gains in transmit and receive diversity were studied. In multi-relay networks, simultaneous transmissions by the relays are in general difficult to handle; towards this end, opportunistic relay selection has been suggested in \cite{BLE} to improve the resource utilization and to reduce the hardware complexity. Stemming from the relay selection concept, many studies have proposed improved selection techniques (see, \emph{e.g.}, \cite{BLE1, KAR, KRI1}).
Traditional HD relaying schemes partition the packet transmission slot into two phases, where the transmission on the Source-Relay $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link happens in the first phase, and the transmission on the Relay-Destination $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ link occurs in the second phase. However, this relaying scheme limits the maximum achievable multiplexing gain to $0.5$, which also results in bandwidth loss.
In order to overcome such multiplexing and bandwidth limitations, several techniques have been proposed in the literature (see, for example, \cite{DIN} and references therein). Among them, the successive relaying scheme in \cite{FAN} incorporates multiple relay nodes and allows concurrent transmissions between source-relay and relay-destination to mimic an ideal full-duplex transmission. However, this scheme targets scenarios with a long distance between the relays and thus inter-relay interference is not considered. An extension of this work is discussed in \cite{chao_ISIT}, where the authors assume that IRI is strong (in co-located or clustered relays) and can always be decoded at the affected nodes; this decoded IRI is exploited in a superposition coding scheme that significantly improves the diversity-multiplexing trade-off performance of the system.
In earlier studies, relays were assumed to lack data buffers and relay selection was mainly based on the $\max-\min$ criterion and its variations (see, for example, \cite{BLE,BLE1,KAR,KRI1}). Here, the relay that receives the source signal is also the one that forwards the signal to the destination. With the adoption of buffer-aided relays, this coupling is broken, since different relays could be selected for transmission and reception, thus allowing increased degrees of freedom. Buffering at the relay nodes has been shown to be a promising solution for cooperative networks and motivates the investigation of new protocols and transmission schemes (see \cite{survey} and \cite{ZLA_CM} for an overview). Ikhlef \emph{et al.} \cite{IKH2} proposed a novel criterion based on $\max-\max$ relay selection (MMRS), in which the relay with the best source-relay $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link is selected for reception and the relay with the best relay-destination $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ link is selected for transmission on separate slots. In \cite{KRICHAR}, at each slot the best link is selected among \emph{all} the available $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links, as a part of the proposed $\mathop{\mathrm{max-link}}$ policy, thus offering an additional degree of freedom to the network, while buffer-aided link selection was studied in topologies with source-destination connectivity (see, \emph{e.g.,} in \cite{CHAR_CL, SHAQ_TWC}), resulting in improved diversity and throughput. More recently, to alleviate the throughput loss of HD relaying, significant focus has been given to minimizing the average packet delay. Towards this end, there have been proposed various approaches: hybrid solutions combining the benefits of MMRS and $\mathop{\mathrm{max-link}}$ \cite{OIWA_TVT}, {the use of broadcasting in the $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link \cite{OIWA_ACCESS, NOM_TCOM2018},} the prioritization of $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ transmissions \cite{NOM_TCOM2018,POUL1,TIAN_TVT} or the selection of the relay with the maximum number of packets residing in its buffer \cite{LIN_TVT}. However, it was shown that a trade-off exists between delay performance and the diversity of transmission as the number of relays with empty or full buffers increases. In \cite{POUL2} delay- and diversity-aware relay selection policies were proposed aiming at reducing the average delay by incorporating the buffer size of the relay nodes into the relay selection process.
To improve throughput and reduce average packet delays, in a number of studies, it was proposed to employ non-orthogonal successive transmissions by the source and a selected relay. In order to recover the HD multiplexing loss, \cite{IKH3} suggests to combine MMRS with successive transmissions (SFD-MMRS). As the proposed topology aims to mimic full-duplex relaying, different relays are selected in the same time slot; however, relays are assumed to be isolated and the effect of IRI is ignored. More practical topologies were studied in \cite{NOM2, NOM_TCOM, SIM_TWC} where IRI exists and is not always possible to be cancelled. For fixed rate transmission, the proposed $\mathop{\mathrm{min-power}}$ solution proposed in \cite{NOM_TCOM}, combining power adaptation and interference cancellation, provides a performance close to the upper bound of SFD-MMRS. In addition, Kim and Bengtsson~\cite{KIM,KIM2} proposed buffer-aided relay selection and beamforming schemes taking the IRI into consideration; they consider a model which can be regarded as an example of relay-assisted device-to-device (D2D) communications, where the source and destination are low-cost devices with a single antenna and the relays comprise more powerful relays with buffers and multiple antennas. Numerical results show that their approach outperforms SFD-MMRS when interference is taken into consideration, and when the number of relays and antennas increases they approach the performance of the interference-free SFD-MMRS, herein called the \emph{ideal} SFD-MMRS. Finally, the use of buffer aided relays has been considered in topologies with full-duplex (FD) capabilities \cite{survey, NOM_FD1, ZLA_CL2016} and decoupled non-orthogonal uplink and downlink transmissions \cite{LIU_TCOM}, showing that throughput can be significantly improved due to the increased scheduling flexibility. Recently, in line with non-orthogonal successive transmissions, there has been considerable attention on non-orthogonal multiple access due to its superior spectral efficiency, and its combination with buffer-aided relaying was inevitable \cite{ZHANG_TCOM2017, NOM_GCW, CAO_TMC2018}.
In many practical considerations (\emph{e.g.}, wireless sensors), the relay nodes are hardware-limited to be HD while the source can be a more powerful wireless device with large buffers and multiple antennas. Although this observation is not always true (\emph{e.g.}, in D2D communications~\cite{KIM,KIM2}), it is a reasonable and common practical scenario. Towards this end, we study a network which consists of a buffer-aided source with a single or multiple antennas, multiple HD buffer-aided relays and a destination. With this setup,
\begin{list4}
\item we are able to approach (achieve) the performance of the ideal SFD-MMRS by adopting a buffer-aided source whose buffer retains replicas of the relay buffers, successfully transmitted data from the source to relays, to facilitate IRI mitigation (cancellation);
\item we relax the assumption of knowing the full CSI and instead, we allow for CSIR and limited feedback from the receiver to the transmitter; under these conditions we propose a relay pair selection scheme that is based on partial phase alignment of signals.
\end{list4}
The rest of the paper is organized as follows. In Section~\ref{sec:model}, the system model is outlined. The proposed relaying schemes for variable and fixed rate are presented in Sections~\ref{sec:policy1} and \ref{sec:policy2}, respectively. The performance of the proposed relaying policies in terms of outage and average throughput, along with comparisons with other state-of-the-art relaying schemes are presented in Section~\ref{sec:numerical}. Finally, conclusions and a discussion on future possible directions are drawn in Section~\ref{sec:conclusions}.
\subsubsection*{Notation}\label{sec:notation}
Vectors are written in bold lower case letters and matrices in bold capital letters. $\mathbb{R}$, $\mathbb{C}$ and $\mathbb{N}$ denote the sets of real numbers, complex and natural numbers, respectively. For a matrix~$\mathbf{A}\in \mathbb{C}^{n\times n}$ ($n\in \mathbb{N}$), $a_{ij}$ denotes the entry in row $i$ and column~$j$. Matrix $\mathbf{I}$ denotes the identity matrix of appropriate dimensions. The trace of a square matrix $\mathbf{A}\in\mathbb{R}^{n\times n}$ is denoted by $\trace(\mathbf{A})$. $(\cdot)^T$ and $(\cdot)^H$ denote the transpose and hermitian transpose operations, respectively; $\|\cdot \|$ denotes the $2$-norm operation. The complex conjugate of a complex number $z$ is denoted by $\overline{z}$; the real and imaginary parts of a complex number $z$ are denoted by $\mathcal{R}(z)$ and $\mathcal{I}(z)$, respectively. We denote by $\mathcal{P}$ the set of all possible relay-pairs in the relay network, and by $|\mathcal{P}|$ its cardinality. A relay pair, denoted by $(R,T)$ belongs to set $\mathcal{P}$ if and only if $R\neq T$, the receiving relay $R$ is not full and the transmitting relay $T$ is not empty.
\section{System model}\label{sec:model}
We consider a cooperative network consisting of a buffer-aided source $S$ with multiple antennas, a set $\mathcal{K} \triangleq \{1, 2, \ldots, K \}$ of $K$ HD decode-and-forward (DF) relays with buffers, and a single destination $D$. Fig.~\ref{fig:model} illustrates a simple example with a buffer-aided two-antenna source $S$, two buffer-aided single-antenna HD DF relays, and a single-antenna destination $D$. To simplify the analysis, we examine the case where connectivity between the source and the destination is established only via the relays and ignore the direct $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ link (as in, \emph{e.g.}, \cite{IKH2,KRICHAR,IKH3,KIM,KIM2}).
The number of data elements in the buffer of relay $R_k$ is denoted by $Q_k$ and its capacity by $Q_{\max}$. For the fixed rate transmission policy, we assume that the packets are transmitted at a fixed rate of $C_0$ bits per channel use (BPCU), and the data of each transmission occupies a single time slot in the buffer.
First, we provide the signals received at relay $R$ and destination $D$. At the destination, at any arbitrary time-slot $n$ the following signal is received:
\begin{align}
y_{D}[n]=h_{TD}x[p]+w_D[n] \; ,
\end{align}
where $x[p]$ is the signal received and stored in a previous time-slot $p$ in the buffer of the now transmitting relay $T$, $h_{TD}$ denotes the channel coefficient from the transmitting relay $T$ to the destination $D$, and $w_{D}[n]$ denotes the additive white Gaussian noise (AWGN) at the destination in the $n$-th time slot, i.e., $w_{D}[n]\sim\mathcal{CN}(0,{\sigma_D^2})$. It must be noted that $x[p]$ was not necessarily received in the just before time-slot (i.e., $p\leq n-1$). At the same time, the reception of the source's signal by relay $R$ is interfered from the transmission of $T$ which forwards a previous signal $x[p]$ to the destination. Thus, $R$ receives
\begin{align}\label{eq:yRgeneral}
y_{R}[n]=\sum_{i\in \mathcal{A}} h_{S_i R} x_{S_i}[n] + h_{TR} x_T[n] +w_R[n] ,
\end{align}
where $\mathcal{A}$ denotes the index set of transmit antennas at the source, i.e., $\mathcal{A}=\{1,2,\ldots,\nu\}$, $h_{S_i R}$ denotes the channel coefficient from the $i$-th transmit antenna at the source to the receiving relay $R$, $h_{TR}$ denotes the channel coefficient from the transmitting relay $T$ to the receiving relay $R$, $x_{S_i}[n]$ denotes the transmitted signal from the $i$-th transmit antenna at the source in the $n$-th time slot, $x_T[n]$ denotes the transmitted signal from the transmitting relay in the $n$-th time slot (i.e., $x_T[n] = x[p]$), and $w_{R}[n]$ denotes the AWGN at the receiving relay in the $n$-th time slot, i.e., $w_{R}[n]\sim\mathcal{CN}(0,{\sigma^2})$.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{system_model_mod_sumin2}
\caption{A simple example of a cooperative network consisting of a {buffer-aided} source $S$ with two antennas ($S_1$ and $S_2$), two HD relays (the receiving relay is denoted by $R$ and the transmitting relay by $T$) and a destination $D$; in this example, $R_1 \equiv R$ and $R_2 \equiv T$. The buffers at $S$ consist basically of replicas of the data queues of the relays; the source has new packets in the source buffer $Q_{S}$ and replicas of the successfully transmitted packets to the relays in a set of copied buffers.}
\label{fig:model}
\end{figure}
The source $S$ is assumed to be saturated (infinite backlog) and hence, it has always data to transmit. The buffering memory at the source is organized into $K$ queues, that basically contain replicas of the data queues of the relays, in order to exploit it for IRI mitigation or cancellation.
The operation is assumed to be divided into time slots. In each time slot, the source and a relay simultaneously transmit their own data to mimic FD relaying (cf. \cite{IKH3,KIM,KIM2,NOM2,NOM_TCOM}). The transmission powers of the source and the $k$-th relay are denoted by $P_S$ and $P_T$, respectively. For notational simplicity, we assume throughout this paper that all devices use a common fixed transmit power level (i.e., $P_S=P_T=P,~\forall T\in\mathcal{K}$), unless otherwise specified. Moreover, we assume that the receivers send short-length error-free acknowledgment/negative-acknowledgment (ACK/NACK) messages over a separate control channel.
We assume narrowband Rayleigh block fading channels. Each channel coefficient is constant during one time slot and varies independently between time slots. For each time slot, the channel coefficient $h_{ij}$ for link $\{i\hspace{-0.09cm}\rightarrow \hspace{-0.09cm} j\}$ follows a circular symmetric complex Gaussian distribution with zero mean and variance $\sigma_{ij}^2$, i.e., $h_{ij}\sim\mathcal{CN}(0,\sigma_{ij}^2)$. Thus, the channel power gain $g_{ij} \triangleq |h_{ij}|^2$ follows an exponential distribution, i.e., $g_{ij}\sim\mathrm{Exp}({\sigma_{ij}^{-2}})$.
In addition, we assume the AWGN at each receiver with variance $\sigma^2$.
\section{Buffer-Aided Relay Selection Based on Buffer-Aided Source Precoding}\label{sec:policy1}
The goal is to design a precoding matrix at the source such that, at each time slot $n$ the $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link for the relay selected to receive a packet from the source maximizes its SINR by transmitting a linear combination of (a) the source's desired signal for the receiving relay $R$ and (b) the signal of the transmitting relay $T$ taken out from the copied buffer at the source. Note that maximizing the SINR of the link is equivalent to minimizing the outage probability of the link. Thus, this precoding design criterion is relevant for both fixed rate and adaptive rate transmissions.
\noindent \textbf{Special case: Source with a single antenna.} The received signal at the receiving relay for the $n$-th time slot, $y_R[n]$, can be expressed as
\begin{align}\label{signals_singleantenna}
y_R[n] &= h_{SR}\underbrace{\big( m_{1}x_S[n]+m_{2}x_T[n]\big)}_{\text{source signal}}+h_{TR} x_T[n] +w_R[n] \nonumber \\
&= h_{SR} \underbrace{\begin{bmatrix}m_1 & m_2\end{bmatrix}}_{\text{precoding matrix}} \underbrace{\begin{bmatrix}x_S[n] & x_T[n]\end{bmatrix}^T}_{\text{data from the source}} +h_{TR} x_T[n] +w_R[n],
\end{align}
where $x_{S_i}[n]$ in Eq.~\eqref{eq:yRgeneral} is written as $m_{1}x_S[n]+m_{2}x_T[n]$, and $m_{1},m_{2} \in \mathbb{C}$. Each transmitting node uses a fixed power $P$ and thus power control issues are not taken into account. The SINR at the receiving relay, denoted by $\Gamma_R$, is given by
\begin{align}
\Gamma_R\triangleq \frac{|h_{SR} {m}_1|^2 P}{|h_{SR} m_2 + h_{TR}|^2 P + \sigma^2} .
\end{align}
Hence, the following optimization problem can be formulated:
\begin{subequations}
\label{OPT0_singleantenna}
\begin{align}
\displaystyle \underset{m_1,m_2}{\max} & ~ \frac{|h_{SR} {m}_1|^2 P}{|h_{SR} m_2 + h_{TR}|^2 P + \sigma^2}\label{0obj1_sa} \\[0.2cm]
\textrm{s.t.}
& ~ |{m}_1|^2 + |{m}_2|^2 \leq 1. \label{0cond2_sa}
\end{align}
\end{subequations}
The solution to optimization problem \eqref{OPT0_singleantenna} will be given for the general case and the solution to this special case can be found by simplifications \emph{mutatis mutandis}.
\noindent \textbf{General case: Source with multiple antennas.} Now, the precoding matrix at the source comprising $\nu$ antennas is $\mathbf{M}\in \mathbb{C}^{\nu \times 2}$. Matrix $\mathbf{M}$ has dimensions $\nu \times 2$, since the transmitted signals are formed as linear combinations of the source signal and the signal transmitted by the active relay. The received signal at the receiving relay for the $n$-th time slot, $y_R[n]$, can be expressed as
\begin{align}\label{signals}
y_R[n]& = \sum_{i\in \mathcal{A}} h_{S_i R}\big( m_{i1}x_S[n]+m_{i2}x_T[n]\big) +h_{TR} x_T[n] +w_R[n],
\end{align}
where $x_{S_i}[n]$ in Eq.~\eqref{eq:yRgeneral} is written as $m_{i1}x_S[n]+m_{i2}x_T[n]$, and $m_{i1},m_{i2} \in \mathbb{C}$.
Defining
\begin{align*}
\mathbf{h}_S^H &\triangleq \begin{bmatrix}h_{S_1R} & h_{S_1R} & \ldots & h_{S_{\nu}R}\end{bmatrix}, \\
\mathbf{m}_1 &\triangleq \begin{bmatrix}m_{11} & m_{21} & \ldots & m_{{\nu}1}\end{bmatrix}^T, \\
\mathbf{m}_2 &\triangleq \begin{bmatrix}m_{12} & m_{22} & \ldots & m_{{\nu}2}\end{bmatrix}^T,
\end{align*}
Eq.~\eqref{signals} can be written as
\begin{align}\label{signals1}
y_R[n] &= \mathbf{h}_S^H \begin{bmatrix}\mathbf{m}_1 & \mathbf{m}_2\end{bmatrix} \underbrace{\begin{bmatrix}x_S[n] & x_T[n]\end{bmatrix}^T}_{\text{data from the source}}+h_{TR} x_T[n] +w_R[n].
\end{align}
\noindent From Eq.~\eqref{signals1}, it can be easily seen that the source node with $\nu$ antennas applies a linear precoding matrix $\mathbf{M}\triangleq \begin{bmatrix}\mathbf{m}_1 & \mathbf{m}_2\end{bmatrix}$ on the transmitted signals such that the IRI is reduced, while at the same time the signal from the source to the intended relay is increased. Each transmitting node uses a fixed power $P$ and thus power control issues are not taken into account. The SINR at the receiving relay, denoted by $\Gamma_R$, is given by
\begin{align}
\Gamma_R\triangleq \frac{\mathbf{h}_S^H \mathbf{m}_1 \mathbf{m}_1^H \mathbf{h}_S P}{|\mathbf{h}_S^H \mathbf{m}_2 + h_{TR}|^2 P + \sigma^2} .
\end{align}
Hence, the following optimization problem can be formulated:
\begin{subequations}
\label{OPT0}
\begin{align}
\displaystyle \underset{\mathbf{m}_1,\mathbf{m}_2}{\max} & ~ \frac{\mathbf{h}_S^H \mathbf{m}_1 \mathbf{m}_1^H \mathbf{h}_S P}{|\mathbf{h}_S^H \mathbf{m}_2 + h_{TR}|^2 P + \sigma^2} \label{0obj1} \\[0.2cm]
\textrm{s.t.}
& ~ \mathbf{m}_1^H \mathbf{m}_1 + \mathbf{m}_2^H \mathbf{m}_2 \leq 1. \label{0cond2}
\end{align}
\end{subequations}
\iffalse
\begin{remark}
Each transmitting node uses a fixed power $P$ and thus power control issues are not taken into account. Note that each antenna $i$, $i \in \{ 1, 2, \ldots, \nu\}$, of the source transmits with power $\mathbb{E}(x_{S_i}x_{S_i}^H)=(|m_{i1}|^2+|m_{i2}|^2)P$. After the precoding, our effective transmitted signal vector is $\mathbf{Mx}$ where $\mathbf{x}=[x_{S_1}, x_{S_2}, \ldots, x_{S_\nu}]^T$. The power of the effective transmitted signal vector is
\begin{align}
\mathrm{tr}(\mathbf{Mx}(\mathbf{Mx})^H) &= \mathrm{tr}(\mathbf{Mx}\mathbf{x}^H\mathbf{M}^H)=\mathrm{tr}(\mathbf{M}(P\mathbf{I})\mathbf{M}^H)\nonumber\\
&=P\mathrm{tr}(\mathbf{M}\mathbf{M}^H)=P,\nonumber
\end{align}
where $\mathbf{xx}^H=P\mathbf{I}$ since $x_i$'s are independent of each other. Hence, $\trace(\mathbf{M}\mathbf{M}^H)\leq 1$.
\end{remark}
\fi
\begin{proposition}\label{propA}
The precoding matrix $\mathbf{M}$ that solves optimization problem~\eqref{OPT0} is given by
\begin{align}
\mathbf{M} = & \frac{1}{\| \mathbf{h}_S \|^2} \begin{bmatrix}
\sqrt{\| \mathbf{h}_S \|^2 - \omega^2 | h_{TR} |^2} \mathbf{h}_S, &
-\omega h_{TR} \mathbf{h}_S \end{bmatrix} \notag \\
= & \frac{\mathbf{h}_S}{\| \mathbf{h}_S \|^2} \begin{bmatrix}
\sqrt{\| \mathbf{h}_S \|^2 - \omega^2 | h_{TR} |^2}, & -\omega h_{TR} \end{bmatrix} ,\label{M:sol}
\end{align}
where
\vspace{-0.4cm}
{\small
\begin{align}\label{omega:sol}
\omega = \left( \frac{1}{2} + \frac{\| \mathbf{h}_S \|^2 + \rho}{2| h_{TR} |^2} \right) - \sqrt{\left( \frac{1}{2} + \frac{\| \mathbf{h}_S \|^2 + \rho}{2| h_{TR} |^2} \right)^2 - \frac{\| \mathbf{h}_S \|^2}{| h_{TR} |^2} }
\end{align}}
and $\rho=\sigma^2/P$.
\end{proposition}
\begin{proof}
See appendix~\ref{A:proof_propA}.
\end{proof}
The last expression in \eqref{M:sol} shows that the optimal $M$
is a combination of standard maximum ratio transmission across the
source antennas, and the single antenna interference suppression solution.
Proposition~\ref{propA} shows that the source does not need to cancel the interference completely in order to achieve the maximum SINR at the receiving relay. This result appears counter-intuitive, since interference is the main source of performance degradation and the SINR should be higher when the IRI is cancelled. However, this occurs due to the limited power at the source. When the power $P$ is high enough, $\rho$ gets very small. Hence, from \eqref{omega:sol},
\begin{align*}
\lim_{\rho\rightarrow 0} \omega &= \left( \frac{1}{2} + \frac{\| \mathbf{h}_S \|^2}{2| h_{TR} |^2} \right) - \sqrt{\left( \frac{1}{2} + \frac{\| \mathbf{h}_S \|^2}{2| h_{TR} |^2} \right)^2 - \frac{\| \mathbf{h}_S \|^2}{| h_{TR} |^2} } \\
& =
\begin{cases}
1, &\text{if }\| \mathbf{h}_S \|^2 \geq | h_{TR} |^2, \\[0.2cm]
\displaystyle \frac{\| \mathbf{h}_S \|^2}{| h_{TR} |^2}, &\text{if }\| \mathbf{h}_S \|^2 < | h_{TR} |^2.
\end{cases}
\end{align*}
which means that interference is essentially cancelled if $\|
\mathbf{h}_S \|^2 \geq | h_{TR} |^2$.
If the source can change its power, then the power levels of the source and the transmitting relay are not necessarily the same. In what follows, we allow the source to adjust its power. The SINR now becomes
\begin{align}
\displaystyle \frac{\mathbf{h}_S^H \mathbf{m}_1 \mathbf{m}_1^H \mathbf{h}_S P_S}{|\sqrt{P_S}\mathbf{h}_S^H \mathbf{m}_2 + \sqrt{P_T}h_{TR}|^2 + \sigma^2}. \label{eq:SINR1}
\end{align}
The following proposition gives the optimal
joint selection of $\omega$ and $P_S$, that maximizes the SINR, as given in \eqref{eq:SINR1}.
\begin{proposition}\label{propB}
When $P$ and $\mathbf{M}$ are considered jointly, then the optimal power $P_S^*$ and $\omega^*$ are given by
\begin{subequations}
\begin{align}
P_S^*&=P_{\max}, \\
\omega^* &= \min\left\{1,{\frac{\|\mathbf{h}_S\|}{| h_{TR} |}}, \omega_1, \omega_2 \right\} ,
\end{align}
where
\begin{align}
\omega_1 &= \frac{P_T| h_{TR} |^2 + \sigma^2}{ \sqrt{P_T P_{\max}}| h_{TR} |^2},\\
\omega_2 &=\frac{P_{\max} \|\mathbf{h}_S\|^2 + P_T| h_{TR} |^2+\sigma^2}{2 \sqrt{P_T P_{\max}|} h_{TR} |^2 } - \nonumber \\
& \hspace{-0cm} \frac{\sqrt{(P_{\max} \|\mathbf{h}_S\|^2 + P_T| h_{TR} |^2+\sigma^2)^2 - 4P_T P_{\max}\|\mathbf{h}_S\|^2 | h_{TR} |^2 }}{2 \sqrt{P_T P_{\max}|} h_{TR} |^2 }.
\end{align}
\end{subequations}
\end{proposition}
\begin{proof}
See appendix~\ref{B:proof_propB}.
\end{proof}
Proposition~\ref{propB} suggests
that in the case which $P_S$ and $\omega$ are jointly optimized, in order to maximize the SINR, the source transmits with $P_{\max}$ and chooses the $\omega$ corresponding to $P_{\max}$.
\iffalse
\begin{remark}\label{remark2}
Another approach would be to fix the transmission rate as $\log_2(1+\gamma)$ and try to minimize the outage probability of the $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link. Then, the optimization problem becomes
\begin{subequations}
\label{OPT_1}
\begin{align}
\displaystyle \underset{\mathbf{m}_1,\mathbf{m}_2}{\min} & ~\mathbb{P} \left( \frac{\mathbf{h}_S^H \mathbf{m}_1 \mathbf{m}_1^H \mathbf{h}_S P}{(\mathbf{h}_S^H \mathbf{m}_2 + h_{TR})(\mathbf{h}_S^H \mathbf{m}_2 + h_{TR})^H P + \sigma^2}<\gamma \right) \label{0obj1} \\[0.2cm]
\textrm{s.t.} & ~
\mathbf{h}_S^H \begin{pmatrix}
\mathbf{m}_1 & \mathbf{m}_2 \\
\end{pmatrix} =
\begin{pmatrix}
\beta & -\omega h_{TR} \\
\end{pmatrix}, \label{0cond1} \\
& ~ \mathbf{m}_1^H \mathbf{m}_1 + \mathbf{m}_2^H \mathbf{m}_2 \leq 1, \label{0cond2} \\
& ~ \omega \in (0,1] . \label{0cond3}
\end{align}
\end{subequations}
It can be easily shown that this is equivalent to maximizing the SINR which leads to maximizing the link rate.
\end{remark}
\fi
\subsection{Adaptive Rate Relay Pair Selection Policy}\label{sec:AR_RS_policy}
Since CSIs of $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links are assumed to be available at the source and relays, adaptive rate transmission, which implies transmission rate is determined by instantaneous CSI at each link, is possible. In this case, the source encodes its own desired codewords based on CSI of the selected $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link and codewords corresponding to the transmitting relay based on CSI of the selected $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ link. Then, the source combines them via precoding. In the adaptive rate transmission, the main objective is to maximize the average end-to-end rate from the source to the destination, which is equivalently expressed as the average received data rate at the destination \cite{KIM2}, i.e.,
\begin{align}
\bar{C} = \underset{{W\to\infty}}{\lim} \frac{1}{W} \sum_{t=1}^{W}C_{T_t^* D}(t),
\label{eq:bar_C}
\end{align}
such that
\begin{align*}
\sum_{t=1}^{W}C_{T_t^* D}(t) \leq \sum_{t=1}^{W}C_{SR_t^*}(t),
\end{align*}
where $W$ denotes a time window length and $C_{SR_t^*}(t)$ and $C_{T_t^*D}(t)$ denote the resulting instantaneous link rates depending on time slot $t$ once the relay pair $(R_t^*,T_t^*)$ is selected. The instantaneous rates are obtained by
\begin{align}
C_{S R}(t) &= \min\left\{ \log_2(1+\gamma_{S R}(t)), Q_{\max} - Q_{R}(t-1) \right\},\\
C_{T D}(t) &= \min\left\{ \log_2(1+\gamma_{T D}(t)), Q_{T}(t-1) \right\},
\end{align}
where $\gamma_{S R}(t)$ and $\gamma_{T D}(t)$ denote the effective SINR/SNR (signal-to-noise-ratio) of $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{T\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links at time slot $t$ after applying the precoding matrix obtained by Proposition~\ref{propA}, i.e.,
\begin{align}
\gamma_{S R}(t) &= \frac{\left(\| \mathbf{h}_S \|^2 - \omega^2 | h_{TR} |^2\right)P}{(1-\omega^2)| h_{TR} |^2 P+\sigma^2} , \\[0.12cm]
\gamma_{T D}(t) &= \frac{|h_{TD}|^2 P}{\sigma^2} ,
\end{align}
where $\omega$ is given in \eqref{omega:sol},
$Q_{\max}$ denotes the maximum length of queue, and $Q_{R}(t-1)$ and $Q_{T}(t-1)$ are the queue lengths of receiving and transmitting relays in BPCU at time slot $(t-1)$. For the selected relay pair $(R^*, T^*)$, the queue lengths are updated at the end of each time slot as
\begin{align}
Q_{R^*}(t) = Q_{R^*}(t-1) + C_{S R^*}(t),\\
Q_{T^*}(t) = Q_{T^*}(t-1) - C_{S T^*}(t).
\end{align}
Hereafter, based on the designed precoding matrix $\mathbf{M}$ in Section~\ref{sec:policy1}, we propose an adaptive rate relay selection policy for maximizing the average end-to-end rate. Applying a Lagrangian relaxation, the objective function to maximize the average end-to-end rate in \eqref{eq:bar_C} for the adaptive rate transmission becomes equivalent to a weighted sum of instantaneous link rates \cite{KIM,KIM2}.
The resulting relay selection scheme has the form
\begin{subequations}
\label{eq:RS_AR}
\begin{align}
\displaystyle \underset{(R, T)}{\max} &~ \delta_R C_{S R}(t) + (1-\delta_T) C_{T D}(t),\\
\textrm{s.t.} &~ R \neq T,~R,T \in \mathcal{K},
\end{align}
\end{subequations}
for predetermined weight factors\footnote{Under i.i.d. channel conditions, the weight factors are approximately identical for all relays and can be reduced as a single weight factor.} $\delta_k\in[0,1],~k\in\mathcal{K}$,
where $C_{S R}(t)$ and $C_{T D}(t)$ denote the instantaneous rates of $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{T\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links at time slot $t$, respectively.
It is worth noting that for finite buffer size, the optimal weight factors can be found using either a subgradient method or a back-pressure algorithm \cite{KIM2}. According to the back-pressure algorithm, the weight factors can be found as $\delta_k = 1 - Q_k(t)/Q_S(t)$ during a training period where $Q_k(t)$ and $Q_S(t)$ denote the buffer occupancies at time $t$ at the $k$-th relay and the source, respectively. The details can be found in \cite{KIM2}.
\begin{remark}
The optimal relay pair selection in \eqref{eq:RS_AR} requires an exhaustive search with $K\times(K-1)$ combinations. Therefore, the computational complexity of optimal relay pair selection is $\mathcal{O}(K^2)$.
\end{remark}
\begin{remark}
The implementation of the proposed scheme requires global CSI of the instantaneous channels and the buffer states. The central unit (\emph{e.g.}, the source in this case) uses this information in order to select the appropriate relay pair. This global CSI requires a continuous feedback for each wireless link as we use a continuous monitoring of the ACK/NACK signaling in order to identify the status of the buffers. Although these implementation issues are beyond the scope of this paper, the proposed scheme can be implementable with the aid of various centralized and distributed CSI acquisition techniques and relay selection approaches as in \cite{BLE} where each relay sets a timer in accordance to the channel quality and through a countdown process the selection of the best relay is performed. Additionally, CSI overhead can be reduced significantly, through distributed-switch-and-stay-combining as in \cite{DSSC1, NOM_CAMAD, DSSC2}.
\end{remark}
\section{Buffer-Aided Relay Selection Based on Buffer-Aided Phase Alignment}\label{sec:policy2}
In this section, we relax the assumption of having knowledge of the full CSI. Instead, we allow for CSIR knowledge with limited feedback, i.e., each receiving node has CSI of the channel it receives data from and it can provide some information to the transmitting node. More specifically, each (receiving) relay feeds back to the source a phase value via a reliable communication link.
This approach is closely related to phase feedback schemes proposed and standardized for MISO transmission (see, for example, \cite{P1, P2, P3, P4}).
The source signal can use this phase value in one of two possible ways:
\begin{itemize}
\item[(a)] to mitigate the interfering signal, so that the overall interference is reduced or even eliminated;
\item[(b)] to amplify the interfering signal, so that it can be decoded and removed from the rest of the received signals.
\end{itemize}
The phase value can be quantized into the desired number of bits, using uniform quantization.
We aim at recovering the multiplexing loss of the network by having the source and a relay to align phases, such that the IRI is reduced. As a result, the proposed phase alignment approach reduces the communication overhead, while introducing a smart quantization that further reduces the complexity of the transmission.
Since there is no full CSI knowledge, as assumed in the scheme proposed in the previous section, no precoding can be applied and a relay-selection policy is proposed based on CSIR knowledge with limited feedback that aims at taking advantage of the multi-antenna source. Selection of the relay-pair is performed by choosing the pair that achieves the maximum end-to-end SINR.
\subsection{Buffer-Aided Phase Alignment (BA-PA)}
Since the source has no CSI for any of the links between the antennas at the source and the relays, equal power allocation across all antennas is a natural approach at the transmitter side. However, for overhead reduction on CSI estimation at the receiver, we consider to use just two of them:
one for transmitting the packet to a relay and the other to mitigate the IRI. The existence of more than two antennas, however, can increase the diversity gain by choosing a subset of antennas based on CSI, while the fact that not all available antennas are included, the overhead for channel information is limited. The same principle can equally well be applied using a single antenna
source node, where one and the same antenna is used both for the
desired packet and the IRI mitigation.
Assuming two antennas used, the receiving signal at $R$ in ~\eqref{eq:yRgeneral} is given by
\begin{align}\label{eq:yR2}
y_{R}[n]&= h_{S_1 R} m_1 x_{S_1}[n] + h_{S_2 R} m_2 x_{S_2}[n] + h_{TR} x[p] +w_R[n] ,
\end{align}
where we set $m_1=1/\sqrt{2}$ and $m_2=e^{j\phi}/\sqrt{2}$.
In each time slot, the signal from the second antenna of the source $x_{S_2} [n]$ is used in one of the following two ways:
\begin{itemize}
\item[{(a)}] to minimize the interference caused by the transmitting relay; this is done by transmitting $x[p]$ (i.e., $x_{S_2} [n]=x[p]$) with a shifted phase such that the interfering signal and the signal from the second antenna is in anti-phase. The optimal phase for this is given by
\begin{align}\label{opt:1_1}
\phi^\star=\arg \min_{\phi} \norm{\frac{h_{S_2 R}}{\sqrt{2}}e^{j\phi} +h_{TR}}^2 ;
\end{align}
\item[{(b)}] to maximize the interference caused by the transmitting relay in order to make the signal strong enough to be decoded first, and hence, eliminate it. The optimal phase for this is given by
\begin{align}\label{opt:1_2}
\phi^\dagger=\arg \max_{\phi} \norm{\frac{h_{S_2 R}}{\sqrt{2}}e^{j\phi} +h_{TR}}^2 .
\end{align}
\end{itemize}
\begin{proposition}\label{propC}
The phase $\phi^\star$ such that the interfering signal at the receiving relay $R_r$ from the transmitted signal $x[p]$ of the transmitting relay $R_t$ is minimized is given by
\begin{align}
e^{j\phi^\star} = -\frac{h_{S_2R}^H h_{TR}}{|h_{S_2 R}^H h_{TR}|} .
\end{align}
Similarly, the phase $\phi^\dagger$ such that the interfering signal at the receiving relay $R_r$ from the transmitted signal $x[p]$ of the transmitting relay $R_t$ is maximized is given by
\begin{align}
e^{j\phi^\dagger} = \frac{h_{S_2R}^H h_{TR}}{|h_{S_2R}^H h_{TR}|} .
\end{align}
\end{proposition}
\begin{proof}
See appendix~\ref{C:proof_propC}.
\end{proof}
Proposition~\ref{propC} gives the expressions for phase alignment for
each of the two approaches considered. By appropriately choosing the
phase shift $\phi$ of the signal from one of the source's antennas,
the source can minimize or maximize the interfering signal in order to
mitigate it or eliminate it completely. Note that the optimal value of
$\phi$ (either $\phi^\star$ or $\phi^\dagger$) can be quantized into
the desired number of bits, using uniform quantization, and be fed
back to the source.
\subsection{Fixed Rate Relay Pair Selection Policy}\label{sec:FR_RS_policy}
Since only CSIR is available, the transmitters use fixed rate transmission. In such a case, the main objective is to minimize the outage probability. Independently of nodal distribution and traffic pattern, a transmission from a transmitter to its corresponding receiver is successful (error-free) if the SINR of the receiver is above or equal to a certain threshold, called the \textit{capture ratio}\footnote{It depends on the modulation and coding schemes as error-correction coding techniques supported by a wireless communication system and corresponds to the required SINR to guarantee the data rate of the application.} $\gamma_0$. Therefore, at the receiving relay $R$ for a successful reception when at the same time relay $T$ is transmitting we require that
\begin{align}\label{SINR_geq1}
\Gamma_{R}^{T}\triangleq \frac{|h_{S_1 R}|^2 P/{2}}{\left|\frac{h_{S_2 R}}{\sqrt{2}}e^{j\phi}+h_{TR}\right|^2 P + \sigma^2} \geq \gamma_0,
\end{align}
and at the destination we require that
\begin{align}\label{SNR_geq1}
\Gamma_{D}^T\triangleq \frac{|h_{TD}|^2 P}{\sigma^2} \geq \gamma_0.
\end{align}
An outage event occurs at the relay $R$ and destination $D$ when $\Gamma_{R}^S < \gamma_0$ and $\Gamma_{D}^T < \gamma_0$, respectively. The outage probability is denoted by $\mathbb{P}(\Gamma_{i}^k < \gamma_0)$, where $i$ represents the receiving node and $k$ the transmitting node.
Each relay $i$ is able to estimate the SINR for each transmitting relay $k$, denoted by $\Gamma_{i}^{k}$, $k\neq i$ (the full pilot protocol needed to the channel estimation is out of the scope of this work). We assume that this information can be communicated to the destination. In addition, the destination node can compute its own SNR due to each of the transmitting relays, denoted by $\Gamma_{D}^{k}$, $k\in\{1,\ldots, K\}$. Finally, we assume that the destination node has buffer state information\footnote{The destination can know the status of the relay buffers by monitoring the ACK/NACK signaling and the identity of the transmitting/receiving relay.} and selects the relays for transmission and reception, based on some performance criterion, \emph{e.g.}, with the maximum end-to-end SINR (as it is defined in \cite{IKH3}), through an error-free feedback channel. Note that by having the destination to take the decision, no global CSI is required at any node.
As we have seen in Proposition~\ref{propC}, the source can minimize the interfering signal or maximize it in order to eliminate it by appropriately choosing the phase shift $\phi$ of the signal from one of the source's antennas. It can be easily deduced that at low IRI it is beneficial to try to remove the interfering signal, whereas at high IRI it is beneficial to amplify the interfering signal and thus eliminate it completely by decoding it first. The receiving relay is able to compute which option gives the highest SINR in each case, since it has knowledge of the channel states and hence, it can decide which phase to feed back to the source at each time slot.
The procedure of the proposed algorithm is as follows: By examining \emph{one-by-one} the possible relay pairs, first we calculate the power of the signal received at {\it D} which is $P_{D}=|h_{TD}|^2 P+\sigma^2$ for an arbitrary relay $T$ with non-empty buffer. The receiving relay $R$ must be different than the transmitting relay and its buffer should not be full. For each candidate relay $i$ for reception, a \emph{feasibility check} for interference cancellation (IC) is performed, i.e.,
$$
\Gamma_{i}^{k} = \frac{\left| h_{R_k R_i}\right|^2 P}{|h_{S_1 R_i}+h_{S_2 R_i}e^{j\phi}|^2 P/{2} + \sigma^2}\geq \gamma_0 .
$$
If IC is feasible, the candidate relay is examined whether SNR at the receiving relay after IC is above the \textit{capture ratio} $\gamma_0$ or not, i.e., once interference is removed \eqref{SINR_geq1} becomes
$$
\Gamma_{R}^{S}\triangleq \frac{|h_{S_1 R}|^2 P/{2}}{\sigma^2} \geq \gamma_0.
$$
If IC is infeasible, interference mitigation (IM) is considered. Hence, it is examined whether SINR at the receiving relay after IM is above the \textit{capture ratio} or not. If the relay denoted by $R$ can provide an SNR/SINR above the capture ratio after IC/IM, it is considered as a candidate receiving relay.
For the selected relay pair $(R^*, T^*)$, the queue lengths are updated at the end of time slot as
\begin{subequations}
\begin{align}
&Q_{R^*}(t) = \min \{Q_{R^*}(t-1) + C_0, Q_{\max}\}, \label{Q_fixed_1} \\
&Q_{T^*}(t) = \max\{Q_{T^*}(t-1) - C_0, 0\}, \label{Q_fixed_2}
\end{align}
\end{subequations}
Note that for fixed rate transmission, the queue length is equivalently modeled as the number of packets in the queue.
\begin{remark}
The relay pair selection policy, as before, requires an exhaustive search with $K\times(K-1)$ combinations imposing a complexity of the order $\mathcal{O}(K^2)$. Note, however, that links with SINR/SNR above the \textit{capture ratio} $\gamma_0$ are suitable for transmission without compromising the performance of the proposed scheme, i.e., it is not necessary to choose the relay pair that provides the maximum end-to-end SINR, as long as the outage event is avoided. Hence, its simplicity allows the selection of a relay pair much faster, provided that the channel conditions are good. This characteristic will allow for more advanced algorithms in the future where the queue length will be a decisive factor on the decision of the relay pair such that certain delay constraints are met. For the time-being a relay pair is selected such that there is no reduction in performance (either due to outage or full/empty buffers).
\end{remark}
\section{Numerical Results}\label{sec:numerical}
We have developed a simulation setup based on the system model description in Section~\ref{sec:model}, in order to evaluate the performance of the proposed relay pair selection schemes with current state-of-the-art for both adaptive and fixed rate transmission cases according to CSI availability at the source.
In the simulations, we assume that the clustered relay configuration ensures $\mathop{\mathrm{i.i.d.}}$ Rayleigh block fading with average channel qualities $\sigma_{SR}^2$, $\sigma_{RR}^2$, and $\sigma_{RD}^2$ for all the $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$, $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$, and $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links, respectively. This assumption simplifies the interpretation of the results and it is used in several studies in the literature \cite{BEL,KRI2,BEN, BEN2}.
The $\mathop{\mathrm{i.i.d.}}$ case and the related analysis can be considered as a useful guideline for more complex asymmetric configurations, in which power control at the source and relay nodes may be used to achieve symmetric channel configuration for better outage performance \cite{ML-AF}.
\subsection{Adaptive Rate Transmission}
For adaptive rate transmission, we evaluate the performance of the proposed BA-SP relay selection (BA-SPRS) scheme in terms of the average end-to-end achievable rate in BPCU.
The following upper bound and state-of-the-art relay selection schemes are considered for the performance comparison.
\begin{itemize}
\item Upper bound: the optimal solution of \eqref{eq:RS_AR} under no IRI
\item HD best relay selection (HD-BRS)~\cite{BLE,KAR}
\item HD hybrid relay selection (HD-HRS)~\cite{IKH2}
\item HD $\mathop{\mathrm{max-link}}$ relay selection (HD-MLRS)~\cite{KRICHAR}
\item ideal (IRI interference is assumed negligible) and non-ideal (IRI interference is taken into consideration) SFD-MMRS~\cite{IKH3}
\end{itemize}
Since in this work we consider multiple antennas only at the source while all conventional schemes above consider a single antenna at the source, we employ maximum ratio transmission (MRT) at the source, which maximizes the effective channel gain at each receiving relay, for all conventional schemes as well for making the comparison fair.
Fig.~\ref{fig:art_rate_snr} shows the average end-to-end rate for varying SNR with two relays and two antennas at the source when $Q_{\mathrm{max}}\to\infty$, $\sigma_{SR}^2=\sigma_{RD}^2=0$ dB with various IRI intensities: the same intensity ($\sigma_{RD}^2=0$ dB), stronger IRI ($\sigma_{RD}^2=3$ dB) and weaker IRI ($\sigma_{RD}^2=-3$ dB).
While the ideal SFD-MMRS scheme almost achieves the upper bound, non-ideal SFD-MMRS significantly degrades the performance since the effect of IRI is severe so that it cannot be neglected. The discrepancy between the proposed BA-SPRS scheme and the upper bound increases with SNR, since it results in increasing IRI. Especially, the gap becomes larger for stronger IRI condition ($\sigma_{RD}^2=3$ dB). This is because perfect IC is impossible as IRI increases. However, when IRI becomes at least $3$ dB below the $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ channel qualities, the proposed scheme almost approaches the upper bound.
The conventional HD schemes achieve almost half the average end-to-end rate of the ideal SFD-MMRS and the proposed BA-SPRS schemes.
Although the HD-MLRS scheme is a more advanced scheme than HD-BRS and HD-HRS schemes in the sense of diversity order, it is the worst for this case with infinite buffer length, since our system setup yields a channel imbalance between the $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links due to multiple antennas at only the source, which gives a higher chance to select the $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link than the $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ link for the HD-MLRS scheme. In other words, for the HD-MLRS scheme, the larger the buffer size, the more source data is buffered at relay buffers, while the throughput is determined by the amount of received data at the destination.
From the results, it is shown that the proposed BA-SPRS scheme always outperforms the non-ideal SFD-MMRS and HD relay selection schemes regardless of SNR and IRI intensity.
\begin{figure}[t]
\centering
\includegraphics[width=0.67\columnwidth]{fig2_ART_rate_SNR-eps-converted-to}
\caption{Average end-to-end rate for varying SNR when $K=2$, $Q_{\mathrm{max}}\to\infty$, $\sigma_{SR}^2=\sigma_{RD}^2=0$ dB, and $\sigma_{RR}^2=-3, 0$, or $3$ dB.}
\label{fig:art_rate_snr}
\end{figure}
Fig.~\ref{fig:art_rate_K} shows the effect of the number of relays. As the number of relays increases, the proposed BA-SPRS scheme quickly approaches the upper bound.
Especially, when $K=3$ and $\sigma_{RR}^2 = 0$ dB, it already almost achieves the upper bound while other schemes can never achieve the upper bound except for the ideal SFD-MMRS scheme neglecting IRI. Therefore, in the case when $K=2$, an additional relay (i.e., making $K=3$) contributes significantly to the performance enhancement with the same IRI intensity.
For stronger IRI ($\sigma_{RR}^2=3$ dB), although there exist some gaps from the upper bound for a small number of relays, it is fast recovered with increasing number of relays even under strong IRI condition and finally converges to the upper bound. Therefore, increasing the number of relays, offering spatial diversity and ignoring hardware and deployment costs, can be a simple solution against strong IRI situations.
For the same reason stated in Fig.~\ref{fig:art_rate_snr}, the conventional HD-MLRS scheme rather degrades the average end-to-end rate with respect to increasing number of relays, since increasing the number of relays causes a similar effect to that of increasing the buffer size.
\begin{figure}[t]
\centering
\includegraphics[width=0.67\columnwidth]{fig3_ART_rate_K-eps-converted-to}
\caption{Average end-to-end rate with increasing the number of relays when $\mathrm{SNR}=20$ dB, $\sigma_{SR}^2=\sigma_{RD}^2=0$ dB, and $\sigma_{RR}^2 = 0$ or $3$ dB.}
\label{fig:art_rate_K}
\end{figure}
Fig.~\ref{fig:art_rate_Qmax} shows the average end-to-end rate for different maximum buffer sizes for $K=3$ and $\mathrm{SNR}=20$ dB, which limits the instantaneous rate for each link. The ideal SFD-MMRS and the proposed BA-SPRS schemes converge to the upper bound as the maximum buffer size increases but they have a cross point at $Q_{\max}\approx25$. The proposed BA-SPRS scheme achieves rather slightly higher average rate when $25<Q_{\max}<5000$ than the ideal SFD-MMRS scheme neglecting IRI. As a result, the proposed BA-SPRS scheme is still effective for the finite buffer size which is a more practical setup. All the schemes converge to their own maximum rate when approximately $Q_{\max}>50$ except for the HD-MLRS scheme. As shown in Figs.~\ref{fig:art_rate_snr} and \ref{fig:art_rate_K}, the HD-MLRS scheme rather degrades the average end-to-end rate as the buffer size increases due to the effect of channel imbalance between $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links. The HD-BRS scheme without using a buffer is not affected from the finite buffer size.
\begin{figure}[t]
\centering
\includegraphics[width=0.67\columnwidth]{fig4_ART_rate_finite_Qmax_logx-eps-converted-to}
\caption{Average end-to-end rate for varying the maximum buffer size $Q_{\mathrm{max}}$ when $K=3$, $\mathrm{SNR}=20$ dB, and $\sigma_{SR}^2=\sigma_{RD}^2=\sigma_{RR}^2 =0$ dB.}
\label{fig:art_rate_Qmax}
\end{figure}
\subsection{Fixed Rate Transmission}
For fixed rate transmission, we evaluate the proposed BA-PA relay selection (BA-PARS) scheme in terms of outage probability and average end-to-end rate. We additionally consider the following state-of-the-art scheme proposed for fixed rate transmission:
\begin{itemize}
\item Buffer-aided successive opportunistic relaying (BA-SOR) with IRI cancellation \cite{NOM2}.
\end{itemize}
For all simulation results with fixed rate transmission, we consider $\mathop{\mathrm{i.i.d.}}$ channel condition with the same IRI intensity, i.e., $\sigma_{SR}^2=\sigma_{RR}^2=\sigma_{RD}^2=0$.
Fig.~\ref{fig:pout_all} shows the outage probability\footnote{While an outage is defined in \cite{IKH3} when a minimum of channel gains of both $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links is less than the capture ratio $\gamma_0$, we define the outage probability as a portion of successfully transmitted packets among the total number of transmitted packets since the previous definition is not rigorous for the case of concurrent transmissions with IRI.} with various SNR values for the transmission rate $C_0=1$ BPCU, three relays ($K=3$), and infinite length of buffer ($Q_{\mathrm{max}}\rightarrow\infty$).
Basically, the HD-BRS scheme has the worst outage performance due to lack of buffering. The HD-HRS scheme, a hybrid mode of HD-BRS and HD-MMRS, is always better than the HD-BRS scheme. Since the HD-MLRS scheme can achieve a full diversity (i.e., $2K$ diversity order) for HD transmission, it shows the best performance except for low SNR region. At low SNR, the ideal SFD-MMRS scheme without taking IRI consideration achieves slightly better performance than the HD-MLRS scheme. However, its outage performance is significantly degraded if IRI is imposed.
The BA-SOR scheme achieves a good performance at low and medium SNR but it becomes bad at high SNR since its relay selection criterion is to maximize the minimum SNR of both $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ and $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links\footnote{In our outage definition, the BA-SOR scheme can be improved at high SNR if it employs to select the best $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ link when the best SNR for $\mathop{\{S\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} R\}}$ link is worse than SNRs of all $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ links, instead of a random selection, since the $\mathop{\{R\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} D\}}$ link separately contributes on the outage event.}.
The proposed BA-PARS scheme achieves similar performance to the BA-SOR scheme at low SNR but it is not degraded at high SNR thanks to a hybrid mode of IC and IM. In addition, assuming a powerful source node such as base station as we stated in Section~II, we depict the case of double power at the source for the proposed BA-PARS scheme, which shows that the proposed BA-PARS scheme can achieve the outage performance of the ideal SFD-MMRS scheme. Hence, if extra power at the source is available, the proposed BA-PARS scheme can provide the best outage performance.
Note that the ideal SFD-MMRS and the double powered BA-PARS schemes achieve a half diversity gain (i.e., $K$ diversity order) of the HD-MLRS scheme but a better power gain at low SNR, since for both schemes, a half rate is required at each link to meet the same transmission rate as the HD-MLRS scheme.
\begin{figure}[t]
\centering
\includegraphics[width=0.67\columnwidth]{fig5_ICC17_fig2_final-eps-converted-to}
\caption{Outage probability for $C_0=1$ BPCU, $K=3$, and $Q_{\mathrm{max}}\rightarrow\infty$. For the proposed BA-PARS scheme, we additionally consider double power at the source (denoted by `[2P]') to show the case of a powerful source node.}
\label{fig:pout_all}
\end{figure}
Fig.~\ref{fig:pout_var_buffer} shows the outage probability of the proposed BA-PARS scheme for varying the maximum buffer size when $C_0=1$ BPCU and $K=3$. As in \cite{IKH3,NOM2}, we assume that half of buffer elements are full at initial phase (in order to reach the steady-state queue lengths quicker). As the maximum buffer size $Q_{\max}$ increases, the outage performance is improved and converges to the case of having buffers of infinite length. The convergence occurs at lower buffer sizes at high SNR than at low SNR, since buffer full/empty events contribute more in outage events at high SNR due to sufficiently good received signal strengths (i.e., outage events due to bad channel conditions occur rarely and outage events are due to buffer full/empty events).
\begin{figure}[t]
\centering
\includegraphics[width=0.67\columnwidth]{fig6_pout_BA-PA_varL-eps-converted-to}
\caption{Outage probability of the proposed BA-PARS scheme for varying the maximum buffer size $Q_{\mathrm{max}}$ when $C_0=1$ BPCU and $K=3$.}
\label{fig:pout_var_buffer}
\end{figure}
Fig.~\ref{fig:rate_all_frt} shows the average end-to-end achievable rate with three different transmission data rates ($C_0=1.5$ and $C_0=2.5$ BPCU) when $K=3$ and $Q_{\mathrm{max}}=10$.
The conventional HD schemes achieve a half of the data rate due to an HD limitation although the HD-MLRS scheme achieves a full diversity in outage performance.
While the ideal SFD-MMRS scheme obtains the best performance which achieves full data rates (1.5 and 2.5 BPCU) as SNR increases, the non-ideal SFD-MMRS scheme is significantly degraded due to IRI, which shows rather worse performance with $C_0 = 2.5$ BPCU than the HD schemes. The BA-SOR scheme can achieve the full data rate with $C_0 = 1.5$ BPCU at high SNR but significantly degrades with higher data rates.
In contrast, the proposed BA-PARS scheme can achieve the full data rates for all the cases such that it guarantees the required data rate if a proper data rate is chosen according to SNR, even if it has some gaps compared to the ideal SFD-MMRS scheme since the source power is split for IC/IM.
Similarly to the outage performance, if double power at the source is available, the proposed BA-PARS scheme can approach the achievable rate of the ideal SFD-MMRS scheme without suffering from IRI.
\begin{figure}[t]
\centering
\includegraphics[width=0.67\columnwidth]{fig7_ICC17_fig4_final-eps-converted-to}
\caption{Average end-to-end rate with two different fixed data rates ($C_0=1.5$ and $C_0=2.5$ BPCU) when $K=3$ and $Q_{\mathrm{max}}=10$.}
\label{fig:rate_all_frt}
\end{figure}
\section{Conclusions and Future Directions}\label{sec:conclusions}
\subsection{Conclusions}
In this work, we present two relay-pair selection policies, depending on the available CSI in the system, that employ a buffer-aided multi-antenna source, a cluster of HD buffer-aided relays and a destination. In the case of global CSI, a linear precoding strategy is applied by the source in order to mitigate IRI. A relay pair is selected, such that the average end-to-end rate is maximized. In the case of CSIR, phase alignment is applied by the source in order to mitigate/cancel IRI. A relay pair is selected, such that the maximum end-to-end SINR is achieved. The benefits of this network deployment are demonstrated via a numerical evaluation, where improved performance is observed with respect to the average end-to-end rate and outage probability, while the conventional non-ideal SFD-MMRS scheme with IRI is significantly degraded.
\subsection{Future directions}
Part of ongoing research is to investigate scenarios where only statistical information is known about the CSI.
It is clear that the two different channel state information cases considered (CSI and CSIR) result in different amounts of overhead. Part of ongoing work is to (a) demonstrate the impacts of this overhead on the proposed policies, and (b) when each case should be considered, given the channel coherence time and required throughput.
\appendices
\section{Proof of Proposition~\ref{propA}}\label{A:proof_propA}
After algebraic manipulations, optimization problem~\eqref{OPT0} can be written as
\begin{subequations}
\label{OPT1}
\begin{align}
\displaystyle \underset{\mathbf{m}_1,\mathbf{m}_2,\beta,\omega}{\max} & ~ \frac{|\beta|^2}{(1-\omega^2)|h_{TR}|^2 + \rho} \label{1obj1} \\[0.2cm]
\textrm{s.t.} & ~
\mathbf{h}_S^H \begin{bmatrix}
\mathbf{m}_1 & \mathbf{m}_2 \\
\end{bmatrix} =
\begin{bmatrix}
\beta & -\omega h_{TR} \\
\end{bmatrix}, \label{1cond1} \\
& ~ \mathbf{m}_1^H \mathbf{m}_1 + \mathbf{m}_2^H \mathbf{m}_2 \leq 1, \label{1cond2}\\
& ~ \omega \in (0,1] . \label{1cond3}
\end{align}
\end{subequations}
where $\rho=\sigma^2/P$, and $\beta$, $\omega$ depend on the choice of $\mathbf{m}_1,\mathbf{m}_2$.
\begin{remark}
By a simple phase shift we observe that we can choose $m_{11}$ and $m_{22}$, such that $\beta$ is a real number; i.e., let $\theta \in [0,2\pi]$, such that $\beta \triangleq \beta' e^{j \theta}$; then, $|\beta|^2 = |\beta' e^{j \theta}|^2= |\beta'|^2$.
\end{remark}
Hence, optimization problem~\eqref{OPT1} can be written as
\begin{subequations}
\label{OPT2}
\begin{align}
\displaystyle \underset{\mathbf{m}_1,\mathbf{m}_2,\beta,\omega}{\max} & ~ \frac{\beta^2}{(1-\omega^2)|h_{TR}|^2 + \rho} \label{2obj1} \\[0.2cm]
\textrm{s.t.} & ~
\mathbf{h}_S^H \begin{bmatrix}
\mathbf{m}_1 & \mathbf{m}_2 \\
\end{bmatrix} =
\begin{bmatrix}
\beta & -\omega h_{TR} \\
\end{bmatrix}, \label{2cond1} \\
& ~ \mathbf{m}_1^H \mathbf{m}_1 + \mathbf{m}_2^H \mathbf{m}_2 \leq 1, \label{2cond2} \\
& ~ \omega \in (0,1] . \label{2cond3}
\end{align}
\end{subequations}
Let $\omega\in (0,1]$ be a fixed value. Then, the problem becomes equivalent to maximizing $\beta^2$, i.e.,
\begin{subequations}
\label{OPT3}
\begin{align}
\displaystyle \underset{\mathbf{m}_1,\mathbf{m}_2,\beta}{\max} & ~ \beta^2 \label{3obj1} \\[0.2cm]
\textrm{s.t.} & ~
\mathbf{h}_S^H \begin{bmatrix}
\mathbf{m}_1 & \mathbf{m}_2 \\
\end{bmatrix} =
\begin{bmatrix}
\beta & -\omega h_{TR} \\
\end{bmatrix}, \label{3cond1} \\
& ~ \mathbf{m}_1^H \mathbf{m}_1 + \mathbf{m}_2^H \mathbf{m}_2 \leq 1. \label{3cond2}
\end{align}
\end{subequations}
Suppose $\mathbf{m}_2^H\mathbf{m}_2 =1-\alpha^2$, $\alpha>0$, where $\alpha$ will be identified later in the proof. Then, by conditions \eqref{3cond1} and \eqref{3cond2} we have
\begin{subequations}
\begin{align}
\mathbf{h}_S^H \mathbf{m}_1 &=\beta, \label{4cond1} \\
\mathbf{m}_1^H \mathbf{m}_1 &\leq \alpha^2. \label{4cond2}
\end{align}
\end{subequations}
Since $\beta$ is maximized, condition~\eqref{4cond2} should be satisfied with equality. This will emerge in the sequel by contradiction. Assume $ \mathbf{m}_1^H \mathbf{m}_1 = \alpha^2 -\varepsilon$ for some $\varepsilon >0$. Then, since both $\alpha^2 -\varepsilon$ and $\beta$ are real numbers we can easily deduce that
\begin{align}
\mathbf{m}_1 & =\frac{\alpha^2-\varepsilon}{\beta} \mathbf{h}_S. \label{eq:1pre}
\end{align}
Substituting \eqref{eq:1pre} into \eqref{4cond1}, $\beta^2 = (\alpha^2 -\varepsilon)\| \mathbf{h}_S \|^2$; $\beta$ is maximized when $\varepsilon =0$, so condition~\eqref{4cond2} is satisfied with equality. Thus,
\begin{align}\label{beta_a}
\beta = \alpha\| \mathbf{h}_S \|,
\end{align}
and
\begin{align}\label{m11a}
\mathbf{m}_1 =\frac{\alpha^2}{\beta}\mathbf{h}_S.
\end{align}
Now, we need to find $\alpha$ and $\mathbf{m}_2$. It is observed in \eqref{beta_a} that $\alpha$ should be maximized in order to maximize $\beta$. Given that $\mathbf{m}_2^H \mathbf{m}_2=1-\alpha^2$, this is equivalent to minimizing $\mathbf{m}_2^H \mathbf{m}_2$. Towards this end, we formulate the following optimization problem:
\begin{subequations}\label{OPT5}
\begin{align}
\displaystyle \min & ~ \mathbf{m}_2^H \mathbf{m}_2 ,\label{5obj1} \\
\textrm{s.t} & ~
\mathbf{h}_S^H \mathbf{m}_2 = -\omega h_{TR}. \label{5cond1}
\end{align}
\end{subequations}
By Cauchy-Schwartz inequality
\begin{align}
|\mathbf{h}_S^H \mathbf{m}_2 |^2 \leq \|\mathbf{h}_S^H \|^2 \| \mathbf{m}_2 \|^2 \label{cauchy},
\end{align}
and by substituting \eqref{5cond1} and \eqref{5obj1} into \eqref{cauchy}, after algebraic manipulation,
\begin{align}
\mathbf{m}_2^H \mathbf{m}_2 \geq \frac{\omega^2 |h_{TR}|^2}{\|\mathbf{h}_S \|^2} \label{cauchy1}.
\end{align}
Minimizing $ \mathbf{m}_2^H \mathbf{m}_2 $ can be achieved when \eqref{cauchy1} holds with equality, i.e.,
\begin{align}
\mathbf{m}_2^H \mathbf{m}_2 = \frac{\omega^2 | h_{TR}|^2}{\|\mathbf{h}_S \|^2} \label{cauchy2}.
\end{align}
Combining \eqref{5cond1} and \eqref{cauchy2} we get
\begin{align}
\left( \mathbf{m}_2^H + \frac{\omega h_{TR}^H }{\|\mathbf{h}_S \|^2}\mathbf{h}_S^H \right) \mathbf{m}_2 =0 \label{cauchy3}.
\end{align}
\begin{comment}
This problem can be equivalently formulated as follows:
\begin{subequations}
\begin{align}
\displaystyle \min & ~ \mathbf{v}^T \mathbf{v} ,\label{obj22} \\
\textrm{s.t} & ~
\mathbf{A} \mathbf{v} = \mathbf{c},
\label{con21}
\end{align}
\label{OPT21}
\end{subequations}
where
\begin{align*}
\mathbf{v}^T &\triangleq \begin{pmatrix}
\mathcal{R}(m_{12}) & \mathcal{I}(m_{12})& \ldots & \mathcal{R}(m_{\nu2})& \mathcal{I}(m_{\nu2})\\
\end{pmatrix}, \\
\mathbf{A} &\triangleq
\begin{pmatrix}
\mathcal{R}(h_{S_1 R}) & -\mathcal{I}(h_{S_1 R}) & \ldots & \mathcal{R}(h_{S_\nu R}) & -\mathcal{I}(h_{S_\nu R}) \\
\mathcal{I}(h_{S_1 R}) & \mathcal{R}(h_{S_1 R}) & \ldots & \mathcal{I}(h_{S_\nu R}) & \mathcal{R}(h_{S_\nu R}) \\
\end{pmatrix}, \\
\mathbf{c}^T &\triangleq \omega \begin{pmatrix}
-\mathcal{R}(h_{TR}) & -\mathcal{I}(h_{TR})\\
\end{pmatrix}.
\end{align*}
This optimization problem is easily shown to be convex and by introducing a real-valued vector Lagrange multiplier $\bm{\lambda}$, the cost function becomes
\begin{align}
L(\mathbf{v},\bm{\lambda})\triangleq \mathbf{v}^T \mathbf{v} + \bm{\lambda}(\mathbf{A} \mathbf{v} - \mathbf{c}).
\end{align}
By the Karush-Kuhn-Tucker (KKT) theorem, the necessary and sufficient conditions for optimality are found. Differentiating w.r.t. $\mathbf{v}$ and $\bm{\lambda}$, respectively, we have
\begin{subequations}
\begin{align}
\bigtriangledown_{\mathbf{v}}L(\mathbf{v},\bm{\lambda}) |_{ \mathbf{v}=\mathbf{v}_\star, \bm{\lambda}=\bm{\lambda}_\star} &= 2\mathbf{v}_\star+ \mathbf{A}^T \bm{\lambda}_\star^T = \mathbf{0}, \label{lag1} \\
\bigtriangledown_{\bm{\lambda}}L(\mathbf{v},\bm{\lambda}) |_{ \mathbf{v}=\mathbf{v}_\star, \bm{\lambda}=\bm{\lambda}_\star} &= \mathbf{A}\mathbf{v}_\star - \mathbf{c}= \mathbf{0}. \label{lag2}
\end{align}
\end{subequations}
From \eqref{lag1} and \eqref{lag2},
\begin{align}
\mathbf{v}_\star = \mathbf{A}^T ( \mathbf{A} \mathbf{A}^T )^{-1}\mathbf{c}.
\end{align}
Note that $\mathbf{A} \mathbf{A}^T = \| \mathbf{h}_S \|^2 \mathbf{I}$.
Hence,
\begin{align}\label{alpha_a}
\alpha^2=1-\mathbf{v}_\star^T\mathbf{v}_\star =1- \mathbf{c}^T ( \mathbf{A} \mathbf{A}^T )^{-1}\mathbf{c}= 1-\omega^2 \frac{| h_{TR} |^2}{\| \mathbf{h}_S \|^2},
\end{align}
and
\begin{align}
\mathbf{v}_\star&=\frac{1}{\| \mathbf{h}_S \|^2}\mathbf{A}^T\mathbf{c} =-\frac{\omega}{\| \mathbf{h}_S \|^2}\begin{pmatrix}
\mathcal{R}({h}^H_{S_1 R} h_{TR}) \\[0.14cm]
\mathcal{I}({h}^H_{S_1 R} h_{TR}) \\[0.14cm]
\vdots\\[0.14cm]
\mathcal{R}({h}^H_{S_\nu R} h_{TR}) \\[0.14cm]
\mathcal{I}({h}^H_{S_\nu R} h_{TR}) \\[0.14cm]
\end{pmatrix}.
\end{align}
\end{comment}
Since $\mathbf{m}_2 \neq \mathbf{0}$, then
\begin{align}
\mathbf{m}_2 =-\omega\frac{h_{TR}}{\| \mathbf{h}_S \|^2}\mathbf{h}_S .
\end{align}
Combining \eqref{beta_a} and \eqref{cauchy2}, the maximum $\beta$ is given by
\begin{align}\label{beta_b}
\beta = \alpha\| \mathbf{h}_S \|= \sqrt{\| \mathbf{h}_S \|^2 - \omega^2 | h_{TR} |^2}.
\end{align}
Substituting $\alpha$ and $\beta$ into \eqref{m11a}, we have
\begin{align}
\mathbf{m}_1 =\frac{ \sqrt{\| \mathbf{h}_S \|^2 - \omega^2 | h_{TR} |^2}}{\| \mathbf{h}_S \|^2}\mathbf{h}_S .
\end{align}
Finally, the precoding matrix $\mathbf{M}$ is given by
\begin{align*}
\mathbf{M} = \frac{1}{\| \mathbf{h}_S \|^2} \begin{bmatrix}
\sqrt{\| \mathbf{h}_S \|^2 - \omega^2 | h_{TR} |^2} \mathbf{h}_S, &
-\omega h_{TR} \mathbf{h}_S \end{bmatrix}.
\end{align*}
Now, we want to find the value of $\omega$ for which the SINR at the receiving relay is maximized.
For both \eqref{cauchy2} and \eqref{beta_b}, it is required that $\omega^2 | h_{TR} |^2 < \| \mathbf{h}_S \|^2$.
We substitute \eqref{beta_b} into optimization~\eqref{OPT2} and for simplicity of notation we denote $a=\|\mathbf{h}_S\|^2$ and $b=|h_{TR}|^2$.
Hence, optimization problem~\eqref{OPT2} is written as
\begin{subequations}
\begin{align}
\displaystyle \underset{\omega}{\max} & ~ f(\omega) \triangleq \frac{a - b\omega^2}{b(1-\omega)^2 + \rho}, \label{OPT_AR_2}\\[0.14cm]
\textrm{s.t.} & ~ 0< \omega \leq \min\left\{1,{\sqrt{\frac{a}{b}}}\right\}, \label{omega:ineq:1}
\end{align}
\end{subequations}
where $\min\{\cdot,\cdot\}$ denotes the minimum of arguments and the right hand side of \eqref{omega:ineq:1} comes from the condition that {$\beta^2$ (and subsequently $f(\omega)$)} is non-negative.
By differentiating $f(\omega)$ with respect to $\omega$,
\begin{align}
\frac{d f(\omega)}{d \omega} &= \frac{-2b\omega \left(b(1-\omega)^2 + \rho\right) + 2b(1-\omega)(a-b\omega^2)}{\left( b(1-\omega)^2 + \rho \right)^2} \nonumber \\
& =\frac{2b(b\omega^2 - (a+b+\rho)\omega + a )}{\left( b(1-\omega)^2 + \rho \right)^2} .
\end{align}
At a turning point, $\frac{d f(\omega)}{d \omega}|_{\omega=\omega^*} = 0$; hence,
\begin{align}\label{eq:2ndorder}
b{\omega^*}^2 - (a+b+\rho)\omega^* + a = 0.
\end{align}
The two roots of \eqref{eq:2ndorder} are obtained by
\begin{align}
\omega_{1,2} = \frac{(a+b+\rho) \pm \sqrt{(a+b+\rho)^2 - 4ab} }{2b}. \label{eq:roots}
\end{align}
First, we verify that the roots~\eqref{eq:roots} have real-values by checking if the second order equation in~\eqref{eq:2ndorder} has a positive discriminant, i.e.,
\begin{align*}
\Delta = (a+b+\rho)^2 - 4ab =(a-b)^2+\rho^2+2a\rho+2b\rho >0.
\end{align*}
Let $\omega_1$ be the smallest root. It can be easily shown that at $\omega_1$ we attain a maximum, whereas at $\omega_2$ we attain a minimum. In addition, {we need to show that any $\omega \geq \omega_2$ does not fulfill inequality constraint~\eqref{omega:ineq:1}, so that the maximum of $f(\omega)$ is not obtained on the boundary. We show this by contradiction. For $a<b$, suppose $\omega_2 < \sqrt{\frac{a}{b}}$. Then,
\begin{align*}
\omega_2 &= \frac{(a+b+\rho) + \sqrt{(a+b+\rho)^2 - 4ab} }{2b} \\
&= \frac{((\sqrt{a}-\sqrt{b})^2 + 2\sqrt{ab}+\rho) + \sqrt{(a+b+\rho)^2 - 4ab} }{2b} \\
& \stackrel{(\phi)}> \frac{2\sqrt{ab}}{2b} = \sqrt{\frac{a}{b}},
\end{align*}
where $(\phi)$ stems from the fact that all the eliminated elements are positive. For $b\leq a$, suppose $\omega_2 < 1$. Then,
\begin{align*}
\omega_2 &= \frac{(a+b+\rho) + \sqrt{(a+b+\rho)^2 - 4ab} }{2b} \\
&= \frac{(a+b+\rho) -2b + \sqrt{(a+b+\rho)^2 - 4ab} }{2b}+1 \\
&= \frac{(a-b+\rho)+ \sqrt{(a+b+\rho)^2 - 4ab} }{2b}+1 \stackrel{(\psi)}>1,
\end{align*}
where $(\psi)$ stems from the fact that all the eliminated elements are all positive.}
Hence, $\omega^*=\omega_1$, given it fulfills inequality constraint~\eqref{omega:ineq:1}. It is easily shown that $\omega_1$ is positive, so we will check if $\omega_1 \leq {\min\left\{1,\sqrt{ \frac{a}{b}} \right\}}$. For $a \geq b$ we need to check if $\omega_1 \leq 1$. Suppose $\omega_1 > 1$; then,
\begin{align*}
& \frac{(a+b+\rho) - \sqrt{(a+b+\rho)^2 - 4ab} }{2b}> 1 \\
\Leftrightarrow~& (a+b+\rho) - 2b > \sqrt{(a+b+\rho)^2 - 4ab}\\
\stackrel{(\xi)}{\Leftrightarrow}~& (a-b+\rho)^2 > (a+b+\rho)^2 - 4ab \\
\Leftrightarrow~& 2(a+\rho)(-2b) > -4ab \\
\Leftrightarrow~& \rho < 0,
\end{align*}
which contradicts the fact that $\rho > 0$. Step $(\xi)$ follows because $a-b+\rho >0$ ($\because a\geq b$). For $a < b$ we need to check if {$\omega_1 \leq\sqrt{ \frac{a}{b}} $}. Suppose {$\omega_1 > \sqrt{\frac{a}{b}}$}; then,
{\begin{align*}
& \frac{(a+b+\rho) - \sqrt{(a+b+\rho)^2 - 4ab} }{2b}>\sqrt{ \frac{a}{b}} \\
\Leftrightarrow~& (a+b+\rho) - 2\sqrt{ab} > \sqrt{(a+b+\rho)^2 - 4ab}\\
\stackrel{(\zeta)}{\Leftrightarrow}~& \big((a+b+\rho) - 2\sqrt{ab}\big)^2 > (a+b+\rho)^2 - 4ab \\
\Leftrightarrow~& ((b+\rho)-a)^2 - ((b+\rho)+a)^2 > - 4ab \\
\Leftrightarrow~& 2(b+\rho)(-2a)> - 4ab \\
\Leftrightarrow~& \rho < 0,
\end{align*}
which contradicts the fact that $\rho > 0$. Step $(\zeta)$ follows because $a+b+\rho -2\sqrt{ab} = (\sqrt{a}-\sqrt{b})^2+\rho>0$}.
\hfill $\Box$
\section{Proof of Proposition~\ref{propB}}\label{B:proof_propB}
It can be easily shown that the choice of $P_S$ changes neither the precoding matrix $\mathbf{M}$ nor the value of $\beta$. Let $c\triangleq \sqrt{P_T}$ and $x\triangleq \sqrt{P_S}$. Then, the optimization problem in which power level selection is also possible can be written as
\begin{subequations}
\begin{align}
\displaystyle \underset{\omega, x}{\max} & ~ h(\omega,x) \triangleq \frac{(a - b\omega^2)x^2}{b(c-\omega x)^2 + \sigma^2}, \label{OPT_AR_3}\\[0.14cm]
\textrm{s.t.} & ~ 0< \omega \leq \min\left\{1,\sqrt{\frac{a}{b}}\right\}, \label{omega:ineq:11} \\
& ~ 0< x \leq M, \label{omega:ineq:12}
\end{align}
\end{subequations}
where $M \triangleq \sqrt{P_{\max}}$. Differentiating $h(\omega,x)$ w.r.t. $x$, the value of $x$ that maximizes $h(\omega,x)$ is given by
\begin{align}
x^* = \frac{bc^2 +\sigma^2}{bc \omega}. \label{omega_x1}
\end{align}
Differentiating $h(\omega,x)$ w.r.t. $\omega$ and by following similar steps to that of Proposition~\ref{propA}, the value of $\omega$ that maximizes $h(\omega,x)$ is given by
\begin{align}
\omega^* = \frac{ax^2+bc^2 +\sigma^2 - \sqrt{(ax^2+bc^2 +\sigma^2)^2 -4abc^2 x^2}}{2bcx}. \label{omega_star1}
\end{align}
The solution of \eqref{omega_x1}-\eqref{omega_star1} gives $\omega^* =0$, suggesting that $x^*=\infty$, i.e., the source uses infinite power. However, power is constrained and this solution is not feasible. By substituting $x^*$ of \eqref{omega_x1} in \eqref{OPT_AR_3}, we obtain
\begin{align}
h(\omega)\triangleq \frac{(a-b\omega^2)(bc^2+\sigma^2)}{b \sigma^2 \omega^2} ,
\end{align}
which is monotonically decreasing with $\omega$. Hence, in order to maximize the SINR with respect to $x$, $\omega$ should be kept as small as possible, while satisfying \eqref{omega_x1}. This means that the value of $x$ that maximizes the SINR is at $x=M$. Let $\omega_a$ be the value computed via \eqref{omega_x1} for $x=M$. For $x=M$, the value of $\omega$ that maximizes the SINR with respect to $\omega$, say $\omega_b$, can be computed via \eqref{omega_star1}. Then,
\begin{itemize}
\item[(a)] if $\omega_b>\omega_a$, one can find $x_b<M$ that increases SINR (since the maximum of the SINR with respect to $x$ is shifted to a lower value); but then, $h(\omega_a, M) > h(\omega_b,x_b)$ which means that if $\omega_b>\omega_a$ the maximum value is always obtained at $h(\omega_a, M)$;
\item[(b)] if $\omega_b<\omega_a$, then $M<x^*$ and the hence $x=M$; it cannot be improved further since $h(\omega, x)$ is an increasing function with respect to $x$ until $x=\frac{bc^2+\sigma^2}{b\omega_b}$. Hence, $\omega_b$ finds the optimal value of the SINR with respect to $\omega$, while it also increases the SINR with respect to $x$.
\end{itemize}
Hence, the source should \emph{always} transmit with $P_{\max}$. In case, $P_{\max}$ is small such that \eqref{omega_x1} is not satisfied and \eqref{omega_star1} does not yield a feasible $\omega$, then the maximum SINR is achieved for $\omega$ on the boundary, i.e., $\omega=\min\{1,\sqrt{\frac{a}{b}}\}$, since $h(\omega, x)$ is an increasing function with respect to $\omega$, for $\omega\leq \omega^*$ (as given by \eqref{omega_star1}).
\hfill $\Box$
\section{Proof for Proposition~\ref{propC}}\label{C:proof_propC}
\noindent By triangle inequality
\begin{align}\label{triangle}
\norm{\frac{h_{S_2R}}{\sqrt{2}}e^{j\phi} +h_{TR}} \geq \norm{ \norm{\frac{h_{S_2R}}{\sqrt{2}}} - \norm{h_{TR}}}.
\end{align}
\noindent The optimization problem
\begin{align}\label{opt:1a}
\min_{\phi} \norm{\frac{h_{S_2R}}{\sqrt{2}}e^{j\phi} +h_{TR}}^2 ,
\end{align}
is minimized when inequality \eqref{triangle} holds with equality; this occurs when $h_{S_2R}$ is in phase with $-h_{TR}$. Let $\phi^\star$ the optimal angle $\phi$ for optimization \eqref{opt:1a}. Since $\norm{e^{j\phi}}=1$, then the minimization yields
\begin{align*}
e^{j\phi^\star}= -\frac{h_{S_2R}^H h_{TR}}{|h_{S_2R}^H h_{TR}|}.
\end{align*}
Similarly, by triangle equality
\begin{align}\label{triangle1}
\norm{\frac{h_{S_2R}}{\sqrt{2}}e^{j\phi} +h_{TR}} \leq \norm{\frac{h_{S_2R}}{\sqrt{2}}} + \norm{h_{TR}}.
\end{align}
\noindent The optimization problem
\begin{align}\label{opt:1b}
\max_{\phi} \norm{\frac{h_{S_2R}}{\sqrt{2}}e^{j\phi} +h_{TR}}^2 ,
\end{align}
is maximized when inequality \eqref{triangle1} holds with equality; this occurs when $h_{S_2R}$ is in phase with $h_{TR}$. Let $\phi^\dagger$ the optimal angle $\phi$ for optimization \eqref{opt:1b}. Since $\norm{e^{j\phi}}=1$, the maximization yields
\begin{align*}
e^{j\phi^\dagger}=\frac{h_{S_2R}^H h_{TR}}{|h_{S_2R}^H h_{TR}|} .
\end{align*}
\begin{comment}
\noindent The optimization problem
\begin{align}\label{opt:1a}
\min_{\phi} \norm{\frac{h_{S_2R}}{\sqrt{2}}e^{j\phi} +h_{TR}}^2 ,
\end{align}
is equivalent to
\begin{align}\label{opt:2a}
\min_{\theta_1, \theta_2 \atop \theta_1^2+\theta_2^2=1} \norm{(c_1+jc_2)(\theta_1+j\theta_2)+(l_1+jl_2)}^2 ,
\end{align}
where $h_{S_2R}=\sqrt{2}(c_1+jc_2)$, $e^{j\phi}=\theta_1+j\theta_2$ and $h_{TR}=l_1+jl_2$. Let
\begin{align*}
L(\theta_1, \theta_2)&\triangleq \norm{(c_1+jc_2)(\theta_1+j\theta_2)+(l_1+jl_2)}^2 \\
& = (c_1\theta_1-c_2\theta_2+l_1)^2 + (c_1\theta_2+c_2\theta_1+l_2)^2.
\end{align*}
The constraint qualification is fulfilled, so we can solve optimization~\eqref{opt:2a} using Lagrange multipliers. By introducing real valued Lagrange multiplier $\lambda$ associated with the constraint $\theta_1^{\star2}+\theta_2^{\star2}=1$, we let
\begin{align*}
L_\lambda (\theta_1, \theta_2, \lambda) \triangleq (c_1\theta_1-c_2\theta_2+l_1)^2 + &(c_1\theta_2+c_2\theta_1+l_2)^2 \\
&+ \lambda(1-\theta_1^2-\theta_2^2).
\end{align*}
Taking the derivative of $L_\lambda (\theta_1, \theta_2, \lambda)$ with respect to $\theta_1$ and $\theta_2$, we get
\begin{subequations}
\begin{align}
\frac{\partial L_\lambda (\theta_1, \theta_2, \lambda)}{\partial \theta_1} = 2c_1(c_1\theta_1-&c_2\theta_2+l_1) \label{eq:partial1} \\
&+ 2c_2(c_1\theta_2+c_2\theta_1+l_2) -2\lambda\theta_1,\nonumber \\
\frac{\partial L_\lambda (\theta_1, \theta_2, \lambda)}{\partial \theta_2} = -2c_2(c_1\theta_1-&c_2\theta_2+l_1) \label{eq:partial2} \\
&+ 2c_1(c_1\theta_2+c_2\theta_1+l_2) -2\lambda\theta_2.\nonumber
\end{align}
\end{subequations}
Invoking the Karush-Kuhn-Tucker (KKT) theorem, the following conditions are necessary and sufficient for optimality.
\begin{subequations}
\begin{align}
\left. \frac{\partial L_\lambda (\theta_1, \theta_2, \lambda)}{\partial \theta_1}\right|_{\theta_1=\theta_1^\star} &= 0 , \\
\left. \frac{\partial L_\lambda (\theta_1, \theta_2, \lambda)}{\partial \theta_2}\right|_{\theta_2=\theta_2^\star} &= 0 , \\
\lambda^\star(1-\theta_1^{\star2}-\theta_2^{\star2}) &=0.
\end{align}
\end{subequations}
When $\lambda^\star=0$, the condition for $\theta_1^{\star2}+\theta_2^{\star2}=1$ is violated. Then, it is necessary that $\lambda^\star\neq 0$. Using the KKT conditions on \eqref{eq:partial1}-\eqref{eq:partial2} and the constraint $\theta_1^{\star2}+\theta_2^{\star2}=1$, we obtain
\begin{align*}
\theta_1^\star =\pm \frac{c_1l_1+c_2 l_2}{\sqrt{(c_1l_1+c_2 l_2)^2+(c_1l_2-c_2 l_1)^2}} =\pm\frac{\mathcal{R}(h_{S_2R}^H h_{TR})}{|h_{S_2R}^H h_{TR}|} ,\\
\theta_2^\star =\pm \frac{c_1l_2-c_2 l_1}{\sqrt{(c_1l_1+c_2 l_2)^2+(c_1l_2-c_2 l_1)^2}} =\pm\frac{\mathcal{I}(h_{S_2R}^H h_{TR})}{|h_{S_2R}^H h_{TR}|} .
\end{align*}
We have two solutions: one corresponds to the minimization and the other to the maximization of the norm $\norm{{h_{S_2R}}e^{j\phi}/{\sqrt{2}} +h_{TR}}^2$. This occurs because the constraint set is compact so the function achieves both its maximum and minimum.
As a result, the minimization yields
\begin{align*}
e^{j\phi^\star}=\theta_1^\star +j\theta_2^\star = -\frac{h_{S_2R}^H h_{TR}}{|h_{S_2R}^H h_{TR}|} ,
\end{align*}
whereas the maximization yields
\begin{align*}
e^{j\phi^\dagger}=\theta_1^\dagger +j\theta_2^\dagger = \frac{h_{S_2R}^H h_{TR}}{|h_{S_2R}^H h_{TR}|} .
\end{align*}
\end{comment}
The proof is now complete.
\hfill $\Box$
\section*{References}
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\label{sec:intro}
Bipartite entanglement entropies in quantum critical systems can harbour universal quantities that characterize the underlying theory.
Originating in the scaling dependence of entropies on the size of the entangled region,
these quantities arise from different geometric features in the entangling boundary.
Such quantities provide both a new perspective on what constitutes a universal number as well as a concrete connection
between seemingly disparate physical theories arising in condensed matter, high energy field theory and gravity.
Amid the growing realization of their conceptual importance is the recognition that little is known about
the number of undiscovered universal quantities, their numerical values,
and their relationship to conventional universality such as that arising from $n$-point correlation functions.
In spacetime dimensions higher than $d$=1+1, this uncertainty is present even for free theories, where the calculations
of entanglement entropies can be technically challenging.\cite{Casini1}
Despite this challenge, the many different types of geometric features
available in higher-dimensional entangling surfaces offers a rich opportunity
to search for new universal quantities in the entanglement entropy.\cite{Casini3,Kallin2013,Bueno1,Bueno3,Witczak-Krempa_2016,Chojnacki_2016}
In addition to giving information about features of the underlying critical theory, understanding these quantities in free theories
is a necessary precursor to their exploration in interacting critical points,\cite{Kallin2014,Fei2015} such as those in real quantum materials or atomic systems.\cite{CuprateTc,BEC_Greiner,MagnetQCP1,MagnetQCP2}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.52]{sinsin6.pdf}
\caption{Contributions to the R\'enyi\ entropy of $(3+1)$-dimensional CFTs from (left) a wedge of opening angle $\phi$ (non-universal) and (right) a trihedral corner parametrized by angles $\theta_1$, $\theta_2$, $\theta_3$ (universal). }
\labell{fig1}
\end{figure}
In this paper, we examine a particular universal quantity that arises in $d=3+1$ spacetime dimensions when the entangling geometry
contains a (cubic) trihedral corner --- see Fig.~\ref{fig1}.
This quantity $v_{\alpha}$ appears as the coefficient of a scaling term with a logarithmic dependence on the size of the entangling boundary. This coefficient has been computed previously in quantum Ising models at the infinite disorder \cite{Kovacs} and Wilson-Fisher fixed points.\cite{Devakul2014}
The latter work observed that $v_{\alpha}$ exhibits a functional dependence on $\alpha$ that is close to that for the corresponding coefficient of a spherical entangling surface.
This observation suggests that it may be possible to understand the trihedral corner coefficient as coming from the singular limit of a smooth interface, {\it i.e.,}\ an eighth of that of the sphere.
Here, we perform a numerical calculation of $v_{\alpha}$ for a massless free scalar field theory and show that while the functional dependence on $\alpha$
indeed closely matches that arising from a smooth sphere, in general both the magnitude and sign of the trihedral coefficient are different.
We suggest future calculations that may provide further insight into the universal content of the trihedral corner term.
\section{R\'{e}nyi\ entropies in $d=3+1$}
\label{sec:renn}
We consider the entanglement corresponding to spatial bipartitions of a physical system into a region $A$ and its complement $\bar A$, separated by a surface $\partial A$.
One can quantify the entanglement for such bipartite systems through the entanglement entropy $\see$ or, more generally, the R\'{e}nyi\ entropies $S_\alpha$.\cite{Renyi1,Renyi2}
For our calculations, $A$ is bounded by a cubic trihedral corner, which is formed by three intersecting orthogonal planes.
The corresponding Hilbert space is bipartitioned such that $\mathcal{H}=\mathcal{H}_{A}\otimes \mathcal{H}_{\bar A}$ and then, given a state $\rho \in \mathcal{H}$, the $\alpha$-R\'{e}nyi\ entropy is defined as
\begin{equation}\label{renyi}
S_\alpha(A) = \frac{1}{1-\alpha} \ln \left[ \Tr \left( \rho_A^{\,\alpha} \right) \right] \, ,
\end{equation}
where $\rho_A = \Tr_{\bar A} \rho$ is the reduced density matrix associated with subregion $A$ and $\alpha$ is the R\'{e}nyi\ index. When Eq.~\req{renyi} can be evaluated for real values of $\alpha$, the entanglement entropy can in turn be obtained as $\lim_{\alpha\rightarrow 1} S_{\alpha}(A)=\see(A)=-\Tr \left( \rho_A \log \rho_A \right)$.
We note that in many interacting systems, the R\'{e}nyi\ entropies can only be studied for positive integer $\alpha \ge 2$, which is accomplished by evaluating a
partition function on a multi-sheeted Riemann surface,\cite{Calabrese1,Calabrese2} as has recently been employed in quantum Monte Carlo simulations\cite{swap} and in experiments on ultra-cold atoms.\cite{Islam2015}
In the following, we begin with general scaling arguments that apply for all values of $\alpha$ for corner singularities in $d=3+1$.
We then take another approach where we explore a smooth regularization of the cube.
We comment on consistencies, inconsistencies and predictions related to these two approaches.
\subsection{General scaling arguments}
When computed for the vacuum state of a local Hamiltonian, R\'{e}nyi\ entropies are dominated by short-distance correlations across the entangling surface $\partial A$.
These local correlations yield the celebrated ``area law" term as the leading contribution.\cite{sorkin,Bombelli1986,Srednicki1993} That is,
\begin{equation}
S_{\alpha}(A)=B_{\alpha}\,{\cal A}/\delta^{d-2}+\cdots\,,
\label{got}
\eeq
where we are considering a quantum field theory living in $d$ spacetime dimensions.
In this expression, the coefficient $B_{\alpha}$ is non-universal, {\it i.e.,}\ regulator dependent, such that it depends on the procedure used to regulate the calculation.
$\cal A$ is the area of the entangling surface $\partial A$ and $\delta$ is a short-distance cutoff, {\it e.g.,}\ the lattice spacing.
In general, the subleading contributions indicated by the ellipsis in Eq.~\reef{got} include further power-law divergences and the corresponding coefficients also depend on the details of the regulator.
However, a regulator-independent coefficient providing well-defined information about the underlying theory appears if there is a subleading contribution that scales logarithmically with $\ell/\delta$, where $\ell$ is a length scale characteristic of the size of the region $A$.
For example, in $d=2+1$, introducing a sharp corner in the entangling surface produces such a logarithmic contribution $S^{\rm univ}_{\alpha,\rm corner}=-a_{\alpha}(\theta)\log(\ell/\delta)$, where the universal coefficient $a_{\alpha}(\theta)$ is a function of the opening angle $\theta$ of the corner and the R\'{e}nyi\ index $\alpha$.\cite{Casini1,Casini2,Casini3,Hirata,Myers:2012vs,Bueno1,Bueno2,Bueno3,Bueno5,Miao2015,Elvang,Kallin2013,Stoudenmire2014,helmes14,devakul14,Kallin2014,Sahoo2016,Helmes2016,DeNobili:2016nmj}
In $d=3+1$, geometric singularities in $\partial A$ can also produce similar universal contributions in $S_{\alpha}(A)$.\cite{Klebanov:2012yf,Myers:2012vs,Bueno4,Devakul2014,Kovacs}
For the present discussion, let us focus on a three-dimensional region $A$ which is a polyhedron auch that the entangling surface $\partial A$ consists of flat polygonal faces, straight edges and sharp corners or vertices. We limit our discussion to the R\'{e}nyi\ entropies in a conformal field theory (CFT), in which case only geometric scales appear in $S_\alpha$.\footnote{In a more general QFT, the dimensionless coefficients $B_\alpha$, $w_\alpha$ and $v_\alpha$ may also depend on the combinations $\mu_i\delta$, where the $\mu_i$ denote various mass scales appearing in the QFT. See, for example, discussion in Ref.~\onlinecite{Myers:2012vs}.} For such a geometry, $S_\alpha(A)$ are expected to scale such that
\begin{eqnarray}
S_{\alpha, \rm poly} &=& B_\alpha \, {\cal A}/{\delta^2} + \sum_i w_\alpha(\phi_i) \, {{L_i}}/{\delta} \label{eq:Spoly} \\
&&{}+ \sum_i v_\alpha(\theta_{i,1},\theta_{i,2},\theta_{i,3})\log\!\left(\ell/\delta \right) + \mathcal{O}((\ell/\delta)^0) \, .
\nonumber
\end{eqnarray}
Beyond the area law term, the first subleading contribution arises from the edges on the boundary of the polyhedron.\cite{Myers:2012vs,Klebanov:2012yf} Here $L_i$ is the length of the $i^{\text{th}}$ edge.\footnote{We examine the functional form of this edge contribution to the R\'{e}nyi\ entropy in more detail in Appendix~\ref{app_edge}.} The (non-universal) coefficients $w_\alpha(\phi_i)$ depend on the opening angle $\phi_i$ between the two faces intersecting at the edge. The logarithmic term in Eq.~\reef{eq:Spoly} arises from the vertices in $\partial A$.
\footnote{In this term, $\ell$ refers to some length scale characteristic of the geometry of region $A$. Note that for the cube in Eq.~\reef{eq:S_octant}, there is a single scale $L$ that defines the area and the edge lengths. In this case, $L$ naturally appears in the logarithmic term.}
As shown in Fig.~\ref{fig1}, the universal coefficient $v_\alpha(\theta_{i,1},\theta_{i,2},\theta_{i,3})$ at the $i^{\text{th}}$ vertex is a function of the angles $\theta_{i,1}$, $\theta_{i,2}$ and $\theta_{i,3}$ between each of the three pairs of (adjacent) edges that end at the vertex.
For the entanglement entropy, $w_1(\phi)\le 0$ follows from strong subadditivity;\footnote{This inequality follows from a simple generalization of the argument in Ref.~\onlinecite{Casini2} to higher dimensions.}
however, there are no known analogous arguments that rigorously fix the sign of either $w_\alpha$ for general $\alpha$ or of $v_\alpha$. However, intuitive arguments can be made to suggest that $w_\alpha\le0$ and $v_\alpha\ge0$.\footnote{For example, we might imagine that the area-law term simply counts Bell-pair correlations across the smooth faces in $\partial A$. Then $w_\alpha\le0$ follows since the edge terms must compensate for over-counting the correlations of the degrees of freedom near the edges.} These signs were confirmed for the explicit examples in Refs.~\onlinecite{Devakul2014,Kovacs} and are reproduced in our calculations below.
This paper focusses on the universal coefficient $v_\alpha$ of the logarithmic contribution to $S_\alpha$
arising from a trihedral corner (formed by three intersecting faces as in Fig.~\ref{fig1}). For simplicity, we examine the cubic case where the opening angles are all $\pi/2$ and thus henceforth we omit from our notation the angular dependence of $v_\alpha$.
For a cube of dimension $L$, Eq.~\reef{eq:Spoly} then simplifies to
\begin{equation}
S_{\alpha, \rm cube} = B_\alpha \, \frac{6L^2}{\delta^2} + w_\alpha \frac{12L}{\delta} + 8v_\alpha\log\!\left(L/\delta \right) + \mathcal{O}((L/\delta)^0) \,.
\label{eq:S_octant}
\end{equation}
We are particularly interested in the dependence of the universal corner coefficient $v_\alpha$ on the R\'{e}nyi\ index $\alpha$.
\subsection{Smooth entangling surfaces}
\label{sec:smoothSurfaces}
As remarked in the introduction, one might hope to understand the trihedral corner coefficient as coming from the singular limit of a smooth entangling surface. Hence,
in this section, we review the structure of universal contributions to the R\'enyi\ entropy for smooth entangling surfaces in $(3+1)$-dimensional CFTs. In this case,\footnote{Implicitly, we are considering the region $A$ to lie on a constant time slice in flat space with $d=4$.} the R\'enyi\ entropy takes the form
\begin{equation}
S_{\alpha, \rm smooth} = B_\alpha \, {\cal A}/{\delta^2} + u_\alpha \log\!\left(\ell/\delta \right) + \mathcal{O}((\ell/\delta)^0) \, ,
\label{smooth}
\eeq
where the universal coefficient $u_{\alpha}$ is determined by the geometry of the entangling surface according to\cite{Fursaev:2012mp,Lee:2014xwa}
\begin{align}
u_{\alpha}=-\int_{\partial A} \frac{d^2y\sqrt{h}}{2\pi} \,\left[ f_a(\alpha)\,\mathcal{R}+f_{b}(\alpha)\,\hat{K}^2\right]\, .
\labell{gamer}
\end{align}
In this expression, $h=\det(h_{ij})$ is the determinant of the induced metric on $\partial A$, $\mathcal{R}$ is the Ricci scalar of this metric, and $\hat{K}^2\equiv K_{ij}^{a}K_{ji}^{a}- (h^{ij}K^{a}_{ji})^2/2$ where $K_{ij}^{a}$ is the extrinsic curvature of the entangling surface.
The two coefficients $f_a(\alpha)$ and $f_b(\alpha)$ in Eq.~\req{gamer} contain universal information that characterizes the underlying CFT.
In particular, in the limit $\alpha\to1$, these functions yield the central charges in the trace anomaly,\cite{Solodukhin:2008dh} {\it i.e.,}\ $f_a\to a$ and $f_b\to c$.
In the case of a massless free scalar, which we study here, these functions are\cite{Fursaev:2012mp,Lee:2014xwa}$^,$\footnote{The result for $f^\mt{scalar}_b(\alpha)$ in Eq.~\reef{name} relies on the equality $f_b(\alpha)=f_c(\alpha)$ where $f_c(\alpha)$ is a third universal coefficient in Eq.~\reef{gamer} proportional to the Weyl curvature of the background spacetime. This equality is known {\it not} to hold for general CFTs,\cite{dong1,dong2} but Ref.~\onlinecite{Lee:2014xwa} provided strong numerical evidence of its validity for the free scalar. Recently, Ref.~\onlinecite{displace} also provided an analytic proof that $f^\mt{scalar}_b(\alpha)=f^\mt{scalar}_c(\alpha)$\label{fuo}}
\begin{equation}\label{name}
f^\mt{scalar}_a(\alpha)=\frac{1}{3}f^\mt{scalar}_b(\alpha)=\frac{(\alpha+1)(\alpha^2+1)}{1440\alpha^3}\, .
\end{equation}
We note that the first contribution in Eq.~\eqref{gamer} is topological such that the integral multiplying $f_a(\alpha)$ yields twice the Euler characteristic $\chi$ of the entangling surface.
In particular, this contribution is the same for a sphere $S^2$ and a cube $C^2$, {\it i.e.,}\ $\chi(S^2) = \chi(C^2) = 2$.
For a sphere, this term represents the only contribution to the universal part of the R\'enyi\ entropy since $\hat{K}^2(S^2)=0$ and so one finds
\begin{equation}
S^{ \rm univ}_\alpha(S^2)= -4\,f_a(\alpha) \log\left(R/\delta\right)\, ,
\labell{game2d3}
\eeq
where $R$ is the radius of the sphere. When the entangling surface is a cube, the curvature is entirely concentrated in the eight corners and hence,
na\"ively, one might set $v_\alpha=-f_a(\alpha)/2$ in Eq.~\eqref{eq:S_octant} for the universal contribution from each individual corner. We examine this idea in more detail below.
\subsection{Smoothed cube}
\label{Scube}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.4]{regcube2.pdf}
\caption{Rounded cube ${\widetilde C}^2$.
The corners and edges are respectively replaced by eighths of a sphere and quarters of a cylinder of radius $R=\varepsilon L$ with $\varepsilon\ll1$.
}
\labell{fig2}
\end{figure}
In Ref.~\onlinecite{Devakul2014}, a numerical estimation of the coefficient $v_\alpha$ for a cubic trihedral corner was obtained for the Ising theory, with a magnitude very close to $1/8$ of the value of the universal coefficient appearing in Eq.~\reef{game2d3}, {\it i.e.,}\ $v_\alpha\simeq f_a(\alpha)/2$.
We comment that this result has the wrong sign compared to $S^{ \rm univ}_\alpha(S^2)$, but since this difference was not noticed at the time,
the result in Ref.~\onlinecite{Devakul2014} led to the suggestion that the trihedral corner coefficient might be understood as coming
from the singular limit of a smooth entangling surface.
In order to explore this,
we consider a smoothed cube in which the edges and vertices have been rounded off and replaced by cylinders and spheres, respectively.
In particular, as shown in Fig.~\ref{fig2}, we consider a rounded cube ${\widetilde C}^2$ where each of the eight corners is replaced by an eighth of a sphere of radius $R=\varepsilon L$ with $\varepsilon\ll1$. Further, each of the twelve edges is replaced by a quarter cylinder of radius $R=\varepsilon L$ and length $L\,(1-2\gamma_0 \varepsilon)$ where $\gamma_0$ is a fixed constant of $\mathcal{O}(1)$. A natural choice for this constant would correspond to $\gamma_0=1$, which fixes the central width of ${\widetilde C}^2$ to be $L$ for all values of $\varepsilon$.\footnote{One can also consider more elaborate schemes. For example, one can choose $\gamma_0$ to fix the area of ${\widetilde C}^2$ such that it matches that of $C^2$ instead, {\it i.e.,}\ one can fix ${\cal A}= 6L^2$ in Eq.~\reef{area}.
Observe that, to within the resolution of the short-distance cutoff $\delta$, any of these choices yields the same cube $C^2$ in the limit $R\to\gamma_1\delta$.}
By design, in the limit $\varepsilon\to0$ we recover the usual cube with sharp edges and corners. Hence, it does not seem \emph{a priori} unreasonable to expect that the trihedral corner coefficient can be extracted from the resulting R\'enyi\ entropy, however, as we show below, the situation is more subtle.
Applying Eqs.~\reef{smooth} and \req{gamer}, the R\'enyi\ entropy of our smoothed cube becomes
\begin{eqnarray}
S_{\alpha}({\widetilde C}^2) &=& B_\alpha \, {\cal A}/{\delta^2} -\left[4f_a(\alpha)+\left(\frac{3}{2\varepsilon}-3\gamma_0\right) f_b(\alpha)\right]
\nonumber\\
&&\qquad\qquad\times\ \log\!\left(L/\delta \right) + \mathcal{O}((L/\delta)^0)\,,
\label{game1hh}
\end{eqnarray}
where the area is given by
\begin{equation}
{\cal A}=6L^2+L^2\varepsilon\,(6\pi-24\gamma_0)+L^2\varepsilon^2 (4\pi-12\pi\gamma_0+24\gamma_0^2) \, .
\label{area}
\eeq
We can see that the limit $\varepsilon\to0$ is problematic in Eq.~\reef{game1hh} since the coefficient of the logarithmic term diverges. To produce regulated ({\it i.e.,}\ finite) R\'enyi\ entropies, we take the limit $\varepsilon \to \gamma_1\delta/L$ or $R\rightarrow \gamma_1\delta$ where $\gamma_1$ is again some $\mathcal{O}(1)$ constant. In this limit, the edges and corners of the cube are still rounded at a scale of the order of the short-distance cutoff $\delta$. In particular,
Eq.~\reef{game1hh} yields
\begin{eqnarray}
S_{\alpha}({\widetilde C}^2) &=& B_\alpha \, \frac{6L^2}{\delta^2} -\frac{3f_{b}(\alpha)}{2\gamma_1}\,\frac{L}\delta\log(L/\delta)
\label{large}\\
&&\quad + w_\alpha \frac{12L}{\delta} + 8v_\alpha\log\!\left(L/\delta \right) + \mathcal{O}((L/\delta)^0) \,,
\nonumber
\end{eqnarray}
where $w_\alpha=-B_{\alpha}\gamma_1(2\gamma_0-\pi/2)$ and
\begin{equation}
v_\alpha=-\frac12 f_a(\alpha)+\frac38\gamma_0\, f_b(\alpha)\,.
\label{larger}
\eeq
This result is problematic in two ways: First, we observe the appearance of a new divergence of the form $L/\delta\,\log(L/\delta)$, which is incompatible with the form expected in Eq.~\eqref{eq:S_octant} (see Appendix~\ref{app_edge}). Second, the coefficient of the logarithmic term is ambiguous because of the appearance of $\gamma_0$ in Eq.~\reef{larger}, {\it i.e.,}\ our desired ``universal'' coefficient depends on the details of the regulator.
A possible resolution of both of these problems is that Eq.~\reef{game1hh}, or Eq.~\reef{smooth}, simply does not describe the R\'{e}nyi\ entropies with sufficient accuracy to take the desired limit. That is, originally $\varepsilon$ is small but independent of the ratio $\delta/L$, and hence any terms of the form $\log\varepsilon$ or $\varepsilon^{-1}\log\varepsilon$ are concealed in the $\mathcal{O}((L/\delta)^0)$ contributions. However, with the limit $\varepsilon \to \gamma_1\delta/L$, such terms emerge to modify the terms explicitly enumerated in Eq.~\reef{large}.
That is, order-one contributions may be building up in the singular limit to restore the universality of the logarithmic coefficient and to cancel the unanticipated $L/\delta\,\log( L/\delta)$ contribution. We expand on this suggestion in the discussion in Section~\ref{discuss}.
Regardless, our smoothed-cube calculation suggests that the universal coefficient $v_\alpha$ is some linear combination of $f_a(\alpha)$ and $f_b(\alpha)$.
Further,
Eq.~\eqref{name} shows that the massless free scalar is a special case where both of these coefficients have the same dependence on the R\'enyi\ index.
Hence, independent of the precise linear combination, the above calculation suggests that
\begin{equation}
\frac{v^{\mt{scalar}}_\alpha}{v^{\mt{scalar}}_1}= \frac{(\alpha+1)(\alpha^2+1)}{4\,\alpha^3}\, .
\labell{game1}
\eeq
Below, we compare this prediction with the $\alpha$-dependence calculated numerically for the free scalar, and show that we find good agreement to within numerical accuracy. Note that this argument predicts that other theories
in general have a different dependence on $\alpha$, depending on the precise linear combination of $f_a(\alpha)$ and $f_b(\alpha)$ appearing in $v_\alpha$.
\section{Numerical methods for the free scalar field}
\label{sec:Methods}
In this section we perform direct calculations of the R\'{e}nyi\ entropies $S_\alpha(A)$ for a free real scalar field on a three-dimensional simple cubic lattice.
At each lattice site $i$, there exists a bosonic field $\phi_i$ and its conjugate momentum $\pi_i$, whose dynamics are controlled by the Hamiltonian
\begin{equation}
H =
\frac{1}{2}\sum_{i=1}^N \left( \pi_i^2 + m^2\phi_i^2 \right)
+ \frac{1}{2} \sum_{\langle i j \rangle} \left( \phi_j - \phi_i \right)^2\, .
\label{eq:Ham_free}
\end{equation}
In this expression, the first sum is over the $N$ lattice sites, while the second is over all nearest-neighbour pairs of sites. $m$ is the mass of the scalar field.
For a lattice in three spatial dimensions, the total number of lattice sites is $N = L_x L_y L_z$, where $L_x$, $L_y$ and $L_z$ are the linear lattice dimensions along $x$, $y$ and $z$ respectively.
This Gaussian theory has the appealing property that $S_\alpha(A)$ can be obtained for any $\alpha$ from knowledge of the two-point functions of $\phi_i$ and $\pi_i$ at lattice points within region $A$ --- see Section~\ref{sec:PeschelTrick}.
In order to isolate the logarithmic corner contribution, we use these calculations along with techniques from the numerical linked-cluster expansion\cite{Rigol2006,Rigol2007_1,Rigol2007_2,Tang2013} (NLCE), as described in Section~\ref{sec:NLCE}.
We focus on the case where the boson is massless ($m=0$) such that the energy spectrum is gapless and we have a scale-invariant critical theory.
However, the techniques described in this section are equally applicable to free bosons of any mass. Further, note that in the following we measure lengths in units of the lattice spacing, which we set to unity for simplicity, {\it i.e.,}\ $\delta=1$.
\subsection{R\'{e}nyi\ entropies for free scalars}
\label{sec:PeschelTrick}
Here we explain how R\'enyi\ entropies for free scalars can be computed from the correlators $\left\langle \phi_i \phi_j \right\rangle$ and $\left\langle \pi_i \pi_j \right\rangle$ as first introduced in Ref.~\onlinecite{Peschel2003}.
The Hamiltonian in Eq.~\eqref{eq:Ham_free} can be written in the more general quadratic (Gaussian) form
\begin{equation}
H =
\frac{1}{2}\sum_{i=1}^N \pi_i^2
+ \frac{1}{2} \sum_{ i j } \phi_i M_{ij} \phi_j ,
\label{eq:Ham_quadratic}
\end{equation}
where $M$ is an $N \times N$ matrix that takes into account the boundary conditions of the finite lattice.
The groundstate two-point correlators are given in terms of this matrix as
\begin{align}
X_{ij} &\equiv \left\langle \phi_i \phi_j \right\rangle = \frac{1}{2} \left( M^{-1/2} \right)_{ij} \label{eq:XP}\, , \\
P_{ij} &\equiv \left\langle \pi_i \pi_j \right\rangle = \frac{1}{2} \left( M^{1/2} \right)_{ij}\, . \nonumber
\end{align}
In order to calculate the R\'{e}nyi\ entropies corresponding to a given region $A$ for a quadratic Hamiltonian, one only needs to calculate $X_{ij}$ and $P_{ij}$ for pairs of sites $i,j \in A$.\cite{Peschel2003}
These region-restricted correlation functions define the matrix $C_A \equiv \sqrt{X_A P_A}$, where $X_A$ and $P_A$ are defined as in Eq.~\eqref{eq:XP} but with $i$ and $j$ labelling the $N_A$ sites in $A$.
The von Neumann and R\'{e}nyi\ entropies are then given in terms of the eigenvalues $\nu_k$ of $C_A$ as\cite{Casini2009}
\begin{align}
S_1 (A) =
\sum_{k=1}^{N_A} \bigg[ & \left( \nu_k + \frac{1}{2} \right) \log \left( \nu_k + \frac{1}{2} \right) \\ \nonumber
&- \left( \nu_k - \frac{1}{2} \right) \log \left( \nu_k - \frac{1}{2} \right) \bigg]\, , \label{eq:vonNeumann_EE_eigvals}
\end{align}
and
\begin{equation}
S_\alpha(A) =
\frac{1}{\alpha-1} \sum_{k=1}^{N_A}
\log \left[ \left( \nu_k + \frac{1}{2} \right)^\alpha - \left( \nu_k - \frac{1}{2} \right)^\alpha \right]\, .\label{eq:Renyi_EE_eigvals}
\end{equation}
We note that the $\left\langle \phi_i \phi_j \right\rangle$ correlators --- and consequently the above entropies --- diverge in the case where the boson is massless ($m=0$) and the lattice has periodic boundary conditions (PBC) in all lattice directions.
Note however that since we are using these expressions as the ``cluster solver'' for the NLCE procedure,
which requires lattice clusters with open boundary conditions, these divergences do not pose a threat.
\subsection{Numerical linked-cluster expansion}
\label{sec:NLCE}
The NLCE is a powerful method that combines measurements of a suitable property on various finite-sized lattice clusters to obtain a sequence of approximations for the corresponding property in the thermodynamic limit $L \to \infty$.
At a given length scale or ``order" $\ell$, this numerical expansion uses sums and differences of finite clusters to systematically cancel off lower-order finite-size and boundary effects.
As a result, at a given order this procedure is capable of accessing longer-range correlations than direct calculations on finite toroidal systems of the same size.
This feature becomes especially advantageous when studying behavior at a critical point where the correlation length diverges.
For our present purposes, the NLCE offers the additional advantage that it can perform calculations in such a way as to isolate the corner contribution to the R\'{e}nyi\ entopies $v_{\alpha}\log(\ell/\delta)$ from both the edge contributions $w_\alpha\cdot (\ell/\delta)$ and the leading area law in Eq.~\eqref{eq:Spoly}.
In this section, we discuss general properties of the NLCE as well as the techniques necessary to isolate this three-dimensional corner contribution at each cluster order.
Analogous isolation techniques have been used with success to study the corner coefficient for various $(2+1)$-dimensional critical systems.\cite{Kallin2013,Kallin2014,Stoudenmire2014,Sahoo2016,Helmes2016}
At the most general level, the NLCE method can be used to study any property $\mathcal{P}$ that is
well defined in the thermodynamic limit (such as an extensive or an intensive property).
The NLCE calculates $\mathcal{P}$ for a lattice system $\mathcal{L}$ by summing
contributions from individual clusters according to
\begin{equation}
\mathcal{P}\left( \mathcal{L} \right) = \sum_{c} W(c)\, ,
\label{eq:NLCE_property}
\end{equation}
where the sum is over all clusters that can be embedded in the lattice
and $W(c)$ is the weight of the cluster.
This weight is defined recursively as
\begin{equation}
W(c) = \mathcal{P}(c) - \sum_{s \in c} W(s)\, ,
\label{eq:NLCE_weight}
\end{equation}
where the sum is over all subclusters $s$ contained in $c$.
This subgraph subtraction procedure elimintates from $W(c)$ the contributions to $\mathcal{P}(c)$ that have already been accounted for in the smaller subclusters.
Since we consider properties that are suitably normalized and have a well-defined thermodynamic limit ({\it i.e.,}\ extensive
or intensive properties), only connected (or linked) clusters have non-zero weight.
Hence the sums in Eqs.~\eqref{eq:NLCE_property} and~\eqref{eq:NLCE_weight} can be restricted to all linked clusters.
\begin{figure*}[ht]
\begin{center}
\def\svgwidth{\textwidth}
\input{lattice3D_cornerSum.pdf_tex}
\end{center}
\caption{The subtraction procedure used to calculate $\mathcal{P}_r(c)$ for the $3 \times 4 \times 4$ cluster $c$ and for a given vertex location $r$ (not pictured).
We add the values of $S_\alpha(A)$ corresponding to the four octants in the top row, subtract the values for the six quadrants in the middle row, and add the values for the three half-planes in the bottom row. We divide the resulting sum by four in order to obtain $\mathcal{P}_r(c)$. }
\label{fig:cuts}
\end{figure*}
In a translationally invariant system, all clusters that are related by translations have the same weights and make identical contributions to an extensive property $\mathcal{P}$.
For a given cluster, there are $N$ such clusters $c$ with the same weight such that the expression for $\mathcal{P}$ reduces to
\begin{equation}
\mathcal{P}\left( \mathcal{L} \right)/N = \sum_{c'} W(c') \, ,
\label{eq:NLCE_property2}
\end{equation}
where the clusters $c'$ are defined modulo translation.
One can often further reduce the number of clusters required in the above sums by exploiting other symmetries of the lattice system.
When the weights are the same for a given class of clusters, then one can write
\begin{equation}
\mathcal{P}\left( \mathcal{L} \right)/N = \sum_{c''} \mathbb{L}(c'') \times W(c'') \,,
\label{eq:NLCE_property3}
\end{equation}
where the sum is now over representative clusters $c''$ from each cluster class.
The quantity $\mathbb{L}(c'')$, called the lattice constant or the embedding factor of the cluster $c''$,
is the number of ways per lattice site that a cluster of class $c''$ can be embedded in the lattice.
So far the graphical basis for the NLCE is the same as for a series expansion in some variable,
such as the inverse temperature $\beta$ in the high temperature series expansion (HTSE).
However, unlike the HTSE where the goal is to maximize the order up to which the expansions are performed,
here the goal is to include contributions from representative clusters of maximal size.
Since the number of possible clusters grows rapidly with order, it is useful to further restrict the types of clusters considered in the expansions.\cite{Kallin2013}
Specifically, in $D$ spatial dimensions, every cluster can be uniquely associated with a $D$-dimensional rectangle (called a cuboid in
3D),
defined as the smallest volume in which the cluster can be fully embedded.
Thus, one can limit the calculations to cuboidal clusters only.
In three spatial dimension, one can then use Eq.~\eqref{eq:NLCE_property3}
with the sum restricted to regular $u_x \times u_y \times u_z$ cuboids with 6 faces, 8 vertices and 12 edges each.
Here $u_x$, $u_y$ and $u_z$ are integer lengths measured in units of the lattice spacing.
The enumeration of such clusters is trivial since their count (or embedding factor) $\mathbb{L}(c'')=1, 3 \text{ or } 6$ just depends on
the symmetry of the cuboid.
The subgraph counts of smaller cuboids in larger ones (needed in Eq.~\eqref{eq:NLCE_weight}) are also trivial. Because of the unique
association of each cluster with a cuboid the subgraph subtraction scheme works using cuboids only and thus the entire
computational burden is on calculating properties of finite cuboids.
Note that clusters in NLCE correspond to actual clusters
that can be embedded in the infinite lattice and therefore there is no periodic boundary condition on the finite clusters.
So far our discussion has focused on extensive properties, which get equal contributions from every region of the lattice.
However, the NLCE method applies equally well to an intensive property $\mathcal{P_\text{int}}$ that is defined with respect to some particular location in the lattice.
In this case, the clusters must be rooted with respect to
the special location where the property is defined.
As a result, the factor of $N$ that led to the simplification of Eq.~\eqref{eq:NLCE_property2} is no longer present.
However, one can create a corresponding extensive quantity by moving the localized quantity to different sites of the lattice and adding up the contributions for different locations.
In a translationally invariant system, the simplying factor of $N$ is then restored and such contributions are all the same so that $\mathcal{P_\text{ext}} = N \mathcal{P_\text{int}}$.
Then the left-hand side of Eqs.~\eqref{eq:NLCE_property2} and~\eqref{eq:NLCE_property3} is $\mathcal{P_\text{ext}} / N = \mathcal{P_\text{int}}$.
Similar to previous calculations in lower dimensions,\cite{Kallin2013,Kallin2014,Stoudenmire2014,Sahoo2016,Helmes2016} we define the desired intensive property $\mathcal{P}$ to be the isolated trihedral vertex contribution to the entanglement entropy $S_\alpha$.
We conceptually imagine a single vertex of interest as arising from an octant of the solid geometry in three spatial dimensions.
Then, each cuboidal cluster $c$ arising in the NLCE calculation is embedded in all possible ways around this vertex in order to create a corresponding extensive quantity as described above.
All possible rotations of each cluster are accounted for by $\mathbb{L}(c)$ in Eq~\eqref{eq:NLCE_property3}.
For a given cuboidal cluster, there are $(u_x -1)( u_y -1)( u_z - 1)$ possible locations $r$ for this vertex within
the cluster, which can be alternatively thought of as translating the cluster with respect to a fixed location of the
corner.
The overall contribution from the cluster is obtained by summing over all of these possible locations such that $N\mathcal{P}(c) = \sum_r \mathcal{P}_r(c)$.
To proceed with the calculation of the entanglement properties, we construct our clusters and subclusters to be regular
$u_x \times u_y \times u_z$ cuboids as described above.
For the corner entanglement entropy, there is no contribution from clusters
where any of $u_x$, $u_y$ or $u_z$ take value one.
We define the length scale (order) of a given cluster to be the maximum of these linear dimensions such that $\ell = \max\{u_x,u_y,u_z\}$.
Each cluster imposes Dirichlet open boundary conditions, with the field $\phi$ constrained to be zero for all lattice sites outside of the cluster.
Finally, in order to isolate the subleading trihedral vertex contribution to the R\'{e}nyi\ entropies, we perform a cluster-by-cluster subtraction procedure.
For each vertex location $r$ within the cluster $c$, we combine the values of $S_\alpha(A)$ corresponding to 13 different bipartitions $\{A, \overline{A}\}$ of the cluster as illustrated in Fig.~\ref{fig:cuts}.
This combination of R\'{e}nyi\ entopies allows us to intrinsically cancel the leading contributions from the area law (which are proportional to $B_\alpha \, {\ell^2}$) and the 90-degree edges (proportional to $ w_\alpha \, {\ell}$) such that $\mathcal{P}_r(c)$ (and, in turn, $\mathcal{P}(c)$) corresponds to
only the trihedral corner contribution to the entropy.
In practice, we take advantage of the symmetries present in the system in order to reduce the number of cluster bipartitions from 13 to 7.
The correlation function methods described in Section~\ref{sec:PeschelTrick} act as our cluster solver such that these methods are used to calculate all needed R\'{e}nyi\ entropies in the above sums for each cluster.
\section{Results}
\label{sec:Results}
In this section, we use the methods outlined in Section~\ref{sec:Methods} to calculate the trihedral corner coefficient in the R\'{e}nyi\ entropies of a massless free scalar in $(3+1)$-dimensions.
Using the NLCE procedure described in Section~\ref{sec:NLCE}, we isolate the corner contribution $\mathcal{P}_\alpha(\ell)$ to the R\'{e}nyi\ entropy $S_\alpha$ by performing calculations on clusters up to order $\ell$ (the maximum linear dimension of a given cluster).
In this section we examine the behavior of $\mathcal{P}_\alpha(\ell)$ as a function of $\ell$ with the goal of studying the vertex coefficient $v_{\alpha}$.
From Eq.~\eqref{eq:Spoly}, we expect for a single vertex,
\begin{equation}
\mathcal{P}_\alpha = v_{\alpha} \log \ell + d_\alpha + \ldots,
\label{eq:P_alpha_scaling}
\end{equation}
where $d_\alpha$ is a subleading constant and the ellipsis indicates additional (unknown) subleading terms that should vanish as $\ell \to \infty$ (or $\delta\to0$). Recall that $\ell$ is measured in units of the lattice spacing.
\begin{figure}[t]
\begin{center}
\includegraphics{fits_alpha1.pdf}
\includegraphics{fits_alpha5.pdf}
\end{center}
\caption{Fits of the corner contribution $\mathcal{P}_\alpha$ to the equation $v_{\alpha} \log \ell + d_\alpha$ for R\'{e}nyi\ indices $\alpha=1$ (top) and $\alpha=5$ (bottom).
The insets illustrate how the coefficients $v_{\alpha}$ extracted from these fits depend on the range of values of $\ell$.
For $\alpha>1$, we perform a second fit to extrapolate to the thermodynamic limit, as explained in the main text.
}
\label{fig:VAlpha_fits}
\end{figure}
\begin{figure*}[tp]
\begin{center}
\includegraphics{cornerCoeffVsAlpha_f1.pdf}
\end{center}
\caption{The normalized logarithmic corner coefficient $v_{\alpha}/v_1$ as a function of the R\'{e}nyi\ index $\alpha$.
These results are based on the fitting procedure illustrated in Fig.~\ref{fig:VAlpha_fits}.
For smooth surfaces, this ratio of logarithmic coefficients is known to behave as in Eq.~\ref{game1}.
}
\label{fig:cornerCoeff}
\end{figure*}
We first investigate what happens if we perform fits of $\mathcal{P}_\alpha$ to the two-parameter function $v_{\alpha} \log \ell + d_\alpha$ (ignoring, for the moment, additional subleading terms).
In Fig.~\ref{fig:VAlpha_fits}, we illustrate such fits for $\alpha = 1$ and $\alpha=5$.
We perform fits over various ranges of the cluster order $\ell$ and find that for $\alpha=1$, the extracted value of $v_1$ is quite stable when this range of $\ell$ values is varied --- indicating that the unknown subleading terms in Eq.~\eqref{eq:P_alpha_scaling} are already negligible for the cluster sizes used in our calculations.
However, for $\alpha >1$, the value of $v_{\alpha}$ increases significantly as higher orders $\ell$ are included in the fit and it is important to consider the effects of subleading terms.
Indeed, at least one source of additional finite-size scaling correction for $\alpha >1$ is known to arise from the conical singularity of the multi-sheeted
Riemann surface,\cite{Sahoo2016}
although the functional form of this correction is only known for $d=1+1$.\cite{Cardy_2010}
In Fig.~\ref{fig:cornerCoeff}, we show the results for $v_{\alpha}/v_1$ versus $\alpha$ as extracted from these various fits.
In order to approximate $v_{\alpha}$ in the thermodynamic limit $\ell \to \infty$, we study the behavior of $v_{\alpha}(\overline{\ell})$ versus $\overline{\ell}$, as illustrated in the insets of Fig.~\ref{fig:VAlpha_fits}.
Here $\overline{\ell}$ is a length scale that characterizes the orders $\ell$ used to extract $v_{\alpha}$ from the two-parameter fits described above.
We choose to define $\overline{\ell}$ as the average order --- such that, for instance, $\overline{\ell}=19$ for the case where orders $\ell = 18$ to 20 are used in the initial fit of $\mathcal{P}_\alpha$ to $v_{\alpha} \log \ell + d_\alpha$.
We could, however, use other definitions of $\overline{\ell}$ such as the minimum or maximum cluster order.
We then extract the behavior of $v_{\alpha}$ for $\overline{\ell} \to \infty$ by fitting $v_{\alpha}(\overline{\ell})$ to the three-parameter function $v_{\alpha}^\infty + p_\alpha/(\overline{\ell} + q_\alpha)$.
Here $v_{\alpha}^\infty$, $p_\alpha$ and $q_\alpha$ are (fitted) constants, where $v_{\alpha}^\infty$ corresponds to $v_{\alpha}$ in the thermodynamic limit and $q_\alpha$ reflects the ambiguity in the definition of $\overline{\ell}$ as described above.
This $\overline{\ell} \to \infty$ extrapolation procedure is used for all $\alpha >1$.
For $\alpha=1$, $v_{\alpha}(\overline{\ell})$ is well-converged as a function of $\overline{\ell}$ and we estimate $v_{\alpha}$ simply from the initial two-parameter fit using the highest orders available.
Fig.~\ref{fig:cornerCoeff} shows that as higher orders are used in the fits, the extracted normalized corner coefficient $v_{\alpha}/v_1$ as a function of $\alpha$ approaches the functional behavior predicted in Eq.~\reef{game1}.
Extrapolating to the infinite-size limit as described above appears to provide good agreement with this functional form, although we are not able to strictly quantify the agreement due to unknown finite-size errors within the NLCE procedure.
Hence, at least qualitatively, the $\alpha$-dependence agrees with that in Eq.~\req{game1}. Further, the latter dependence is identical to the behaviour of the universal coefficient appearing for a sphere, {\it i.e.,}\ $u_{\alpha}(S^2)/u_1(S^2)
=(\alpha+1)(\alpha^2+1)/(4\alpha^3)$. Hence this aspect of our results matches the observation in Ref.~\onlinecite{Devakul2014}, but we would also like to know how the overall coefficient in $v_{\alpha}$ relates to that of the spherical boundary, $u_\alpha(S^2)/8=-f_a(\alpha)/2$. Our result for $v_1$ reads
\begin{equation}
v_1=+0.00286\, ,
\eeq
which differs both in magnitude and sign from the coefficient for an eighth of a sphere, {\it i.e.,}\ $u_1(S^2)/8=-1/720\simeq -0.00139$.
More generally, if we fit our results for $v_{\alpha}$ versus $\alpha$ to the form $\xi\, u_\alpha(S^2)/8$ for constant $\xi$,
we find that
\begin{equation} \label{sss}
v_{\alpha} \simeq \xi \, \left[-\frac12\,f_a(\alpha)\right]\, ,\quad \text{with } \xi=-2.06\, .
\eeq
Hence, our trihedral corner result presents a very similar $\alpha$-dependence, but differs significantly in magnitude, and also in sign, from the result corresponding to an eighth of the sphere.
\section{Discussion} \label{discuss}
In this paper, we have performed numerical linked-cluster calculations to evaluate the universal coefficient $v_\alpha$ produced by a cubic trihedral corner in the R\'{e}nyi\ entropy of a massless free scalar field in $3+1$ dimensions.
Our results suggest that the dependence of this coefficient on the R\'enyi\ index $\alpha$ is well approximated by $v_{\alpha}/v_1 = (\alpha+1)(\alpha^2+1)/(4\alpha^3)$, which is the functional form expected for the universal coefficient appearing for a smooth spherical entangling surface.
However, despite this similarity, our numerical results demonstrate that the magnitude and the sign of the universal coefficient for an eighth of a sphere does not match $v_{\alpha}$, as had been suggested in Ref.~\onlinecite{Devakul2014}.
The $\alpha$-dependence of the universal coefficient for general CFTs in the case of smooth surfaces is controlled by two independent functions, $f_a(\alpha)$ and $f_b(\alpha)$. In the special case of a massless free scalar, these functions are proportional to each other, as shown in Eq.~\req{name}.
We have attempted to express the trihedral vertex coefficient $v_{\alpha}$ in terms of these two functions by using the general result valid for smooth surfaces in Eq.~\req{gamer} with a smoothed model of a cube in Section \ref{Scube}.
However, this calculation gave a result that was problematic in two respects.
First, it contained an unphysical $L/\delta\,\log( L/\delta)$ contribution, and second, the coefficient of the logarithmic term was regulator dependent.
As noted, we expect that these problems arise since we have not properly accounted for the $\mathcal{O}((L/\delta)^0)$ terms in Eq.~\reef{game1hh}.
In the singular limit $\varepsilon\to\gamma_1\delta/L$, some of these overlooked contributions build up to restore the desired universality of the logarithmic coefficient and to eliminate the unphysical $L/\delta\,\log(L/\delta)$ term.
The phenomenon where lower order ({\it i.e.,}\ less divergent) contributions in the R\'{e}nyi\ entropy can build up to produce universal terms with a stronger divergence in a limit where the entangling surface becomes singular has been explicitly seen in Ref.~\onlinecite{Bueno4}. In particular, this effect was described for the appearance of a sharp corner in $2+1$ dimensions, and of a conical singularity in $3+1$ dimensions.
In the first case, small deformations of a circular entangling surface produce a universal contribution to the entanglement entropy that scales as $(L/\delta)^0$.\cite{Mezei14}
These universal contributions can be evaluated for each Fourier mode on the circle, and when the Fourier modes are combined to produce a sharp corner, the sum of these contributions gives rise to a $\log(\ell/\delta)$ term.\cite{Bueno4}$^,$\footnote{These calculations are easily extended beyond $\alpha=1$ to general R\'{e}nyi\ entropies using the techniques introduced in Ref.~\onlinecite{displace}.}
In the second example, the logarithmic contribution in Eq.~\req{gamer} becomes a $\log^2(\ell/\delta)$ term when a conical singularity appears in an otherwise smooth spherical entangling surface.
However, there is an important difference between these two situations and the case of the trihedral corner considered here.
Both for the corner in $2+1$ dimensions and for the cone in $3+1$ dimensions, the order of the divergence corresponding to the universal term is higher for the singular surface than for the initial smooth surface. Hence, in the two cases studied in Ref.~\onlinecite{Bueno4}, the singular deformation changes the order of the divergence corresponding to the universal contribution, and hence the lower order contributions which are building up are in fact the universal contributions for the smooth entangling surfaces. Interestingly, this is not what happens for the trihedral corner, whose universal contribution is logarithmic, just like for a smooth entangling surface. Therefore, while it is natural to expect that some dependence on $f_{a}(\alpha)$ and $f_b(\alpha)$ survives in $v_{\alpha}$, it is also plausible that new contributions hidden in the $\mathcal{O}((L/\delta)^0)$ terms contribute.
It would be interesting to evaluate $v_{\alpha}$ for other CFTs in which $f_a(\alpha)$ and $f_b(\alpha)$ are independent ({\it i.e.,}\ rather than being proportional to one another, as in the free scalar theory) to gain a better understanding of the universal character of $v_{\alpha}$. On the one hand, the above discussion suggests that it may be premature to think that $v_{\alpha}$ is fully controlled by $f_a(\alpha)$ and $f_b(\alpha)$ alone. On the other hand, the $\alpha$ dependence of our numerical results is consistent with the trihedral coefficient being a simple linear combination
\begin{equation}
v_\alpha = \beta_a\,f_a(\alpha)+\beta_b\,f_b(\alpha)\,,
\label{linear}
\eeq
where $\beta_{a,b}$ are unspecified constants. Hence let us examine the latter possibility further. First, we substitute the relation that $f_b(\alpha)=3\,f_a(\alpha)$, which holds for a massless free scalar, into Eq.~\reef{linear}. Then combining the resulting expression with the fit in Eq.~\reef{sss} yields
\begin{equation}
\beta_a+3\,\beta_b\simeq 1.03\,.
\eeq
Further, we recall that $f_a(\alpha)$ multiplies a topological term in Eq.~\reef{gamer}. If we assume that the topological nature of the $f_a(\alpha)$ contribution survives for the trihedral coefficient, we would find
\begin{equation}
\beta_a=-\frac12\,,\qquad\beta_b\simeq0.51\,.
\label{guess}
\eeq
The numerical value of the second coefficient is remarkably close to being $1/2$ and hence these somewhat speculative steps are pointing to a rather simple result:
\begin{equation}
v_\alpha \overset{?}{=} \frac12\big[f_b(\alpha)-f_a(\alpha)\big]\,.
\label{guess2}
\eeq
Using the functions in Eq.~\reef{name} for the massless free scalar, the above expression predicts
\begin{equation}
{\rm free\ scalar:}\quad v_\alpha\overset{?}{=} \frac{(\alpha+1)(\alpha^2+1)}{1440\alpha^3}\,.
\eeq
Now the simplest CFT in which $f_a(\alpha)$ and $f_b(\alpha)$ are not proportional to one another is a free Weyl fermion. In this theory, the two functions are given by\cite{Fursaev:2012mp,Lee:2014xwa}$^,$\footnote{As in the case of the scalar, the result for $f^\mt{Weyl}_b(\alpha)$ in Eq.~\reef{name2} relies on the equality $f_b(\alpha)=f_c(\alpha)$. Ref.~\onlinecite{Lee:2014xwa} provided strong numerical evidence that this equality holds for the free fermion, but no analytic proof has yet been constructed in this case.}
\begin{eqnarray}
f^\mt{Weyl}_a(\alpha)&=&\frac{(\alpha+1)(37\alpha^2+7)}{5760\alpha^3}\, ,
\label{name2}\\
f^\mt{Weyl}_b(\alpha)&=&\frac{(\alpha+1)(17\alpha^2+7)}{1920\alpha^3}\, .
\nonumber
\end{eqnarray}
Combining these expressions with Eq.~\reef{guess2} then yields the surprisingly simple prediction:
\begin{equation}
{\rm Weyl\ fermion:}\quad v_\alpha \overset{?}{=} \frac{7}{4}\,\frac{(\alpha+1)(\alpha^2+1)}{1440\alpha^3}\,.
\label{name3}
\eeq
While this prediction relies on a number of unproven steps, it yields a very satisfying result. Namely, that the $\alpha$-dependence of $v_\alpha/v_1$ is {\it identical} for the free scalar and for the free fermion. We are currently extending our calculations to evaluate the trihedral corner coefficient $v_\alpha$ for the free fermion.
We expect that with the numerical accuracy achieved here, it will be straightforward to distinguish the scaling in Eq.~\reef{name3} from, {\it e.g.,}\ that of $f^\mt{Weyl}_a(\alpha)$ or $f^\mt{Weyl}_b(\alpha)$ alone.
Matching Eq.~\reef{name3} would provide strong evidence that $v_{\alpha}$ is fully determined by $f_a(\alpha)$ and $f_b(\alpha)$, and by Eq.~\reef{guess2} in particular.
On the other hand, disproving Eq.~\reef{name3} would suggest the trihedral corner provides new universal information beyond these two functions.
\begin{acknowledgments}
We are thankful to Grigory Bednik, Horacio Casini, Johannes Helmes, Veronika Hubeny,
Bohdan Kulchytskyy, Max Metlitski, Sharmistha Sahoo, and especially William Witczak-Krempa for stimulating discussions.
Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research \& Innovation. LEHS gratefully acknowledges funding from the Ontario Graduate Scholarship. The work of PB was supported by a postdoctoral fellowship from the Fund for Scientific Research -- Flanders (FWO).
PB also acknowledges support from the Delta ITP Visitors Programme. PB is grateful to the organizers of the ``It from Qubit Summer School'' held at the Perimeter Institute for Theoretical Physics and to the organizers of the ``Quantum Matter, Spacetime and Information'' conference held at the Yukawa Institute for Theoretical Physics (YITP) at Kyoto University.
The work of RRPS is supported by the US National Science Foundation
grant number DMR-1306048.
RGM is supported in part by funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) and a Canada Research Chair.
RCM is supported in part by funding from NSERC, from the Canadian Institute for Advanced Research and from the Simons Foundation through the ``It from Qubit'' collaboration.
\end{acknowledgments}
|
\subsection{Introduction}
The dynamics of a quantum system coupled to an external environment (so-called ``open'' quantum system) are a result of the intricate interplay between the coherent evolution of the quantum degrees of freedom, the structure of the environment and the form and strength of the system-environment coupling. From quantum cavities \cite{ cavity1, cavity2, cavity3} and superconducting qubits \cite{supercond1,supercond2,supercond3,supercond4,supercond5,supercond6} , through quantum dots \cite{quantumdot1,quantumdot2,quantumdot3,quantumdot4,Gruss2016,Ajisaka2013}, molecular junctions \cite{Peskin2016, Peskin2010, Thingna2016, Segal2016, Ajisaka2015, Wang2015, Zelovich2015, Elenewski2017,Esposito2010,Hod2016, Rudge2016, Dubi2013} and cold atoms \cite{coldatoms1,coldatoms2,coldatoms3,coldatoms4} , to excitons traveling in photosynthetic complexes \cite{Mohseni2014,Lambert2013,Fleming2011,Scholes2005}, open quantum systems show dynamics which can be far richer and more surprising than their coherent (environment-free) counterparts.
Recent ideas of designing a quantum state by engineering a specific environment \cite{KrausPRA2008,BarreiroNatPhys2010,Kienzler2015,Verstraete2009,
Diehl2008Nat,Bardyn2013NJP,Vorverg2013PRL} or by a periodic modulation of the open system's Hamiltonian \cite{Volokitinarxiv2016,Hartmannarxiv2016,Peskin2016} open a path to new forms of control over quantum systems.
Combining these two concepts of time-periodic modulations and environment (bath) engineering, here we study the dynamics of open quantum systems where the environment thermodynamic parameters are periodically modulated. Even though there exists formally exact methods of treating such set-ups \cite{exact1,exact2,exact3,exact4,exact5,exact6,exact7,exact8,exact9}, they are generally quite difficult to use, bein computationally demanding. To simplify calculations, often the adiabatic approximation is used \cite{adiabatic1, adiabatic2, adiabatic3}. The adiabatic approximation assumes the environment parameters are modulated very slowly such that the system is always at a non-equilibrium steady state with the instantaneous environment. This approximation fails to capture dynamical effects which arise due to the time-periodic modulations. In this work, we therefore seek to go beyond the adiabatic approximation.
For weak system-bath coupling, a combination of Quantum Master Equation methods (typically of the Lindblad form) with Floquet theory is often used to go beyond the adiabatic approximation \cite{floquet1,floquet2,floquet3,floquet4,floquet5,floquet6}. However, Floquet theory has the drawback of converting a finite-dimensional problem to an infinite dimensional one. Consequently, most interesting results from Floquet theory are often obtained as a perturbation expansion in terms of inverse drive frequency, giving results primarily for high frequency driving.
Moreover, from calculations in the absence of time-periodic modulations, the commonly used phenomenological Lindblad Quantum Master Equations are known to have several drawbacks, especially in an out-of-equilibrium situation \cite{Purkayastha2016,QME1,QME2,QME3}. It has been recently shown that microscopically derived Redfield Quantum Master Equation (RQME) under Born-Markov approximation can be used to overcome the drawbacks of phenomenological Lindblad equations and to obtain correct results up-to leading order in system-bath coupling \cite{Purkayastha2016}. To our knowledge, there has been no work in extending the RQME to the case when thermodynamic parameters of the baths are periodically modulated.
In this work, we derive a Born-Markov approximated Redfield theory for a quantum system consisting of non-interacting particles (bosons or fermions) on a lattice of arbitrary dimension and geometry, weakly coupled to multiple periodically-modulated baths. The baths are modeled by non-interacting particles (bosons or fermions) with infinite degrees of freedom. For a lattice of $N$ sites, this approach leads to a closed system of $N^2$ linear differential equations which can be easily solved numerically in time domain, thereby avoiding the drawback of infinite dimensional theory like Floquet theory. We apply our formalism to a generic example of an open quantum system, namely a double quantum-dot coupled to two fermionic reservoirs, characterized by periodically modulated chemical potentials. The current through the system depends on both the chemical potential modulation properties and the chemical potential difference (i.e. voltage bias) between the electrodes. We show that even when there is no bias between the electrodes the dynamics of the system are non-trivial and reveal the internal structure of the junction (i.e. the internal spectrum of the dot and the asymmetry in the junction). When there is a voltage bias between the electrodes, the modulation induces unusual hysteresis behavior. These predictions can in principle be tested experimentally, demonstrating the applicability of our formalism to realistic open quantum systems.
\subsection{Set-up and protocol}\label{B}
Consider a general non-interacting tight-binding Hamiltonians for the whole system+bath set-up :
\begin{align}
\label{model_H}
\mathcal{H} &= \mathcal{H}_S + \mathcal{H}_B + \mathcal{H}_{SB}~,~~{\rm where} \\
~\mathcal{H}_S &= \sum_{\ell=1}^N H^{(s)}_{\ell m} \hat{a}_\ell^{\dagger} \hat{a}_m~,~~~
\mathcal{H}_B = \sum_{\ell=1}^N\sum_{r=1}^\infty \Omega_{\ell r} \hat{B}_{\ell r}^{\dagger} \hat{B}_{\ell r}~, \nonumber\\
\mathcal{H}_{SB} &= \varepsilon\sum_{\ell=1}^N \sum_{r} (\kappa_{\ell r} \hat{B}_{\ell r}^{\dagger} \hat{a}_\ell + \kappa_{\ell r}^* \hat{a}_\ell^{\dagger} \hat{B}_{\ell r})~.
\end{align}
$H^{(s)}$ is a Hermitian matrix and $\hat{a}_\ell$ correspond to fermionic (bosonic) annihilation operators defined respectively on $\ell$th lattice point of system and $\hat{B}_{\ell r}$ to those of baths attached to the $\ell$th point.
The baths have infinite degrees of freedom. $\varepsilon$ is a dimensionless parameter that controls system bath coupling, so that $\{\kappa_{\ell r}\}$ have dimensions of energy. Here each lattice point is connected to a bath. By setting system-bath coupling to zero at an arbitrary number of lattice points, any geometry of system-bath connections can be achieved. We also introduce the bath spectral functions:
\begin{equation}
\label{J}
{J}_{\ell}(\omega)=2\pi \sum_r \mid \kappa_{\ell r} \mid^2 \delta(\omega - \Omega_{\ell r})
\end{equation}
Note that the full system+bath Hamiltonian (\ref{model_H}) is time-independent. In the conventional dc bias case, for weak system-bath coupling, the baths are always assumed to be in thermal equilibrium with their own corresponding temperature and chemical potential. In other words, $Tr_B(\hat{B}_{\ell r}^{\dagger} \hat{B}_{\ell r})$, where $Tr_B(..)$ implies trace taken over bath degrees of freedom, is given by the corresponding fermi (bose) distribution $\Big [ e^{-\beta_\ell(\Omega_{\ell r}-\mu_\ell)}\pm1 \Big ]^{-1}$. Here, we are interested in the case when the inverse temperature $\beta_\ell$ and chemical potential $\mu_\ell$ are periodic functions of time. At this stage, one may want to model such a set-up by deriving a quantum master equation for the density matrix of the system for the time independent case and making the fermi (bose) distributions time periodic `by hand'. However, the regime of validity of such an approach will not be clear. So, instead, in the following, we assume a protocol, and make physical assumptions so that such an equation may be reached. This shows how and under what conditions such a situation can arise physically.
We now describe the protocol. In the following, $\chi$ is the full density matrix of system+bath, $\rho_B$ is the density matrix of the bath, and $\rho$ is density matrix of the system obtained by tracing $\chi$ over bath variables. Let us, for the time being, assume one bath. The protocol can readily be generalized to multiple baths.
The protocol for a single bath is as follows :
a) At time $t=0$, $\chi = \rho \otimes \rho_B(0)$, with $\rho=\rho_0(0)$, which is some arbitrary initial system state, and $\rho_B(0) = {exp[-\beta(0)(\mathcal{H}_B-\mu(0)\mathcal{N})]}/{Z(\tau_S)}$. That is, the inital state is a product state of an arbitrary system state and a thermal bath state.
b) We evolve the system for a time $\tau_D$. After time $\tau_D$, the system reaches the state $\rho_0(\tau_D)$. Note that during this time, the temperatures and chemical potentials of the bath has not changed.
c) At $t=\tau_D$, we restart the entire system+bath setup with the initial state $\chi = \rho \otimes \rho_B(\tau_D)$, with $\rho=\rho_1(0)=\rho_0(\tau_D)$ and $\rho_B(\tau_D) = {exp[-\beta(\tau_D)(\mathcal{H}_B-\mu(\tau_D)\mathcal{N})]}/{Z(\tau_D)}$. That is, at time $\tau_D$, the bath is changed into the thermal state with new inverse temperature and chemical potential $\beta(\tau_D), \mu(\tau_D)$, and the full system+bath state is again taken as the product state. This step implicitly assumes weak-system bath coupling. This is because, under weak system-bath coupling, we can assume that the bath is hardly affected by the system, and to leading order in system-bath coupling, the full density matrix is in product form.
d) We again let the system density matrix evolve under this new bath for time $\tau_D$ starting from the new initial state. Again, after time $\tau_D$, we restart the entire set-up with initial state $\chi = \rho(0) \otimes \rho_B(2\tau_D)$, with $\rho=\rho_2(0)=\rho_1(\tau_D)$ and $\rho_B(2\tau_D) = {exp[-\beta(2\tau_D)(\mathcal{H}_B-\mu(2\tau_D)\mathcal{N})]}/{Z(2\tau_D)}$.
This protocol continues for a very long time and we will be mostly interested in the long time dynamics of this process.
In the limit of very small $\tau_D$, the above protocol gives a nearly continuous evolution of temperatures and chemical potentials of the baths. In particular, if the temperature and chemical potential vary periodically with a period $T$, and $\tau_D \ll T$, then the above protocol describes dynamics of a system evolving under a continuous periodic drive from the temperature and chemical potential of the bath.
To make analytical progress with this problem, we need to make one more assumption, the Markov assumption. Let $\tau_B$ be the characteristic time-scale of relaxation of the effects of system-bath coupling on the bath. Then, we assume that $\tau_B \ll \tau_D$. It is only under this condition that microscopically derived Markovian Quantum Master Equations can be applied. So, if $\tau_{expt}$ be experimentally the smallest time scale, most of our following discussion holds if the following condition on time-scales is maintained :
\begin{align}
\label{t_cond1}
\tau_B \ll \tau_D \ll \tau_{expt} \ll T
\end{align}
It can be shown that $\tau_B$ depends on the temperature of the bath, $\tau_B \sim \beta$ (see Appendix~\ref{Appendix_B}). Hence, our analytical discussion below will not be valid at extremely low temperatures. It is only valid when temperature is large enough so that above condition on time scales can be satisfied.
For multiple baths, the above protocol is followed for each bath. $\tau_B$ is then taken as the largest of the relaxation times of the baths, so that $\tau_D$ is much larger than relaxation time scales of all baths. When the system is driven by multiple baths according to our protocol, we call it an ac drive process. Correspondingly, the usual case of having a time independent temperature or chemical potential difference between baths is called a dc bias process.
The above protocol, along with the condition (Eq.~\ref{t_cond1}) on time scales, breaks down the ac drive process into steps of time independent processes where Born-Markov approximation can be applied. For such time independent processes, the Born-Makov approximated Redfield Qunatum Master Equation (RQME), as well as the evolution equation for two point correlation functions from the RQME have been derived for Hamiltonian (\ref{model_H}) recently \cite{Purkayastha2016} (see Appendix~\ref{Appendix_A}). Going through the above protocol with the RQME only has the effect of making the fermi (bose) distribution functions time-dependent with the instantaneous temperatures and chemical potentials (see Appendix~\ref{Appendix_C}). The final result, as desired, will be same as that obtained by `putting by hand' time dependent fermi of bose distribution functions in the RQME derived without such time dependence. However, it is important to note that, the description via such an equation would not be possible if each step of our protocol would not satisfy the Markov condition $\tau_B \ll \tau_D$. Hence, only when system-bath coupling is weak and the condition on time scales in Eq.~\ref{t_cond1} is satisfied, can our set-up and protocol be described by such an equation. This crucial fact would not be clear in a `putting by hand' approach. Also, for small systems, the protocol can easily be exactly (i.e, without Markov approx) simulated with finite but large baths, and hence allows for numerical validation of results.
\subsection{Redfield equation for the correlation matrix (single particle density matrix)}\label{C}
Since the system is quadratic, various physical observables like current and occupation can be directly calculated from the two-point correlation functions. Using the results in Ref.~\cite{Peskin2016} (which are also re-derived in Appendix), we can readily write down the evolution equation for two-point correlation functions for our set-up. For this, it is convenient to go to the eigenbasis of the system Hamiltonian. Let $c$ be the unitary matrix which diagonalizes $H^{(S)}$, i.e., \begin{equation}
c^\dagger H^{(S)} c= \omega^{(D)}~,
\end{equation}
where $c^\dagger c=I$ and $\omega^{(D)}$ is a diagonal matrix with elements $\omega_\nu$. Then we also define new operators $\{A_\alpha\}$ through the transformation \begin{equation}
\hat{a}_{\ell} = \sum_{\alpha=1}^N c_{\ell \alpha} \hat{A}_\alpha~.
\label{Aop}
\end{equation}
Thus $\hat{A}_\alpha$ is the annihilation operator for the $\alpha$th eigen-mode with energy $\omega_\alpha$.
The evolution equation for the equal time two point correlation functions $C_{\alpha\nu}(t) = \langle \hat{A}_\alpha^\dagger(t) \hat{A}_\nu(t) \rangle = Tr\big( \rho\hat{A}_\alpha^\dagger(t) \hat{A}_\nu(t) \big) $ can be derived by following the derivation of the Redfield equation for the density matrix $\rho$ \cite{breuer2007theory,CarmichaelBook,Nitzan2006} and substituting it back to the definition of the correlation functions \cite{Purkayastha2016}(see Appendix~\ref{Appendix_A}). The protocol, along with Born-Markov approximation, only makes the fermi (bose) distributions time-dependent (see Appendix~\ref{Appendix_B}). The resulting equation is
\begin{align}
\label{C_differential}
&\frac{dC_{\alpha\nu}}{dt} = \Big [ i\omega_\alpha C_{\alpha\nu}(t) -\varepsilon^2 \sum_{\sigma=1}^N C_{\alpha\sigma}(t) v_{\nu \sigma }+(\alpha \leftrightarrow \nu )^\dagger \Big ]\nonumber \\
&+ \varepsilon^2 Q_{\alpha\nu}(t)~, \\
&~~{\rm where}, \nonumber \\
& v_{\alpha \nu}=f_{\alpha\nu}(\omega_{\nu})-if_{\alpha\nu}^{\Delta}(\omega_{\nu}) \nonumber \\
& Q_{\alpha\nu}(t) = [F_{\nu\alpha}(\omega_\alpha,t)-iF_{\nu\alpha}^{\Delta}(\omega_\alpha,t)+(\alpha \leftrightarrow \nu )^*] \nonumber \\
&f_{\alpha\nu}(\omega) = \sum_{{\ell}=1}^Nc_{\ell \alpha}^*c_{\ell \nu} \frac{J_{\ell}(\omega)}{2}, \nonumber \\
& F_{\alpha\nu}(\omega,t) = \sum_{{\ell}=1}^N c_{\ell \alpha}^*c_{\ell \nu} \frac{J_{\ell}(\omega)n_{\ell}(\omega,t)}{2} \nonumber
\end{align}
with $f_{\alpha\nu}^{\Delta}(\omega) = \mathcal{P} \int \frac{}{}\frac{d\omega^{\prime} f_{\alpha\nu}(\omega^{\prime})}{\pi(\omega^{\prime}-\omega)}, \hspace{2pt}
F_{\alpha\nu}^{\Delta}(\omega,t) = ~ \mathcal{P} \int \frac{}{}\frac{d\omega^{\prime} F_{\alpha\nu}(\omega^{\prime},t)}{\pi(\omega^{\prime}-\omega)}$, where $\mathcal{P}$ denotes principal value, $n_\ell(\omega,t)~=~\Big [ e^{-\beta_\ell(t)(\omega-\mu_\ell(t))}\pm1 \Big ]^{-1}$ is the instantaneous fermi or bose distribution function and $J_\ell(\omega)$ is the spectral function (Eq.~\ref{J}) of bath (lead) attached to $\ell$th site of the system. Weak system-bath coupling requires $\varepsilon^2 J_{\ell}(\omega) \ll \{\omega_{\alpha} \}$.
Eq.~\ref{C_differential} gives a closed set of $N^2$ linear differential equations. The matrix with elements $C_{\alpha\nu}$ is sometimes also called the single particle density matrix of the system. Writing $C$ and $Q$ as vectors, this equation can be cast in the form
\begin{align}
\label{C_vector_differential}
&\frac{dC}{dt}=-MC+ \varepsilon^2 Q(t),\\
&{\rm where} \nonumber \\
& M = \mathbb{I}_N\otimes G^* + G \otimes \mathbb{I}_N, \nonumber \\
& G_{\alpha\nu} = \varepsilon^2 v_{\alpha\nu}^* - i\omega_{\nu}\delta_{\alpha\nu} \nonumber
\end{align}
$\mathbb{I}_N$ is the $N$ dimensional identity matrix and $\otimes$ denotes Kronecker product. Eq.~\ref{C_vector_differential} has the formal solution
\begin{align}
\label{C_vec_t}
C(t) = e^{-Mt}C(0) + \varepsilon^2\int_0^t dt^\prime e^{-M(t-t^\prime)}Q(t^\prime)
\end{align}
Note that the matrix $M$ has no time dependence and hence is the same matrix as would appear in the dc problem. If under dc bias the system reaches a steady state after a long time, the real part of eigenvalues of the matrix $M$ has to be positive. We will assume this henceforth.
We wish to look at the long time properties of Eq.~\ref{C_vec_t}.
Let our ac drive be periodic with a time period $T$. Then, for integer $r$,
\begin{align}
Q(t+rT) = Q(t), \hspace{5pt} r\in \mathbb{Z}
\end{align}
We break up the observation time $t$ into steps of $T$, so that, for integer $m$,
\begin{align}
t=mT+\tau \hspace{5pt} m\in \mathbb{Z}
\end{align}
Then Eq.~\ref{C_vec_t} can be written as
\begin{align}
&C(mT+\tau) = \varepsilon^2\sum_{r=1}^m \Big ( e^{-MrT} \Big ) \int_0^T dt^\prime e^{-M(\tau-t^\prime)}Q(t^\prime) \nonumber \\
&+\varepsilon^2\int_0^\tau dt^\prime e^{-M(\tau-t^\prime)}Q(t^\prime)+e^{-M(mT+\tau)}C(0)
\end{align}
Now, since we have assumed that the real part of all the eigenvalues of $M$ are positive, we can perform the sum in above equation. Also, we are interested in the long time limit, $m \gg 1$. Hence we have
\begin{align}
\label{C_vec_steady}
&C(mT+\tau) = \varepsilon^2(e^{MT}-\mathbb{I}_{N^2})^{-1} \int_0^T dt^\prime e^{-M(\tau-t^\prime)}Q(t^\prime) \nonumber \\
&+\varepsilon^2\int_0^\tau dt^\prime e^{-M(\tau-t^\prime)}Q(t^\prime)
\end{align}
Note that, in the long time limit, the RHS is independent of $m$, and is also independent of the initial condition. This means that in the long time limit, the correlation functions settle down in a periodic state, with period same as the ac drive. This is consistent with Floquet theory. However, note that, Floquet theory was not used explicitly to derive this result. Eq.~\ref{C_vec_steady} is the central result of the manuscript.
\subsection{Resonances and Currents}\label{D}
Now, let us go back to Eq.~\ref{C_differential}. Since in the long time limit the periods of $C(t)$ and $Q(t)$ are same, we can perform a Fourier series expansion :
\begin{align}
C_{\alpha\nu}(t) = \sum_{p=-\infty}^\infty C_{\alpha\nu}^p e^{ip\Omega_0 t}, \hspace{2pt} Q_{\alpha\nu}(t) = \sum_{p=-\infty}^\infty Q_{\alpha\nu}^p e^{ip\Omega_0 t}
\end{align}
where $\Omega_0=\frac{2\pi}{T}$. Substituting in Eq.~\ref{C_differential}, we obtain the following equation for each Fourier mode,
\begin{align}
\label{C_Fourier}
&i(\omega_\alpha -\omega_\nu-p\Omega_0) C_{\alpha\nu}^p + \varepsilon^2 Q_{\alpha\nu}^p \nonumber \\
& -\varepsilon^2 \Big [ \sum_{\sigma=1}^N C_{\alpha\sigma}^p v_{\nu \sigma }+C_{\sigma\nu}^p v_{\alpha\sigma }^*~ \Big ] =0
\end{align}
This immediately gives us two important results. First, we note that for the time-average correlation functions, $\overline{C}_{\alpha\nu}(t) = \frac{1}{T}\int_t^{t+T} dt^\prime C_{\alpha\nu}(t^\prime)=C_{\alpha\nu}^0$, and similarly for $Q_{\alpha\nu}$. With this we find that the equation averaged over one time period gives exactly the steady state equation for a dc bias given by the time-period averaged bose or fermi distribution functions $\overline{n_\ell(\omega,t)}$.
Second, we see that when $\omega_\alpha -\omega_\nu=p\Omega_0$, Eq.~\ref{C_Fourier} becomes independent of system-bath coupling $\varepsilon^2$. This is the phenomenon of resonance. At resonance, the leading term of system correlation functions do not depend on system bath coupling. So all system properties, like current between two sites inside the system, which were otherwise proportional to $\varepsilon^2$ in the leading term, become larger by orders of magnitude.
Even though system properties become large at resonance, the current from any of the baths still remain small. To see this, we need the expression for current from the baths in terms of the correlation functions. The expression for particle current is obtained from the continuity equation for conservation of particles
\begin{align}
\label{current_B}
&\frac{d}{dt}\Big (\sum_{\alpha=1}^N C_{\alpha\alpha} \Big ) = \sum_{\ell} I_{B^{(\ell)}\rightarrow \ell}
\end{align}
$I_{B^{(\ell)}\rightarrow \ell}$ is the particle current from bath attached to the $\ell$th site. Under dc bias, in steady state, the LHS of above equation is zero, and hence the currents from the baths are equal. However, under ac drive, even at long time, the LHS is not zero, and thus, instantaneous currents from the baths can be different. The summation on the right runs over all lattice sites connected to bath. The expressions for the currents from the baths obtained from above continuity equations and Eq.~\ref{C_differential} are
\begin{align}
& I_{B^{(\ell)}\rightarrow \ell} = \varepsilon^2 \Big [ \sum_{\alpha=1}^N Q_{\alpha}^{(\ell)}(t) - \sum_{\sigma,\alpha=1}^N C_{\alpha\sigma}(t)\big(v_{\alpha\sigma}^{(\ell)} + v_{\sigma\alpha}^{(\ell)*}\big) \Big ]
\end{align}
where
\begin{align}
&Q_{\alpha}^{(\ell)}(t) = |c_{\alpha \ell}|^2 \mathfrak{J}_\ell(\omega_\alpha)n_\ell(\omega_\alpha) \nonumber \\
&v_{\alpha\sigma}^{(\ell)} = c_{\alpha \ell}^* c_{\sigma \ell} \Big(\frac{\mathfrak{J}_\ell(\omega_\sigma)}{2}- i \mathcal{P} \int \frac{d\omega}{2\pi}\frac{\mathfrak{J}_\ell(\omega)}{\omega-\omega_\sigma} \Big )
\end{align}
These expressions show that the currents from the baths are explicitly proportional to $\varepsilon^2$. So even when the system correlation functions are independent of $\varepsilon^2$ in the leading order, the currents from the baths are still $O(\varepsilon^2)$. However, this is not true for particle current between two sites which reside \emph{within} the system.
The current between $\ell$th and $\ell+1$th lattice sites of the system is given by
\begin{align}
\label{current_S}
I_{\ell \rightarrow \ell+1} = 2 \hspace{2pt} Im \Big ( H^{(s)}_{\ell \hspace{2pt} \ell+1} \hat{a}_\ell^{\dagger} \hat{a}_{\ell+1} \Big )
\end{align}
This current is not explicitly proportional to $\varepsilon^2$. At resonance, it is independent of $\varepsilon^2$ in the leading order. Thus, at resonance, the current between two adjacent lattice points in the system can be much larger than the current from the baths. On the other hand, the time-period averaged current, which corresponds to steady state of a dc bias, is same inside the system as from the baths.
The difference in frequency between adjacent resonances, which is given by
\begin{align}
\frac{\omega_\alpha-\omega_\nu}{p} -\frac{\omega_\alpha-\omega_\nu}{p+1} = \frac{\omega_\alpha-\omega_\nu}{p(p+1)} \sim \frac{1}{p^2}
\end{align}
decreases as $1/p^2$. Hence small driving frequencies are always close to resonance with one of the higher modes (large $p$) of the steady state oscillations. So, for small driving frequencies, there is not much difference between the resonance and off-resonance condition. The values of the correlation functions, however, may not be so large as the first resonance ($p=1$). This is because, the weight of the driving signal at higher modes may decrease.
It is also interesting to note that none of the expressions for currents can be reduced to the form of difference between fermi (bose) distributions of the various baths. Therefore, even when all the baths are driven by the exactly same time dependent temperature or chemical potential (symmetric ac drive), so that there is no instantaneous temperature or chemical potential difference, there can be an instantaneous current, both between the system and the baths and inside the system. At resonance, the internal currents may have a large amplitude (compared to the system-bath currents). The time-period averaged current is, of course, zero in this case.
It is important to state that we have \textit{not} made the ubiquitous adiabatic approximation $T \gg t_{steady}$, where $t_{steady}$ is the time to reach steady state, which corresponds to the smallest real part of eigenvalues of the matrix $M$ in Eq.~\ref{C_vector_differential}. If adiabatic approximation were made, then the expressions for the correlation functions at any time would be given by the dc-bias steady state results with the fermi (bose) distributions given by the instantaneous temperatures and chemical potentials. In that case, the expressions for currents \textit{would have} reduced to the form of difference between fermi (bose) distributions of the various baths, and \textit{no instantaneous current} would have been seen in the symmetric ac drive case.
Under dc bias, the steady state can be quite easily obtained for non-interacting open systems described by Hamiltonian of the form Eq.~\ref{model_H} by exact methods (i.e, not under Born-Markov approximation). Transient dynamics of approach to steady state, on the other hand, is generally quite difficult to calculate from such methods. From our protocol, it can be seen that when, $T \lesssim t_{steady}$ (i.e, when adiabatic approximation is not valid), it is the transient behaviour of approach to steady state for dc bias that becomes important to describe the ac driven case. Therefore, such exact methods are difficult to use in our ac drive set-up. However, as has been recently established, RQME gives the correct time dynamics of approach to steady state under dc bias of two-point correlation functions for our set-up as long as Born-Markov approximation is satisfied \cite{Purkayastha2016}. This fact has allowed us to use RQME for ac drive case to go beyond the adiabatic approximation easily.
Recently, another method based on a quantum master equation called the Discrete Lioville von-Neumann (DLvN) method has been formulated \cite{DLvN1,DLvN2,DLvN3}, which has been shown to give accurate results for time-dependent behaviour \cite{DLvN3}. However, in this method one models the leads via finite but large size Hamiltonians, and needs to solve explicitly for the leads along with the system. As a result, if there are $N_L$ sites in left lead, $N_R$ sites in the right lead and $N$ sites in the system, then one needs to solve for a closed set of $(N_L + N_R + N)^2$ linear differential equations. In comparison, using our RQME method, the leads are considered implicitly in the spectral functions (Eq.~\ref{J}) and one needs to solve only for a closed set of $N^2$ linear differential equations (see Eqs.~\ref{C_differential} and \ref{C_vector_differential}).
RQME has the drawback of not being completely positive because of not being of the Linblad form. As a result, under certain initial conditions, it may lead to a non-physical state at short times. In fact, the rigorous microscopic derivation of the dc bias Redfield equation for quadratic baths requires the observation time $t$ to satisfy $t\gg \tau_B$ (see Appendix~\ref{Appendix_A}). Thus, it is only in this regime of observation times that the results should be valid. As a consequence, one expects that, problems regarding non-positivity show up only at short times ($t\leq \tau_B$), where the Redfield equation itself is not justified \cite{Eastham2016}. Based on this, it is expected that, in our ac drive set-up, when the condition on time scales in Eq.~\ref{t_cond1} is satisfied, problems regarding positivity will not arise. This, however, is not a completely rigorous statement, and requires further investigation. In Ref \cite{Purkayastha2016}, the long time behaviour of the correlation functions for the dc bias case of our set-up have been shown to be given correctly by the Redfield equation. Since our protocol breaks down the ac drive process into a series of such dc drive steps, this gives further support regarding validity of our results. But, as mentioned before, since $\tau_B$ is inversely proportional to temperature (see Appendix~\ref{Appendix_B}), our theory may not be valid at extremely low temperatures.
\subsection{A detailed example: the $N=2$ case}\label{E}
All the above results are quite general and hold for non-interacting system of bosons or fermions in lattice of any dimension and geometry. In the following, to validate the theory, as well as to better understand the physics of such ac drive, we apply this theory to a simple model of two fermionic lattice sites, each connected to their own baths. This generic model can be used to describe physical systems such as driven double quantum dots\cite{Stehlik2016} or single-molecule junctions with two molecular moieties (e.g. biphenyl-dithiol molecular junctions \cite{Venkataraman2006}).
We consider the following specific two-site system coupled to baths which are one-dimensional chains:
\begin{align}
\hat{\mathcal{H}}_S &= \omega_{0} (\hat{a}_1^{\dagger} \hat{a}_1+\hat{a}_2^{\dagger} \hat{a}_2) + g(\hat{a}_1^{\dagger} \hat{a}_{2} + \hat{a}_{2}^{\dagger} \hat{a}_1)~, \nonumber \\
\hat{\mathcal{H}}_B^{(\ell)}&= t_B (\sum_{s=1}^\infty \hat{b}_s^{\ell\dagger}\hat{b}_{s+1}^\ell+h.c.), \hspace{5pt} \hat{\mathcal{H}}_B = \hat{\mathcal{H}}_B^{(1)}+\hat{\mathcal{H}}_B^{(2)}~,\nonumber \\
\hat{\mathcal{H}}_{SB} &= \varepsilon\gamma_1 (\hat{a}_1^{\dagger}\hat{b}_{1}^1+ h.c.) + \varepsilon\gamma_2 (\hat{a}_2^{\dagger}\hat{b}_{1}^2+ h.c.)~, \label{ham2S}
\end{align}
where the operators are all fermionic and $\hat{b}_{s}^\ell$ is the annihilation operator of the $s$th bath site of the $\ell$th bath. The eigenmodes of the system are given by $\hat{A}_1 = {(\hat{a}_1-\hat{a}_2)}/{\sqrt{2}}$, $\hat{A}_2 = {(\hat{a}_1+\hat{a}_2)}/{\sqrt{2}}$ with eigenvalues $\omega_1 = \omega_0 - g$, $\omega_2 = \omega_0 + g$. The bath spectral functions, defined in Eq.~\ref{J}, can be obtained explicitly by going to eigenmodes of the baths and are given by \cite{Purkayastha2016,Peskin2016} (see Appendix~\ref{Appendix_D})
\begin{equation}
\mathfrak{J}_\ell(\omega) = \Gamma_\ell\sqrt{1-\left(\frac{\omega}{2t_B}\right)^2}~, \hspace{5pt} \Gamma_\ell=\frac{2\gamma_\ell^2}{t_B}~.
\label{Jform}
\end{equation}
$\omega_0 \gg (\varepsilon^2/t_B)$ so that QME can be applied while the parameter $g$ can be varied freely. The two baths are taken at the same temperature. Inverse temperature $\beta$ is taken to be constant and the ac drive is given by periodic chemical potential :
\begin{align}
n_\ell(\omega,t) = [e^{\beta(\omega-\mu_\ell(t))}+ 1]^{-1}, \hspace{5pt} \mu_\ell(t+rT)=\mu_\ell(t) \hspace{2pt} \forall r\in \mathbb{Z}
\end{align}
We choose the drive frequency as
\begin{align}
\label{p_def}
\Omega_0 = \frac{\omega_2 - \omega_1}{p} = \frac{2g}{p}
\end{align}
If $p$ is integer, the system is at one of the resonances. If $p$ is not an integer, it is away from resonance.
We look at the particle current in the ac driven steady state. The expressions for the currents are given in Eq.~\ref{current_B},~\ref{current_S}. For the $N=2$ case, the expression for $I_{1\rightarrow 2}$ simplifies to $I_{1\rightarrow 2}= 2g\hspace{2pt} Im (C_{12})$.
Two different ac drives are considered:
(i) symmetric ac drive or zero voltage drive :
\begin{align}
\label{pot_num_0}
\mu_1 = \mu_2 = V_0 \sin(\Omega_0 t)
\end{align}
(ii) asymmetric ac drive :
\begin{align}
\label{pot_num_1}
\mu_1 = V_1 \cos(\Omega_0 t), \hspace{5pt} \mu_2 = V_2 \sin(\Omega_0 t)
\end{align}
\begin{figure}[t]
\includegraphics[width=9cm,height=9cm]{simulations_2.pdf}
\caption{Internal current $I_{1\rightarrow_2}$ as a function of time, for the two cases (i) for symmetric ac drive: $\mu_1=\mu_2=V_0 \cos(\Omega_0 t)$ (top panel), (ii) for asymmetric ac drive :$\mu_1 = V_1 \cos(\Omega_0 t), \hspace{5pt} \mu_2 = V_2 \sin(\Omega_0 t)$ (bottom panel). The figure compares numerically exact simulation results for current inside the system (stars) with those obtained from our theory Eq.~\ref{C_vec_t} for all times (solid line), and Eq.~\ref{C_vec_steady} for long times (dashed line). The near perfect agreement validates our theory. Other parameters $\Omega_0=2g, g=0.5,t_B=200,\omega_0=1,\beta_1=\beta_2=0.1,\Gamma_1=0.01,\Gamma_2=0.09$. All times are measured in units of $\omega_0^{-1}$, all energies are measured in units of $\hbar \omega_0$.}
\label{fig:one_simu}
\end{figure}
\begin{figure*}[t]
\includegraphics[scale=0.55]{symmetric_2.pdf}
\caption{Various currents under symmetric ac drive ($\mu_1~=~\mu_2~=~V_0\sin(\Omega_0 t)$). Panel (a), (b), (c) describes the set-up at first resonance ($\Omega_0=\omega_2-\omega_1=2g$). Panel (a) shows long time results of $I_{1\rightarrow 2}^{s}$ and currents from the baths ($I_{B^{(1)} \rightarrow 1}^{s}$,$I_{B^{(2)} \rightarrow 2}^{s}$) with asymmetric system-bath coupling ($\Gamma_1 \neq \Gamma_2$) as a function of time over one time period of the ac drive. $I_{B^{(1)} \rightarrow 1}^{s}$,$I_{B^{(2)} \rightarrow 2}^{s}$ are much smaller than $I_{1\rightarrow 2}^{s}$. Panel (b) shows the behavior of the maximum currents with $V_0$ for asymmetric system-bath coupling. The maximum currents increase with $V_0$ and finally saturates. Panel (c) shows behavior of the maximum currents with the degree of asymmetry ($\Gamma_1/\Gamma_2$). The maximum current inside the system ($I_{1\rightarrow 2}^{max}$) decreases with increase in degree of asymmetry, and becomes zero for symmetric coupling. In contrast, maximum current from baths ($I_{B^{(1)} \rightarrow 1}^{max}$,$I_{B^{(2)} \rightarrow 2}^{max}$) remain non-zero and become same for symmetric coupling. This shows a stark difference between symmetric and asymmetric system-bath coupling, and gives an experimental way to determine asymmetry of system-bath coupling. Panel (d) shows the maximum current in the system as a function of $p=2g/\Omega_0$ (Eq.~\ref{p_def}) at two different temperatures. $p$ being integer corresponds to resonances. The first resonance peak is very strong. Higher resonance peaks are much weaker, and are washed out by increasing temperature. Other parameters $g=0.5,t_B=200,\omega_0=1,\beta_1=\beta_2=0.1$. All times are measured in units of $\omega_0^{-1}$, all energies are measured in units of $\hbar \omega_0$. } \label{fig:sym_drive}
\end{figure*}
\subsubsection{Comparison between analytic formula and exact numerics}
Before proceeding to elucidate the physics dominating the driven system, we wish to validate the analytic derivation. For this we compare results obtained from Eq.~\ref{C_vec_t} with that obtained from full numerical simulation of our protocol. For small systems like the two-site case considered here, the protocol for our ac drive set-up can be simulated exactly with finite but large baths. For each time-independent step of our protocol, we choose a bath of finite size with bath correlations satisfying fermi distributions and evolve the full system+bath Hamiltonian $\hat{\mathcal{H}}$ using unitary Hamiltonian dynamics. Let us collectively denote by ``$d$'' a column vector with all annihilation operators of both system and baths. The full Hamiltonian can be written as $\hat{\mathcal{H}}=\sum_{i,j}H_{i j} d^\dagger_i d_j$ where $i$ now refers to either system or bath sites. If $ D= \langle d d^{\dagger} \rangle$ denotes the full correlation matrix of system and baths, its time evolution is given by $D(t)=e^{i H t} D e^{-i H t}$. The bath correlations are then changed according to the protocol, and the process is repeated. Various observables like current inside the system calculated using this exact numerical simulation can be compared with that obtained from Eq.~\ref{C_vec_t}, thus providing a way to validate our theory. Note that the numerical simulation does not take into consideration the `Markov' condition on time scales given in Eq.~\ref{t_cond1}. It holds even when Eq.~\ref{t_cond1} is not respected. Thus it allows us to check the validity of the crucial assumption on time scales required for our analytical treatment.
Fig.~\ref{fig:one_simu} shows the results for current inside the system $I_{1\rightarrow 2}^{s}$ and $I_{1\rightarrow 2}^{a}$ for cases (i) Eq.~\ref{pot_num_0} and (ii) Eq.~\ref{pot_num_1} respectively (superscripts in $I$ stand for symmetric and asymmetric currents), as obtained from exact numerics as well as from our theory. Eq.~\ref{C_vec_steady} has been used to obtain the long time result, while Eq.~\ref{C_vec_t} has been used to get the result at all times, showing approach to the long time dynamics. Numerical simulation has been done with baths of size $256$ sites, which are large enough to have negligible finite-size effects. The near perfect agreement with exact numerical simulations validate our theory. For this plot, the drive frequency is chosen to be $\Omega_0=\omega_1-\omega_2=2g$, so that the set-up is at the first resonance. The near perfect match occurs for other frequencies also as long as Eq.~\ref{t_cond1} is satisfied. For our choice of parameters $\beta\sim 0.1$ satisfies Eq.~\ref{t_cond1}. The initial condition for plots shown in the figure corresponds to no particle in the system, but the agreement with numerical simulations has been checked for other initial conditions (like randomly chosen initial values of correlation functions etc.) also.
Note that numerical validation was only possible owing to the small size of the system, which allowed for using reasonably sized finite baths. For larger system sizes, much larger baths will be required and the set-up will not be amenable to numerical simulation. However, the theory can be easily used for much larger systems connected to infinite baths.
Having validated the theory, we now look at the physics of the long time dynamics for both the cases.
\subsubsection{Symmetric - or zero voltage - ac drive}
First, we look at long time dynamics of the symmetric drive (Eq.~\ref{pot_num_0}). In this case, both chemical potentials are the exact same sinusoidal so that there is no instantaneous voltage difference, i.e, $V(t)=\mu_1(t)-\mu_2(t)=0$. Even though there is no voltage difference, as discussed before, because of being a driven system, there can still be an instantaneous current. Moreover, because of presence of displacement current, the current from the left bath, current in the system, current from the right bath can be different.
Panel (a) of Fig.~\ref{fig:sym_drive} shows currents from the left (right) bath to site 1(2), $I^s_{B^{(1)}\rightarrow 2}~(I^s_{B^{(2)}\rightarrow 2})$, and current inside the system over one time period for the case of first resonance ($p=1$ in Eq.~\ref{p_def}). As expected from our discussion of resonance, the current inside the system is much larger than the currents from the baths. A physical explanation for the non-zero instantaneous current for zero voltage drive can be given as follows.
Let $V_0>\omega_2$ (see Eq.~\ref{pot_num_0}). As the chemical potential varies, the Fermi energy of the particles in the bath varies. When it exceeds $\omega_2$, particles flow into the system from the baths. When the Fermi energy of the particles in the bath is smaller than $\omega_1$, particles flow out of the system into the baths. If the time to reach steady state is larger than the time period of the drive, this transient behavior is observed, which leads to the instantaneous current.
It is clear from this argument, that the instantaneous current increases with increase in $V_0$. Particularly, if $V_0<\omega_1$, there will be a small instantaneous current. This behaviour is shown in panel (b) of Fig.~\ref{fig:sym_drive} which shows the maximum instantaneous currents from the baths and inside the system as a function of $V_0$. The current increases continuously with $V_0$ and finally saturates. The saturation occurs because there are only two eigen-energy levels of the system. Actually, because of the fermionic nature of the set-up, one would expect steps or kinks at the positions $V_0=\omega_1$ and $V_0=\omega_2$. However, such behaviour is not observed because temperature is not low enough. On the other hand, it is important to note that the time-period averaged current is zero always because it is proportional to the difference between time-period averaged fermi distributions of the two baths.
We note that the internal current, while strictly speaking cannot be measured directly, nevertheless may have physical consequences. Specifically, such internal current may lead to local heating of the junction (due to Joule heating). This can lead for instance, to a breakdown of the system (in molecular junctions) or to heating-induced observable changes in current (for a double quantum dot).
Panel (c) of Fig.~\ref{fig:sym_drive} shows a more interesting effect. The plot shows the maximum instantaneous current from the baths and inside the system as a function of asymmetry $\Gamma_1/\Gamma_2$ of the system-bath coupling (note that the chemical potentials are still symmetric). The set-up is still at first resonance. The maximum current inside the system decreases with decrease in asymmetry, and vanishes for the fully symmetric junction, $\Gamma_1/\Gamma_2=1$. At the same time, in the symmetric point the maximum currents from left and right baths are equal. This can be physically explained as follows. If the baths are identically coupled as well as identically driven, the particles come into the system at exactly same rate leading to same current from both baths. Since the particles are fermions, if one particle occupy each site, there can be no current in between the two sites due to Pauli exclusion principle. This leads to the fact that if rate of inflow of particles from both baths is same, there is no current between the two sites. Thus, current between the two sites of the system comes from a mismatch between rate of inflow of particles from the baths. Therefore, current inside the system increases with increase in asymmetry of system-bath coupling.
Panel (d) of Fig.~\ref{fig:sym_drive} shows plots of the maximum current inside the system for asymmetric system-bath coupling with $p$ (as defined in Eq.~\ref{p_def}) for two different temperatures. When $p$ is equal to an integer, the set-up is at resonance. Thus, from our previous discussion, the maximum current inside the system should show peaks at integer values of $p$. The first resonance peak is strong. However, lower frequency resonance peaks are much weaker. This is because the driving signal itself has a small contribution from such modes. Moreover, we see that higher temperature washes out the lower frequency resonances.
We point out that asymmetry in molecular junctions may be a key ingredient in deciphering its electronic transport properties and in designing single-molecule devices (see, e.g., \cite{Warner2015,Zhang2015,Wang2014}). Our results demonstrate that measuring the time-dependent current (for zero voltage bias but time-dependent voltages) is a direct way to measure asymmetry in molecular junctions, which can serve (in parallel to usual transport measurements) for junction characterization.
\begin{figure}[t]
\includegraphics[width=9cm,height=8cm]{asymmetric_2.pdf}
\caption{Panel (a) shows long time results of currents vs voltage $V=\mu_1-\mu_2$ over one time period of ac drive ($V_1=50,V_2=5$). The system is driven by $\mu_1~=V_1\cos(\Omega_0 t)$, $\mu_2~=~V_2\sin(\Omega_0 t)$, system-bath coupling is asymmetric ($\Gamma_1=0.01,\Gamma_2=0.09$) and system is at first resonance ($\Omega_0=\omega_2-\omega_1=2g$). The currents show hysteretic behavior, due to the dynamical nature of the charge transport in and out of the system. Because of the resonance, the current inside the system is much greater than currents from the baths. Panel (b) shows the time-period-averaged current as a function of difference between amplitudes of drives $V_1-V_2$. ($V_2=5$, $V_1$ is varied.) The current increases first and then reaches a plateau. The averaged current is much smaller than the maximum instantaneous current. The dotted lines between the two panels highlights the difference in scale of the plots. Other parameters $g=0.5,t_B=200,\omega_0=1,\beta_1=\beta_2=0.1$. All times are measured in units of $\omega_0^{-1}$, all energies are measured in units of $\hbar \omega_0$.} \label{fig:asym}
\end{figure}
\subsubsection{Asymmetric ac drive}
Having discussed the physics of long time dynamics of the symmetric ac drive, we move to the case of the asymmetric ac drive.
Consider the situation where the left chemical potential and right chemical potentials vary sinusoidally out of phase with same frequency but have different amplitude in general.
Unlike the symmetric ac drive case, here the voltage across the system $V=\mu_1-\mu_2$ is non-zero. For $V_1\neq V_2$. The difference between time-period averaged Fermi distributions of the two baths is not zero in this case. Hence, there is a net time-period averaged current through the system. Fig.~\ref{fig:asym} shows plots of currents from the baths and current in the system at first resonance. Panel (a) shows current vs voltage ($V=\mu_1-\mu_2$) curve over one time period. We observe an interesting hysteresis behavior. Also, because of resonance, the current inside the system is much greater than current from the baths. This effect would not have been seen in the adiabatic limit. Panel (b) shows the time-period averaged current as a function of $V_1-V_2$. The current increases, and finally saturates (similar to $I_{max}$ vs $V_0$ curve for the symmetric drive case). Note that the time-averaged current is much smaller than the maximal instantaneous currents. This implies that perhaps time-dependent signals can be measured even when the average currents are small and below the noise level.
The hysteresis in I-V curve for asymmetric system bath coupling can potentially have device applications. This behavior depends intricately on the phase difference between the two drives, the amplitude difference of the two drives, as well as on the asymmetry of system-bath coupling. These dependences are quite complicated and a detailed investigation of the hysteresis behavior will be taken up in a future work.
The two site set-up described above is experimentally realizable using quantum dots. However, the bottleneck experimental parameter for observation of most of the above effects is frequency of the drive $\Omega_0$. Our most interesting results are close to resonance $\Omega_0=2g$. The maximum frequency of ac drive currently experimentally realizable is $\sim 10 GHz$ \cite{natcomm2016}. This means to observe the above effects $g\sim 0.01 meV$. The baths must have much wider bandwidths than system energy scales, which can be easily arranged experimentally. As evidenced by Eq.~\ref{Jform}, having wide bandwidths will automatically realize the weak system-bath coupling. Temperature $\sim 1mK$ (which is equivalent to $\beta\sim 0.1$) will be consistent with the Markov approximation. These experimental parameters may be challenging but not impossible. Our study thus points to new rich physics in such experimental domain.
\subsection{Conclusion}
We have formulated a theory of an out-of-equilibrium open quantum system set-up where the system is connected to multiple baths and is driven by periodically varying thermodynamic parameters (temperature or chemical potential) of the baths. The key assumptions in our theory are weak system-bath coupling (Born approx) and that the relaxation time scale of the bath is much smaller than the time period of the drive (Markov approx). We do not make any assumption on the relaxation time scales of the system. This has allowed us to go beyond the adiabatic approximation where the system is always assumed to be in the steady state with the instantaneous bath. Our theory works for systems of non-interacting particles (i.e, system Hamiltonian is quadratic) bilinearly coupled to multiple baths (also quadratic Hamiltonian) in a lattice of any dimension or geometry (the particles may be all bosonic or all fermionic, boson-fermion couplings are not bilinear). Even though there exists formally exact methods of treating such periodically driven microscopic models, they are generally quite difficult to calculate beyond the adiabatic regime. On the other hand, phenomenological models based on Lindblad equations have been shown to be either quite restrictive or to have severe shortcomings. Unlike these, under Born-Markov approximation, our theory gives a much simpler, more transparent and much less restrictive results, providing a complete set of linear differential equations for equal time two-point correlation functions from which various physical observables (for example currents and populations) can be directly obtained. Also various physical effects can be directly read-off from such evolution equation for correlation functions. The long time dynamics of the system are readily shown to be periodic with the same period as the drive. The occurrence of resonance when the frequency of the periodic drive equals the difference of two eigen-energies of the system is also easily seen from the equations.
We have tested our theory by applying it to the case of two fermionic sites with hopping between them, weakly coupled to two different baths at same constant temperature but with sinusoidally driven chemical potentials. In this we have considered two cases. We have validated our theory by comparing with exact numerics done with finite but large baths for both the cases. The first case is when the sinusoidal drive is symmetric, i.e, both chemical potentials vary in an exactly same manner so that there is no instantaneous voltage difference. In this case, even though there is no net time-period averaged current, there is an instantaneous current, which can become quite large at resonance. Furthermore, inside the system, the instantaneous current depends on asymmetry of the system-bath coupling and becomes zero when system-bath coupling is symmetric. This gives an experimental way of detecting the asymmetry of system-bath coupling. The maximum instantaneous current increases with amplitude of drive and shows resonance peaks. Increase in temperature washes out the higher resonance peaks.
In the other case, the chemical potentials of the baths have an amplitude difference, as well as a phase difference. Due to amplitude difference, the net time-period averaged current is non-zero and increases with increase in amplitude difference. We find interesting hysteresis in the I-V curves, a detailed investigation of which will be taken up in a future work.
Two most important experimentally relevant insights from our results, especially for studying electronic transport in molecular junctions or through quantum dots, are : (i) time-dependent measurements of current for symmetric oscillating voltages (with zero instantaneous voltage bias) can point to the degree of asymmetry in the system, and (ii) under certain conditions, time-dependent currents can exceed time-averaged currents by several orders of magnitude, and can therefore be detected even when the average current is below the measurement threshold.
Since the formalism is for non-interacting fermions (bosons) on a lattice, bilinearly coupled to fermionic (bosonic) baths, interactions can be built in at a mean field level. Thus our theory can be used to describe a wide range of experimental set-ups (such as molecular junctions, quantum dots, cavity-QED expts, cold atoms, cavity optomechanics etc.). However, the theory cannot be applied to extremely low temperatures, because of the Markov approximation. The relaxation time scale of a bath can be shown to be inversely proportional to the temperature of the bath. Hence, at extremely low temperatures, the Markov approximation would be violated. Also, our formalism cannot treat cases where system-bath coupling is not bilinear, such as fermions connected to bosonic baths and vice-versa. Further work is required for go beyond these limitations. Other future work includes treating interacting systems beyond mean-field, as well as incorporating the effects of dephasing or local vibrations.
\vspace{100pt}
\paragraph*{Acknowledgements: } Y.D. acknowledges support from the Israel Science Fund (Grant No. 1256/14). A.P would like to thank Prof Abhishek Dhar for very insightful discussions, and also for funding support from his grant under the Indo-Israel joint research project No.6-8/2014(IC).
|
\section{Introduction}
Iron-based superconductors are prototypes of moderately correlated\cite{qazilbash,yang} and multiorbital
systems exhibiting unconventional superconductivity(SC).\cite{stewart} A close
proximity of the superconducting phase to the antiferromagnetic (AFM) phase\cite{avci, hussey} indicates
a crucial role of the spin fluctuations like in the cuprate superconductors. The pairing mediated by the
spin-fluctuations has been a subject of intense investigation both theoretically and
experimentally.\cite{scalapino,dahm} An experimental signature of the role of spin-fluctuations, for instance, can
be obtained in the form of change in magnetic-exchange energy involved
in the transition from the normal to the superconducting state, which can then be compared with
the superconducting condensation energy. The magnetic-exchange energy, on the other hand, can be
estimated from the measurements of the spin excitations that yields information
regarding the exchange coupling present in the system.\cite{zhou,wang} Therefore, it is of much importance to comprehend
various characteristics of the spin excitations for gaining insight into the mechanism
of the Cooper pair formation.
Parent compounds of iron pnictides exhibit a spin-density wave state
with ordering wavevector ${\bf Q} = (\pi, 0)$, which involves parallel ferromagnetic chains running along
$y$ direction while being coupled to each other antiferromagnetically along $x$.\cite{zhao} This
particular spin arrangement arises due to the significant nesting present between the circular
hole pockets at $\Gamma$ and the elliptical electron pockets at X in the Fermi surfaces (FSs) of
the unfolded Brillouin zone corresponding to one Fe atom per unit
cell.\cite{mazin,singh,haule,yi,kondo,yi1,brouet,kordyuk} Existence of the Fermi surfaces in
the SDW state,\cite{yi,yi1} metallicity\cite{chuang,nakajima} and
small magnetic moments\cite{cruz,huang} with largest being $\approx$ 1$\mu_B$ found in 122 series supports the
nesting based scenario. However, experimental investigation carried out by the inelastic neutron scattering
(INS) measurements reveals remarkably high-energy scale of the spin-wave excitations, which are
sharp, highly dispersive, and can extend up to energy $\sim$ 200 meV.\cite{ewings,diallo,zhao1,harriger,ewings1}
The spin-wave dispersion can be
described within conventional Heisenberg model with highly anisotropic exchange couplings.\cite{yao} However, such a
description suffers from the limitation that it
cannot explain the spin-wave damping resulting from
the particle-hole excitations in the metallic SDW state. The limitation may be
overcome by considering additional terms which
account for the bandstructure and coupling between the electron spin and the local spin through
Hund's coupling.\cite{lv} Various studies give an estimate of the intraorbital Coulomb interaction ($U$) to be
$U$/$W \approx$ 0.25, where $W$ is the bandwidth.\cite{yang,anisimov} Therefore, plausibly a completely
itinerant approach is best suited to describe the spin excitations in
these materials. Within the latter approach, excitonic\cite{brydon} and orbital\cite{nimisha,knolle} models
have been frequently employed to
investigate the spin-wave excitations. A comparative analysis of these two models while considering a two-orbital model
has also been carried out, where excitations were found to be heavily damped away from the ordering wavevector ${\bf Q}$ in
the latter.\cite{knolle}
Standard on site interaction includes two important parameters - intraorbital Coulomb
interaction $U$ and Hund's coupling $J$. The correlation effect due to $U$ involves the suppression of charge
fluctuations whereas that due to $J$ pertains to the unscreened high-spin state on the neighboring sites
resulting in the correlations. Different values of $U/J$ have often been used by different groups based on
the different estimates from different methods. For instance, a combination of the constrained
random-phase approximation and the maximally localized
Wannier function yields $J/U \sim$ 0.14 \cite{miyake} whereas work based on the dynamical mean-field
theory estimates is $J/U \sim 0.25$.\cite{ishida} Similarly, experiments also provide varying estimates for
the same.\cite{schafgans,yang} With regard to various properties of this material, a comparative role
of these two interaction parameters is of strong current interest.
$U$ vs $J$ phase diagrams in this direction is an important step.\cite{luo} Furthermore, in the two-orbital model, it has been illustrated that $J$ plays an important role in
stabilizing the doped SDW state against long-wavelength fluctuations through the generation of
additional ferromagnetic spin coupling involving the inter-orbital susceptibility.\cite{nimisha}
The spin excitations have been studied recently within the five-orbital model
as function of $U$ but with a fixed Hund's coupling $J$.\cite{kaneshita,kovacic} Well-defined branches extending
upto high energy were obtained for those values of $U$ that led to the magnetic moment $m \sim 1$ or larger. An
important issue that has not attracted much attention is the role of $J$ in the various features of
spin-wave excitations such as dispersion, anisotropy and the way spin-wave spectral weight is
distributed.
In this paper, we examine the role of Hund's coupling in the spin-wave excitations of ($\pi, 0$) SDW state of undoped
iron pnictides within a five-orbital tight-binding model. The interaction parameters are
chosen in such a way that a fixed magnetic moment $m \approx 1$ results in the self-consistent SDW state, which
is motivated by the observed magnetic moments in 122 compounds. we find that $J$
is crucial for (i) the sharp and well-defined excitations
up to high energy, (ii) the anisotropy in the excitations around X, and
(iii) the fact that the spin-wave spectral weight is concentrated near $\sim$ 200meV.
The plan of the paper is as follows. In section II, a mean-field description of the ($\pi, 0$) SDW as well as the
strategy to calculate spin-wave excitations is presented. In section III, results on the spin-wave dispersion
along high-symmetry direction for
different combination of intraorbital Coulomb interaction and Hund's coupling are presented. The way spin-wave spectral
function behaves as a function of the interaction parameters is also discussed. Finally, conclusions are presented in
section IV.
\section{Theory}
The kinetic part of the model Hamiltonian that we consider is
\begin{equation}
\mathcal{H}_0 = \sum_{\k}\sum_{\mu,\nu}\sum_{\sigma} \varepsilon_{\k}^{\mu\nu}
d_{\k \mu\sigma}^\dagger d_{\k \nu\sigma} + \text{H.c.},
\end{equation}
where the operator $d_{{\bf k} \mu \sigma}^\dagger$ ($d_{{\bf k} \mu \sigma}$) creates (destroys)
an electron in the $\mu$-th orbital with spin $\sigma$ and momentum $\k$, and
$\varepsilon_{\k}^{\mu\nu}$ are the hopping elements\cite{ikeda} from orbital $\mu$ to $\nu$. The orbitals $\mu$
and $\nu$ belong to the set of five $d$-orbitals $d_{xz}$, $d_{yz}$, $d_{xy}$, $d_{x^2-y^2}$, and $d_{3z^2-r^2}$.
Standard onsite Coulomb interactions in the Hamiltonian
\begin{eqnarray}
\mathcal{H}_{int} &=& U \sum_{{\bf i},\mu} n_{{\bf i}\mu \uparrow} n_{{\bf i}\mu \downarrow} + (U' -
\frac{J}{2}) \sum_{{\bf i}, \mu<\nu} n_{{\bf i} \mu} n_{{\bf i} \nu} \nonumber \\
&-& 2 J \sum_{{\bf i}, \mu<\nu} {\bf{S_{{\bf i} \mu}}} \cdot {\bf{S_{{\bf i} \nu}}} + J \sum_{{\bf i}, \mu<\nu, \sigma}
d_{{\bf i} \mu \sigma}^{\dagger}d_{{\bf i} \mu \bar{\sigma}}^{\dagger}d_{{\bf i} \nu \bar{\sigma}}
d_{{\bf i} \nu \sigma} \nonumber\\
\label{int}
\end{eqnarray}
include the intraorbital (interorbital) Coulomb interaction term as the first (second) term.
The last two terms represent the Hund’s coupling and the pair
hopping energy, respectively.
The Hamiltonian in the ($\pi, 0$) SDW state after mean-field approximation is obtained as
\begin{equation}
{\mathcal{H}}_{\k} =
\sum_{\k \sigma}\Psi^{\dagger}_{{\bf k} \sigma}
\begin{pmatrix}
\hat{\varepsilon}_{\k}+\hat{N} \,& \,{\rm sgn}\bar{\sigma}\hat{\Delta} \\
{\rm sgn}\bar{\sigma}\hat{\Delta} \,& \,\hat{\varepsilon}_{\bf {k+Q}}+\hat{N}
\end{pmatrix}
\Psi_{{\bf k} \sigma},
\ee
where $\Psi^{\dagger}_{{\bf k} \sigma} = (c^{\dagger}_{{\k}1\sigma},....,c^{\dagger}_{{\k}5\sigma},
{c}^{\dagger}_{{\k}\bar{1}\sigma},....,{c}^{\dagger}_{\k\bar{5}\sigma})$ with $\bar{c}^{\dagger}_{{\k}\bar{\mu}\sigma}$ =
$c^{\dagger}_{{\k+{\bf Q}}\mu\sigma}$. Matrix elements of matrices $\hat{M}$ and $\hat{N}$ are
\begin{eqnarray}
2\Delta_{\mu\mu} &=& Um_{\mu\mu}+J\sum_{\mu \ne \nu}m_{\nu\nu} \nonumber\\
2\Delta_{\mu\nu} &=& Jm_{\mu\nu}+(U-2J)m_{\nu\mu}
\end{eqnarray}
and
\begin{eqnarray}
2N_{\mu\mu} &=& Un_{\mu\mu}+(2U-5J)\sum_{\mu \ne \nu}n_{\nu\nu} \nonumber\\
2N_{\mu\nu} &=& Jn_{\mu\nu}+(4J-U)n_{\nu\mu},
\end{eqnarray}
where
\begin{equation}
n_{\mu\nu} = \sum_{\k \sigma} \langle c^{\dagger}_{\k \mu \sigma}c_{\k \nu \sigma}\rangle, \,\,\,
m_{\mu\nu} = \sum_{\k \sigma} \langle c^{\dagger}_{\k \bar{\mu} \sigma}c_{\k \nu \sigma}\rangle .
\ee
Multiorbital transverse-spin susceptibility is defined as
\begin{eqnarray}
&&\chi_{\alpha \beta, \mu \nu}({\bf q},{\bf q}^{\prime},i\omega_n) \nonumber\\
&&=\frac{1}{\beta} \int^{\beta}_0{d\tau e^{i \omega_{n}\tau}\langle T_\tau
[{ S}^{+}_{\alpha \beta}({\bf q}, \tau) {S}^{-}_{\nu \mu} (-{\bf q}^{\prime}, 0)]\rangle}.
\end{eqnarray}
Thus,
\begin{eqnarray}
&&\chi^{}_{\alpha \beta, \mu \nu}({\bf q},{\bf q},i\omega_n) \nonumber\\
&&= \sum_{\k,i\omega^{\prime}_n} G^{0\uparrow}_{\alpha \mu}
(\k+{\bf q},i\omega^{\prime}_n+i\omega_n)G^{0\downarrow}_{\nu \beta} (\k,i\omega^{\prime}_n).
\end{eqnarray}
The components of the spin operator in Eq. 7 are given by
\begin{equation}
{\cal S}^{i}_{{\bf q}}= \sum_{\bf k} \sum_{\sigma \sigma^{\prime}} \sum_{\mu \mu^{\prime}} d^{\dagger}_{\mu \sigma}(\k+{\bf q})
E_{\mu \mu^{\prime}} \sigma^{i}_{\sigma \sigma^{\prime}} d_{\mu^{\prime} \sigma^{\prime}}(\k),
\ee
where $i = x,y,z$, $\hat{E}$ is a 5$\times$5 identity matrix corresponding to the orbital bases, and $\sigma^i$s are the Pauli matrices for
the spin degree of freedom.
An element of the transverse-spin susceptibility
in the SDW state is
\begin{eqnarray}
{\chi}^{0}_{\alpha \beta, \mu \nu} = \chi^{}_{\alpha \beta, \mu \nu}
+\chi^{}_{\bar{\alpha} \beta, \bar{\mu} \nu}+\chi^{}_{\alpha \bar{\beta}, \mu \bar{\nu}}
+\chi^{}_{\bar{\alpha} \bar{\beta}, \bar{\mu} \bar{\nu}},
\end{eqnarray}
Then, the susceptibility matrix can be written as
\begin{equation}
\hat{{\chi}}^0({\bf q},i\omega_n) =
\begin{pmatrix}
\hat{{\chi}}^0({\bf q},{\bf q},i\omega_n) \,& \,\hat{{\chi}}^0({\bf q},{\bf q}+{\bf Q},i\omega_n) \\
\hat{{\chi}}^0({\bf q}+{\bf Q},{\bf q},i\omega_n) \,& \,\hat{{\chi}}^0({\bf q}+{\bf Q},{\bf q}+{\bf Q},i\omega_n)
\end{pmatrix},
\ee
where $\hat{{\chi}}^0({\bf q},{\bf q},i\omega_n)$ and others are $n^2\times n^2$ matrices.
Physical transverse-spin susceptibility corresponding to the spin operators defined by Eq. 8 is
\begin{equation}
{\chi}^{ps}({\bf q}, i\omega_n) = \sum_{\alpha \mu} {\chi}^0_{\alpha \alpha, \mu \mu} ({\bf q}, {\bf q}, i\omega_n).
\ee
Interaction effects are incorporated within the random-phase approximation (RPA) so that the
spin susceptibility is given by
\begin{equation}
\hat{{\chi}}_{\rm RPA}({\bf q}, i \omega_n) = (\hat{{\bf 1}} - \hat{U}\hat{{\chi}}^0({\bf q},i \omega_n))^{-1} \hat{{\chi}}^0({\bf q},i \omega_n).
\ee
Here, $\hat{{\bf 1}}$ is a $2n^2 \times 2n^2$ identity matrix and the elements of block diagonal matrix
$\hat{U}$ are
\begin{eqnarray}
&&{U}_{\mu_1 \mu_2;\mu_3 \mu_4} \nonumber\\
&&= \left\{
\begin{array}{@{\,} l @{\,} c}
U \,\,& (\mu_1=\mu_2=\mu_3=\mu_4)\\
U-2J \,\, & (\mu_1=\mu_2\ne\mu_3=\mu_4)\\
J \,\,& (\mu_1 = \mu_2\ne \mu_3 = \mu_4)\\
J \,\, & (\mu_1=\mu_4\ne \mu_2=\mu_3)\\
0 \,\,& (\mathrm{otherwise})
\end{array}\right., \nonumber\\
\end{eqnarray}
where $U^{\prime}$ = $U$ - $2J$ has been used as required by the rotational invariance.
Analytic continuation $i \omega_n \rightarrow \omega + i \eta$ in all the calculations
described below is performed with $\eta$ as 0.002eV. The unit of energy eV is used throughout
unless stated otherwise.
\begin{figure}
\begin{center}
\vspace*{-2mm}
\psfig{figure=fig1a.eps,width=88mm,angle=0}
\vspace*{-0mm}
\psfig{figure=fig1b.eps,width=88mm,angle=0}
\vspace*{-0mm}
\psfig{figure=fig1c.eps,width=88mm,angle=0}
\vspace*{5mm}
\psfig{figure=fig1d.eps,width=88mm,angle=0}
\vspace*{-13mm}
\end{center}
\caption{Imaginary part of the RPA-level physical spin susceptibility Im$\chi^{ps}_{\rm RPA}({\bf q}, \omega)$
along the high-symmetry directions for the set of interaction parameters chosen
in such a way that $m = 1$ in each case. (a) $U = 1.1, J = 0.247U$ (b) $U = 1.3, J = 0.167U$ (c) $U = 1.5, J = 0.110U$, and
$U = 1.7,$ and $J = 0.730U$.}
\label{rpadisp}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-2mm}
\hspace*{0mm}
\psfig{figure=fig2a.eps,width=88mm,angle=0}
\psfig{figure=fig2b.eps,width=88mm,angle=0}
\psfig{figure=fig2c.eps,width=88mm,angle=0}
\vspace*{5mm}
\psfig{figure=fig2d.eps,width=88mm,angle=0}
\vspace*{-13mm}
\end{center}
\caption{Imaginary part of the bare physical spin susceptibility Im$\chi^{ps}_{}({\bf q}, \omega)$ along
the high-symmetry directions for the set of interaction parameters
as in Fig. \ref{rpadisp}.}
\label{baredisp}
\end{figure}
\section{Results}
\begin{figure}
\begin{center}
\vspace*{-4mm}
\hspace*{0mm}
\psfig{figure=fig3.eps,width=66.0mm,angle=-90}
\vspace*{-15mm}
\end{center}
\caption{FSs obtained for the two sets of interaction parameters (a) $U = 1.1, J = 0.247U$ and (b)
$U = 1.7, J = 0.073U$. Magnetic moment $m = 1$ in each case. In both the cases, FSs
include several pockets near $\Gamma$ with a difference in the sizes.}
\label{fs}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-4mm}
\hspace*{-4mm}
\psfig{figure=fig4.eps,width=90.0mm,angle=0}
\vspace*{-10mm}
\end{center}
\caption{Electron dispersions for (a) $U = 1.1, J = 0.247U$ and (b)
$U = 1.7, J = 0.073U$ in the high-symmetry directions. Magnetic moment $m = 1$ in each case.}
\label{disp}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-4mm}
\hspace*{0mm}
\psfig{figure=fig5.eps,width=90.0mm,angle=0}
\vspace*{-12mm}
\end{center}
\caption{Constant energy cuts of Im$\chi^{ps}_{\rm RPA}({\bf q}, \omega)$ for $U = 1.1, J = 0.247U$ at
energies (a) 50 meV, (b) 100 meV, (c) 150 meV, and (d) 200 meV.}
\label{cut1}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-4mm}
\hspace*{0mm}
\psfig{figure=fig6.eps,width=90.0mm,angle=0}
\vspace*{-12mm}
\end{center}
\caption{Constant energy cuts of Im$\chi^{ps}_{\rm RPA}({\bf q}, \omega)$ for $U = 1.7, J = 0.073U$ at
energies (a) 50 meV, (b) 100 meV, (c) 150 meV, and (d) 200 meV.}
\label{cut2}
\end{figure}
\begin{figure*}
\begin{center}
\vspace*{-4mm}
\hspace*{0mm}
\psfig{figure=fig7.eps,width=160mm,angle=0}
\vspace*{-5mm}
\end{center}
\caption{Spin-wave spectral functions for the cases when (a) magnetic moment $m = 1$ whereas $U$ and $J$ are set
constant in (b) and (c), respectively.}
\label{spect}
\end{figure*}
\begin{table}[ht]
\caption{Orbital resolved magnetic moments and charge densities for the two cases $U = 1.1$, $J = 0.247U$
and $U = 1.7$, $J = 0.073U$ denoted by subscripts 1 and 2, respectively.}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand{\arraystretch}{1}
\begin{tabular}{c|c c c c c c c c c c c c c c c c}
\hline\hline
$ $ & $d_{3x^2-r^2}$ & $d_{xz}$ & $d_{yz}$ & $d_{xy}$ & $d_{x^2-y^2}$ \\ [0.5ex]
$n_1$ &$1.457$ & $1.230$ & $1.165$ & $0.982$ & $1.165$ \\
$n_2$ &$1.574$ & $1.309$ & $1.189$ & $0.834$ & $1.094$ \\
$m_1$ &$0.160$ & $0.155$ & $0.252$ & $0.338$ & $0.098$ \\
$m_2$ &$0.099$ & $0.185$ & $0.197$ & $0.422$ & $0.075$ \\
[1ex]
\hline
\end{tabular
\vspace{-2mm}
\end{table}
\begin{table}[ht]
\caption{ $\sum_{{\bf q},\omega}$Im $\chi^{sp}_{\alpha \alpha, \beta \beta}$ (${\bf q}, \omega$) at RPA-level for
$U = 1.1$, $J = 0.247U$.}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand{\arraystretch}{1}
\begin{tabular}{c|c c c c c c c c c c c c c c c c}
\hline\hline
$\alpha/\beta $ & $d_{3x^2-r^2}$ & $d_{xz}$ & $d_{yz}$ & $d_{xy}$ & $d_{x^2-y^2}$ \\ [0.5ex]
$d_{3x^2-r^2}$ &$0.030$ & $0.023$ & $0.030$ & $0.039$ & $0.019$ \\
$d_{xz}$ & $0.023$ & $0.076$ & $0.033$ & $0.028$ & $0.026$ \\
$d_{yz}$ & $0.030$ & $0.033$ & $0.101$ & $0.032$ & $0.032$ \\
$d_{xy}$ & $0.039$ & $0.028$ & $0.032$ & $0.170$ & $0.036$ \\
$d_{x^2-y^2}$ & $0.019$ & $0.026$ & $0.032$ & $0.036$ & $0.026$ \\
[1ex]
\hline
\end{tabular
\vspace{-2mm}
\end{table}
\begin{table}[ht]
\caption{Same as in Table II but for $U = 1.7$, $J = 0.073U$.}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand{\arraystretch}{1}
\begin{tabular}{c|c c c c c c c c c c c c c c c c}
\hline\hline
$\alpha/\beta $ & $d_{3x^2-r^2}$ & $d_{xz}$ & $d_{yz}$ & $d_{xy}$ & $d_{x^2-y^2}$ \\ [0.5ex]
$d_{3x^2-r^2}$ &$0.010$ & $0.011$ & $0.031$ & $0.017$ & $0.009$ \\
$d_{xz}$ & $0.011$ & $0.085$ & $0.021$ & $0.003$ & $0.019$ \\
$d_{yz}$ & $0.031$ & $0.021$ & $0.240$ & $0.000$ & $0.050$ \\
$d_{xy}$ & $0.017$ & $0.003$ & $0.000$ & $0.250$ & $0.027$ \\
$d_{x^2-y^2}$ & $0.009$ & $0.019$ & $0.050$ & $0.027$ & $0.040$ \\
[1ex]
\hline
\end{tabular
\vspace{-2mm}
\end{table}
Fig. \ref{rpadisp} shows the imaginary part of RPA susceptibility Im$\chi^{ps}_{\rm RPA}({\bf q}, \omega)$ along the
high-symmetry directions Y-$\Gamma$-X-M calculated in SDW state for values of
$U$ starting from 1.1 and increasing in step of 0.2 while $J$ is chosen
so that magnetic moment $m = 1$ with a maximum deviation only upto $\approx$ 2$\%$. We find a
strong dependence of several features of the self-consistent mean-field state on $J$ as seen from the Table I, where the difference in
the orbital-resolved magnetizations for the two set of parameters $U = 1.1$, $J = 0.247U$ and $U = 1.7$, $J = 0.073U$
can be as large as $\approx$ 50$\%$ (for $d_{3x^2-r^2}$ orbital) despite the same magnetic moment. A significant
difference in the two cases can be noticed for all the orbitals except $d_{x^2-y^2}$. Further, orbital-resolved
magnetization decreases for $d_{xz}$ and $d_{xy}$ and increases for the remaining orbitals, when $J$ decreases.
Well-defined spin-wave excitations are obtained along $\Gamma$-X and in a part of X-M, where they can extend
up to $\approx$ 0.2eV for $U = 1.1$, $J = 0.247U$ as shown in the Fig. \ref{rpadisp}(a). Contributions to
the excitations are mainly from the intraorbital susceptibilities particularly corresponding to
the orbitals $d_{yz}$, $d_{yz}$ and $d_{xy}$ with latter contributing the most as
seen from the Table II and Table III. Tables
show integrated spectral weight $\sum_{{\bf q},\omega}$Im$\chi^{ps}_{\rm RPA}({\bf q}, \omega)$ with the
upper cutoff for the summation over $\omega$ chosen as $\omega_u = 0.5$eV. When $J$ is decreased, contributions
due to the interorbital susceptibilities also decrease whereas the contribution from the intraorbital
susceptibility corresponding to $d_{yz}$ orbital increases and becomes similar in magnitude to that
corresponding to $d_{xy}$.
The excitations become increasingly broad and
diffusive when $J$ is small. However, exactly opposite happens
near M, where they become rather sharp and non-diffusive. At the same time, energy of
the spin-wave excitations near M increases. Heavy damping is present due to the metallicity
of the SDW state as also reflected in the imaginary part of bare spin susceptibilities which appear gapless
near $\Gamma$ but are gapped in other regions (Fig. \ref{baredisp}). The gap decreases in most part of
the high-symmetry directions on decreasing $J$ except near
M, where it increases on the contrary. Ungapped imaginary part of the bare susceptibility derives
from the fact that FSs exist for all the sets of parameters considered with the Fermi pockets being
clustered around $\Gamma$ (Fig. \ref{fs}).
Behavior of the spin-excitations near M for smaller $J$ indicate sharpness may not be always plausible particularly in
the multiorbital systems like pnictides, where many bands are located near the Fermi level.
In order to understand the broadening especially near M when $J$ is large, it is useful to examine the reconstructed
bands in the SDW state, wherefrom it can be seen
that a part of the lowest lying partially-filled band gets lowered further near ($0, \pi$) on increasing $J$.
Finally, it is in the immediate vicinity of the Fermi level for $U = 1.7$, $J = 0.073U$ as shown in Fig. \ref{disp}
and may affect the imaginary part of the bare and RPA-level susceptibilities significantly, which
are ungapped and heavily damped, respectively. At the same
time, other portions of the bands remain largely unaffected. Thus, the unusual feature of spin-wave excitations near M may
arise due to a subtle interplay between the roles of interaction and the bandstructure.
Fig. 5 and 6 show the energy cuts for $U = 1.1, J = 0.247U$ and
$U = 1.7, J = 0.073U$, respectively. For the former case, anisotropy in the form of
elliptical structure of excitations around X can be seen upto a energy as high as 200meV, where the major axis is oriented
along $\Gamma$-Y. Excitations around M are rather broad and less intense with nature of
anisotropy being similar to that around X. That changes quickly near 100meV when the structure around M becomes extended
along $\Gamma$-X. In contrast, anisotropy around
X is relatively weak for the latter set of interaction parameters and is strong only in a narrow energy window
around $\omega = 150$meV for M. Moreover, the elliptical structure around X is also absent. Thus, anisotropy
near ${\bf Q}$ particularly in the elliptical form is very sensitive to $J$. In both cases, the
anisotropy continues to exist upto high energy. As revealed also by the
INS measurements, a significant anisotropy with elliptical shape is present
in the spin-wave excitations around X.\cite{diallo,zhao}
Fig. \ref{spect} shows the spin-wave spectral functions calculated for three different cases. When the magnetic
moment $m \approx 1$ and a suitable set of interaction parameters is chosen, there is noticeable shift in the
spectral weight towards low-energy region upon increasing $U$ (Fig. \ref{spect} (a)). The hump-like peak
structure located near 250meV for $U = 1.1, J = 0.247U$, which
is in agreement with the experiments, \cite{zhao1} relocates near 125meV $U = 1.7, J = 0.073U$. Another important factor that
can be responsible for the spectral weight shifting towards low-energy region is doping of holes or electrons, which will
be discussed elsewhere. On the other hand, transfer of the spectral weight towards
lower energy is significant in other two cases where one of the two interaction parameters is increased while keeping the other constant.
\section{Conclusions}
To conclude, we have investigated the role of Hund's coupling in the spin-wave excitations of the SDW
state of undoped iron pnictides by using a realistic electronic structure within a five-orbital model. We find that the Hund's coupling at the higher end of the range of various theoretical and experimental estimates ($J \sim U/4$) is
required for the sharp and well-defined spin-wave dispersion in most part of the
high-symmetry directions for a given magnetization. Not only that a similar value of Hund's coupling
is also crucial for the elliptical
shape of the anisotropy around ${\bf Q}$ = ($\pi$, 0) in the spin-wave excitations as well as
for the spectral weight to be concentrated near energy $\gtrsim$ 200meV. Thus, our study highlights the essential role of Hund’s coupling
in describing the experimentally observed features of spin-wave excitations in the SDW state of undoped iron pnictides.
\section{Acknowledgement}
We acknowledge the use of HPC clusters at HRI.
|
\section{Introduction}
A magnetic skyrmion in magnets is swirling spin texture that behaves as a particle with a long lifetime \cite{Skyrme61,Skyrme1962556,Bogdanov89} due to the topological protection. They were observed experimentally recently in chiral magnets where the inversion symmetry is broken. \cite{Muhlbauer2009,Yu2010a} For their unique topological properties and long lifetime, skyrmions have attracted considerable interests as possible information carriers. Skyrmions can be driven by various external fields, such as electric current \cite{Jonietz2010,Yu2012,Schulz2012}, magnetic/electric field gradient \cite{White2012,PhysRevLett.113.107203}, thermal gradient \cite{Kong2013,Lin2014PRL,Mochizuki2014} and magnon current \cite{PhysRevB.90.094423} etc. The ability to manipulate skyrmions by an electric current is especially attractive because this implies immediately that skyrmions can be used in spintronic devices. \cite{Fert2013,nagaosa_topological_2013} Moreover the threshold current to make skyrmions mobile is weak, thanks to the smooth spin texture associated with the skyrmion.
For applications, it is crucial to understand the dynamics of skyrmion in response to an external drive. The equation of motion of a rigid skyrmion in two dimensions (in the $x$-$y$ plane) has been obtained by Thiele long time ago. \cite{Thiele72} It has the following form
\begin{align}\label{eq0}
\mathbf{G}_T\times \mathbf{v}+D_T \mathbf{v}=\mathbf{F}_T,
\end{align}
where $\mathbf{v}$ is the skyrmion velocity. The first term describes gyromotion with the gyrovector $\mathbf{G}_T$ perpendicular to the $x$-$y$ plane, and the second term is the damping. Here $D_T\ll G_T$ and the skyrmion moves almost perpendicular to the external force $\mathbf{F}_T$. In the Thiele's collective coordinate approach, the skyrmion texture $\mathbf{S}(\mathbf{r})$ is assumed to be rigid and it moves as a whole, $\mathbf{S}[\mathbf{r}-\mathbf{R}(t)]$, with $\mathbf{R}(t)$ representing the translational motion of a skyrmion. It has been demonstrated in numerical simulations that the Thiele's equation of motion correctly captures the skyrmion dynamics driven by a spin transfer torque induced by a dc current. \cite{Iwasaki2013,szlin13skyrmion1,szlin13skyrmion2} Nevertheless, it is unknown to date how to justify the rigid skyrmion approximation.
A skyrmion is a collective excitation of spin texture and thus has internal degrees of freedom. \cite{Lin_internal_2014,PhysRevB.90.094423} In the presence of an external drive, the skyrmion can be deformed when the magnon modes are excited by the external drive. In the continuum approximation where the system preserves the translational symmetry, there is a Goldstone mode in the characteristic deformations of the skyrmion corresponding to the translational motion of a skyrmion. \cite{Everschor11,Everschor12,PhysRevB.58.R8889} The Thiele's equation of motion includes only the Goldstone mode. In principle, the deformations associated with other magnon modes can also be involved in the motion of a skyrmion. To go beyond the Thiele's approach, one needs to take all the magnon excitations into account. One can introduce corrections to the Thiele's equation of motion, such as mass and gyrodamping, and then fit the generalized Thiele's equation of motion to the skyrmion trajectory obtained from direct simulation of the spin dynamics based on the Landau-Lifshitz-Gilbert equation. In this way the parameters in the Thiele equation of motion can be extracted. \cite{PhysRevB.50.12711,PhysRevB.90.174434,buttner_dynamics_2015} Similar to the original Thiele's equation of motion, the generalized Thiele's equation in prior is not justified. A theory to describe the skyrmion dynamics by treating all the magnon modes on an equal footing thus is required.
In this work, we present a linear theory for the skyrmion dynamics by taking all the magnon modes into account. Specifically, we consider a small oscillation of a skyrmion subjected to a weak oscillating magnetic field gradient, spin transfer torque and spin Hall torque. To do that, we first define the skyrmion center as its topological charge center. We then express the dynamics of the skyrmion center in terms of the magnon modes. We find that only the magnon modes with an angular momentum $|m|=1$ are responsible for the dynamics of skyrmion center. The retardation or inertia of skyrmions is due to the magnon modes in the continuum. For a skyrmion driven by a field gradient or spin transfer torque, the inertia is negligible, which justifies the Thiele's rigid skyrmion approximation. For a spin Hall torque, the motion of skyrmions is originated from the coupling between the magnon continuum and current. The inertia is significant in this case. The inertia can be quantified by measuring the phase shift between the skyrmion velocity and the driving field.
\section{Model and Bloch skyrmion solution}\label{Sec2}
We consider the following Hamiltonian density for spins $\mathbf{S}(\mathbf{r})$ in two dimensional space \cite{Bogdanov89}
\begin{equation}\label{eq1}
{\cal H} = \frac{{{J_{\mathrm{ex}}}}}{2}\mathop \sum \limits_{\mu = x,y} {\left( {{\partial _\mu }{\bf{S}}} \right)^2} + D{\bf{S}}\cdot\nabla \times {\bf{S}} - {\bf{B}}\cdot{\bf{S}},
\end{equation}
which successfully captures many experimental observations in chiral magnets. Here $\mathbf{S}$ is a unit vector representing the spin direction, $J_{\mathrm{ex}}$ is the exchange interaction, $D$ is the Dzyaloshinskii-Moriya (DM) interaction \cite{Dzyaloshinsky1958,Moriya60,Moriya60b} and $\mathbf{B}=B\hat{z}$ with a unit vector $\hat{z}$ is the external magnetic field perpendicular to the plane. We have neglected the weak dipolar interaction. The skyrmion size is much bigger than the spin lattice constant, and this justifies the continuum approximation in Eq. \eqref{eq1}. We renormalize the length in unit of $J_{\mathrm{ex}}/D$ and energy density and $B$ in unit of $D^2/J_{\mathrm{ex}}$. Then Eq. \eqref{eq1} takes a dimensionless form. We will use dimensionless quantities in the following derivations. The Hamiltonian Eq. \eqref{eq1} supports the Bloch skyrmion solution. The results for the N\'{e}el skyrmions will be discussed in Sec. VI.
We focus on a single skyrmion in the ferromagnetic background, which is a metastable state for $B>0.55$. \cite{Bogdanov94,Lin_internal_2014,PhysRevB.90.094423} Because of the rotational symmetry, it is more convenient to work in the polar coordinate $\mathbf{r}=(r,\ \phi)$. The skyrmion solution is $\mathbf{S}_0=(\cos\varphi\sin\theta,\ \sin\varphi\sin\theta, \ \cos\theta)$ with $\varphi=\phi+\pi/2$ and $\theta(r)$ being a function of $r$ only. The stationary skyrmion solution is obtained by minimizing $\mathcal{H}$ with respect to $\theta$, and we have equation for $\theta(r)$
\begin{equation}\label{eq2}
\cos \left( {2\theta } \right) + \frac{1}{{2r}}\sin(2\theta) + B r\sin\theta - \left( {{\partial _r}\theta + 1} \right) - r\partial _r^2\theta = 0.
\end{equation}
We solve Eq. \eqref{eq2} using the relaxation method to find $\theta(r)$. The skyrmion structure and the results for $\theta(r)$ and $\partial_r\theta(r)$ at $B=0.8$ are displayed in Fig. \ref{f1}.
\begin{figure}[b]
\psfig{figure=fig1.pdf,width=\columnwidth}
\caption{(color online) Profiles of $\theta(r)$ and $\partial_r\theta(r)$ for a skyrmion in the ferromagnetic background. Inset is a schematic view of a skyrmion. Here $B=0.8$.
} \label{f1}
\end{figure}
\section{Eigenmodes analysis}
We calculate the eigenmodes of a skyrmion in the ferromagnetic background, following the approach in Ref. \onlinecite{PhysRevB.90.094423}. We introduce a local coordinate system with the local $z$ axis along the spin direction $\mathbf{S}$. The spin representation in the lab coordinate and the local coordinate are related by the following subsequent rotation operations in the lab frame: rotation along the $z$ axis by $\phi_0$, rotation along the $y$ axis by $\theta$ and rotation along the $z$ axis by $\varphi$. We choose $\phi_0=\pi/2$. Then the spin in the local coordinate $\mathbf{L}=(L_X,\ L_Y,\ L_Z)$ is related to that in the lab frame $\mathbf{S}$ according to $\mathbf{S}=\hat{O}\mathbf{L}$, where \cite{Lin_internal_2014}
\begin{equation}\label{eq4}
\hat{O}= \left(
\begin{array}{ccc}
-\text{sin$\varphi $} & -\text{cos$\varphi $} \text{cos$\theta $} & \text{cos$\varphi $} \text{sin$\theta $} \\
\text{cos$\varphi $} & -\text{sin$\varphi $} \text{cos$\theta $} & \text{sin$\varphi $} \text{sin$\theta $} \\
0 & \text{sin$\theta $} & \text{cos$\theta $} \\
\end{array}
\right).
\end{equation}
In the stationary state $\bar{L}_{X}=\bar{L}_Y=0$ and $\bar{L}_Z=1$. We consider small deviations $\mathbf{L}$ from the stationary state and introduce the magnon fields
\begin{equation}\label{eq5}
\psi=\frac{L_X+i L_Y}{\sqrt{2}}, \ \ \ \psi^*=\frac{L_X-i L_Y}{\sqrt{2}},
\end{equation}
and $L_Z=1-\psi\psi^*$ with $|\psi|\ll 1$. Expanding the Hamiltonian to the second order in $\psi$, we obtain
\begin{equation}\label{eq6}
\mathcal{H}_\psi=\frac{1}{2}\hat{\psi}^\dagger\mathcal{H}_\psi\hat{\psi},
\end{equation}
\begin{equation}\label{eq7}
\mathcal{H}_\psi=(-\nabla^2+V_0)I_0-2\sigma_z\left(\frac{\cos\theta}{r^2}-\frac{\sin\theta}{r}\right)i\partial_\phi+V_1\sigma_x,
\end{equation}
with $\sigma_i$ ($i=x,\ y,\ z$) being the Pauli matrices and $I_0$ the unit matrix. Here
\begin{equation}\label{eq8}
\hat{\psi}^\dagger=(\psi^*,\ \ \psi),
\end{equation}
\begin{equation}\label{eq9}
{V_0} = \frac{{1 + 3\cos \left( {2\theta } \right)}}{{4{r^2}}} - \frac{{3\sin \left( {2\theta } \right)}}{{2r}} + {B_z}\cos\theta - {\partial _r}\theta - \frac{1}{2}{\left( {{\partial _r}\theta } \right)^2},
\end{equation}
\begin{equation}\label{eq10}
{V_1} = \frac{{{{\sin }^2}\theta }}{{2{r^2}}} + \frac{{\sin \left( {2\theta } \right)}}{{2r}} - {\partial _r}\theta - \frac{1}{2}{\left( {{\partial _r}\theta } \right)^2}.
\end{equation}
Equations \eqref{eq6}-\eqref{eq10} were first derived in Ref. \onlinecite{PhysRevB.90.094423}. Note that with the definition of magnon wave function in Eq. \eqref{eq5}, there is a minus sign in front of $\sigma_z$ in Eq. \eqref{eq7} which is different from the results in Ref. \onlinecite{PhysRevB.90.094423}. The eigenmodes are determined by the equation
\begin{equation}\label{eq11}
-i\sigma_z\partial_t\hat{\psi}=\mathcal{H}_\psi\hat{\psi}.
\end{equation}
This equation for magnons has the form of the Schr\"{o}dinger equation in the presence of a centrosymmetric potential. We can introduce an angular momentum $m$ with $\psi=\psi_m(r,t)\exp(i m\phi)$ to classify the eigenmodes. The two components in $\hat{\psi}$ are related by complex conjugation because the magnetic moment $\mathbf{S}$ is real. This indicates that the matrix equation in Eq. \eqref{eq11} is reductant. Indeed $\mathcal{H}_\psi$ has particle-hole symmetry, $\mathcal{H}_\psi=\sigma_x K\mathcal{H}_\psi K \sigma_x$ with $K$ being the complex conjugate operator. This means that if $\exp[i(\omega t+m\phi)]\hat{\eta}_m$, with $\hat{\eta}_m^\dagger\equiv (\eta_1^*, \ \ \eta_2^*)$, solves Eq. \eqref{eq11}, then $\exp[-i(\omega t+m\phi)]\sigma_x K\hat{\eta}_m$ also solves Eq. \eqref{eq11}. Then $\hat{\psi}$ can be obtained by linear superposition of the two symmetry-related solutions
\begin{equation}\label{eq12}
\hat{\psi}_m=b \exp[i(\omega t+m\phi)]\hat{\eta}_m+b^*\exp[-i(\omega t+m\phi)]\sigma_x K\hat{\eta}_m.
\end{equation}
For $\hat{\psi}_m$ in Eq. \eqref{eq12}, its two components are complex conjugate to each other. Here $\hat{\eta}$ is determined by the eigenvalue problem
\begin{equation}\label{eq13}
\omega_m\sigma_z\hat{\eta}_m=\mathcal{H}_\psi\hat{\eta}_m.
\end{equation}
To solve Eq. \eqref{eq13}, we use the Bessel functions $J_m(k r)$ as basis to represent the matrix $\mathcal{H}_\psi$ as detailed in the Appendix A. From $\hat{\eta}_m$, we know $\mathbf{L}_m$ from Eq. \eqref{eq5}. Then we obtain the eigenmodes $\tilde{\mathbf{S}}_m$ in the lab frame through rotation, $\tilde{\mathbf{S}}_m=\hat{O}\mathbf{L}_m$.
In the high energy limit $\omega\gg \omega_g$, with $\omega_g=B$ being the magnon gap, we can use magnon momentum $k$ to label the magnon mode. The eigenfrequency in this limit is $\omega_m(k)=k^2+\omega_g$ and the eigenmodes are $\hat{\eta}_m^\dagger=(1,\ 0)J_m(k r)$. When $\omega$ is comparable to $\omega_g$, the momentum $k$ is not a good quantum number because the presence of the skyrmion breaks the translational invariance for magnons.
In the continuum limit adopted in Eq. \eqref{eq1}, valid when the skyrmion size is much bigger than the spin lattice constant, the mode corresponding to the translational motion of a skyrmion is a Goldstone mode. It has $|m|=1$ and the magnon mode in the lab frame is
\begin{align}\label{eq14}
{\tilde{\mathbf{S}}_{|m| = 1,j=1}} = \frac{{{{i}m}{\partial _x}{\mathbf{S}_0} - {\partial _y}{\mathbf{S}_0}}}{{\sqrt 2 \sqrt {\int{{d}}{r^2}{{\left( {{\partial _x}{\mathbf{S}_0}} \right)}^2}} }}.
\end{align}
The imaginary and real part of ${\tilde{\mathbf{S}}_{|m| = 1}}$ correspond to the translation motion along the $x$ and $y$ direction respectively.
\section{Linear response to external drive}
We proceed to calculate the response of skyrmion to external drive in terms of the eigenmodes. We consider the Landau-Lifshitz-Gilbert equation of motion for $\mathbf{S}$
\begin{equation}\label{eq15}
\partial_t\mathbf{S} =-\mathbf{S}\times \mathbf{H}_{\mathrm{eff}}+\alpha \mathbf{S}\times \partial_t\mathbf{S} +\mathbf{\Gamma},
\end{equation}
where $\alpha$ is the Gilbert damping, $\mathbf{H}_{\mathrm{eff}}\equiv -\delta\mathcal{H}/\delta\mathbf{S}$ is the effective magnetic field and $\mathbf{\Gamma}$ is the torque due to external fields. Here the time is in unit of $J_{\mathrm{ex}}/(\gamma D^2)$ with $\gamma$ the gyromagnetic ratio. We study the linear response of a skyrmion to a weak torque, $|\mathbf{\Gamma|}\ll 1$. We will work in the lab frame here. We consider a small oscillation of the skyrmion center around the equilibrium position. The linear response $\mathbf{S}=\mathbf{S}_0+\tilde{\mathbf{S}}$ is governed by
\begin{equation}\label{eq16}
{\partial _t}\tilde{\mathbf{S}}(t) = \hat{\mathcal{H}}_S\tilde{\mathbf{S}} + \alpha {\mathbf{S}_0} \times {\partial _t}\tilde{\mathbf{S}} +\mathbf{ \Gamma}.
\end{equation}
Here $\hat{\mathcal{H}_S}$ and $\tilde{\mathbf{S}}$ are connected to $\mathcal{H}_\psi$ and $\hat{\psi}$ through rotation, Eq. \eqref{eq4}. Expanding $\tilde{\mathbf{S}}$ in terms of the eigenmodes, we have
\begin{equation}\label{eq17}
\tilde{\mathbf{S}}(t) = \mathrm{Re}\left[\exp(i\omega t)\sum_{m, j}a_{m, j}\tilde{\mathbf{S}}_{m,j}\right].
\end{equation}
The index $j$ labels the modes at a given $m$. Substituting Eq. \eqref{eq17} into Eq. \eqref{eq16} and projecting into the $\tilde{\mathbf{S}}_{m,j}$ mode, we obtain
\begin{align}
a_{m,j}(\omega-\omega_{m,j})+i\omega\sum_{j'}\alpha_{m, j, j'}a_{m,j'}=F_{m,j},
\end{align}
where $F_{m,j}$ represents the coupling between the external torque and the eigenmodes
\begin{equation}
F_{m,j}(\omega) =-i\int dr^2 \tilde{\mathbf{S}}_{m,j}^\dagger\cdot \mathbf{\Gamma},
\end{equation}
and $\alpha_{m, j,j'}$ is the damping coefficient for different modes
\begin{align}\label{eqalphamj}
\alpha_{m,j,j'} =i\alpha \int dr^2 \tilde{\mathbf{S}}_{m,j}^\dagger\cdot\left(\mathbf{S}_0\times \tilde{\mathbf{S}}_{m,j'} \right)\nonumber\\
=i2\pi\alpha\int dr r[L_{Y;m, j}^*(r)L_{X; m,j'}(r)-L_{X;m, j}^*(r)L_{Y;m, j'}(r)].
\end{align}
As shown in the Appendix B, the diagonal elements $|\alpha_{m,j,j}|$ are much larger than the off-diagonal elements $|\alpha_{m,j,j'}|$. For $\alpha\ll 1$, we can neglect the off-diagonal elements of $\alpha_{m,j,j'}$ and take $\alpha_{m,j,j'}=\alpha_{m,j}\delta_{j,j'}$, with $\delta_{j,j'}$ being the Kronecker delta function. For $\omega_{m, j}\gg \omega_g$, we have $\alpha_{m,j}=1$. We then obtain $a_m$ in the frequency domain
\begin{equation}\label{eqamj1}
a_{m,j}(\omega) =\frac{F_{m,j}}{\omega-\omega_{m,j}+i \alpha_{m, j}\omega}.
\end{equation}
To obtain the equation of motion for a skyrmion as a particle, we need to define its center. We use the definition based on its topological charge density
\begin{equation}\label{eq21}
{\bf{R}} (t)= \frac{{\int d{r^2}{\bf{r}}\mathbf{S}\cdot\left( {{\partial _x}\mathbf{S} \times {\partial _y}\mathbf{S}} \right)}}{{\int d{r^2}\mathbf{S}\cdot\left( {{\partial _x}\mathbf{S} \times {\partial _y}\mathbf{S}} \right)}}=\frac{\int d{r^2}{\bf{r}}\mathbf{S}\cdot\left( {{\partial _x}\mathbf{S} \times {\partial _y}\mathbf{S}} \right)}{4\pi N_s},
\end{equation}
where $N_s=-1$ is the skyrmion topological charge for the skyrmion shown in Fig. \ref{f1}, which is invariant with respect to small perturbations.
Taking the time derivative of $\mathbf{R}(t)$ and expanding in terms of the eigenmodes, we obtain
\begin{equation}\label{eq22}
{\mathbf{R}}(\omega)=\sum_{m, j}a_{m,j}(\omega)\mathbf{W}_{m,j}(\omega),
\end{equation}
with
\begin{align}\label{eq23}
\begin{split}
\mathbf{W}_{m,j}=\frac{1}{{4\pi {N_s}}}\int {d}{r^2}{\bf{r}}\left[ {{\tilde S}_{m,k}}\cdot\left( {{\partial _x}{S_0} \times {\partial _y}{S_0}} \right) \right. \\
\left. + {S_0}\cdot\left( {{\partial _x}{{\tilde S}_{m,k}} \times {\partial _y}{S_0}} \right) + {S_0}\cdot\left( {{\partial _x}{S_0} \times {\partial _y}{{\tilde S}_{m,k}}} \right) \right]\\
=\zeta_{m, j}(\hat{x}+i m \hat{y})\delta_{|m|, 1},
\end{split}
\end{align}
and
\begin{align}
\zeta_{m, j}=\frac{1}{4 N_s}\int dr r \left[ i m L_{X;m, j}\partial_r \theta-\partial_r\left(L_{Y; m, j} \sin\theta \right)\right]\delta_{|m|, 1}.
\end{align}
One important observation is that \emph{only the modes with $|m|=1$ couple to the skyrmion center motion}. The lowest mode with $|m|=1$ is the Goldstone mode corresponding to the translational motion of a rigid skyrmion. Other modes with $|m|=1$ lie in the magnon continuum. For a rigid skyrmion, the response to external drive is instantaneous, i.e. the inertia of a skyrmion is absent. Therefore the inertia of skyrmion is contributed from the excitation of the magnon continuum. Note that the linear analysis in Eq. \eqref{eq16} is valid when $\mathbf{R}(\omega)$ is smaller than the skyrmion size.
An alternative definition of the skyrmion center that is more relevant to experiments is based on the out-of-plane component of the spin
\begin{align}
{\bf{R}}' = \frac{{\int {d}{r^2}{\bf{r}}\left( {{S_z} - 1} \right)}}{{\int {{d}}{r^2}\left( {{S_z} - 1} \right)}}.
\end{align}
The equation of motion for ${\bf{R}}' $ to the linear order in perturbation has the same expressions as those in Eqs. \eqref{eq22} and \eqref{eq23}, except for $\zeta_{m, j}$,
\begin{align}
\zeta_{m, j}=\frac{\pi}{\int dr^2(S_{0, z}-1)}\int dr r^2 L_{Y; m, j} \sin\theta \delta_{|m|, 1}.
\end{align}
In the present work, we use the skyrmion center defined in Eq. \eqref{eq21}.
The external torque $\mathbf{\Gamma}$ is proportional to some control parameters $\mathbf{P}$, such as electric current density or magnetic field gradient. The response of velocity $\mathbf{v}=i\omega \mathbf{R}(\omega)$ to $\mathbf{P}$ can be expressed in a matrix form
\begin{align}\label{eq25}
\left(
\begin{array}{c}
v_x \\
v_y \\
\end{array}
\right)=\left(
\begin{array}{cc}
\chi _{11} & \chi _{12} \\
\chi _{21} & \chi _{22} \\
\end{array}
\right) \left(
\begin{array}{c}
P_x \\
P_y \\
\end{array}
\right).
\end{align}
The response is isotropic under the spatial rotation of $\mathbf{P}$ and $\mathbf{v}$, which requires $\chi_{22}=\chi_{11}$ and $\chi_{12}=-\chi_{21}$. Here $\chi_{ij}$ can be complex and the skyrmion trajectory is generally an ellipse. The skyrmion inertia manifests itself in the phase shift between the drive $\mathbf{P}(\omega)$ and velocity $\mathbf{v}(\omega)$. We can define a longitudinal phase shift $\Theta_{\parallel}$ and transverse phase shift $\Theta_{\perp}$,
\begin{align}
\tan\Theta_{\parallel}\equiv \frac{\mathrm{Im}[\chi_{11}]}{\mathrm{Re}[\chi_{11}]}, \ \ \ \tan\Theta_{\perp}\equiv \frac{\mathrm{Im}[\chi_{21}]}{\mathrm{Re}[\chi_{21}]}.
\end{align}
To cast Eq. \eqref{eq25} into the standard Thiele's form, we can introduce a generalized force $\mathbf{f}$ from $\mathbf{P}$, and rewrite Eq. \eqref{eq25} into
\begin{align}\label{eq26}
\left(
\begin{array}{cc}
\mathcal{D} +i \omega m & \mathcal{G}-i \omega \mathcal{A} \\
- \mathcal{G}+i \omega \mathcal{A} & \mathcal{D} +i \omega m \\
\end{array}
\right)\left(
\begin{array}{c}
v_x \\
v_y \\
\end{array}
\right)=\left(
\begin{array}{c}
f_x \\
f_y \\
\end{array}
\right).
\end{align}
Here $ \mathcal{D}\propto \alpha$ is the damping, $m$ is the mass, $\mathcal{G}$ is the gyrocoupling and $\mathcal{A}\propto \alpha$ is the gyrodamping. All those quantities are frequency dependent. The quantities at the left-hand side of Eq. \eqref{eq26} depend on the definition of the generalized force $\mathbf{f}$ at the right-hand side. In our discussions, we will focus on Eq. \eqref{eq25} since it captures fully the dynamics of a skyrmion.
In the limit of $\omega=0$, only the Goldstone modes with $|m|=1$ and $j=1$ contribute to the summation in Eq. \eqref{eq22}. In this limit, the rigid skyrmion approximation employed in the Thiele's collective coordinate approach becomes exact. From the Goldstone modes in Eq. \eqref{eq14}, we obtain $\alpha_{|m|=1,j=1}=-m\alpha/\kappa$ and $W_{|m=1|, j=1}=-(m i\hat{x}-\hat{y})/\sqrt{8\pi\kappa}$, where the skyrmion form factor $\kappa$ is
\begin{align}
\kappa=\int dr^2(\partial_{\mu} S_0)^2/(4\pi),
\end{align}
with $\mu=x,\ y$. Here $\kappa$ is of the order of unity, $\kappa\sim 1$. The equation of motion becomes
\begin{align}
\mathbf{v}(\omega)=\sum_{m=\pm1, j=1}\frac{i F_{m, j} (-m i\hat{x}+\hat{y})}{\sqrt{8\pi\kappa}(1-i m\alpha/\kappa)}.
\end{align}
\section{Applications}
In the following, we calculate the response matrix $\chi_{ij}$ for a skyrmion driven by a magnetic field gradient, spin transfer torque and spin Hall torque separately.
\begin{figure}[t]
\psfig{figure=fig1bb.pdf,width=\columnwidth}
\caption{(color online) Schematic view of a skyrmion in chiral magnets subjected to different drives: (a) a linear magnetic field gradient, (b) spin transfer torque and (c) spin Hall torque. The crosses in (c) represent the spin accumulation and spin current normal to the interface due to the spin Hall effect.
} \label{f1bb}
\end{figure}
\subsection{Magnetic field gradient}
We first consider the skyrmion motion driven by a magnetic field gradient as shown in Fig. \ref{f1bb} (a). This has been studied by numerical simulations recently \cite{moon_magnetic_2015,moon_skyrmion_2016}. We assume that the field is along the $z$ axis and the gradient is along the $x$ direction. The total magnetic field is $\mathbf{B}_T=B\hat{z}+\Delta_B^{(x)} r\cos\phi \hat{z}$. The magnetic field in the first term stabilizes a stationary skyrmion structure and that in the second term drives the skyrmion into motion. The unit of the field gradient is $D^3/(J_{\mathrm{ex}}^2 M_s)$ with $M_s$ the saturation field. The torque is then given by
\begin{align}
\mathbf{\Gamma} =-\mathbf{S}_0\times \left(\Delta_B^{(x)} r\cos\phi \hat{z}\right).
\end{align}
The coupling coefficient $F_{m,j}$ is
\begin{align}
F_{m,j}=-i \Delta_B^{(x)} \pi \delta_{|m|,1} \int dr r^2 L_{X;m,j} \sin\theta.
\end{align}
Here $F_{m, j}$ is nonzero only for $|m|=1$. In the rigid skyrmion approximation or $\omega\rightarrow0$, we have
\begin{align}
F_{|m|=1,j=1}=\frac{i \pi\Delta_B^{(x)}}{\sqrt{8\pi\kappa}}\int dr r (\sin\theta)^2.
\end{align}
The equation of motion is
\begin{align}\label{eq32}
\mathbf{v}(\omega)=-\frac{\Delta_B^{(x)}(\alpha \hat{x}+\kappa \hat{y})}{4 \left(\alpha ^2+\kappa ^2\right)}\int dr r (\sin\theta)^2.
\end{align}
For $\alpha\ll1$, the skyrmion moves almost perpendicular to the magnetic field gradient. There is additional velocity component antiparallel to the field gradient due to the damping.
The external drive in Eq. \eqref{eq25} is the field gradient $P_{\mu}=\Delta_B^{(\mu)}$. To include all the magnon modes, we calculate numerically the response matrix $\chi_{ij}$ and the results are shown in Fig. \ref{f2}. At $\omega=0$, $\chi_{ij}$ is real and the Thiele's equation of motion in Eq. \eqref{eq32} is recovered. We can see that $\mathrm{Im}[\chi_{ij}]$ is much smaller than $\mathrm{Re}[{\chi_{ij}}]$, because $\chi_{ij}$ is mainly contributed from the Goldstone modes. The phase shift near the magnon gap is $\Theta_{\parallel}\sim 10^{-3}$ and $\Theta_{\perp}\sim 10^{-4}$. The response is almost instantaneous and the inertia is weak. As shown in Fig. \ref{fa1}, the magnon modes associated with the Goldstone modes with $|m|=1$, $j=1$ are localized in the skyrmion, while the extended modes with $j>1$ only have small weight around the skyrmion. Therefore $|F_{|m|=1, j=1}|\gg |F_{|m|=1, j>1}|$ and $|\mathbf{W}_{|m|=1, j=1}|\gg |\mathbf{W}_{|m|=1, j>1}|$.
\begin{figure}[t]
\psfig{figure=fig2.pdf,width=\columnwidth}
\caption{(color online) Real and imaginary part of $\chi_{11}$ and $\chi_{21}$ in the case of a skyrmion driven by a magnetic field gradient. Here $B=0.8$ and $\alpha=0.2$.
} \label{f2}
\end{figure}
\begin{figure}[b]
\psfig{figure=fig3.pdf,width=\columnwidth}
\caption{(color online) Real and imaginary part of $\chi_{11}$ and $\chi_{21}$ in the case of a skyrmion driven by a spin transfer torque. Here $B=0.8$, $\alpha=0.1$ and $\beta=0.2$.
} \label{f3}
\end{figure}
\subsection{Spin transfer torque}
Here we study the skyrmion motion driven by a spin transfer torque as shown in Fig. \ref{f1bb} (b). We inject an electric current into to a chiral magnet. The electric current is polarized automatically by the skyrmion spin texture because of the Hund's coupling, which results in a spin current. For a nonuniform spin texture, the polarization of spin current changes in space. To conserve magnetic moment, there is transfer of magnetic moment between the spin current and localized spin texture, which generates a torque acting on the localized moments. The spin transfer torque can be expressed as \cite{Slonczewski1996,Tatara2008}
\begin{align}
\mathbf{\Gamma} = \left[ {{\partial _x}{\mathbf{S}_0} - \beta {\mathbf{S}_0} \times \left( {{\partial _x}{\mathbf{S}_0}} \right)} \right]{J_x}.
\end{align}
The current has unit of $2eD/(\hbar P_s)$ and is assumed to be along the $x$ direction, $\mathbf{J}=J_x\hat{x}$. The dimensionless parameter $P_s$ is the spin polarization factor. In the first approximation, the direction of spin current polarization is always parallel to the local magnetization vector. This contribution, called adiabatic spin transfer torque, is described by the first term at the right-hand side. The second term proportional to $\beta\ll 1$ describes the nonadiabatic and dissipative spin transfer torque, originating from the spatial mistracking of polarization of spin current and localized magnetization. \cite{PhysRevLett.93.127204} We then compute $F_{m,j}=-iJ_x(\tau_{m,j}^{\mathrm{STT}}+\beta \tau_{m,j}^{\mathrm{N}})$, with
\begin{align}
\tau_{m,j}^{\mathrm{STT}}=-\pi\int dr\left[ -i m L_{X;m,j}^*\sin(\theta)+r L_{Y;m,j}^*\partial_r\theta\right]\delta_{|m|,1}\delta_{j,1},
\end{align}
\begin{align}
\tau_{m,j}^{\mathrm{N}}=-\pi\int dr\left[ i m L_{Y;m,j}^*\sin(\theta)+r L_{X;m,j}^*\partial_r\theta\right]\delta_{|m|,1}.
\end{align}
The coupling coefficient $F_{m,j}$ vanishes for $|m|\neq 1$. Furthermore, for the adiabatic spin transfer torque, $\tau_{m,j}^{\mathrm{STT}}$, the current couples with the translational mode $\partial_x\mathbf{S}_0$. Therefore $\tau_{m,j}^{\mathrm{STT}}$ is nonzero only for the translational mode with $j=1$. The inertia of a skyrmion arises from the contribution of the nonadiabatic spin transfer torque, $\tau_{m,j}^{\mathrm{N}}$, and damping.
For a rigid skyrmion valid when $\omega\rightarrow0$, we have
\begin{align}
F_{|m=1|,j=1}=J_x\left(-m\sqrt{2\pi\kappa}+i\sqrt{\frac{2\pi}{\kappa}}\beta\right),
\end{align}
The equation of motion becomes
\begin{align}\label{eq37}
\mathbf{v}(\omega)=\left[-\frac{\kappa^2+\alpha\beta}{\alpha^2+\kappa^2}\hat{x}+\frac{(\alpha-\beta)\kappa}{\alpha^2+\kappa^2}\hat{y}\right]J_x(\omega),
\end{align}
which reproduces the well known Thiele's equation of motion, Eq. \eqref{eq0}. It also describes the \emph{full} equation of motion when $\beta=0$ because the adiabatic spin transfer only couples with the translational mode. In real materials, $\alpha\ll 1$ and $\beta\ll 1$, the skyrmion moves almost antiparallel to current, but there is a transverse motion due to the damping. We can define a Hall angle for the skyrmion motion $\tan\theta_H\equiv |v_y|/|v_x|=|\alpha-\beta|\kappa/(\kappa^2+\alpha\beta)$. For $\alpha=\beta$, Eq. \eqref{eq15} is Galilean invariant for a rigid skyrmion and the skyrmion moves exactly antiparallel to the current.
For $\beta>0$ and to deal with the skyrmion distortion, we compute numerically $\chi_{11}$ and $\chi_{21}$ at $B=0.8$ by taking the modes with $j>1$ into account, and the results are displayed in Fig. \ref{f3}. Here $\mathbf{P}$ in Eq. \eqref{eq25} is $\mathbf{P}=\mathbf{J}$. We have $\mathrm{Re}[\chi_{11}]\approx -({\kappa^2+\alpha\beta})/({\alpha^2+\kappa^2})$, $\mathrm{Re}[\chi_{21}]\approx {(\alpha-\beta)\kappa}/({\alpha^2+\kappa^2})$. The phase shift near the magnon gap is $\Theta_{\parallel}\sim 10^{-9}$ and $\Theta_{\perp}\sim 10^{-7}$. The contributions from other modes are negligible and the dynamics of a skyrmion is described by the Thiele's equation Eq. \eqref{eq37}. Therefore the response of a skyrmion to a spin transfer torque is instantaneous and the inertia of a skyrmion is negligible, compared to the viscous and the Magnus force. This justified the rigid skyrmion approximation used in Thiele's approach in this case.
\subsection{Spin Hall torque}
\begin{figure}[b]
\psfig{figure=fig4.pdf,width=\columnwidth}
\caption{(color online) Real and imaginary part of $\chi_{11}$ and $\chi_{21}$ in the case of a skyrmion driven by a spin Hall torque. Here $B=0.8$ and $\alpha=0.2$.
} \label{f4}
\end{figure}
We consider a bilayer system with a chiral magnet atop of a heavy metal, such as Ta and Pd, as depicted in Fig. \ref{f1bb} (c). We then apply a current in the heavy metal. Because of the spin Hall effect, there is spin current normal to the interface, which generates a torque in the chiral magnet. This is called the spin Hall torque and is given by \cite{emori_current-driven_2013,ryu_chiral_2013,PhysRevB.90.184427,liu_spin-torque_2012}
\begin{align}
\mathbf{\Gamma}={{\bf{S}}_0} \times \left( {{{\bf{S}}_0} \times \left( {\hat z \times {\bf{J}}} \right)} \right).
\end{align}
The current is in unit of $2 D^2 e d/(J_{\mathrm{ex}}\hbar\theta_{\mathrm{SH}})$ with $\theta_{\mathrm{SH}}$ being the spin Hall angle and $d$ the film thickness. For a current along the $x$ direction, $\mathbf{J}=J_x\hat{x}$, the coupling coefficient $F_{m,j}$ becomes
\begin{align}
F_{m,j}=-i\pi J_x\delta_{|m|,1} \int dr r (L_{Y;m,j}^* \cos\theta-i m L_{X;m,j}^*).
\end{align}
In the rigid skyrmion approximation or $\omega\rightarrow0$,
\begin{align}
F_{|m|=1,j=1}=-i\int dr^2\left[\frac{{{{-i}m}{\partial _x}{\mathbf{S}_0} - {\partial _y}{\mathbf{S}_0}}}{{\sqrt 2 \sqrt {\int{{d}}{r^2}{{\left( {{\partial _x}{\mathbf{S}_0}} \right)}^2}} }}\right]\cdot\mathbf{\Gamma}.
\end{align}
We obtain $F_{|m|=1,j=1}=0$ because the skyrmion has rotational symmetry. Therefore the rigid skyrmion does not couple to the spin Hall torque when the damping is neglected. The torque couples to the skyrmion motion through excitation of the magnon modes in the continuum and the weak off diagonal damping term in Eq. \eqref{eqalphamj}. Therefore the spin Hall torque is less efficient in driving the skyrmion compared to that of spin transfer torque. Here we provide an estimate on the current density required in order to achieve the same velocity for the cases with spin transfer torque and spin Hall torque. For the spin transfer torque, the velocity is almost frequency independent for a given current; while for the spin Hall torque, the velocity is maximal at a given current when the frequency is near the magnon gap. Taking the optimal velocity for the spin Hall torque, to attain the same velocity, the required current density for both torques is about $J_{\mathrm{SHE}}\theta_{\mathrm{SH}}J_{\mathrm{ex}}/J_{\mathrm{STT}} d D P_s\sim 10^3$, where $J_{\mathrm{SHE}}$ ($J_{\mathrm{STT}}$) is the current density in the case of spin Hall (transfer) torque. Therefore $J_{\mathrm{SHE}}$ is larger than $J_{\mathrm{STT}}$ by several orders of magnitude for materials with $\theta_{\mathrm{SH}}J_{\mathrm{ex}}/ d D P_s\sim 1$. This is consistent with the experimental observations that the skyrmion velocity at a given current in the case of spin transfer torque is much bigger than that in the case of spin Hall torque \cite{Schulz2012,Jiang17072015,jiang_direct_2017}, although different pinning strength in these systems may also partially account for the difference.
In the present case, $\mathbf{P}$ in Eq. \eqref{eq25} is $\mathbf{P}=\mathbf{J}$ .The numerical results of $\chi_{ij}$ are presented in Fig. \ref{f4}. Since $\chi_{21}\sim i \chi_{11} $, the trajectory is roughly a circle. One prominent feature is that $\chi_{ij}(\omega)$ develops a resonance-like feature around the magnon gap, $\omega\approx \omega_g$. This can be understood as follows. For a given angular momentum $|m|=1$, the density of state of magnon, $\rho(\omega)=1/\sqrt{\omega-\omega_g}$ , diverges at $\omega_g$. The equation of motion of a skyrmion by including all magnon excitations is
\begin{align}\label{eq40}
{\mathbf{R}}(\omega)=\sum_{|m|=1, j}\frac{F_{m,j} \mathbf{W}_{m,j}}{\omega-\omega_{m,j}+i \alpha_{m, j}\omega} \nonumber\\
\approx F(\omega_g) \mathbf{W}(\omega_g)\int_{\omega_g}^{+\infty} d\omega \rho(\omega)\frac{1}{\omega-\omega_{m,j}+i \alpha(\omega_g)\omega} \nonumber\\
=2F(\omega_g) \mathbf{W}(\omega_g)\frac{1}{\sqrt{\omega-\omega_g+ i \alpha(\omega_g)\omega}},
\end{align}
where $F(\omega_g)$, $\mathbf{W}(\omega_g)$ and $\alpha(\omega_g)$ are the corresponding quantities evaluated at the magnon gap. From Eq. \eqref{eq40}, it is clear that $\chi_{ij}$ develops peaks around $\omega\approx \omega_g$, in agreement with the results in Fig. \ref{f4}.
\section{N\'{e}el skyrmion}
Skyrmion can also be stabilized at interfaces of heterostructure, where the inversion symmetry is broken explicitly \cite{Heinze2011,Fert2013,Romming2013}. The skyrmion at interface has helicity of $0$ and is called N\'{e}el skyrmion. The effective Hamiltonian supporting N\'{e}el skyrmions in two dimensions, $(x,\ y)$, can be written as \cite{thiaville_dynamics_2012,PhysRevB.90.184427}
\begin{equation}\label{Neeleq1}
\mathcal{H}_n= \frac{{{J_{\mathrm{ex}}}}}{2}\mathop \sum \limits_{\mu = x,y} {\left( {{\partial _\mu }{\bf{n}}} \right)^2}+D[n_z(\nabla\cdot \mathbf{n})-(\mathbf{n}\cdot\nabla)n_z]-\mathbf{B}\cdot\mathbf{n}.
\end{equation}
The Bloch skyrmion described by Eq. \eqref{eq1} can be obtained by a global rotation of spin associated with the N\'{e}el skyrmion along the magnetic field direction by $\pi/2$, i.e. $\mathbf{S}=\hat{O}_n \mathbf{n}$, with the spin rotation operator
\begin{equation}\label{Neeleq2}
\hat{O}_n= \left(
\begin{array}{ccc}
0 & -1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1 \\
\end{array}
\right).
\end{equation}
One can verify that $\mathcal{H}_n$ can be obtained from $\mathcal{H}$ in Eq. \eqref{eq1} by the same global rotation of spins. Therefore the eigenfrequencies of the internal modes for the Bloch and N\'{e}el skyrmions are identical, and the eigenmodes for two skyrmion textures are related by a spin rotation.
One can also connect the equation of motion for the Bloch skyrmion to that of the N\'{e}el skyrmion by the same spin rotation. For a Bloch skyrmion driven by a magnetic field gradient with polarization perpendicular to the layer, spin transfer torque and spin Hall torque, the equation of motion for spins is
\begin{align}\label{Neeleq3}
{\partial _t}{\bf{S}} = - {\bf{S}} \times {{\bf{H}}_{\mathrm{eff}}}(\mathbf{S}) + \alpha {\bf{S}} \times {\partial _t}{\bf{S}} + \left( {\mathbf{J}_{\mathrm{STT}}\cdot\nabla } \right){\bf{S}} \nonumber\\
- \beta \mathbf{S} \times \left( {\mathbf{J}_\mathrm{STT}\cdot\nabla } \right){\bf{S}} + {\bf{S}} \times \left[ {{\bf{S}} \times \left( {{\bf{\hat z}} \times {{\bf{J}}_{\mathrm{SHE}}}} \right)} \right].
\end{align}
After the spin rotation, the equation of motion for spins associated with a N\'{e}el skyrmion has the same form as Eq. \eqref{Neeleq3} if we replace the effective field by $\mathcal{H}_{\mathrm{eff}}(\mathbf{n})\equiv-\delta \mathcal{H}_n/\delta \mathbf{n}$ and $\mathbf{J}_{\mathrm{SHE}}\leftarrow \hat{O}_n^{-1} \mathbf{J}_{\mathrm{SHE}} $. Therefore the results derived for the Bloch skyrmions can be applied directly to the N\'{e}el skyrmions.
\section{Discussions and conclusions}\label{Sec4}
The current can generate a magnetic field gradient which affects the skyrmion trajectory. The magnetic field gradient depends on the sample geometry. In our linear approximation, we need to superpose the trajectory due to the current on the trajectory due to the induced field gradient. Meanwhile, the skyrmion motion distorts magnetic field outside of the magnetic plate, which contributes to the skyrmion kinetic energy. In the adiabatic approximation, a moving skyrmion induces an electric field $\nabla\times {\bf E}=(1/c)\dot{ {\bf B}}=-(\mathbf{v}\cdot\nabla) {\bf B}/c$ according to the Faraday's law. As a result its kinetic energy increases as $\int d^2r {\bf E}^2/(8\pi)=M_{\rm em}\dot{{\bf R}}^2$ which determines the electromagnetic part of the skyrmion mass tensor $M_{\rm em}$. Such an inertia effect has relativistic origin and is proportional to $(v/c)^2$. Since the maximal skyrmion velocity, $10^2\ \mathrm{m/s}$, \cite{szlin13skyrmion1} is much smaller than the light velocity, $10^8\ \mathrm{m/s}$, the correction to skyrmion dynamics due to the electric field outside the magnetic plate is negligible.
Let us compare our results on the inertia of a skyrmion to those in literatures. The mass of a skyrmion bubble was calculated and measured in Refs. \onlinecite{Makhfudz2012,buttner_dynamics_2015} and they found that the mass term is important in order to reproduce the skyrmion trajectory. In the calculations \cite{Makhfudz2012}, the authors calculate the mass from the edge modes of the bubble. In experiments \cite{buttner_dynamics_2015}, the mass was obtained by fitting the experimentally measured trajectory determined from the spin polarization to equation of motion of a skyrmion as a particle with a mass term, assuming a hormonic potential for the skyrmion bubble. In Refs. \onlinecite{Makhfudz2012,buttner_dynamics_2015}, the skyrmion bubbles are stabilized by the dipolar interactions, and the magnon excitations are expected to be quite different from those of a skyrmion in chiral magnets. This may be the origin of the difference between our results the those in Refs. \onlinecite{Makhfudz2012,buttner_dynamics_2015}. This also suggests that the skyrmions in chiral magnets are more rigid and are advantageous for applications. The inertia of a skyrmion in chiral magnets was studied recently in Ref. \onlinecite{PhysRevB.90.174434}. They calculated numerically the trajectory of a skyrmion from the Landau-Lifshitz-Gilbert equation with a thermal noise, where the skyrmion diffuses in the sample. They then fit the trajectory to the equation of motion for a skyrmion as a particle according to Eq. \eqref{eq26}, from which they extracted the mass and gyrocoupling coefficient. They found that the inertia is important to describe the skyrmion diffusion. For a skyrmion driven by a spin transfer torque, they found the response of a skyrmion to current is almost instantaneous, consistent with our results. It is unclear the importance of nonlinearity in their calculations.
The dynamics of skyrmions depends on its internal modes. The analysis here is applicable for skyrmions stabilized by the DM interaction, where the helicity of skyrmions is not a Goldstone mode. Skyrmions can also exist in inversion-symmetric magnets with competing interactions \cite{PhysRevLett.108.017206,leonov_multiply_2015,PhysRevB.93.064430,PhysRevB.93.184413}. Due to the preservation of the inversion symmetry, the helicity of skyrmion is also a Goldstone mode, in addition to the mode associated with translational motion. The dynamics can be very different, as has been demonstrated recently in Refs. \onlinecite{PhysRevB.93.064430,zhang_skyrmions_2017}. Skyrmions with higher topological charge can also be stabilized \cite{PhysRevB.95.094423}. The internal modes are different for skyrmions with higher topological charge, and the dynamics of these skyrmions requires a separate study.
To conclude, we have developed a linear theory to calculate the equation of motion of a skyrmion by taking all the magnon modes into account. We calculate the skyrmion velocity in response to external drives, such as a magnetic field gradient, spin transfer torque and spin Hall torque. The skyrmion dynamics is only governed by the magnon modes with an angular momentum $|m|=1$. The inertia of a skyrmion is contributed from the magnon continuum. For a skyrmion driven by a magnetic field gradient or spin transfer torque, the dynamics is dominated by the Goldstone modes corresponding to the translational motion of a skyrmion because the eigen function of the Goldstone modes have maximal weight around the skyrmion center, while the modes in the magnon continuum only has very little weight around the skyrmion. The response of a skyrmion is instantaneous and our calculations justify the rigid skyrmion approximation employed in the Thiele's collective coordinate approach. In the case of a spin Hall torque, the skyrmion motion couples to the torque through the magnons in the continuum and the Gilbert damping. Since the magnon density of state diverges around the magnon gap, there are resonances in the response functions around the magnon gap. The trajectory of a skyrmion is an ellipse for a skyrmion driven by a spin Hall torque at finite frequencies. The inertia of a skyrmion can be quantified in experiments by measuring the phase shift between the external drive and velocity. For applications, it is desirable for skyrmion to respond to an external drive without delay or retardation even at high frequencies. This can be achieved by driving a skyrmion with a spin transfer torque. Our results establish the connection between the skyrmion dynamics and its magnon spectrum, and shed new light on the skyrmion dynamics.
\begin{acknowledgments}
The author is indebted to Lev N. Bulaevskii, Avadh Saxena, Cristian D. Batista and Satoru Hayami for helpful discussions. This work was carried out under the auspices of the U.S. DOE Contract No. DE-AC52-06NA25396 through the LDRD program.
\end{acknowledgments}
|
\section{Introduction}\label{sec:intro}
The mechanical interaction between a moving mirror and the radiation field has always been an interesting issue of study. This is not only due to its practical purposes but also to its representation of a fundamental system in quantum optics. Besides the change of the zero point energy of the quantum vacuum provoked by static boundary conditions a second, there is a yet even more fascinating feature of the quantum vacuum arising when considering dynamical boundaries conditions. The presence of moving boundaries leads to a non stable vacuum electromagnetic state, resulting in the generation of real photons, which is an amazing demonstration of the existence of quantum vacuum fluctuations of quantum electrodynamics (QED), referred to in the literature as the dynamical Casimir effect (DCE) \cite{reviews} or motion-induced radiation. DCE is a common name ascribed to the processes in which photons are generated from vacuum due to the external time variation of boundary conditions for some field \cite{12,13,15}. For the usual electromagnetic case this corresponds to the fast motion of a mirror or modulation of the dielectric properties of the mirror/ intra-cavity medium \cite{16}.
Research in the field has mainly concentrated on one-dimensional models and a few works with more realistic three dimensional models \cite{fulling,dodonov1,mundarian,dodonov2,Ruser}. Since the amount of radiation generated is very small, much attention has been paid to the study of one-dimensional model for which the effect is enhanced, as for example through the parametric resonance condition. The main difference between one and three-dimensional cavities is that, while in one dimension the cavity's frequency spectrum is equidistant and leads to strong intermode interactions, in three dimensions the spectrum is in general non equidistant, and only a few modes may be coupled \cite{crocce,dodonov3}.
A cavity made of two perfectly parallel reflecting mirrors, one of which oscillates with a mechanical frequency equal to a multiple of the fundamental of the static cavity (while the other one is at rest) is a typical case where the mentioned enhancement takes place. In most works this problem has been analytically studied through a perturbative expansion of the equations of motion of the field in terms of the small oscillation amplitude to find an approximative solution at short times. Although the direct measurement of radiation generated by moving mirrors is an important experimental challenge, it was asserted \cite{Wilson} that photon creation induced by time-dependent boundary conditions has been observed experimentally in superconducting circuits. This experiment consists of a coplanar waveguide terminated by a superconducting quantum interference device (SQUID), upon which a time-dependent magnetic flux is applied. A related experiment involving a Josephson metamaterial embedded in a microwave cavity has been described in Ref. \cite{Paraoanu}. These experiments stimulated new theoretical research on role of dynamical Casimir physics in quantum information processing, quantum simulations and engineering of nonclassical states of light and matter \cite{19,20,21,22,23}. There are also ongoing experiments aimed at measuring the photon creation induced by the time-dependent conductivity of a semiconductor slab enclosed by an electromagnetic cavity \cite{Bragio}, as well as proposals based on the use of high frequency resonators to produce the photons, and ultracold atoms to detect the created photons via superradiance \cite{Kim}.
Recent studies have indicated that DCE could be implemented even using a single two-level atom (qubit) with time-dependent parameters, such as the transition frequency or the atom-field coupling strength \cite{5, 25,26,27,28}. Generation of excitation from vacuum occurs due to the counter-rotating terms in the Rabi Hamiltonian, which for many years had been neglected under the Rotating Wave Approximation. On the other side,
it has been proved that an ensemble of two-level atoms collectively coupled to the electromagnetic field of a cavity, driven at low frequencies and close to a quantum phase transition, stimulates the production of photons from the vacuum. This paves the way to an effective simulation of the DCE \cite{vedral}.
The case of cavities with two moving mirrors has also been considered due to a perturbative treatment \cite{chinos} and also using a technique inspired in the renormalisation group method, where the solution to the set of generalised Moores' equations is valid both for short and long times, improving perturbative approaches \cite{DalvityDiego}. Authors considered two mirrors oscillating resonantly at the same frequency, allowing for different amplitudes and dephasing between mirrors in one dimensional cavities.
As we shall see, the radiation induced strongly depends on the relation among the amplitudes, the frequency and the different phase in the walls oscillations. As predicted in \cite{DalvityDiego}, we show that for some relations among the variables, there is constructive interference which leads to an exponential growth of particles inside the cavity. For some other relations, there is destructive interference and hence no vacuum radiation. We also show that our solution accounts for other physical solutions (non perturbative regime for example) as for one oscillating wall.
In this paper we will present a detailed numerical analysis of the particle creation rate, along with analytical considerations on
the cases mentioned above.
The paper is organised as follows. In Sec. \ref{bc} we shall present the equation of motion for the field modes of the electromagnetic field. In Sec. \ref{onewall}, we begin by analysing the photon creation in the case where one wall of the cavity is at rest and the other one oscillates with a multiple frequency of the fundamental of the static cavity. Therein, we shall consider a three dimension as well as one dimensional cavities. In Sec. \ref{2wall}, we concentrate on the case of two oscillating mirrors, one at each end of the cavity. We focus on the study of different cases of dephasing between the oscillating walls, i.e. zero and $\pi$-dephasing movements. Finally, in Sec.\ref{conclusiones} we make our conclusions.
\section{Boundary Conditions}
\label{bc}
We consider a rectangular cavity formed by perfectly conducting walls with dimensions $L_x$ ($L_y$ and $L_z$ as well if we consider a three dimensional cavity). The mirrors placed at $x=L(t)$
and $x=R(t)$ are at rest for $t<0$ and begin to move at $t=0$ following a given trajectory $L(t)$ and $R(t)$ respectively.
We assume these trajectories as prescribed for the problem and that they work as a time-dependent boundary condition for the field.
We start with the field operator $A(x,t)$ for the vector potential which satisfies the wave equation $\Box \vec{A}=0$. In terms of the creation
$a_k^{\dagger}$ and annihilation $a_k$
operators, the field operator can be expressed as:
\begin{equation}
A(x,t)= \sum_k^{\infty} \bigg[ \hat{a}_k \psi_k (x,t) + \hat{a}_k^{\dagger} \psi_k^*(x,t) \bigg].
\end{equation}
In the previous equation, $\psi_k(x,t)$ are the mode functions of the field and are chosen so as to satisfy the boundary conditions, i.e. $\psi_k(L(t),t)=0$ and $\psi_k(R(t),t)=0$.
We shall firstly consider the instantaneous mode basis, for the $1+1$ field \cite{comentario}:
\begin{equation}
\phi_k(x,t)=\sqrt{\frac{2}{R-L}} \sin \left[\frac{k \pi (x-L)}{R-L} \right],
\end{equation}
and write each mode as:
\begin{equation}
\psi_k(x,t)=\sum_{m=1}^\infty Q_{k m}(t) \phi_m (x,t),
\end{equation}
where $m$ is a positive integer.
By considering small amplitude motions of the walls, we can write their trajectories as given by:
\begin{equation}
L(t) = A_L \epsilon_L \sin(\Omega_L t + \phi_L)
\end{equation}
and
\begin{equation}
R(t)=L_0 - A_R \epsilon_R \sin(\phi_R) + A_R \epsilon_R \sin (\Omega_R t + \phi_R),
\end{equation}
where $L_0$ is the cavity length in the static situation, and $\epsilon_R$, $\epsilon_L$
are small (dimensionless) parameters which characterise the small deviations of the walls from the initial static positions, and
$A_L$ and $A_R$ are amplitudes. The mirrors can oscillate in phase or not, depending on the values of $\phi_L$ and $\phi_R$.
By inserting the expansion of field modes into the wave equation and integrating over spatial dimensions leads to the equation of motion for the canonical variables expressed as:
\begin{widetext}
\begin{eqnarray}
\ddot {Q}_m ^{(n)} + \omega^2_m(t) Q_m^{(n)} &=& \frac{1}{L_0}\sum_s b_{ms} \dot{Q}_s^{(n)} +\frac{1}{4 L_0^2} \sum_s \bigg
\{ g_{ms} + (3 (\dot R - \dot L)^2 + 2 L_0 (\ddot R- \ddot L) ) a_{ms} +
(r_{ms}^1+ r_{ms}^2) \bigg\} Q_s^{(n)},
\label{ecmov}
\end{eqnarray}
\end{widetext}
where, in the general 3+1 dimensions case,
$\omega_m(t)= \sqrt{\frac{(m \pi)^2}{(R(t)-L(t))^2}+ k_{\parallel}^2}$ is the mode frequency. The coefficients in the equation
are defined as follows:
\begin{widetext}
\begin{equation*}
a_{ms}= \left\lbrace
\begin{array}{l}
1 \text{ if } m = s \\
0 \text{ otherwise} \\
\end{array}
\right.
\end{equation*}
\begin{equation*}
b_{ms}= \left\lbrace
\begin{array}{l}
\frac{4 m s }{(m^2-s^2)} (\dot R (-1)^{m+s}-\dot L) \text{ if } m \neq s \\
0 \text{ otherwise} \\
\end{array}
\right.
\end{equation*}
\begin{equation*}
g_{ms}= \left\lbrace
\begin{array}{l}
\frac{m s }{(m^2-s^2)} \left[24 \dot L (\dot R- \dot L)-24 \dot R (\dot R-\dot L) (-1)^{m+s}+8 L_0 ( \ddot R(-1)^{m+s}- \ddot L\right]
\text{ if } m \neq s \\
3 (\dot L - \dot R)^2)-2 (L_0 (\ddot R - \ddot L) \text{ if } m=s \\
\end{array}
\right.
\end{equation*}
\begin{equation*}
r_{ms}^1= \left\lbrace
\begin{array}{l}
\frac{16 m^3 s }{(m^2-s^2)^2} \dot L (\dot R- \dot L)(-1+(-1)^{m+s}) \text{ if } m \neq s \\
2 m^2 \pi^2 \dot L (\ddot R + \ddot L) \text{ if } m=s \\
\end{array}
\right.
\end{equation*}
\begin{equation*}
r_{ms}^2= \left\lbrace
\begin{array}{l}
\frac{-16 m^3 s }{(m^2-s^2)^2} ( \dot R-\dot L) [\dot L+(\dot L-2 \dot R)(-1)^{m+s} ]\text{ if } m \neq s \\
\frac{2}{3} (\dot R -\dot L) (3 (\dot L-\dot R)+ m^2 \pi^2 (\dot R + \dot L)) \text{ if } m=s \\
\end{array}
\right.
\end{equation*}
\end{widetext}
\section{Numerical Method}
In this Section we describe the numerical method used for solving the equation of motion of the field modes determined by Eq.(\ref{ecmov}).
In order to solve the equation of motion of the $n$ modes, we perform a change of variables in order to obtain a new system of equations:
\begin{eqnarray}
\dot{Q}_m &=& U_m, \nonumber \\
\dot{U}_m &=& -\omega_m^2(t) Q_m + \sum_s S_{m s} (t) Q_{s},
\label{modelo}
\end{eqnarray}
where $S_{ms}(t)$ is a bracket proportional to $Q_s$ in Eq.(\ref{ecmov}).
We have dropped the supra-index for simplicity.
The initial conditions, specified for each field mode in all cases are:
\begin{equation}
Q_k^{(n)}(0)=\frac{1}{\sqrt{2 \omega_n} }\delta_{k,n}; ~~~~~~\dot{Q}_k^{(n)}(0)= -i \sqrt{\frac{\omega_n}{2} }\delta_{k,n};
\end{equation}
which indicate that the field modes and their derivatives are continuous at $t=0$, as long as $R(t)$, $L(t)$, $\dot{R}(t)$ and $\dot{L}(t)$ are smooth functions.
For a time dependence of the boundary, either $R(t)$ or $L(t)$, which is not sufficiently smooth, (for example discontinuities in its time-derivative), one may expect spurious particle creation.
We have used an integration scheme based on a fourth order Runge-Kutta-Merson numerical method between $t=0$ and a maximum time $t_{\rm max}>0$.
In all cases, the moving walls are at rest at $t=0$ and then, the perturbation is turned on for times between $0<t<t_F$, with $t_F<t_{\rm max}$, where the walls remains static again (this can be applied to either one or two moving walls). For times $t<0$ and $t> t_F$, the cavity is a static one and we know the set of orthonormal functions. The quantisation of the system is straightforward through creation and annihilation operators:
\begin{eqnarray}
Q_n (t<0) &=& \frac{1}{\sqrt{2 \omega_n}} ( \hat{a}_n e^{-i \omega_n t} + \hat{a}_n^{\dagger} e^{i \omega_n t} ), \label{freemode}
\end{eqnarray}
with frequency $\omega_n(t)=1/L_0\sqrt{(\pi n)^2 + k_{\parallel}^2}$, where $L_0$ is the initial length of the cavity and $k_{\parallel}$ is associated to the non-dynamical dimensions of de cavity ($L_y$ and $L_z$).
The time-independent annihilation and creation operators $\hat{a}_n$, $\hat{a}^{\dagger}_ n$ associated with the particle notion for $t \leq 0$ are subject to the commutation relations $[\hat {a}_n, \hat {a}_m]=[\hat {a}_n^{\dagger}, \hat {a}_m^{\dagger}]=0$ and $[\hat {a}_n, \hat {a}_m^{\dagger}]=\delta_{nm}$. The initial vacuum state $|0,t \leq 0 \rangle$ is defined by:
\begin{equation}
a^{\dagger}_{n} |0,t \leq 0 \rangle =0 ~~~\forall ~~n.
\end{equation}
When the cavity dynamics is switched on at $t=0$ and the walls follow the prescribed trajectory $L(t)$ and $R(t)$, the field modes are coupled. Then, the $Q_n$ can be written as:
\begin{equation}
Q_n(t \geq 0) = \sum_m \frac{1}{\sqrt{2\omega_m}} ( \hat{a}_m \epsilon_n^m(t) + \hat{a}_m^{\dagger} \epsilon_n^*(t)),
\end{equation}
with complex functions $\epsilon_n^m(t)$ that satisfy the equation of modes. When the motion ceases and the walls are at rest again for $t>t_F$, $Q_n(t)$ can be expressed again as:
\begin{eqnarray}
Q_n (t \geq t_F) &=& \frac{1}{\sqrt{2 \omega_n^{1}}} ( \hat{A}_n e^{-i \omega_n^1(t-t_F)} + \hat{A}_n^{\dagger} e^{i \omega_n^1 (t-t_F)} ),
\end{eqnarray}
with $ \omega_n^{1} = \omega_n^{1}(t \geq t_F)$ and the annihilation and creation operators $\hat{A}_n$ and $\hat{A}_m^{\dagger}$ corresponding to the particle notion for
$t \geq t_F$. The final vacuum state $|0, t \geq t_F \rangle$ is defined by:
\begin{equation}
A^{\dagger}_{n} |0,t \leq t_F \rangle =0 ~~~\forall ~~n.
\end{equation}
As expected, the initial state particle operators $\hat{a}_n$ and $\hat{a}_n^{\dagger}$ are linked to the final state operators $\hat{A}_n$ and $\hat{A}_n^{\dagger}$ by a Bogolubov transformation $\hat{A}_n= \sum_m (A_{mn}(t_F) \hat{a}_m + B_{mn}^*(t_F) \hat{a}_m^{\dagger} )$.
The total number of particles created in a mode $n$ during the motion of the wall is given by the expectation value of the particle number operator $\hat{A}_n^{\dagger} \hat{A}_n$ associated with the particle notion for $t \geq t_F$ with respect to the initial vacuum state:
\begin{equation}
N_n(t_F)=\langle 0, t\leq |A^{\dagger}_n A^{\dagger}_n |0, t\leq 0 \rangle = \sum_m |B_{m n }(t_F)|^2.
\label{Nparticulas}
\end{equation}
In order to obtain the numerical results presented in the following Sections we proceed in the following way. Two cut-off parameters $\Lambda$ (for the field modes considered) and $\Lambda_m$ (for the number of canonical variables considered) are introduced to make the system of differential equations finite and suitable for a numerical treatment. The system of $n \times m$ coupled differential equations is then evolved numerically from $t = 0$ up to a final time $t_F$ and the expectation value of Eq.(\ref{Nparticulas}) is calculated for several times in between. By doing so we interpret $t_F$ as a continuous variable such that Eq.(\ref{Nparticulas}) becomes a continuous function of time. Consequently, the stability of the numerical solutions with respect to the cut-offs has to be ensured. In particular $\Lambda$ will be chosen such that the numerical results for the number of particles created in single modes are stable. In most cases, it is enough to choose $\Lambda_m=\Lambda$.
In our units, the spectral modes $k_n=\Omega_n$ are given in units of $1/L_0$ ($k_n L_0$ is dimensionless) and consequently time is measured in units of $L_0$.
\section {One moving mirror}
\label{onewall}
We can start by studying the photon creation when only one wall is moving following $R(t)$ and the other one is at rest in $x=0$ (for example, by setting $\epsilon_L=0$ and $\epsilon_R \neq 0$). In this case, we can take any value of $\phi_R$, say $\phi_R=0$. If we excite the cavity with an external frequency $\Omega_R$ such that $\Omega_R= 2 \omega_1$, we shall produce parametric resonance induced by the moving mirror at $x=R(t)$. We can consider either a one dimensional cavity or a three dimensional one, since this approach can be applied to either $1+1$ or $3+1$ dimensions, by taking into account that in the latter case, the ``moving walls" are in the x-direction, while the field satisfies Dirichlet boundary conditions in the other walls of the cavity
($y$ and $z$ directions). The important difference between one- and higher-dimensional cavities is that the frequency spectrum in only one spatial dimension is equidistant while it is in general non-equidistant for more spatial dimensions. An equidistant spectrum yields strong intermode coupling whereas in case of a non-equidistant spectrum only a few or even more modes may be coupled allowing for exponential photon creation in a resonantly vibrating three-dimensional cavity.
For both Dirichlet boundary conditions, the eigenfrequencies inside the cavity satisfy the following condition:
\begin{equation}
\omega_n=\frac{1}{L_0}\sqrt{(\pi n)^2 + M^2},
\label{dirichletfrequencies}
\end{equation}
where $n$ is natural number and we have set $M^2 \equiv {k_\parallel}^2$. If the field is massless (which corresponds to a one dimensional cavity), then the spectrum is equidistant, i.e. the difference between two consecutive eigenfrequencies is constant. Otherwise, if $M$ has arbitrary nonzero values, the spectrum is non-equidistant, corresponding to the one of a three-dimensional cavity.
\subsubsection{Three-Dimensional cavity}
As we have stressed at the end of the previous Section, when considering a three dimensional cavity, the parallel component of the wave number $k_{\parallel}=\pi \sqrt{(n_y/L_y)^2+ (n_z/L_y)^2}$ can be associated with the non-dynamical cavity dimensions, can be identified with the ``mass" of a massive field \cite{Ruser}. Consequently, the number of TE-mode photons created in a three-dimensional cavity equals the number of scalar particles of mass $k_{\parallel}$ created in a one-dimensional cavity. Then, we perform the simulations by considering $\Omega=\Omega_R= 2 \omega_1$ and different values of a ``mass" $M$ for simulating the particle creation in a three-dimensonal cavity.
\begin{figure}[h!]
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=7.5cm]{casomasivo1pared.eps}\end{center}
\end{minipage}
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=7.5cm]{casonomasivo1pared.eps}
\end{center}
\end{minipage}
\caption{(Color online) (a) Left: Behaviour of the $\vert{B}_1\vert^2$ coefficient as a function of the dimensionless time, for different values of $M$ under the perturbation of $\Omega= 2 \omega_1$. The grey solid (upper) curve represents the analytical solution for the number of particles created when $M=10$, while the blue solid (lower) curve is the numerical solution obtained for the right-moving mirror. (b) Right: Behaviour of the $\vert {B}_1\vert^2$ coefficient as a function of the dimensionless time, for smaller values of $M$, i.e. $M=0.01$, $M=0.05$ and $M=1$.
Parameters used: $\epsilon=0.001$, $\Lambda=10$. }
\label{caso1pared}
\end{figure}
In Fig.\ref{caso1pared}, we can see the behaviour of the $\vert {B}\vert^2$ coefficient of the mode $n= 1$ of the field for different values of $M$, say $M=1$, $M=5$ and $M=10$. It
is expected that the particle creation would be exponential as the frequency spectrum becomes more non-equidistant. This is achieved, more evidently, for bigger values of $M \gg \pi/L_0$.
In a resonant vibrating cavity, the number of TE-mode photons created in the resonant mode $n$ increases exponentially in time as Ref \cite{Ruser}:
\begin{equation}
N_n(t) = \sinh \bigg(\frac{n^2 \pi^2 \epsilon_R t}{2\Omega L_0^2}\bigg)^2.
\label{ruser}
\end{equation}
We can see in Fig.\ref{caso1pared} the expected exponential behaviour as the mass increases. The grey solid curve (parallel to $M=10$) represents $N_1(t)$ as computed for $M=10$ in \cite{Ruser} using Eq.(\ref{ruser}). We can also see that the behaviour is similar though in our cavity it is attenuated by the multiplicative factor included in the relationship between the number of created particles $N_n$ and the coefficient $\vert B_n\vert^2$. As we decrease the value of $M$, we reach a region of values where we can neglect the mass term inside the frequency definition and obtain an equidistant spectrum. In Fig.\ref{caso1pared} (b) we can see that the behaviour induced by the resonant frequency $\Omega= 2 \omega_1$ when the $M \rightarrow 0$ changes considerably and can not be fitted by Eq.(\ref{ruser}) any longer.
\subsubsection{One dimensional cavity}
When studying the case $M = 0$ the numerical results should converge towards the well known results for the massless case where all modes are coupled. In this case, we know that an eigenfrequency $k$ is defined by $\omega_n = (n \pi/L_0)$. Then, if we excite the system with $\Omega= 2 \omega_1$, we should obtain a quadratic behaviour for the field mode at times $\Omega t << 1/\epsilon$. For a later time, we should expect a linear behaviour. In Fig.\ref{caso1paredM001}, we present the behaviour of $\vert {B}_1\vert^2$ for a value of $\epsilon =0.01$. For very short times, we can observe a quadratic behaviour and then a linear one. However, for values bigger than $\Omega t \gg 1/\epsilon$ we can see that the growth becomes exponential. The exponential behaviour for longer times is equivalently to what has been found in Ref. \cite{nosotros}. The coupling to a big number of field modes derives in an exponential growth at longer times, in the very non-perturbative regime.
\begin{figure}[h!]
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=7.5cm]{caso1paredeps01M001.eps}
\end{center}
\end{minipage}
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=7.5cm]{caso1paredeps001M001.eps}
\end{center}
\end{minipage}
\caption{(Color online) In both plots, we can see the behaviour of the $\vert {B_1}\vert^2$ coefficient of the mode $\omega_1$ of the field as a function of the dimensionless time. For times smaller than $1/\epsilon$, the behaviour can be fitted by a quadratic curve, while for times $\Omega t \sim 1/\epsilon$ with a linear in time curve. For very long times, we find an exponential behaviour as shown in Ref. \cite{nosotros}. (a) Right: we use $\epsilon_R=0.01$ and. (b) Left: we set $\epsilon_R=0.001$. Parameters used: $M=0$, $\Lambda=10$. }
\label{caso1paredM001}
\end{figure}
For equidistant spectrum, we can see that the coupling between an infinite number of modes leads to a quadratic growth in time of the number of at short timescales ($\sim 1/(\Omega\epsilon_R)$) and a linear one in the long time limit ($\sim 1/(\Omega^2\epsilon_R^2)$) as expected.
In Fig. 2(a) and 2(b) we show the dependence upon $\epsilon$ of the quadratic and linear regimes. For example $\epsilon_R=0.001$, we can observe the same behaviour but in a different temporal scale, as shown in Fig. \ref{caso1paredM001}(b).
We can even compute the energy density for the case of a moving mirror for a one-dimensional field. In this case, the frequency spectrum is equidistant and the energy grows quadratically for short times, as in Fig.\ref{caso1paredM001}. In Fig.\ref{energia1pared}, we show the energy behaviour for an $\epsilon_R=0.001$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{energia1parede001.eps}
\caption{(Color online) Energy density for a moving wall located at $x=R(t)$ as a function of the dimensionless time. The density of energy grows quadratically with time
in the one dimensional case of a one moving mirror. Parameters: $\epsilon=0.001$, $\Lambda=10$. In our units, energy is measeured in units of $1/L_0$.}
\label{energia1pared}
\end{center}
\end{figure}
In order to show that the behaviour at very long times is much enhanced by the finiteness of the number of modes, we show the comparison for different values of $\Lambda$ in Fig.\ref{comparacionmodos}. Therein, we can see in a solid blue line the simulation for a cavity containing 10 field modes. The red dashed line representes a 50-modes field cavity while in the dotted black line, the cavity contains 100 modes. The upper curves represent the $\vert {B}_1\vert^2$ coefficient, while the slower ones are for $\vert{B}_2\vert^2$. It is evident that at short times, all cavities yield the same results, but at very long times it is wiser to consider a bigger number of modes. However the qualitative behaviour remains the same, we obtain a lower exponential growth as the number of modes considered increases.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=8.cm]{comparacionmodos.eps}
\caption{(Color online) In this plot, we have used $\epsilon=0.01$, $\Lambda=10$ for the blue solid curves, $\Lambda=50$ for the red dashed curves and $\Lambda=100$ for the black dotted ones. In the upper values of the $\vert B\vert^2$ axis we plot the $\vert B_1\vert^2$ for the mode $n = 1$ of the field, while in the lower part of the axis is the $\vert B_2\vert^2$ of the field as functions of the dimensionless time. We need to use a big number of modes $\Lambda$ in order to get total convergence into the final values. This plots show the dependance with $\Lambda$ and tendency to exponential growing of created particles.}
\label{comparacionmodos}
\end{center}
\end{figure}
In this Section we have mainly proved numerically previous results and investigated some regions beyond the analytical studies. In the following Section we will analyse the more general case of two moving mirrors.
\section {Two moving mirrors}
\label{2wall}
In Ref.\cite{DalvityDiego} an unified analytic treatment of the dynamical Casimir effect in a one dimensional resonantly oscillating cavity for arbitrary amplitude and dephasing has been presented. Therein, it has been shown that for certain cases there is destructive interference and no radiation is generated. For others, there is constructive interference and motion-induced photons appears. When this takes
place, the way the energy and number of created photons inside the cavity grow in time depends on the relation among several parameters. For certain motions the growth of the energy density is exponential and for some others it is a power law. In this Section, we shall simulate the photon creation of two moving walls, for different situations. We will assume the case where $M \ll (n \pi)/(R-L)$ in order to re-obtain the results presented in Ref.\cite{DalvityDiego}, i.e. a one dimensional cavity. We will also consider the three-dimensional case by including a mass term in the frequency of the field in a non-perturbative treatment.
\subsubsection {Non-particle induced modes}
For a particular case of equal amplitudes $\epsilon_R=\epsilon_L$ and excitation frequency $\Omega_R=\Omega_L= \omega_n$, where $n$ indicates the mode field, the energy inside the cavity oscillates around the static Casimir value and there is no motion induced radiation. This was first reported in \cite{chinos}, for a perturbative treatment. It can be showed that for an even value of $n$ and $\phi_R=0$ or odd value of $n$ and $\phi_R=\pi$,
\begin{figure}[h!]
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{casophipiq5.eps}
\end{center}
\end{minipage}
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{casophi0q4.eps}
\end{center}
\end{minipage}
\caption{(Color online) (a) Left: $\vert{B}\vert^2$ coefficient as a function of the dimensionless time, for different field modes: $\vert{B}_3\vert^2$ (red dotted line), $\vert{B}_5\vert^2$ (blue dashed line) and $\vert{B}_7\vert^2$ (blue solid line) under $\Omega_R=\Omega_L= \omega_5$ and $\phi_R=\pi$.
(b) Right: $\vert {B}\vert^2$ coefficient for different field modes: $\vert{B}_3\vert^2$, $\vert{B}_4\vert^2$ and $\vert{B}_5\vert^2$ for $\Omega_R=\Omega_L= \omega_4$ and $\phi_R=0$.
Parameters used: $\phi_L = 0$, $\epsilon=0.01$, $\Lambda=10$. }
\label{noseexcitan}
\end{figure}
there is destructive interference among the two-moving mirrors \cite{DalvityDiego} (assuming $\phi_L=0$ in all cases). This particular behaviour is represented in
Fig.\ref{noseexcitan} (a) and (b), where we show destructive interference and no creation of particles.
\subsubsection {Dephased mirrors}
In this Section we follow the study of the cavity with two moving mirrors, analysing the case where the mirrors move in dephased trajectories. In particular, in Fig. \ref{comparacionphis} we show different cases in which particle creation, for odd modes, depends on the different values of the phase $\phi_R$ for a fixed value of $\phi_L = 0$. In this general case, the number of created particles grows exponentially with time for dephased motion and oscillates non-exponentially for the particular case of $\phi_L = 0$ and $\phi_R = \pi/2$ (solid pink line in Fig. \ref{comparacionphis}), when the excitation is $\Omega = \omega_5$. This case is in agreement with the result obtained in the previous Section since, it is equivalent to the total destructive interference example where there is no particle creation. We have also included cases of trajectories in phase (translational oscillation mode) as $\phi_L = 0$ and $\phi_R = 0$ (dotted grey line in Fig. \ref{comparacionphis}) and $\phi_L = \phi_R = \pi/4$ (solid red line in Fig. \ref{comparacionphis}) in which the behaviour is also exponential as we will see in the following Section. Furthermore, in the blue dashed line we show $\phi_L = 0$ and $\phi_R = \pi/4$ and also $\phi_L = 0$ and $\phi_R = 0.35 \pi$ in the dot-dashed black line. Both examples show an exponential growth in time of the number of particles created with excitation $\Omega = \omega_5$ for mode $n = 3$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=8.cm]{casoq5distintosphi.eps}
\caption{(Color online) We present the behaviour of an odd field mode $|B_3|^2$ as a function of the dimensionless time, for different values of $\phi_R$ under a perturbation $\Omega=\omega_5$. In the case of $\phi_L= \phi_R=0$ (dotted grey line), we see there is another slope (solid red line) which corresponds to a dephasing of $\phi_L=\phi_R = \pi/4$. Both cases are examples of a translational oscillation mode, a case which is considered separately in the following Section. In the blue dashed line we present $\phi_L = 0$ and $\phi_R = \pi/4$ and $\phi_L = 0$ and $\phi_R = 0.35 \pi$ in the dot-dashed black line. Both examples show exponential grow with time of the number of particles created with excitation $\Omega = w_5$ for mode $n= 3$. Parameters used:$\epsilon=0.01$, $\Lambda=10$.}
\label{comparacionphis}
\end{center}
\end{figure}
\subsubsection {Translational modes}
For the particular situation that $\epsilon_R=\epsilon_L = \epsilon$, $\Omega_R=\Omega_L$ and $\phi_L = \phi_R=0$, the cavity oscillates
as a whole. In this case, the mechanical length is kept constant and is pictorially called ``electromagnetic shaker" \cite{DalvityDiego}.
As in the previous examples, it is well known that, due to parametric resonance, a naive perturbative solution of Eq. (\ref{ecmov}), in powers of $\epsilon$, breaks down after a short amount of time. In order to find an analytical solution valid for longer times one can use the Multiple Scale Analysis (MSA) technique \cite{crocce,msa}. The MSA provides us with a simple technique equivalent to summing the most secular terms to all order in the perturbative treatment. In this way, it is possible to get a solution valid for a period of time longer than the perturbative case. We shall introduce a second time scale $\tau = \epsilon t$, and write Eq.(\ref{freemode}) as:
\begin{equation}Q_n(t,\tau) = \frac{A_n(\tau)}{\sqrt{2 \omega_n}} e^{-i \omega_n t} + \frac{B_n(\tau)}{\sqrt{2 \omega_n}}
e^{i \omega_n t}, \label{msamode}
\end{equation}
where the functions $A_n$ and $B_n$ are slowly varying and contain the cumulative resonant effects. To obtain differential equations for them, we insert this ansatz into Eq. (\ref{ecmov}), expand in powers of $\epsilon$ up to first order, and neglect second derivatives of $A_n$ and $B_n$.
The basic idea of MSA is to impose the condition that any term on the right-hand side of the previous equation with a time dependency of the form $e^{\pm i \omega_nt}$ must vanish. If not, these terms would be in resonance with the left-hand-side term and secularities would appear. We will follow this procedure in order to get some analytical predictions about particle creation in the translational mode. After imposing the requirement that no term $e^{\pm i \omega_nt}$ appear, we get:
\begin{eqnarray} \frac{dA_n}{d\tau} &=& \frac{A_0\Omega}{4L_0 \omega_n} \sum_m \omega_m {\tilde b}^{(1)}_{nm} \left\{\left[ \delta(\Omega + \omega_m - \omega_n) + \delta(\Omega- \omega_m + \omega_n) \right]A_m - \left[ \delta(\Omega - \omega_m - \omega_n) + \delta(\Omega + \omega_m + \omega_n) \right] B_m\right\} \nonumber \\
&-& \frac{A}{16L_0^2 \omega_n} \sum_m {\tilde g}^{(1)}_{nm} \left\{\left[ \delta(\Omega + \omega_m - \omega_n) - \delta(\Omega- \omega_m + \omega_n) \right]A_m + \left[ \delta(\Omega - \omega_m - \omega_n) - \delta(\Omega + \omega_m + \omega_n) \right] B_m \right\} \label{Adt}
\end{eqnarray}
\begin{eqnarray} \frac{dB_n}{d\tau} &=& - \frac{A_0\Omega}{4L_0 \omega_n} \sum_m \omega_m {\tilde b}^{(1)}_{nm} \left\{\left[ \delta(\Omega + \omega_m + \omega_n) + \delta(\Omega- \omega_m - \omega_n) \right]A_m - \left[ \delta(\Omega - \omega_m + \omega_n) + \delta(\Omega + \omega_m - \omega_n) \right] B_m\right\} \nonumber \\
&+& \frac{A}{16L_0^2 \omega_n} \sum_m {\tilde g}^{(1)}_{nm} \left\{\left[ \delta(\Omega + \omega_m + \omega_n) - \delta(\Omega- \omega_m - \omega_n) \right]A_m + \left[ \delta(\Omega - \omega_m + \omega_n) - \delta(\Omega + \omega_m - \omega_n) \right] B_m \right\}, \label{Bdt}
\end{eqnarray}
where $A_0$ is the amplitude of the mirrors displacement and the first order coefficients are given by:
\begin{eqnarray}{\tilde b}^{(1)}_{nm} &=& \frac{4 n m }{n^2 - m^2} \left[(-1)^{n+m} - 1\right] \nonumber \\
{\tilde g}^{(1)}_{nm} &=& - \frac{n m }{n^2 - m^2} 8 L_0 \Omega^2 \left[(-1)^{n+m} - 1\right].
\label{coeflineales}\end{eqnarray}
In these equations, there is not a time-dependent frequency term, from which one can get one resonant mode evolution when exciting the
system with $\Omega = 2\omega_n$. Therefore, in this case, the perturbation part of the mode
equation is given in terms of an infinite sum of resonant terms. It is easy to see that, in the small amplitude of the perturbative regime, we will obtain an exponentially in time amount of created particles when $\Omega = \vert \omega_n \pm \omega_m \vert$ (with ``n+m" an odd number from Eq.(\ref{coeflineales})). Therefore, only
those modes $n$ and $m$ will be parametrically excited \cite{crocce, nosotros}; while other modes will not be excited. In order to set a simple example, we can suppose that the external frequency is given by $\Omega = \omega_2 + \omega_3$. In this case, the linear equations are given by:
\begin{eqnarray} \ddot A_2 - \left(\frac{6A_0\Omega}{5L_0}\right)^2 \frac{(\omega_3 - \omega_2)^2}{\omega_2 \omega_3} A_2 &=& 0, \\
\ddot B_2 - \left(\frac{6A_0\Omega}{5L_0}\right)^2 \frac{(\omega_3 - \omega_2)^2}{\omega_2 \omega_3} B_2 &=& 0,
\end{eqnarray}
and the solution for the $B_2$ coefficient is: $B_2(t) \sim \exp( \Gamma \epsilon t)$ where the rate $\Gamma = (6A_0/5L_0) (\omega_3^2 - \omega_2^2)/\sqrt{\omega_2 \omega_3}$. This will produce that $N_2$ be exponential with time.
\begin{figure}[h!]
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{M10n3n2eps01.eps}
\end{center}
\end{minipage}
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{M50n3n2eps001.eps}
\end{center}
\end{minipage}
\caption{(Color online) (a) Left: solid lines for $|B_8|^2$ and $|B_5|^2$, the others lines correspond to $|B_3|^2$, $|B_2|^2$ and $|B_1|^2$. Parameters used: $\epsilon=0.01$, $\Lambda=10$, $M=10$. (b) Right: All modes are parametrically excited for a bigger value of the mass. Parameters used: $\epsilon=0.001$, $\Lambda=10$, $M=50$. Excitation frequency $\Omega = \omega_2 + \omega_3$.}
\label{traslacional2}
\end{figure}
The same conclusion could be obtained for the mode $n= 3$ and that would yield $N_3$ to be also an exponential function of time.
All other modes different to $n= 2$ and $n = 3$, are not parametrically excited in the perturbative regime under MSA.
However, for given values of $\epsilon$ and $M$, the perturbation in the Eqs. (\ref{Adt}) and (\ref{Bdt}) is large, and the perturbative and MSA approaches are not longer valid. Nonetheless, our numerical evidence goes beyond the perturbative regime and the MSA improvement itself and we can study the mode fields at longer times. In Fig. \ref{traslacional2}, we show an example of non-perturbative result in which we can find an exponentially growing number of created particles for mostly (or all) modes, after exciting the system with $\Omega = \omega_2 + \omega_3$ . This is in principle a non-expected result from the perturbative approach mentioned above.
In Fig.\ref{traslacional2} (a), we can see the behaviour of some field modes for the excitation $\Omega = \omega_2 + \omega_3$, where some modes oscillates while other are already excited for time $\Omega t \sim 1/\epsilon$. In Fig.\ref{traslacional2} (b), for a bigger value of mass, we can see that all modes are parametrically excited for the same time scale, even though $\epsilon$ is smaller.
\subsubsection {Breathing modes}
We next consider the case in which $\epsilon_R=\epsilon_L$, $\Omega_R=\Omega_L= \omega_n$ and $\phi_R=\pi$ ($\phi_L=0$), representing the situation in which the mirrors oscillate symmetrically with respect to the center of the cavity. The mechanical length changes periodically as an ``antishaker", in relation to the previous example in which the cavity moves as a whole. This is simulated in Fig.\ref{caso1phipiq4}. We show the blue dashed line for the mode $n=2$, the dot-dashed magenta line for $n=4$ and the solid line for $n=6$. The odd modes, such as $n=1$ and $n=3$, activate at longer times as shown in Fig.\ref{caso1phipiq4}(a).
In Fig.\ref{caso1phipiq4}, we are exciting the system by $\Omega_R=\Omega_L= \omega_4$ and $\phi_L=\pi$ with a negligible mass, yielding $\omega_4= 4 \pi/L_0$. Therein, we can observe that all even modes are getting excited while the odd modes become excited at longer times (they appeared at times $\Omega t \sim 500$). At the end, all modes become exponentially excited in a non-perturbative example.
\begin{figure}[h!]
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{casophipiq4.eps}
\end{center}
\end{minipage}
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{Energiaphipiq4.eps}
\end{center}
\end{minipage}
\caption{(Color online) (a) Left: For the breathing modes, we have at left: $\vert B\vert^2$ coefficient versus time for different field modes $n= 2$ ($\vert B_2\vert^2$), $n = 4$ ($\vert B_4\vert^2$), and $ n = 6$
($\vert B_6\vert^2$) under $\Omega_R=\Omega_L= \omega_4$ and $\phi_L=\pi$. Field modes $n = 1$ ($\vert B_1\vert^2$) and $n = 3$ ($\vert B_3\vert^2$) seem to get excited at longer times. (b) Right: Energy density as a function of the dimensionless time. Parameters used: $\epsilon=0.01$, $\Lambda=10$. In our units, energy is measeured in units of $1/L_0$.}
\label{caso1phipiq4}
\end{figure}
\subsubsection {Three dimensional cavities with two moving mirrors}
We can even study particle creation in the three dimensional cavity with two moving mirrors. Then, we start increasing the value of the mass, and see how the explosive cocktail of Fig. \ref{tresD} starts changing considerably. In Fig.\ref{tresD}, we show the behaviour of the coefficient for field mode $n=1$, $\vert{B}_1\vert^2$ under a perturbation defined by $\Omega_R=\Omega_L= \omega_5$ and $\phi_L=0$ for different values of the mass value: $M=0.01$ (dotted line), $M=1$ (dashed) and $M=5$ (solid). In the right side of the figure, we show the energy as a function of time for the same values of the parameters. Therein, we see that for $M=0.01$, the energy has an exponential behaviour while for $M=5$ it is not.
It is easy to see in Fig.\ref{tresD} the different results for the number of created particles between one and three dimension cavities. In fact, it is important to note in Fig. \ref{tresD} (a) that the smaller the mass value, the bigger the growth of the coefficient
$\vert{B}_1\vert^2$. Then, one dimensional cavities with two oscillating mirrors produce a bigger excitation of modes than the
corresponding cubic cavity in three dimensions. The same mass hierarchy can be seen in Fig. \ref{tresD} (b) for the energy density inside the cavities (one dimensional cavity when $M\rightarrow 0$, and three dimensional ones for bigger $M$).
\begin{figure}[!h]
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{B1phi0q5distintasmasas.eps}
\end{center}
\end{minipage}
\begin{minipage}{8.5cm}
\begin{center}
\includegraphics[width=8cm]{energiaphi0q5distintasmasas.eps}
\end{center}
\end{minipage}
\caption{(Color online) (a) Left: ${B}_1$ coefficient for field mode 1 under $\Omega_R=\Omega_L= \omega_5$ and $\phi_L=0$ for different values of the mass: $M=0.01$ (red dots), $M=1$ (blue dashes) and $M=5$ (black solid curve). (b) Right: Energy as a function of time. Parameters used: $\epsilon=0.01$, $\Lambda=10$. In our units, energy is measeured in units of $1/L_0$.}
\label{tresD}
\end{figure}
\section{Conclusions}
\label{conclusiones}
In this paper we have presented a detailed numerical analysis of the particle creation for a quantum field in a cavity with two perfectly conducting moving mirrors. This approach was applied to a one dimensional as well as to a three-dimensional cavity box. We have derived the equation of motion of the field modes and numerically evaluated the Bogolubov transformation between {\it in} and {\it out}-states. From Bogolubov coefficients, we were able to numerically calculate the number of created particles after the mirrors stop moving and return to the
unperturbed position.
In the case of a one moving wall at $x=R(t)$, while the wall at $x=0$ remains at rest, we have recovered the very well known results. In that case, we have showed that the rate of particle production depends strongly on whether the frequency spectrum is equidistant or not, obtaining the correct behaviour with respect to the number of created particles. A quadratic behaviour for short timescale and a linear growing for larger temporal scales when the spectrum is equidistant. We also found that, beyond perturbative predictions, the dependance to the number of created particles is exponential with time for very large times. In the case of three-dimensional cavities with one moving mirror, we have shown that our results are in
agreement with Literature for the perturvatibe or MSA regions, and have provided numerical evidence of the bahaviour of the number of created particles even in the non-perturvatibe case.
In the case of two moving walls, we have shown that the rate of particle creation depends strongly on the relation among the amplitudes, the frequency and the phase difference in the mirrors' oscillations. We have shown that in some cases there are constructive interference leading to an exponential growth of particles inside the cavity, while for other relations there exist destructive interference with no vacuum radiation.
We have gone further all analytical studies by considering a massive field and computing the energy density inside the cavity.
We also study non-perturbative regimes for translational modes obtaining that all modes into the cavity grows exponentially when the amplitude of the perturbation in the mode equation is large (compared with perturbative expansions analysis). Then, we report an exponentially growing number of created particles when exciting with external frequencies $\Omega = \vert \omega_n \pm \omega_m\vert$ (with $n+m$ an odd number).
We will present a further analysis of the translational mode oscillation with analytical and more numerical support elsewhere \cite{next}.
\section*{Acknowledgements}
This work was supported by ANPCyT, CONICET, and Universidad de Buenos Aires (UBA). We thanks F.D. Mazzitelli for useful comments.
|
\section{Introduction}
\label{Intro}
Dwarf galaxies are currently among the less evolved galactic environments:
they are places where active star formation is most likely to be found, and
where metallicity is low, and even very low.
Dwarfs are therefore systems of low stellar mass and low metallicity.
These characteristics have recently deepened the interest in
studying star-forming dwarfs
as ideal laboratories to possibly
anticipate some details of star formation at high redshift.
Metallicity (the mass fraction of heavy elements in the interstellar medium; ISM) is a key parameter when studying galaxy evolution, and it determines the dust mass fraction in the gas.
Dust forms from the available heavy elements from supernova core-collapse and the outflows from low-mass stars
\citep[see for example][]{TodiniFerrara:2001,Gomez:2012a,Gomez:2012b,Indebetouw:2014,Rowlands:2014,Matsuura:2015}.
The dust cycle in the ISM is a crucial piece in the process of star formation.
Star-forming regions are embedded in the densest pieces of the ISM, the so-called molecular clouds (MCs).
Dust grains therein are heated by energetic UV photons that are emitted by young stars \citep{Kennicutt:1998,Kennicutt:2009},
which are particularly susceptible to be absorbed and reemitted in the IR spectral range.
Dust absorption of UV photons in MCs retards molecule dissociation. In this way,
it is an essential piece in the process of ISM cooling, thus facilitating star formation.
Dust in the diffuse gas (cirrus) absorbs light from older stars. It is a matter of debate whether MCs
are also UV photon emitters, or if no UV radiation can escape from them
\citep[see for example][]{Boquien:2011,Bendo:2012}.
Dust emission from dwarf galaxies differs considerably from that of more metal-rich galaxies.
They harbor warmer dust, have lower dust-to-gas ratios, present lower polycyclic aromatic hydrocarbon (PAH) band emission, have broader IR spectral energy distribution (SED) peaks, and show a flattening of the far-IR/submm slope
(the so-called submm excess).
The all-sky {\it IRAS} infrared telescope,
which revolutionized our idea of the infrared Universe,
was
the first to detect warmer dust in galaxies
\citep[e.g.,][]{Helou:1986,Melisse:1994,Galliano:2003}.
It was followed up by the {\it Spitzer} mission 20 years later, which enabled a
deeper
study of the precise sources found previously by {\it IRAS}, and it ratified its warm dust findings
\citep[e.g.,][]{Galliano:2005,Rosenberg:2006,Cannon:2006,Galametz:2009}.
More recent investigations have reported broad and flat IR SEDs in dwarf galaxies, further showing that all the peculiarities described previously are more dramatic
in the most actively star-forming dwarfs
\citep[e.g.,][]{Boselli:2010, Boselli:2012, Smith:2012a, Ciesla:2014}.
However, no very low-metallicity ($\rm 12 + log(O/H)\lesssim 8.0$) galaxies were the subjects of these studies.
An important step forward has been made thanks to the {\it Herschel} guaranteed time key
program on dwarf galaxies, the so-called Dwarf Galaxy Survey \citep[DGS,][]{Madden13},
which includes the galaxies with the lowest metallicity in the local Universe.
Subsequent systematic studies \citep{RR13,RR14,RR15} essentially confirmed previous results.
In addition, the {\it Herschel} Astrophysical Terahertz Large Area Survey \citep[$H$-ATLAS, ][]{Eales10} provided an unbiased view of the dusty Universe.
In particular, a dust-selected sample of local galaxies was made available for the first time, analyzed at high resolution with the sensitivity of {\it Herschel}.
This is the {\it Herschel}-ATLAS Phase-1 Limited-Extent Spatial Survey,
\citep[HAPLESS][]{Clark:2015},which spans a range of $\rm 7.4 < M_* < 11.3 \log_{10} M_{\odot}$ in stellar masses,
therefore containing dwarf galaxies. The analysis of this survey revealed the diverse IR properties of local dust-selected galaxies,
including again SED IR peak broadening and slope flattening.
In these pioneering analyses of dwarf galaxies it is generally assumed
that these characteristics may be evidence of either an additional warm dust component or a lower-than-Galactic dust emissivity index taking $\beta<2.0$ values, or both.
Specifically, flux densities are modeled as modified black bodies following
\begin{equation}
F_{\nu} \propto(\frac{\lambda}{\lambda_0})^{-\beta_{\rm}} B_{\nu}(\lambda,T),
\label{ModBB}
\end{equation}
where $\beta$ is the effective slope in the FIR-submm region,
$\lambda_0 \approx 100 \mu$m is the reference wavelength,
and
$B_{\nu}(\lambda,T)$ is the Planck function.
In order to fit the data, observers therefore change their modeling in the FIR-submm range relative to the standard fixed $\beta=2$,
which is appropriate for normal galaxies
\citep[e.g.,][]{Boselli2012, RR13}
and/or add new ad hoc additional dust components to it \citep[double black bodies at different temperatures, see e.g.,][]{Galametz:2012, RR14,RR15,Clark:2015}.
In particular, this latter method is needed
to fit the IR peak broadening that is observed at wavelengths lower than $\lambda \sim 70\,\mu$m. This excess of mid-IR (MIR) emission cannot be accounted for through a unique modified black body fitting.
For example, HAPLESS analyses find that the flux density modeling through Eq. \ref{ModBB}
with fixed $\beta=2$ underestimates fluxes at 100 and 500 $\mu$m,
while overestimating them at 160 $\mu$m. Therefore these analyses use a two-temperature component with fixed $\beta=2$.
On the other hand, \citet{RR15} use a second warmer dust component to cure
their FIR-submm dust modeling of a lack of photons in the MIR range of some DGS galaxies, when necessary.
These empirical modelings have provided very valuable estimates of
dust masses, temperatures, and emissivity indices of their analyzed galaxies.
A step further to solve the problem of dwarf galaxy SED peculiarities can be gained
by analyzing dwarf emission in a controlled
scheme, that is to say, when some information
is available about
the nature and intrisic SEDs of the dust heating sources and on the dust components'
(cirrus + molecular clouds) properties
and space configurations.
This is the case of hydrodynamical simulations.
Two different approaches, both requiring completion with a radiative transfer code to calculate dust effects on the SEDs, can be followed in this regard.
A first procedure is through simulations that include dust formation and evolution in a self-consistent
manner
\citep[improving one and two-zone models, e.g.,][in that they provide spatial structure]{Dwek98, Lisenfeld98, HirashitaFerrara2002, Calura:2008, Asano:2013}.
These models
\citep[e.g.,][]{Bekki:2013, Bekki:2015a, Bekki:2015b, McKinnon2016, Aoyama2017}
include dust formation from stellar evolution and supernova (SN) explosions,
dust evolution processes such as grain growth and destruction in the ISM, and dust effects
on galaxy evolution, such as radiation fields effects, star formation, and
dust-enhanced H$_2$ formation.
The effects of dust in galaxy evolution
could be important, and studying them self-consistently through hydrodynamical
simulations is a very good approach. This emerging method is still in its infancy, and
considerable work remains to be done.
Ultimately it will provide a refined estimate of dust properties (e.g., composition,
size distribution, and dust-to-gas mass ratio) in terms of evolutionary paths,
and also in terms of redshift. At the moment, this method gives different possible solutions that
need to be tuned against observational data.
A second possibility is to study dust effects on the SEDs of simulated galaxies
from the outputs of standard hydrodynamical simulations that do not include dust formation and evolution,
but nevertheless produce realistic galaxies
at given redshifts.
In order to study the SEDs of local (i.e., at redshift $z$=0) dwarf galaxies,
their dust properties at $z$=0 are needed.
These can be well approximated
by methods other than using simulations where dust formation and evolution is self-consistently implemented, for example,
following the extensively used models
by \citet{Weingartnera:2001, Draine:2003, Draine:2007} or \citet{Zubko04}, among others,
which have been thoroughly examined to reproduce observational data.
A few codes exist that interface with the outputs of
hydrodynamical simulations to solve the radiative transfer through
dust and
predict a multiwavelength SED for simulated
galaxies.
For example,
SKIRT and its updated version \citep{Baes:2002,Baes:2005,Baes:2011,Steinacker:2013},
SUNRISE
\citep{Jonsson:2004, Jonsson:2006, Jonsson:2009},
RADISHE
\citep{Chakrabarti:2008, Chakrabarti:2009},
and Art2
\citep{Li:2007, Li:2008, Yajima:2012}
all use Monte Carlo
techniques to follow the radiation of photons through the
ISM and calculate a global radiation field and
hence the dust reemission.
In addition, SUNRISE
includes the treatment of star-forming regions using the dust
and photoionization code MAPPINGSIII \citep{Groves:2008}.
In a somewhat different context, the dust radiative-transfer code
GRASIL \citep[][hereafter S98 and S99, respectively]{Silva:1998,Silva:1999} has been used highly
successfully for many years,
in particular in combination with semianalytical models of galaxy formation such as GALFORM
\citep{Cole:2000, Granato:2000, Baugh:2005,
Lacey:2008}, MORGANA \citep{Monaco:2007,
Fontanot:2008, Fontanot:2009}, and ABC \citep{Granato:2004, Silva:2005, Lapi:2006, Cook:2009}.
The 3D version of GRASIL, GRASIL-3D \citep[see][hereafter DT14]{Dominguez:2014}, can
be interfaced with the outputs of hydrodynamical codes of galaxy formation,
producing observation-like multiwavelength SEDs of simulated galaxies while
inheriting GRASIL code strengths.
Following GRASIL, some GRASIL-3D particular strengths relative
to the codes quoted above can be summarized as follows.
i) The radiative transfer is not solved
through Monte Carlo methods, but in a grid.
ii)
The code separately treats the radiative transfer in molecular
clouds (MCs) and in the diffuse cirrus component, whose dust compositions are
different.
We note that MCs are not resolved in simulations, and therefore a subresolution modeling is introduced.
iii) It takes into account the fact that younger stars are associated
with denser ISM environments by means of an age-dependent dust reprocessing of stellar
populations (note that GRASIL has been the first model to do
so), mimicking through the $t_0$
parameter the time young stars are enshrouded within MCs before their destruction.
iv) It includes a detailed
non-equilibrium calculation for dust grains with diameter
smaller than $a_{flu} \sim 250$ \AA,
allowing for
a proper treatment of PAH features, which dominate the MIR in
some cases.
The aim of this work is to provide a physically based explanation to the particular
emission patterns dwarf galaxies show in the IR-submm range\footnote{We do not intend to provide fittings to observed individual dwarf galaxy IR SEDs, since
the precise baryon 3D distributions and star formation histories of simulated galaxies do not have free parameters.
}.
As described above, these can be summarized as \citep{RR13}
(1) a broadening of the IR peak of the SED, which implies a warmer dust component;
(2) an excess of emission in the submm ($\sim$500 $\mu$m) that causes a flattening of the submm/FIR slope;
and (3) a very low intensity of PAH emission features.
This is done by studying the IR SEDs of a sample of 27 star-forming dwarf galaxies that are identified in a hydrodynamical simulation of the local Universe.
This simulation does not include dust evolution self-consistently, but as we show, it produces realistic dwarf galaxies.
The SEDs have been obtained with GRASIL-3D.
The paper is organized as follows:
Section \ref{method} describes the simulation and the important characteristics of the GRASIL-3D radiative transfer code.
Sections \ref{sample} and \ref{sec:obs} present the sample of simulated star-forming dwarf galaxies and the observational galaxies with which the results are compared.
Sections \ref{sec:prop} and \ref{sec:emi} show the general properties and the IR and submm emission of the sample of simulated galaxies.
In Sect. \ref{sec:disc} we discuss the results with regard to the physics that drives the existence of an additional warm dust component and the effects of varying
the free parameters of
GRASIL-3D.
Finally, our study is summarized in Sect. \ref{conclu}.
\section{Method} \label{method}
\subsection{Simulation}
\subsubsection{Initial conditions and run: the CLUES project}
\label{CLUES}
The sample of simulated dwarf galaxies is taken from a single simulation with initial conditions from the Constrained Local UniversE Simulations (CLUES) project\footnote{\tt http://clues-project.org/}
\citep{gottloeber10, yepes14},
where peculiar velocities obtained from catalogs are imposed as constraints on the initial conditions in order to simulate a cosmological volume that is representative of our local Universe. In these simulations structures like the Virgo cluster are always reproduced and the Local Group forms in the correct environment.
On smaller scales, the distribution of structure is essentially random.
Several dark matter-only realizations are run until a Local Group analog (a Milky Way-M31 binary group, hereafter LG) is found. Then the zoom-in technique is applied, which means that the LG region is resimulated with baryons and at a higher resolution. The resimulation includes 4096$^3$ effective particles in a sphere with radius $\sim 2h^{-1}$Mpc around the LG. The mass resolution of particles is m$_{\rm star}$=1.3$\times$10$^4$M$_{\odot}$, m$_{\rm gas}$=1.8$\times$10$^4$M$_{\odot}$, and m$_{\rm dm}$=2.9$\times$10$^5$M$_{\odot}$, and the gravitational softening lengths are $\epsilon_{\rm bar}$=223pc between baryons and $\epsilon_{\rm dm}$=486pc between dark matter particles.
The cosmology used is $\Lambda$CDM with WMAP3 parameters, that is, $H_0=73$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\rm m}=0.24$, $\Omega_{\Lambda}=0.76$, $\Omega_{\rm baryon}=0.04$, and $\sigma_8=0.76$.
\subsubsection{GASOLINE}\label{gasoline}
This CLUES simulation has been evolved using the parallel N-body+SPH tree-code \verb,GASOLINE, \citep{wadsley04}, which includes gas hydrodynamics and cooling, star formation, energy feedback, and metal enrichment to model structure formation.
We briefly describe here the most important implementations of this well-tested code (for details see \citealt{stinson06} and \citealt{governato10}).
The simulation
follows the physics used in \citet{governato10} and \citet{guedes11}.
Feedback recipes are tuned to match the stellar-to-halo mass relation at one stellar mass.
When gas becomes cold and dense, stars are formed according to a Schmidt law with a star formation rate $\propto\rho^{1.5}$. When these stars die, they release energy and metals into the surrounding ISM.
The total mass and number of stars that explode as SNe
is calculated using stellar lifetime calculations from \citet{Raiteri96}.
It is assumed that
only stars between 8 and 40 M$_\odot$ explode as SNII,
and the mass of a binary system that can eventually explode as SNIa is between 3-16 M$_\odot$.
Stars more
massive than 40 M$_\odot$ are assumed to either collapse into black
holes or explode as SNIb. Regardless of this, only few stars
form with masses greater than 40 M$_\odot$, thus the impact on the
feedback is minimal.
The energy feedback by SNII is implemented by means of the blastwave formalism \citep{stinson06}, where
a fixed fraction of the canonical $10^{51}$ erg per SN explosion is released to the ISM.
The metals produced in these stars are released as the main-sequence progenitors die,
and the metals are distributed to the gas within the
blast radius. Iron and oxygen
are produced in SNII according to the analytic fits used in \citet{Raiteri96}
using the yields from \citet{Woosley95}:
M$_{\rm Fe}$ =2.802 $\times$ 10$^{-4}$M$^{1.864}_*$ and
M$_{\rm O}$ =4.586 $\times$ 10$^{-4}$M$^{2.721}_*$.
Feedback from SNIa also follows the \citet{Raiteri96} model. Each SNIa produces 0.63 M$_\odot$ iron and
0.13 M$_\odot$ oxygen \citep{Thielemann86} and the
metals are ejected into the nearest gas particle.
Furthermore, the significant feedback contribution of stellar winds is also taken into account.
Stars with masses of between 1 and 8
M$_\odot$ return a fraction of their initial mass,
which is determined
using the function derived by \citet{Weidemann87}. The returned gas
has the same metallicity as the star particle. Metal diffusion between gas particles is implemented as described in \citet{shen10}.
Finally, \verb,GASOLINE, also accounts for the effect of a uniform background radiation field on the ionization and excitation state of the gas.
The formation and evolution of dwarfs in the LG neighborhood has been analyzed in zoomed simulations that are based on the same initial conditions as used in our paper, but performed with the GADGET code \citep{BenitezLlambay13, BenitezLlambay15, BenitezLlambay16}.
This particular CLUES simulation has been studied in \citet{SantosSantos16}, where it was shown that its intermediate- to high-mass disk-like galaxies have a correct mass distribution and angular momentum content.
\subsection{GRASIL-3D}
\label{GRASIL3D}
To develop the GRASIL-3D radiative-transfer code, DT14 have followed the main
characteristics and scheme of GRASIL. We briefly recall the main features
here. For a full description we refer to DT14, S98, and S99.
To solve the radiative transfer,
GRASIL-3D uses a Cartesian grid whose cell size is set by the smoothing length
used in the simulation code.
The gas is subdivided into a dense phase (fraction $f_{mc}$
of the total mass of gas) associated with young stars
(star-forming molecular clouds, MCs) and a diffuse phase
(cirrus) where more evolved (free) stars and MCs
are placed.
Young stars leave their parent clouds on a timescale $t_0$.
The MCs are represented as spherical clouds with
a certain
optical depth
containing an inner stellar source of radiation, whose
radiative transfer through the MC is calculated following \citet{Granato:1994}. The radiative
transfer of the radiation emerging from MCs and from free
stars is then computed through the cirrus dust.
\subsubsection{Gas fractions in MCs and cirrus}
The fraction of gas in MCs, $f_{mc}$, is
calculated locally by implementing a
subresolution model based on the assumption that MCs are defined by a density threshold $\rho_{mc, thres}$
\citep[e.g.,][]{Elmegreen:2002,Hennebelle:2012}, and using a theoretical log-normal probability distribution function (PDF) for the gas
densities.
The PDF is characterized by a density parameter $\rho_0$ and a dispersion $\sigma$.
They define the PDF volume-averaged density as
$ <\rho>_V \equiv\rho_0 e^{\sigma^2/2}$,
which we in turn identify with the SPH density of the \textit{i-th} gas particle (simulation output),
$ <\rho>_V = \rho_{\rm gas}(\vec{r_i}) $.
In this way,
$\rho_0$ can be calculated for each gas particle,
leaving only two parameters to control
the calculation of $f_{mc}$: the gas density threshold $\rho_{mc, thres}$, and the PDF dispersion $\sigma$.
The diffuse gas fraction of a given particle is then $(1 - f_{mc})$.
This PDF approach is
suggested by small-scale ($\sim$ 1 kpc) simulations
\citep[see, for example,][]{WadaNorman:2007,Federrath:2008,Brunt:2009,Hennebelle:2012}
and is supported by recent high-resolution {\it Herschel} observations of nearby MCs \citep[e.g.,][]{Schneider12}.
Values for $\rho_{mc, thres}$ vary in the literature.
For example,
$\rho_{mc, thres}$ = 100 H nuclei cm$^{-3} \approx$ 3.3 M$_{\odot}$ pc$^{-3}$ in \citet{Tasker:2009},
and $= 35$ H nuclei cm$^{-3} \approx 1$ M$_{\odot}$ pc$^{-3}$
in \citet{BallesterosP:1999,BallesterosP_Scalo:1999}.
Values for $\sigma$ have been calculated to
range from 2.36 to 3.012 in \citet{Wada:2007},
while \citet{Tasker:2009} give $\sigma$ = 2.0.
In DT14 the range for the density threshold
is taken to be $\rho_{mc,thres}$ = 10 - 100 H nuclei cm$^{-3} \approx 3.3 \times 10^8 - 3.3 \times 10^9 $M$_{\odot}$ kpc$^{-3}$, while the PDF dispersion takes either 2.0 or 3.0.
These are the values used here.
A useful check for the choice of these parameters can
be made by comparing
the final calculated average HI and H$_2$ contents of the galaxy or sample of galaxies, with observations.
\subsubsection{Stellar emission}
The luminosity of the young stellar
populations placed inside each active MC
is modeled in GRASIL
by linearly decreasing the fraction $f$ of the stellar populations' energy radiated inside the cloud with its age $t$:
\begin{equation}
f(t) = \left\{ \begin{array}{ll}
1 & \textrm{ $t \le t_0$} \\
2 -t/t_0 & \textrm{ $t_0 < t \le 2 t_0$ } \\
0 & \textrm{$t > 2 t_0$}
\end{array} \right.
\label{fracyoung}
\end{equation}
Here $t_0$ is a free parameter representing the time taken for stars to escape the MCs where they were born,
and it mimics MC destruction by young stars.
The same approach is adopted in GRASIL-3D.
In this way,
only stars with ages<2$\times t_0$ are eligible to form part of the MC heating engine, with probability $f(t)$.
Observationally, $t_0$ takes values on the order of the lifetime of the most massive stars, from 3 to 100 Myr, and it varies with the density of the surrounding ISM, being higher in the densest environments.
Typical values were found by
comparison to local observations in S98, yielding $\sim$
2.5 to 8 Myr for normal spiral galaxies, and between 18
and 50 Myr for starburst galaxies.
The contributions to the
heating of cirrus dust come from free stellar particles with ages
$t > 2 t_0$, as well as from stars with $t > t_0$ supplying $1- f(t)$ of
their radiation.
\subsubsection{Dust-to-gas mass ratio}
\label{DGratio}
The dust content in MCs and in the cirrus is computed from a dust-to-gas mass ratio $\delta$.
In this work, $\delta$ has been set to vary with
gas metallicity Z$_{gas}$
following the broken power-law in terms of oxygen abundances proposed in Table 1 of \citet{RR14}.
This approach
is needed
to fit the
very low
dust masses observed in low-metallicity (12+log(O/H)$\lesssim$8.0) environments or galaxies
\citep{Lisenfeld98, Galliano:2005, Hunt2005, Galliano:2008, Engelbracht2008, Galametz2011, Galliano:2011, Izotov:2014},
and it is consistent with dust and PAH evolution models \citep[e.g.,][]{Hirashita:2002, Calura:2008, GallianoDwek:2008, Asano:2013, Seok:2014}.
Since in GRASIL-3D each particle has
a different gas mass and metallicity, the dust-to-gas mass ratio is a local quantity that varies from cell to cell.
Adopting 12+log(O/H)$_\odot$ = 8.69 \citep[i.e. Z$_\odot$ =0.014,][]{Asplund09} solar values,
we obtain
\begin{eqnarray}
\delta(Z_{gas, k}) = \frac{D}{G} = \frac{1}{10^{a_1}}\left( \frac{Z_{gas, k}}{Z_\odot}\right)^{\alpha_1} \quad \rm for\ Z>Z_t \\
\delta(Z_{gas, k}) = \frac{D}{G} = \frac{1}{10^{a_2}}\left( \frac{Z_{gas, k}}{Z_\odot}\right)^{\alpha_2} \quad \rm for\ Z\leq Z_t \nonumber
\label{dustfrac}
\end{eqnarray}
where $\alpha_1$=1.0, $a_1$=2.21, $\alpha_2$=3.0, $a_2$=0.77,
and $Z_{gas, k}$ is the mass fraction of gas metals at the $k$-th grid cell.
The break in the metallicity dependence occurs at Z$\rm_t$=0.0026 (i.e., 12+log(O/H)=7.96),
and the result is given in solar units, where $\delta_\odot$=D/G$_\odot$=1/162 according to \citet{Zubko04} for a dust composition like ours (see next subsection).
We adopt a slope of 1 at high metallicities ($\alpha_1=1$), which has been shown in many studies to adequately represent this regime
\citep{James2002, Draine:2007, Galliano:2008, Leroy:2011},
and a slope of $\alpha_2=3.0$ at low metallicities
(we do not use exactly the same value as that given in \citet{RR14} but a simpler value that is inside their errors).
This broken power-law scaling of $\delta$ with metallicity gives satisfactory final dust masses compared to observational estimates, as we show in Sect. \ref{dgratio},
while using a linear scaling ($\alpha_1 = \alpha_2 = 1$) at any metallicity does not work so well.
However, we note that
the problem of the D/G vs Z assignation at low metallicity is still not solved because of
the uncertainties in determining both the dust and H$_2$ masses of observed very low-metallicity galaxies.
The dust-to-gas mass ratio is also relevant for calculating the radiation transfer through MCs since it determines the optical depth of the cloud:
\begin{equation}
\tau_{mc} \propto \delta(Z_{gas, k}) \frac{m_{mc}}{r_{mc}^2},
\label{taumc}
\end{equation}
where $m_{mc}$ and $r_{mc}$ are the mass and radius of individual MCs, respectively, both free parameters in GRASIL-3D.
The proportionality constant depends on the dust model.
Since both $\tau$ and the luminosity from the central source can vary in each grid cell, the radiative transfer is calculated separately for each of them even if $t_0$ is taken to be constant.
\subsubsection{Dust model}\label{dustmodel}
The dust in GRASIL-3D is assumed to consist of a mixture of
different grain types:
carbonaceous and silicate spherical grains,
where graphite grains smaller than $a_{flu}=$250$\AA$ are considered as PAHs.
These in turn consist of a mixture of neutral and ionized particles.
This simple composition has been shown to appropriately reproduce observational constraints coming from the average interstellar extinction, thermal IR emission, and interstellar abundances in the Galaxy, LMC and SMC \citep{Weingartnera:2001,Zubko04}.
A different size distribution is adopted for each grain type, following the functional form given in \citet{Weingartnera:2001}, and updated in \citet{Draine:2007}.
The optical properties come from \citet{Laor:1993} and \citet{Li:2001}.
A different dust composition is assumed in the diffuse ISM and in MCs.
The difference resides in the
PAH abundance or index $q_{\rm PAH}$, defined as
the percentage of dust mass contributed by PAHs containing less than $10^3$ C atoms.
Previous studies have discretized this index into seven values
\citep[see][]{Weingartnera:2001, Draineb:2007, Draine:2007},
each characterizing a specific dust model, where the highest value (4.58\%) matches the Galactic dust.
Here, we use a low value of $q_{\rm PAH}$ = 1.12\% for the cirrus dust \citep[as in model MW3.1\_10 from][]{Draine:2007, Draineb:2007},
as suggested by observations of low-metallicity galaxies that show very low PAH band emission
\citep{Engelbracht2005, Ohalloran2006, Draineb:2007, Smith2007b, Gordon2008, Rosenberg:2008, Hunt2010, WuR2011, Ciesla:2014}
and by PAH evolution models \citep{Seok:2014, Bekki:2013}.
This has been implemented globally for all grid cells, and we consider it a suitable approximation since
the use of one value for $q_{\rm PAH}$ or another only affects the
intensity of the particular mid-infrared PAH emission features, and not
the far-infrared luminosity.
In MCs however, the PAH abundance is decreased 1000 times relative to the diffuse component, as tuned by \citet{Vega:2005} to fit MIR properties of local actively star-forming galaxies.
The differences between using this low $q_{\rm PAH}$ cirrus value or the standard Galactic value are discussed in Sect. \ref{ParVar}.
The temperature distribution is defined for each grain type (silicate, graphite, or PAH) by its particular size distribution. In the case of grains with sizes $a> a_{flu}$, each interval in grain size corresponds to emission from a
modified black body of a given temperature $T$, with an emissivity index of $\beta$=2 defining its FIR/submm slope.
For PAH grains, in contrast, no unique temperature is defined.
PAHs of a given size are not able to reach temperature equilibrium due to stochastic heating, and span instead
a wide distribution of temperatures \citep{Guhathakurta:1989, Siebenmorgen:92}.
The final emission of the MC and cirrus dust components in a given grid cell is the sum of all the temperature contributions coming from each grain size interval corresponding to each grain type.
Throughout the paper we use the respective maxima location of the
MC and cirrus dust components
as an estimate of their effective global temperatures.
However, it is worth noting that
the typically used `cold' or `warm' terms for dust emission
do not strictly apply for GRASIL-3D SEDs
(nor for real galaxies)
as they represent
a very simplified description of a more complex situation.
\subsubsection{Choice of GRASIL-3D free parameter values}
According to the previous subsections,
the free parameters in GRASIL-3D are
i) the escape timescale from MCs, $t_0$; ii) the threshold density for MCs, $\rho_{mc, thres}$;
iii) the log-normal PDF dispersion parameter, $\sigma$; and iv) the ${m_{mc}}/{r_{mc}^2}$
combination, determining the optical depth of individual MCs.
It is worthwhile to recall that the simulation provides the geometry of each galaxy, the star formation history (SFH), metal enrichment history, and stellar and gas fractions, in such a way that no further parameters are needed to describe them.
The following set of parameter values have been used and
considered as of reference in this study:
$t_0$=40 Myr,
$\rho_{mc,thres}=3.3 \times 10^{9}$ M$_\odot$kpc$^{-3}$, $\sigma$=2., $m_{mc}$=10$^6$ M$_\odot$, and $r_{mc}$=14 pc.
The choice of a high $t_0$ value comes from following the results of S98 for star-bursting galaxies,
which seems a good approximation since as we show below, our simulated dwarfs are particularly picked according to their recent star formation.
The PDF parameters chosen provide the best match to observational data of H$_2$ molecular gas masses of nearby low-mass galaxies when using a metallicity-dependent CO-to-H$_2$ conversion factor (see Sects. \ref{SFRH-H2} and \ref{HIH2mass}). Finally, the mass and radii of single molecular clouds are set to the values that best match observed SEDs of local galaxies in S98. These numbers are within the observational estimates from MCs in our Galaxy.
In Sect. \ref{ParVar} we discuss the effects of parameter variation on the resulting SEDs in detail.
GRASIL-3D has a general applicability to the outputs of
simulated galaxies, and it has been used in a variety of contexts
to analyze their SEDs. For example,
DT14 have analyzed normal spiral galaxies and compared the resulting SEDs to different observational surveys.
\citet{Obreja:2014} have tested star formation rate tracers used by observers against mass-weighted
quantities, obtaining a satisfactory agreement.
\citet{Granato:2015} have analyzed the early phases of galaxy cluster formation in hydrodynamical simulations,
comparing their IR emission to observations.
\citet{Buck2016} have studied the nature of clumps in simulated massive disk galaxies.
\citet{Goz2016} have evaluated the possible correlation of IR luminosity in \textit{Spitzer} and \textit{Herschel} bands with several galaxy properties.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics[scale=0.6]{figure1.pdf}}
\caption{Stellar mass distributions of the CLUES total (black outline), dwarf (blue pattern), and star-forming dwarf (red) galaxy samples. }
\label{FigMass}%
\end{figure}
\begin{figure}
{\includegraphics[width=\linewidth]{figure2.pdf}}
\caption{Distance to the CLUES Local Group barycenter versus the velocity relative to this barycenter. The color-coding is the same as in Fig. \ref{FigMass}. The vertical dashed line indicates the observational zero-velocity surface of the
Local Group (R$_{\rm LG}$ = 1060 $\pm$ 70 kpc) as derived in \citet{McConnachie12}.}
\label{FigDvsV}%
\end{figure}
\section{Sample of simulated star-forming dwarf galaxies} \label{sample}
The galaxies from the CLUES simulation were identified with the Amiga Halo Finder (\texttt{AHF}; \citealt{knollmann09}), where halo masses are defined as the mass inside a sphere containing $\Delta_{\rm vir}\simeq$350 times the cosmic background matter density.
This search has produced 139 highly resolved dark matter halos, 120 of which host
bound stellar mass objects ranging 10$^6-$10$^9$M$_\odot$, to which we asign the label "dwarf". These may or may not have gas particles, and can be either isolated or satellites.
In this work we are interested in the IR-submm emission of dwarf galaxies.
As is well known,
dust grains absorb the ultraviolet light emitted by very young stars and reemit it at longer wavelengths.
We therefore focus on those dwarfs that have recent star formation; in particular, dwarfs that contain stellar particles (stellar populations) with ages lower than 80 Myr
(this is twice the $t_0$ value we chose).
This leaves a final sample of 27 dwarf galaxies, all of which contain gas. Their general properties are given in Table \ref{Table1}.
The mass distributions of the CLUES total, dwarf, and star-forming dwarf galaxy samples are given in Fig. \ref{FigMass}.
We note that
the fraction of galaxies in the LG with recent star formation is consistent with observations \citep{Mateo:1998,Tolstoy:2009, Dolphin:2005}.
In Fig. \ref{FigDvsV} we plot the velocity of each galaxy relative to the center of mass of the simulated LG, that is, the barycenter of the Milky Way-M31 pair, versus their distance to this barycenter. A vertical dashed line indicates the zero-velocity surface of the LG (R$_{\rm LG}$ = 1060 $\pm$ 70 kpc) as derived in \citet{McConnachie12}.
Galaxies are labeled with their ID number (column 1, Table \ref{Table1}) to allow following them in all figures throughout the paper.
When compared to real data \citep[see Fig. 5]{McConnachie12},
this plot gives us a first
satisfactory
insight into the simulated LG structure, and in particular, into the galaxy
distribution. Considering the local volume to have a radial extent of $\sim$3 Mpc, we find galaxies that are true LG members, galaxies that are close but not yet bounded, and galaxies that are outer nearby neighbors, rising in velocity with distance as expected due to the Hubble flow.
The total number of simulated LG members is comparable to the observed number.
The galaxies of this sample present differences in their SFHs, gas masses, sizes, and morphologies, the latter being either irregular, spheroidal, or disky.
Furthermore, the majority of the local star-forming dwarfs we studied are isolated,
except for
three (numbers 86, 104, and 150) that are satellites of other more massive galaxies.
Some of the general properties of the sample, listed in Table \ref{Table1}, are studied in Sect. \ref{sec:prop}.
\section{Observational samples to compare our results with}\label{sec:obs}
We report on two types of comparisons with observational data. First, we examine whether the simulated dwarfs have global properties that agree with observational samples, for which we focus on the SFHs, HI and H$_2$ gas contents, dust masses, and metallicities because these properties have the strongest influence on the resulting SEDs. Second, we compare the IR-submm emission of our simulated sample with observed galaxies that span the same stellar mass and metallicity ranges. This second comparison is the core of this paper, and it is made possible by the recent \textit{Herschel} observations mentioned in Sect. \ref{Intro}.
\subsection{ HI and H$_2$ gas content, star-formation rate, and gas metallicity}
\label{SFRH-H2}
We compare the HI and H$_2$ masses of the simulated dwarf galaxies to data from the observational studies described below.
For all of these galaxy samples
the H$_2$ content or molecular gas masses are calculated by assuming a constant CO-to-H$_2$
conversion factor with
the Galactic value: X$_{\rm CO;MW} =2.0 \times 10^{20}$ cm$^{-2}$ (K km s$^{-1}$)$^{-1}$.
However,
we note that it has long been argued
that CO may not trace the molecular gas in dwarf galaxies: the conversion factor may be metallicity dependent
\citep[see e.g.,][]{Taylor98, Wilson:1995,Boselli:2002, Hunt2005, Leroy:2011,Schruba:2012, Bolatto13},
and the use of the standard Galactic value could understimate the molecular gas fraction in low-metallicity galaxies.
Some of the following samples also provide the H$_2$ masses obtained using a conversion factor of this type.
We therefore use these samples
in Sect. \ref{HIH2mass}
to study the
different GRASIL-3D parametrizations needed to match the molecular gas masses obtained
using a constant and a metallicity-dependent conversion factor.\\
\paragraph{The Dwarf Galaxy Survey \citep[DGS,][]{Madden13}.}
DGS is a sample of star-forming dwarf galaxies with metallicities ranging from $Z \sim 0.03$
to $0.55\,Z_{\odot}$ and stellar masses covering $ 3 \times 10^6$ to $3 \times 10^{10}$ M$_{\odot}$
\citep{Madden14}. HI and H$_2$ masses are given in \citet{RR14}, where they compiled data from \citet{Madden13}.
HI masses were corrected to account for the aperture size used to obtain the dust SED.
\citet{RR14} also provide the H$_2$ masses of the DGS galaxies obtained using a metallicity-dependent conversion factor of the form $ X_{\rm CO,Z}\propto Z^{-2}$, following \citet{Schruba:2012}.
Dust masses for DGS are given in \citet{RR15}, where they represent the best fit to data following the dust SED model from \citet{Galliano:2011}.
\paragraph{The Key Insights on Nearby Galaxies: a Far-Infrared Survey with \textit{Herschel} \citep[KINGFISH,][]{Kennicutt11}.}
This sample consists of 61 galaxies, most of them spirals with some early-type and dwarf galaxies.
Their stellar mass range is $\rm 1.7 \times 10^7-4.8 \times 10^{11} M_{\odot}$
peaking at $\rm M_* \sim 5.0 \times 10^{10} M_{\odot}$, and metallicities reach up to Z $\sim$ 1.20 Z$_{\odot}$.
No particular dust emission features have been found in this sample \citep{RR15}. We therefore use it as a reference with which to compare the specifities of dust emission in lower mass galaxies. HI and H$_2$ masses of KINGFISH galaxies are also given in \citet{RR14}, where they compiled data from \citet{Draineb:2007}.
\citet{RR14} additionally provide the H$_2$ masses obtained using a metallicity-dependent conversion factor of the form $ X_{\rm CO,Z}\propto Z^{-2}$.
Dust masses for KINGFISH are estimated in \citet{RR15}.
\paragraph{The \textit{Herschel}-ATLAS Phase-1 Limited-Extent Spatial Survey \citep[HAPLESS,][]{Clark:2015}.}
This is the first 250 $\mu$m blind survey selected from the \textit{Herschel} Astrophysical Terahertz Large Area Survey \citep[H-ATLAS,][]{Eales10}
on the basis of their (higher) dust mass content. It consists of 42 nearby galaxies, spanning a range in stellar mass from $ 2.5 \times 10^7$ to $2 \times 10^{11}$ M$_{\odot}$
with a peak at M$_* \simeq 3 \times 10^8$ M$_{\odot} $. They are the most actively star-forming galaxies in H-ATLAS,
with $ -11.8 <$ sSFR $< -8.9 \log_{10}$ yr$^{-1}$, and seem to be in an early stage of converting their gas into stars. HI masses are calculated from the highest-resolution 21 cm
observations available for each galaxy, coming either from the HI Parkes All Sky
Survey \citep[HIPASS,][]{Meyer:2004}, the Arecibo Legacy
Fast ALFA Survey \citep[ALFALFA, ][]{Haynes:2011}, or from other published literature sources.
\paragraph{\citealt{Leroy08}.}
These authors compiled the available data for 23 nearby star-forming galaxies to study star formation rates (SFRs) and efficiencies. Half of them are HI-dominated low-mass ($\rm <10^{10}\,M_\odot$) galaxies and the other half are large spirals. The HI and H$_2$ masses are calculated from integrating the atomic (THINGS, \citealt{Walter08}) and molecular (HERACLES, \citealt{Leroy09} or BIMA SONG, \citealt{Helfer03}) hydrogen surface densities within 1.5 r$_{25}$.
\paragraph{\citealt{Groves15}.}
The correlation between
the total gas mass and the IR luminosity is studied for a
sample of 36 nearby (D$\sim$10 Mpc) galaxies observed
with the KINGFISH, THINGS, and HERACLES surveys.
These galaxies range from 10$^{6.5}$ to 10$^{10.5}$ M$_\odot$ in stellar mass and have disk-like
morphologies.
Molecular gas masses are calculated using HERACLES data by assuming a fixed-to-Galactic-value CO-to-H$_2$ conversion factor and a fixed line ratio of CO(2-1) to CO(1-0) of R$_{21}$ = 0.8 .
\paragraph{APEX Low-redshift Legacy Survey for MOlecular Gas \citep[ALLSMOG,][]{Bothwell:2014}.}
ALLSMOG is designed to observe the CO(2-1) emission line with the APEX
telescope in a sample of local galaxies (0.01 $< z <$ 0.03). These galaxies are from the SDSS data release 7, and their stellar masses are in the
10$^{8.5} - 10^{10}$ M$_\odot$
range.
HI masses are taken
from either HIPASS, ALFALFA,
or from the large collection of HI observations
assembled by \citet{Springob:2005}.
\citet{Bothwell:2014} also provide for this sample the results of the H$_2$ masses calculated using the metallicity-dependent CO-to-H$_2$ conversion factor of \citet{Wolfire10}.
\paragraph{The Faint Irregular Galaxies GMRT Survey \citep[FIGGS,][]{Begum08}.}
This is a HI imaging survey of 65 extremely
faint nearby dwarf irregular galaxies observed with the Giant Meterwave Radio Telescope. It is a subsample of the \citet{Karachentsev04} catalog of galaxies within a distance of 10 Mpc.
HI masses were derived from global HI profiles.
\\
We note that
several galaxies are in common in these samples,
and the atomic and molecular hydrogen masses each reference gives can be different for a same object. We kept all values in our analysis to account for the scatter coming from the discrepancy in these mass determinations.
\\
The SFRs and metallicities of the CLUES star-forming dwarfs were compared to those of galaxies from DGS and KINGFISH.
For these two surveys, SFRs were estimated considering two of the most widely used SFR tracers, the far-ultraviolet (FUV) and the H$\alpha$ luminosities, corrected for
dust attenuation, using either total-infrared (TIR) or 24 $\mu$m luminosities \citep{Kennicutt09, Calzetti10, Hao11, Kennicutt12}. We refer to Section 2.6.1 in \citet{RR15} for details of the estimations.
The metallicities in both samples have been determined in
\citet{Madden13} and \citet{Kennicutt11}
using the calibration from \citet{Pilyugin05}.
\subsection{IR-submm emission}
We compare our IR-submm emission results to data from DGS, KINGFISH, and HAPLESS.
All three galaxy samples have been observed with \textit{Herschel} in the 70, 100, and 160 $\mu$m PACS
and 250, 350 and 500 $\mu$m SPIRE bands \citep[see][respectively]{RR15,Clark:2015,Dale12}.
Mid-infrared data from IRAC 8$\mu$m and MIPS 24$\mu$m \textit{Spitzer} bands for DGS and KINGFISH \citep[see][respectively]{RR15, Dale:2005},
are also studied.
We highlight that GRASIL-3D allows the precise calculation of the luminosities of the simulated galaxies as seen through a very wide range of telescopic filters (with their particular transmission and calibration), allowing for an accurate comparison with the above observational data.
\section{Results: General properties of the sample}\label{sec:prop}
We briefly present here the general properties of
the CLUES star-forming dwarf galaxy sample,
namely, their SFHs, neutral hydrogen mass content, and metallicities, to demonstrate that they constitute a realistic dwarf galaxy sample with properties comparable to observed properties.
The molecular H$_2$ gas masses as calculated by GRASIL-3D are also compared to observational data to justify the choice of $\rho_{mc,thres}$ made.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure3.pdf}}
\caption{Star formation rate (upper panel) and specific star formation rate (lower panel) of the CLUES star-forming dwarf galaxies compared to observational data from the DGS and KINGFISH galaxy samples.}
\label{Figsfr}
\end{figure}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure4.pdf}}
\caption{ M$_{\rm HI}$/M$_*$ (upper panel) and M$_{\rm H2}$/M$_*$ (middle and lower panels) ratios versus stellar mass. CLUES simulated galaxies are shown as red squares, while the various observational samples take the symbols shown in the legend. HI and H$_2$ masses are calculated by GRASIL-3D based on a log-normal PDF for gas densities.
The middle and lower panels differ in the assumption for the observed galaxies of a metallicity-dependent or constant CO-to-H$_2$ conversion factor, respectively, which in GRASIL-3D means the use of different density threshold values $\rho_{mc,thres}$ in order to match observational data. }
\label{FigHcontent}%
\end{figure}
\subsection{Star formation rate history}
The
SFRs at $z=0$
of the CLUES dwarfs were calculated using
the mass of stars born in the past 100 Myr, a time interval that is consistent with SFR tracers such as FUV and
TIR luminosities, used
by the observational samples we compare our results to.
Their star formation rate histories (SFHs)
are depicted in Fig. \ref{FigSFH}
as histograms where each bin represents 100 Myr.
They show
significant diversity, some of them having formed many stars recently, while others have not.
In general, all the simulated dwarf galaxies present a bursty SFH, with an average of $\sim$2.5 important bursty episodes per galaxy, which are generally short lived ($\sim$200 Myr).
In Fig. \ref{Figsfr} we show the SFR and specific SFR (sSFR) at $z=0$ versus stellar mass, compared to observational data from DGS and KINGFISH. Simulated galaxies follow the observed trend with stellar mass in both cases.
We note that a few CLUES galaxies
fall slightly below their observational counterparts at M$_{*}\approx$2-5$\times$10$^{8}$M$_\odot$,
indicating that they host fewer young stars.
This may be due to environmental effects.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure5.pdf}}
\caption{Stellar mass compared to gas metallicity in terms of the oxygen abundance. The CLUES sample of star-forming dwarf galaxies lies in the low-metallicity region overlapping DGS data.}
\label{Figmet}%
\end{figure}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure6.pdf}}
\caption{Metallicity distribution of the CLUES sample of star-forming dwarf galaxies, compared to that of the DGS and KINGFISH surveys.}
\label{Figmethist}%
\end{figure}
\subsection{HI and H$_2$ gas contents} \label{HIH2mass}
To ignore the spurious effects of hot ionized (HII) gas, we only provide GRASIL-3D with gas particles with temperatures lower than 2$\times$10$^4$K.
For the simulated dwarfs that contain hot gas (50\% of the star-forming sample), this temperature cut cleanly separates the cold and hot gas phases in a temperature histogram.
Therefore we can safely assert that
M$_{\rm cold gas}$=M$_{\rm HI}$+M$_{\rm H2}$+M$_{\rm He}$+M$_{\rm Z}$, where M$_{\rm He}$ and M$_{\rm Z}$ are the mass fractions in helium and in metals, respectively.
Concerning atomic and molecular gas fractions,
as explained previously, in GRASIL-3D the fraction of gas in the form of molecular clouds $f_{mc}$ is derived under the assumption that MCs are defined by a density threshold, $\rho_{mc,thres}$, and by a probability distribution function (PDF) for the gas distribution at subresolution scales with dispersion $\sigma$.
When the MC mass is known,
HI and H$_2$ masses are derived by correcting for the fraction of mass in helium and
ignoring the fraction of mass in metals.
We show in Fig. \ref{FigHcontent} the HI and H$_2$ gas contents of
the simulated galaxies
compared to observational data inside the dwarf stellar mass range.
The upper panel proves that the fraction of HI gas mass per total stellar mass in simulated galaxies agrees with observational data.
Regarding the fraction of molecular gas H$_2$, the middle and lower panels show observational data assuming a metallicity-dependent and a constant-Galactic CO-to-H$_2$ conversion factor, respectively.
The gas density threshold of reference used in this work ($\rho_{mc,thres}$=3.3$\times$10$^{9}$ M$_\odot$ kpc$^{-3}$) provides an acceptable match to the observational M$_{\rm H2}$ data resulting from assuming a metallicity-dependent conversion factor (middle panel); however, a value one order of magnitude higher is required to fit observed data that assume a constant conversion factor (lower panel).
This translates into a decrease of M$_{\rm H2}$ in CLUES galaxies of an average factor of $\approx 4$, as can be observed by comparing the middle and lower panels of Fig. \ref{FigHcontent}.
In the following sections, the infrared luminosity results are shown for these two different gas density threshold values.
As we show, qualitatively the final conclusions of this study are unchanged
when using either of these two $\rho_{mc,thres}$ parametrizations.
We therefore do not attempt to
complicate the picture by
considering different methods for estimating the molecular gas fraction or taking physical processes into account that may affect the formation of H$_2$ on dust grains (such as interstellar radiation fields or dust density).
\subsection{Gas metallicity}
Among other metals, GASOLINE traces the evolution of oxygen in gas and stellar particles, which is produced as a yield of SNII and SNIa explosions (see Sect. \ref{gasoline}).
The 12+log(O/H) metallicity for each object was computed by averaging the oxygen mass fraction from cold gas particles inside 2$\times$R$_{\rm half}$, where R$_{\rm half}$ is the median radial distance of gas particles from the center of the galaxy.
We show the mass-metallicity relation of CLUES galaxies
in Fig. \ref{Figmet}. In general, they follow the observational relation, although a few of them with LogM$_{*} \sim$8.4 have lower metallicities than the data.
Furthermore, Fig. \ref{Figmethist} displays the metallicity distribution.
CLUES dwarfs are certainly low-metallicity galaxies: their 12+log(O/H) values range between 7.0 and 8.0 dex,
with a low-metallicity tail showing a very similar number distribution to that of DGS.
{
\subsection{Dust-to-gas mass ratio}\label{dgratio}
Figure \ref{DtoG_metmstar} presents the total (i.e., in a given galaxy) D/G ratio versus 12+log(O/H) gas metallicity colored by stellar mass.
CLUES galaxies at very low metallicities show dispersion in their dust masses, as is observed.
We note, however, that all observed galaxies with metallicities below 12+log(O/H)$\lesssim 7.7$ (and many other at higher metallicities) have both uncertain
i) H$_2$ masses \citep[the absence of data has been replaced by an M$_{\rm H2}$ correction][]{RR14}
and
ii) dust masses, \citep[due to non-detections of the galaxy at wavelengths larger than 160 $\mu$m][]{RR15}.
More observational work is needed to surpass these uncertainties.
We recall that we have assumed a broken power-law dependence of D/G on Z metallicity (see Eq. \ref{dustfrac}) to calculate the CLUES dwarf dust masses.
The dust masses and the consequences on the final SEDs that result by assuming a linear dependence of D/G on Z instead are shown in Sect. \ref{sec:varDG}.
\\
\noindent
Based on this analysis, we can conclude that the properties of the sample of CLUES simulated star-forming dwarfs are consistent with observations, and we proceed to the SED analysis.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure7.pdf}}
\caption{
Dust-to-gas mass ratio versus gas metallicity in terms of oxygen abundances. Data points are colored by stellar mass.
We assume in this study a broken power-law dependency between the D/G ratio and Z metallicity \citep{RR14} in each grid cell (Eq. \ref{dustfrac}).
Taking the uncertainty in the determination of both M$_{\rm H2}$ and M$_{\rm dust}$ for all observed galaxies with 12+log(O/H)$ \lesssim 7.7$ into account,
we consider that this approach provides satisfactory estimates of the total dust masses of the CLUES dwarf sample as compared to DGS and KINGFISH.
}
\label{DtoG_metmstar}%
\end{figure}
\section{Results: The diversity of emission in the IR-submm range} \label{sec:emi}
\subsection{SEDs}
To explore how CLUES star-forming dwarfs behave
regarding their IR-submm emission, we show
in Fig. \ref{FigSEDs} their SEDs as calculated with GRASIL-3D.
The intrinsic stellar emission is shown in red, the MC emission in blue, the emission from cirrus in green, and the black curve gives the angle-averaged total emission.
The SEDs show a surprising variety, mimicking the different characteristics shown in DGS and HAPLESS SEDs as described by \citet{RR15} and \citet{Clark:2015}, respectively.
Specific examples of SEDs with clearly defined patterns are
i) IR peak broadening in
numbers 34, 104, 149, and 164;
ii) submm excess and slope flattening in numbers 26, 32, 50, and 86;
and iii) low PAH emission in
numbers 26, 32, 53, 116, and 125.
Other galaxies show less clear patterns and/or mixed properties, for example,
numbers 23, 34, 154, and 93.
Finally, in three cases none of these features are apparent: numbers 103, 87, and 186.
Contrary to observational data, where no SED decomposition into different dust components is directly available,
GRASIL-3D provides the MC and cirrus emissions.
Fig. \ref{FigSEDs} shows that
two distinct dust components can be distinguished in the simulated galaxies: a colder (cirrus, green) and
a warmer component (MCs, blue).\footnote{
We recall that the terms `cold' and `warm' refer to the location of the respective emission maxima
(sum of the emissions of grains of different types, sizes and temperatures),
therefore they refer to a temperature that is only effective (see Sect. \ref{dustmodel}). }
}
In this way, a first answer to the question of the double dust component SED in dwarf galaxies
suggested in numerous studies
\citep[e.g.,][]{Bendo2010, Bendo2012, Bendo2012b, Boquien2011, RR13}
is naturally given by our modeling: the two components are the emission from
PAH-depleted
dust in MCs that is heated by young stars, and from dust in the cirrus that is heated by more evolved stars.
Each of these two SED components is
shaped by their respective
dust grain
total effective
temperatures
(determined by the location of the emission maxima) and the flux intensities at their
respective maxima.
The final IR-submm SED is the addition of these two (MC and cirrus) contributions.
According to the relative positions of the two maxima and the respective flux values there,
a continuum of possibilities emerges from this combination,
with extremal behaviors occuring when one component dominates over the other.
In the next subsections the implications
of these edge situations and those of the intermediate possibilities
are described and discussed in detail.
To better quantify the comparisons with observations, this is done through color-color analyses.
\subsection{IR peak broadening} \label{sect:broad}
While the SEDs of larger and metal-richer galaxies peak at around 100-160 $\mu$m
\citep[as most KINGFISH SEDs show, see Fig. 13,][]{RR15}, dwarf galaxies in general present a broad IR SED with a wide maximum.
This behavior is reproduced in the GRASIL-3D SEDs of the CLUES dwarf galaxies. In particular,
two different broadening features are visible in Fig. \ref{FigSEDs}, having an equivalent in the
SEDs of DGS galaxies
\citep[see Fig. 12,][]{RR15}.
\begin{enumerate}
\item
In some cases, the MC emission is much stronger (and effectively hotter) than that of the cirrus
in such a way that it hides PAH emission and produces a broad maximum that typically peaks in the 20 - 40 $\mu$m range.
This is the case for numbers 26, 32, 50, 116, and 164,
and
to a lesser extent for
numbers 86, 150, and 364.
A typical observational counterpart is DGS galaxy HS0017+1055.
\item
A different characteristic situation arises when both
the MC and cirrus emissions have similar intensities at their
(separated) maxima, typically at $\lambda \sim 40$ and $\sim 160\,\mu$m, respectively.
In this case, the combined emission is rather flat between the two maxima,
or with a slight slope if the maxima heights are somewhat dissimilar.
Typical examples of this situation are numbers 29, 53, 69, 104, 154, 93, 105, 113, and 125.
On the observational side, DGS galaxy Pox186 is an example, and most SEDs of the HAPLESS galaxies
\citep[Fig. A2 of][]{Clark:2015} could reflect this situation.
\end{enumerate}
To quantify these behaviors, we show in
Fig. \ref{FigBroad} the PACS/PACS
$\nu$L$_{70}$/$\nu$L$_{100}$ vs $\nu$L$_{100}$/$\nu$L$_{160}$ color-color diagram,
which traces the peak of the SED \citep{RR13}.
CLUES star-forming dwarfs (filled squares) are colored according to the amount of energy absorbed by their MCs per unit dust mass in MCs,
(E$_{\rm abs}$/M$_{\rm dust}$)$_{\rm MC}$
(see colorbar).
We assume that the energy absorbed by MCs is equal to their infrared luminosity (i.e., energy emitted), which we compute by integrating the MC curve between 4 and 2000 $\mu$m.
In the upper and lower panels we compare
CLUES dwarf galaxies
to the DGS (crosses) and KINGFISH (circles) samples,
and in the middle panel
we compare
to HAPLESS galaxies (triangles).
We show the HAPLESS
sample separately from DGS and KINGFISH because as explained before, these galaxies were selected for their
high dust content and are therefore a biased sample.
The upper and lower panels differ in the $t_0$ value used.
Contrasting symbols represent galaxies
within the $\rm 10^6<M_*(M_{\odot})<10^9$ stellar mass range, while faint symbols stand for
more massive galaxies.
The two columns in the figure show results using different values for the gas density threshold that determines the molecular gas mass $\rho_{mc,thres}$, 3.3$\times$10$^9$ or 3.3$\times$10$^{10}$ M$_\odot$ kpc$^{-3}$. As explained in Sect. \ref{HIH2mass}, the first matches observational data assuming a metallicity-dependent CO-to-H$_2$ conversion factor, while the higher parameter is required for fitting H$_2$ masses calculated assuming the constant Galactic value.
The upper panels show
most KINGFISH high-mass galaxies gathered in the lower left corner of the figure where $\nu$L$_{70}$/$\nu$L$_{100}<$1 and $\nu$L$_{100}$/$\nu$L$_{160}\sim$1, because they peak at high
wavelengths.
A few CLUES dwarfs whose MCs are poorly heated by young stars
(blue squares,
numbers 70, 76, 87, 103, and 186) have SEDs that also peak at around 100-160 $\mu$m and thus appear in the same location.
However,
most of the simulated, DGS and KINGFISH galaxies
inside the "dwarf" stellar mass range, show a different behavior.
On one hand,
those that show a maximum emission in the 20 - 40 $\mu$m range
(simulated dwarfs with dominant MC emission)
consequently have $\nu$L$_{70}$/$\nu$L$_{100}>$1 and $\nu$L$_{100}$/$\nu$L$_{160}>$1, which places them in the upper right part of the diagram.
We note that DGS galaxies whose SED data have been fit in \citet{RR15} with an additional emission component as a MIR modified black body (marked with open circles) appear in this same area.
On the other hand,
galaxies
showing an almost flat behavior in the $\sim 40-160 \,\mu$m range of their SEDs
are clustered closer to the $\nu$L$_{70}$/$\nu$L$_{100}= \nu$L$_{100}$/$\nu$L$_{160} =1$ position of the diagram.
To reinforce our interpretation that the warmer effective temperatures of MC dust grains in CLUES star-forming dwarfs
are due to the higher energy absorption of
this dust (per unit mass), we decreased the energy input from young stars to $t_0 = 5$ Myr, and
repeated
the SED calculation.
The results are shown in the lower panels of Fig. \ref{FigBroad}.
As expected, energy absorption by MCs decreased for most simulated galaxies, in some cases to the extent that
their MC emission faded away
to reveal the cirrus component
(dark blue squares).
As a result, they gather where the massive KINGFISH galaxies lie.
We therefore postulate that IR peak broadening in DGS galaxies, regardless of the peak position,
has the same origin: the multi-temperature warm dust emission in MCs, heated by young stars.
The two broadening types come from the
relative intensity of the MCs and the cirrus maxima. The stronger the MC emission
(we recall that the maxima appears in the 20 - 40 $\mu$m range), the more displaced
toward the right upper corner in the
$\nu$L$_{70}$/$\nu$L$_{100}$ vs $\nu$L$_{100}$/$\nu$L$_{160}$ color-color diagram, tracing the peak of the SED.
In the middle panels of Fig. \ref{FigBroad} we compare our results to
HAPLESS\footnote{Note that here
the available data are L$_{60}$ instead of L$_{70}$.} dwarf galaxies,
which are biased against low dust mass content.
These galaxies therefore tend to have low values of the energy absorbed by MCs
per unit dust mass, $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$.
Their SEDs are not expected to show MC emission
dominance, but rather broad maxima typically in the range between $\lambda \sim 40$ and $\sim 160\,\mu$m, as is the case.
Finally, we compare the CLUES results shown in the left and right columns of Fig. \ref{FigBroad} for different values of $\rho_{mc,thres}$.
%
The only difference is observed in the color of the representative points,
which take higher values of (E$_{\rm abs}$/M$_{\rm dust}$)$_{\rm MC}$
in the case of an approximately
four times
lower M$_{\rm H2}$
($\rho_{mc,thres}$=3.3$\times$10$^{10}$, right column),
as expected: a lower molecular gas mass (which implies a lower amount of dust) absorbs the same energy coming from young stars.
Nonetheless, the general arrangement of the points across the diagram does not vary,
and therefore
our conclusions remain valid.
\subsection{Submillimeter excess and slope flattening}
\label{Submm-excess}
Figure \ref{FigExcess} shows the PACS/SPIRE
$\nu$L$_{100}$/$\nu$L$_{250}$ vs $\nu$L$_{250}$/$\nu$L$_{500}$ color-color diagram.
The symbol and color-coding are the same as in Fig. \ref{FigBroad}.
This diagram reflects the variation in emissivity index $\beta$, which is
understood as a measure of the final FIR/submm slope of the SED (see Eq. \ref{ModBB}).
Physically, it is an intrinsic optical property of grains: a typical value of $\beta$=2.0 is found for Galactic grains and is commonly used to model the SEDs of high-mass galaxies, showing a good agreement with data.
In contrast, dwarf galaxies have been reported to present flatter slopes, reflecting an excess of emission at submm wavelengths.
In this study, both the MC and cirrus components were modeled assuming $\beta$=2.0,
but their combination yields a final effective slope --or $\beta_{\rm eff}$-- that in general departs from this value.
Figure \ref{FigExcess} shows three curves depicting the theoretical \textit{Herschel} luminosity ratios obtained assuming modified black bodies of different fixed emissivity indices, which we accordingly also call $\beta_{\rm eff}$, drawn from Fig. 10 in \citet{RR13}.
While the bulk of
high-mass galaxies
follow the higher $\beta_{\rm eff}$
(i.e., with no submm excess found),
the observed dwarf galaxies present lower $\nu$L$_{250}$/$\nu$L$_{500}$ ratios, occupying
the $\beta_{\rm eff}<1.5$ region.
We note in particular the presence of the DGS galaxies marked in \citet{RR13} as having an excess of emission at 500 $\mu$m
with respect to modified black body $\beta=2$ fits to the data
(open squares).
This indicates a general flattening of the submm SED slopes of these galaxies as compared to the SEDs of more massive and metal-rich galaxies.
This same behavior is also observed in the SEDs of the CLUES star-forming dwarf galaxies.
When we analyze the figure in more detail,
we see that CLUES dwarfs with an IR slope flattening caused by a dominant MC emission
(i.e., numbers 26, 32, 50, 116, 164, 86, 150, and 364)
appear between the $\beta_{\rm eff}=1.0$ and $1.5$ lines and toward the upper right corner,
where the inequalities
$\nu$L$_{100} \gg \nu$L$_{250}$
and $\nu$L$_{250} \gg \nu$L$_{500}$ clearly hold.
On the other hand, when IR peak broadening is instead caused by a similar MC and cirrus peak intensity,
the $\nu$L$_{100}$ luminosity is similar to, and in most cases only slightly higher than, $\nu$L$_{250}$,
placing these cases close to the dotted line that represents the
$\nu$L$_{100}$= $\nu$L$_{250}$
equality.
Finally, CLUES galaxies with (almost) bare cirrus emission (dark blue squares, numbers 103, 76, 87, and 186)
are the closest to the dotted line and also have the lowest $\nu$L$_{250}$/$\nu$L$_{500}$ ratios.
As in Fig. \ref{FigBroad},
the upper and lower panels of Fig. \ref{FigExcess} differ in the MC escape timescale used in the GRASIL-3D run: $t_0=40$ Myr in the upper panels; $t_0=5$ Myr in the lower panels.
When we use this lower value, far fewer CLUES galaxies have stars young enough to live in MCs, and thus
the energy absorbed is very low (dark blue, see color bar).
Their SEDs
change to roughly present only the cirrus dust component, steepening
their submm slopes and bringing
their representative points
between the lines corresponding to the $\beta_{\rm eff}=$1.5 and 2.0 fits, where massive
galaxies lie.
We conclude that the heating of MCs by young stars drives
the SED flattening around the 500 $\mu$m SPIRE band.
HAPLESS galaxies (middle panels of Fig. \ref{FigExcess}) show the same general behaviour as DGS or KINGFISH,
but we note that they tend to take lower $\nu$L$_{100}$/$\nu$L$_{250}$ values than DGS and KINGFISH because they mostly present two different emission components \citep[see Fig. A2 of][]{Clark:2015}, and it is rare for one of them to dominate over the other.
Comparing the two columns of Fig. \ref{FigExcess},
we again observe the subtle increase in
the $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ parameter
(according to the color bar) when using $\rho_{mc,thres}$=3.3$\times$10$^{10}$ M$_\odot$ kpc$^{-3}$, as a result of the lower MC dust mass.
Since a lower amount of MC dust grains absorbs the same energy coming from young stars, MCs acquire a higher global effective temperature.
On the other hand, the cirrus phase has gained the difference in dust mass,
consequently decreasing its global temperature. These two temperature variations render a greater
separation between both SED components.
As a result, with a higher value of $\rho_{mc,thres}$, the emergence of an excess of submm emission thanks to the combined contribution of the emission of the two dust components is more
likely.
This accounts for the slight displacement of points toward higher $\nu$L$_{250}$/$\nu$L$_{500}$ values, which
is noticeable (barely) only in the case of simulated dwarf galaxies with
similar MC and cirrus emission intensities.
The global differences in the PACS/SPIRE diagram are nonetheless insignificant, therefore we can conclude that the interpretation of our results is independent of the gas density threshold value used,
which translates in that it is independent of an
average factor of 4 higher or lower M$_{\rm H2}$ in the galaxy.
The effects of varying $\rho_{mc,thres}$ as well as the PDF dispersion $\sigma$ are explored in detail in Sect. \ref{ParVar}.
\subsection{PAH emission}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure8.pdf}}
\caption{Stellar mass versus the $\nu$L$_{8.0}$/$\nu$L$_{24}$ ratio, which reflects the strengths of PAH emissions. DGS galaxies show \textit{IRAS} 22 $\mu$m data instead of MIPS 24 $\mu$m.}
\label{FigPAH}%
\end{figure}
The PAH band emission in dwarf galaxy SEDs appears to be very weak,
which has suggested the use of dust models with a low PAH abundance in order to explain the data.
Here we use one such model, characterized by $q_{\rm PAH}=$1.12\%,
a suitable value for low-metallicity galaxies according to \citet{Draineb:2007} (see Sect. \ref{dustmodel}).
The $\nu$L$_{8.0}$/$\nu$L$_{24}$ ratio is a good estimator of the PAH
peak intensities because
it compares the strength of one of the most prominent peaks of the aromatic features (at $\lambda\approx8.6 \,\mu$m) and the luminosity
immediately outside the PAH emission domain.
Figure \ref{FigPAH} shows this IRAC 8$\mu$m/MIPS 24$\mu$m \textit{Spitzer} luminosity ratio compared to stellar mass.
The majority of the CLUES star-forming dwarfs show very low PAH emission, the same as has been reported for DGS \citep{RR15}.
As we noted in Sect. \ref{sect:broad},
MC emission can i) completely hide PAH bands when the maximum of the MC emission is higher than the maximum of the cirrus component,
or ii) partially hide PAH bands when both MC and cirrus maxima tend to have comparable strengths.
Only dwarfs
number 91, 87, 186, 103, and 76 in Fig. \ref{FigPAH}
show a mid-infrared emission
where
$\nu$L$_{8.0}>\nu$L$_{24}$.
These are precisely the cases in which MC emission is poorer
as a result of a low energy absorption from young stars
(dark blue squares in the upper panels of Figs. \ref{FigBroad} and Fig. \ref{FigExcess}). This situation prevents the concealment of PAH cirrus emission.
We propose that the weakness or lack of PAH emission in DGS galaxies
could be caused by the combined contribution of
i) a low PAH abundance and ii) warm dust emission from PAH-devoid MCs that partially or completely hides the cirrus PAH features.
We note that DGS galaxies show \textit{IRAS} 22 $\mu$m data instead of MIPS 24 $\mu$m. CLUES luminosities at 22 $\mu$m are similar to those at 24 $\mu$m, only differing for the few galaxies mentioned
above,
where L$_{22}>$L$_{24}$.
We therefore expect the few DGS galaxies with $\nu$L$_{8.0}$/$\nu$L$_{22}>1$
to slightly shift upward if the L$_{24}$ data were available. This does not affect our results since these points
would still be consistent with the rest of data inside the same stellar mass range.
The rest of DGS galaxies would remain unaltered or even shift downward.
\section{Discussion} \label{sec:disc}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure9.pdf}}
\caption{Correlations of the $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ ratio with distance to diagram positions representative of zero injection of energy into molecular clouds: the $\beta$=2 line in Fig. \ref{FigExcess} and the lower left corner of Fig. \ref{FigBroad}. Pearson correlation coefficients are given in each case.
}
\label{Broad-Corr}%
\end{figure}
\subsection{Nature of the heating engine }
\label{Nature-cold-warm}
According to Fig. \ref{FigSEDs}, the cirrus components of the
simulated dwarf galaxies peak between $100 - 200 \,\mu$m,
wavelengths at which usually the total SEDs of more massive and metal-rich galaxies peak.
Figure \ref{FigSEDs} also shows that the emission from MCs is effectively hotter than usual, showing in addition a wider diversity of shapes among the different galaxies.
In GRASIL-3D modeling, radiation
from evolved
stars heats the diffuse cirrus component where the bulk of dust resides.
In contrast, MCs are heated by low $\lambda$ emissions from young stars
embedded in MCs.
Therefore the amount of energy young stars inject into MCs, $E_{\rm abs,MC}$, is expected to be a key parameter to understand
the particularities in the IR-submm region of dwarf galaxy SEDs.
However, the MC grain temperature is most likely correlated to $E_{\rm abs,MC}$ per unit dust mass,
$(E_{\rm abs}/M_{\rm dust})_{\rm MC}$,
as we show below.
For this reason, simulated results are color-coded by this quantity in Figs. \ref{FigBroad}, \ref{FigExcess}, and \ref{FigPAH}.
In Fig. \ref{Broad-Corr} we show a measure of
the
energy injection effects
by representing the correlation between the $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ parameter and the
distance to points of zero injection in Figs. \ref{FigBroad} and \ref{FigExcess}.
These points are the particular locations in each color-color diagram where galaxies lie when they are scarcely heated by young stars (appearing in dark blue color): the lower left corner of Fig. \ref{FigBroad} and the $\beta_{\rm eff}$=2 curve in Fig. \ref{FigExcess}.
The upper panel of Fig. \ref{Broad-Corr} shows the distance from each CLUES dwarf
in Fig. \ref{FigExcess}
to the
$\beta_{\rm eff}$=2 curve.
Simulated dwarfs with the maximum
$(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ ratio are farther away from
this line than those with lower ratio values,
the correlation having a Pearson coefficient of
r=0.92.
When we only use $E_{\rm abs, MC}$,
that is, when we do not divide by the total MC dust mass,
the correlation decreases to r=0.83.
The lower panel of Fig. \ref{Broad-Corr} shows the distance from the position of each simulated dwarf in Fig. \ref{FigBroad} to the
lower left corner of this diagram (approximately (0.65,0.40)),
and again a clear correlation is visible,
with
r=0.93.
When only $E_{\rm abs,MC}$ is used, the scattering increases
and yields r=0.79.
We
finally note that
although not depicted, concerning PAH emission, low $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ ratio values are correlated to high PAH emission, as can be deduced from Fig. \ref{FigPAH}.
CLUES dwarf galaxies with low or very low PAH emissions correspond to
high $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ ratios.
Our findings are independent of the $\rho_{mc,thres}$ value:
with the high-density threshold value ($\rho_{mc,thres}=3.3 \times 10^{10}$ M$_\odot$kpc$^{-3}$) the same clear correlations appear, improving when $E_{\rm abs}$ is taken by unit dust mass.
These results corroborate that the simulated galaxies' global MC effective temperature,
which
defines the maxima location and
drives the variability in the final IR-submm SED shapes,
is adequately captured by the $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ ratio.
\subsection{Dependence of the results on the dust-to-gas mass assignation versus Z}
\label{sec:varDG}
As described in Sect. \ref{DGratio},
for this particular study on dwarf galaxies we have assumed in GRASIL-3D a
local
dependence of the dust-to-gas mass ratio on metallicity following the broken power-law proposed in Table 1 of
\citet{RR14}.
This scaling
consists of two power-laws, the low-metallicity end
having a steeper slope that provides
a lower total dust mass to such environments, as
suggested observationally and by dust evolution models.
For metal-richer galactic regions, in particular with metallicities
12+log(O/H)$\gtrsim8.0$,
the standard convention of scaling the D/G ratio linearly with metallicity \citep["reference scaling" in][]{RR14} has been adopted, which is long known to satisfactorily fit data
of normal metal-richer galaxies.
We computed the dust masses, SEDs, and IR-submm luminosities that result from assuming
this reference scaling at all metallicities.
Figure
\ref{Figdglinear}
shows the dust-to-gas mass ratios obtained in this case.
Galaxies with 12+log(O/H)$\lesssim8.0$ show no dispersion and
end up with an excessive amount of dust as compared to data.
However, the SEDs obtained with this dependence show that eventually this does not translate in a great difference in the final luminosities.
In fact, only a few galaxies that have particularly very low global metallicities show perceptible changes in their SEDs (presenting a slightly stronger MC emission due to the now higher dust content), while the rest stay nearly the same.
The reason for this low variation of the final SED
is that
the grid cells that contribute the most to the emission
of a given galaxy
are those
with a higher dust content and thus a higher metallicity.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure10.pdf}}
\caption{
Dust-to-gas mass ratio versus gas metallicity in terms of oxygen abundances. Data points are colored by stellar mass. A linear scaling is assumed between the D/G mass ratio and the metallicity inside each grid cell.
}
\label{Figdglinear}%
\end{figure}
\subsection{Dependence of the results on the dust model for cirrus}
Spectra of low-metallicity galaxies show extremely low PAH band emission as compared to more metal-rich galaxies.
The fact that the most prominent peaks in the mid-infrared region are due to single-photon heating of the smallest PAH grains
has led modelers to consider the use of different dust models containing fewer very small PAH grains in order to match the data.
In particular, \citet{Draineb:2007}
proposed seven different models that are
characterized by a different PAH index $q_{\rm PAH}$ each, ranging from 0.47\% to 4.58\%, where lower values are more appropriate for low-metallicity galaxies, and the highest value corresponds to a Galactic abundance.
In Fig. \ref{Figvarqpah} we show the differences in the SED of galaxy number 150 (and specifically in its cirrus emission) between adopting the standard Galactic value of 4.58\% and the value of reference used in this work, 1.12\%.
PAH bands in the latter case are indeed mitigated, while with a high $q_{\rm PAH}$ they gain clear visibility.
We note that the effect would be much weaker in the case of dwarfs with dominating MC emission in the MIR region.
\subsection{Dependence of the results on parameter choices}
\label{ParVar}
Table 2 of DT14 shows the combinations of GRASIL-3D parameters used for their study of normal disk-like galaxies. All values proposed are within observationally measured ranges. Here we discuss the variation of
\begin{itemize}
\item $t_0$, the escape timescale from MCs;
\item $\rho_{mc,thres}$ and $\sigma$, the gas density threshold and the parameter that governs the log-normal PDF function to calculate the gas fraction in MCs; and
\item $\tau_{mc}$, the optical depth of MCs, which is dependent on the values of the mass and radii of single molecular clouds ($m_{mc}$ and $r_{mc}$).
\end{itemize}
\subsubsection{Escape timescale from molecular clouds $t_0$}
To vary $t_0$ means
to vary the amount and spectral distribution of the stellar energy that heats the MCs.
We show in Fig. \ref{FigVart0} the SEDs obtained for galaxy number 26 with $t_0$=5 Myr and $t_0$=40 Myr.
The higher the $t_0$ value, the higher the fraction of the total available stellar energy that heats the MCs.
Since the mass of MCs remains the same but now it absorbs more energy, the global effective MC grain temperature increases. As a result, the peak of the MC component broadens in the mid-infrared region, covering the possible PAH cirrus emission.
With a low value of $t_0$, the percentage of galaxies in our sample with SEDs presenting the characteristic features observed in DGS, HAPLESS, and dwarf KINGFISH galaxies
(i.e., IR peak broadening, submm slope flattening, and low PAH emission)
decreases drastically, as has been shown before in the lower panels of Figs. \ref{FigBroad} and \ref{FigExcess}.
Within GRASIL-3D modeling,
high $t_0$ values are therefore needed to account for dwarf SED properties.
This is consistent with results of S98 and S99, who showed that a high $t_0$ value is necessary to reproduce the SEDs of star-bursting galaxies.
\subsubsection{MC density threshold $\rho_{mc,thres}$ and PDF dispersion $\sigma$}
%
This pair of parameters control the cirrus (HI) and MC (H$_2$) masses, regardless of the value of the MC escape timescale $t_0$.
We have shown the subtle consequences of increasing $\rho_{mc,thres}$ in Figs. \ref{FigBroad} and \ref{FigExcess}, which can be summarized in a slight additional warming of the MC component
(per unit MC dust mass) as a result of a lower MC dust mass.
In this section we show specific examples of the different SEDs obtained with different parametrizations.
Maintaining a constant energy injection, $t_0=40$ Myr,
we compare in Fig. \ref{FigVarPDF} four combinations of these parameters ($\rho_{mc,thres}$ [M$_{\odot}$ kpc$^{-3}$], $\sigma$), in particular,
the evolution in the SED of galaxy number 26
when using the following pairs of values:
(3.3$\times$10$^8$, 3), (3.3$\times$10$^9$, 3), (3.3$\times$10$^9$, 2), and (3.3$\times$10$^{10}$, 2).
This evolution
from left to right
corresponds to a progressive MC (cirrus) mass decrement (increment)
inducing a global increment (decrement) of MC (cirrus) dust temperature.
Both effects produce a gradual separation of the two dust components,
which gives rise to a broader total IR peak and to a flatter submm slope.
In particular, for
a fixed value of $\sigma$=3, the MC mass of galaxy number 26 is reduced by a factor of 1.4 from $\rho_{mc,thres}$=3.3$\times$10$^8$ M$_{\odot}$ kpc$^{-3}$ to 3.3$\times$10$^9$ M$_{\odot}$ kpc$^{-3}$, while with $\sigma$=2, the MC mass is reduced by a factor of 3.8 from $\rho_{mc,thres}$=3.3$\times$10$^9$ M$_{\odot}$ kpc$^{-3}$ to 3.3$\times$10$^{10}$ M$_{\odot}$ kpc$^{-3}$.
We found similar results for all the simulated dwarfs in the sample.
Furthermore, the central panels of Fig. \ref{FigVarPDF} show that the effects of reducing the PDF dispersion $\sigma$ from 3 to 2 are qualitatively the same (and quantitavely similar) to increasing the density threshold
one order of magnitude: MC mass and emission decreases (by a factor of 2.8 for galaxy number 26) while the cirrus mass and emission increases.
We note that although the SED of galaxy number 26 with $\rho_{mc,thres}$=3.3$\times$10$^8$ M$_{\odot}$ kpc$^{-3}$ seems to satisfy all observed dwarf SED characteristics, the eventual H$_2$ mass is at the limit of agreeing with observational data using a metallicity-dependent CO-to-H$_2$ conversion factor, and it does not agree with data using a constant factor (see the middle and lower panels of Fig. \ref{HIH2mass}).
If the remaining galaxies in our sample change by a factor similar to galaxy number 26,
not only would it lead to an excess of molecular gas, but to a great shortage of HI neutral gas. Therefore we do not consider this low $\rho_{mc,thres}$=3.3$\times$10$^8$ M$_{\odot}$ kpc$^{-3}$ value in our study.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure11.pdf}}
\caption{
Comparison of SEDs of CLUES galaxy number 150 using different values for the PAH abundance or PAH index $q_{\rm PAH}$ appropriate either for the Galaxy (4.58\%) or for low-metallicity galaxies (1.12\%).
}
\label{Figvarqpah}%
\end{figure}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figure12.pdf}}
\caption{Comparison of SEDs of CLUES galaxy number 26 using different values for the $t_0$ GRASIL-3D parameter.
}
\label{FigVart0}%
\end{figure}
\subsubsection{Masses and radii of molecular clouds}
The values of $m_{mc}$ and $r_{mc}$ are constrained observationally to range typically between 10$^5$-10$^6$ M$_\odot$ and 10-50 pc for normal spiral galaxies, as discussed in S98, S99, and DT14.
We have studied the impact on the SED of varying
the mass of single MCs within its allowed range, while fixing the radius at 14 pc. This approximation is consistent since it its their combined response ($m_{mc}/r_{mc}^2$) what matters, affecting the total emission through the optical depth of MCs $\tau_{mc}$.
We calculated the SEDs of galaxy number 26 taking $m_{mc}$=10$^6$, 3$\times$10$^5$ and 10$^5$ M$_\odot$, and the rest of the parameters as usual.
Figure \ref{FigVarMMC} shows that
as $m_{mc}$ decreases, $\tau_{mc}$ decreases (which means MCs absorb less energy per unit length).
With extremely low $m_{mc}$
a situation can be reached
where UV photons escape from the MCs to the ISM, heating the cirrus.
This would occur at the expense of the MIR emission of MCs, which gradually become unable to hide the
already low
PAH
emission from the cirrus.
In addition, in this case, the
emission and temperature of the cirrus would increase, favoring PAH band detectability.
We therefore
used $m_{mc}$ $=$10$^6$ M$_\odot$ here
to account for dwarfs with a very broad IR peak in the 20 - 40 $\mu$m range
and no detected PAH emission features.
\begin{figure*}
\resizebox{\hsize}{!}{\includegraphics{figure13.pdf}}
\caption{Comparison of SEDs of galaxy number 26 obtained with different values of the density threshold $\rho_{mc,thres}$ and the PDF dispersion $\sigma$.}
\label{FigVarPDF}%
\end{figure*}
\section{Summary and conclusions} \label{conclu}
The IR-submm emission of low-metallicity dwarf galaxies considerably differs from that of more metal-rich dwarf galaxies.
Signatures of warmer dust in dwarf galaxies were first found with \textit{IRAS} and were later on confirmed with \textit{Spitzer}.
More recently, an important step forward in the analysis of their IR-submm emission was made thanks to
the \textit{Herschel} guaranteed time key program
The Dwarf Galaxy Survey
\citep[DGS,][]{RR13}.
The surprising variety of SED shapes found therein \citep[see also HAPLESS results, ][]{Clark:2015}
can be summarized as
(1) a broadening of the IR peak of the SED, implying an additional warmer dust component;
(2) an excess of emission in the submm ($\sim$500 $\mu$m) that causes a flattening of the submm/FIR slope;
and (3) a very low intensity of PAH emission features.
To fit these particular SED patterns, observers change their modeling relative to the standard
fixed $\beta=2$ in the FIR-submm range \citep[see Eq. \ref{ModBB} and ][]{RR13}
and/or add new ad hoc additional dust components to it \citep{RR14,RR15,Clark:2015}.
With the aim of going a step further and providing a physical explanation to these particular
emission features, the SEDs of a sample of 27 simulated star-forming
dwarf galaxies were analyzed in detail and compared to DGS, HAPLESS, and dwarf KINGFISH galaxies.
The sample of simulated dwarf galaxies comes from a single simulation run with
the GASOLINE code \citep{wadsley04},
out of initial conditions provided by the CLUES project
(Constrained Local UniversE Simulations, \citealt{gottloeber10,yepes14}).
These simulated dwarfs have stellar masses, star formation rates at $z=0$,
HI and H$_2$ contents, and metallicities that satisfactorily reproduce observational dwarf galaxy data.
The IR-submm properties of the simulated sample were obtained from their
SEDs, calculated using the GRASIL-3D radiation transfer code \citep{Dominguez:2014}.
The particular strengths of GRASIL-3D compared to other codes
can be summarized as follows: i) the radiative transfer is solved in a grid; ii) it is designed to separately treat the radiative transfer in molecular clouds (MCs) and in the diffuse cirrus component,
whose dust compositions are different;
iii) it takes into account that younger stars are associated with denser ISM environments by means of an age-dependent dust-reprocessing of stellar populations;
iv) it includes a detailed non-equilibrium calculation for dust grains with diameter smaller
than $a_{flu} \sim 250$ \AA, as required.
In GRASIL-3D modeling, by MCs we mean the densest subvolumes of the gas density distribution that surround young stars and absorb their high-energy photons.
These emissions are reemitted in the IR-submm spectral range, with the radiative transfer
equation solved with techniques appropriate for thick media \citep{Granato:1994}.
Thanks to the separate treatment of the radiative transfer in MCs and in the cirrus, the IR-submm SEDs
of the simulated CLUES dwarfs can be directly decomposed into their MC and cirrus
contributions. It has been found that these two components have
clearly separated emission maxima,
with that of the cirrus
spanning some 50 $\mu$m around $\lambda = 150\,\mu$m,
a range where the maxima of normal galaxies often lie,
while that of MCs appears at lower $\lambda$
values and with a variety of shapes. The respective maxima imply that
dust grains in MCs reach higher global effective temperatures than their counterparts in the cirrus.
A second important element causing the diversity of emission features is the relative intensities of the MC and the cirrus emission at their maxima.
Two particular situations arise when i) one component (generally MC emission) clearly dominates or when ii)
both components have similar heights at their respective maxima.
Of course a continuum of possibilities exists in between these two, describing SEDs in Fig. \ref{FigSEDs}.
When MC emission dominates,
a broad
maximum typically appears
in the 20 - 40 $\mu$m range,
providing MIR photons, and (partially) hiding PAH emissions in some cases.
This can explain features 1 and 3 described above.
Although less common, there are also some examples where cirrus emission dominates the SED and the characteristics of more massive metal-rich galaxies are recovered.
A different situation arises when both the MC and cirrus emissions have similar intensities at their
(separated) maxima, typically at $\lambda \sim 40$ and $\sim 160\,\mu$m, respectively.
In this case, their combined emission is rather flat between the two maxima,
and PAH emission hiding is less likely.
To quantify these behaviors, an analysis of the PACS/PACS
$\nu$L$_{70}$/$\nu$L$_{100}$ vs $\nu$L$_{100}$/$\nu$L$_{160}$
color-color diagram, which traces the peak of the SED,
was performed. Simulated galaxies with SEDs belonging to the previous situations appear
segregated in this plot, with intermediate situations at intermediate plot locations.
In addition, we found that the driving parameter that causes the segregation
is the amount of energy young stars inject into MCs
per unit dust mass
$(E_{\rm abs}/M_{\rm dust})_{\rm MC}$,
which takes higher values in the case of dominant MC emission (i.e., a broader IR peak).
Some simulated dwarfs show a flattening of the slope of their SED at submm wavelengths, as compared to that of more massive and metal-rich galaxies.
These simulated dwarfs are precisely those whose
MC and cirrus contributions
are more separated from each other,
their combination
giving an effective excess of emission in the submm region (point 2 above).
It is important to remark that
although we used a fixed value of $\beta = 2$
to model the individual MC and cirrus emissions,
a lower final effective slope, $\beta_{\rm eff}$, appears when they blend.
To further explore this effect,
we used
the PACS/SPIRE
$\nu$L$_{100}$/$\nu$L$_{250}$ vs $\nu$L$_{250}$/$\nu$L$_{500}$
color-color diagram,
which best reflects the variations of
$\beta_{\rm eff}$.
By comparing our results to
the theoretical luminosity ratios that would be obtained assuming a modified black body of different fixed $\beta_{\rm eff}$ \citep[Fig. 10,][]{RR13},
we found that
CLUES dwarfs
lie in between the $\beta_{\rm eff}$ = 1.5 and 1.0 lines, and
off
the $\beta_{\rm eff}$ = 2 line.
A clear trend with the
$(E_{\rm abs}/M_{\rm dust})_{\rm MC}$
ratio
is visible, with higher ratio dwarfs lying farther away from the
$\beta_{\rm eff}$ = 2 line (i.e., representing no submm excess) than those with lower ratio.
This is a strong suggestion that the $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ parameter
adequately describes and drives the submm excess shown by some CLUES dwarfs
and, presumably, that of DGS, HAPLESS, and dwarf KINGFISH galaxies.
This suggestion is corroborated by Fig. \ref{Broad-Corr}.
As said above, a high $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ ratio implies
MC emission in the MIR range, potentially hiding PAH band emissions.
To be quantitative, we plotted the $\nu$L$_{8.0}$/$\nu$L$_{24}$ ratios versus the corresponding stellar masses and
$(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ ratios.
The result is that
despite the low PAH abundance of the dust model adopted in this study (characterized by a PAH index of $q_{\rm PAH}$=1.12\%),
the lower the $(E_{\rm abs}/M_{\rm dust})_{\rm MC}$ value, the more prominent
the PAH band emission appears in the SED. This result strongly supports the idea that within our modeling,
MC dust grains heated by young star light emit in the MIR range and hide the possible residual cirrus PAH band emission.
\begin{figure*}
\resizebox{\hsize}{!}{\includegraphics{figure14.pdf}}
\caption{Comparison of SEDs of CLUES galaxy number 26 using different values of the mass of a single molecular cloud $m_{mc}$.}
\label{FigVarMMC}%
\end{figure*}
GRASIL-3D results depend upon
i) the implementation of the dust-to-gas mass ratio dependence on metallicity,
ii) the dust model, and
iii) internal modeling parameters.
Concerning i),
a broken power-law dependence of the dust-to-gas mass ratio on metallicity has been set
as suggested by dust evolution models
in order to better match observed low-metallicity dwarf galaxy dust masses.
This implementation does not produce an important effect on the final SEDs with respect to a linear scaling because of the higher influence of high-metallicity grid cells on the final total luminosities.
Regarding ii),
a simple dust model consisting of silicates, graphites, and carbonaceous PAH grains was assumed in both MCs and the cirrus, with the size distributions for each grain type following \citet{Weingartnera:2001} and \citet{Draine:2007}.
The very low PAH band emission observed in dwarf galaxies has compelled modelers to adopt dust models with low PAH abundances. In this study we also assumed such a model, characterized by a PAH index $q_{\rm PAH}$ of 1.12\%. We showed that although MC dominant emission in dwarf galaxies with very recent star formation can hide the possible PAH emission, which minimizes the ultimate relevance of this parameter, the use of a standard-Galactic value for this index ($q_{\rm PAH}$=4.58\%) would lead to an overly easy visibility of PAH bands in the SEDs of a substantial fraction of the galaxies of our sample, which does not agree with DGS or dwarf KINGFISH statistics.
We also note
the importance of the prescription used in this work with respect to the PAH abundance in MCs, where it has been
drastically lowered with respect to the cirrus \citep[following][]{Vega:2005} in order to show no PAH band emission.
%
iii)
Finally,
in order for the SEDs of the simulated dwarf galaxies to recover the particular features of DGS, HAPLESS, and dwarf KINGFISH galaxies,
the GRASIL-3D parameter space had to be somewhat limited.
In particular, the lack of PAH emission detections demands
high values for $t_0$,
(the time young stars are enshrouded within MCs)
and
excludes low values of the masses of individual MCs.
The effects on the SED of varying the parameters determining the total MC mass in a given galaxy
($\rho_{mc, thres}$ and $\sigma$) are not that remarkable,
but their values are limited by the final HI and H$_2$ masses obtained.
Different values for the gas density threshold $\rho_{mc, thres}$ are required
in order to better match observational molecular gas mass data obtained by means of a
metallicity-dependent or a constant-Galactic CO-to-H$_2$ conversion factor. In particular, a value one order of magnitude higher is required for the latter case.
To summarize, the analysis of the SEDs of simulated star-forming dwarf galaxies, calculated with
GRASIL-3D, and their comparison to real data from DGS, HAPLESS, and KINGFISH dwarfs, allows us to conclude the following.
\begin{itemize}
\item The SEDs of simulated dwarf galaxies naturally reproduce the particular spectral features that observed dwarf galaxies show in the IR-submm region.
\item In this spectral range, the SEDs receive two contributions: one from dust in the diffuse gas component (cirrus),
and a second contribution from dust in the densest component (MCs).
\item
The emission maxima of the MC and cirrus components
are clearly separated, with the former typically at $\lambda \sim 40\,\mu$m and the latter at $\lambda \sim 160\,\mu$m.
The intensities of MCs and cirrus at their maxima can also differ.
\item These two variables (maxima separation and relative intensity of the two dust components at their maxima)
are responsible for shaping the diversity of IR-submm SEDs of dwarf galaxies.
\item More specifically, broad peaks in the MIR region appear when MCs dominate the total emission,
while SEDs with flat configurations between $\lambda \sim 40-160\,\mu$m
result when the MC and cirrus emission components have similar peak intensities.
\item
The submm excess can be explained as a result of the combination of the separated MC and cirrus contributions, with no need of imposing a $\beta \neq 2$ value.
\item The lack of detected PAH emission is the result of
i) a low proportion of PAHs in the total dust in the cirrus, plus ii)
a wealth of MIR emission by hot dust
grains within PAH-devoid MCs, hiding the possible residual cirrus PAH emission.
\item The driving parameter adequately describing these particular features is the
amount of energy injected per unit dust mass in MCs. The energy comes from young star emissions.
\item The lower this parameter, the less apparent the particular features.
In the limit of very low parameter values, the higher-mass higher-metallicity galaxy behavior is recovered.
%
\end{itemize}
Our conclusions are expected to remain valid when numerical models
of dust formation and evolution are used, for example, if dust formation and evolution
had been self-consistently implemented in the simulation. The reason is that
the relevant regions of the parameter space consistent with observational
data were explored with enough detail so as to cover the possible results
of numerical models at $z=0$.
Therefore, the GRASIL-3D two-component dust model gives a sound physical interpretation of the emission of low-mass low-metallicity (dwarf)
galaxies, with dust grains within MCs potentially providing a wealth of MIR photons (i.e., the warm dust
component). Combined with the emissions from colder grains within the cirrus, SEDs in the IR-submm
range are obtained that satisfactorily agree with the particular patterns observed at these wavelengths.
\begin{acknowledgements}
We would like to thank the referee, whose insightful comments
have helped to improve the quality of this paper.
We thank S. Gottloeber and G. Yepes for a careful reading of the manuscript and their comments.
We thank the CLUES collaboration for providing the initial conditions for the simulation analyzed in this work.
This work was partially supported through MINECO/FEDER (Spain) grants AYA2012-31101 and
AYA2015-63810-P.
ISS thanks financial support through the first grant.
CB also thanks the Ramon y Cajal program.
We acknowledge the Centro de Computaci\'on Cient\'ifica (Universidad Aut\'onoma de Madrid,
Red Espa\~nola de Supercomputaci\'on) for computational support.
The CLUES simulations were performed and analyzed at the High Performance Computing Center Stuttgart (HLRS).
We thank DEISA for access to computing resources through DECI projects SIMU-LU and SIMUGAL-LU and the generous allocation of resources
from STFC’s DiRAC Facility (COSMOS: Galactic Archaeology),
the DEISA consortium, co-funded through EU FP6
project RI-031513 and the FP7 project RI-222919 (through the
DEISA Extreme Computing Initiative), the PRACE-2IP Project (FP7 RI-283493).
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{i}
Stationary solutions of an evolutionary system provide an important piece of information concerning the behavior in the long run.
For systems with background in classical fluid mechanics, stationary solutions typically minimize the entropy production and play the role
of an attractor, at least for energetically insulated fluid flows, see e.g. \cite{FeiPr}.
{The principal question arising in the context of randomly driven systems} is the existence of a (stochastic) steady state solution for the system. Earlier results in this direction concern the incompressible case: Flandoli \cite{Fl1} proved existence of an invariant measure by the ``remote start'' method in 2D case. This result has been extended in a few papers, for instance in Goldys and Maslowski \cite{GoMa1}, \cite{GoMa2} where existence of invariant measure by the method of embedded Markov chain theory verifying also the exponential speed of convergence to invariant measure. A different approach has been adopted by Hairer and Mattingly \cite{HM06} in which case a slightly weaker convergence result (implying however the uniqueness of invariant measure) has been shown under much weaker conditions on the nondegeneracy of the noise.
In the paper \cite{BMO} the existence of invariant measure is proved for 2D Navier-Stokes equation on unbounded domain by a compactness method in the (weak) bw-topology.
In 3D much less is known in the case of incompressible fluid. The problems here appear already on the level of Markov property induced by the equation. Transition Markov semigroup has been constructed in the papers by Da Prato and Debussche \cite{DaPDe1}, \cite{DaPDe2}, provided the noise term is sufficiently rough in space. However, even the problem of uniqueness of transition Markov semigroup remains open. A different approach was adopted by Flandoli and Romito \cite{FlaRom} who used the classical Stroock-Varadhan type argument to find a suitable Markov selection and construct the semigroup. The transition semigroup is shown to be be exponentially ergodic (under appropriate conditions on the noise term) by the same arguments as in \cite{GoMa1}. However, the uniqueness of the Markov transition semigroup has not been proved so far.
In absence of the Markov property (i.e. in the situation when the concept of invariant measure as a steady state is not well defined) it is possible to work directly with stationary solutions, i.e. with solutions which are strictly stationary stochastic processes. In the pioneering paper Flandoli and Gatarek \cite{FlaGat} existence of such stationary solution has been shown in the 3D incompressible case by means of finite-dimensional approximations.
{To our best knowledge, no relevant results on large time behavior have been achieved so far for compressible stochastic fluid flows,
where the situation is much more complex.
To fill at least partially this gap,} we examine the class of stationary
solutions for a stochastically driven Navier-Stokes system:
\Cbox{Cgrey}{
\begin{eqnarray}
\label{i1}
{\rm d} \varrho + {\rm div}_x (\varrho \vc{u}) \,{\rm d} t &=& 0, \\
\label{i2}
{\rm d} (\varrho \vc{u}) + {\rm div}_x (\varrho \vc{u} \otimes \vc{u}) \,{\rm d} t + \nabla_x p(\varrho) \,{\rm d} t &=& {\rm div}_x \mathbb{S}(\nabla_x \vc{u}) \,{\rm d} t + \mathbb{G}(\varrho, \varrho \vc{u}) \ {\rm d} {W}, \\
\label{i3}
\mathbb{S} (\nabla_x \vc{u}) &=& \mu \left( \nabla_x \vc{u} + \nabla_x^t \vc{u} - \frac{2}{3} {\rm div}_x \vc{u} \mathbb{I} \right) + \eta {\rm div}_x \vc{u} \mathbb{I},
\end{eqnarray}
}
\noindent
where $\varrho = \varrho(t,x)$ is the mass density and $\vc{u} = \vc{u}(t,x)$ the macroscopic velocity of a compressible viscous fluids contained in a physical domain
$\mathcal O \subset \mathbb{R}^3$. Here the symbol $p=p(\varrho)$ denotes the pressure, typically given by the isentropic state equation
\begin{equation} \label{i3a}
p(\varrho) = a \varrho^\gamma, \ a > 0,
\end{equation}
and $\mathbb{S}$ is the viscous stress tensor determined by Newton's rheological law (\ref{i3}), with viscosity coefficients $\mu > 0$, $\eta \geq 0$.
The stochastic driving force is represented by the stochastic differential of the form
\begin{equation*} \label{i3b}
\mathbb{G}(\varrho,\varrho\mathbf u)\, {\rm d} W = \sum_{k = 1}^\infty \vc{G}_k(x,\varrho,\varrho\mathbf u)\, {\rm d} W_k,
\end{equation*}
where $W=(W_k)_{k\in\mathbb{N}}$ is a cylindrical Wiener process specified in Section \ref{s} below.
As pointed out above,
in contrast with the frequently studied \emph{incompressible} Navier-Stokes system, the problems related to the dynamics of
\emph{compressible} fluid flows driven by stochastic forcing are basically open.
First existence results were based on a suitable transformation formula that allows to reduce the problem to a random system of PDEs: The stochastic integral does no longer appear and deterministic methods are applicable, see \cite{MR1760377} for the 1D case, \cite{MR1807944} for a rather special periodic 2D. Finally, the work by Feireisl, Maslowski, Novotn\'y \cite{MR2997374} deals with the 3D case.
The first ``truly'' stochastic existence result for the compressible Navier-Stokes system perturbed by a general nonlinear multiplicative noise was obtained by Breit, Hofmanov\'a \cite{BrHo}. The existence of the so-called finite energy weak martingale solutions in three space dimensions with periodic boundary conditions was established. Extension of this result to the zero Dirichlet boundary conditions then appeared in \cite{2015arXiv150400951S,MR3385137}. For completeness, let us also mention \cite{BrFeHo2015,BrFeHo2015A,BrFeHo2016} where further results appeared, namely, a singular limit, the so-called relative energy inequality and the local existence of strong solutions, respectively.
{Our goal is to establish the existence of global--in--time solutions to system (\ref{i1})--(\ref{i3}) that are stationary in the stochastic sense.
To this end, we use a \emph{direct} method based on the four layer approximation scheme developed in
\cite{BrHo} inspired by \cite{feireisl1}. More specifically, the stationary solutions are constructed at the very basic approximation level. The final result is obtained by means of a combination of deterministic and stochastic compactness methods.}
To be more precise, the equations are regularized {by adding artificial viscosity and an artificial pressure term to the momentum equation (\ref{i2}).}
Thus one is led to study the following approximate system
\begin{align}\label{eq:approx}
\begin{aligned}
\mathrm{d} \varrho+{\rm div}_x(\varrho\mathbf u)\mathrm{d} t&=\varepsilon\Delta\varrho\,\mathrm{d} t,\\
\mathrm{d}(\varrho\mathbf u)+\big[{\rm div}_x(\varrho\mathbf u\otimes\mathbf u)
+a\nabla \varrho^\gamma+\delta\nabla\varrho^\Gamma \big]\mathrm{d} t&= \varepsilon \Delta (\varrho \vc{u}) + {\rm div}_x\mathbb{ S}(\nabla_x \vc{u})\,\mathrm{d} t+\mathbb{G}(\varrho,\varrho\mathbf u) \,\mathrm{d} W,
\end{aligned}
\end{align}
where $\Gamma>\max\{\frac{9}{2},\gamma\}$. {For technical reasons, explained in detail in \cite{feireisl1},
the two limits $\varepsilon \to 0$, $\delta\to 0$ must be distinguished and performed in this order.}
{To find stationary solutions for \eqref{eq:approx} with $\varepsilon>0$ and $\delta>0$ fixed, two additional approximation layers are needed.} {Namely, a suitable Faedo-Galerkin approximation of \eqref{eq:approx} of dimension $N\in\mathbb{N}$, with certain truncations of various nonlinear terms (corresponding to a parameter $R \in\mathbb{N}$).}
Letting $R\rightarrow\infty$ gives a unique solution to the Faedo-Galerkin approximation.
The passage to the limit as $N\rightarrow\infty$ yields existence of a solution to \eqref{eq:approx}.
Except for the first passage to the limit, we always employ the stochastic compactness method. However, due to the delicate structure of \eqref{i1}--\eqref{i3} it is necessary to work with weak topologies and therefore we are lead to Jakubowski's generalization of the classical Skorokhod representation theorem \cite[Theorem 2]{jakubow}. It applies to a large class of topological spaces, the so-called quasi-Polish spaces, including (but not limited to) separable Banach spaces equipped with weak topologies.
Another important ingredient of the proof is then the identification of the limit in the nonlinear terms. To be more precise, two main difficulties arise. First, the passage to the limit in the terms that depend nonlinearly on $\varrho$ (i.e. the pressure term and the stochastic integral) cannot be performed directly since strong convergence of the approximate densities does not follow from the compactness argument. This issue appears already in the deterministic setting and is overcome by a technique based on regularity of the effective viscous flux introduced by Lions \cite{LI4}. A suitable stochastic version of this method was developed in \cite{BrHo} to treat \eqref{i1}--\eqref{i3}. {Note, however, that the stationary problem is rather different from the initial--value problem, where
compactness of the initial density field can be incorporated by a suitable choice of the initial data. Here, in analogy with the deterministic
approach developed in \cite{EF56}, compactness of the denisty must be recovered from stationarity of the flow.}
The second difficulty one has to face arises in the passage of the limit in the stochastic integral. Indeed, one has to deal with a sequence of stochastic integrals driven by a sequence of Wiener processes. One possibility is to pass to the limit directly and such technical convergence results appeared in several works (see \cite{bensoussan} or \cite{krylov}), a detailed proof can be found in \cite{debussche1}. Another way is to show that the limit process is a martingale, identify its quadratic variation and apply an integral representation theorem for martingales, if available (see \cite{daprato}). The existence theory for \eqref{i1}--\eqref{i3} developed in \cite{BrHo} relies on neither of those and follows a rather new general and elementary method that was introduced in \cite{on1} and already generalized to different settings.
The main goal of the present paper is {to show the existence of stationary solutions to \eqref{i1}--\eqref{i3}
in the framework of weak martingale solutions introduced in \cite{BrHo}. Although the multi-level approximation procedure is
identical with that used in \cite{BrHo}, the uniform estimates necessary for the existence theory are in general not suitable} to study the long-time behavior of the system. They are based on the application of Gronwall's lemma and therefore grow exponentially with the final time $T$. Hence, the major challenge is to derive new estimates which are uniform with respect to all the approximation parameters as well as in $T$. This is the heart of the paper. Let us point out that the standard methods used for the incompressible system, as for instance in \cite{FlaGat}, \cite{FlaRom}, are not applicable in the compressible case. Indeed system \eqref{i1}--\eqref{i3} is of mixed hyperbolic-parabolic type and the dissipation term does not contain the density. Consequently, the forcing terms on the right-hand side of the energy balance cannot be absorbed in the dissipative term appearing on the left-hand side {in an obvious straightforward manner.}
Furthermore, it does not seem to be possible to find {universal estimates that would be uniform} in all the parameters $R,N,\varepsilon,\delta$ as well as in $T$.
Instead, during each approximation step we develop new estimates which are then used for the particular passage to the limit at hand. More precisely, at the {starting} level, that is for fixed parameters $R,N\in\mathbb{N},$ $\varepsilon,\delta>0$, we show existence, uniqueness and continuous dependence on the initial condition. Thus, the resulting system is Markovian and the transition semigroup is Feller. Consequently, the existence of invariant measures can be shown with the help of the standard Krylov--Bogoliubov method in the infinite-dimensional setting. This generates a family of approximate stationary solutions. Note that we loose uniqueness already after the first passage to the limit (in $R$). Hence the usual Krylov--Bogoliubov approach cannot be employed anymore, and even the concept of invariant measure becomes ambiguous. To overcome this problem we construct stationary solutions on the next level as limits of the corresponding approximate stationary solutions from the previous level.
At each approximation step, there are essentially three necessary estimates: for the energy, the velocity and the pressure. At the deepest level, we are able to obtain the first two estimates uniformly in $R,N$ but the third one depends on all the parameters $R,N,\varepsilon,\delta$ and is therefore not suitable for any limit procedure. The key observation is that these estimates may be significantly improved if we take stationarity into account. Therefore, working directly with stationary solutions given by the Krylov--Bogoliubov method, we derive an estimate for the energy as well as the velocity which is uniform in all the approximation parameters. The estimate for the pressure is more delicate and has to be reproved at each level by applying a suitable test function to \eqref{i1}--\eqref{i3}.
The proof is then concluded by performing the limit for vanishing approximation parameters based on a combination of deterministic and probabilistic tools, similarly to \cite{BrHo}.
{
It is remarkable that our result holds for the same range of the adiabatic exponent $\gamma > \frac{3}{2}$ as in the nowadays available existence theory. Note that the relevant deterministic problem, namely the existence of bounded absorbing sets and attractors require a rather inconvenient technical restriction
$\gamma > \frac{5}{3}$, see \cite{EF56}, \cite{FP14}. Indeed, consider the iconic example of the driving force $\varrho \vc{f}(x) {\rm d} W$ in (\ref{i2}).
If we replace it by the deterministic forcing $\varrho \vc{f}(x) {\rm d} t$, then, to the best of our knowledge, it is not known if the global-in-time weak solutions
remain uniformly bounded for $t \to \infty$ for $\gamma$ in the physically relevant range $1 \leq \gamma \leq 5/3$. On the other hand, the
stochastic forcing $\varrho \vc{f}(x) {\rm d} W$ gives rise to stationary solutions for any $\gamma > 3/2$ as shown in Theorem \ref{Tm1}. The reason is
the cancelations of certain terms in the energy balance due to stochastic averaging. We therefore observe a kind of regularizing effect
due to the presence of noise.
Note, however, that the growth conditions imposed on the diffusion coefficients $\mathbb{G}(\varrho, \varrho \vc{u})$ appearing in the driving term
are more restrictive than in \cite{BrHo}. }
The rest of the paper is devoted to the proof of {existence of a stationary solution to the compressible Navier--Stokes system as stated in} Theorem \ref{Tm1}
{below}. The precise setting is given in Section \ref{m}. In Section \ref{b}, we introduce the basic finite-dimensional approximation and construct
a family of approximate solutions adapting the standard Krylov--Bogoliubov method. In Section \ref{d}, we develop global-in-time estimates for stationary solutions and pass to the limit $R\to\infty$ and $N\to\infty$. Section \ref{L} is devoted to the vanishing viscosity limit, i.e. $\varepsilon\to0$. Finally, in Section \ref{P}, we perform the limit for vanishing artificial pressure, i.e. $\delta\to0$, obtaining the desired stationary solution, the existence of which is claimed in Theorem \ref{Tm1}.
\section{Mathematical framework}
\label{m}
\subsection{Boundary conditions}
Although the boundary conditions in the real world applications may be quite complicated and of substantial influence on the fluid motion, our goal is to focus on the
effect of stochastic perturbations imposed through stochastic volume forces. Accordingly, we
consider the periodic boundary conditions, where the physical domain may be identified with the flat torus
\[
\mathbb{T}^3 \equiv \left( [-1,1]|_{\{ -1,1 \}} \right)^3.
\]
On the other hand, however, our method leans essentially on the dissipative effect of viscosity represented by $\mathbb{S}$ in (\ref{i2}).
In particular, it is convenient to keep a kind of Korn--Poincar\' e inequality in force. Following the idea of Ebin \cite{Eb}, we consider the
{physically relevant} \emph{complete slip} conditions
\color{black}
\Cbox{Cgrey}{
\begin{equation} \label{i4}
\vc{u} \cdot \vc{n}|_{\partial \mathcal O} = 0, \ [\mathbb{S}(\nabla_x \vc{u}) \cdot \vc{n}] \times \vc{n}|_{\partial \mathcal O} = 0
\end{equation}
imposed on the boundary of the cube
\[
\mathcal{O} = [0,1 ]^3.
\]
\color{black}
}
The crucial observation is that
the constraint (\ref{i4}) is automatically satisfied by \emph{periodic} functions $\varrho$, $\vc{u}$ defined on torus $\mathbb{T}^3$ and belonging to the symmetry class
\begin{equation} \label{i6}
\begin{split}
\varrho(t, -x) &= \varrho(t,x) \qquad \ x \in \mathbb{T}^3,\\
u^i(t, \cdot, - x_i, \cdot) &= - u^{i}(t, \cdot, x_i, \cdot)\qquad i = 1, 2, 3 , \\
u^i(t, \cdot, - x_j, \cdot) &= u^i(t, \cdot, x_j, \cdot) \qquad i \ne j, \ i,j=1, 2, 3,
\end{split}
\end{equation}
cf. \cite{Eb}. In such a way, we may eliminate the problems connected with the presence of physical boundary by considering
periodic functions defined on $\mathbb{T}^3$ and belonging, in addition, to the symmetry class (\ref{i6}). Note that
for $\vc{u}$ in the class (\ref{i6}), we have Korn--Poincar\' e inequality
\begin{equation} \label{i7}
\intO{ \mathbb{S} (\nabla_x \vc{u}) : \nabla_x \vc{u} } \geq c_{KP} \| \vc{u} \|^2_{W^{1,2}(\mathbb{T}^3; \mathbb{R}^3)}.
\end{equation}
\color{black}
In addition, we prescribe the total mass
\Cbox{Cgrey}{
\begin{equation} \label{i5}
\intO{ \varrho(t,x) } = M_0,\qquad t\in[0,\infty),
\end{equation}
where $M_0 > 0$ is {a deterministic constant.}
}
The assumption that $M_0$ is deterministic is taken for simplicity, in order to avoid unnecessary technicalities. A more general case of random $M_0$ satisfying
\begin{equation}\label{i5aa}
\underline m\leq M_0 \leq \overline m \qquad\mathbb{P}\text{-a.s.}
\end{equation}
for some deterministic constants $\underline m,\overline m\in (0,\infty)$ can also be considered. In that case, one would prescribe the law of $M_0$ such that \eqref{i5aa} holds.
\subsection{Stochastic setting}
\label{s}
{We consider a cylindrical $( \mathfrak{F}_t )_{t \geq 0}$-Wiener process defined on a stochastic basis
\[
\big( \Omega, \mathfrak{F}, ( \mathfrak{F}_t )_{t \geq 0} ,\prst \big),
\]
with a probability space $(\Omega,\mathfrak{F},\mathbb{P})$, and a right-continuous complete filtration
$( \mathfrak{F}_t )_{t \geq 0}$.
Formally, it is given by $W(t)=\sum_{k\geq1}e_k W_k(t) $ with $(W_k)_{k\in\mathbb{N}}$ being mutually independent real-valued standard Wiener processes relative to $(\mathscr{F}_t)_{t\geq0}$. Here $(e_k)_{k\in\mathbb{N}}$ denotes a complete orthonormal system in a sepa\-rable Hilbert space $\mathfrak{U}$ (e.g. $\mathfrak{U}=L^2(\mathbb{T}^3)$ would be a natural choice).}
The stochastic integral in \eqref{i2} is understood in the following sense
\[
\int \mathbb{G}(\varrho,\varrho\vc{u}) \,{\rm d} W =\sum_{k = 1}^\infty \int \mathbb{G}_k(x, \varrho, \varrho\vc{u})e_k \,{\rm d} W_k=: \sum_{k = 1}^\infty \int \vc{G}_k(x, \varrho, \varrho\vc{u}) \,{\rm d} W_k,
\]
where the one-dimensional summands on the right-hand side are standard It\^o-type stochastic integrals. In agreement with (\ref{i6}), we suppose that the functions $\vc{G}_k = \vc{G}_k(x, \rho, \vc{q})$ satisfy
\begin{equation} \label{i8}
\begin{split}
G^i_k (\cdot, -x_i, \cdot, - q^i, \cdot ) &= - G^i_k (\cdot, x_i, \cdot, q^i ,\cdot), \qquad i=1,2, 3,\\
G^i_k (\cdot, -x_j, \cdot, - q^j , \cdot) &= G^i_k (\cdot, x_j, \cdot, q^j, \cdot), \qquad i \ne j, \ i,j=1, 2, 3.
\end{split}
\end{equation}
\begin{Remark} \label{RC1}
The meaning of (\ref{i8}) is to keep {the spatially periodic solutions} in the symmetry class (\ref{i6}) as long as the initial data belong to
(\ref{i6}) $\mathbb{P}$-a.s.
\end{Remark}
Finally, we define the auxiliary space $\mathfrak{U}_0\supset\mathfrak{U}$ via
$$\mathfrak{U}_0=\bigg\{v=\sum_{k\geq1}\alpha_k e_k;\;\sum_{k\geq1}\frac{\alpha_k^2}{k^2}<\infty\bigg\},$$
endowed with the norm
$$\|v\|^2_{\mathfrak{U}_0}=\sum_{k\geq1}\frac{\alpha_k^2}{k^2},\qquad v=\sum_{k\geq1}\alpha_k e_k.$$
Note that the embedding $\mathfrak{U}\hookrightarrow\mathfrak{U}_0$ is Hilbert-Schmidt. Moreover, trajectories of $W$ are $\mathbb{P}$-a.s. in $C([0,T];\mathfrak{U}_0)$ (see \cite{daprato}). {For simplicity of the presentation, we often identify $\mathbb{G}(\varrho,\varrho\vc{u})$ as a Hilbert-Schmidt operator on $\mathfrak{U}$ with the sequence $\{\mathbf{G}_k(\varrho,\varrho\vc{u})\}_{k\in\mathbb{N}}$ as an element of $\ell^2$.}
\color{black}
\subsection{Main result}
We use the concept of weak martingale solution introduced in \cite{BrHo}.
In accordance with the available {\it a priori} bounds provided by the energy estimates, a suitable state space for
$[\varrho, \varrho \vc{u}]$ is taken
\[
\varrho \in L^\gamma(\mathbb{T}^3),\ \varrho \vc{u} \in L^{\frac{2 \gamma}{\gamma + 1}}(\mathbb{T}^3; \mathbb{R}^3),
\]
where $\gamma$ is the adiabatic exponent in the state equation (\ref{i3a}). Accordingly, we consider initial laws
$\Lambda$ defined on the Borel $\sigma$-algebra of the product space $L^\gamma (\mathbb{T}^3) \times L^{\frac{2 \gamma}{\gamma + 1}}(\mathbb{T}^3; \mathbb{R}^3)$.
\begin{Definition} \label{Dm1}
A quantity
\[
\left[ \big( \Omega, \mathfrak{F}, ( \mathfrak{F}_t )_{t \geq 0} ,\prst \big) ; \varrho, \vc{u}, W \right]
\]
is called a \emph{weak martingale solution} to problem (\ref{i1})--(\ref{i3}) {in $[0,T]$} with the initial law $\Lambda$ provided:
\begin{itemize}
\item
$\big( \Omega, \mathfrak{F}, ( \mathfrak{F}_t )_{t \geq 0} ,\prst \big)$ is a stochastic basis with a complete right-continuous filtration;
\item $W$ is an $( \mathfrak{F}_t )_{t \geq 0}$-cylindrical Wiener process;
\item the density $\varrho$ satisfies $\varrho \geq 0$, $t \mapsto \left< \varrho(t), \psi \right> \in C([0,T])$ for any
$\psi \in C^\infty(\mathbb{T}^3)$
$\mathbb{P}$-a.s., the function $t \mapsto \left< \varrho(t), \psi \right>$
{is $(\mathfrak{F}_t)$-adapted},
and
\begin{equation} \label{mom1}
\expe{ \sup_{t \in [0,T]} \| \varrho(t) \|^n_{L^\gamma(\mathbb{T}^3)} } < \infty \ \mbox{{for a certain}}\ n > 1;
\end{equation}
\item {the velocity field $\vc{u} \in L^2(\Omega \times (0,T); W^{1,2}(\mathbb{T}^3;\mathbb{R}^3))$ satisfies}
\begin{equation*} \label{mom2}
\expe{ \left( \int_0^T \| \vc{u}(t) \|^2_{W^{1,2}(\mathbb{T}^3; \mathbb{R}^3)} \ \,{\rm d} t \right)^n } < \infty\ \mbox{\textcolor{black}{for a certain}}\ n > 1;
\end{equation*}
\item the momentum $\varrho \vc{u}$ satisfies $t \mapsto \left< \varrho \vc{u}(t), \phi \right> \in C([0,T])$ for any $\phi \in C^\infty(\mathbb{T}^3;\mathbb{R}^3)$
$\mathbb{P}$-a.s., the function $t \mapsto \left< \varrho \vc{u}(t), \phi \right>$ \textcolor{black}{is $(\mathfrak{F}_t)$-adapted},
\begin{equation} \label{mom3}
\expe{ \sup_{t \in [0,T]} \left\| \varrho \vc{u}(t) \right\|^n_{L^{\frac{2 \gamma}{\gamma + 1}}{(\mathbb{T}^3;\mathbb{R}^3)} }} < \infty\ \mbox{\textcolor{black}{for a certain}}\ n > 1;
\end{equation}
\item $\Lambda=\mathbb{P}\circ \left( \varrho(0), \varrho \vc{u} (0) \right)^{-1} $,
\item $\mathbb{G}(\varrho,\varrho\vc{u})=\{\mathbf{G}_k(\varrho,\varrho\vc{u})\}_{k\in\mathbb{N}}\in L^2(\Omega\times (0,T),\mathcal{P},\mathrm{d}\mathbb{P}\otimes\,{\rm d} t ;\ell^2(W^{-b,2}(\mathbb{T}^3;\mathbb{R}^3)))$ for some $b>\frac{3}{2}$, where $\mathcal{P}$ denotes the progressively measurable $\sigma$-field associated to $(\mathfrak{F}_t)_{t\geq 0}$;
\item for all test functions $\psi \in C^\infty(\mathbb{T}^3)$, $\phi \in C^\infty(\mathbb{T}^3; \mathbb{R}^3)$ and all $t \in [0,T]$ it holds $\mathbb{P}$-a.s.
\begin{eqnarray}
\nonumber
{\rm d} \intO{ \varrho \psi } &=& \intO{ \varrho \vc{u} \cdot \nabla_x \psi } \,{\rm d} t ,
\\
\nonumber {\rm d} \intO{ \varrho \vc{u} \cdot \phi } &=& \intO{ \Big[ \varrho \vc{u} \otimes \vc{u} : \nabla_x \phi - \tn{S}(\nabla_x \vc{u}) : \nabla_x \phi
+ p(\varrho) {\rm div}_x \phi \Big] } \,{\rm d} t \\ \nonumber
&&\quad+ \sum_{k=1}^\infty\intO{ \mathbf{G}_k (\varrho, \varrho \vc{u}) \cdot \phi } \, {\rm d} W_k.
\end{eqnarray}
\end{itemize}
\end{Definition}
\begin{Remark} \label{mR1}
In addition to Definition \ref{Dm1}, we say that $[\varrho, \vc{u}]$ satisfy the complete slip boundary conditions (\ref{i4}), \textcolor{black}{if $[\varrho(t, \cdot), \varrho \vc{u}(t, \cdot)]$} belong
to the symmetry class (\ref{i6}) for any $t \in [0,T]$ $\mathbb{P}$-a.s.
\end{Remark}
\begin{Remark} \label{mR1bis}
\textcolor{black}{
Note that the statement about progressive measurability of the diffusion coefficients $\mathbb{G}(\varrho, \varrho \vc{u})$ is introduced for completeness, and,
as a matter of fact, can be deduced from the (weak) progressive measurability of $\varrho$ and $\varrho \vc{u}$, see \cite{BrHo}.}
\end{Remark}
\begin{Remark} \label{mR1bis+}
\textcolor{black}{
In contrast to the existence theory developed in \cite{BrHo}, the moments in (\ref{mom1})--(\ref{mom3}) are bounded up to a certain positive order $n$
rather then for \emph{all} $n > 1$ as in \cite{BrHo}. This is because the integrability of the moments for the initial--value problem is controlled by the initial data.
}
\end{Remark}
\begin{Remark} \label{mR2}
Similarly to
\cite{BrFeHo2015A}, \textcolor{black}{we consider the class of \emph{dissipative} martingale solutions satisfying}, in addition to the stipulations specified in Definition \ref{Dm1},
an energy inequality. \textcolor{black}{Indeed some form of energy balance will be used
at every step of the construction of the stationary solution.} As a result, the stationary solution we obtain is also a dissipative martingale solution
in the sense of \cite{BrFeHo2015A}. In addition, as in \cite{BrHo}, the equation of continuity \eqref{i1} is satisfied in the renormalized sense
\begin{eqnarray}
{\rm d}\intO{b(\varrho)\psi} &=& \intO{ b(\varrho)\vc{u} \cdot\nabla_x\psi }\,{\rm d} t - \intO{ \big(b'(\varrho)\varrho-b(\varrho) \big){\rm div}_x\vc{u} \,\psi }\,{\rm d} t
\end{eqnarray}
for every $\psi\in C^\infty(\mathbb{T}^3)$, and every $b\in C^1([0,\infty))$ with $b'(z)=0$ for $z\geq M_b$ for some constant $M_b>0$. This is an essential tool to pass to the limit in the nonlinear pressure.
\end{Remark}
Due to the \textcolor{black}{specific} structure of the Navier-Stokes system \eqref{i1}--\eqref{i3},
a concept of stationarity must chosen accordingly.
\textcolor{black}{We recall the standard definition of stationarity for \emph{continuous} processes ranging
in the Sobolev space $W^{k,p}$.}
\begin{Definition}\label{D2}
Let $k\in\mathbb{N}_0$, $p\in[1,\infty)$ and let $\mathbf{U}=\{\mathbf{U}(t);t\in[0,\infty)\}$ be an $W^{k,p}(\mathbb{T}^3)$-valued measurable stochastic process. We say that $\mathbf{U}$ is \emph{stationary} on $W^{k,p}(\mathbb{T}^3)$ provided the joint laws
$$\mathcal{L}(\mathbf{U}(t_1+\tau),\dots, \mathbf{U}(t_n+\tau)),\quad \mathcal{L}(\mathbf{U}(t_1),\dots, \mathbf{U}(t_n))$$
on $[W^{k,p}(\mathbb{T}^3)]^n$ coincide for all $\tau\geq0$, for all $t_1,\dots,t_n\in [0,\infty)$.
\end{Definition}
However, we observe that according to Definition \ref{Dm1}, the velocity $\vc{u}$ is not a stochastic process in the classical sense. Indeed, its trajectories belong to $L^2(0,T;W^{1,2}(\mathbb{T}^3;\mathbb{R}^3))$, i.e. are only defined almost everywhere in time. Therefore, even though the above definition of stationarity can be used for $[\varrho,\varrho\vc{u}]$, it is not suitable to describe stationarity of $\vc{u}$. To overcome this flaw, we consider solutions as random variables ranging in the
space $L^q_{\rm loc}([0, \infty);W^{k,p}(\mathbb{T}^3))$ as follows.
\begin{Definition}\label{D1}
Let $k\in\mathbb{N}_0$, $p,q\in[1,\infty)$ and let $\mathbf{U}$ be an $L^q_{\rm loc}([0,\infty);W^{k,p}(\mathbb{T}^3))$-valued random variable. Let $\mathcal S_\tau$ be the time shift on the space of trajectories given by $\mathcal{S}_\tau \mathbf{U}(t)=\mathbf{U}(t+\tau).$
We say that $\mathbf{U}$ is \emph{stationary} on $L^q_{\rm loc}([0,\infty);W^{k,p}(\mathbb{T}^3))$ provided the laws
$\mathcal{L}(\mathcal{S}_\tau\mathbf{U}),$ $ \mathcal{L}(\mathbf{U})$
on $L^q_{\rm loc}([0,\infty);W^{k,p}(\mathbb{T}^3))$ coincide for all $\tau\geq0$.
\end{Definition}
{Note that as Lemma \ref{lem:s} shows, it is actually sufficient to consider Definition \ref{D1} for $q=1$.}
As a matter of fact, the two concepts of stationarity {introduced
in Definition \ref{D2} and Definition \ref{D1}} are equivalent as soon as the stochastic process in question is
continuous in time; or alternatively, if it is weakly continuous and satisfies a suitable uniform bound. Proofs of these statements are provided in Lemma \ref{l:equivD12} and Corollary \ref{l:equivD123} below. Furthermore, it can be shown that both notions of stationarity are stable under weak convergence, see Lemma \ref{lem:stac} and Lemma \ref{lem:stac2}.
Motivated by Definition \ref{D1}, we adapt the concept
of stationarity introduced in the context of incompressible viscous fluids by Romito \cite{Romi}, cf. also the approach proposed by
It\^{o} and Nisio \cite{ItNi}.
\begin{Definition}\label{Dm2}
A weak martingale solution $[\varrho,\vc{u}, W]$ to \eqref{i1}--\eqref{i3} is called {\em stationary} provided
the joint law of the time shift
$
\left[\mathcal{S}_\tau\varrho, \mathcal{S}_\tau\vc{u}, \mathcal{S}_\tau W - W \right]$ on
$$ L^1_{\rm loc}([0, \infty); L^\gamma(\mathbb{T}^3)) \times L^1_{\rm loc}([0, \infty); W^{1,2}(\mathbb{T}^3; \mathbb{R}^3))
\times C([0, \infty); \mathfrak{U}_0 )
$$
is independent of $\tau \geq 0$.
\end{Definition}
\begin{Remark}
{In accordance with the previous discussion, if $[\varrho, \vc{u}, W]$ is a stationary martingale solution of the Navier--Stokes system
(\ref{i1})--(\ref{i3}) in the sense of Definition \ref{Dm2}, then the process $[\varrho,\varrho\mathbf u]$ is stationary on $L^\gamma(\mathbb{T}^3)\times L^\frac{2\gamma}{\gamma+1}(\mathbb{T}^3;\mathbb{R}^3)$ in the sense of Definition \ref{D2};} whereas for $\vc{u}$ we only have stationarity on $L^2_{\text{loc}}([0,\infty);W^{1,2}(\mathbb{T}^3;\mathbb{R}^3))$ in the sense of Definition \ref{D1}.
\end{Remark}
The following theorem is the main result of the present paper. For notational simplicity, we restrict ourselves to the most difficult
and {physically relevant} case of three space dimensions.
However, our result extends to the two- and mono-dimensional case as well, even under the weaker assumption {$\gamma>1$ and
$\gamma \geq 1$, respectively.}
\Cbox{Cgrey}{
\begin{Theorem} \label{Tm1}
Let $M_0 \in(0,\infty)$ be given.
Let $p = p(\varrho)$ be given by \eqref{i3a} with $\gamma > \frac{3}{2}$. Suppose that the
diffusion coefficients $\vc{G}_k$ belong to the symmetry class \eqref{i8} and there exist functions $\mathbf{g}_k\in C(\mathbb{T}^3 \times [0, \infty) \times \mathbb{R}^3; \mathbb{R}^3)$ and $\alpha_k\geq0$, $k\in\mathbb{N}$, such that
\begin{equation} \label{m1}
\begin{split}
\vc{G}_k(x, \rho, \vc{q}) &= \rho \vc{g}_k (x, \rho, \vc{q}),\\
\left|\nabla_{\rho, \vc{q}} \vc{g}_k (x, \rho, \vc{q}) \right| +
|\vc{g}_k (x, \rho, \vc{q})| &\leq \alpha_k, \qquad \sum_{k =1}^\infty \alpha^2_k = G < \infty.
\end{split}
\end{equation}
Then problem \eqref{i1}--\eqref{i3}, \eqref{i4}, \eqref{i5} admits a {stationary martingale solution $[\varrho, \vc{u}, W]$.}
\end{Theorem}
}
Note that if for instance $\mathbf{G}_k(x,\rho,0)=0$ for all $x\in\mathbb{T}^3$, $\rho\in [0,\infty)$ and $k\in\mathbb{N}$, then \eqref{i1}--\eqref{i3} admits a trivial stationary solution, namely, $\vc{u}\equiv 0$ and $\varrho\equiv \text{const}.$ Nevertheless, Theorem \ref{Tm1} applies to more general diffusion coefficients $\mathbf{G}_k$ where such trivial solutions do not exist.
\begin{Remark}
Let us briefly discuss the noise term in the equation (\ref{i2}). Technically, $W$ is a cylindrical Wiener process. However, note that our approach covers also the standard case of distributed (space-dependent) noise under very natural conditions.
More specifically, consider the equation (\ref{i2}) written formally as
$$ \frac{ (\varrho \vc{u})}{\,{\rm d} t } (t,x) + {\rm div}_x (\varrho \vc{u} \otimes \vc{u}) (t,x) + \nabla_x p(\varrho) (t,x) = {\rm div}_x \mathbb{S}(\nabla_x \vc{u}) (t,x) + \sigma (x, \varrho (t,x), \varrho \vc{u} (t,x)) \frac{{\rm d} W}{\,{\rm d} t }(t,x),$$
for $(t,x)\in (0,\infty) \times \mathbb{T}^3$, where the noise intensity $\sigma$ takes the form $\sigma (x,\rho, \vc{q})= \rho \sigma _1(x,\rho, \vc{q})$ and $\sigma _1 \in C^1(\mathbb{T}^3\times [0,\infty) \times \mathbb{R}^3; \mathbb{R}^3)$. Furthermore, $ \frac{{\rm d} W}{\,{\rm d} t }$ stands for a white in time, space-dependent noise, which is considered to be a formal derivative of an $L^2 (\mathbb{T}^3; \mathbb{R}^3)$-valued Wiener process. Denote by $\Sigma$ the (trace class) incremental covariance of $W$; obviously there exists an orthonormal basis $(f_k)_{k\in\mathbb{N}}$ in $L^2 (\mathbb{T}^3; \mathbb{R}^3)$ and a sequence $(\lambda _k)_{k\in\mathbb{N}},\ \lambda_k\ge 0,$ such that
$$ \Sigma f_k = \lambda _k f_k,\qquad W(t,x) = \sum_{k=1}^\infty \sqrt{\lambda _k}f_k(x)W_k(t),\qquad \sum_{k=1}^\infty \lambda _k <\infty,$$
where $(W_k)_{k\in\mathbb{N}}$ is a sequence of independent, standard scalar Wiener processes. In such case Definition~\ref{Dm1} yields a natural concept of rigorous solution to the system (\ref{i1})--(\ref{i3}) if we set
$$\vc{g}_k (x, \rho, \vc{q}) = \sigma _1(x, \rho, \vc{q}) f_k(x) \sqrt{\lambda _k},\qquad k\in \mathbb{N}.$$
If $\sigma_1$ is bounded, globally Lipschitz in $\rho, \ \vc{q}$, {uniformly in} $x$, $f_k \in C(\mathbb{T}^3;\mathbb{R}^3)$ and $f_k$, are bounded on $\mathbb{T}^3$, uniformly in $k\in \mathbb{N}$, then the noise term satisfies the condition (\ref{m1}), thereby Theorem \ref{Tm1} is applicable to the present case.
\end{Remark}
{The rest of the paper is devoted to the proof of Theorem \ref{Tm1}.}
\section{Basic finite-dimensional approximation}
\label{b}
In this section, we introduce the zero-level approximate system to \eqref{i1}-\eqref{i3} and study its long-time behavior for suitable initial data
{belonging to the symmetry class (\ref{i6}). More precisely, based on an energy estimate, Proposition \ref{prop:en}, and bounds for the density, Lemma \ref{lem:1608}, we apply the Krylov--Bogoliubov method to deduce the existence of an invariant measure.
{We point out that in accordance with hypothesis (\ref{i8}), the solutions
can be constructed to be
spatially periodic solutions, i.e. they belong to the symmetry class (\ref{i6}), as long as
the initial data belong to the same class (\ref{i6}).} We always tacitly assume this fact without specifying it explicitly in the future.
}
{Let}
\begin{align*}
\begin{aligned}
H_N&=\bigg\{\mathbf{w}=[\mathbf{w}_1,\mathbf{w}_2,\mathbf{w}_3]:\ \mathbf{w}_i=\sum_{|{\bf m}|\leq N} a_{\bf m}[\mathbf{w}_i] \exp\left( {\rm i} {{\bf m}} \cdot x \right) ,\
|\vc{m}| \leq N \bigg\}
\end{aligned}
\end{align*}
{be the space of trigonometric polynomials of order $N$, endowed with the Hilbert structure of the Lebesgue space $L^2(\mathbb{T}^3; \mathbb{R}^3)$, and let $\|\cdot\|_{H_N}$ denote the corresponding norm.}
Let
\[
\Pi_N : L^2(\mathbb{T}^3; \mathbb{R}^3) \to H_N
\]
be the associated $L^2$-orthogonal projection. Note that the following holds
\begin{align}\label{eq:project}
\|\Pi_N v\|_{L^p(\mathbb{T}^3;\mathbb{R}^3)}\leq\,c_p\|v\|_{L^p(\mathbb{T}^3;\mathbb{R}^3)}\qquad\forall v\in L^p(\mathbb{T}^3;\mathbb{R}^3),
\end{align}
and
$$\Pi_N v\to v\qquad\text{in}\qquad L^p(\mathbb{T}^3;\mathbb{R}^3),$$
for any $p\in(1,\infty)$, cf. \cite[Chapter 3]{Gra}.
\subsection{Approximate field equations}
Fix $R\in\mathbb{N}$, $N\in\mathbb{N}$, $\varepsilon > 0$, $\delta > 0$ and let {$\Gamma>\max\{\tfrac{9}{2},\gamma\}$}. The approximate solutions $\varrho = \varrho_N$, $\vc{u} = \vc{u}_N$, $\vc{u}_N(t) \in H_N$ for any $t$, are constructed to satisfy the following system of equations
\begin{equation} \label{b1}
\begin{split}
{\rm d} \varrho &+ {\rm div}_x (\varrho [\vc{u}]_R ) \,{\rm d} t = \varepsilon \Delta \varrho \,{\rm d} t - 2 \varepsilon \varrho \,{\rm d} t + H \left( \frac{1}{M_0} \intO{ \varrho } \right) \,{\rm d} t ,\\
{\rm d} \intO{ \varrho \vc{u} \cdot \varphi} &- \intO{ \varrho [\vc{u}]_R \otimes \vc{u} : \nabla_x \varphi } \,{\rm d} t
- \intO{ a \varrho^\gamma H ( \| \vc{u} \|_{H_N} - R ) {\rm div}_x \varphi } \,{\rm d} t \\
&= - \intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \varphi} \,{\rm d} t + \sum_{k=1}^\infty\intO{ \varrho\,\Pi_N\mathbf{g}_k(\varrho,\varrho\mathbf{u}) \cdot \varphi } \ {\rm d} W_k \\
&\hspace{-.5cm}+ \varepsilon \intO{ \varrho \vc{u} \cdot \Delta \varphi } \,{\rm d} t - 2 \varepsilon \intO{ \varrho \vc{u} \cdot \varphi } \ \,{\rm d} t + \delta \intO{ \varrho^\Gamma
H(\| u \|_{H_N} - R ){\rm div}_x \varphi } \,{\rm d} t ,
\end{split}
\end{equation}
for any test function $\varphi \in H_N$, where
\[
[u]_R = H \left( \|\vc{u} \|_{{H_N}} - R \right) \vc{u}
\]
with
\[
H \in C^\infty(\mathbb{R}), \quad H = \left\{ \begin{array}{l} 1 \ \mbox{on}\ (-\infty, 0],\\ \mbox{a decreasing function on}\
( 0 , 1), \\ 0 \ \mbox{on}\ [1, \infty). \end{array} \right.
\]
Note that the basic approximate system \eqref{b1} is not the same as the one from \cite{BrHo}, cf. \eqref{eq:approx}. To be more precise, in order to obtain global-in-time estimates we are forced to include two more ``stabilizing'' terms in the continuity equation and to modify the momentum equation accordingly. Nevertheless, similarly to \cite[Section 3]{BrHo}, it can be shown that problem
(\ref{b1}) admits a unique strong pathwise solution for any initial data $[\varrho_0,(\varrho\vc{u})_0]$ satifying, for some $\nu>0$,
\begin{equation} \label{b1a}
\begin{split}
&\varrho_0 \in C^{2 + \nu}(\mathbb T^3), \ 0 < \underline{\varrho} < \varrho_0 < \Ov{\varrho}, \
(\varrho \vc{u})_0 \in {C^2(\mathbb T^3; \mathbb{R}^3)}\ \mathbb{P}\mbox{-a.s.},\\
&\expe{ \left( \intO{ \left[ \frac{| (\varrho \vc{u})_0 |^2}{\varrho_0} + \frac{a}{\gamma- 1} \varrho_0^\gamma + \frac{\delta}{\Gamma - 1} \varrho_0^\Gamma \right] } \right)^n } \leq c(n)\ \ \text{for all } 1\leq n<\infty.
\end{split}
\end{equation}
where $\underline{\varrho}$, $\Ov{\varrho}$ are deterministic constants, and
where the associated initial value of $\vc{u}$ is uniquely determined by
\begin{equation*
\vc{u}_0 \in H_N, \ \intO{ \varrho_0 \vc{u}_0 \cdot \varphi } = \intO{ (\varrho \vc{u})_0 \cdot \varphi } \ \ \mbox{for all} \ \varphi \in H_N.
\end{equation*}
\subsection{Basic energy estimates}
The energy estimates established in \cite[Section 3]{BrHo} are not well-suited for the construction of stationary solutions. Indeed, the application of Gronwall's Lemma leads to an exponentially (in time) growing right hand side.
In this subsection we derive improved energy estimates which overcome this problem and hold true uniformly in $t$. However, it is important to note that at this stage of the proof, we are not able to obtain estimates independent of all the approximation parameters, namely, the following bounds blow up as $\varepsilon\to0$. The necessary uniform estimates for the passage to the limit in $\varepsilon$ will be derived directly for stationary solutions in Section \ref{d}.
\begin{Proposition}\label{prop:en}
Let $(\varrho,\mathbf{u})$ be a solution to \eqref{b1} starting from
\begin{equation} \label{b13}
\varrho_0 =1 ,\quad (\varrho\vc{u})_0=\vc{u}_0=0.
\end{equation}
Then the following bounds hold true.
\begin{equation} \label{b14b}
\expe{ \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] (\tau, \cdot) }
\right)^n } \leq
c\left( n,\varepsilon, G \right),\ n\in\mathbb{N},
\end{equation}
\begin{equation} \label{b14}
\frac{1}{T} \expe{ \int_0^T \left( \| \vc{u} \|^2_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)} + \frac{2a\varepsilon}{\gamma} |\nabla_x \varrho^{\gamma/2} |^2_{L^2(\mathbb T^3; \mathbb{R}^3)} +
\frac{2\delta \varepsilon}{\Gamma} |\nabla_x \varrho^{\Gamma/2} |^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \right) \,{\rm d} t } \leq c\left( \varepsilon, G \right).
\end{equation}
\end{Proposition}
\begin{proof}
Applying It\^{o}'s chain rule to \eqref{b1} we deduce the basic energy balance
\begin{equation} \label{b2}
\begin{split}
&{\rm d} \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] }
+ 2 \varepsilon \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a\gamma }{\gamma - 1} \varrho^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho^{\Gamma} \right] } \,{\rm d} t \\
&\quad+ \intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \vc{u} } \,{\rm d} t
+ \varepsilon \intO{ \varrho |\nabla_x \vc{u}|^2 } \,{\rm d} t + \varepsilon \intO{ \left( a \gamma \varrho^{\gamma - 2} + \delta \varrho^{\Gamma - 2} \right) |\nabla_x \varrho |^2 } \,{\rm d} t
\\
&\quad+
\varepsilon \intO{ \frac{1}{2} H\left( \frac{1}{M_0} \intO{ \varrho } \right) |\vc{u}|^2 }\,{\rm d} t
\\
& = \sum_{k=1}^\infty\intO{ \varrho\,\Pi_N\mathbf{g}_k(\varrho,\varrho\vc{u}) \cdot \vc{u} } \, {\rm d} W_k + \frac{1}{2} \sum_{k=1}^\infty\intO{ \frac{1}{\varrho} |\varrho \Pi_N\vc{g}_k(\varrho,\varrho\vc{u})|^2 } \,{\rm d} t
\\
&\quad+ H\left( \frac{1}{M_0} \intO{ \varrho } \right) \intO{ \left( \frac{a \gamma}{\gamma - 1} \varrho^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho^{\Gamma - 1}
\right) } \,{\rm d} t ,
\end{split}
\end{equation}
we refer the reader to \cite[Proposition 3.1]{BrHo} for details. In view of hypothesis (\ref{m1}) and the continuity of $\Pi_N$ \eqref{eq:project}, we have
\begin{align}\label{b4}
\begin{aligned}
\sum_{k=1}^\infty\intO{ \frac{1}{\varrho} |\varrho\Pi_N\vc{g}_k(\varrho,\varrho\vc{u})|^2 }&\leq c \|\varrho\|_{L^\gamma(\mathbb{T}^3)}\sum_{k=1}^\infty\|\mathbf{g}_k(\varrho,\varrho\vc{u})\|^2_{L^{2\gamma'}(\mathbb{T}^3;\mathbb{R}^3)}\\
& \leq c \|\varrho\|_{L^\gamma(\mathbb{T}^3)}\sum_{k=1}^\infty\|\mathbf{g}_k(\varrho,\varrho\vc{u})\|^2_{L^\infty(\mathbb{T}^3;\mathbb{R}^3)} \leq c(G)\|\varrho\|_{L^\gamma(\mathbb{T}^3)},
\end{aligned}
\end{align}
where $\tfrac{1}{\gamma}+\tfrac{1}{\gamma'}=1$.
Remark that the function $\hat{\varrho} = \intO{ \varrho }$ satisfies the (deterministic) ODE
\begin{equation} \label{b3}
\frac{{\rm d}}{\,{\rm d} t } \hat{\varrho} = - 2 \varepsilon \hat{\varrho} + H \left( \frac{\hat{\varrho}}{M_0} \right).
\end{equation}
In particular, the function $\hat{\varrho}$ is bounded by a constant depending solely on the initial mass $M_0$.
Taking expectation in (\ref{b2}) leads to
\begin{equation*
\begin{split}
&\frac{{\rm d}}{\,{\rm d} t } \expe{ \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } }
+ 2 \varepsilon \expe{ \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a \gamma}{\gamma-1} \varrho^\gamma + \frac{\delta \Gamma}{\Gamma-1} \varrho^\Gamma \right] } } \\
&\quad+
\expe{ \intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \vc{u} }
} + \varepsilon \expe{ \intO{ \left( a \gamma \varrho^{\gamma - 2} + \delta \Gamma \varrho^{\Gamma - 2} \right) |\nabla_x \varrho |^2 } }
\\
& \leq c\left( G \right)\mathbb{E}\|\varrho\|_{L^\gamma(\mathbb{T}^3)} + \expe{ H\left( \frac{1}{M_0} \intO{ \varrho } \right) \intO{ \left( \frac{a \gamma}{\gamma - 1} \varrho^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho^{\Gamma - 1}
\right) } }.
\end{split}
\end{equation*}
Now, we observe that both terms on the right hand side can be estimated by the weighted Young inequality and then absorbed in the second term on the left hand side. This readily implies \eqref{b14b} for $n=1$ with an $\varepsilon$-dependent constant on the right hand side that blows up as $\varepsilon\to0$.
In addition, keeping (\ref{b13}) in mind and applying the Korn--Poincar\' e inequality (\ref{i7}), we deduce the estimate for the ergodic averages \eqref{b14}.
As the next step, we apply the It\^o formula to (\ref{b2}) to obtain, for $n\in\mathbb{N}$,
\begin{align} \label{b4a}
\begin{aligned}
&{\rm d} \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^n
+ 2 \varepsilon n \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a\gamma }{\gamma - 1} \varrho^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^n \,{\rm d} t \\
&\quad+ n \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}
\intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \vc{u} } \,{\rm d} t \\
&\quad+ \varepsilon n \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}\intO{ \varrho |\nabla_x \vc{u}|^2 } \,{\rm d} t \\
&\quad+ \varepsilon n \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}\intO{ \left( a \gamma \varrho^{\gamma - 2} +\delta \varrho^{\Gamma - 2} \right) |\nabla_x \varrho |^2 } \,{\rm d} t
\\
&\quad+
\varepsilon n \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}\intO{ \frac{1}{2} H\left( \frac{1}{M_0} \intO{ \varrho } \right) |\vc{u}|^2 }\,{\rm d} t \\
& = n \sum_{k=1}^\infty\left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1} \intO{ \varrho\Pi_N\mathbf{g}_k(\varrho,\varrho\vc{u}) \cdot \vc{u} } \, {\rm d} W_k \\
&\quad+ \frac{n}{2}
\left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}
\sum_{k=1}^\infty\intO{ \frac{1}{\varrho} |\varrho\Pi_N\vc{g}_k(\varrho,\varrho\vc{u})|^2 } \,{\rm d} t
\\
&\quad+n \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}
\times \\
&\quad\quad\times H\left( \frac{1}{M_0} \intO{ \varrho } \right) \intO{ \left( \frac{a \gamma}{\gamma - 1} \varrho^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho^{\Gamma - 1}
\right) } \ \,{\rm d} t \\
&\quad+ \frac{n (n-1)}{2} \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-2}\sum_{k=1}^\infty\left( \intO{ \varrho\Pi_N\vc{g}_k(\varrho,\varrho\vc{u}) \cdot \vc{u} } \right)^2 \,{\rm d} t \\
&=:\mathcal K.
\end{aligned}
\end{align}
By virtue of (\ref{m1}) and the continuity of $\Pi_N$ \eqref{eq:project},
\begin{equation} \label{b4b}
\begin{split}
\sum_{k=1}^\infty\left( \intO{ \varrho\Pi_N\vc{g}_k(\varrho,\varrho\vc{u}) \cdot \vc{u} } \right)^2 &\leq \sum_{k=1}^\infty \left\| \sqrt{\varrho}\,\Pi_N\vc{g}_k(\varrho,\varrho\vc{u}) \right\|^2_{L^2(\mathbb T^3;\mathbb{R}^3)} \| \sqrt{\varrho} \vc{u} \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \\ &\leq
c\sum_{k=1}^\infty \|\varrho\|_{L^\gamma(\mathbb T^3)}\|\Pi_N\mathbf{g}_k(\varrho,\varrho\vc{u})\|^2_{L^{2\gamma'}(\mathbb T^3;\mathbb{R}^3)} \| \sqrt{\varrho} \vc{u} \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \\
&\leq
c\sum_{k=1}^\infty \|\varrho\|_{L^\gamma(\mathbb{T}^3)}\|\mathbf{g}_k(\varrho,\varrho\vc{u})\|^2_{L^{2\gamma'}(\mathbb T^3;\mathbb{R}^3)} \| \sqrt{\varrho} \vc{u} \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \\
&\leq
c(G)\|\varrho\|_{L^\gamma(\mathbb T^3)} \| \sqrt{\varrho} \vc{u} \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \\
&\leq
c(G)\|\varrho\|_{L^\gamma(\mathbb{T}^3)} \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] }.
\end{split}
\end{equation}
Therefore, passing to expectations, the right hand side of \eqref{b4a} may be estimated by
\begin{align*}
\mathbb{E}\mathcal K&\leq
n \left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}
\intO{ \left( \frac{a \gamma}{\gamma - 1} \varrho^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho^{\Gamma - 1}
\right) } \ \,{\rm d} t \\
&\quad+c\left( n, G \right)\mathbb{E}\left( \intO{ \left[ \frac{1}{2} \varrho |\vc{u}|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho^{\Gamma} \right] } \right)^{n-1}
\|\varrho\|_{L^\gamma(\mathbb{T}^3)}\,{\rm d} t .
\end{align*}
Now, after application of the weighted Young inequality, both these terms can be absorbed in the second term on the left hand side of \eqref{b4a}, yielding a constant that blows up as $\varepsilon\to0$.
Hence we may infer \eqref{b14b} for any solution of (\ref{b1}) starting from regular initial data (\ref{b1a}).
\end{proof}
\subsection{Regularity of the density}
Making use of the of the additional damping terms in the first equation in \eqref{b1}, we are able to show strong statements about the regularity of the solution depending on the parameters.
\begin{Lemma}\label{lem:1608}
Let $\mathbf{u}\in C([0,\infty);H_N)$.
Let $\varrho$ be a classical solution to
\begin{align}\label{eq:1508}
\partial_t\varrho + {\rm div}_x (\varrho [\vc{u}]_R ) = \varepsilon \Delta \varrho - 2 \varepsilon \varrho + H \left( \frac{1}{M_0} \intO{ \varrho } \right)
\end{align}
with $\varrho(0)\in C^{2+\nu}(\mathbb T^3)$ such that $\varrho(0)>0$ and $\int_{\mathbb T^3} \varrho(0)\,{\rm d} {x}\leq \overline m$.\begin{itemize}
\item[(a)] Then we have
\begin{equation} \label{b23}
\| \varrho (\tau, \cdot) \|_{W^{k,p} (\mathbb T^3)} \leq c(\overline m,k,p, N, R, \varepsilon)\quad \forall \tau\geq1
\end{equation}
for all $k\in\mathbb N$ and $p<\infty$.
\item[(b)]
There exists a (deterministic) constant $\underline{\varrho} = \underline{\varrho}(\overline m, N, R, \varepsilon) > 0$ such that
\begin{equation} \label{b21}
\varrho(\tau,\cdot) \geq \underline{\varrho} \quad \forall\tau\geq1.
\end{equation}
\end{itemize}
In particular, the constants are independent of $\mathbf{u}$.
\end{Lemma}
\begin{proof}
We start with equation (\ref{b3}) for the density averages that is independent of $\mathbf{u}$. Since (\ref{b3}) is a first order (deterministic) ODE an easy observation shows
\begin{align}\label{eq:1508b}
\hat{\varrho}(t)\rightarrow M_\varepsilon\quad \text{as}\quad t\rightarrow\infty,
\end{align}
where $M_\varepsilon>0$ is the unique solution to the equation $2\varepsilon M_\varepsilon=H\big(\frac{M_\varepsilon}{M_0}\big)$. The convergence above is uniform in the sense that for every $\k>0$ there is $T=T(\overline{m},\varepsilon,\k)$ deterministic such that
$|\hat{\varrho}(t)-M_\varepsilon|<\k$ for all $t\geq T$.
The next step is to show that $\varrho$ is uniformly bounded from below as claimed in (b).
Returning to the equation of continuity, we have
\[
\partial_t \varrho - \varepsilon \Delta \varrho + \nabla_x \varrho [\vc{u} ]_R = - (2 \varepsilon + {\rm div}_x [\vc{u}]_R ) \varrho + H\left( \frac{1}{M_0} \hat{\varrho} \right).
\]
Seeing that
\[
|{\rm div}_x [\vc{u}]_R | \leq D(R,N)
\]
for some constant $D(R,N)$,
we may use the comparison principle to deduce that
\[
\varrho (t, \cdot) \geq \underline{\varrho}(t),
\]
where $\underline{\varrho}$ solves the equation
\begin{equation}\label{b111}
\frac{{\rm d} \underline{\varrho}}{\,{\rm d} t } = - \underline{\varrho} ( 2 \varepsilon + D(R,N) ) + H\left( \frac{1}{M_0} \hat{\varrho}\right),
\quad 0 < \underline{\varrho}(0)\leq \inf_{\mathbb T^3} \varrho(0).
\end{equation}
In accordance with \eqref{eq:1508b} we have
\[
H\left( \frac{1}{M_0} \hat{\varrho}(t)\right) \to H\left( \frac{M_\varepsilon}{M_0} \right)=2\varepsilon M_\varepsilon>0 \quad \mbox{as}\quad t \to \infty.
\]
Since any solution to \eqref{b111} is asymptotically stabilized towards this equilibrium, we conclude that $\hat\varrho(t)>0$ for any $t>0$ and
\[
\underline{\varrho} (t) \to \frac{ H\left( \frac{M_\varepsilon}{M_0} \right) }{2 \varepsilon + D(R,N)} \quad \mbox{as}\quad t \to \infty
\]
and finally \eqref{b21} follows.
Now we are going to prove part (a). First, note that \eqref{eq:1508b}
implies
\begin{align}\label{eq:1508c}
\hat{\varrho}(t)=\|\varrho(t)\|_{L^1(\mathbb T^3)}\leq\,c(\overline{m}).
\end{align}
We apply maximal regularity theory (see e.g. \cite{HP}) to the equation \eqref{eq:1508} to obtain
\begin{align*}
\|&\partial_t\varrho\|_{L^2(T,T+1;W^{-2,q}(\mathbb T^3))}+\|\Delta\varrho\|_{L^2(T,T+1;W^{-2,q}(\mathbb T^3))}\\&\leq \,c\,\Big(\|\varrho(T)\|_{W^{-1,q}(\mathbb{T}^3)}+\|{\rm div}_x(\varrho[\mathbf{u}]_R)\|_{L^2(T,T+1;W^{-2,q}(\mathbb T^3))}+\Big\|H \Big( \tfrac{1}{M_0} \intO{ \varrho } \Big)\Big\|_{L^2(T,T+1;W^{-2,q}(\mathbb{T}^3))}\Big)
\end{align*}
where $q$ is chosen such $1<q<3/2$. Since $L^1(\mathbb T^3)\hookrightarrow W^{-1,q}$ and we have \eqref{eq:1508c}
\begin{align*}
\|&\partial_t\varrho\|_{L^2(T,T+1;W^{-2,q}(\mathbb T^3))}+\|\varrho\|_{L^2(T,T+1;L^{q}(\mathbb{T}^3))}\\&\leq \,c\,\Big(\|\varrho(T)\|_{W^{-1,q}(\mathbb T^3)}+\|\varrho\|_{L^2(T,T+1;W^{-1,q}(\mathbb T^3))}+1\Big)\\
&\leq \,c\,\Big(\|\varrho(T)\|_{L^1(\mathbb T^3)}+\|\varrho\|_{L^2(T,T+1;L^1(\mathbb T^3))}+1\Big)\\
&\leq\,c\big(\|\varrho\|_{L^\infty(T,T+1;L^1(\mathbb T^3))}+1\big)\leq\,c,
\end{align*}
where $c$ depends on $R$ and $\varepsilon$ but is independent of $T$. Consequently, there is $\tau=\tau(T)\in[T,T+1]$ such that
$\varrho(\tau)$ is bounded in $L^q(\mathbb T^3)$ independently of $T$. A similar argument as above shows
\begin{align*}
\|&\partial_t\varrho\|_{L^2(\tau,\tau+1;W^{-1,q}(\mathbb T^3))}+\|\varrho\|_{L^2(T,T+1;W^{1,q}(\mathbb{T}^3))}\\&\leq \,c\,\Big(\|\varrho(\tau)\|_{L^{q}(\mathbb T^3)}+\|\varrho\|_{L^2(T,T+1;L^{q}(\mathbb T^3))}+1\Big)\leq \,c.
\end{align*}
So we have
\begin{align*}
\varrho\in L^2(T,T+1;W^{1,q}(\mathbb T^3))
\end{align*}
with a bound independent of $T$. Now, we can bootstrap the argument to obtain the claim.
\end{proof}
\subsection{Approximate invariant measures}
\label{ssec:i}
With estimates (\ref{b14b}), (\ref{b14}), \eqref{b23} at hand, we are ready to apply the method of Krylov--Bogoliubov \cite[Section 3.1]{daPrZa} to construct an invariant measure for system (\ref{b1})
with fixed parameters $R$, $N$, $\varepsilon$, and $\delta$. For $\underline{r}>0$ we define the set
$$\mathcal R=\mathcal R_{\underline{r}}=\{(r,\mathbf{v})\in C^{2+\nu}(\mathbb T^3)\times H_N;\,\,\underline{r}^{-1}\leq r\leq \underline{r},\,\,\|\nabla r\|_{L^\infty(\mathbb{T}^3)}\leq\underline{r}\}.$$
It will be the state space for solutions to \eqref{b1}. By $C_b(\mathcal R)$ we denote the space of continuous bounded functions on $\mathcal R$.
First of all, we remark that the approximate system \eqref{b1} can be solved using the Banach fixed point theorem as in \cite[Section 3]{BrHo}.
In what follows, for an $\mathfrak{F}_s$-measurable $\mathcal{R}$-valued random variable $\eta$, we denote by $\mathbf{U}^\eta_{s,t}=(\varrho^\eta_{s,t},\vc{u}^\eta_{s,t})$ the solution of \eqref{b1} at time $t$ starting at time $s$ from the initial condition $\eta$. If $s=0$ then we write simply $\mathbf{U}^\eta_{t}$. We obtain the following result.
\begin{Theorem}
There is $\underline{r}$ large enough such that the following holds. Let $0\leq s<t$ be given.
Let $\eta$ be an $\mathfrak{F}_s$-measurable $\mathcal R$-valued initial condition. Then there exists $\mathbf{U}_s^\eta=(\varrho^\eta_{s},\mathbf{u}^\eta_s)\in L^2(\Omega;C([s,t];\mathcal R))$ which is the unique strong pathwise solution to \eqref{b1} starting from $\eta$ at time $s$. In addition, if $\eta_1$, $\eta_2$ are two such initial conditions then there is $\beta\in(0,2)$ such that
\begin{align}\label{eq:cont-dep}
\begin{aligned}
&\mathbb{E}\big\|\mathbf{U}^{\eta_1}_{s,t}-\mathbf{U}^{\eta_2}_{s,t}\big\|^2_{\mathcal R}\leq C(t-s,R,N,\varepsilon,\delta)\,\mathbb{E}\|\eta_1-\eta_2\|^\beta_{\mathcal R}.
\end{aligned}
\end{align}
\end{Theorem}
\begin{proof}
The existence of the unique strong pathwise solution was established in \cite[Section 3]{BrHo}. In addition, by means of Lemma \ref{lem:1608}, the solution belongs to $L^2(\Omega;C([s,t];\mathcal R))$ if we choose $\underline{r}$ large enough. Following \cite[Section 3]{BrHo} we obtain
\begin{align*}
\mathbb{E}\big\|\mathbf{u}^{\eta_1}_{s,t}-\mathbf{u}^{\eta_2}_{s,t}\big\|^2_{H_N}\leq \,\mathbb{E}\sup_{s\leq\sigma\leq t}\big\|\mathbf{u}^{\eta_1}_{s,\sigma}-\mathbf{u}^{\eta_2}_{s,\sigma}\big\|^2_{H_N}\leq C(t-s,R,N,\varepsilon,\delta)\,\mathbb{E}\|\eta_1-\eta_2\|^2_{\mathcal R}.
\end{align*}
Moreover, \cite[Lemma 2.2]{feireisl1} implies
\begin{align*}
&\sup_{s\leq\sigma\leq t}\big\|\varrho^{\eta_1}_{s,\sigma}-\varrho^{\eta_2}_{s,\sigma}\big\|_{W^{1,2}(\mathbb T^3)}\leq C(t-s,R,N,\varepsilon,\delta)\sup_{s\leq\sigma\leq t}\|\mathbf{u}^{\eta_1}_{s,\sigma}-\mathbf{u}^{\eta_2}_{s,\sigma}\|_{H_N}
\end{align*}
$\mathbb P$-a.s. and hence
\begin{align*}
\mathbb{E}\big\|\varrho^{\eta_1}_{s,t}-\varrho^{\eta_2}_{s,t}\big\|^\beta_{W^{1,2}(\mathbb T^3)}\leq C(t-s,R,N,\varepsilon,\delta)\,\mathbb{E}\|\eta_1-\eta_2\|^\beta_{\mathcal R}.
\end{align*}
for any $\beta>0$.
In order to obtain the final estimate we choose $l\in\mathbb N$ such that $W^{l,2}(\mathbb T^3)\hookrightarrow C^{2+\nu}(\mathbb T^3)$ and interpolate $W^{l,2}(\mathbb T^3)$ between $W^{l+1,2}(\mathbb T^3)$ and $W^{1,2}(\mathbb T^3)$. Using Lemma \ref{lem:1608} this implies
for some $\beta\in(0,2)$
\begin{align*}
\mathbb{E}\big\|\varrho^{\eta_1}_{s,t}-\varrho^{\eta_2}_{s,t}\big\|^2_{C^{2+\nu}(\mathbb T^3)}&\leq\,c\, \mathbb{E}\big\|\varrho^{\eta_1}_{s,t}-\varrho^{\eta_2}_{s,t}\big\|^2_{W^{l,2}(\mathbb T^3)}\\
&\leq\,c\, \mathbb{E}\big\|\varrho^{\eta_1}_{s,t}-\varrho^{\eta_2}_{s,t}\big\|^\beta_{W^{1,2}(\mathbb T^3)}\big\|\varrho^{\eta_1}_{s,t}-\varrho^{\eta_2}_{s,t}\big\|_{W^{l+1,2}(\mathbb T^3)}^{2-\beta}\\
&\leq C(t-s,R,N,\varepsilon,\delta)\,\mathbb{E}\|\eta_1-\eta_2\|^\beta_{\mathcal R}.
\end{align*}
\end{proof}
Let us now define the operators $P_t$ by
$$(P_t\varphi)(\eta):=\mathbb{E}\big[\varphi\big(\mathbf{U}^\eta_t\big)\big]\qquad \varphi\in C_b(\mathcal{R}).$$
\begin{Corollary}
The equation \eqref{b1} defines a Feller Markov process, that is, $P_t:C_b(\mathcal{R})\to C_b(\mathcal{R})$ and
\begin{equation}\label{eq:markov}
\mathbb{E} [\varphi (\mathbf{U}^\eta_{t + s}) | \mathfrak{F}_t ] = (P_s
\varphi) (\mathbf{U}^\eta_t) \qquad \forall \varphi \in C_b (\mathcal{R}), \hspace{1em} \forall
\eta \in H, \hspace{1em} \forall t, s > 0.
\end{equation}
Besides, the semigroup property $P_{t+s}=P_t\circ P_s$ holds true.
\end{Corollary}
\begin{proof}
The Feller property $P_t:C_b(\mathcal{R})\to C_b(\mathcal{R})$ is an immediate consequence of \eqref{eq:cont-dep} and the dominated convergence theorem.
In order to establish the Markov property \eqref{eq:markov}, we shall prove that
\[ \mathbb{E} [\varphi (\mathbf{U}^\eta_{t + s}) Z] =\mathbb{E} [(P_s \varphi) (\mathbf{U}^\eta_t)
Z] \qquad \forall Z \in \mathfrak{F}_t . \]
By uniqueness
\[ \mathbf{U}^\eta_{t + s} = \mathbf{U}^{\mathbf{U}^\eta_t}_{t, t + s} \qquad \mathbb{P}\text{-a.s.}. \]
It is therefore sufficient to show that
\[ \mathbb{E} [\varphi (\mathbf{U}^{\mathbf{V}}_{t, t + s}) Z] =\mathbb{E} [(P_s \varphi)
(\mathbf{V}) Z] \]
holds true for every $\mathfrak{F}_t$-measurable random variable $\mathbf{V}$. By
approximation (one uses dominated convergence and the fact that $\mathbf{V}_n
\rightarrow \mathbf{V}$ in $E$ implies $P_t \varphi (\mathbf{V}_n) \rightarrow P_t
\varphi (\mathbf{V})$ in $\mathbb{R}$ a.s.), it is enough to prove it for random
variables $\mathbf{V} = \sum_{i = 1}^k \mathbf{V}^i \mathbf{1}_{A^i}$ where $\mathbf{V}^i \in \mathcal{R}$ are deterministic and $(A^i)\subset \mathfrak{F}_t$ is a collection of disjoint sets such that $\cup_i A^i=\Omega$. Consequently, it is enough to prove it for every
deterministic $\mathbf{V}\in E$. Now, the random variable $\mathbf{U}^{\mathbf{V}}_{t, t + s}$ depends
only on the increments of the Brownian motion between $t$ and $t + s$, hence
it is independent of $\mathfrak{F}_t .$ Therefore
\[ \mathbb{E} [\varphi (\mathbf{U}^{\mathbf{V}}_{t, t + s}) Z] =\mathbb{E} [\varphi
(\mathbf{U}^{\mathbf{V}}_{t, t + s})] \mathbb{E} [Z] . \]
Since $\mathbf{U}^{\mathbf{V}}_{t, t + s}$ has the same law as $\mathbf{U}^{\mathbf{V}}_s$ by uniqueness,
we have
\[ \mathbb{E} [\varphi (\mathbf{U}^{\mathbf{V}}_{t, t + s}) Z] =\mathbb{E} [\varphi
(\mathbf{U}^{\mathbf{V}}_s)] \mathbb{E} [Z] = P_s \varphi (\mathbf{V}) \mathbb{E} [Z]
=\mathbb{E} [P_s \varphi (\mathbf{V}) Z] \]
and the proof of \eqref{eq:markov} is complete.
Taking expectation in \eqref{eq:markov} we get on the one hand
\[ \mathbb{E} [\mathbb{E} [\varphi (\mathbf{U}^\eta_{t + s}) | \mathfrak{F}_t
]] =\mathbb{E} [\varphi (\mathbf{U}^\eta_{t + s})] = (P_{t + s} \varphi) (\eta)
\]
and on the other hand
\[ \mathbb{E} [(P_s \varphi) (\mathbf{U}^\eta_t)] = (P_t (P_s \varphi)) (\eta) . \]
Thus the semigroup property follows.
\end{proof}
For an $\mathfrak{F}_0$-measurable random variable $\eta\in \mathcal{R}$, let $\mu_{t,\eta}$ denote the law of $\mathbf{U}_t^\eta$. If the law of $\eta$ is $\mu$ then it follows from the definition of the operator $P_t$ that $\mu_{t,\eta}=P_t^*\mu$. For the application of the Krylov--Bogoliubov method, we shall prove the following result.
\begin{Proposition}
Let the initial condition be given by \eqref{b13}, that is $\eta\equiv(1,0)\in\mathcal{R}$.
Then the set of laws
$$\left\{\frac{1}{T}\int_0^{T}\mu_{s,\eta}\,\mathrm{d} s;\;T>0\right\}$$
is tight on $\mathcal{R}$.
\end{Proposition}
\begin{proof}
Recall that $\mu_{s,\eta}$ are laws on the space $\mathcal{R}$. In particular, the second component is finite dimensional whereas the first one not.
Let $\mu_{s,\eta}^\varrho$ and $\mu_{s,\eta}^\mathbf{u}$ denote the marginals of $\mu_{s,\eta}$ corresponding respectively to the first and second component of the solution. That is, $\mu_{s,\eta}^\varrho$ is the law of $\varrho^\eta_s$ on $C^{2+\nu}(\mathbb{T}^3)$ and $\mu_{s,\eta}^\mathbf{u}$ is the law of $\mathbf{u}^\eta_s$ on $H_N$. It is then enough to establish tightness of both following sets separately:
\begin{equation}\label{tight}
\left\{\frac{1}{T}\int_0^{T}\mu^\mathbf{u}_{s,\eta}\,\mathrm{d} s;\;T>0\right\},\qquad \left\{\frac{1}{T}\int_0^{T}\mu^\varrho_{s,\eta}\,\mathrm{d} s;\;T>0\right\}.
\end{equation}
As a consequence of \eqref{b14} and the equivalence of norms on $H_N$ we have
\begin{align*}
\frac{1}{T}\mathbb{E}\bigg[\int_0^T\|\mathbf{u}^\eta_t\|_{H_N}^2\,{\rm d} t \bigg]\leq c(N,\varepsilon, G).
\end{align*}
Consequently, for compact sets
$$B_R:=\left\{\mathbf{v}\in H_N;\, \|\mathbf{v}\|_{H_N}\leq R\right\}\subset H_N$$
by means of Chebyshev inequality we obtain
\begin{align*}
\frac{1}{T}\int_0^T\mu_{s,\eta}^\mathbf{u}(B^c_R)\,\mathrm{d} s&=\frac{1}{T}\int_0^T\mathbb{P}(\|\mathbf{u}^\eta_s\|_{H_N}>R)\,\mathrm{d} s\leq \frac{1}{R^2}\frac{1}{T}\mathbb{E}\bigg[\int_0^T\|\mathbf{u}^\eta_t\|_{H_N}^2\,{\rm d} t \bigg],
\end{align*}
which in turn implies the tightness of the first set in \eqref{tight}.
In order to establish tightness in the second component, we define
$$B_R:=\left\{r\in W^{k,p}(\mathbb{T}^3);\,\|r\|_{W^{k,p}(\mathbb{T}^3)}\leq R\right\}.$$
For $p\in(1,\infty)$ and $k\in\mathbb{N}$ sufficiently large this is a compact set in $C^{2+\nu}(\mathbb{T}^3)$ hence making use of \eqref{b23} we have
\begin{align*}
\frac{1}{T}\int_0^T\mu_{s,\eta}^\varrho(B^c_R)\,\mathrm{d} s&=\frac{1}{T}\int_0^T\mathbb{P}(\|\varrho^\eta_s\|_{W^{k,p}}>R)\,\mathrm{d} s\leq \frac{1}{R}\sup_{t\geq 0}\mathbb{E}\|\varrho^\eta_t\|_{W^{k,p}}
\end{align*}
and the desired tightness follows.
\end{proof}
Finally, the Krylov--Bogoliubov theorem \cite[Theorem 3.1.1]{daPrZa} applies and yields the following.
\begin{Corollary}\label{cor:inv}
Fix $R,\,N\in\mathbb{N}$, $\varepsilon,\,\delta>0$. Then there exists an invariant measure $\mathcal{L}_{\varrho,\vc{u}}$ for the dynamics given by \eqref{b1}. In addition, as consequence of \eqref{eq:1508b},
$$\mathcal{L}_{\varrho,\vc{u}}[r\geq\underline\varrho]=1,\qquad \mathcal{L}_{\varrho,\vc{u}}\bigg[\int_{\mathbb T^3}r\,{\rm d} {x}=M_\varepsilon\bigg]=1.$$
\end{Corollary}
\section{First limit procedures: $R\to\infty$, $N\to\infty$}
\label{d}
The existence of an invariant measure for the zero level approximate problem \eqref{b1} implies the existence of a stationary solution $[\varrho_R,\mathbf{u}_R]$. Our ultimate goal is to perform successively the limits for $R \to \infty$, $N \to \infty$,
$\varepsilon \to 0$, and finally $\delta \to 0$. Even though this may look like a straightforward modification of the arguments used in the existence proof
\cite{BrHo}, there are several new aspects that must be handled. First of all, the uniform bounds used in the existence proof \cite{BrHo} are controlled by the initial data. This is not the case for the stationary solution for which the
``initial value'' is not {\it a priori} known and the necessary estimates must be deduced from the energy balance (\ref{b4a}) using the fact that the solution
possesses the same law at any time. Moreover, the estimates derived in the previous section, that is, Proposition \ref{prop:en} and Lemma \ref{lem:1608} do not hold independently of the approximation parameters $R,N,\varepsilon,\delta$ and are therefore not suitable for the limit procedure.
In addition, since the point-values of the density are not compact, the proof of the strong convergence of the approximate densities based on continuity of the effective viscous flux must be modified.
Let $[\varrho_{R}, \vc{u}_{R}]$ be a solution of the approximate problem (\ref{b1}) whose law at (every) time $t$ is given by the invariant measure
$\mathcal{L}_{\varrho_{R}, \vc{u}_{R}}$ constructed in Corollary \ref{cor:inv}. As the first step, we show a new uniform bound for $[\varrho_{R}, \vc{u}_{R}]$ that can be deduce from the energy balance (\ref{b4a}). Note that at this stage, the estimate still blows up as $\varepsilon\to0$.
\begin{Proposition}\label{prop:90}
Let $[\varrho_{R}, \vc{u}_{R}]$ be a stationary solution to \eqref{b1} given by the invariant measure from Corollary \ref{cor:inv}. Then we have for all $n\in\mathbb{N}$ and a.e. $t\in(0,\infty)$
\begin{equation} \label{d51x}
\begin{split}
& \expe{ \left( \intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a\gamma }{\gamma - 1} \varrho_{R}^\gamma+ \frac{\delta \Gamma }{\Gamma - 1} \varrho_{R}^{\Gamma} \right] } \right)^n } \leq c(n, G,\varepsilon),
\end{split}
\end{equation}
$$
\expe{\|\vc{u}_R\|_{W^{1,2}(\mathbb{T}^3)}^2}\leq c(G,\varepsilon).
$$
\end{Proposition}
\begin{proof}
After taking expectations in (\ref{b4a}), we observe that due to stationarity of $[\varrho_R,\vc{u}_R]$, the first term is constant in time, thus its time derivative vanishes. This is a consequence of Corollary \ref{cor:00}. By the same reasoning we may ultimately omit the time integrals in all the remaining expressions. Then we apply (\ref{b4}), (\ref{b4b}) to estimate the terms coming from the stochastic integral and obtain
\begin{equation*
\begin{split}
&\varepsilon \expe{ \left( \intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a\gamma }{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho_{R}^{\Gamma} \right] } \right)^n } \\
&\quad+ \expe{ \left( \intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a}{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{R}^{\Gamma} \right] } \right)^{n-1}
\intO{ \mathbb{S}(\nabla_x \vc{u}_{R}) : \nabla_x \vc{u}_{R} } } \\
&\quad+ \varepsilon \expe{ \left( \intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a}{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{R}^{\Gamma} \right] } \right)^{n-1}\intO{ \left( a \gamma \varrho_{R}^{\gamma - 2} +\delta \varrho_{R}^{\Gamma - 2} \right) |\nabla_x \varrho_{R} |^2 } }
\\
&\leq c(n,G) \left( \intO{ \left[ \frac{1}{2} \varrho_R |\vc{u}_R|^2 + \frac{a}{\gamma - 1} \varrho_R^\gamma + \frac{\delta}{\Gamma - 1} \varrho_R^{\Gamma} \right] } \right)^{n-1}
\intO{ \left( \frac{a \gamma}{\gamma - 1} \varrho_R^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho_R^{\Gamma - 1}
\right) } \\
&\quad+c\left( n, G \right)\mathbb{E}\left( \intO{ \left[ \frac{1}{2} \varrho_R |\vc{u}_R|^2 + \frac{a}{\gamma - 1} \varrho_R^\gamma + \frac{\delta}{\Gamma - 1} \varrho_R^{\Gamma} \right] } \right)^{n-1}
\|\varrho_R\|_{L^\gamma(\mathbb{T}^3)}.
\end{split}
\end{equation*}
The application of the weighted Young inequality allows to absorb both terms on the right hand side into the first term on the left hand side. The claim follows.
\end{proof}
\begin{Proposition}\label{prop:1608a}
Let $[\varrho_{R}, \vc{u}_{R}]$ be a stationary solution to \eqref{b1} given by the invariant measure from Corollary \ref{cor:inv}. Then we have for all $n\in\mathbb{N}$, a.e. $T\in(0,\infty)$ and $\tau>0$
\begin{equation} \label{b2xy}
\begin{split}
&\expe{\bigg(\sup_{t\in[T,T+\tau]}\intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a}{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{R}^{\Gamma} \right] }\bigg)^n}\\
&\quad
+ 2 \varepsilon\,\expe{ \bigg(\int_T^{T+\tau}\intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a\gamma }{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho_{R}^{\Gamma} \right] } \,{\rm d} t \bigg)^n}\\
&\quad+ \expe{\bigg(\int_T^{T+\tau} \| \vc{u}_{R}\|^2_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)} \,{\rm d} t \bigg)^n}
+ \varepsilon \,\expe{\bigg(\int_T^{T+\tau}\intO{ \left( a \gamma \varrho_{R}^{\gamma - 2} + \delta \varrho_{R}^{\Gamma - 2} \right) |\nabla_x \varrho_{R} |^2 } \,{\rm d} t \bigg)^n}
\\
& \leq c(n,G,\varepsilon,\tau),
\end{split}
\end{equation}
where the constant on the right hand side does not depend on $T$.
\end{Proposition}
\begin{proof}
Taking the $n$-th power and expectation in the energy balance \eqref{b2} and applying \eqref{b4}, \eqref{b4b} and the Korn--Poincar\'e inequality, we deduce
\begin{equation} \label{b2x}
\begin{split}
&\expe{\sup_{t\in[T,T+\tau]}\intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a}{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{R}^{\Gamma} \right] }}^n\\
&\quad
+ 2 \varepsilon\,\expe{ \int_T^{T+\tau}\intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a\gamma }{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho_{R}^{\Gamma} \right] } \,{\rm d} t }^n\\
&\quad+ \expe{\int_T^{T+\tau} \| \vc{u}_{R}\|_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)} \,{\rm d} t }^n
+ \varepsilon \,\expe{\int_T^{T+\tau}\intO{ \left( a \gamma \varrho_{R}^{\gamma - 2} + \delta \varrho_{R}^{\Gamma - 2} \right) |\nabla_x \varrho_{R} |^2 } \,{\rm d} t }^n
\\
& \leq c(n)\, \expe{\intO{ \left[ \frac{1}{2} \varrho_{R}(T) |\vc{u}_{R}(T)|^2 + \frac{a}{\gamma - 1} \varrho_{R}^\gamma(T) + \frac{\delta}{\Gamma - 1} \varrho_{R}^{\Gamma}(T) \right] } }^n\\
&\quad+ c(n,G)\, \expe{\int_T^{T+\tau}\|\varrho_{R}\|_{L^\gamma(\mathbb{T}^3)}\int_{\mathbb{T}^3}\varrho_{R}|\vc{u}_{R}|^2\,\mathrm{d} x\,\mathrm{d} t}^\frac{n}{2}+c(n,G)\,\expe{\int_T^{T+\tau}\|\varrho_{R}\|_{L^\gamma(\mathbb{T}^3)}\,{\rm d} t }^n
\\
&\quad+ c(n)\,\expe{\int_T^{T+\tau} \intO{ \left( \frac{a \gamma}{\gamma - 1} \varrho_{R}^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho_{R}^{\Gamma - 1}
\right) } \,{\rm d} t }^n.
\end{split}
\end{equation}
The first term on the right hand side can be estimated due to \eqref{d51x} by a constant $c(n,G,\varepsilon)$. The third term on the right hand side can be estimated by Young's inequality as follows
\begin{align}\label{b2xx}
\begin{aligned}
&\expe{\int_T^{T+\tau}\|\varrho_{R}\|_{L^\gamma(\mathbb{T}^3)}\,{\rm d} t }^n\\
&\qquad\leq \frac{\varepsilon}{2}\,\expe{\int_T^{T+\tau}\intO{ \left[ \frac{1}{2} \varrho_{R} |\vc{u}_{R}|^2 + \frac{a\gamma }{\gamma - 1} \varrho_{R}^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho_{R}^{\Gamma} \right] } \,{\rm d} t }^n +c(n,\varepsilon,\tau)
\end{aligned}
\end{align}
and then absorbed into the second term on the left hand side of \eqref{b2x}. A similar approach applies to the last term on the right hand side of \eqref{b2x}. For the remaining term we write
\begin{align*}
&\expe{\int_T^{T+\tau}\|\varrho\|_{L^\gamma(\mathbb{T}^3)}\int_{\mathbb{T}^3}\varrho_{R}|\vc{u}_{R}|^2\,\mathrm{d} x\,\mathrm{d} t}^\frac{n}{2}
\leq \expe{\sup_{t\in[T,T+\tau]}\intO{\frac{1}{2}\varrho_{R}|\vc{u}_{R}|^2 } \int_T^{T+\tau} \|\varrho_{R}\|_{L^\gamma(\mathbb{T}^3)}\,\mathrm{d} t }^\frac{n}{2}\\
&\leq \kappa\,\expe {\sup_{t\in[T,T+\tau]}\intO{\frac{1}{2}\varrho_R|\vc{u}_R|^2}}^n + c(\kappa)\,\expe{ \int_{T}^{T+\tau} \|\varrho_R\|_{L^\gamma(\mathbb{T}^3)}\,{\rm d} t }^n,
\end{align*}
where the last term can be again estimated as in \eqref{b2xx}. Choosing $\k$ sufficiently small yields the claim.
\end{proof}
In view of the uniform bounds provided by Proposition \ref{prop:1608a} , for fixed $\varepsilon,\delta > 0$, the asymptotic limits for $R \to \infty$ and $N \to \infty$ can be carried over exactly as for the initial value problem in \cite[Section 3, Section 4]{BrHo}.
In the limit, we obtain the following approximate system.
\begin{itemize}
\item {\bf Regularized equation of continuity.}
\begin{equation} \label{L1}
\begin{split}
\int_{0}^{\infty} &\intO{ \left[ \varrho \partial_t \varphi + \varrho \vc{u} \cdot \nabla_x \varphi \right] } \,{\rm d} t \\ &= \varepsilon \int_0^\infty \intO{ \left[\nabla_x \varrho \cdot \nabla_x \varphi
- 2 \varrho \varphi \right] }\,{\rm d} t - 2\varepsilon \int_0^\infty \intO{ M_{\varepsilon} \varphi } \,{\rm d} t
\end{split}
\end{equation}
for any $\varphi \in C^\infty_c((0, \infty) \times \mathbb{T}^3)$ $\mathbb{P}$-a.s.
\item {\bf Regularized momentum equation.}
\begin{equation} \label{L2}
\begin{split}
\int_0^\infty \partial_t \psi\intO{ \varrho \vc{u} \cdot \varphi} \,{\rm d} t &+ \int_0^\infty\psi \intO{ \varrho \vc{u} \otimes \vc{u} : \nabla_x \varphi } \,{\rm d} t
+ \int_0^\infty \psi \intO{ (a \varrho^\gamma+\delta\varrho^\beta) {\rm div}_x \varphi } \,{\rm d} t \\
&- \int_0^\infty\psi \intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \varphi} \,{\rm d} t - \varepsilon \int_0^\infty \psi \intO{ \varrho \vc{u} \cdot \Delta \varphi } \,{\rm d} t \\&- 2 \varepsilon \int_0^\infty \psi \intO{ \varrho \vc{u} \cdot \varphi } \ \,{\rm d} t
=- \sum_{k=1}^{\infty}\int_0^\infty \psi \intO{ \mathbf{G}_k \cdot \varphi } \ {\rm d} W_k
\end{split}
\end{equation}
for any $\psi \in C^\infty_c((0, \infty))$, $\varphi \in C^\infty(\mathbb{T}^3; \mathbb{R}^3)$ $\mathbb{P}$-a.s.
\end{itemize}
To summarize, we deduce the following.
\begin{Proposition}\label{prop:1608b}
Let $\varepsilon,\delta>0$ be given. Then there exists a stationary weak martingale solution $[\varrho_\varepsilon,\mathbf{u}_\varepsilon]$ to \eqref{L1}--\eqref{L2}.
In addition, for $n\in\mathbb{N}$ and every $\psi\in C^\infty_c((0,\infty))$, $\psi\geq 0$, the following generalized energy inequality holds true
\begin{align} \label{b2a}
\begin{aligned}
&-\int_0^\infty\partial_t\psi \left( \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^n\,{\rm d} t \\
&\quad+ 2 \varepsilon n \int_0^\infty\psi\left( \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a\gamma }{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^n \,{\rm d} t \\
&\quad+ n \int_0^\infty\psi\left( \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^{n-1}
\intO{ \mathbb{S}(\nabla_x \vc{u}_\varepsilon) : \nabla_x \vc{u}_\varepsilon } \,{\rm d} t \\
& \leq n\sum_{k=1}^\infty \int_0^\infty\psi\left( \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^{n-1} \intO{ \varrho_\varepsilon\,\mathbf{g}_k(\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon) \cdot \vc{u}_\varepsilon } \, {\rm d} W_k \\
&\quad+ \frac{n}{2} \int_0^\infty\psi
\left( \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^{n-1}
\sum_{k=1}^\infty\intO{ \frac{1}{\varrho_\varepsilon} |\varrho_\varepsilon\,\vc{g}_k(\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon)|^2 } \,{\rm d} t
\\
&\quad+n \int_0^\infty\psi\left( \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^{n-1}
\times \\
&\qquad\qquad\times H\left( \frac{M_\varepsilon}{M_0} \right) \intO{ \left[ \frac{a \gamma}{\gamma - 1} \varrho_\varepsilon^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho_\varepsilon^{\Gamma - 1}
\right] } \ \,{\rm d} t \\
&\quad+ \frac{n (n-1)}{2} \int_0^\infty\psi\left( \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^{n-2}\times\\
&\qquad\qquad\times\sum_{k=1}^\infty\left( \intO{ \varrho_\varepsilon\,\vc{g}_k(\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon) \cdot \vc{u}_\varepsilon } \right)^2 \,{\rm d} t .
\end{aligned}
\end{align}
\end{Proposition}
\begin{proof}
The proof follows the lines of \cite[Section 3, Section 4]{BrHo}. The first passage to the limit as $R\to\infty$ relies on a stopping time argument from \cite[Subsection 3.1]{BrHo} whereas the limit $N\to\infty$ employs the stochastic compactness method based on the Jakubowski-Skorokhod representation theorem \cite[Theorem 2]{jakubow} as in \cite[Section 4]{BrHo}. We point out that all the necessary estimates in \cite[Section 3, Section 4]{BrHo} come from the energy balance, which is controlled by the initial condition. In the present construction, the bound for the initial energy is replaced by the estimate \eqref{d51x} which holds true due to stationarity. Apart from that, the only difference to \cite{BrHo} is that we have to deal with path spaces containing an unbounded time interval, that is
\begin{align*}
L^q_{\text{loc}}([0,\infty);X),\quad (L^q_{\text{loc}}([0,\infty);X),w),\quad C_{\text{loc}}([0,\infty);(X,w)),
\end{align*}
where $q\in(1,\infty)$ and $X$ is a reflexive separable Banach space. Recall that $L^q_{\text{loc}}([0,\infty);X)$ is a separable metric space given by
\begin{align*}
(f,g)\mapsto \sum_{M\in\mathbb{N}}2^{-M}\big(\|f-g\|_{L^q(0,M;X)}\wedge 1\big),
\end{align*}
and a set $\mathcal{K}\subset L^q_{\text{loc}}([0,\infty);X)$ is compact provided the set
$$\mathcal{K}_M:=\{f|_{[0,M]};\,f\in\mathcal{K}\}\subset L^q(0,M;X)$$
is compact for every $M\in\mathbb{N}$. On the other hand, the remaining two spaces are (generally) nonmetrizable locally convex topological vector spaces, generated by the seminorms
$$
f\mapsto \int_0^M \langle f(t), g(t) \rangle_X\,{\rm d} t ,\qquad M\in\mathbb{N},\,g\in L^{q'}(0,\infty; X^*),\,\tfrac{1}{q}+\tfrac{1}{q'}=1,
$$
and
$$
f\mapsto \sup_{t\in[0,M]}\langle f(t), g \rangle_X,\qquad M\in\mathbb{N},\,g\in X^*,
$$
respectively. As above, a set $\mathcal{K}$ is compact provided it's restriction to each interval $[0,M]$ is compact in $(L^q(0,M;X),w)$ and $C([0,M];(X,w))$, respectively.
Furthermore, it can be seen that these topological spaces belong to the class of the so-called quasi-Polish spaces, where the Jakubowski-Skorokhod theorem \cite[Theorem 2]{jakubow} applies. Indeed, in these spaces there exists a countable family of continuous functions that separate points. The proof of tightness of the corresponding laws in the current setting is therefore reduced to exactly the same method as in \cite[Section 4]{BrHo}. Note that the key estimate was \cite[(4.1)]{BrHo}, which is replaced by \eqref{b2xy}. Consequently, the passage to the limit follows the lines of \cite[Section 4]{BrHo}. In addition, Lemma \ref{lem:stac} and Lemma \ref{lem:stac2} show that this limit procedure preserves stationarity defined in Definition \ref{D2} and Definition \ref{D1}. Hence the limit solution is stationary.
Finally, we obtain \eqref{b2a} by passing to the limit in \eqref{b4a}. The passage to the limit in the stochastic integral can be justified for instance with help of \cite[Lemma 2.1]{debussche1}.
\end{proof}
\begin{Remark}
Note that for $n=1$ the generalized energy inequality \eqref{b2a} corresponds to the usual energy inequality established in \cite{BrFeHo2015A}. The higher order version for $n\in\mathbb{N}$ is new and employed in order to obtain an analog of Proposition \ref{prop:1608a} suitable for the subsequent limit procedures $\varepsilon\to0$ and $\d\to0$ in Section \ref{L} and Section \ref{P}.
\end{Remark}
\section{Vanishing viscosity limit}
\label{L}
Our goal in this section is to perform the passage to the limit as $\varepsilon\to0$. This represents the most critical and delicate part of our construction. Remark that after completing this limit procedure we have already proved existence of stationary solutions to the stochastic Navier-Stokes system for compressible fluids -- under an additional assumption upon the adiabatic exponent $\gamma$. The last passage to the limit presented in Section \ref{P} is then devoted to weakening this additional assumption.
We point out that the key results needed for the previous limit procedure in Section \ref{d}, namely, Proposition \ref{prop:90} and consequently Proposition \ref{prop:1608a}, depend on $\varepsilon$. Furthermore, it turns out that the global in time energy estimates uniform in $\varepsilon$ and $\d$ are very delicate. On the contrary, in the existence proof in \cite{BrHo}, the basic energy estimate \cite[Proposition 3.1]{BrHo} established on the first approximation level held true uniformly in all the approximation parameters. Consequently, no further manipulations with the energy inequality were needed.
This brought significant technical simplifications in comparison to the present construction of stationary solutions. To be more precise,
due to the fact that already after the passage to the limit $N\to\infty$, the energy balance is violated and has to be replaced by an inequality. In other words, one cannot justify the application of It\^o's formula anymore and it is necessary to establish a more general version of the energy inequality, cf. \eqref{b2a}.
Recall from \cite[Section 5]{BrHo}, that in addition to the usual energy estimate \cite[(5.2)]{BrHo}, a higher integrability of the density \cite[(5.9)]{BrHo} was necessary in order to justify the compactness argument. Nevertheless, as in the deterministic setting it was not possible to obtain strong convergence of the approximate densities directly. Consequently, the identification of the limit proceeded in two steps. First, the passage to the limit in the approximate system was done but the expressions with nonlinear dependence on the density could not be identified. Second, a stochastic analog of the effective viscous flux method originally due to Lions \cite{LI4} allowed to prove strong convergence of the densities and hence to complete the proof.
Let us begin with an estimate for the velocity.
\begin{Proposition}\label{prop:1608}
Let $[\varrho_{\varepsilon}, \vc{u}_{\varepsilon}]$ be the stationary solution to \eqref{L1}--\eqref{L2} constructed in Proposition \ref{prop:1608b}. Then for a.e. $t\in(0,\infty)$
\begin{equation} \label{d1}
\expe{
\| \vc{u}_\varepsilon \|^2_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)} } \leq c(G,M_0),
\end{equation}
\begin{equation} \label{d12a}
\begin{split}
&\expe{\intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] }
\left\| \vc{u}_\varepsilon \right\|^2_{W^{1,2} (\mathbb T^3; \mathbb{R}^3)} }\\
&\qquad\leq c(G,M_0)\, \expe{ \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } }+c(M_0).
\end{split}
\end{equation}
\end{Proposition}
\begin{proof}
Taking expectation in the energy inequality \eqref{b2a}, we observe that due to stationarity of $[\varrho_{\varepsilon}, \vc{u}_{\varepsilon}]$, the first term is constant in time, thus its time derivative vanishes. We recall that $M_\varepsilon\leq c(M_0)$ and using \eqref{m1} we estimate
\begin{align}\label{b4n}
\begin{aligned}
\sum_{k=1}^\infty\intO{ \frac{1}{\varrho_\varepsilon} |\varrho_\varepsilon\vc{g}_k(\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon)|^2 }&\leq c(G) \int_{\mathbb{T}^3}\varrho_\varepsilon \,{\rm d} {x}\leq c(G,M_0),
\end{aligned}
\end{align}
and
\begin{equation}\label{b4bc}
\begin{split}
\sum_{k=1}^\infty\left( \intO{ \varrho_\varepsilon\vc{g}_k(\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon) \cdot \vc{u}_\varepsilon } \right)^2 &\leq \sum_{k=1}^\infty \left\| \sqrt{\varrho_\varepsilon}\vc{g}_k(\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon) \right\|^2_{L^2(\mathbb T^3;\mathbb{R}^3)} \| \sqrt{\varrho_\varepsilon} \vc{u}_\varepsilon \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \\
&\leq
c(G,M_0)\intO{ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2}.
\end{split}
\end{equation}
which leads to
\begin{equation} \label{d51}
\begin{split}
& 2 \varepsilon \,\mathbb{E}\bigg[\bigg(\intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a\gamma }{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] }\bigg)^n\bigg]\\
&\quad + \mathbb{E}\bigg[\left( \intO{ \left[ \frac{1}{2} \varrho_{\varepsilon} |\vc{u}_{\varepsilon}|^2 + \frac{a}{\gamma - 1} \varrho_{\varepsilon}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\varepsilon}^{\Gamma} \right] } \right)^{n-1}\intO{ \mathbb{S}(\nabla_x \vc{u}_\varepsilon) : \nabla_x \vc{u}_\varepsilon } \bigg]\\
& \leq c(G,M_0)\,\mathbb{E}\bigg[\left( \intO{ \left[ \frac{1}{2} \varrho_{\varepsilon} |\vc{u}_{\varepsilon}|^2 + \frac{a}{\gamma - 1} \varrho_{\varepsilon}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\varepsilon}^{\Gamma} \right] } \right)^{n-1}\bigg]\\
&\quad+ \mathbb{E}\bigg[\left( \intO{ \left[ \frac{1}{2} \varrho_{\varepsilon} |\vc{u}_{\varepsilon}|^2 + \frac{a}{\gamma - 1} \varrho_{\varepsilon}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\varepsilon}^{\Gamma} \right] } \right)^{n-1}\\
&\qquad\qquad\times H\left( \frac{1}{M_0} \intO{ \varrho_\varepsilon } \right) \intO{ \left( \frac{a \gamma}{\gamma - 1} \varrho_\varepsilon^{\gamma - 1} + \frac{\delta \Gamma}{\Gamma - 1} \varrho_\varepsilon^{\Gamma - 1}
\right) } \bigg].
\end{split}
\end{equation}
Moreover, it follows as a consequence of Corollary \ref{cor:inv} that
\[
H \left( \frac{1}{M_0} \intO{ \varrho_\varepsilon } \right) = H \left( \frac{M_{\varepsilon}}{M_0} \right) = 2 \varepsilon M_{\varepsilon}.
\]
Hence, setting $n=1$ and applying the Korn--Poincar\'e inequality yields \eqref{d1}, since the second term on the right hand side in \eqref{d51} can be absorbed in the first term on the left hand side. Setting $n=2$ in \eqref{d51} we deduce \eqref{d12a}.
\end{proof}
We point out that the corresponding bound for the energy which can be obtained from \eqref{d51}, i.e.
\begin{align*}
\varepsilon \expe{ \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_{\varepsilon}|^2 + \frac{a\gamma }{\gamma - 1} \varrho_{\varepsilon}^\gamma + \frac{\delta \Gamma }{\Gamma - 1} \varrho_{\varepsilon}^{\Gamma} \right] } }\leq c(G,M_0)
\end{align*}
still depends on $\varepsilon$ and is therefore not suitable for the passage to the limit $\varepsilon\to 0$. As the next step, we derive an improved estimate for the energy as well as for the pressure.
\begin{Proposition}\label{prop:1708}
Let $[\varrho_\varepsilon,\vc{u}_\varepsilon]$ be the stationary solution to \eqref{L1}--\eqref{L2} constructed in Proposition \ref{prop:1608b}. Then the following uniform bound holds true for a.e. $t\in(0,\infty)$
\begin{equation} \label{d131}
\expe{ \intO{ \left[ a \varrho_\varepsilon^{\gamma +1} + \delta \varrho_\varepsilon^{\Gamma + 1} +
\frac{1}{3}\varrho_\varepsilon^{2} |\vc{u}_\varepsilon|^2 \right]} } \leq c(\delta,G,M_0).
\end{equation}
In addition, if $s\in \big(1,\frac{\Gamma+1}{\Gamma-1}\wedge\frac{2(\gamma+1)}{\gamma+2}\big]$
then for a.e. $T > 0$ and $\tau>0$
\begin{equation} \label{d151}
\begin{split}
&\expe{ \left( \sup_{t \in [T, T + \tau] } \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^s }\\
&\qquad +\expe{\bigg(\int_T^{T+\tau}\|\vc{u}_\varepsilon\|_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)}^2\,{\rm d} t \bigg)^s}\leq c(\tau,\delta, M_0, G, s),
\end{split}
\end{equation}
where the constant is independent of $T$.
\end{Proposition}
\begin{proof}
Our goal is to use
the quantity
$\nabla_x\Delta_x^{-1} \left[ \varrho_\varepsilon - M_\varepsilon \right]$
as test functions in the momentum equation,
where $\Delta_x$ is the periodic Laplacian. We apply It\^{o}'s formula to the functional $f(\rho,\mathbf{q})=\int_{\mathbb{T}^3} \mathbf{q}\cdot\Delta_x^{-1}\nabla_x\rho\,{\rm d} {x}.$
This step can be made rigorous by mollifying the equation, see \cite[Section 5]{BrHo}.
After a rather tedious but straightforward manipulations,
we deduce from (\ref{L1}) and \eqref{L2} that
\begin{equation} \label{d41}
\begin{split}
\int_T^{T+1} &\intO{ \left( a \varrho_\varepsilon^{\gamma + 1} + \delta \varrho_\varepsilon^{\Gamma + 1} \right) } \ \,{\rm d} t + \int_T^{T+1} \intO{ \frac{1}{3} \varrho_\varepsilon^{2} |\vc{u}_\varepsilon|^2 } \ \,{\rm d} t \\
&=
M_\varepsilon \int_T^{T+1} \intO{ \left( a \varrho_\varepsilon^\gamma +
\delta \varrho_\varepsilon^\Gamma \right) } \,{\rm d} t + \frac{1}{3} M_\varepsilon \int_T^{T+1} \intO{ \varrho_\varepsilon |\vc{u}_\varepsilon|^2 } \,{\rm d} t \\
&\qquad+
\int_T^{T+1} \intO{ \left(\frac{4}{3} \mu+\eta \right) {\rm div}_x \vc{u}_\varepsilon \ \varrho_\varepsilon } \,{\rm d} t \\
&\qquad+ 2\varepsilon \int_T^{T+1} \intO{ \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta_x^{-1} \left[
\varrho_\varepsilon -M_\varepsilon \right] } \,{\rm d} t +\varepsilon \int_T^{T+1} \intO{ \vr_\varepsilon^2 {\rm div}_x \vc{u}_\varepsilon } \,{\rm d} t \\
&\qquad - \int_T^{T+1} \intO{ \left( \varrho_\varepsilon \vc{u}_\varepsilon \otimes \vc{u}_\varepsilon - \frac{1}{3} \vr_\varepsilon |\vc{u}_\varepsilon |^2 \mathbb{I} \right) : \nabla_x \Delta^{-1}_x
\nabla_x \varrho_\varepsilon } \,{\rm d} t \\
&\qquad+
\left[ \intO{ \varrho_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1}_x \left[ \varrho_\varepsilon - M_\varepsilon \right] } \right]_{t = T}^{t = T+1} \\
&\qquad+ \int_T^{T+1} \intO{ \varrho_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta_x^{-1} \left[ {\rm div}_x (\varrho_\varepsilon \vc{u}_\varepsilon ) \right] } \,{\rm d} t \\
&\qquad- \sum_{k=1}^\infty\int_T^{T+1} \intO{ \mathbf{G}_k(\vr_\varepsilon,\vr_\varepsilon\vc{u}_\varepsilon) \cdot \nabla_x \Delta_x^{-1} \left[ \varrho_\varepsilon - M_\varepsilon\right] } \ {\rm d} W_k.
\end{split}
\end{equation}
Note that in the above the second term on the left hand side, the second term on the right hand side and the second summand on the fifth line were added artificially and they cancel out. Passing to expectations in (\ref{d41}) and keeping in mind that the processes are stationary we deduce
\begin{equation*
\begin{split}
&\expe{ \intO{ \left[ a \varrho_\varepsilon^{\gamma +1} + \delta \varrho_\varepsilon^{\Gamma + 1} +
\frac{1}{3}\varrho_\varepsilon^{2} |\vc{u}_\varepsilon|^2 \right]} }\\
& \leq
c(M_0) \,\expe{ \intO{ \Big( \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^\Gamma \Big) } } \\
&\quad - \expe{ \intO{ \Big( \varrho_\varepsilon \vc{u}_\varepsilon \otimes \vc{u}_\varepsilon - \frac{1}{3} \varrho_\varepsilon |\vc{u}_\varepsilon |^2 \mathbb{I} \Big) : \nabla_x \Delta^{-1}_x \nabla_x \varrho_\varepsilon } } \\
&\quad+
\expe{ \intO{ \Big(\frac{4}{3}\mu+\eta \Big) {\rm div}_x \vc{u}_\varepsilon \ \varrho_\varepsilon }\,{\rm d} t }+ \expe{ \int_T^{T+1} \intO{ \varrho_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1}_x {\rm div}_x (\varrho_\varepsilon \vc{u}_\varepsilon ) }}\\
&\quad+ 2 \varepsilon \,\expe{ \intO{ \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta_x^{-1} \left[
\varrho_\varepsilon -M_\varepsilon \right] } } +\varepsilon\,\expe{ \intO{ \vr_\varepsilon^2 {\rm div}_x\vc{u}_\varepsilon } }\\
&=:(I)+(II)+(III)+(IV)+(V)+(VI).
\end{split}
\end{equation*}
Now, we estimate each term separately. By Young's inequality we obtain for every $\kappa>0$
\begin{align*}
(I)&\leq\,\kappa\, \expe{ \intO{ \Big( \frac{1}{3} \varrho_\varepsilon^2 |\vc{u}_\varepsilon|^2 + a\varrho_\varepsilon^{\gamma+1} + \delta\varrho_\varepsilon^{\Gamma+1} \Big) } }
+c(\kappa,M_0)\, \expe{ \int_{\mathbb T^3}\Big( |\vc{u}_\varepsilon|^2 + 1 \Big) \,{\rm d} {x} }\\
&\leq\,\kappa\, \expe{ \intO{ \Big( \frac{1}{3} \varrho_\varepsilon^2 |\vc{u}_\varepsilon|^2 + a\varrho_\varepsilon^{\gamma+1} + \delta\varrho_\varepsilon^{\Gamma+1} \Big) } }+c(\kappa,G,M_0),
\end{align*}
using the uniform bound (\ref{d1}).
In order to control the remaining integrals on the right hand side, we first use H\" older's inequality to obtain
\begin{equation*
\begin{split}
(II)
&\leq \,c\, \expe{\| \sqrt{\varrho_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_\varepsilon \|_{L^6(\mathbb T^3; \mathbb{R}^3)} \left\| \sqrt{\varrho_\varepsilon} \,\nabla_x \Delta_x^{-1} \nabla_x \varrho_\varepsilon \right\|_{L^3(\mathbb T^3; \mathbb{R}^{3\times 3})}} \\
&\leq c
\left( \expe{\| \sqrt{\varrho_\varepsilon} \vc{u}_\varepsilon \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_\varepsilon \|^2_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)}} + \expe{\left\| \sqrt{\varrho_\varepsilon}\, \nabla_x \Delta^{-1}_x \nabla_x \varrho_\varepsilon \right\|^2_{L^{3 }(\mathbb T^3; \mathbb{R}^{3\times 3})}} \right).
\end{split}
\end{equation*}
Furthermore, since $\Gamma\geq9/2$, we have
\begin{align*}
&\expe{\left\| {\varrho_\varepsilon}^{1/2} \nabla_x \Delta^{-1}_x \nabla_x\varrho_\varepsilon \right\|^2_{L^3(\mathbb T^3; \mathbb{R}^{3 \times 3})}}
\leq \expe{\left\| \varrho_\varepsilon \right\|_{L^{\frac{9}{2}}(\mathbb T^3)} \left\| \nabla_x \Delta^{-1}_x \nabla_x \varrho_\varepsilon \right\|^2_{L^\frac{9}{2}(\mathbb T^3; \mathbb{R}^{3 \times 3})}}\\
&\leq\,c\,\expe{\left\| \varrho_\varepsilon \right\|^3_{L^{\frac{9}{2}}(\mathbb T^3)}}\,{\rm d} t \leq\,c\,\mathbb E \bigg[\left\| \varrho_\varepsilon \right\|^3_{L^{\Gamma}(\mathbb T^3)}\bigg]\leq\,\kappa\delta\expe{\left\| \varrho_\varepsilon \right\|^\Gamma_{L^{\Gamma}(\mathbb T^3)}}+c(\kappa,\delta).
\end{align*}
Note that we also used the continuity of $\nabla_x \Delta^{-1}_x \nabla_x$
and Young's inequality.
Similarly, we can estimate
\begin{align*}
(IV) &\leq\expe{
\|\vc{u}_\varepsilon \|_{L^6(\mathbb T^3; \mathbb{R}^3)} \| \varrho_\varepsilon \|_{L^{3}(\mathbb T^3)} \left\| \nabla_x \Delta^{-1}_x {\rm div}_x \left( \varrho_\varepsilon\vc{u}_\varepsilon \right)
\right\|_{L^2(\mathbb T^3; \mathbb{R}^{3 \times 3})}} \\
&\leq\,c\,\expe{
\|\vc{u}_\varepsilon \|_{L^6(\mathbb T^3; \mathbb{R}^3)} \| \varrho_\varepsilon \|_{L^{3}(\mathbb T^3)} \left\|\varrho_\varepsilon\vc{u}_\varepsilon
\right\|_{L^2(\mathbb T^3; \mathbb{R}^{3})}}\\
&\leq\,c\,\expe{
\|\vc{u}_\varepsilon \|_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)}^2 \| \varrho_\varepsilon \|_{L^{3}(\mathbb T^3; \mathbb{R}^3)}^2 }\\
&\leq\,c\,\expe{
\|\vc{u}_\varepsilon \|_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)}^2 \| \varrho_\varepsilon \|_{L^{\Gamma}(\mathbb T^3; \mathbb{R}^3)}^\Gamma}+c(G,M_0)
\end{align*}
using (\ref{d1}). We also have that
\begin{align*}
(III)&\leq\,\kappa\delta\,\expe{\|\varrho_\varepsilon\|_{L^2(\mathbb{T}^3)}^2}+c(\kappa,\delta)\expe{\|\nabla\mathbf{u}_\varepsilon\|^2_{L^2(\mathbb{T}^3;\mathbb{R}^3)}}\\
&\leq\,\kappa\delta\,\expe{\|\varrho_\varepsilon\|_{L^\Gamma(\mathbb{T}^3)}^\Gamma}+c(\kappa,\delta,G,M_0)
\end{align*}
as well as
\begin{align*}
(VI)&\leq\,\kappa\delta\,\mathbb{E}\|\varrho_\varepsilon\|_{L^\Gamma(\mathbb{T}^3)}^\Gamma+c(\kappa,\delta,G,M_0).
\end{align*}
Finally, continuity of $\nabla_x\Delta_x^{-1}$ and (\ref{d1}) imply
\begin{align*}
(V)&\leq\,\kappa\delta\,\bigg(\mathbb{E}\|\varrho_\varepsilon\|_{L^4(\mathbb{T}^3)}^4+\mathbb{E}\|\nabla_x\Delta_x^{-1}[\varrho_\varepsilon-M_\varepsilon]\|_{L^4(\mathbb{T}^3)}^4\,{\rm d} {x}\bigg)+c(\kappa,\delta)\|\mathbf{u}_\varepsilon\|_{L^2(\mathbb{T}^3;\mathbb{R}^3)}^2\\
&\leq\,c\,\kappa\delta\,\mathbb{E}\|\varrho_\varepsilon\|_{L^\Gamma(\mathbb{T}^3)}^\Gamma+c(\kappa,\delta,G,M_0)
\end{align*}
Summing up the inequalities above, choosing $\kappa$ small enough and using stationarity, we obtain
\begin{equation*
\begin{split}
\mathbb E\bigg[& \intO{ a \varrho_\varepsilon^{\gamma + 1} + \delta \varrho_\varepsilon^{\Gamma +1}} + \intO{ \frac{1}{3} \varrho_\varepsilon^{2} |\vc{u}_\varepsilon|^2 } \bigg] \\
&\leq \expe{ \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^\Gamma \right] } \,\| \vc{u}_\varepsilon \|_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)}^2 } + c(\delta,G,M_0).
\end{split}
\end{equation*}
Thus, due to \eqref{d12a} and Young's inequality, we may conclude that the stationary solution $[\varrho_\varepsilon,\vc{u}_\varepsilon]$ admits the uniform bound \eqref{d131} as well as
\begin{equation} \label{d141}
\expe{ \left(1 + \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^\Gamma \right] } \right) \| \vc{u}_\varepsilon \|^2_{W^{1,2} (\mathbb T^3; \mathbb{R}^3)} } \leq c(\delta,G,M_0).
\end{equation}
Finally, let us show \eqref{d151}.
To this end, we may go back to the energy inequality (\ref{b2a}) for $n=1$, obtaining
\[
\begin{split}
&\expe{ \left( \sup_{t \in [T, T + \tau] } \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^s }
+\expe{\bigg(\int_T^{T+\tau}\|\vc{u}_\varepsilon\|_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)}^2\,{\rm d} t \bigg)^s}\\
&\qquad\leq c(s)\,\expe{ \sup_{t \in [T, T + \tau] } \left|\sum_{k=1}^\infty\int_{T}^{t} \intO{ \varrho_\varepsilon\,\mathbf{g}_k(\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon) \cdot \vc{u}_\varepsilon } \,{\rm d} W_k \right|^s }\\
&\qquad\quad +c(s)\,\expe{\sup_{t \in [T, T + \tau] } \bigg|\sum_{k=1}^\infty\int_T^{t}\int_{\mathbb{T}^3}\frac{1}{\varrho_\varepsilon}|{\varrho_\varepsilon}\mathbf g_k({\varrho_\varepsilon},{\varrho_\varepsilon}\vc{u}_\varepsilon)|^2\,\mathrm{d} x\,\mathrm{d} r\bigg|^s}\\
&\qquad\quad+ c(s)\,\expe{\sup_{t \in [T, T + \tau] } \bigg|\int_T^{t} \int_{\mathbb{T}^3}\varrho_\varepsilon^{\gamma-1}+\varrho_\varepsilon^{\Gamma-1}\,\mathrm{d} x\,\mathrm{d} r\bigg|^s}.
\end{split}
\]
The first term on the right hand side is estimated using the Burkholder-Davis-Gundy inequality and (\ref{b4bc}); the second term using \eqref{b4n}. We deduce that
\[
\begin{split}
&\expe{ \left( \sup_{t \in [T, T + \tau] } \intO{ \left[ \frac{1}{2} \varrho_\varepsilon |\vc{u}_\varepsilon|^2 + \frac{a}{\gamma - 1} \varrho_\varepsilon^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\varepsilon^{\Gamma} \right] } \right)^s } \\
& \quad\leq c(s,\tau,M_0,G)\left(1+\mathbb{E}\bigg[\bigg(\int_T^{T+\tau}\int_{\mathbb{T}^3}\varrho_\varepsilon|\vc{u}_\varepsilon|^2\,\mathrm{d} x\,\mathrm{d} r\bigg)^\frac{s}{2}\bigg] + \expe{ \bigg|\int_T^{T+\tau} \int_{\mathbb{T}^3}\varrho_\varepsilon^{\gamma-1}+\varrho_\varepsilon^{\Gamma-1}\,\mathrm{d} x\,\mathrm{d} r\bigg|^s}\right).
\end{split}
\]
Now, by H\" older's inequality, stationarity, \eqref{d141} and \eqref{d131}, for $s\in (1,2)$,
\begin{equation}\label{e4}
\begin{split}
&\mathbb{E}\left( \int_T^{T+\tau}\intO{ \varrho_\varepsilon |\vc{u}_\varepsilon |^2 } \,\mathrm{d} r\right)^\frac{s}{2}\\
&\qquad\leq c(\tau,s)\,\mathbb{E}\bigg(\int_T^{T+\tau}\left( \| \sqrt{\varrho_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_\varepsilon \|_{L^6(\mathbb T^3; \mathbb{R}^3)} \right)^s \| \sqrt{\varrho_\varepsilon} \|_{L^{3}(\mathbb T^3)}^s\,{\rm d} t \bigg)^\frac{1}{2}\\
&\qquad\leq c(\tau,s)\, \mathbb{E}\bigg(\int_T^{T+\tau} \| \sqrt{\varrho_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3; \mathbb{R}^3)}^2 \| \vc{u}_\varepsilon \|_{L^6(\mathbb T^3; \mathbb{R}^3)}^2+\|\sqrt{\varrho_\varepsilon}\|_{L^3(\mathbb{T}^3)}^\frac{2s}{2-s}\,{\rm d} t \bigg)^\frac{1}{2}\\
&\qquad\leq c(\tau,s) \left(\expe{ \| \sqrt{\varrho_\varepsilon} \vc{u}_\varepsilon \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_\varepsilon \|_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)}^2 }+ \mathbb{E}\|\varrho_\varepsilon\|_{L^\gamma(\mathbb{T}^3)}^{\frac{s}{2 -{s}}}\right)\\
&\qquad\leq c(\tau,s,\d,G,M_0)\left(1+ \mathbb{E}\left( \intO{ \varrho_\varepsilon^{\gamma+1} } \right)^{\frac{s}{(\gamma+1)( 2 -{s})}} \right)\\
&\qquad\leq c(\tau,s,\d,G,M_0)\left(1+ \mathbb{E}\intO{ \varrho_\varepsilon^{\gamma+1} } \right)\leq c(\tau,s,\d,G,M_0)
\end{split}
\end{equation}
provided $s\leq \tfrac{2(\gamma+1)}{\gamma+2}$. Similarly,
\[
\bigg(\int_{\mathbb{T}^3}\varrho_\varepsilon^{\gamma-1}+\varrho_\varepsilon^{\Gamma-1}\,\mathrm{d} x\bigg)^s\leq 1+\int_{\mathbb{T}^3}\varrho_\varepsilon^{\gamma+1}+\varrho_\varepsilon^{\Gamma+1}\,\mathrm{d} x,
\]
provided $s\leq\tfrac{\Gamma+1}{\Gamma-1}$. Consequently, \eqref{d151} follows due to \eqref{d131}.
\end{proof}
With Proposition \ref{prop:1608} and Proposition \ref{prop:1708} at hand, we are able to follow the compactness argument of \cite[Section 5.1]{BrHo}.
To be more precise, as $\varepsilon\rightarrow0$ we aim at constructing stationary solutions to the following system.
\begin{itemize}
\item {\bf Equation of continuity.}
\begin{equation} \label{L1'}
\begin{split}
\int_{0}^{\infty} \intO{ \left[ \varrho \partial_t \varphi + \varrho \vc{u} \cdot \nabla_x \varphi \right] } \,{\rm d} t =0
\end{split}
\end{equation}
for any $\varphi \in C^\infty_c((0, \infty) \times \mathbb{T}^3)$ $\mathbb{P}$-a.s.
\item {\bf Regularized momentum equation.}
\begin{equation} \label{L2'}
\begin{split}
\int_0^\infty \partial_t \psi\intO{ \varrho \vc{u} \cdot \varphi} \,{\rm d} t &+ \int_0^\infty \psi\intO{ \varrho \vc{u} \otimes \vc{u} : \nabla_x \varphi } \,{\rm d} t
+ \int_0^\infty \psi\intO{ (a \varrho^\gamma+\delta\varrho^\Gamma) {\rm div}_x \varphi } \,{\rm d} t \\
&- \int_0^\infty \psi\intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \varphi} \,{\rm d} t
=-\sum_{k=1}^\infty \int_0^\infty \psi \intO{ \mathbf{G}_k(\varrho,\varrho\vc{u}) \cdot \varphi } \ {\rm d} W_k
\end{split}
\end{equation}
for any $\psi \in C^\infty_c((0, \infty))$, $\varphi \in C^\infty(\mathbb{T}^3; \mathbb{R}^3)$ $\mathbb{P}$-a.s.
\end{itemize}
Note that unlike the energy estimate in \cite{BrHo}, the bound \eqref{d151} only gives limited moment estimates, i.e. $s$ cannot be arbitrarily large. Nevertheless, \eqref{d151} is sufficient to perform the passage to the limit. We also point out that the assumption \eqref{m1} on the noise coefficients is actually stronger than the one in \cite{BrHo}, and consequently the convergence of the stochastic integral is more straightforward.
We deduce the following.
\begin{Proposition}\label{prop:1608b'}
Let $\delta>0$ be given. Then there exists a stationary solution $[\varrho_\delta,\mathbf{u}_\delta]$ to \eqref{L1'}--\eqref{L2'}. Moreover, we have the estimates
\begin{equation} \label{d1'}
\expe{
\| \vc{u}_\delta(t) \|^2_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)} } \leq c(G,M_0),
\end{equation}
and
\begin{equation} \label{d51'}
\begin{split}
&\expe{ \left( \intO{ \left[ \frac{1}{2} \varrho_{\delta} |\vc{u}_{\delta}|^2 + \frac{a}{\gamma - 1} \varrho_{\delta}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\delta}^{\Gamma} \right] } \right)
\intO{ \| \vc{u}_{\delta}\|^2_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)} } }\\
&\leq c(G,M_0)
\expe{
\intO{ \left[ \frac{1}{2} \varrho_{\delta} |\vc{u}_{\delta}|^2 + \frac{a}{\gamma - 1} \varrho_{\delta}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\delta}^{\Gamma} \right] } } + c(M_0),
\end{split}
\end{equation}
for a.e. $t\in(0,\infty)$.
In addition, the equation of continuity \eqref{L1'} holds true in the renormalized sense and for all $\psi\in C^\infty_c ((0,\infty))$, $\psi\geq0$, the following energy inequality holds true
\begin{align} \label{b2ac}
\begin{aligned}
&-\int_0^\infty\partial_t\psi \left( \intO{ \left[ \frac{1}{2} \varrho_\d |\vc{u}_\d|^2 + \frac{a}{\gamma - 1} \varrho_\d^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\d^{\Gamma} \right] } \right)\,{\rm d} t \\
&\quad+ \int_0^\infty\psi
\intO{ \mathbb{S}(\nabla_x \vc{u}_\d) : \nabla_x \vc{u}_\d } \,{\rm d} t \\
& \leq \sum_{k=1}^\infty \int_0^\infty\psi \intO{ \varrho_\d\,\mathbf{g}_k(\varrho_\d,\varrho_\d\vc{u}_\d) \cdot \vc{u}_\d } \, {\rm d} W_k \\
&\quad+ \frac{1}{2} \int_0^\infty\psi
\sum_{k=1}^\infty\intO{ \frac{1}{\varrho_\d} |\varrho_\d\,\vc{g}_k(\varrho_\d,\varrho_\d\vc{u}_\d)|^2 } \,{\rm d} t .
\end{aligned}
\end{align}
\end{Proposition}
\begin{proof}
First, we proceed as in \cite[Section 5.1]{BrHo} and establish the necessary tightness of the joint law of $[\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon,\vc{u}_\varepsilon,W]$. The only difference is that the corresponding path spaces are replaced by their local-in-time analogs as discussed in the proof of Proposition \ref{prop:1608b}. Consequently, the Jakubowski-Skorokhod theorem applies and we obtain a new family of martingale solutions, still denoted by
$[\varrho_\varepsilon,\varrho_\varepsilon\vc{u}_\varepsilon,\vc{u}_\varepsilon,W]$, obeying the same laws and converging in probability with respect to a new basis, still denoted by
$\big( \Omega, \mathfrak{F}, ( \mathfrak{F}_t )_{t \geq 0} ,\prst \big)$.
In addition, the limit satisfies
\begin{equation} \label{L3}
\int_{0}^{\infty} \intO{ \left[ \varrho \partial_t \varphi + \varrho \vc{u} \cdot \nabla_x \varphi \right] } \,{\rm d} t = 0, \qquad \intO{ \varrho } = M_0,
\end{equation}
for any $\varphi \in C^\infty_c((0, \infty) \times \mathbb{T}^3)$ $\mathbb{P}$-a.s.,
\begin{equation} \label{L4}
\begin{split}
\int_0^\infty \partial_t \psi \intO{ \varrho \vc{u} \cdot \varphi} \,{\rm d} t &+ \int_0^\infty\psi \intO{ \varrho \vc{u} \otimes \vc{u} : \nabla_x \varphi } \,{\rm d} t
+ \int_0^\infty \psi\intO{ \big(a \Ov{\varrho^\gamma} +\d \Ov{\varrho^\Gamma}\big) {\rm div}_x \varphi } \,{\rm d} t \\
&- \int_0^\infty \psi\intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \varphi} \,{\rm d} t =- \int_0^\infty \psi {\rm d} M_\varphi
\end{split}
\end{equation}
for any $\psi \in C^\infty_c((0, \infty))$, $\varphi \in C^\infty(\mathbb{T}^3; \mathbb{R}^3)$ $\mathbb{P}$-a.s. Here $M_\varphi$ is a square integrable martingale and the bars denote the corresponding weak limits with respect to $t,x$. For details, we refer to \cite[Proposition 5.6]{BrHo}.
In addition, $\varrho$ satisfies the renormalized equation of continuity. That is
\begin{align}\label{eq:rendelta}
\partial_t b({\varrho})&+{\rm div}_x\big(b({\varrho}){\mathbf{u}}\big)
+\big(b'({\varrho}){\varrho}-b({\varrho})\big){\rm div}_x{\mathbf{u}}
=0
\end{align}
in the sense of distribution on $(0,\infty)\times\mathbb{T}^3$ for every $b\in C^1([0,\infty))$ with $b'(z)=0$ for $z\geq M_b$ for some constant $M_b>0$. However, as discussed in \cite[Remark 1.1]{feireisl1}, the assumption on $b'$ can be weakened to
$$|b'(z)z|\leq c(z^\theta+z^\frac{\gamma}{2})\quad\text{for all } z>0 \text{ and some }\theta\in(0,\tfrac{\gamma}{2}).$$
This in particular includes the function $b(z)=z\log z$ employed below.
In order to complete the proof, it is enough to show strong convergence of the densities as in \cite[Section 5.2]{BrHo}. More specifically, we prove that
\begin{equation} \label{L5}
\limsup_{\varepsilon \to 0} \expe{ \| \vr_\varepsilon - \varrho \|^{\Gamma + 1}_{L^{\Gamma + 1}(\mathbb T^3)} }\leq
\limsup_{\varepsilon \to 0} \expe{ \intO{ \left( \vr_\varepsilon^{\Gamma + 1} - \Ov{\varrho^\Gamma} \varrho \right) } } \leq 0 \ \mbox{for any} \ t > 0.
\end{equation}
Note that the first inequality follows from the algebraic inequality which holds true
\[
(A - B)^{\Gamma + 1} = (A - B)^\Gamma (A-B) \leq\, (A^\Gamma - B^\Gamma)(A - B) \ \mbox{whenever}\ A,B \geq 0.
\]
In order to see {the rightmost inequality} in (\ref{L5}) we use the method of Lions \cite{LI4} based on regularity of the effective viscous flux. More specifically, mimicking
the technique from the proof of Proposition \ref{prop:1708}, we derive from \eqref{L1}--\eqref{L2} the identity
\begin{equation}\label{L6a}
\begin{split}
\int_T^{T+1} &\intO{ \Big(a \vr_\varepsilon^{\gamma + 1}+\delta \vr_\varepsilon^{\Gamma + 1}\Big) } \ \,{\rm d} t =M_\varepsilon
\int_T^{T+1} \intO{ \left( a \vr_\varepsilon^\gamma+\delta \vr_\varepsilon^{\Gamma } \right) } \,{\rm d} t \\
&+ \int_T^{T+1} \intO{\Big( \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} {\rm div}_x (\vr_\varepsilon \vc{u}_\varepsilon ) - \vr_\varepsilon \vc{u}_\varepsilon \otimes \vc{u}_\varepsilon : \nabla_x \Delta^{-1}_x
\nabla_x \vr_\varepsilon \Big)} \,{\rm d} t \\
&+
\int_T^{T+1} \intO{ \left(\frac{4}{3} \mu+\eta \right) {\rm div}_x \vc{u}_\varepsilon \ \vr_\varepsilon } \,{\rm d} t \\
&+ 2\varepsilon \int_T^{T+1} \intO{ \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} \left[
\vr_\varepsilon - M_\varepsilon \right] } \,{\rm d} t + \varepsilon \int_T^{T+1} \intO{ \vr_\varepsilon^2 {\rm div}_x \vc{u}_\varepsilon } \,{\rm d} t \\
&+
\left[ \intO{ \varrho \vc{u} \cdot \nabla_x \Delta^{-1}_x \left[ \vr_\varepsilon - M_\varepsilon \right] } \right]_{t = T}^{t = T+1} \\
&- \sum_{k=1}^\infty\int_T^{T+1} \intO{ \mathbf{G}_k(\vr_\varepsilon,\vr_\varepsilon\vc{u}_\varepsilon) \cdot \nabla_x \Delta^{-1} \left[ \vr_\varepsilon - M_\varepsilon \right] } \ {\rm d} W_k.
\end{split}
\end{equation}
In addition, since $\vr_\varepsilon$ satisfies the equation of continuity in the strong sense, the application of the commutator lemma in the spirit of \cite{DL} yields
\begin{align*
\mathrm{d} (\vr_\varepsilon\log\vr_\varepsilon)=-{\rm div}_x\big(\vr_\varepsilon\log\vr_\varepsilon\,{\mathbf{u}}_\varepsilon\big)-\vr_\varepsilon\,{\rm div}_x\vc{u}_\varepsilon+\varepsilon\Delta( \vr_\varepsilon\log\vr_\varepsilon)-\varepsilon \frac{|\nabla\vr_\varepsilon|^2}{\vr_\varepsilon}.
\end{align*}
Inserting this into \eqref{L6a} implies
\begin{equation} \label{L6}
\begin{split}
\int_T^{T+1} &\intO{ \Big(a \vr_\varepsilon^{\gamma + 1}+\delta \vr_\varepsilon^{\Gamma + 1}\Big) } \ \,{\rm d} t =
M_\varepsilon\int_T^{T+1}\intO{ \Big(a \vr_\varepsilon^\gamma+\delta \vr_\varepsilon^{\Gamma}\Big) } \,{\rm d} t \\
&+ \int_T^{T+1} \intO{\Big( \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} {\rm div}_x (\vr_\varepsilon \vc{u}_\varepsilon ) - \vr_\varepsilon \vc{u}_\varepsilon \otimes \vc{u}_\varepsilon : \nabla_x\Delta^{-1}_x
\nabla_x \vr_\varepsilon \Big)} \,{\rm d} t \\
&
- \left( \frac{4}{3} \mu+\eta \right) \left[ \intO{ \vr_\varepsilon \log (\vr_\varepsilon) } \right]_{t = T}^{t = T+1} - \varepsilon \left( \frac{4}{3} \mu+\eta \right)\int_T^{T+1} \intO{ \frac{|\nabla_x \vr_\varepsilon|^2}{\vr_\varepsilon} } \,{\rm d} t \\
&+ 2\varepsilon \int_T^{T+1} \intO{ \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} \left[
\vr_\varepsilon - M_\varepsilon \right] } \,{\rm d} t + \varepsilon \int_T^{T+1} \intO{ \vr_\varepsilon^2 {\rm div}_x \vc{u}_\varepsilon } \,{\rm d} t \\
&+
\left[ \intO{ \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1}_x \left[ \vr_\varepsilon - M_\varepsilon \right] } \right]_{t = T}^{t = T+1} \\
&- \sum_{k=1}^\infty\int_T^{T+1} \intO{ \vc{G}_k(\vr_\varepsilon,\vr_\varepsilon\vc{u}_\varepsilon) \cdot \nabla_x \Delta^{-1} \left[ \vr_\varepsilon - M_\varepsilon\right] } \ {\rm d} W_k.
\end{split}
\end{equation}
Similarly, as the limit density $\varrho$ also satisfies the renormalized equation of continuity \eqref{eq:rendelta}, we deduce choosing $b(z)=z\log z$ that
\begin{align*
\mathrm{d} (\varrho\log\varrho)=-{\rm div}_x\big(\varrho\log\varrho\,{\mathbf{u}}\big)-\varrho\,{\rm div}_x\vc{u}
\end{align*}
holds true in the sense of distributions.
Therefore, we obtain from the limit equations (\ref{L3}), (\ref{L4}) that\begin{equation*
\begin{split}
\int_T^{T+1}& \intO{ \left(a \Ov{\varrho^{\gamma}} +\delta\Ov{\varrho^\Gamma} \right)\varrho } \ \,{\rm d} t =M_0
\int_T^{T+1} \intO{ \left( a \Ov{\varrho^\gamma}+\delta\Ov{\varrho^\Gamma} \right) } \,{\rm d} t \\
&+ \int_T^{T+1} \intO{ \Big(\varrho \vc{u} \cdot \nabla_x \Delta^{-1} {\rm div}_x (\varrho \vc{u} ) - \varrho \vc{u} \otimes \vc{u} : \nabla_x\Delta^{-1}_x
\nabla_x\varrho \Big) } \,{\rm d} t \\
&
- \left( \frac{4}{3} \mu+\eta \right) \left[ \intO{ \varrho \log \varrho} \right]_{t = T}^{t = T+1} \\
&+
\left[ \intO{ \varrho \vc{u} \cdot \nabla_x \Delta^{-1}_x \left[ \varrho - M_0 \right] } \right]_{t = T}^{t = T+1} - \int_T^{T+1} {\rm d} M_\Phi,
\end{split}
\end{equation*}
with
\[
\Phi = \nabla_x \Delta^{-1}_x \left[ \varrho - M_0\right].
\]
Thus passing to expectations and using the fact that the processes are stationary, we get
\begin{align}
\nonumber
\mathbb{E}\bigg[ &\int_T^{T+1}\intO{\Big(a \vr_\varepsilon^{\gamma + 1}+\delta\vr_\varepsilon^\Gamma\Big)\vr_\varepsilon } \,{\rm d} t \bigg] \leq
M_{\varepsilon} \,\expe{ \int_T^{T+1}\intO{ \Big(a \vr_\varepsilon^\gamma+\delta\vr_\varepsilon^\Gamma\Big) } \,{\rm d} t } \\
\label{2702}&+ \expe{ \int_T^{T+1}\intO{\Big( \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} {\rm div}_x (\vr_\varepsilon \vc{u}_\varepsilon ) - \vr_\varepsilon \vc{u}_\varepsilon \otimes \vc{u}_\varepsilon : \nabla_x\Delta^{-1}_x
\nabla_x \vr_\varepsilon \Big) } \,{\rm d} t } \\
&+ 2 \varepsilon \,\expe{ \intO{ \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} \left[
\vr_\varepsilon - M_\varepsilon \right] } } + \varepsilon \,\expe{ \intO{ \vr_\varepsilon^2 {\rm div}_x \vc{u}_\varepsilon } }.\nonumber
\end{align}
Note that the inequality is due to the fact that we are not able to pass to the limit in the fourth term on the right hand side of \eqref{L6} and we can only use its negativity.
Similarly we obtain
\begin{align}\nonumber
&\expe{ \int_T^{T+1}\intO{ \Big(a \Ov{\varrho^{\gamma}}+\delta\Ov{\varrho^\Gamma}\Big) \varrho }\,{\rm d} t } =
M_0\, \expe{ \int_T^{T+1}\intO{ \left( a \Ov{\varrho^\gamma}+\delta\Ov{\varrho^\Gamma} \right) } \,{\rm d} t } \\
&+ \expe{ \int_T^{T+1} \intO{ \Big(\varrho \vc{u} \cdot \nabla_x \Delta^{-1} {\rm div}_x (\varrho \vc{u} ) - \varrho \vc{u} \otimes \vc{u} : \nabla_x\Delta^{-1}_x
\nabla_x \varrho \Big) }\,{\rm d} t }.\label{2702b}
\end{align}
Note that the $\varepsilon$-terms in \eqref{2702} vanish due to Proposition \ref{prop:1708} and we have $M_\varepsilon\rightarrow M_0$ as $\varepsilon\to0$.
Consequently, the desired conclusion (\ref{L5}) follows as soon as we observe that
\begin{equation} \label{L8}
\begin{split}
&\lim_{\varepsilon \to 0} \expe{ \int_T^{T+1} \intO{ \Big(\vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} {\rm div}_x (\vr_\varepsilon \vc{u}_\varepsilon ) - \vr_\varepsilon \vc{u}_\varepsilon \otimes \vc{u}_\varepsilon : \nabla_x \Delta^{-1}_x
\nabla_x \vr_\varepsilon \Big) } \,{\rm d} t }\\
&= \expe{ \int_T^{T+1} \intO{ \Big(\varrho \vc{u} \cdot \nabla_x \Delta^{-1} {\rm div}_x (\varrho \vc{u} ) - \varrho \vc{u} \otimes \vc{u} : \nabla_x \Delta^{-1}_x
\nabla_x \varrho \Big) } \,{\rm d} t }.
\end{split}
\end{equation}
In fact, \eqref{L8} in combination with \eqref{2702} and \eqref{2702b} implies
\begin{align*}
\limsup_{\varepsilon\to0}\,\expe{ \int_T^{T+1}\intO{\Big(a \vr_\varepsilon^{\gamma }+\delta\vr_\varepsilon^\Gamma\Big)\vr_\varepsilon } \,{\rm d} t }\leq \expe{\int_T^{T+1} \intO{ \Big(a \Ov{\varrho^{\gamma}}+\delta\Ov{\varrho^\Gamma}\Big) \varrho } \,{\rm d} t }
\end{align*}
which shows strong convergence of $\vr_\varepsilon$ by monotonicity arguments.
Relation \eqref{L8} can be established by compensated compactness arguments (applied $\mathbb{P}$-a.s.) if we show that the expressions under expectations are $\mathbb{P}$-equi-integrable.
Considering the two summands separately and using continuity of $\nabla_x \Delta^{-1}_x \nabla_x$, we have
\[
\begin{split}
\left| \intO{ \vr_\varepsilon \vc{u}_\varepsilon \cdot \nabla_x \Delta^{-1} {\rm div}_x (\vr_\varepsilon \vc{u}_\varepsilon ) } \right| &\leq \,
c \| \sqrt{\vr_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3)} \| \sqrt{\vr_\varepsilon} \|_{L^{2 \Gamma}(\mathbb T^3)} \| \nabla_x \Delta^{-1}_x \nabla_x\vr_\varepsilon\vc{u}_\varepsilon \|_{L^{\frac{2\Gamma}{\Gamma-1}}(\mathbb T^3; \mathbb{R}^3)}\\
&\leq \,
c \| \sqrt{\vr_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3)} \| \sqrt{\vr_\varepsilon} \|_{L^{2 \Gamma}(\mathbb T^3)} \| \vr_\varepsilon\vc{u}_\varepsilon \|_{L^{\frac{2\Gamma}{\Gamma-1}}(\mathbb T^3; \mathbb{R}^3)}\\
&\leq
c \| \sqrt{\vr_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3)} \| \sqrt{\vr_\varepsilon} \|_{L^{2 \Gamma}(\mathbb T^3)}\| \vc{u}_\varepsilon \|_{L^{6}(\mathbb T^3; \mathbb{R}^3)} \| \vr_\varepsilon \|_{L^\Gamma (\mathbb T^3) }\\
&\leq
c \| \sqrt{\vr_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3)} \| \sqrt{\vr_\varepsilon} \|_{L^{2 \Gamma}(\mathbb T^3)} \| \vc{u}_\varepsilon \|_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)} \| \vr_\varepsilon \|_{L^\Gamma (\mathbb T^3) },
\end{split}
\]
as $\Gamma \geq \frac{9}{2}$. Similarly, we have
\[
\begin{split}
\left| \intO{ \vr_\varepsilon \vc{u}_\varepsilon \otimes \vc{u}_\varepsilon : \nabla_x \Delta^{-1}_x \nabla_x \vr_\varepsilon } \right| &\leq
c \| \sqrt{\vr_\varepsilon} \vc{u}_\varepsilon \|_{L^2(\mathbb T^3)} \| \vc{u} \|_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)} \| \sqrt{\vr_\varepsilon} \|_{L^{2 \Gamma}(\mathbb T^3)} \| \vr_\varepsilon \|_{L^\Gamma (\mathbb T^3) }.
\end{split}
\]
Here, in accordance with (\ref{d141}),
\[
\expe{ \| \sqrt{\vr_\varepsilon} \vc{u}_\varepsilon \|^2_{L^2(\mathbb T^3)} \| \vc{u} \|^2_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)} } \leq c(\delta,G,M_0),
\]
while, by virtue of (\ref{d131}),
\[
\| \sqrt{\vr_\varepsilon} \|_{L^{2 \Gamma}(\mathbb T^3)} \| \vr_\varepsilon \|_{L^\Gamma (\mathbb T^3) } = \| \vr_\varepsilon \|_{L^\Gamma (\mathbb T^3)}^{\frac{3}{2}} \in L^q(\Omega),\qquad
q = \frac{2 \Gamma }{3} > 2.
\]
We have shown (\ref{L5}); whence strong convergence of $\vr_\varepsilon$. Consequently, as in \cite[Section 5.2]{BrHo}, we may identify the nonlinear terms in \eqref{L4} and hence $[\varrho,\vc{u}]$ is a weak martingale solution to \eqref{L1'}--\eqref{L2'}. Stationarity then follows by Lemma \ref{lem:stac} and Lemma \ref{lem:stac2}. The estimate \eqref{d1'} and \eqref{d51'}, respectively, is obtained by weak lower semicontinuity from \eqref{d1} and \eqref{d12a}, respectively, since the constants were uniform in $\varepsilon$. The same arguments give the energy inequality \eqref{b2ac}. Note that the passage to the limit in the stochastic integral can be justified for instance with help of \cite[Lemma 2.1]{debussche1}.
\end{proof}
\begin{Remark}\label{rem:1808}
It is important to note that there is an essential difference between the strong convergence of the density in the existence theory, see \cite[Section 5.2]{BrHo}, and the above proof. More specifically, the existence theory requires compactness of the initial data which is not available in the present setting. Instead the fact that the solution is stationary must be used.
\end{Remark}
\section{Vanishing artificial pressure limit}
\label{P}
As the final step of the proof of our main result, Theorem \ref{Tm1}, it remains to perform the last limit procedure, that is, $\d\to0$. Recall that according to Proposition \ref{prop:1608b'}, the stationary solutions constructed in the previous section already satisfy the uniform bounds \eqref{d1'} and \eqref{d51'}. Nevertheless, the pressure estimate as well as the estimate for the energy and velocity from Proposition \ref{prop:1708} all blow up as $\d$ vanishes. Therefore, in order to apply the compactness argument from \cite[Section 6]{BrHo} it is necessary to improve these estimates. The rest of the construction then proceeds exactly as in \cite[Section 6.1--6.3]{BrHo}.
\begin{Proposition}\label{prop:1808}
Let $[\varrho_\d,\vc{u}_\d]$ be the stationary solution to \eqref{L1'}--\eqref{L2'} constructed in Proposition~\ref{prop:1608b'}. Then the following uniform bound holds true for some $\a>0$ and a.e. $t\in(0,\infty)$
\begin{equation} \label{d13}
\expe{ \intO{ \bigg[ a \varrho_{\delta}^{\gamma + \alpha} + \delta \varrho_{\delta}^{\Gamma + \alpha} +
\varrho_{\delta}^{1 + \alpha} |\vc{u}_{\delta}|^2 \bigg]} } \leq c(G,M_0),
\end{equation}
In addition, for some $s>1$
and for a.e. $T > 0$ and $\tau>0$
\begin{equation} \label{d151'}
\begin{split}
&\expe{ \left( \sup_{t \in [T, T + \tau] } \intO{ \left[ \frac{1}{2} \varrho_\d |\vc{u}_\d|^2 + \frac{a}{\gamma - 1} \varrho_\d^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\d^{\Gamma} \right] } \right)^s }\\
&\qquad +\expe{\bigg(\int_T^{T+\tau}\|\vc{u}_\d\|_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)}^2\,{\rm d} t \bigg)^s}\leq c(\tau, M_0, G, s),
\end{split}
\end{equation}
where the constant is independent of $T$.
\end{Proposition}
\begin{proof}
As far as the pressure estimates are concerned we use the test function
$$\nabla_x\Delta_x^{-1} \bigg[\varrho^\alpha-\int_{\mathbb T^3}\varrho^\alpha\,{\rm d} {x}\bigg],\quad \alpha>0.$$
We obtain
after a rather tedious but straightforward manipulation the following analogue of \eqref{d41}
\begin{equation} \label{d2}
\begin{split}
&\int_T^{T+1} \intO{ \left( a \varrho_\delta^{\gamma + \alpha} + \delta \varrho_\delta^{\Gamma + \alpha} \right) } \ \,{\rm d} t + \int_T^{T+1} \intO{ \frac{1}{3} \varrho_\delta^{1 + \alpha} |\vc{u}|^2 } \ \,{\rm d} t \\
&=
\int_T^{T+1} \left( \intO{ \left( a \varrho_\delta^\gamma +
\delta \varrho_\delta^\Gamma \right) } \intO{ \varrho_\delta^\alpha } \right) \,{\rm d} t + \frac{1}{3} \int_T^{T+1}
\left( \intO{ \varrho_\delta |\vc{u}_\delta|^2 } \,{\rm d} {x} \intO{ \varrho_\delta^\alpha } \right) \,{\rm d} t \\
&\qquad+
\int_T^{T+1} \intO{ \left(\frac{4}{3} \mu+\eta \right) {\rm div}_x \vc{u}_\delta \ \varrho_\delta^\alpha } \,{\rm d} t \\
&\qquad - \int_T^{T+1} \intO{ \left( \varrho_\delta \vc{u}_\delta \otimes \vc{u}_\delta - \frac{1}{3} \varrho_\delta |\vc{u}_\delta |^2 \mathbb{I} \right) : \nabla_x \Delta^{-1}_x
\nabla_x \left[ \varrho_\delta^\alpha \right] } \,{\rm d} t \\
&\qquad+
\left[ \intO{ \varrho_\delta \vc{u}_\delta \cdot \nabla_x \Delta^{-1}_x \left[ \varrho_\delta^\alpha - \intO{\varrho_\delta^\alpha } \right] } \right]_{t = T}^{t = T+1} \\
&\qquad- \int_T^{T+1} \intO{ \varrho_\delta \vc{u}_\delta \cdot \nabla_x \Delta^{-1} [ {\rm d} (\varrho_\delta^\alpha) ] } \\
&\qquad- \sum_{k=1}^\infty\int_T^{T+1} \intO{ \vc{G}_k(\varrho_\delta,\varrho_\delta \vc{u}_\delta) \cdot \nabla_x \Delta^{-1} \left[ \varrho_\delta^\alpha - \intO{\varrho_\delta^\alpha } \right] } \ {\rm d} W_k.
\end{split}
\end{equation}
Next, we evoke the renormalized equation of continuity \eqref{eq:rendelta}
\begin{equation*
{\rm d} \varrho_\delta^{\alpha} + {\rm div}_x (\varrho_\delta^\alpha \vc{u}_\delta ) \,{\rm d} t + (\alpha - 1) \varrho_\delta^\alpha {\rm div}_x \vc{u}_\delta \,{\rm d} t = 0
\end{equation*}
deducing from (\ref{d2})
\begin{equation} \label{d4}
\begin{split}
&\int_T^{T+1} \intO{ \left( a \varrho_\delta^{\gamma + \alpha} + \delta \varrho_\delta^{\Gamma + \alpha} \right) } \ \,{\rm d} t + \int_T^{T+1} \intO{ \frac{1}{3} \varrho_\delta^{1 + \alpha} |\vc{u}_\delta|^2 } \ \,{\rm d} t \\
&=
\int_T^{T+1} \left( \intO{ \left( a \varrho_\delta^\gamma +
\delta \varrho^\Gamma \right) } \intO{ \varrho_\delta^\alpha } \right) \,{\rm d} t + \frac{1}{3} \int_T^{T+1}
\left( \intO{ \varrho_\delta |\vc{u}_\delta|^2 } \intO{ \varrho_\delta^\alpha } \right) \,{\rm d} t \\
&\qquad+
\int_T^{T+1} \intO{ \left(\frac{4}{3} \mu+\eta \right) {\rm div}_x \vc{u}_\delta \ \varrho_\delta^\alpha } \,{\rm d} t \\
&\qquad - \int_T^{T+1} \intO{ \left( \varrho_\delta \vc{u}_\delta \otimes \vc{u}_\delta - \frac{1}{3} \varrho |\vc{u}_\delta |^2 \mathbb{I} \right) : \nabla_x \Delta^{-1}_x
\nabla_x \left[ \varrho_\delta^\alpha \right] } \,{\rm d} t \\
&\qquad+
\left[ \intO{ \varrho_\delta \vc{u}_\delta \cdot \nabla_x \Delta^{-1}_x \left[ \varrho_\delta^\alpha - \intO{\varrho_\delta^\alpha } \right] } \right]_{t = T}^{t = T+1} \\
&\qquad+ \int_T^{T+1} \intO{ \varrho_\delta \vc{u}_\delta \cdot \nabla_x \Delta_x^{-1} \left[ {\rm div}_x (\varrho_\delta^\alpha \vc{u}_\delta ) + (\alpha - 1) \varrho_\delta^\alpha {\rm div}_x \vc{u}_\delta \right] } \,{\rm d} t \\
&\qquad- \sum_{k=1}^\infty\int_T^{T+1} \intO{ \vc{G}_k(\varrho_\d,\varrho_\d\vc{u}_\d) \cdot \nabla_x \Delta^{-1}_x \left[ \varrho_\delta^\alpha - \intO{\varrho_\delta^\alpha } \right] } \ {\rm d} W_k.
\end{split}
\end{equation}
Before proceeding, we make the assumption that $0 < \alpha < 1/3$, which implies in particular
\begin{equation*
\left| \intO{ \varrho_\delta^\alpha } \right| \leq c(M_0), \qquad \left\| \nabla_x\Delta_x^{-1} \left[ \varrho_\delta^\alpha - \intO{\varrho_\delta^\alpha } \right] \right\|_{L^\infty(\mathbb T^3;\mathbb{R}^3)} \leq c(M_0)
\end{equation*}
using H\"older's inequality, Sobolev's embedding and continuity of $\nabla_x\Delta_x^{-1}\nabla_x$.
Passing to expectations in (\ref{d4}) and keeping in mind that the processes are stationary we deduce
\begin{equation} \label{Dd5}
\begin{split}
&\expe{ \intO{ \bigg[ a \varrho_{\delta}^{\gamma + \alpha} + \delta \varrho_{\delta}^{\Gamma + \alpha} +
\varrho_{\delta}^{1 + \alpha} |\vc{u}_{\delta}|^2 \bigg]} } \\
& \leq
c(M_0) \left( \expe{ \intO{ \left[ \frac{1}{2} \varrho_\delta |\vc{u}_\delta|^2 + \frac{a}{\gamma - 1} \varrho^\gamma + \frac{\delta}{\Gamma - 1} \varrho_\delta^\Gamma \right] } } + 1 \right) \\
&\qquad+
\expe{ \intO{ \left( \frac{4}{3}\mu+\eta \right) {\rm div}_x \vc{u}_\delta \ \varrho_\delta^\alpha } } \\
& \qquad+ \expe{ \intO{ \left( \varrho_\delta \vc{u}_\delta \otimes \vc{u}_\delta - \frac{1}{3} \varrho_\delta |\vc{u}_\delta |^2 \mathbb{I} \right) : \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_\delta^\alpha \right] } } \\
&\qquad + \expe{ \intO{ \varrho_\delta \vc{u}_\delta \cdot \nabla_x \Delta^{-1}_x [ {\rm div}_x (\varrho_\delta^\alpha \vc{u}_\delta ) + (\alpha - 1) \varrho_\delta^\alpha {\rm div}_x \vc{u}_\delta ] }}.
\end{split}
\end{equation}
Moreover, using the uniform bound (\ref{d1'}) we may further reduce (\ref{Dd5}) to
\begin{equation*
\begin{split}
&\expe{ \intO{ \bigg[ a \varrho_{\delta}^{\gamma + \alpha} + \delta \varrho_{\delta}^{\Gamma + \alpha} +
\varrho_{\delta}^{1 + \alpha} |\vc{u}_{\delta}|^2 \bigg]} } \\
& \leq \expe{ \intO{ \left( \varrho_\delta \vc{u}_\delta \otimes \vc{u}_\delta - \frac{1}{3} \varrho_\delta |\vc{u}_\delta |^2 \mathbb{I} \right) : \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_\delta^\alpha \right] } } \\
&\qquad+ \expe{ \intO{ \varrho_\delta \vc{u}_\delta \cdot \nabla_x \Delta^{-1}_x [ {\rm div}_x (\varrho_\delta^\alpha \vc{u}_\delta ) + (\alpha - 1) \varrho_\delta^\alpha {\rm div}_x \vc{u}_\delta ] }} + c(G,M_0).
\end{split}
\end{equation*}
Note that we applied Young's inequality to the first and second term on the right-hand side of \eqref{Dd5} and in order to absorb the arising term eventually.
To control the remaining integrals on the right hand side, we first use H\" older's inequality to obtain
\begin{equation*
\begin{split}
&\left| \expe{ \intO{ \left( \varrho_\delta \vc{u}_\delta \otimes \vc{u}_\delta - \frac{1}{3} \varrho_\delta |\vc{u}_\delta |^2 \mathbb{I} \right) : \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_\delta^\alpha \right] } } \right|\\
&\leq c\,\expe{ \| \sqrt{\varrho_\delta} \vc{u}_\delta \|_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_\delta \|_{L^6(\mathbb T^3; \mathbb{R}^3)} \left\| \sqrt{\varrho_\d}\, \nabla_x \Delta^{-1} \nabla_x \left[ \varrho_\delta^\alpha \right] \right\|_{L^3(\mathbb T^3; \mathbb{R}^{3\times 3})} }\\
&\leq c
\left( \expe{\| \sqrt{\varrho} \vc{u}_\delta \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_\delta \|^2_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)}} +\expe{ \left\| \sqrt{\varrho_\d}\, \nabla_x \Delta^{-1} \nabla_x \left[ \varrho_\delta^\alpha \right] \right\|^2_{L^{3 }(\mathbb T^3; \mathbb{R}^{3\times 3})} } \right).
\end{split}
\end{equation*}
Furthermore, we have
\[
\left\| \sqrt{\varrho_\d}\, \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_\delta^\alpha \right] \right\|^2_{L^3(\mathbb{T}^3; \mathbb{R}^{3 \times 3})}
\leq \left\| \sqrt{\varrho_\delta} \right\|^2_{L^{2 \gamma}(\mathbb T^3)} \left\| \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_\delta^\alpha \right] \right\|^2_{L^q(\mathbb T^3; \mathbb{R}^{3 \times 3})}
\]
\[
\frac{1}{2 \gamma} + \frac{1}{q} = \frac{1}{3}, \ \gamma > \frac{3}{2}.
\]
Now, we choose $\alpha > 0$ so small that $\alpha q \leq 1$ to conclude that
\begin{equation*
\left\| \sqrt{\varrho_\d}\, {\rm div}_x \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_\delta^\alpha \right] \right\|^2_{L^3(\mathbb T^3; \mathbb{R}^{3 \times 3})}
\leq c(M_0) \|\varrho_\d\|_{L^\gamma(\mathbb{T}^3)} .
\end{equation*}
Similarly, we can handle
\begin{equation*
\begin{split}
&\left| \intO{ \varrho_\delta \vc{u}_\delta \cdot \nabla_x \Delta^{-1}_x {\rm div}_x [ \varrho_\delta^\alpha \vc{u}_\delta ] } \right|\\
& \leq
\| \sqrt{\varrho_\delta} \vc{u}_\delta \|_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \sqrt{\varrho_\delta} \|_{L^{2 \gamma}(\mathbb T^3)} \left\| \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_\delta^\alpha \vc{u}_\delta \right]
\right\|_{L^q(\mathbb T^3; \mathbb{R}^{3 \times 3})}
\end{split}
\end{equation*}
where
\[
\frac{1}{2} + \frac{1}{2 \gamma} + \frac{1}{q} = 1, \ \mbox{in particular}\ q < 6 \ \mbox{if}\ \gamma > \frac{3}{2},
\]
and where
\begin{equation*
\left\| \nabla_x \Delta^{-1}_x \nabla_x \left[ \varrho_{\delta}^\alpha \vc{u} \right]
\right\|_{L^q(\mathbb T^3; \mathbb{R}^{3\times 3})} \leq \| \varrho_{\delta}^\alpha \vc{u}_\delta \|_{L^q(\mathbb T^3; \mathbb{R}^3)} \leq \| \vc{u}_\delta \|_{L^6(\mathbb T^3; \mathbb{R}^3)} \| \varrho_{\delta}^\alpha \|_{L^s(\mathbb T^3)},
\ \frac{1}{6} + \frac{1}{s} = \frac{1}{q}.
\end{equation*}
Taking $\alpha s \leq 1$ we get, similarly to the above,
\begin{equation*
\begin{split}
&\left| \intO{ \varrho_{\delta} \vc{u}_\delta \cdot \nabla_x \Delta^{-1}_x {\rm div}_x [ \varrho_{\delta}^\alpha \vc{u}_\delta ] } \right|\\
& \leq c (M_0) \left( \| \sqrt{\varrho_{\delta}} \vc{u}_\delta \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)}
\| \vc{u}_\delta \|^2_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)} + \| \varrho_{\delta} \|_{L^\gamma(\mathbb T^3; \mathbb{R}^3)} \right).
\end{split}
\end{equation*}
Finally,
\[
\begin{split}
&\left| \intO{\varrho_{\delta} \vc{u}_\delta \cdot \nabla_x \Delta^{-1}_x [ \varrho_{\delta}^\alpha {\rm div}_x \vc{u}_\delta ] } \right|\\
&\leq \| \sqrt{\varrho_{\delta}} \|_{L^{2 \gamma}(\mathbb T^3; \mathbb{R}^3)} \| \sqrt{\varrho_{\delta}} \vc{u}_\delta \|_{L^2(\mathbb T^3; \mathbb{R}^3)}
\left\| \nabla_x \Delta^{-1}_x [\varrho_{\delta}^\alpha {\rm div}_x \vc{u}_\delta ] \right\|_{L^q(\mathbb T^3; \mathbb{R}^3)} \\
& \leq
\frac{1}{2} \left( \| \varrho_{\delta} \|_{L^\gamma(\mathbb T^3)} + \| \sqrt{\varrho_{\delta} \vc{u}} \|^2_{L^2(\mathbb T^3;\mathbb{R}^3)} \left\| \nabla_x \Delta^{-1}_x [\varrho_{\delta}^\alpha {\rm div}_x \vc{u}_\delta ] \right\|_{L^q(\mathbb T^3; \mathbb{R}^3)}^2\right),
\end{split}
\]
where
\[
\frac{1}{2\gamma} + \frac{1}{2} + \frac{1}{q} = 1, \ \ q < 6 \ \mbox{if}\ \gamma > \frac{3}{2}.
\]
As the $\nabla_x \Delta_x^{-1}$-operator gains one derivative, we get, by means of the standard Sobolev embedding,
\[
\left\| \nabla_x \Delta_x^{-1}[\varrho_{\delta}^\alpha {\rm div}_x \vc{u}_\delta ] \right\|_{L^q(\mathbb T^3; \mathbb{R}^3)} \leq \| \varrho_{\delta}^\alpha {\rm div}_x \vc{u}_\delta \|_{L^r(\mathbb T^3)}, \ r < 2.
\]
Thus, similarly to the previous steps, we may conclude that
\begin{equation*
\left| \intO{\varrho_{\delta} \vc{u}_{\delta} \cdot \nabla_x [ \varrho_{\delta}^\alpha {\rm div}_x \vc{u}_{\delta} ] } \right|
\leq c (M_0) \left( \| \sqrt{\varrho_{\delta}} \vc{u}_{\delta} \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)}
\| \vc{u}_{\delta} \|^2_{W^{1,2}_0(\mathbb T^3; \mathbb{R}^3)} + \| \varrho_{\delta} \|_{L^\gamma(\mathbb T^3; \mathbb{R}^3)} \right).
\end{equation*}
Summing up the above estimates we obtain
\begin{equation*
\begin{split}
&\expe{ \intO{ a \varrho_{\delta}^{\gamma + \alpha} + \delta \varrho_{\delta}^{\Gamma + \alpha}+\frac{1}{3} \varrho_{\delta}^{1 + \alpha} |\vc{u}_{\delta}|^2 } } \\
&\leq \expe{ \intO{ \left[ \frac{1}{2} \varrho_{\delta} |\vc{u}_{\delta}|^2 + \frac{a}{\gamma - 1} \varrho_{\delta}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\delta}^\Gamma \right] } \intO{ \| \vc{u}_{\delta} \|_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)}^2 } } + c(G,M_0),
\end{split}
\end{equation*}
where we absorbed the term $\| \varrho_{\delta} \|_{L^\gamma(\mathbb T^3; \mathbb{R}^3)}$ in the left-hand side.
We close the estimates by evoking (\ref{d51'}).
Thus we may conclude that any global in time stationary solutions admit the uniform bound \eqref{d13} as well as
\begin{equation} \label{d14}
\expe{ \left(1 + \intO{ \left[ \frac{1}{2} \varrho_{\delta} |\vc{u}_{\delta}|^2 + \frac{a}{\gamma - 1} \varrho_{\delta}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\delta}^\Gamma \right] } \right) \| \vc{u}_{\delta} \|^2_{W^{1,2} (\mathbb T^3; \mathbb{R}^3)} } \leq c(G,M_0).
\end{equation}
Finally, we claim that
\[
\expe{ \left( \intO{ \left[ \frac{1}{2} \varrho_{\delta} |\vc{u}_{\delta}|^2 + \frac{a}{\gamma - 1} \varrho_{\delta}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\delta}^{\Gamma} \right] } \right)^s } \leq
c
\]
for a certain $s = s(\alpha) > 1$. Indeed the $\varrho_{\delta}$-dependent terms can be estimated directly by (\ref{d13})
while, by H\" older inequality and Sobolev embedding,
\[
\begin{split}
\left( \intO{ \varrho_{\delta} |\vc{u}_{\delta} |^2 } \right)^s &\leq \left( \| \sqrt{\varrho_{\delta}} \vc{u}_{\delta} \|_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_{\delta} \|_{L^6(\mathbb T^3; \mathbb{R}^3)} \right)^s \| \sqrt{\varrho_{\delta}} \|_{L^{3}(\mathbb T^3)}^s\\
&\leq c \left[ \| \sqrt{\varrho_{\delta}} \vc{u}_{\delta} \|^2_{L^2(\mathbb T^3; \mathbb{R}^3)} \| \vc{u}_{\delta} \|_{W^{1,2}(\mathbb T^3; \mathbb{R}^3)}^2 + \left( \intO{ \varrho_{\delta}^\gamma } \right)^{\frac{1}{\gamma( 2 -s)}} \right].
\end{split}
\]
Note that we also took into account $\gamma>\frac{3}{2}$. This can be estimated by \eqref{d14} provided $s<2-\frac{1}{\gamma}$. The term with $\varrho_{\delta}^\gamma$ (and $\delta\varrho_{\delta}^\Gamma$) is estimated by Jensen's inequality.
Now we go back to the energy inequality (\ref{b2ac}). Due to \eqref{b4n} we obtain after taking the power $s$ and the supremum in time and expectation
\[
\begin{split}
&\expe{ \left( \sup_{t \in [T, T + \tau] } \intO{ \left[ \frac{1}{2} \varrho_{\delta} |\vc{u}_{\delta}|^2 + \frac{a}{\gamma - 1} \varrho_{\delta}^\gamma + \frac{\delta}{\Gamma - 1} \varrho_{\delta}^{\Gamma} \right] } \right)^s }+\expe{\bigg(\int_T^{T+\tau}\|\vc{u}_\d\|_{W^{1,2}(\mathbb{T}^3;\mathbb{R}^3)}^2\,{\rm d} t \bigg)^s} \\ &\leq c( G,M_0) + \expe{ \sup_{t \in [T, T + \tau] } \left|\sum_{k=1}^\infty\int_{T}^{t} \intO{ \vc{G}_k(\varrho_\d,\varrho_\d\vc{u}_\d) \cdot \vc{u}_{\delta} } \, {\rm d} W_k \right|^s } .
\end{split}
\]
{Note that all terms are well-defined by \eqref{d14}.}
Here, the second term on the right hand side is controlled by (\ref{b4bc}) and the Burkholder-Davis-Gundy inequality similarly to \eqref{e4} as follows
\begin{align*}
&\expe{ \sup_{t \in [T, T + \tau] } \left|\sum_{k=1}^\infty\int_{T}^{t} \intO{ \vc{G}_k(\varrho_\d,\varrho_\d\vc{u}_\d) \cdot \vc{u}_{\delta} } \, {\rm d} W_k \right|^s }\\
&\leq c(s,M_0)\,\expe{\bigg(\int_T^{T+\tau} \bigg(\intO{\frac{1}{2}\varrho_\d|\vc{u}_\d|^2}\bigg)^\frac{s}{2}\,{\rm d} t },
\end{align*}
which can be again estimated by \eqref{d13}.
We therefore conclude that \eqref{d151'} holds true.
for a.e. $T > 0$, where the constant depends on $\tau$ but it is independent of $T$.
\end{proof}
Finally, we have all in hand in order to complete the proof of Theorem \ref{Tm1}.
\begin{proof}[Proof of Theorem \ref{Tm1}]
We follow the lines of \cite[Section 6]{BrHo}. In view of Proposition \ref{prop:1808}, we are able to apply the Jakubowski-Skorokhod representation theorem and obtain convergence of $[\varrho_\d,\vc{u}_\d]$ (in fact, we obtain a new family of martingale solutions defined on a new probability space but keep the original notation for simplicity) to a stationary weak martingale solution of
\begin{equation*
\int_{0}^{\infty} \intO{ \left[ \varrho \partial_t \varphi + \varrho \vc{u} \cdot \nabla_x \varphi \right] } \,{\rm d} t = 0, \qquad \intO{ \varrho } = M_0,
\end{equation*}
for any $\varphi \in C^\infty_c((0, \infty) \times \mathbb{T}^3)$ $\mathbb{P}$-a.s.,
\begin{equation*
\begin{split}
\int_0^\infty \partial_t \psi \intO{ \varrho \vc{u} \cdot \varphi} \,{\rm d} t &+ \int_0^\infty\psi \intO{ \varrho \vc{u} \otimes \vc{u} : \nabla_x \varphi } \,{\rm d} t
+ \int_0^\infty \psi\intO{ a \Ov{\varrho^\gamma} {\rm div}_x \varphi } \,{\rm d} t \\
&- \int_0^\infty \psi\intO{ \mathbb{S}(\nabla_x \vc{u}) : \nabla_x \varphi} \,{\rm d} t =- \int_0^\infty \psi {\rm d} M_\varphi
\end{split}
\end{equation*}
for any $\psi \in C^\infty_c((0, \infty))$, $\varphi \in C^\infty(\mathbb{T}^3; \mathbb{R}^3)$ $\mathbb{P}$-a.s. Here $M_\varphi$ is a square integrable martingale and the bars denote the corresponding weak limits.
In addition, $\varrho$ satisfies the renormalized equation of continuity.
In order to identify the nonlinear density dependent terms, we keep Remark \ref{rem:1808} in mind and apply the effective viscous flux method in the same way as in \cite[Section 6.1--6.3]{BrHo}, which then completes the proof.
Note that similarly to Section \ref{L}, even the limited moment estimates from Proposition \ref{prop:1808} are sufficient for the passage to the limit.
\end{proof}
|
\section{Introduction}\label{sec:Introduction}
\input{sec_introduction.tex}
\section{Symmetry principles of quantum liquid crystals}\label{sec:Symmetry principles of quantum liquid crystals}
\input{sec_symmetry.tex}
\section{Preliminaries}\label{sec:Preliminaries}
\input{sec_preliminaries.tex}
\section{Dual elasticity in three dimensions}\label{sec:Dual elasticity in three dimensions}
\input{sec_dualelasticity.tex}
\section{Dislocation-mediated quantum melting}\label{sec:Dislocation-mediated quantum melting}
\input{sec_quantummelting.tex}
\section{Nematic phases}\label{sec:nematic}
\input{sec_nematic.tex}
\section{Smectic phases}\label{sec:smectic}
\input{sec_smectic.tex}
\section{Columnar phases}\label{sec:columnar}
\input{sec_columnar.tex}
\section{Charged quantum liquid crystals}\label{sec:Charged quantum liquid crystals}
\input{sec_charged.tex}
\section{Conclusions}\label{sec:Conclusions}
\input{sec_conclusions.tex}
\begin{acknowledgments}
A.J.B. is supported by the MEXT-Supported Program for the Strategic Research Foundation at Private Universities ``Topological Science'' (Grant No. S1511006). This work was supported by the Netherlands foundation for Fundamental Research of Matter (FOM).
\end{acknowledgments}
\subsection{Charged dual elasticity}\label{subsec:Charged dual elasticity}
In QLC2D we have shown how to incorporate the electromagnetic interactions in quantum elasticity, leading to the coarse-grained long-wavelength description of the Wigner crystal. This qualifies to be textbook material. One profits here optimally from the stress representation of elasticity, yielding a highly transparent description. Photons are the carriers of the EM force and one would better dualize the phonons into stress photons so that apples are compared to apples. The outcome is a simple linear mode-coupling affair.
The starting point is the coupling between displacements $u^a$ and the electromagnetic field, which is derived in QCL2D to be of form,
\begin{equation}
S_\mathrm{EM} = - \int \mathrm{d} \tau \mathrm{d}^D x \; j^\imath_\mu A^\imath_\mu \equiv - \int \mathrm{d} \tau \mathrm{d}^D x \; \mathcal{A}^a_\mu \partial_\mu u^a,\label{eq:EM coupling}
\end{equation}
where
\begin{align}
j^\imath_\mathfrak{t} &= - \mathrm{i} n e^* c_\mathrm{T} \partial_m u^m, &
j^\imath_m &= \mathrm{i} n e^* c_\mathrm{T} \partial_\mathfrak{t} u^m,\\
A^\imath_\mathfrak{t} &= \mathrm{i} \frac{1}{c_\mathrm{T}} V, &
A^\imath_m &= A_m,
\end{align}
and defining the vector potential that couples to displacements,
\begin{align}
\mathcal{A}^a_\mu &= n e^* c_\mathrm{T} (\delta_{\mu a} A^\imath_\mathfrak{t} - \delta_{\mu \mathfrak{t}} A^\imath_a).
\end{align}
Here the label $^\imath$ denotes that the temporal components are rescaled by a factor of $\mathrm{i}$ to get Euclidean products in imaginary time, e.g. $V^\imath = - \mathrm{i} V$ for the scalar Coulomb potential. Furthermore $n$ is the density of (charged) particles and $e^*$ is the charge of the constituent particles (e.g., $e^*= 2e$ for Cooper pairs of electrons).
In the EM coupling term, the usual electromagnetic gauge invariance $A^{\imath}_{\mu} \to A^{\imath}_{\mu}+ \partial_{\mu} \lambda$ is equivalent to conservation of particle number $\partial_{\mu} j^{\imath}_{\mu} = 0$, which we already identified as the glide constraint, i.e. the conservation of charge for a charged elastic medium. Therefore EM gauge invariance is guaranteed in the path integral for the dual gauge fields via the glide constraint, even in the dislocation-condensed phases.
For a charged elastic medium, the EM interaction term Eq.~\eqref{eq:EM coupling} simply has to be added
to the original quantum-elasticity action Eq.~\eqref{eq:solid action}.
After this, the EM field $A_\mu$ will just be `carried along' in the strain-stress duality transformation explained in Sec.~\ref{sec:Dual elasticity in three dimensions}.
In this way, the EM field $A_\mu$ will just be `carried along' in the duality. This results in an expression describing a simple linear coupling between the stress fields and the EM gauge fields~\cite{QLC2D}. The dual Lagrangian is
\begin{align}
\mathcal{L}^\mathrm{EM}_\mathrm{dual} &= \mathcal{L}^\mathrm{neutral}_\mathrm{dual} + \mathcal{L}_\mathrm{Meissner} + \mathcal{L}_\mathrm{int} + \mathcal{L}_\mathrm{Maxw},\label{eq:general dual EM Lagrangian}\\
\mathcal{L}_{\rm Meissner}&=\tfrac{1}{2} \mathcal{A}^{a \dagger}_m C^{-1}_{mn ab} \mathcal{A}^b_n + \frac{1}{2} \frac{1}{\mu}\mathcal{A}^{a \dagger}_\mathfrak{t} \mathcal{A}^a_\mathfrak{t}, \nonumber\\
&= \tfrac{1}{2} \varepsilon_0\omega_\mathrm{p}^2 \Big( (A^\imath_a)^2 + \tfrac{\mu}{\kappa} (A^\imath_\mathfrak{t})^2 \Big).\label{eq:Meissner in hiding}\\
\mathcal{L}_\mathrm{int} &=
- \sigma^a_m C^{-1}_{mn a b} \mathcal{A}_n^b - \frac{1}{\mu} \sigma^a_\mathfrak{t} \mathcal{A}^a_\mathfrak{t} \nonumber\\
&= -\frac{n e^* c_\mathrm{T}}{D \kappa} \sigma^a_a A^\imath_\mathfrak{t} + \frac{ne^* c_\mathrm{T}}{\mu} \sigma^a_\mathfrak{t} A_a. \label{eq:EM-stress interaction}
\end{align}
Here $\mathcal{L}^\mathrm{neutral}_\mathrm{dual}$ is just the stress Lagrangian of the neutral elastic medium, for instance Eq.~\ref{eq:dual solid Lagrangian}, and $\mathcal{L}_\mathrm{Maxw} = \frac{1}{4\mu_0} (\partial_\mu A_\nu - \partial_\nu A_\mu)^2$, where $\mu_0$ is the magnetic constant. The term $\mathcal{L}_\mathrm{Meissner}$, which has the form of a Meissner term giving a mass to the photon field, arises automatically in the duality construction; in the solid it is exactly canceled by elastic terms while in the liquid crystals it can remain to cause the real Meissner effect indicative of superconductivity. In this sense, the Meissner effect is already ``lying in wait'' in the crystal to become manifest when shear rigidity is destroyed~\cite{ZaanenNussinovMukhin04,QLC2D}.
Eq.~\eqref{eq:EM-stress interaction} is valid in all dimensions. However, when we wish to express the stress tensors in terms of dual stress gauge fields, in 3+1D one encounters the two-form fields:
$\sigma^a_\mu = \epsilon_{\mu\nu\kappa\lambda}\partial_\nu \tfrac{1}{2} b^a_{\mu\nu}$. Upon substitution one finds,
\begin{align}
\mathcal{L}_\mathrm{int} &= \tfrac{1}{4} b^{a\dagger}_{\mu\nu} g^a_{\mu\nu,\lambda} A^\imath_\lambda + \tfrac{1}{4} A^{\imath\dagger}_\lambda g^{a\dagger}_{\lambda,\mu\nu} b^a_{\mu\nu}\nonumber\\
&= \tfrac{1}{4}\tilde{b}^{a\dagger}_{\mu\nu} \tilde{g}^a_{\mu\nu,\lambda} A^\imath_\lambda + \tfrac{1}{4} A^{\imath\dagger}_\lambda \tilde{g}^{a\dagger}_{\lambda,\mu\nu} \tilde{b}^a_{\mu\nu}
\end{align}
where $g^{a\dagger}_{\lambda,\mu\nu} = (g^a_{\mu\nu,\lambda})^*$, and as explained in Eq.~\eqref{eq:tilde fields definition} the variables with a tilde are rescaled with the dislocation velocity $c_\mathrm{d}$. The $\tilde{g}^{a}_{\lambda,\mu\nu}$ are momentum-dependent coefficients tabulated in matrix form as
\begin{widetext}
\begin{align}
\tilde{b}^{a\dagger}_{\mu\nu} \tilde{g}^a_{\mu\nu,\lambda} A^\imath_\lambda &=
\begin{pmatrix} \tilde{b}^{ \dagger}_{1-} & \tilde{b}^{\mathrm{L}\dagger}_{\mathrm{R} \mathrm{S}} &
\tilde{b}^{\mathrm{R} \dagger}_{\mathrm{R} \mathrm{S} } & \tilde{b}^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}}
\end{pmatrix}
ne^* \begin{pmatrix}
-\frac{\sqrt{2}}{D\kappa} c_\mathrm{d} \tilde{p} & 0 & 0 & 0 \\
-\mathrm{i} \frac{1}{D\kappa} \omega_n & \frac{1}{\mu} c_\mathrm{T} q & 0 & 0 \\
0 & 0 & \frac{1 }{\mu} c_\mathrm{T} q & 0 \\
0& 0 & 0 & \frac{1}{\mu} c_\mathrm{T} q
\end{pmatrix}
\begin{pmatrix}
A_\mathfrak{t} \\ A_\mathrm{L} \\ A_\mathrm{R} \\ A_\mathrm{S}
\end{pmatrix}.\label{eq:b-A coupling matrix tilde}
\end{align}
\end{widetext}
This follows the pattern set in the 2+1D case: collecting the two-form indices $\mu,\nu$ in a vector, the coupling to the EM field is captured by just a matrix, and is linear. The other components of $b^a_{\mu\nu}$ do not couple to $A_\mu$. From this form, we can immediately infer that the $b^\mathrm{R}_{1\mathrm{R}}$--$b^\mathrm{S}_{1\mathrm{S}}$ sector ($\mathrm{T} 3$) does not couple to electromagnetic fields. The `longitudinal'
rotational Goldstone mode of the quantum nematic with its vanishing propagator in the solid resides in this sector and is therefore completely invisible to electromagnetic means. The matrices $g^a_{\mu\nu,\lambda}$ are all that is necessary to calculate the EM response given the dual stress propagators of the neutral elastic medium, as we shall see now.
To this end, observe that the general form of the stress part of the Lagrangian is
$\mathcal{L} = \tfrac{1}{8} \tilde{b}^{a\dagger}_{\mu\nu} ( G^{-1})^{ab}_{\mu\nu,\kappa\lambda} \tilde{b}^b_{\kappa\lambda}$. Define the matrix $G$ via $(G^{-1})^{ab}_{\mu\nu,\kappa\lambda} G^{bc}_{\nu\rho,\lambda\sigma} = \delta_{ac}\delta_{\mu\rho}\delta_{\kappa\sigma}$. The stress gauge fields can now be integrated out, and we learn how these dress the electromagnetic fields:
\begin{align}\label{eq:stress contribution to EM fields}
\mathcal{L}_{\rm para} &= \tfrac{1}{8} \tilde{b}^{a\dagger}_{\mu\nu} (G^{-1})^{ab}_{\mu\nu,\kappa\lambda} \tilde{b}^b_{\kappa\lambda} + \tfrac{1}{4} \tilde{b}^{a\dagger}_{\mu\nu} \tilde{g}^a_{\mu\nu,\lambda} A^{\imath}_\lambda + \tfrac{1}{4} A^{\imath\dagger}_\lambda \tilde{g}^{\dagger a}_{\lambda,\mu\nu} \tilde{b}^a_{\mu\nu} \nonumber\\
&= - \tfrac{1}{8} A^{\imath \dagger}_\rho \tilde{g}^{a\dagger}_{\rho,\mu\nu} G^{ab}_{\mu\nu,\kappa\lambda} \tilde{g}^b_{\kappa \lambda,\sigma} A^\imath_\sigma.
\end{align}
This must be added to the `diamagnetic' or `Meissner' contribution already present in the dual Lagrangian Eq.~\eqref{eq:Meissner in hiding}. As usual, the shear velocity $c_\mathrm{T}^2 = \mu / \rho$ while the plasmon frequency $\omega_\mathrm{p}$ is defined as
\begin{equation}
\omega_\mathrm{p}^2 = \frac{(n e^*)^2}{\rho \varepsilon_0}.
\end{equation}
Finally, $\varepsilon_0$ is the dielectric constant in units of $[\varepsilon_0] = \frac{\mathrm{C}^2}{\mathrm{J}\;\mathrm{m}}$.
Using Eq.~\eqref{eq:stress contribution to EM fields} the only extra input that is needed are the stress propagators $G^{ab}_{\mu\nu,\kappa\lambda}$ for the {\em neutral} elastic medium and we can directly compute the
electromagnetic propagator $\langle A^\dagger_\mu A_\nu \rangle$ enumerating the electromagnetic response, where the elastic medium just translates into the usual photon self-energy.
The effective EM action due to the medium is of the form $\mathcal{L}^{\mathrm{EM}}_{\mathrm{medium}} = \mathcal{L}_\mathrm{para} + \mathcal{L}_\mathrm{Meissner}$, to be added to the vacuum Maxwell action which in our units reads:
\begin{equation}
\mathcal{L}_\mathrm{Maxw} = \tfrac{1}{2}\varepsilon_0 c_\mathrm{T}^2 q^2 \lvert A^\imath_\mathfrak{t} \rvert^2 + \tfrac{1}{2}\varepsilon_0 (\omega_n^2 + c_l^2 q^2) \big( \lvert A^\imath_\mathrm{R} \rvert^2 + \lvert A^\imath_\mathrm{S} \rvert^2 \big).
\label{eq:Maxwell action}
\end{equation}
Here we have taken the Coulomb gauge fix for the EM field, removing $A_\mathrm{L}$. The shear velocity $c_\mathrm{T}$ shows up because of our definition of the temporal components, and $c_l$ is the speed of light. The full effective EM action $\mathcal{L}^\mathrm{EM}_{\mathrm{eff,medium}} = \mathcal{L}_\mathrm{Maxw} + \mathcal{L}^{\mathrm{EM}}_{\mathrm{medium}}$ and it is straightforward to compute the full portfolio of EM response from this action, as we will now show.
With the knowledge that the electromagnetic response can be straightforwardly derived from the neutral elastic propagator, we can translate these results into observable quantities. We will work in real time and real frequencies $\omega$ for the remainder of this section. Given the photon propagator that follows from $\mathcal{L}^\mathrm{EM}_{\mathrm{eff,medium}}$, we can define the photon self-energy $\Pi_{mn}(\omega,q)$ via:
\begin{align}
\langle A^{\dagger}_m(\omega, q) A_n(-\omega,-q) \rangle = \frac{1}{\varepsilon_0}\frac{1}{\omega^2 - c_l^2 q^2 -\Pi_{mn} (\omega, q)}.
\label{photonpropdef}
\end{align}
For the diagonal components of the self-energy we write $\Pi_m = \Pi_{mm}$ (no sum).
The conductivity tensor $\hat{\sigma}_{mn} (\omega,q)$ and the dielectric function $\hat{\varepsilon}_{mn}(\omega,q)$ are defined as:
\begin{align}
\hat{\sigma}_{mn} (\omega,q) &= -\mathrm{i} \omega \left(\hat{\varepsilon}_{mn}(\omega,q) -\varepsilon_0\right),\nonumber\\
&= \varepsilon_0\frac{\mathrm{i}}{\omega} \Pi_{mn} (\omega,q).
\label{eq:optical conductivity}
\end{align}
Furthermore, in QLC2D we showed it is useful to define an energy- and momentum-dependent penetration depth $\lambda_m(\omega,q)$ which characterizes the screening (exponential decay) of the photon component $A_m$ in the medium, which is either due to the skin effect or to the Meissner effect, as follows:
\begin{equation} \label{eq:penetration depth from self-energy}
\lambda_m(\omega,q) = \frac{c_l}{\mathrm{Im} \sqrt{\omega^2 - \Pi_{mm}}}.
\end{equation}
When the penetration depth stays finite in the limit $\omega \to 0$ we will interpret this as a genuine Meissner effect indicating the presence of superconductivity.
For the longitudinal EM response, we calculate the photon propagator in the Coulomb gauge to find from Eq.~\eqref{eq:stress contribution to EM fields} and \eqref{eq:stress gauge field components} that $ \langle A^{\dagger}_{\mathfrak{t}}(\omega,q) A_{\mathfrak{t}}(-\omega,-q) \rangle_\mathrm{para} = \varepsilon_0 \omega_\mathrm{p}^2 \frac{1}{(D \kappa)^2}\langle \sigma^a_a \; \sigma^b_b \rangle$. In fact the total propagator including the Meissner term is simply proportional to the longitudinal propagator. This is immediately derived from Eq.~\eqref{eq:EM coupling}, and we find~\cite{QLC2D}
\begin{equation}
\langle A^{\dagger}_{\mathfrak{t}}(\omega,q) A_{\mathfrak{t}}(-\omega,-q) \rangle = \varepsilon_0 \omega_\mathrm{p}^2 \mu G_\mathrm{L}.
\end{equation}
This implies that the longitudinal EM response just coincides with the longitudinal elastic propagator of the medium. For this reason we shall not address this response much further below. We only wish to emphasize that the special features of quantum liquid crystals, such as the appearance in the longitudinal dielectric function of a second, gapped pole due to the dislocation condensate next to the ordinary plasmon, is only noticeable at finite momentum. All these poles have vanishing spectral weight for momentum going to zero. For this reason, finite-momentum spectroscopy such as {\em electron energy-loss spectroscopy} (EELS) is the only way to observe these features. In fact, new machines which have proper resolution at finite momentum seem to be coming online at present~\cite{VigEtAl15}.
For the transverse EM response, all the quantities in Eqs.~\eqref{eq:optical conductivity},\eqref{eq:penetration depth from self-energy} are simply related. We shall give results in the form of poles of the photon propagator itself, the transverse conductivity or the penetration depth, whichever provides the clearest picture.
\subsection{The electrodynamics of the Wigner crystal}
As we already alluded to above, with our initial assumptions, the theory of quantum elasticity of charged matter describes a generic isotropic ``Wigner'' crystal formed from charged bosonic constituents. Nevertheless, in the ordered crystalline phase, the wisdoms in this section are entirely general: they apply as well to a `conventional' Wigner crystal of electrons as to any other medium of this kind. Surely, the dualization can be carried out only for a crystal formed out of bosons. In the condensed-matter context one could envisage a crystal formed from ``preformed'' Cooper pairs that are subsequently subjected to quantum melting. This may be of relevance to e.g. the `stripy' charge order found in cuprate superconductors~\cite{FradkinKivelsonTranquada15,BergFradkinKivelson09,HamidianEtAl16}.
In QLC2D we derived the electrodynamics of the isotropic Wigner crystal using the duality transformation detailed in Sec.~\ref{subsec:Charged dual elasticity}. It is a rather classic subject but to our surprise we found several
novelties, it seems all related to the remarkably efficiency of the stress formalism. The longitudinal channel is according to expectations: the longitudinal phonon acquires a plasmon gap --- it is not quite the standard plasmon
because it is still propagating with the longitudinal phonon velocity which is larger than the sound velocity of a liquid ($c_\mathrm{L}$ vs. $c_\kappa$). The longitudinal optical conductivity is characterized by an infinitely sharp Drude peak at zero frequency, indicating
that the Wigner crystal is a perfect conductor (but not a superconductor, see below). This has to be the case since our elastic medium is supposed to live in a perfect Galilean continuum where all matter at finite density is perfectly conducting since momentum is conserved. In impurity language: the crystal as a whole is `unpinned' and can be `set in motion' by an infinitesimal external potential.
The surprises are in the transverse response. The finite-frequency EM photon exerts a shear force on the crystal which responds by a reactive restoring force encapsulated by the transverse phonons/stress photons
expelling the photon: the Wigner crystal is characterized by a skin
depth $\lambda(\omega,\mathbf{q})$. What is the difference with a superconductor where the same photon ``acquires a mass''? The stress formalism exhibits an elegant view on this affair. As we already announced in the previous section, a literal Meissner term is present in the effective action: Eq.~\eqref{eq:Meissner in hiding}. However, in order to determine the response of the medium there is a second term due to the stress photons in Eq.~\eqref{eq:stress contribution to EM fields}. One finds that at precisely zero
frequency $\omega \to 0$, this produces a term that exactly cancels the bare Meissner term. In the DC limit the penetration (skin) depth becomes infinite and static magnetic fields can penetrate the whole solid. In the quantum liquid crystals on the other hand, the transverse stress photons acquire a mass because of the dual dislocation condensate. The effect is that the cancellation is no longer complete and static magnetic fields are expelled, proving that these are indeed conventional superconductors as well exhibiting a regular Meissner effect.
Last but not least, what is happening to the transverse phonons of the neutral crystal when we couple in electromagnetism? The transverse motions are surely exempted from the plasmon mechanism, but these do correspond to fluctuating electrical dipoles. Surprisingly, the dispersion of the `dressed' transverse phonons turns out to become {\em quadratic} in the charged crystal. Keeping in mind that we consider $D+1$-dimensional electromagnetism coupled to a $D+1$-dimensional medium one expects that these behaviors should not depend on dimensionality. This is indeed the case, as we now will demonstrate. Consider the 3+1D elasticity Lagrangian
Eq.~\eqref{eq:solid stress gauge field Lagrangian} and the $G^{ab}_{\mu\nu,\kappa\lambda}$ follow immediately by inverting the Gaussian kernel. By substituting these in Eq.~\eqref{eq:stress contribution to EM fields}we obtain
$\mathcal{L}_\mathrm{para}$. Adding $\mathcal{L}_\mathrm{Meissner}$ from Eq.~\eqref{eq:Meissner in hiding} the contribution of the elastic medium to the EM action is obtained which reads in the Coulomb gauge ($A_\mathrm{L} = 0$),
\begin{align}
\mathcal{L}^\mathrm{EM}_\mathrm{solid}
&= \tfrac{1}{2}\varepsilon_0 \omega_\mathrm{p}^2 \Big( \frac{c_\mathrm{T}^2 q^2}{\omega_n^2 + c_\mathrm{L}^2 q^2} \lvert A^\imath_\mathfrak{t} \rvert^2 \nonumber\\
&\phantom{mmmmmm}+ \frac{\omega_n^2 }{\omega_n^2 + c_\mathrm{T}^2 q^2} \big( \lvert A^\imath_\mathrm{R}\rvert^2 +\lvert A^\imath_\mathrm{S}\rvert^2\big) \Big).
\end{align}
Taking into account the Maxwell action Eq.~\eqref{eq:Maxwell action} the full photon propagator is straightforwardly obtained. This yields two degenerate transverse propagators $\langle A^{\imath\dagger}_\mathrm{R}\; A^\imath_\mathrm{R} \rangle$ and $\langle A^{\imath\dagger}_\mathrm{S}\; A^\imath_\mathrm{S} \rangle$. These are like in 2+1D characterized by two poles with dispersion relations at small momentum:
\begin{align}
\omega_{1} & = \sqrt{ (\omega^2_\mathrm{p} + (c_l^2+c^2_\mathrm{T})q^2} \simeq \sqrt{\omega_\mathrm{p}^2+c_l^2 q^2} + \mathcal{O}(q^2),\nonumber \\
\omega_{2} & = \frac{c_\mathrm{T} c_l q^2}{\sqrt{ (\omega^2_\mathrm{p} + (c_l^2+c^2_\mathrm{T})q^2)}} \simeq \frac{c_l c_\mathrm{T}}{\omega_\mathrm{p}} q^2 + \mathcal{O}(q^4).
\label{Xtallongwave}
\end{align}
assuming $c_l \gg c_\mathrm{T}$. The first mode is the familiar plasma-polariton characterized the plasmon energy $\omega_\mathrm{p}$. The second mode is massless and characterized by a quadratic dispersion, the surprise
we already encountered in 2+1D. The take home message is that except for the fact that there are two photon polarizations 3+1D (instead a single one in 2+1D) everything else is independent of dimensionality.
\subsection{The superconducting nematic}\label{subsec:The superconducting nematic}
This dimensionality-independence of the electromagnetic effects extends also to the quantum liquid crystals. We refer to QCL2D for a detailed discussion of the electromagnetism in the quantum nematics. As we
already discussed above, a highlight is the way that the Meissner effect shows up, unambiguously proving that the charged dual stress superconductor is at the same time a regular superconductor. This important matter is easy to verify to hold in 3+1D as well. The penetration depth can be computed from Eq.~\eqref{eq:penetration depth from self-energy}. The full expression is lengthy but
the main interest is in the $\omega =0$ limit that determines whether we are dealing with a Meissner phase:
\begin{equation}
\lambda_{\mathrm{R},\mathrm{S}}^\mathrm{nem} (0,q) = \frac{c_l}{\omega_\mathrm{p}} \sqrt{1 + 2 c_\mathrm{T}^2 q^2/\Omega^2} \equiv \lambda_\mathrm{L} \sqrt{1 + \lambda_\mathrm{s}^2 q^2}.
\end{equation}
Here we have defined the London penetration depth as $\lambda_\mathrm{L} = c_l / \omega_\mathrm{p}$ and the `shear penetration depth' as $\lambda_\mathrm{s} = \sqrt{2} c_\mathrm{T} / \Omega$. At $q= 0$, the penetration depth is
exactly $\lambda_\mathrm{L}$, demonstrating that the quantum nematic is a superconductor. These results are identical to those obtained in 2+1D\cite{QLC2D}. Since the quantum nematic is isotropic, the result holds for both $\mathrm{R},\mathrm{S}$-polarizations of the photon field.
There is a lot more going on, such as the way that the massive shear photons acquire electromagnetic weight and thereby becoming measurable in principle by electromagnetic experiments. However, this is by and large rather independent
of dimensionality as we just discussed for the Wigner crystal itself.
These propagating massive modes are of course special to our ``maximally-correlated limit''. This is different for the massless modes which behave universally, independent of the degree of the microscopic correlations. We already saw this at work in the Wigner crystal: the quadratic dispersion of the dressed transverse phonon has to apply universally. Besides the sound mode, the neutral quantum nematic is characterized by the rotational Goldstone bosons/torque modes
as we demonstrated in Sec.~\ref{subsec:Collective modes in the quantum nematic}. As we repeatedly emphasized the sound mode acts as the phase mode of the superconductor ``eating the photon'', and this works in the same way as in 2+1D. However, more is going on in the rotational Goldstone sector in 3+1D as compared to the 2+1D case. Instead of the single transverse torque mode of the latter, in three dimensions symmetry just imposes that there are three modes. As we discussed, these are polarized like phonons: there are two degenerate transverse modes but in addition there is now also a `longitudinal rotational phonon'. The qualitative novelty is that this longitudinal mode does {\em not} couple to electromagnetic fields, as we mentioned
below Eq.~\eqref{eq:b-A coupling matrix tilde}. Imagining that our quantum liquid crystals would be formed from electrons, this has the interesting consequence that this mode cannot be observed directly since the only way to apply external forces is through its electrical charge. The two 'transverse rotational phonons' are coupled to the two transverse photon polarizations, and show up as massless poles in the photon propagator, identical to the 2+1D case. For instance, in the optical conductivity one finds a zero-frequency Drude peak, the massless rotational phonon and a gapped shear mode~\cite{QLC2D}.
\begin{figure*}[t]
\includegraphics[width=7cm]{smecticxy_meissner_R}
\hspace{1cm}
\includegraphics[width=7cm]{smecticxy_meissner_S}
\caption{Penetration depth in the smectic (solid black lines) at $q=0.5$ as a function of energy $\omega$ for $\eta = 0$, left: $\mathrm{R}$-polarization; right: $\mathrm{S}$-polarization. For comparison we have plotted the same quantity for the Wigner crystal (dashed red line).
For $\eta = 0$, the $\mathrm{R}$-polarization is within the liquid plane, and we find a finite penetration depth at $\omega \to 0$, indicative of the Meissner effect. Conversely, the $\mathrm{S}$-polarization is perpendicular to the liquid plane, and the penetration depth appears only at finite energy, like the screening of AC magnetic fields in an ordinary conductor (skin effect). Nevertheless, the screening in this direction is stronger resulting in a shorter penetration depth than that of the Wigner crystal. The smectic shows very anisotropic superconductivity. As an aside, the divergence in the left plot is the resonance of the rotational Goldstone mode --- in general all poles of the photon propagator also show up in this `dynamic' penetration depth.
}\label{fig:xy smectic Meissner}
\end{figure*}
\subsection{2D superconductivity in the smectics}\label{subsec:2D superconductivity in the smectics}
The story of the above repeats itself in the smectic and the columnar phases. In QCL2D we analyzed the electromagnetic responses of the 2+1D quantum smectics in detail, finding out that the already intricate mixture of solid-like and
liquid-like behaviors gets further enriched when electromagnetism is added. Like for the solid and the nematic, the 3+1D case is by and large a further variation on the same theme, where the novelties are related to the fact that we now are dealing with either a `2+1D solid' with an `extra quantum-liquid direction' (the columnar phase), or a `2+1D quantum liquid' with an `extra solid direction': the smectic phase. There is a lot to explore here but we have not found anything that is qualitatively new. We leave it therefore for follow-up work in the future.
However, there is one aspect that deserves closer consideration given that it is highly relevant for the gross physics of these systems. As we discussed in Sec.~\ref{sec:smectic}, the liquid planes of the 3+1D smectics behave in many regards like 2+1D quantum nematics. The important implication is that in these directions the charged quantum smectics should be like two-dimensional superconductors while a Wigner crystal, i.e. perfect metal, behavior should occur in the `solid direction'. Let us demonstrate here that is indeed the case.
Let us first focus on the transverse conductivity, given by Eq.~\eqref{eq:optical conductivity}. We use the same simplifications as in Sec.~\ref{sec:smectic} and we first consider momentum within the liquid plane, i.e. $\eta = 0$. The transverse conductivities are independent of the angle $\zeta$ within the plane due to the axial symmetry and read:
\begin{align}
\hat{\sigma}_{\mathrm{R}} &= \mathrm{i} \omega_\mathrm{p}^2 \frac{1}{\omega} \frac{ \omega^2 (\omega^2 - c_\mathrm{d}^2 q^2) - \Omega^2 (\omega^2 - \frac{1}{2} c_\mathrm{d}^2 q^2)}{(\omega^2 - c_\mathrm{T}^2 q^2 )(\omega^2 - c_\mathrm{d}^2 q^2) - \Omega^2 (\omega^2 - \frac{1}{2} c_\mathrm{d}^2 q^2)},\\
\hat{\sigma}_{\mathrm{S}} &= \mathrm{i} \omega_\mathrm{p}^2 \frac{ \omega (\omega^2 - c_\mathrm{d}^2 q^2 - \Omega^2)}{(\omega^2 - c_\mathrm{T}^2 q^2 )(\omega^2 - c_\mathrm{d}^2 q^2) - \frac{1}{2}\Omega^2 \omega^2 }.
\end{align}
This gives away the mixed character of the smectic. The first line addresses the response within the liquid plane and it is identical to that of the nematic showing three poles: the $\omega =0$ perfect conductor Drude peak, the gapped shear mode and the massless rotational Goldstone mode. The second line is the response orthogonal to the liquid plane. The transverse Drude peak is lacking, instead showing the presence of the gapped condensate mode and in addition a quadratically dispersing undulation mode.
What happens to the superconductivity? We computed the penetration depth for photons with polarization $\mathrm{R}$ (in the liquid plane) and $\mathrm{S}$ (orthogonal to the liquid plane), respectively, and the result is plotted in Fig.~\ref{fig:xy smectic Meissner}. For the $\mathrm{R}$-polarization we find a finite penetration depth at $\omega =0$, and therefore the system is in a Meissner phase. On the other hand, the penetration depth for the $\mathrm{S}$-polarization only sets in at finite energy as is usual for the ordinary, metallic skin effect, such that static magnetic fields can penetrate. This gives away that this smectic behaves like a strictly 2D superconductor where photon polarizations are screened in the liquid direction but not along
the solid direction. Nevertheless, the skin effect for the $\mathrm{S}$-polarization is enhanced as compared to the Wigner crystal with no dislocation condensate at all.
The anisotropic superconductivity can also be studied as function of the interrogation angle $\eta$, showing that again there is only a Meissner response precisely in the liquid plane. Upon tilting away from the planar direction the penetration depth diverges at zero frequency, while the skin depth decreases rapidly at finite frequencies. All of this confirms our earlier statements, that superfluidity and superconductivity exist whenever shear rigidity is lost. A similar effect takes place in the columnar phases, where there are one-dimensional superconducting lines.
\subsection{Higgs term and glide constraint}
The starting point is Eq.~\eqref{eq:dislocation Higgs term} where the sum over $a$ is removed and we select only a single Burgers direction $x$, which can be done without loss of generality when starting from an isotropic solid.
As before, we choose the dislocation Lorenz gauge fix Eq.~\eqref{eq:dislocation Lorenz gauge fix} and disregard the decoupled condensate phase degrees of freedom. As for the glide constraint, following the patterns of the earlier sections, we find the contributions
\begin{align}
\mathcal{L}^{(x)}_\mathrm{glide} = - 2 \frac{\Omega^2}{4 c_\mathrm{T}^2 \mu} \frac{ \tilde{p}^2}{\frac{1}{c_\mathrm{d}^2} \omega_n^2 + q_x^2 } \left\lvert \tilde{b}^x_{yz} \right\rvert^2,
\end{align}
and similar for the $y$- and $z$-condensates.
The Higgs term for the columnar $x$-condensates is:
\begin{align}
\mathcal{L}^{(x)}_\mathrm{Higgs}
&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu} \Big[
\lvert \tilde{b}^x_{1\mathrm{R}} \rvert^2 + \lvert \tilde{b}^x_{1\mathrm{S}} \rvert^2 + \lvert \tilde{b}^x_{\mathrm{R}\mathrm{S}} \rvert^2 \nonumber\\
&\phantom{mmnmm}
- \frac{\tilde{p}^2}{ \frac{1}{c_\mathrm{d}^2} \omega_n^2 + q_x^2} \lvert \tilde{b}^x_{yz}\rvert^2 \Big],
\label{eq:dislocation Higgs term columnar x}
\end{align}
and similar permutations of $x,y,z$ for the other choices $y$ and $z$.
\begin{figure}
\includegraphics[width=8cm]{columnar_velocity.pdf}
\caption{Velocities of the massless modes in the columnar phase as a function of interrogation angle $\eta$ from Eq.~\eqref{eq:columnar L T1 massless dispersion}. Here we have chosen a representative value of the Poisson ration $\nu = 0.3$. The velocity of the longitudinal mode varies from that of the longitudinal phonon $c_\mathrm{L}$ at $\eta =0 ,\pi/2$ while it is reduced to $c_\mathrm{i} = \sqrt{c_\mathrm{L}^2 - c_\mathrm{T}^2}$ at $\eta =\pi/4$. The velocity of the transverse mode is maximal at $\eta =\pi/4$ where it reaches that of the transverse phonon $c_\mathrm{T}$, while it is reduced to zero at $\eta =0,\pi/2$, where it also has vanishing pole strength. This pattern is identical to that of the 2D smectic~\cite{QLC2D}. }\label{fig:columnar velocity}
\end{figure}
\subsection{Collective modes in the quantum columnar phases}
As mentioned before, departing from an isotropic solid there is axial symmetry around the liquid direction, and therefore we can capture the response of the columnar phase by choosing translational symmetry restoration in the $x$-direction, and setting the azimuthal angle $\zeta= 0$ without loss of generality. The angle $\eta$ again interpolates between momentum completely in the liquid ($\eta =0$) and completely in the solid direction ($\eta = \pi/2$).
Like for the smectics, the general form of the propagators is complicated and it is not insightful to write them down explicitly. We shall study them through the dispersion relation of their poles and the visualization of their spectral functions.
We find that for general angle $\eta$, the sectors $G_{\mathrm{L}}$ and $G_{\mathrm{T} 1}$, and the sectors $G_{\mathrm{T} 2}$ and $G_{\mathrm{T} 3}$ are mixed and share the same poles (although the pole strengths are different). For $G_{\mathrm{L}}$ and $G_{\mathrm{T} 1}$ there is one massive pole and two massless poles, with dispersion relations,
\begin{align}
(\omega^{\mathrm{L}, \mathrm{T} 1}_{1,2})^2 &= \frac{1 - \nu \pm \sqrt{\nu^2 - (1-\nu) \sin^2 2 \eta}}{-1 + 2\nu} c_\mathrm{T} q + \ldots ,\label{eq:columnar L T1 massless dispersion}\\
(\omega^{\mathrm{L}, \mathrm{T} 1}_3)^2 &= \frac{1}{2}\Omega^2 + \big(c_\mathrm{T}^2 + c_\mathrm{d}^2 \cos^2 \eta) \big) q^2 + \ldots .
\end{align}
The last mode is related to gapped shear from the liquid direction while the massless modes interpolate between the longitudinal and transverse phonons in the solid-like directions. The velocity of these massless poles as a function of $\eta$ at a representative value of $\nu = 0.3$ is plotted in Fig.~\ref{fig:columnar velocity}. It mimics the 2D smectic~\cite{QLC2D}: the longitudinal pole has maximum velocity at $\eta = 0,\pi/2$ and minimum velocity $c_{\rm min} = \sqrt{c_\mathrm{L}^2 - c_\mathrm{T}^2}$ at $\eta = \pi/4$. The transverse pole has vanishing pole strength and vanishing velocity at $\eta =0,\pi/2$ and a maximum velocity of $c_\mathrm{T}$ at $\eta =\pi/4$. As we mentioned in QLC2D, remarkably at $\eta = \pi/4$ the ultimate `solid $\times$ liquid' behavior is attained: the shear mode has the full transverse velocity like a transverse phonon, while the longitudinal mode has the pure compression velocity $c^\mathrm{2D}_\kappa = \sqrt{ (c^\mathrm{2D}_\mathrm{L})^2 - c_\mathrm{T}^2}$.
The propagators $G_{\mathrm{T} 2}$ and $G_{\mathrm{T} 3}$ share their poles, although the pole strengths differ. They each have one massive and one massless pole, with dispersion relations:
\begin{align}
\omega^{\mathrm{T} 2,3}_1 &= \sin \eta\; c_\mathrm{T} q + \ldots ,\\
\omega^{\mathrm{T} 2,3}_{2} &= \frac{1}{\sqrt{2}} \Omega + \frac{c_\mathrm{d}^2 + \cos^2 \eta\; c_\mathrm{T}^2}{\sqrt{2} \Omega} q^2 + \ldots .
\end{align}
The first mode is the second transverse phonon, the velocity of which vanishes at $\eta = 0$ and $\eta =\pi/2$, see below. The gapped mode is a condensate mode `partially' coupled with the transverse phonon depending on the angle $\eta$.
\begin{table*}
\begin{tabular}{clclc}
\multicolumn{3}{c}{general $\eta$}\\
\toprule
sector & & massless & &massive \\
\hline
$\mathrm{L}$ & \rdelim\}{2}{1mm}[] & \multirow{2}{*}{Eq.~\eqref{eq:columnar L T1 massless dispersion}} & \rdelim\}{2}{1mm}[] & \multirow{2}{*}{$\frac{1}{\sqrt{2}}\Omega + \frac{\cos^2 \eta\; c_\mathrm{d}^2 + c_\mathrm{T}^2 }{\sqrt{2}\Omega} q^2$} \\
$\mathrm{T} 1$ & & & & \\
$\mathrm{T} 2$ &\rdelim\}{2}{1mm}[] & \multirow{2}{*}{$\sin \eta \;c_\mathrm{T} q$} & \rdelim\}{2}{1mm}[]& \multirow{2}{*}{$\frac{1}{\sqrt{2}}\Omega + \frac{c_\mathrm{d}^2 + \cos^2 \eta\; c_\mathrm{T}^2}{\sqrt{2}\Omega} q^2$} \\
$\mathrm{T} 3$ & & & & \\
\hline
total & & 3 & & 2 \\
\botrule
\end{tabular}
\hfill
\begin{tabular}{cccc}
\multicolumn{3}{c}{$\eta=0$}\\
\toprule
sector & massless & &massive \\
\hline
$\mathrm{L}$ & $c_\mathrm{L} q$ & & - \\
$\mathrm{T} 1$ &$\frac{1}{\sqrt{2}\Omega} c_\mathrm{d} c_\mathrm{T} q^2$ & & $\frac{1}{\sqrt{2}}\Omega + \frac{c_\mathrm{d}^2 + c_\mathrm{T}^2}{\sqrt{2} \Omega} q^2 $\\
$\mathrm{T} 2$ &$\frac{1}{\sqrt{2}\Omega} c_\mathrm{d} c_\mathrm{T} q^2$ & & $\frac{1}{\sqrt{2}}\Omega + \frac{c_\mathrm{d}^2 + c_\mathrm{T}^2}{\sqrt{2} \Omega} q^2 $\\
$\mathrm{T} 3$ & - & & - \\
\hline
total & 3 & & 2 \\
\botrule
\end{tabular}
\hfill
\begin{tabular}{cccc}
\multicolumn{3}{c}{$\eta=\pi/2$}\\
\toprule
sector & massless & &massive \\
\hline
$\mathrm{L}$ & $c_\mathrm{L} q$ & & - \\
$\mathrm{T} 1$ & - & & $\frac{1}{\sqrt{2}} \Omega + \frac{c_\mathrm{T}}{\sqrt{2} \Omega} q $ \\
$\mathrm{T} 2$ & $c_\mathrm{T} q$ & & - \\
$\mathrm{T} 3$ & - & & $\frac{1}{\sqrt{2}} \Omega + \frac{c_\mathrm{d}}{\sqrt{2} \Omega} q$ \\
\hline
total & 2 & & 2 \\
\botrule
\end{tabular}
\caption{Collective modes in the columnar phases. Indicated are the dispersion relations to lowest orders in momentum. For translational symmetry restoration in the $x$-directions, $G_\mathrm{L}$ and $G_{\mathrm{T} 1}$ share their poles as do $G_{\mathrm{T} 2}$ and $G_{\mathrm{T} 3}$. The first pair contain two massless modes which extrapolate to the a longitudinal and transverse phonon, and a gapped shear mode, which is identical to the 2+1D smectic. The second pair contain a second transverse phonon whose velocity increases smoothly from zero to $c_\mathrm{T}$ as $\eta$ goes from $0$ to $\pi/2$. There is also a gapped mode which has contribution from the dislocation condensate as well as from gapped shear. At $\eta =0$ we find two undulation modes in the $G_{\mathrm{T} 1,2}$-sectors, while $G_{\mathrm{T} 3}$ vanishes. At $\eta = \pi/2$, there is a 2D solid in the ${\mathrm{L}}$--${\mathrm{T} 2}$-sectors, while the responses in the other two sectors are gapped.}\label{table:columnar spectrum}
\end{table*}
For the special angles $\eta = 0$ and $\eta = \pi/2$ the propagators obtain a simple form. For momentum in the liquid direction $\eta =0$, we find that $G_{\mathrm{L}}$ is like a pure solid, Eq.~\eqref{eq:longitudinal propagator}, while $G_{\mathrm{T} 1}$ and $G_{\mathrm{T} 2}$ both take exactly the smectic form Eq.~\eqref{eq:GT1 eta0 smectic}. The third transverse propagator $G_{\mathrm{T} 3}$ vanishes identically. The columnar phase looks just like a 2D smectic since the two transverse phonons are gapped in same way due to the axial symmetry. For momentum in the solid direction $\eta = \pi/2$, $G_\mathrm{L}$ and $G_{\mathrm{T} 2}$ are the pure solid longitudinal resp. transverse propagators Eqs.~\eqref{eq:longitudinal propagator},\eqref{eq:transverse propagator}, while $G_{\mathrm{T} 1}$ is again identical to the smectic and is a pure gapped shear mode as in Eq.~\eqref{eq:smectic pi/2 gapped shear}. The third sector $G_{\mathrm{T} 3}$ contains a pure gapped condensate mode. Indeed, we do not find a rotational Goldstone mode here, as expected.
\subsection{Dislocation worldsheets}\label{subsec:Dislocation worldsheets}
Before we dualize the action of the solid, let us first discuss dislocation lines in the imaginary time setting of the quantum theory. The dislocation density $J^a_n(x)$ in Eq.~\eqref{eq:dislocation density definition} is a static quantity. In spacetime, the dislocation line along $n$ with Burgers vector $a$ can move in direction $\mu$ where $\mu$ contains both temporal and spatial components. The dislocation line $J^a_n$ traces out a {\em worldsheet} $J^a_{\mu\nu}$ in spacetime, see Fig.~\ref{fig:dislocation worldsheet}. The density of the line is represented by $J^a_{tn} = J^a_n$ and the flow or current in direction $m$ of the line along $n$ is represented by $J^a_{mn}$. The worldsheet element $J^a_{\mu\nu}(x)$ at $x$ is a two-form quantity in the differential geometry sense, cf. Sec.~\ref{subsec:Two-form gauge fields}, and is antisymmetric in its indices $\mu,\nu$. The {\em dislocation worldsheet element} is defined by
\begin{equation}\label{eq:dislocation current definition}
J^a_{\mu\nu} (x) = \epsilon_{\mu\nu\kappa\lambda} \partial_\kappa \partial_\lambda\tfrac{1}{2} u^a(x).
\end{equation}
Here $\epsilon_{\mu\nu\kappa\lambda}$ is the completely antisymmetric Levi-Civita symbol and $\epsilon_{txyz} = 1$. This is the 3+1D generalization of Eq.~\eqref{eq:dislocation density definition}. Note that the Burgers vector $a$ is always spatial, since fields are always smooth in the time direction. Therefore Lorentz symmetry is still badly broken; close to the critical point, we are at most dealing with an emergent relativistic theory where the `speed of light' is actually a material speed such as the phonon velocity $c_\mathrm{T}$. We will call $J^a_{\mu\nu}$ the {\em dislocation current} that couples to the dual stress gauge field $b^a_{\mu\nu}$ defined below, in analogy with the particle current $j_\mu$ sourcing a vector gauge field $A_\mu$ in Maxwell electrodynamics. By definition $J^a_{t n}$ is an edge dislocation if $a \neq n$ and a screw dislocation if $a = n$. Note that because of the antisymmetry in the lower indices $J^a_{an}$ (no sum) equivalently represents the current in direction $a$ of an edge dislocation or the opposite of the current in direction $n$ of a screw dislocation.
\begin{figure}
\includegraphics[width=7cm]{dislocation_worldsheet.png}
\caption{Section of a dislocation worldsheet, depicting a dislocation line (red) moving in time. In blue is represented the surface element $J^a_{\mu\nu}(X)$ at point $X$ with Burgers vector $B^a$.}\label{fig:dislocation worldsheet}
\end{figure}
In the bulk of the solid and in absence of disclinations the dislocation current is conserved:
\begin{equation}\label{eq:dislocation conservation}
\partial_\mu J^a_{\mu\nu} = 0 \quad \forall a,\nu.
\end{equation}
This implies that a dislocation line cannot begin or end in the material: dislocation lines must be closed loops, and dislocation worldsheets must be closed surfaces.
The worldsheet picture of defect lines can be very useful even in condensed matter physics. For instance, we derived all important dynamical electromagnetic effects of Abrikosov vortices in superconductors by regarding them as worldsheets in Ref.~\onlinecite{BeekmanZaanen11}.
In Sec.~\ref{subsec:Dislocations and disclinations} we mentioned the {\em glide constraint} which states that edge dislocations can only move in the direction of their Burgers vector. The precise statement in terms of the dislocation currents was derived in Ref.~\onlinecite{CvetkovicNussinovZaanen06}:
\begin{equation}\label{eq:glide constraint}
\epsilon_{tamn} J^a_{mn} = 0.
\end{equation}
Recall that $J^a_{mn}$ is the current in direction $m$ of the dislocation line along $n$ with Burgers vector $a$. Eq.~\eqref{eq:glide constraint} states that edge dislocations ($a\neq n$) cannot move in the direction orthogonal to the Burgers vector. For screw dislocations ($a = n$) there is no such constraint: their motion perpendicular to the dislocation line does not involve addition or removal of constituent particles. The glide constraint is in fact a consequence of conservation of particle number. This can be seen by inserting the definition Eq.~\eqref{eq:dislocation current definition}:
\begin{align}
0 = \epsilon_{tamn} \epsilon_{mn\kappa\lambda} \partial_\kappa \partial_\lambda u^a(x) = (\partial_\mathfrak{t} \partial_a - \partial_a \partial_\mathfrak{t}) u^a = 0.
\end{align}
To lowest order, fluctuations of the mass density $\rho(x)$ are $\rho_0 \partial_a u^a$ while the mass current is $j_a(x) = \rho_0 \partial_t u^a$. Thus the glide constraint is equivalent to the conservation law $\partial_\mathfrak{t} \rho + \partial_a j_a =0$~\cite{CvetkovicNussinovZaanen06,QLC2D}. The glide constraint is active during dislocation condensation, and amazingly turns out to protect the compression mode, in turn related to the conservation law, from obtaining a dual Higgs mass, see Sec.~\ref{sec:Dislocation-mediated quantum melting}.
There is a similar generalization for the disclination worldsheet, defined by
\begin{equation}\label{eq:disclination worldsheet}
\Theta^c_{\mu\nu} (x) = \epsilon_{\mu \nu \kappa\lambda } \partial_\kappa \partial_\lambda \tfrac{1}{2} \epsilon^{cab} \omega^{ab}(x).
\end{equation}
The interpretation is that $\Theta^c_{tn}(x)$ is the disclination density at $x$ of the line along $n$ with rotational plane orthogonal to $c$, while $\Theta^c_{mn}(x)$ is the flow or current of that line in direction $m$. For $\Theta^c_{tn}(x)$, if $n = c$ it is the density of a segment of a {\em wedge} disclination while it is the density of an {\em twist} disclination if $n \neq c$. A closed disclination line will typically be of wedge or twist nature at different positions.
In the presence of disclinations, the dislocation current is no longer conserved. Instead of Eq.~\eqref{eq:dislocation conservation} we have~\cite{Kleinert89b,QLC2D}
\begin{equation}\label{eq:dislocation disclination interdepence}
\partial_\mu J^a_{\mu\nu} = \epsilon_{\mathfrak{t} abc} \Theta^c_{b\nu}.
\end{equation}
This equation implies that a disclination line can source dislocations. If $\nu = \mathfrak{t}$, then the left-hand side is the divergence of dislocation density, which can only be non-zero if the dislocation line ends. In other words, a dislocation line can end on a static twist disclination line. If $\nu = n$, the right-hand side denotes the current or flow of a disclination line. Then this equation implies that a moving disclination leaves dislocations in its wake~\cite{QLC2D,KlemanFriedel08}. This equation is a consequence of the fact that translations and rotations are not independent. In the space group the point group operations (including rotations) are in semidirect relation with the translations; locally, a rotation is equivalent to two {\em finite} translations, which in topological context turns into the statement that disclinations can be formed from a finite density of dislocations with equal Burgers vector. As we will discuss in more detail later, the liquid crystals can be topologically defined
by insisting that disclinations are massive (i.e. absent) which in turn implies that even locally the Burgers vectors have to be antiparallel in the dislocation condensate, since a finite `Burgers vector magnetization' is the same as a finite disclination density in the vacuum.
Since we are treating temporal and spatial dimensions on the same footing, a velocity is needed to compare quantities with different units. Certainly, we are in the idealized limit devoid of interstitials, disorder and other influences that could
dissipate the motion of the phonons and the topological defects. Everything moves ballistically without scattering or drag. In the vortex--boson duality~\cite{Franz07,BeekmanSadriZaanen11}, upon approaching the quantum critical point, scale invariance sets in as well as emergent Lorentz invariance. There can be only one velocity governing both sides of the phase transition: the velocity associated with the vortices coincides precisely with the phase velocity of the superfluid. In the same vein, the velocity associated with the defect-condensate is also the phase velocity. In the present context of elasticity, the same argument holds in principle, but there is a complication in the form of the glide constraint which restricts the motion of edge dislocations. Since screw dislocations do not suffer from this, Friedel argued that the dislocation speed should equal the material speed, given by the transverse or shear velocity $c_\mathrm{T}$~\cite{Friedel64}. In real-world solids, there is some evidence that edge dislocations can move `transonically' with speeds up to the longitudinal velocity $c_\mathrm{L}$, see e.g. Ref.~\onlinecite{RuestesEtAl15}. A difficulty is that the arguments for emergent Lorentz invariance become precise near the
continuous quantum phase transition, while deep in the solid `irrelevant' operators such as the nature of the chemical bond may become important; e.g. dislocations in covalent solids are immobile while simple metals are malleable
because of the rather isotropic nature of their electronic binding forces. Weak--strong dualities acquire their universal meaning in any case only close to the continuous quantum phase transition where one should become insensitive
to microscopic details. In principle, the scale of the characteristic velocity of the dislocations and the dislocation condensate is therefore assumed to be the transverse phonon velocity and the only complication arises from the glide
constraint. This also suggests that the velocity of edge dislocations $c_\mathrm{e}$ and of screw dislocations $c_\mathrm{s}$ can a priori be different.
In QLC2D we found that it is actually very helpful to treat the velocity of dislocations as different from the shear velocity, since it enables one to track the degrees of freedom originating in the dislocation condensate. On the other hand, below in Sec.~\ref{subsubsec:Dislocation worldsheet condensation} we will find that differing velocities for edge and screw dislocations considerably complicate the computations. Although our formalism is able in principle to handle the general case, with the exception of the end of Sec.~\ref{subsubsec:Dislocation worldsheet condensation} we will set $c_\mathrm{e} = c_\mathrm{s} \equiv c_\mathrm{d}$ as the uniform dislocation velocity, which in turn should be of the order of the shear velocity $c_\mathrm{T}$.
\subsection{Two-form gauge fields}\label{subsec:Two-form gauge fields}
Given a conserved current in four dimensions $j_\mu$, the associated conservation law (continuity equation) $\partial_\mu j_\mu = 0$ can be imposed by expressing the current as the four-curl of a two-form gauge field $b_{\kappa\lambda}$ (sometimes called Kalb--Ramond field~\cite{KalbRamond74})
\begin{equation}\label{eq:two-form gauge field definition}
j_\mu(x) = \epsilon_{\mu\nu\kappa\lambda} \partial_\nu \tfrac{1}{2} b_{\kappa\lambda} (x).
\end{equation}
The only relevant components of $b_{\kappa\lambda}$ are those antisymmetric in $\kappa,\lambda$, and this defines an antisymmetric two-form field. By expressing physical quantities in terms of $b_{\kappa\lambda}$ the constraint $\partial_\mu j_\mu=0$ is thereby identically satisfied. In addition, the field $b_{\kappa\lambda}$ is a gauge field since the addition of the gradient of an arbitrary smooth vector field $\varepsilon_\lambda(x)$ leaves the current $j_\mu$ in Eq.~\eqref{eq:two-form gauge field definition} invariant:
\begin{equation}\label{eq:two-form gauge transformations}
b_{\kappa\lambda} (x) \to b_{\kappa\lambda}(x) + \partial_\kappa \varepsilon_\lambda(x) - \partial_\lambda \varepsilon_\kappa(x).
\end{equation}
Let us now focus on the counting of the physical, propagating degrees of freedom of a two-form gauge field. First, recall that the number of propagating degrees of freedom of a one-form gauge field $A_\mu(x)$ (e.g. the photon field in electromagnetism) equals the number of spatially transverse components. In $D$ space dimensions, there is one longitudinal and $D-1$ transverse directions. Accordingly, such a one-form field represents one physical `photon'
in 2+1 dimensions and two such degrees of freedom in 3+1 dimensions. A two-form gauge field has instead two indices, and the components representing physical, propagating degrees of freedom (`photons') are the ones for which both indices are transverse. In 3+1 dimensions, there are two transverse directions, and since the field is antisymmetric in its two indices, there is a only one independent component with purely transverse indices. In $D$ space dimensions,
the number of propagating physical degrees of freedom is
\begin{equation}
\begin{pmatrix} D-1 \\ 2 \end{pmatrix} = \frac{(D-1)!}{2! (D-3)!} = \frac{(D-1)(D-2)}{2}.
\end{equation}
This result is only valid for $D \ge 3$. In particular, in 3+1 dimensions, the field $b_{\kappa\lambda}$ represents only a single physical `photon'. As we will see later, each phonon independently dualizes into a two-form gauge field,
corresponding to one such a physical propagating `stress photon'.
Another way to establish the correct number of degrees of freedom is to consider the gauge volume due to the gauge transformations Eq.~\eqref{eq:two-form gauge transformations}. The field $b_{\mu\nu}$ in 3+1 dimensions has six independent components due to its antisymmetry. However, the components that transform under gauge transformations do not correspond to physical degrees of freedom, and can be removed by a suitable gauge fixing. Since adding the gradient of an arbitrary scalar field $\varepsilon_\mu(x) \to \varepsilon_\mu(x) + \partial_\mu \xi(x)$ leads to the exact same gauge transformation Eq.~\eqref{eq:two-form gauge transformations}, there is a redundancy within the specification of $\varepsilon_\mu$. This is sometimes called ``gauge-in-the-gauge''. According to the theory of constraints in dynamical systems~\cite{HenneauxTeitelboim92}, the gauge transformation field $\varepsilon_\mu$ consists of three gauge components $\varepsilon_m$ and their time derivatives to function as 6 gauge parameters to be used for gauge fixing. However, due to the redundancy just mentioned we can fix only $6-1=5$ components, leaving one component as a propagating degree of freedom.
\subsection{Stress gauge fields}\label{subsec:Stress gauge fields}
The original displacement field $u^a$ still features in Eq.~\eqref{eq:dual action} as a result of the Legendre transformation~\cite{QLC2D}. In order to eliminate it in favor of the stresses, we need to perform the path integral over this field by properly taking into account the defects. Topological defects are singularities in the displacement field, see Sec.~\ref{subsec:Dislocations and disclinations}. Therefore we split $u^a$ in a smooth and a singular part:
\begin{equation}\label{eq:displacement field smooth singular}
u^a(x) = u^a_\mathrm{smooth} + u^a_\mathrm{sing}.
\end{equation}
On the smooth part we are allowed to perform integration by parts, to subsequently integrate out $u^a_\mathrm{smooth}$ in the path integral as a Lagrange multiplier for the constraint
\begin{equation}\label{eq:stress conservation}
\partial_\mu \sigma^a_\mu=0.
\end{equation}
This constraint corresponds with the conservation of stress --- external stresses will appear on the RHS. This can be implemented
by expressing the stress tensor as the four-curl of a two-form gauge field, see Sec~\ref{subsec:Two-form gauge fields}:
\begin{equation}\label{eq:dual stress gauge field definition}
\sigma^a_\mu = \epsilon_{\mu\nu\kappa\lambda} \partial_\nu \tfrac{1}{2} b^a_{\kappa\lambda}.
\end{equation}
At this point we specialized to 3+1 dimensions; in 2+1D one would be dealing instead with one-form stress gauge fields, see QLC2D. We call $b^a_{\mu\nu}$ the {\em dual stress gauge field}. As before, $b^a_{\kappa\lambda}$ is antisymmetric in $\kappa \leftrightarrow \lambda$,
while the factor of $\tfrac{1}{2}$ is inserted for later convenience.
The stress tensor $\sigma^a_\mu$ (the `field strength') is invariant under the set of gauge transformations, cf. Eq.~\eqref{eq:two-form gauge transformations},
\begin{equation}\label{eq:stress gauge transformation}
b^a_{\kappa\lambda} (x) \to b^a_{\kappa\lambda}(x) + \partial_\kappa \varepsilon^a_\lambda(x) - \partial_\lambda \varepsilon^a_\kappa(x).
\end{equation}
Here $\varepsilon^a_\lambda$ are three independent arbitrary vector fields, one for each $a = x,y,z$. Transforming to the Fourier-Matsubara coordinates $(0,1,\mathrm{R},\mathrm{S})$ for the spacetime indices (see appendix~\ref{sec:Fourier space coordinate systems}), the components $\sigma^a_0$ are removed by stress conservation. Furthermore, the remaining components are in one-to-one relation with the stress gauge field components:
\begin{align}
\sigma^a_\mathfrak{t} &= q b^a_{\mathrm{R}\mathrm{S}}, &
\sigma^a_\mathrm{L} &= \mathrm{i} \frac{\omega_n}{c_\mathrm{T}} b^a_{\mathrm{R} \mathrm{S}}, &
\sigma^a_\mathrm{R} &= p\; b^a_{1 \mathrm{S}}, &
\sigma^a_\mathrm{S} &= p\; b^a_{1\mathrm{R}}. \label{eq:stress gauge field components}
\end{align}
The reason is that the components $(\mathrm{R}\mathrm{S})$, $(1\mathrm{R})$, $(1\mathrm{S})$ are gauge-invariant, while all other components $(01)$, $(0\mathrm{R})$, $(0\mathrm{S})$ are pure gauge and do not contribute the stress tensor in Eq.~\eqref{eq:dual stress gauge field definition}. This is equivalent to imposing the Lorenz gauge fix $\partial_\mu b^a_{\mu\nu} = 0\ \forall\, \nu,a$.
We still have to deal with the singular displacement field $u^a_\mathrm{sing}$. Using the definition Eq.~\eqref{eq:dual stress gauge field definition} we derive,
\begin{align}\label{eq:stress dislocation minimal coupling}
\mathrm{i} \sigma^a_\mu \partial_\mu u^a_\mathrm{sing} &= \mathrm{i} (\epsilon_{\mu\nu\kappa\lambda} \partial_\nu \tfrac{1}{2} b^a_{\kappa\lambda})( \partial_\mu u^a_\mathrm{sing})\nonumber\\
&= \mathrm{i} b^a_{\kappa\lambda}\tfrac{1}{2}\epsilon_{\mu\nu\kappa\lambda} \partial_\mu \partial_\nu u^a_\mathrm{sing} = \mathrm{i} b^a_{\kappa\lambda} J^a_{\kappa\lambda}.
\end{align}
Here we performed integration by parts on $b^a_{\kappa\lambda}$ which is smooth everywhere, and used the definition of the dislocation worldsheet Eq.~\eqref{eq:dislocation current definition}.
It follows that the topological defects source the dual gauge fields, though these sources are now worldsheets the gauge fields themselves are of the two-form kind. In 2+1D
dislocations are instead `particles' (worldlines in 2+1D) and just source the stress gauge fields in the same way as electrically charged particles source electromagnetic gauge fields.
As discussed at length in QLC2D, the essence of stress--strain duality is that one translates the way that the elastic medium vibrates (phonons) into its capacity to propagate forces between the dislocations which are the natural
internal sources of stress. In the language of forces, the theory just takes the shape of a gauge theory: in this sense, phonons turn into literal (stress) photons. This stress gauge theory formulation is highly convenient in the
further development of the duality since it revolves around matter formed from the dislocations, and the gauge theory is the easy way to describe the interactions between the dislocations. In 2+1D quantum elasticity turns
into a `flavored' form of electromagnetism, but that is a coincidence characteristic for this dimension. A scalar field theory in 3+1D is not dual to a one-form gauge field (electromagnetism); instead we find the dual two-form gauge theory. Furthermore we just learned that it counts the physical degrees of freedom correctly (``1 phonon $=$ 1 stress photon'') at the same time doing justice to the fact that the sources of internal stress are worldsheets, Eq.~\eqref{eq:stress dislocation minimal coupling}.
For systems with one broken symmetry generator (like the broken $U(1)$-symmetry of superfluid), there is a single two-form field, as explained in Sec.~\ref{subsec:Two-form gauge fields}. In solids, there are $D$ broken generators when breaking $\mathbb{R}^D \to \mathbb{Z}^D$, and accordingly there are $D$ phonons turning into $D$ `flavors' of two-form gauge fields. This goes hand-in-hand with the topological charge of dislocations, the Burgers vector $B^a$, which is clearly not simply an integer winding number. The symmetry between the flavor index $a$ and the spacetime indices $\kappa$, $\lambda$ in $b^a_{\kappa\lambda}$ is far from perfect, and in fact it is often useful to regard them as completely separated. However, as can be inferred from Eq.~\eqref{eq:dual solid Lagrangian}, there is some `mixing' between the Burgers and spacetime `sectors'.
Finally, let us rewrite the partition sum Eq.~\eqref{eq:dual partition sum} in terms of the dual gauge fields:
\begin{align}
\mathcal{Z}_\mathrm{solid} = \int \mathcal{D}b^a_{\kappa\lambda} \mathcal{D} J^a_{\kappa\lambda} \mathcal{F}(b^a_{\kappa\lambda})\; \mathrm{e}^{-\int \mathrm{d} \tau \mathrm{d}^3x\; \mathcal{L}_\mathrm{dual}}.
\end{align}
Here $\mathcal{F}(b^a_{\kappa\lambda})$ is a gauge-fixing factor enforcing for instance a Lorenz gauge fix $\mathcal{F}(b^a_{\mu\nu}) = \delta(\partial_{\mu} b^a_{\mu\nu})$. At this stage of the development,
the action $\mathcal{L}_{\rm dual}$ describes describes the quantum mechanics of a isolated dislocations and
$\mathcal{D} J^a_{\kappa\lambda}$ denotes the sum over worldsheet elements of closed worldsheets associated with dilute dislocation--antidislocation loops on the time slice. The summation over the dislocation worldsheets in the path-integral cannot be performed and the action $\mathcal{L}_{\rm dual}$ should not be confused with the description of a dense system (`foam' in spacetime) of dislocations which is the subject of string field theory
More specifically, $\mathcal{L}_\mathrm{dual}$ is the sum of Eqs.~\eqref{eq:solid stress Lagrangian components} and \eqref{eq:second gradient Lagrangian in stress tensor} in which we have substituted Eq.~\eqref{eq:stress gauge field components} to obtain the contributions:
\begin{widetext}
\begin{align}
\mathcal{L}_\mathrm{dual} &= \mathcal{L}_{\mathrm{T} 1} + \mathcal{L}_{\mathrm{T} 2} + \mathcal{L}_{\mathrm{T} 3} + \mathcal{L}_{\mathrm{L} 1} + \mathcal{L}_{\mathrm{L} 2} + \mathrm{i} b^{a\dagger}_{\kappa\lambda} J^a_{\kappa\lambda},\label{eq:solid stress gauge field Lagrangian}\\
\mathcal{L}_{\mathrm{T} 1} &= \frac{1}{8\mu}
\begin{pmatrix}b^{\mathrm{L}\dagger}_{1 \mathrm{S}} & b^{\mathrm{R}\dagger}_{\mathrm{R}\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
p^2(1 + \frac{1}{\ell^2 q^2}) & \mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n p(1 - \frac{1}{\ell^2 q^2}) \\
-\mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n p(1 - \frac{1}{\ell^2 q^2}) & \frac{1}{c_\mathrm{T}^2}\omega_n^2(1 + \frac{1}{\ell^2 q^2}) + 4 q^2
\end{pmatrix}
\begin{pmatrix}b^{\mathrm{L}}_{1 \mathrm{S}} \\ b^{\mathrm{R}}_{\mathrm{R}\mathrm{S}} \end{pmatrix},\label{eq:L1} \\
\mathcal{L}_{\mathrm{T} 2} &= \frac{1}{8\mu}
\begin{pmatrix}b^{\mathrm{L}\dagger}_{1 \mathrm{R}} & b^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
p^2(1 + \frac{1}{\ell^2 q^2}) & \mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n p(1 - \frac{1}{\ell^2 q^2}) \\
-\mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n p(1 - \frac{1}{\ell^2 q^2}) & \frac{1}{c_\mathrm{T}^2}\omega_n^2(1 + \frac{1}{\ell^2 q^2}) + 4 q^2
\end{pmatrix}
\begin{pmatrix}b^{\mathrm{L}}_{1 \mathrm{R}} \\ b^{\mathrm{S}}_{\mathrm{R}\mathrm{S}} \end{pmatrix}, \\
\mathcal{L}_{\mathrm{T} 3}&= \frac{1}{8\mu}
\begin{pmatrix}b^{\mathrm{R}\dagger}_{1 \mathrm{R}} & b^{\mathrm{S}\dagger}_{1\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
p^2(1 + \frac{1}{\ell^2 q^2}) & p^2(1 - \frac{1}{\ell^2 q^2}) \\
p^2(1 - \frac{1}{\ell^2 q^2}) & p^2(1 + \frac{1}{\ell^2 q^2})
\end{pmatrix}
\begin{pmatrix}b^{\mathrm{R}}_{1 \mathrm{R}} \\ b^{\mathrm{S}}_{1 \mathrm{S}} \end{pmatrix},\label{eq:L3}\\
\mathcal{L}_{\mathrm{L} 1} &= \frac{1}{8\mu} \frac{2}{1+\nu}
\begin{pmatrix}b^\dagger_{1 -} & b^{\mathrm{L} \dagger}_{\mathrm{R}\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
(1-\nu) p^2 & \mathrm{i}\sqrt{2} \nu \frac{1}{c_\mathrm{T}}\omega_n p \\
-\mathrm{i} \sqrt{2}\nu \frac{1}{c_\mathrm{T}}\omega_n p & \frac{1}{c_\mathrm{T}^2}\omega_n^2 + 2 (1 + \nu) q^2
\end{pmatrix}
\begin{pmatrix}b_{1 -}\\ b^{\mathrm{L} }_{\mathrm{R}\mathrm{S}} \end{pmatrix}. \label{eq:solid Lagrangian gauge field L1}\\
\mathcal{L}_{\mathrm{L} 2} &= \frac{1}{4\mu} p^2 \lvert b_{1+} \rvert^2\label{eq:solid Lagrangian gauge field L2}
\end{align}
\end{widetext}
Here $b^{a\dagger}_{\kappa\lambda} J^a_{\kappa\lambda}$ is a shorthand for $\frac{1}{2} b^{a\dagger}_{\kappa\lambda} J^a_{\kappa\lambda} + \frac{1}{2} J^{a\dagger}_{\kappa\lambda} b^a_{\kappa\lambda}$, and in analogy to Eq.~\eqref{eq:sigma plus minus definition} we have defined
\begin{align}
b_{1+} &= \frac{1}{\sqrt{2}} (b^\mathrm{R}_{1\mathrm{S}} + b^\mathrm{S}_{1\mathrm{R}}), &
b_{1-} &= \frac{1}{\sqrt{2}} (b^\mathrm{R}_{1\mathrm{S}} - b^\mathrm{S}_{1\mathrm{R}}).
\end{align}
\subsection{The interpretation of stress components}\label{subsec:Interpretation of stress components}
We will derive the propagators of Sec.~\ref{subsec:Field-theoretic elasticity} on the dual side. Let us however first find out what can be learned regarding the spectrum of excitations from the Lagrangians Eqs.~\eqref{eq:solid stress Lagrangian components}, \eqref{eq:solid stress gauge field Lagrangian}. At first, there are 12 components $\sigma^a_\mu$ as $\mu = \mathfrak{t},x,y,z$ and $a = x,y,z$. Let us first consider the static, 3+0D limit where the three temporal components are absent: $\sigma^a_\mathfrak{t} = 0$. Stress conservation $\partial_m \sigma^a_m = -q \sigma^a_\mathrm{L} =0$ removes another three components. Furthermore we have the three Ehrenfest constraints $\sigma^a_m = \sigma^m_a$, so that the only physical components are $\sigma^\mathrm{R}_\mathrm{S} + \sigma^\mathrm{S}_\mathrm{R}, \sigma_+$ and $\sigma_-$, and these must be interpreted as static elastic `Coulomb' forces. From Eq.~\eqref{eq:solid stress Lagrangian components} we can see that the first two have a correlation function proportional to the shear modulus $\mu$ and correspond to shear forces, while in this limit the Lagrangian $\mathcal{L}_{\mathrm{L} 1}$ reduces to
\begin{align}\label{eq:electric shear stress 3+0D}
\mathcal{L}_{\mathrm{L} 1} \text{(3+0D)} &= \frac{1}{4\mu} \frac{1 -\nu}{1 +\nu} \lvert \sigma_- \rvert^2.
\end{align}
This involves the compression modulus $\kappa$ through the Poisson ratio $\nu$, and this contribution represents the compressional force. In Ref.~\cite{ZaanenNussinovMukhin04}, in analogy to electromagnetism, this force was called `electric', related to longitudinal stress, while the other two are purely transverse and can be called `magnetic'. These equations reproduce the results of classical elasticity, see for instance Ref.~\onlinecite{Kleinert89b}. Note that the sectors $\mathcal{L}_{\mathrm{T} 1}$ and $\mathcal{L}_{\mathrm{T} 2}$ are completely absent in this limit.
What happens when we include quantum dynamics by adding $\sigma^a_\tau$? We still have three conservation laws and three Ehrenfest constraints: there are $12 -3 -3 = 6$ physical components. Three of these are the same elastic Coulomb forces as above although they get modified at finite energies, while the three new degrees of freedom are the propagating phonons, two transverse and one longitudinal. The conservation laws are now $\partial_\mu \sigma^a_\mu = 0$ as in Eq.~\eqref{eq:stress conservation}, so the components $\sigma^a_\mathfrak{t}$ and $\sigma^a_\mathrm{L}$ together represent one phonon per $a$. Once the Ehrenfest constraints are lifted through adding second-order terms Eq.~\eqref{eq:second gradient Lagrangian in stress tensor}, three more forces are added, which fall off exponentially with length scale $\ell$. These are short-ranged, rotational forces, and can be classified as either longitudinal or transverse as explained in Sec.~\ref{subsec:Rotational elasticity}. Together, there are 9 physical degrees of freedom, three of which are the propagating phonons. The breakdown of this classification for the sectors of Eq.~\eqref{eq:solid stress Lagrangian components} is given in Table~\ref{table:stress degrees of freedom}.
\newlength{\oldcolsep}
\setlength{\oldcolsep}{\tabcolsep}
\setlength{\tabcolsep}{.5em}
\begin{table}
\begin{tabular}{ccccc}
\toprule
sector & & phonon & shear force & rotational force \\
\cline{1-1} \cline{3-5}
$\mathrm{T} 1$ & & transverse & -- & transverse \\
$\mathrm{T} 2$ & & transverse & -- & transverse \\
$\mathrm{T} 3$ & & -- & magnetic & longitudinal \\
$\mathrm{L} 1$ & & longitudinal & electric & -- \\
$\mathrm{L} 2$ & & -- & magnetic & -- \\
\botrule
\end{tabular}
\caption{Classification of stress degrees of freedom of the 3+1D isotropic solid into the five sectors identified in Eq.~\eqref{eq:solid stress Lagrangian components}. We distinguish three types: phonons, which are propagating degrees of freedom; shear forces which have a Coulomb-like nature; and rotational forces which are only present once the Ehrenfest constraints are softened and which fall off exponentially. }\label{table:stress degrees of freedom}
\end{table}
\setlength{\tabcolsep}{\oldcolsep}
\begin{table}
\begin{tabular}{ccl}
\toprule
component & spin & interpretation \\
\hline
$-\sigma^\mathrm{L}_\mathrm{L} + \sigma^\mathrm{R}_\mathrm{R} - \sigma^\mathrm{S}_\mathrm{S}$ & (0,0) & longitudinal phonon \\
\rule{0pt}{3ex}
$\sigma^\mathrm{R}_\mathrm{S} -\sigma^\mathrm{S}_\mathrm{R}$ & (1,0) & longitudinal rotational force \\
$\sigma^\mathrm{R}_\mathrm{L} -\sigma^\mathrm{L}_\mathrm{R}$ & (1,1) & transverse rotational force \\
$\sigma^\mathrm{S}_\mathrm{L} -\sigma^\mathrm{L}_\mathrm{S}$ & (1,-1) & transverse rotational force \\
\rule{0pt}{3ex}
$2\sigma^\mathrm{L}_\mathrm{L} + \sigma^\mathrm{R}_\mathrm{R} - \sigma^\mathrm{S}_\mathrm{S}$ & (2,0) & electric shear\\
$\sigma^\mathrm{L}_\mathrm{R} + \sigma^\mathrm{R}_\mathrm{L}$ & (2,1) & transverse phonon \\
$\sigma^\mathrm{L}_\mathrm{S} + \sigma^\mathrm{S}_\mathrm{L}$ & (2,-1) & transverse phonon \\
$\sigma^\mathrm{R}_\mathrm{R} + \sigma^\mathrm{S}_\mathrm{S}$ & (2,2) & magnetic shear\\
$\sigma^\mathrm{R}_\mathrm{S} + \sigma^\mathrm{S}_\mathrm{R}$ & (2,-2) & magnetic shear\\
\botrule
\end{tabular}
\caption{Helicity decomposition of the 3+0D stress tensor with respect to the momentum $\mathbf{q}$ which sets the longitudinal direction $\mathrm{L}$, where it is understood that the phonons will emerge once dynamics is added. The components are not yet normalized as to not clutter the notation.}\label{table:stress helicity decomposition}
\end{table}
Even more insight can be gained by the so-called helicity decomposition of the stress tensor~\cite{Kleinert89b}. Again we restrict ourselves to 3+0D where the temporal components $\sigma^a_\mathfrak{t}$ are absent. However, we will keep around the longitudinal components $\sigma^a_\mathrm{L}$ which are subject to stress conservation, because we know these will contain the phonons once we switch on the time axis. The stress tensor $\sigma^a_m$ can now be decomposed by regarding its behavior under spatial 3-rotations. Since the rotation group generates the angular momentum operators, this classification can be assigned `spin quantum numbers' $(s,m)$. Because $\sigma^a_m$ is a 2-tensor, the `total spin' number $s$ takes values in $0,1,2$, while the `magnetic' number takes values in $(-s, \ldots, s)$. The $s$-sectors where also employed in Eqs.~\eqref{eq:P0}--\eqref{eq:P2}. Performing the helicity decomposition, one ends up with Table~\ref{table:stress helicity decomposition}.
The single spin-0 component is invariant under rotations; this is clearly the compressional, longitudinal phonon. The three spin-1 components are absent when the Ehrenfest constraint is imposed, so these correspond to the rotational forces. In the spin-2 sector, the $m=\pm 1$-components are the two transverse phonons, while the other three are the shear Coulomb forces. The $s=2,m=0$-component is the `electric shear' related to longitudinal stress, while the $m= \pm 2$-components are the `magnetic' shear forces. In a forthcoming article, we will show that this spin-2 character becomes very apparent when an elastic medium is coupled to (linearized) gravity.
Up to know we have considered classification of the stress tensor components. As we already emphasized, a key insight of vortex--boson duality is that the dual gauge fields themselves have a direct physical meaning as the mediators of interactions between the topological defects. This can be seen by looking at the Lagrangian Eq.~\eqref{eq:solid stress gauge field Lagrangian}, which is of the form
\begin{equation}\label{eq:general dual Langrangian Coulomb phase}
\mathcal{L}_\mathrm{dual} = (\partial b )^2 + \mathrm{i} b \cdot J.
\end{equation}
reducing to the Maxwell action of electromagnetism in 2+1D~\cite{CvetkovicZaanen06a,QLC2D} while in 3+1D one runs into the two-form gauge theory which we showed elsewhere to be
very convenient dealing with interactions between superfluid and Abrikosov vortices~\cite{BeekmanSadriZaanen11,BeekmanZaanen11}. It is now natural
to focus on the stress gauge field propagators, which can be determined in the usual way from the generating functional with external (dislocation) sources~\cite{QLC2D}
\begin{equation}
\langle b^{a \dagger}_{\mu\nu} \; b^b_{\kappa\lambda} \rangle = \frac{1}{\mathcal{Z}[0]} \frac{\delta}{\delta J^{b\dagger}_{\kappa\lambda} } \frac{\delta}{\delta J^{a\phantom{\dagger}}_{\mu\nu}} \mathcal{Z}[J] \Big\rvert_{J=0}.
\end{equation}
This amounts to integrating out the $b$-fields in Eq.~\eqref{eq:solid stress gauge field Lagrangian}, which boils down to inverting the matrices, since the Lagrangian is quadratic.
It is now insightful to use the Coulomb gauge fix $\partial_m b^a_{m \nu} = 0$ instead of the Lorenz gauge fix, since it removes all occurrences of the $\mathrm{L}$-components in $\mu$, $\nu$.
The Lagrangian in this gauge fix is easily obtained from Eq.~\eqref{eq:solid stress gauge field Lagrangian} by substituting $p b^a_{1 \nu} = -q b^a_{\mathfrak{t} \nu}$ according to Eq.~\eqref{eq:tLRS to 01RS transformation}. This also implies $p b^{1 -} = -q b^{\mathfrak{t} -}$. After this substitution and inverting the matrices, we read off the diagonal components to find for the longitudinal sector $\mathcal{L}_{\mathrm{L} 1}$:
\begin{align}
\langle b^\dagger_{\mathfrak{t} -} \; b_{\mathfrak{t} -} \rangle &= \frac{\mu}{q^2} \frac{2}{1-2\nu} \frac{\omega^2 - (1+\nu)c_\mathrm{T}^2 q^2}{\omega^2 - c_\mathrm{L}^2 q^2},\label{eq:stress gauge field propagator electric shear}\\
\langle b^{\mathrm{L} \dagger}_{\mathrm{R}\mathrm{S}} \; b^\mathrm{L}_{\mathrm{R}\mathrm{S}} \rangle &= -\mu \frac{ c_\mathrm{L}^2}{\omega^2 - c_\mathrm{L}^2 q^2 }.\label{eq:stress gauge field propagator longitudinal phonon}
\end{align}
where we have performed the Wick rotation to real time $\omega_n \to \mathrm{i} \omega - \delta$ (ignoring the infinitesimal $\delta$).
The second equation obviously describes the propagating longitudinal phonon with velocity $c_\mathrm{L}$. In the limit $\omega \to \infty$ the first equation is proportional to $1/q^2$, indicating that this is an instantaneous force like the Coulomb force. In the static limit $\omega \to 0$, the propagator is $2\mu \frac{1- \nu}{1+\nu} \frac{1}{q^2}$, consistent with Eq.~\eqref{eq:electric shear stress 3+0D}.
This was already identified in QLC2D: the static force between edge dislocation sources is carried by the temporal components of the dual stress gauge field in the longitudinal sector (carrying a transverse Burgers index). The novelty in 3+1D is that there are two more, `magnetic' shear forces, carried by $b_{\mathfrak{t} +}$ resp. $b^\mathrm{R}_{\mathfrak{t} \mathrm{R}} + b^\mathrm{S}_{\mathfrak{t} \mathrm{S}}$ in the $\mathrm{L} 2$- resp. $\mathrm{T} 3$-sectors. Their propagator is simply $\frac{2\mu}{q^2}$.
Let us now turn to the transverse sectors $\mathcal{L}_{\mathrm{T} 1, \mathrm{T} 2, \mathrm{T} 3}$. Integrating out the dual stress gauge fields in the Coulomb gauge fix yields the propagators,
\begin{align}
\langle b^{\mathrm{L}\dagger}_{\mathfrak{t} \mathrm{R}} \; b^\mathrm{L}_{\mathfrak{t} \mathrm{R}} \rangle = \langle b^{\mathrm{L}\dagger}_{\mathfrak{t} \mathrm{S}} \; b^\mathrm{L}_{\mathfrak{t} \mathrm{S}} \rangle
&= \mu \frac{1}{q^2} \frac{ \omega^2 (1 + \ell^2 q^2) - 4 c_\mathrm{T}^2 \ell^2 q^4}{\omega^2 - c_\mathrm{T}^2 q^2(1+ \ell^2 q^2)},\label{eq:stress gauge field propagator rotational force}\\
\langle b^{\mathrm{R}\dagger}_{\mathrm{R}\mathrm{S}} \; b^\mathrm{R}_{\mathrm{R} \mathrm{S}} \rangle = \langle b^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}} \; b^\mathrm{S}_{\mathrm{R} \mathrm{S}} \rangle
&= -\mu \frac{c_\mathrm{T}^2(1 + \ell^2 q^2)}{\omega^2 - c_\mathrm{T}^2 q^2(1+\ell^2q^2) },\label{eq:stress gauge field propagator transverse phonons}\\
\langle b^{\mathrm{R}\dagger}_{\mathfrak{t} \mathrm{R}} \; b^\mathrm{R}_{\mathfrak{t} \mathrm{R}} \rangle = \langle b^{\mathrm{S}\dagger}_{\mathfrak{t} \mathrm{S}} \; b^\mathrm{S}_{\mathfrak{t} \mathrm{S}} \rangle
&= \mu \frac{1 + \ell^2 q^2}{q^2}.\label{eq:stress gauge field propagator magnetic shear}
\end{align}
The second equation describes the two transverse phonons propagating with velocity $c_\mathrm{T}$. The third equation, stemming from the $\mathrm{T} 3$-sector shows a static shear force, that persists in the limit $\ell \to 0$; it is the partner of the magnetic shear force in the $\mathrm{L} 2$-sector. However, for $q \gg 1/\ell$, this propagator does not vanish but becomes proportional to $\ell^2$, showing that at short length scales it probes the second-order, rotational elasticity. Therefore the $T3$-sector contains magnetic shear and a rotational force. The first equation, in the limit $\omega \to \infty$, reduces to the third equation, and one would be led to think it also contains a shear force next to a rotational force. That this is not the case can be seen in the limit $\omega \to 0$, where we have
\begin{equation}
\langle b^{\mathrm{L}\dagger}_{\mathfrak{t} \mathrm{R}} \; b^\mathrm{L}_{\mathfrak{t} \mathrm{S}} \rangle = \langle b^{\mathrm{L}\dagger}_{\mathfrak{t} \mathrm{S}} \; b^\mathrm{L}_{\mathfrak{t} \mathrm{R}} \rangle (\omega \to 0) = 4\mu \frac{1}{q^2 + \frac{1}{\ell^2}}.\label{eq:rotational force suppression}
\end{equation}
These Coulomb forces decay exponentially on the {\em rotational length scale} $\ell$ of Eq.~\eqref{eq:isotropic solid second gradient energy}. In the limit $\ell \to 0$ of a crystal, these forces disappear altogether. This result is also found in 2+1D (see QLC2D) where one such short-ranged force is present which we called the {\em rotational force} since it is only present when local rotations are permitted in the elastic medium. All in all, we end up with the classification in Table~\ref{table:stress degrees of freedom}.
\subsection{Dual propagator relations}\label{subsec:Dual propagator relations}
Given the dual action, let us now derive the strain (phonon) propagators $G_\mathrm{L}$ and $G_\mathrm{T}$ from Eqs.~\eqref{eq:longitudinal propagator}, \eqref{eq:transverse propagator} but this time using the stress photon formalism.
One first has to translate these strain propagators to the stress gauge field language. This is more subtle than one may naively expect, and it was for the first time accomplished in 2+1D in Ref.~\onlinecite{ZaanenNussinovMukhin04}.
The main issue is that the external sources in the path integral used to derive the two-point functions must themselves be carried through the dualization procedure. Here we will introduce a method for the transverse
propagator which is slightly modified compared to the 2+1D case~\cite{QLC2D}.
Let us first consider the longitudinal propagator Eq.~\eqref{eq:longitudinal propagator}, defined as,
\begin{align}
G_\mathrm{L} &= \langle \partial_a u^{a} \; \partial_b u^b \rangle
= \frac{1}{Z[0]} \frac{\delta}{\delta \mathcal{K}} \frac{\delta}{\delta \mathcal{K}} Z[\mathcal{K}] \big\rvert_{\mathcal{K} =0},\label{eq:longitudinal propagator from path integral}\\
\mathcal{Z}[\mathcal{K}] &= \int \mathcal{D} u^a \exp(-\mathcal{S}_\mathrm{solid}-\int \mathrm{d} \tau \mathrm{d}^D x\; \mathcal{K} \partial_a u^a\big).
\end{align}
According to Eq.~\eqref{eq:stress tensor definition}, the stress tensor changes under the introduction of $\mathcal{K}$. It only affects the compression component,
\begin{equation}
P^{(0)}_{mnab} \sigma_n^b = -\mathrm{i} D \kappa P^{(0)}_{mnab} \partial_n u^b -\mathrm{i} \mathcal{K} \delta_{ma}, \label {eq:P0 stress with source}
\end{equation}
In the dualization procedure, the dual Lagrangian is modified from Eq.~\eqref{eq:dual solid Lagrangian} to~\cite{QLC2D},
\begin{align}
\mathcal{L}_\mathrm{dual} = \frac{1}{2} \sigma^a_\mu C^{-1}_{\mu\nu ab} \sigma^b_\nu - \frac{1}{2\kappa} \mathcal{K}^2 + \mathrm{i} \frac{1}{D\kappa}\mathcal{K} \sigma^a_a + \mathrm{i} \sigma^a_\mu \partial_\mu u^a.
\end{align}
From this expression and Eq.~\eqref{eq:longitudinal propagator from path integral} we find
\begin{equation}\label{eq:longitudinal propagator stress propagator relation}
G_\mathrm{L} = \frac{1}{\kappa} - \frac{1}{(D\kappa)^2} \langle \sigma^{a}_a\; \sigma^b_b \rangle.
\end{equation}
Using Eq.~\eqref{eq:stress gauge field components} the compressional stress can be written as,
\begin{align}
\sigma^a_a = -\sigma^\mathrm{L}_\mathrm{L} + \sqrt{2} \sigma_-
= -\mathrm{i}\frac{\omega_n}{c_\mathrm{T}} b^\mathrm{L}_{\mathrm{R}\mathrm{S}} + \sqrt{2} p\; b_{1-}.
\end{align}
Note that $G_\mathrm{L}$ only depends on $\mathcal{L}_{\mathrm{L} 1}$ Eq.~\eqref{eq:solid Lagrangian gauge field L1} and not $\mathcal{L}_{\mathrm{L} 2}$ Eq.~\eqref{eq:solid Lagrangian gauge field L2}. In terms of the stress gauge fields, the explicit expression for the stress term in the longitudinal propagator becomes
\begin{align}
\langle \sigma^{a}_a\; \sigma^b_b \rangle &=
\frac{\omega_n^2}{c_\mathrm{T}^2} \langle b^{\mathrm{L}\dagger}_{\mathrm{R} \mathrm{S}}\; b^\mathrm{L}_{\mathrm{R} \mathrm{S}} \rangle
+ 2p^2 \langle b^\dagger_{ 1 -}\; b_{1 -} \rangle \nonumber \\
& \phantom{m} + \mathrm{i} \sqrt{2} \frac{\omega_n}{c_\mathrm{T}} p \big( \langle b^{\mathrm{L}\dagger}_{\mathrm{R} \mathrm{S}}\; b_{1 -} \rangle
- \langle b^\dagger_{1 -}\; b^{\mathrm{L}}_{\mathrm{R} \mathrm{S}} \rangle \big)
.\label{eq:compression stress propagator}
\end{align}
The great advantage of this expression is that it is valid both in the ordered (solid) but also the disordered (quantum liquid crystal) phases.
The propagators of the stress gauge fields can be obtained as usual from the generating functional with external sources $\mathcal{Z}[J^a_{\kappa\lambda}]$, i.e. integrating out the dual stress fields and taking functional derivatives with respect to $J^a_{\kappa\lambda}$. Since the Lagrangian is Gaussian in $b^a_{\kappa\lambda}$, this amounts to inverting the matrix in Eq.~\eqref{eq:solid Lagrangian gauge field L1}. It can be verified that after this inversion and inserting the correct contributions in Eq.~\eqref{eq:compression stress propagator} and then Eq.~\eqref{eq:longitudinal propagator stress propagator relation}, we obtain the correct expression Eq.~\eqref{eq:longitudinal propagator} (ignoring the contributions from second-gradient elasticity)
We emphasize again that although this dual route is much more laborious in case of the solid, it can be taken in the quantum liquid crystals as well.
The derivation of the transverse propagator is along the same lines. For 2+1 dimensions, there is only one transverse phonon and the longitudinal and transverse sectors there are rather similar~\cite{QLC2D}. In 3+1 dimensions however, there are two transverse phonons instead of one. More importantly, there are three rotational planes instead of one. In fact, inspecting Eq.~\eqref{eq:transverse propagator}, the transverse propagator is a sum over three terms
\begin{equation}
G_\mathrm{T} = 2 \sum_{ab} \langle \omega^{ab} \; \omega^{ab} \rangle = 4 \sum_c \langle \omega^c \; \omega^c \rangle.
\end{equation}
It will turn out to be quite convenient to consider these three terms separately in terms of the Fourier space components for the index $c$. Let us therefore consider
\begin{align}
G_\mathrm{T} &= G_{\mathrm{T} 1} + G_{\mathrm{T} 2} + G_{\mathrm{T} 3} \nonumber\\
&= 4 \langle \omega^{\mathrm{S}\dagger} \; \omega^\mathrm{S} \rangle + 4 \langle \omega^{\mathrm{R}\dagger} \; \omega^\mathrm{R} \rangle +4 \langle \omega^{\mathrm{L}\dagger} \; \omega^\mathrm{L} \rangle.\label{eq:transverse propagator from torque correlators}
\end{align}
We will see that the labels $\mathrm{T} 1$, $\mathrm{T} 2$, $\mathrm{T} 3$ match those of the solid Lagrangian Eqs.~\eqref{eq:L1}--\eqref{eq:L3}. These terms can in fact be straightforwardly computed calculated in terms of the strains, Eq.~\eqref{eq:displacement propagator}. The result is,
\begin{align}
G_{\mathrm{T} 1} &= G_{\mathrm{T} 2} = \frac{1}{\mu} \frac{ c_\mathrm{T}^2 q^2}{\omega_n^2 + c_\mathrm{T}^2 q^2(1 + \ell^2 q^2)},\label{eq:GT1 GT2 solid}\\
G_{\mathrm{T} 3} &= 0.\label{eq:GT3 solid}
\end{align}
$G_{\mathrm{T} 1}$, $G_{\mathrm{T} 2}$ describe the transverse phonons in the $\mathrm{R}$- resp. $\mathrm{S}$-direction while $G_{\mathrm{T} 3}$ is just unphysical. In Sec.~\ref{sec:nematic} we will show that in the isotropic nematic each of these three propagators will feature a massless rotational Goldstone mode. To compare the result of Eqs.~\eqref{eq:GT1 GT2 solid}, \eqref{eq:GT3 solid} with the definitions Eq.~\eqref{eq:transverse propagator from torque correlators}: for the component $\omega^\mathrm{L}$ the momentum is perpendicular to the rotational plane. Apparently the propagator between a source and sink of the rotational field which are separated from each other perpendicular to the rotational plane vanishes identically, while forces can be
exchanged when they are separated within the rotational plane. These latter forces are obviously just shear forces. In the presence of second-order elasticity Eq.~\eqref{eq:second gradient Lagrangian in stress tensor}, all three propagators are finite, although the interactions associated with these second-order terms are short ranged.
To derive the dual relations for the transverse propagators, let us add sources $\mathcal{J}^c$ coupling to each of the rotation fields $\omega^c$. Similar to Eq.~\eqref{eq:longitudinal propagator from path integral}, we then have
\begin{align}
4\langle \omega^c \; \omega^{c} \rangle
&= \frac{1}{\mathcal{Z}[0]} \frac{\delta}{\delta \mathcal{J}^{c}} \frac{\delta}{\delta \mathcal{J}^c} \mathcal{Z}[\mathcal{J}] \big\rvert_{\mathcal{J} =0}, \ {\text{no sum } c} \label{eq:transverse propagator from path integral}\\
\mathcal{Z}[\mathcal{J}] &= \int \mathrm{D} \omega^c \exp\big(-\mathcal{S}_\mathrm{solid} - \int \mathrm{d} \tau \mathrm{d}^D x\; 2 \mathcal{J}^c \omega^c \big).
\end{align}
Since the antisymmetric components of $\sigma^a_m$ are absent from first-order elasticity, we need second-gradient terms to derive the dual propagators. Starting from Eq.~\eqref{eq:second-gradient elasticity} with the source term in Eq.~\eqref{eq:transverse propagator from path integral} added, the torque stress is modified from Eq.~\eqref{eq:torque stress definition} to
\begin{equation}
\tau^c_m = -\mathrm{i} 4 \mu \ell^2 \partial_m\omega^c + 2\mathrm{i} \frac{\partial_m}{(\partial_n)^2} \mathcal{J}^c.
\end{equation}
Here we used the formal equality $1 = (\partial_m/\partial_n^2) \partial_m$ and integration by parts. The terms in the dual Lagrangian involving torque stress become,
\begin{align}
\mathcal{L}^{(2)}_\mathrm{dual}[J] &= \frac{1}{4\mu \ell^2} \Big( \frac{1}{2} \lvert \tau^c_m \rvert^2
- 2\mathrm{i} \frac{\partial_m \tau^c_m}{q^2} \mathcal{J}^c - 2 \mathcal{J}^{c\dagger} \frac{1}{q^2} \mathcal{J}^c
\Big) \nonumber\\
&\phantom{mm} +\mathrm{i} \tau^c_m \partial_m \omega^c.
\end{align}
We find subsequently from Eq.~\eqref{eq:transverse propagator from path integral} the identities for each $c$ (no sum over $c$):
\begin{equation}
4\langle \omega^c \; \omega^{c} \rangle
= \frac{1}{ \mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \langle \partial_m \tau^c_m\; \partial_n \tau^{c }_n \rangle.
\end{equation}
Finally we can substitute the Ehrenfest constraints in the presence of torque stress Eq.~\eqref{eq:Ehrenfest constraint torque stress} to find for each component $c$ separately,
\begin{equation}
4\langle \omega^c \; \omega^{c } \rangle
= \frac{1}{ \mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \langle \epsilon_{cma} \sigma^a_m\; \epsilon_{cnb} \sigma^{b}_n \rangle.\label{eq:transverse propagator stress propagator relation}
\end{equation}
This expression is valid even in the limit $\ell \to 0$: the factors of $\ell$ will cancel out to leave a non-divergent expression. Substituting the dual stress gauge field using Eq.~\eqref{eq:stress gauge field components},
\begin{widetext}
\begin{align}
G_{\mathrm{T} 1} &= \frac{1}{\mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \Big[
p^2 \langle b^{\mathrm{L}\dagger}_{1 \mathrm{S}} \; b^{\mathrm{L}}_{1 \mathrm{S}} \rangle + \frac{\omega_n^2}{c_\mathrm{T}^2} \langle b^{\mathrm{R}\dagger}_{\mathrm{R}\mathrm{S}}\; b^\mathrm{R}_{\mathrm{R} \mathrm{S}}\rangle + \mathrm{i} \frac{\omega_n}{c_\mathrm{T}} p \Big( \langle b^{\mathrm{R}\dagger}_{\mathrm{R}\mathrm{S}} \; b^\mathrm{L}_{1 \mathrm{S}}\rangle - \langle b^{\mathrm{L}\dagger}_{1\mathrm{S}}\; b^\mathrm{R}_{\mathrm{R} \mathrm{S}}\rangle \Big)
\Big],\nonumber\\
G_{\mathrm{T} 2} &= \frac{1}{\mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \Big[
p^2 \langle b^{\mathrm{L}\dagger}_{1 \mathrm{R}}\; b^{\mathrm{L}}_{1 \mathrm{R}} \rangle + \frac{\omega_n^2}{c_\mathrm{T}^2} \langle b^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}}\; b^\mathrm{S}_{\mathrm{R}\mathrm{S}}\rangle + \mathrm{i} \frac{\omega_n}{c_\mathrm{T}} p \Big( \langle b^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}} \; b^\mathrm{L}_{1 \mathrm{R}}\rangle - \langle b^{\mathrm{L}\dagger}_{1\mathrm{R}}\; b^\mathrm{S}_{\mathrm{R}\mathrm{S}}\rangle \Big)
\Big],\nonumber\\
G_{\mathrm{T} 3} &= \frac{1}{\mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \Big[ p^2 \langle b^{\mathrm{R}\dagger}_{1 \mathrm{R}} \; b^{\mathrm{R}}_{1 \mathrm{R}} \rangle + p^2 \langle b^{\mathrm{S}\dagger}_{1 \mathrm{S}} \; b^\mathrm{S}_{1\mathrm{S} }\rangle - p^2 \langle b^{\mathrm{S}\dagger}_{1\mathrm{S}} \; b^\mathrm{R}_{1 \mathrm{R}}\rangle - p^2\langle b^{\mathrm{R}\dagger}_{1\mathrm{R}}\; b^\mathrm{S}_{1\mathrm{S} }\rangle \Big].\label{eq:transverse propagator gauge field}
\end{align}
\end{widetext}
This makes clear why we choose these specific labels. In the isotropic solid, the propagator $G_{\mathrm{T} 1}$ is a result of the Lagrangian contribution $\mathcal{L}_{\mathrm{T} 1}$ etc. while the longitudinal propagator is related to $\mathcal{L}_{\mathrm{L} 1}$. Again, it can be verified directly by inverting the matrices in Eqs.~\eqref{eq:L1}--\eqref{eq:L3} that these expressions reproduce Eqs.~\eqref{eq:GT1 GT2 solid}, \eqref{eq:GT3 solid}.
\begin{figure*}
\begin{center}
\includegraphics[width=7.7cm]{3DsolT.png}
\hfill
\includegraphics[width=7.7cm]{3DsolL.png}\\
{\centering
\includegraphics[scale=0.6]{legends_solid.pdf}
}\\
\caption{Spectral functions Eq.~\eqref{eq:spectral function definition} (left: transverse; right: longitudinal) of the isotropic solid in units of the inverse shear modulus $1/\mu \equiv 1$, with Poisson ratio $\nu = 0.2$. The width of the poles is artificial and denotes the relative pole strengths: these ideal poles are actually infinitely sharp. The propagating modes, phonons, are massless (zero energy as $q \to 0$) and have linear dispersion $\omega = cq$. The velocities are the transverse $c_\mathrm{T}$ resp. longitudinal velocity $c_\mathrm{L}$. The pole in the transverse sector is doubly degenerate as there are two transverse phonons.}\label{fig:solid spectral functions}
\end{center}
\end{figure*}
Upon Wick rotating to real time $ \omega_n \to \mathrm{i} \omega - \delta$ one can compute the spectral functions
\cite{QLC2D},
\begin{equation}\label{eq:spectral function definition}
S (\omega,q) = \mathrm{Im}\; G (\mathrm{i} \omega -\delta, q).
\end{equation}
where $G = G(\omega_n ,q)$ is the particular propagator under investigation. The propagating modes show up as poles in the spectral function and for future reference we plot the spectral functions $G_\mathrm{L}$ and $G_{\mathrm{T} 1} = G_{\mathrm{T} 2}$ in Fig.~\ref{fig:solid spectral functions}. One infers the presence of the longitudinal and transverse phonon poles, linearly dispersing with velocities $c_\mathrm{L}$ and $c_\mathrm{T}$, respectively.
\subsection{Torque stress gauge fields}\label{subsec:Torque stress gauge fields}
We have seen that the stress gauge fields $b^a_{\kappa\lambda}$ are dual to the displacement fields $u^a$, coupling minimally to the dislocation sources $J^a_{\kappa\lambda}$. One can wonder whether such a dualization
exists as well for the rotation fields $\omega^c$ of Eq.~\eqref{eq:rotation field definition}. The answer is affirmative. In Eq.~\eqref{eq:torque stress definition} we have already seen that in the presence of higher-order elasticity terms the momentum canonical to the rotation field is the torque stress $\tau^c_m$. It is as well possible to consider the fate of torque stress in linear elasticity when $\ell = 0$. Define~\cite{BeekmanWuCvetkovicZaanen13,QLC2D}:
\begin{equation}\label{eq:linear torque stress definition}
\tau^c_\mu = \epsilon_{cba} \epsilon_{b\mu \kappa\lambda}\tfrac{1}{2} b^a_{\kappa\lambda}.
\end{equation}
The reader should be careful that this definition of torque stress and Eq.~\eqref{eq:torque stress definition} are of a different origin and they should not be interchanged. Notice that $ \tau^c_\mu $ is not gauge invariant under the gauge transformations Eq.~\eqref{eq:stress gauge transformation}. This is not a problem, since in the present formulation it is at first simply a collection of stress gauge fields. On a deeper level, torque stress is not unambiguously defined as long as there is a rigidity against linear shear stress. In the nematic phases, this latter rigidity is lost and torque stress becomes a genuine physical force, see Sec.~\ref{subsec:Torque stress in the quantum nematic}. Once more, this stems from the fact that translations and rotations are intertwined by their semidirect relationship. With the definition Eq.~\eqref{eq:linear torque stress definition}, the Ehrenfest constraints $\epsilon_{cma} \sigma^a_m = 0$ turn into dynamical constraints
\begin{equation}\label{eq:torque stress conservation}
\partial_\mu \tau^c_\mu = 0,\qquad \forall \ c.
\end{equation}
This can be seen by taking the divergence of Eq.~\eqref{eq:linear torque stress definition} and using the definition of the stress gauge fields Eq.~\eqref{eq:dual stress gauge field definition}. In other words, for a conserved torque stress
the linear stress tensor is symmetric in its spatial indices. A non-zero torque stress will induce antisymmetric linear stress. As long as external torque stress is absent, we can enforce it explicitly by defining torque stress as the curl of the {\em torque stress gauge field} $h^c_{\kappa\lambda}$:
\begin{equation}\label{eq:torque stress gauge field definition}
\tau^c_\mu = \epsilon_{\mu\nu\kappa\lambda} \partial_\nu \tfrac{1}{2} h^c_{\kappa\lambda}.
\end{equation}
Writing this out in components we find
\begin{align}\label{eq:torque stress gauge field components}
\tau^c_\mathfrak{t} &= q h^c_{\mathrm{R}\mathrm{S}}, &
\tau^c_\mathrm{L} &= \mathrm{i} \frac{\omega_n}{c_\mathrm{T}} h^c_{\mathrm{R}\mathrm{S}}, &
\tau^c_\mathrm{R} &= p h^c_{1\mathrm{S}}, &
\tau^c_\mathrm{S} &= p h^c_{1\mathrm{R}}.
\end{align}
The torque stress tensor is invariant under the gauge transformations
\begin{equation}
h^c_{\kappa\lambda}(x) \to h^c_{\kappa\lambda}(x) + \partial_\kappa \chi^c_\lambda(x) - \partial_\lambda \chi^c_\kappa(x),
\end{equation}
where $\chi^c_\kappa$ is any arbitrary Burgers-flavored vector field. However, there is still the initial ambiguity of defining the torque stress through Eq.~\eqref{eq:linear torque stress definition}
We can now mimic the derivation Eqs.~\eqref{eq:displacement field smooth singular} and \eqref{eq:stress dislocation minimal coupling} to find that the torque stress gauge field couples minimally to the disclination current Eq.~\eqref{eq:disclination worldsheet}:
\begin{align}
\mathrm{i} \tau^c_\mu \partial_\mu \omega^c_\mathrm{sing} &= \mathrm{i} (\epsilon_{\mu\nu\kappa\lambda} \partial_\nu \tfrac{1}{2} h^c_{\kappa\lambda})( \partial_\mu \omega^c_\mathrm{sing})\nonumber\\
&= \mathrm{i} h^c_{\kappa\lambda}\tfrac{1}{2}\epsilon_{\mu\nu\kappa\lambda} \partial_\mu \partial_\nu \omega^c_\mathrm{sing} = \mathrm{i} h^c_{\kappa\lambda} \Theta^c_{\kappa\lambda}.\label{eq:torque stress disclination minimal coupling}
\end{align}
The interpretation is accordingly: the torque stress gauge fields mediate interactions between disclination sources. It is therefore possible to examine the nature of interactions between disclinations even in the solid, and this is
quite insightful. One can use Eq.~\eqref{eq:linear torque stress definition} and then Eq.~\eqref{eq:torque stress gauge field definition} to express the relevant part of the elasticity Lagrangian in terms of the torque stress gauge field, to integrate out these gauge fields and take functional derivatives with respect to the disclination sources.
Write out Eq.~\eqref{eq:linear torque stress definition} component-by-component:
\begin{align}
\tau^\mathrm{L}_\mathfrak{t} &= b^\mathrm{R}_{\mathrm{L} \mathrm{R}} - b^\mathrm{S}_{\mathrm{L}\mathrm{S}}, &
\tau^\mathrm{R}_\mathfrak{t} &= b^\mathrm{L}_{\mathrm{L} \mathrm{R}} - b^\mathrm{S}_{\mathrm{R}\mathrm{S}}, &
\tau^\mathrm{S}_\mathfrak{t} &= b^\mathrm{L}_{\mathrm{L} \mathrm{S}} - b^\mathrm{R}_{\mathrm{R}\mathrm{S}},\nonumber\\
\tau^\mathrm{L}_\mathrm{L} &= b^\mathrm{R}_{\mathfrak{t} \mathrm{R}} - b^\mathrm{S}_{\mathfrak{t}\mathrm{S}}, &
\tau^\mathrm{R}_\mathrm{L} &= b^\mathrm{L}_{\mathfrak{t} \mathrm{R}}, &
\tau^\mathrm{S}_\mathrm{L} &= b^\mathrm{L}_{\mathfrak{t} \mathrm{S}},\nonumber\\
\tau^\mathrm{L}_\mathrm{R} &= b^\mathrm{R}_{\mathfrak{t} \mathrm{L}} , &
\tau^\mathrm{R}_\mathrm{R} &= b^\mathrm{L}_{\mathfrak{t} \mathrm{L}} + b^\mathrm{S}_{\mathfrak{t}\mathrm{S}}, &
\tau^\mathrm{S}_\mathrm{R} &= b^\mathrm{R}_{\mathfrak{t} \mathrm{S}},\nonumber\\
\tau^\mathrm{L}_\mathrm{S} &= b^\mathrm{S}_{\mathfrak{t} \mathrm{L}} , &
\tau^\mathrm{R}_\mathrm{S} &= b^\mathrm{S}_{\mathfrak{t} \mathrm{R}}, &
\tau^\mathrm{S}_\mathrm{S} &= -b^\mathrm{L}_{\mathfrak{t} \mathrm{L}} + b^\mathrm{R}_{\mathfrak{t}\mathrm{R}}.
\label{eq:torque stress in stress gauge field components}
\end{align}
Not all components of $b^a_{\kappa\lambda}$ have inverse relations with respect to $\tau^a_{\mu}$. Note that there now appear some components $b^E_{\mathfrak{t} \mathrm{L}}$ which are completely absent in linear elasticity. These will become important in the dislocation condensates.
To simplify the derivation, let us impose the Coulomb gauge for the stress gauge fields: $\partial_k b^a_{k\lambda} = - q b^a_{\mathrm{L} \lambda} =0 \ \forall a,\lambda$. The Lagrangian Eq.~\eqref{eq:solid stress gauge field Lagrangian} is easily converted using the relation $b^a_{1 \lambda} = - \frac{q}{p} b^a_{\mathfrak{t} \lambda}$ which holds in the Coulomb gauge. Furthermore, using Eq.~\eqref{eq:torque stress in stress gauge field components} in the Coulomb gauge and Eq.~\eqref{eq:torque stress gauge field components} we find the explicit relations:
\begin{align}
b^\mathrm{R}_{\mathrm{R}\mathrm{S}} &= - \tau^\mathrm{S}_\mathfrak{t} &&= - q h^\mathrm{S}_{\mathrm{R}\mathrm{S}},&\nonumber\\
b^\mathrm{L}_{\mathfrak{t} \mathrm{S}} &= \tau^\mathrm{S}_\mathrm{L} &&= \mathrm{i} \frac{\omega}{c_\mathrm{T}} h^\mathrm{S}_{\mathrm{R}\mathrm{S}},\nonumber\\
b^\mathrm{S}_{\mathrm{R}\mathrm{S}} &= - \tau^\mathrm{R}_\mathfrak{t} &&= - q h^\mathrm{R}_{\mathrm{R}\mathrm{S}},\nonumber\\
b^\mathrm{L}_{\mathfrak{t} \mathrm{R}} &= \tau^\mathrm{R}_\mathrm{L} &&= \mathrm{i} \frac{\omega}{c_\mathrm{T}} h^\mathrm{R}_{\mathrm{R}\mathrm{S}},\nonumber\\
b^\mathrm{R}_{\mathfrak{t} \mathrm{R}} + b^\mathrm{S}_{\mathfrak{t} \mathrm{S}} &= \tau^\mathrm{R}_\mathrm{R} + \tau^\mathrm{S}_\mathrm{S} &&= p h^\mathrm{R}_{1\mathrm{S}} + p h^\mathrm{S}_{1\mathrm{R}},\nonumber\\
b^\mathrm{R}_{\mathfrak{t} \mathrm{R}} - b^\mathrm{S}_{\mathfrak{t} \mathrm{S}} &= \tau^\mathrm{L}_\mathrm{L} &&= \mathrm{i} \frac{\omega}{c_\mathrm{T}} h^\mathrm{L}_{\mathrm{R}\mathrm{S}}.
\end{align}
We can now express the transverse sectors of the elasticity Lagrangian directly in torque stress gauge fields, finding:
\begin{align}
\mathcal{L}_{\mathrm{T} 1} &= \frac{1}{2\mu}q^2 p^2 \left\lvert h^\mathrm{S}_{\mathrm{R}\mathrm{S}} \right\rvert^2 ,\nonumber\\
\mathcal{L}_{\mathrm{T} 2} &= \frac{1}{2\mu}q^2 p^2 \left\lvert h^\mathrm{R}_{\mathrm{R}\mathrm{S}} \right\rvert^2,\nonumber\\
\mathcal{L}_{\mathrm{T} 3} &= \frac{1}{2\mu} q^2 p^2 \left \lvert h^\mathrm{R}_{1\mathrm{S}} + h^\mathrm{S}_{1 \mathrm{R}} \right\rvert^2.
\label{eq:solid Lagrangian torque stress gauge field}
\end{align}
Notice that the terms from second-order elasticity $\propto \ell^2$ drop out. This is because by definition Eq.~\eqref{eq:second gradient Lagrangian in stress tensor} these are proportional to $\partial_\mu \tau^c_\mu$, and as such should vanish if we use the torque stress gauge fields Eq.~\eqref{eq:torque stress gauge field definition}.
These expressions have the following physical meaning: the torque stress gauge fields mediate interactions between disclination sources. As usual, the purely transverse components $h^c_{\mathrm{R}\mathrm{S}}$ represent the propagating components, while the temporal components $h^c_{\mathfrak{t} \mathrm{R}}$, $h^c_{\mathfrak{t} \mathrm{S}}$ (which are equal to $-\frac{q}{p}h^c_{1 \mathrm{R}}$, $-\frac{q}{p}h^c_{1 \mathrm{S}}$ in the Coulomb gauge) represent static `Coulomb' forces. Due to the additional factor of $q^2$ in Eq.~\eqref{eq:solid Lagrangian torque stress gauge field}, the propagators behave as $\langle h^\mathrm{R}_{\mathrm{R} \mathrm{S}} h^{\mathrm{R}\dagger}_{\mathrm{R} \mathrm{S}}\rangle \propto 1/q^2 p^2$ etc., falling off much quicker than $1/p^2$. In 3D, this implies that the energy
of a disclination--antidisclination pair increases quadratically with their separation. This is physically the same phenomenon as the confinement of quarks,
i.e. that the gluons of QCD cannot occur in isolation below the confinement scale $\Lambda_{\rm QCD}$, with the only difference that in QCD the energy between the quark sources increases linearly. In the same vein one can claim that the `torque photons' are confined in the solid at scales less than the rotational confinement scale, which in the solid is associated with the rotational stiffness length scale $\ell$~\cite{BeekmanWuCvetkovicZaanen13,QLC2D}. In Sec.~\ref{sec:nematic} we will see that these rotational forces and associated rotational Goldstone modes are deconfined in the dislocation condensates,
turning into mediators of physical long-range interactions between the (deconfined) disclinations. Notice that the component related to the third rotational Goldstone mode $h^\mathrm{L}_{\mathrm{R}\mathrm{S}}$ is completely absent in the solid. It should originate from the components $\tau^\mathrm{L}_\mathfrak{t}$ and $\tau^\mathrm{L}_\mathrm{L}$, which are present only in the second-gradient term in $\mathcal{L}_{\mathrm{T} 3}$. In Sec.~\ref{sec:nematic} we will see that a Goldstone mode does emerge in this sector in the nematic phase.
Finally, it is also possible to substitute $b_{\mathfrak{t}\pm} = p (h^\mathrm{R}_{1\mathrm{R}} \pm h^\mathrm{S}_{1 \mathrm{S}})$ in the longitudinal sectors $\mathcal{L}_{\mathrm{L} 1,2}$. This shows that also the Coulomb forces mediated by these components are confined. The component $b^\mathrm{L}_{\mathrm{R}\mathrm{S}}$ representing the longitudinal phonon has no counterpart in the torque sector.
\subsection{Quantum liquid crystals: the context}
Liquid crystals are ``mesophases'' of matter with a ``vestigial'' pattern of spontaneous symmetry breaking arising at intermediate temperatures or coupling: rotational symmetry is broken while translational invariance partially or completely persists. Classical liquid crystals are formed from highly anisotropic molecular constituents which, upon cooling from the liquid phase, can order their respective orientations while maintaining translational freedom. Only at lower temperatures crystallization sets in. These forms of matter have been known for about a century, and their theoretical description was established by De Gennes and many others~\cite{DeGennesProst95, SinghDunmur02, ChaikinLubensky00}. Starting from the opposite side, it was long realized that dislocations (the topological defects associated with translational order) are responsible for material degradation and even melting of solids~\cite{Friedel64}. Berezinskii, Kosterlitz and Thouless (BKT) in their landmark papers already suggested that unbinding of dislocations and disclinations (rotational topological defects) will lead to the disordering of two-dimensional solids~\cite{KosterlitzThouless72,KosterlitzThouless73}, he theory of which was further developed and refined by Nelson, Halperin and Young~\cite{HalperinNelson78,NelsonHalperin79,Young79}. We will refer to the topological melting driven by dislocation unbinding as the {\em KTNHY transition}. Here it was also predicted that an intermediate phase exists as a result of the exclusive proliferation of dislocations in a triangular 2D crystal, dubbed the {\em hexatic liquid crystal}. Translational symmetry is fully restored but the rotational symmetry remains broken down to the $C_6$ point group characterizing the triangular crystal.
Almost two decades later, Kivelson, Fradkin and Emery~\cite{KivelsonFradkinEmery98} proposed that the spatial ordering of electrons in strongly-correlated electron systems, as realized in
underdoped high-$T_\mathrm{c}$ superconductors, could feature symmetry properties analogous to classical liquid crystals. The stripe `crystalline' order is now destroyed at zero temperature by quantum
fluctuations in the form of proliferating dislocations, such that on macroscopic length scales the system forms a nematic quantum fluid (superconductor) which maintains however the orientational preference of the stripe electronic crystal.
This signaled the birth of the subject of quantum liquid crystals. Quite some empirical support was found since then supporting the existence of such forms of quantum liquid crystals. This includes direct evidences for the existence
of quantum nematic order in underdoped cuprates, likely related
to the original context of fluctuating stripes~\cite{AndoEtAl02, Vojta09, OganesyanKivelsonFradkin01, BorziEtAl07,HinkovEtAl08, FradkinEtAl10,FradkinKivelson10,Fradkin12}. This theme flourished in the context of the iron superconductors where quite some evidence surfaced for the prominent role of orientational symmetry breaking driven by the
electron system as being central to their physics~\cite{ChuangEtAl10,ChuEtAl12,FernandesChubukovSchmalian14}. An ambiguity in these condensed matter systems is that the crystal formed by the atoms is already breaking space translations and rotations while the electron and ion systems are coupled.
The quasi two-dimensional electron systems in the iron and copper superconductors are typically realized in tetragonal square lattices where the rotational symmetry is broken to a point group characterized by a fourfold axis. This fourfold symmetry is broken to an orthorhombic crystal structure characterized by a two-fold rotational symmetry $C_4 \to C_2$, dubbed the ``Ising nematic phase''. Given that symmetry-wise the purely electronic and crystalline tendencies to lower the point group symmetry cannot be distinguished, one does face a degree of ambiguity that cannot be avoided, giving rise to ongoing debates about the origin of the electronic nematicity in these materials~\cite{FernandesChubukovSchmalian14}.
Inspired by the initial suggestion by Kivelson {\em et al.} one of the authors (J.Z.) initiated a program to extend the KTNHY topological melting ideas to the quantum realms, initially in two space dimensions.
The emphasis has been here all along on the fundamental, theoretical side based on the symmetries and associated defects. The main restriction is that it only deals with matter formed from bosons: the constructions rest on the machinery of statistical
physics being mobilized in the $D+1$ dimensional Euclidean spacetime, turning into the quantum physics of bosons after Wick rotation. This matter lives in the Galilean continuum and the point of departure
is the spontaneous breaking of space translations and rotations into a crystal. The KTNHY transition is just one particular example of a Kramers--Wannier (or weak--strong) duality and it was found out in the 1980s how to extend this to
three dimensions when dealing with Abelian symmetries. In the context of crystalline elasticity one can rest on strain--stress duality, where phonon degrees of freedom are mapped to {\em dual stress gauge fields}.
This amounts to a generalization of the famous vortex--boson or Abelian-Higgs duality, as pioneered by Kleinert~\cite{Kleinert89b}. Using the well-known mapping of a $D$-dimensional quantum to a $D+1$-dimensional classical system, the 2+1D quantum liquid crystals were investigated by stress-strain duality starting with Ref.~\onlinecite{ZaanenNussinovMukhin04}. The relation is essentially the same as the KTNHY case in two
dimensions. One first establishes the structure of the weak--strong duality by focusing on the minimal $U(1)$-case associated with vortex melting, to then extend it to the richer theater of the space groups
underlying the crystalline symmetry breaking, profiting from the fact that the restoration of translational invariance by dislocations is associated with an Abelian symmetry.
The essence is that this duality language is geared to describe the physics of a quantum fluid (in fact, a superfluid or superconductor) that is in the limit of {\em maximal} correlation, being as close to the solid as possible. Only the collective excitations are important here. It is assumed at the onset that the particles forming the crystal continue to be bound: the `construction material' of the quantum liquid consists of local crystalline order supporting phonons disrupted by a low density
of topological defects: the dislocations. At length scales smaller than the distance between the dislocations the liquid behaves still like the solid. However, at larger distances the translational symmetry is restored
by a condensate formed out of the quantized dislocations. To a certain degree the liquid crystal aspect is a convenience.
The {\em Bose condensate of dislocations} restoring the translational symmetry is straightforwardly described in terms of a ``dual stress superconductor''. The rotational topological defects, disclinations, that restore the rotational symmetry, are just harder to deal with technically and by ``keeping disclinations out of the vacuum'' rotational symmetry continues to be broken, describing the quantum liquid crystal.
The isotropic quantum fluid is realized when these disclinations proliferate as well~\cite{Kleinert83}.
This program resulted in a series of papers that gradually exposed the quite extraordinary physics of such maximally correlated quantum liquid
crystals in 2+1 dimensions~\cite{ZaanenNussinovMukhin04,KleinertZaanen04,Cvetkovic06,CvetkovicNussinovZaanen06,CvetkovicZaanen06a,CvetkovicZaanen06b,CvetkovicNussinovMukhinZaanen08,ZaanenBeekman12,BeekmanWuCvetkovicZaanen13,LiuEtAl15}.
Recently, we have written an extensive review that comprehensively details the dual gauge field theory of these quantum liquid crystals in two dimensions~\cite{QLC2D}, to which we shall hereafter refer as QLC2D.
The present work is the extension of this theory to three spatial dimensions and we recommend the novice to the subject to have a close look at QLC2D first. We will often refer back to those results, while we do not hesitate to skip
derivations and explanations provided there when these are representative for the way things work in 3+1D as well. We also refer the reader to the introduction of QLC2D for more background on the history of and the physical interest in quantum liquid crystals.
\subsection{From two to three dimensions: weak--strong duality and the string condensate}\label{subsec:From two to three dimensions}
Our universe has three spatial dimensions and therefore the most natural quantum states of matter are formed in 3+1 dimensions. The generalization of the theory to 3+1D has been quite an ordeal ---
we are even not completely confident that the solution we present here is really watertight. Wherein lies the difficulty?
This is rooted in the fundamentals of Abelian weak--strong dualities which are very well understood in both 1+1/2D (KT topological melting) and 2+1D/3D (Abelian-Higgs duality~\cite{FisherLee89,Kleinert89a,KiometzisKleinertSchakel95,HerbutTessanovic96,CvetkovicZaanen06a,NguyenSudbo99,HoveSudbo00,HoveMoSudbo00,SmisethSmorgravSudbo04,SmisethEtAl05,SmorgravEtAl05}) while it is much less settled in 3+1D for
quite deep reasons. At the heart of these dualities is the notion that given a particular form of spontaneous symmetry breaking, the unique agents associated with restoring the symmetry are the topological excitations.
Let us first consider a broken global $U(1)$-symmetry, where the vortex is the topological workhorse.
In the zero-temperature ordered phase these only occur in the form of bound vortex--antivortex pairs since a single
free vortex suffices to destroy long-range order. In 1+1D they are point-like entities (instantons) in spacetime having a logarithmic interaction, subjected to the famous BKT vortex-unbinding transition. In
2+1D vortices are `particle-like' objects characterized by worldlines forming closed loops in spacetime in the ordered phase. At the quantum phase transition these loops `blow out', forming a {\em tangle of worldlines} corresponding
to a Bose condensate of vortices. In the ordered phases vortices are subjected to long-range interactions which work in exactly the same way in this particular dimension as electromagnetic interactions, namely by coupling to vector gauge fields. In the disordered phase, this gauged vortex condensate is therefore a {\em dual superconductor} (Higgs phase). In the context of quantum elasticity, the dislocations take the role of vortices forming the {\em dual stress superconductor}. There is however much more additional structure and the outcome is the rich world described in QLC2D.
The complication coming in at 3+1D is that dislocations (or vortices) are `line-like', forming loops in space that trace out worldsheets, not worldlines, in spacetime. In other words, they are {\em strings}. In 2+1D we are dealing with an ordinary Bose condensate of particles, constructed using the second-quantization procedure. Second quantization is however not applicable to strings in 3+1D and a fool-proof procedure to write down the effective field theory associated with the `foam' formed in spacetime from proliferated dislocation strings is just not available. Here we have to rely on a guess based on symmetry considerations that was first proposed by Rey~\cite{Rey89} in the context of fundamental string field theory. Let us present here a crude sketch of the essence of this affair in the minimal setting of the
Abelian-Higgs/vortex duality associated with the topological melting of the superfluid.
The point of departure is the relativistic Josephson action $\mathcal{L} \sim (\partial_{\mu} \varphi)^2$ describing the phase mode of the superfluid $\varphi$ in imaginary time.
The $U(1)$-field is compact and vortices arise as the topological excitations. The elementary dualization in 2+1D maps the phase mode $\varphi$ onto a vector gauge field $a_\mu$ and the vortex onto a particle current $J^\mathrm{V}_\mu$, while the action is recast as $f_{\mu \nu} f_{\mu \nu} + a_{\mu} J^\mathrm{V}_\mu$. This describes the worldlines of isolated vortices in terms of the vortex current $J^\mathrm{V}_\mu$, being subjected to a long-range interaction mediated by an effective $U(1)$-gauge field $a_{\mu}$ with field strength $f_{\mu \nu} = \partial_mu a_\nu - \partial_\nu a_\mu$. This is identical to electrodynamics in this particular dimension; one may interpret the superfluid as the Coulomb phase of an electromagnetic system sourced by conventional currents $J^\mathrm{V}_\mu$. The gauge fields $a_{\mu}$ arise as a way to impose the conservation of the supercurrent (field strength): $j_{\mu} = \epsilon_{\mu\kappa\lambda} f_{\kappa\lambda}$ is conserved $\partial_{\mu} j_{\mu} = 0$ when the original phase field $\psi$ is smooth. This continuity equation can be identically imposed by parameterizing the currents in terms of the gauge fields as $j_{\mu} = \epsilon_{\mu \nu \lambda} \partial_{\nu} a_{\lambda}$, and $a_\mu$ is directly sourced by the vortex currents $J^\mathrm{V}_\mu$.
The duality is easily extended in this ordered, Coulomb phase to 3+1D. The only difference is that one has to invoke {\em two-form} gauge fields $b_{\mu \nu}$. Namely, the supercurrent continuity equation $\partial_\mu j_\mu = 0$ is imposed by expressing it as the `four-curl' of a two-form field: $j_{\mu} = \epsilon_{\mu \nu \kappa \lambda} \partial_{\nu} b_{\kappa \lambda}$. At the same time, the vortex is a worldsheet in spacetime, parametrized by $J^\mathrm{V}_{\mu\nu}$. The action for an isolated piece of vortex world sheet has the form $\mathcal{L} \sim h_{\mu \nu \kappa}h_{\mu \nu \kappa} + b_{\mu \nu} J^\mathrm{V}_{\mu \nu}$, where $h_{\mu \nu \kappa} = \epsilon_{\mu\nu\kappa\lambda} j_\lambda$ is the field strength associated
with the gauge field $b_{\mu \nu}$. This is well known in string theory where such two-form fields arise naturally and are known as Kalb--Ramond fields~\cite{KalbRamond74}.
This dual description of the ordered phase is only the beginning of the story. We have just summarized the dual version of the interaction between isolated vortices deep in the ordered, superfluid phase. Towards the disordering quantum phase transition, in 2+1D vortex worldline loops grow and proliferate (vortices condense). This disordered state is relativistic {\em superconductor} (Higgs phase) formed out of vortex matter. Namely, the dual gauge fields $a_\mu$ couple minimally to a complex scalar field $\Phi = | \Phi | \mathrm{e}^{\mathrm{i} \phi}$, representing the second-quantized collective vortex condensate degrees of freedom. In the London limit where the amplitude $|\Phi|$ is frozen, this leads to the Ginzburg--Landau form $\mathcal{L} \sim |\Phi|^2 (\partial_\mu \phi - a_{\mu})^2 + f_{\mu \nu} f_{\mu \nu}$.
It is here that the great difficulty of the duality in 3+1D is found. The vortex strings of 3+1D proliferate (condense) into a `foam' of worldsheets in spacetime, and the question arises: what is the universal form of the effective action describing such a `string condensate'? This is a fundamental problem: the construction of string field theory. As a matter of fact, presently it is just not known how to generalize second quantization to stringy degrees of freedom. One can however rely on symmetry. Deep in the dual superconductor, the minimal coupling principle appears to insist that there is only a single consistent way of writing a Josephson action. As Rey pointed out~\cite{Rey89}, see also Ref.~\onlinecite{Franz07}, the two-form gauge field $b_{\mu \nu}$ has to be Higgsed completely and this is accomplished by a Lagrangian of the form $ \mathcal{L} \sim |\Phi|^2( \partial_{\mu} \phi_{\nu} - \partial_{\nu} \phi_{\mu} - b_{\mu \nu})^2$. One is now led to accept that the `string foam' is characterized by a vector-valued phase field $\phi_{\nu}$, having more degrees of freedom than the simple scalar $\phi$ in 2+1D.
As we discussed elsewhere, problems of principle arise with this construction in the context of this disordered superfluid/dual superconductor in 3+1D~\cite{BeekmanSadriZaanen11}. The dual superconductor can be interpreted as a boson-Mott insulator and it appears that the vectorial phase field $\phi_{\nu}$ {\em overcounts} the number of degrees of freedom. The Anderson-Higgs mechanism transfers the condensate degrees of freedom to the longitudinal polarizations of the photon (dual gauge) field. The scalar field $\phi$ has one degree of freedom but the vectorial phase field $\phi_\nu$ contains two degrees of freedom that, together with the single Goldstone mode of the superfluid, end up forming a triplet of degenerate massive modes in the 3+1D disordered superfluid. Conversely, the boson-Mott insulator is known to possess two massive propagating modes, the ``doublon and holon'' excitations. We proposed a resolution to repair this overcounting~\cite{BeekmanSadriZaanen11,BeekmanZaanen12}.
How does this play out in the current context of quantum liquid crystals? As we will see below, translational symmetry can be restored `one direction at a time', and the disorder field theory consists basically of three more-or-less independent $U(1)$-fields. These cause the shear degrees of freedom to be gapped, leading to the `liquid behavior' of liquid crystals. Furthermore, up to three rotational Goldstone modes emerge once translational symmetry is restored.
All these degrees of freedom are a priori accommodated in the ordinary, linear stress operators of elasticity---these are not the condensate phase degrees of freedom that are transferred by the Anderson--Higgs mechanism to the longitudinal polarizations of the dual gauge field. However, we benefit from the additional structure of elasticity, which contains not only linear stress, the canonical conjugate to displacements, but also torque stress, which is conjugate to local rotations. Torque stress cannot be unambiguously defined as long as shear rigidity is present, but it becomes a good physical quantity in the quantum liquid crystals. We find below that the condensate phase degrees of freedom do leave their mark on torque stresses. As we shall identify in Sec.~\ref{subsec:Torque stress in the quantum nematic} there are components, corresponding to the longitudinal two-form gauge fields, which are visible in the torque stress linear response. This is not only a clear sign that the problems outlined in Ref.~\onlinecite{BeekmanSadriZaanen11} do not arise, but also a great, and possibly first, way to test the existence of a condensate of the form proposed in Ref.~\onlinecite{Rey89} in condensed matter.
\subsection{Overview and summary of results}
As we just argued, assuming that we can rely on the minimal coupling construction for the `stringy' condensate of the dual stress superconductor, the theory of the quantum liquid crystals in 3+1D becomes a as-straightforward-as-possible
generalization of this physics in 2+1D. We have accordingly organized this paper closely following the 2+1D template~\cite{QLC2D}. In the next three chapters we set the stage by reviewing general symmetry principles, and
generalities of elasticity theory as of relevance to the remainder. In the remaining sections we will then develop step-by-step the theory of the various forms of quantum liquid-crystalline order.
The main difference in three dimensions is the nature of rotational symmetry; its ramifications for the universal
features associated with the order parameter theory will be reviewed in Sec.~\ref{sec:Symmetry principles of quantum liquid crystals}.
For empirical reasons, nearly all nematic liquid crystals of the soft matter tradition are of a very special kind: the uniaxial nematics formed from the `rod-like molecules' that orient their long axis in the same direction. As we will briefly review in the next section, these are only a part of a very large class of {\em generalized nematics} characterized by the $O(3)$ rotational symmetry of isotropic three-dimensional space, broken
down to some point group. In two dimensions all rotational proper point groups are Abelian while in 3D the point groups are generally non-Abelian. As a consequence the order parameter theory of these 3D generalized nematics is a very rich and complex affair~\cite{LiuEtAl16b,LiuEtAl16,NissinenEtAl16}. The uniaxial nematic has the point-group symmetry $D_{\infty\mathrm{h}}$, which breaks only two out of three rotational symmetries and the proper rotational part of which is Abelian; it is therefore not a good representative of rotational symmetry breaking in three dimensions.
In order to render the duality construction as simple and transparent as possible we depart from
a maximally symmetric setting: the `isotropic nematic'. In 2+1D this is literally realized by the hexatic liquid crystal, where one departs from a triangular crystal characterized by isotropic elasticity as far as its long-distance properties are concerned, and this isotropic nature is carried over to the `quantum hexatic'. In 3+1D there is no space group associated with isotropic elasticity. Instead one can consider a cubic crystal and assert that the cubic anisotropies can be approximately ignored: this is our point of departure. The $O_\mathrm{h}$ point group of the cubic crystal is however non-Abelian with far-reaching consequences for disclination defects. Nevertheless, as long as we are not interested in condensation of disclinations into the liquid (superfluid) phase, these complications can be ignored. The `isotropic quantum nematic' breaks three rotational symmetries and should carry three rotational Goldstone modes, which we shall verify explicitly with dual gauge fields. As we already discovered in QLC2D, smectic type phases have a particular elegant description in the duality setting in terms of a partial condensation of dislocations. As we will further elucidate in this section, in 3+1D this implies that both quantum smectic and columnar phases arise naturally.
In Sec.~\ref{sec:Preliminaries} we review some basic material: the field theory of quantum elasticity, stress--strain duality, rotational elasticity and static topological defect lines in solids. Quantum elasticity is just the classical theory of elasticity with an added quantum kinetic energy in imaginary time, promoted to the path integral formulation of the quantum partition function. This is a linear theory of deformations that simply describes acoustic phonons. Usually elasticity theory is expressed in term of strain fields but by employing stress--strain duality it can be formulated as well in terms of stress tensors, which are in turn the field strengths in the dual-gauge-field-theoretical formulations in the remainder. The theory governing the low-energy excitations of a translationally symmetric but rotationally rigid medium can be called {\em rotational elasticity}, which is shortly reviewed. The topological defects, the agents destroying the crystalline order of the solid state, are dislocations and disclinations with Burgers resp. Frank vectors as topological charge.
\begin{figure*}
\subfloat[crystal -- $\mathbb{Z}^3$]{\includegraphics[height=3cm]{crystal.png}\label{subfig:crystal}}
\hfill
\subfloat[columnar -- $\mathbb{Z}^2 \times \mathbb{R}$]{\includegraphics[height=3cm]{columnar.png}\label{subfig:columnar}}
\hfill
\subfloat[smectic -- $\mathbb{Z} \times \mathbb{R}^2$]{\includegraphics[height=3cm]{smectic.png}\label{subfig:smectic}}
\hfill
\subfloat[nematic -- $\mathbb{R}^3$]{\includegraphics[height=3cm]{nematic.png}\label{subfig:nematic}}
\caption{
Sketch of the symmetry of the solid and liquid crystals. Features in red denote translational symmetry, liquid-like in that direction. The translation group is also indicated in the captions. Rotational symmetry is broken to a discrete point group $\bar{P}$ in all cases. \protect\subref{subfig:crystal} For simplicity we start with a cubic crystal with translational symmetry completely broken down to a discrete subgroup. \protect\subref{subfig:columnar} Restoring translations in one direction leads to a regular 2D array of 1D liquid lines, the columnar phase. In the superconductivity jargon, this is the stripe (``river of charge'') phase. \protect\subref{subfig:smectic} Restoring translations in two dimensions yields the smectic phase; the liquid plane has the features of a 2D nematic. This is the (highly simplified) envisaged scenario in high-$T_\mathrm{c}$ superconductors. \protect\subref{subfig:nematic} Restoring translations all directions leads to a generalized nematic phase; the remaining anisotropy depends on the details of the rotational symmetry breaking. In this work we only consider explicitly the simplified case of the `isotropic' nematic, with only a single rotational modulus, see Sec.~\ref{subsec:Torque stress in the quantum nematic}. }\label{fig:phases}
\end{figure*}
Resting on the seminal work of Kleinert~\cite{Kleinert89b}, we found that in 2+1D the theory can be rewritten in terms of {\em stress gauge fields} that enumerate the capacity of the solid medium to propagate forces between external stresses as well as the internal stresses sourced by the dislocations~\cite{QLC2D}. This gauge theory corresponds to a `flavored' version of quantum electrodynamics in 2+1D, in terms of the usual one-form $U(1)$-gauge fields identifying phonons with `stress photons'. This is drastically different in 3+1D, which we shall extensively explain in Sec.~\ref{sec:Dual elasticity in three dimensions}. The topological defects are now worldsheets in spacetime. Since these act as stress sources, the gauge fields that propagate the stress are two-form gauge fields of the kind encountered in string theory. By working through the two-form gauge field formalism we do show that at least on the level of description of the elastic medium,
the correct phonon propagators are impeccably reproduced: compare Eqs.~\eqref{eq:longitudinal propagator}, \eqref{eq:transverse propagator} with Eqs.~\eqref{eq:longitudinal propagator stress propagator relation}, \eqref{eq:transverse propagator gauge field}.
Sec.~\ref{sec:Dislocation-mediated quantum melting} is the core of the development in this paper. The quantum liquid crystals are described as solids subjected to a proliferation (condensation) of dislocations. In 2+1D this is in principle
a straightforward affair because the dislocations are fundamentally like bosonic particles and the tangle of dislocation lines in spacetime is just a Bose condensate that is `charged' under the stress gauge fields: this is a plain Higgs
condensate and the quantum liquid crystals are therefore called {\em stress superconductors} similar to the dual superconductors in the context of the Abelian-Higgs duality~\cite{NguyenSudbo99,HoveSudbo00,HoveMoSudbo00}.
As we discussed in Sec.~\ref{subsec:From two to three dimensions}, this path gets slippery in 3+1D because we have now to rely on an effective field theory description of the `string foam' formed in spacetime by the proliferation of the dislocations. This section will be devoted to a careful formulation of the Higgs action, with the bottom line that all gauge field components obtain a Higgs gap as usual. We also highlight the complications encountered in the construction of the
dislocation condensate that were already on the foreground in the 2+1D case~\cite{QLC2D} which straightforwardly generalize to 3+1D: the {\em glide} and {\em Ehrenfest constraints} as well as the population of distinct Burgers vectors that is behind the difference between the columnar-, smectic- and nematic-type orders, see Fig.~\ref{fig:phases}.
The machinery is now in place and can be unleashed on the various kinds of quantum liquid crystals. We start with the {\em quantum nematic} order in Sec.~\ref{sec:nematic}. This is defined as a condensate where all
Burgers vector directions contribute equally, completely restoring the translational symmetry while the rotational symmetry is still broken. Resting on the prescription of Sec.~\ref{sec:Dislocation-mediated quantum melting} we find that this 3+1D quantum nematic shares all the traits of the 2+1D version. This acts as a sanity check confirming that the `Higgsing' of Sec.~\ref{sec:Dislocation-mediated quantum melting} does make sense. As in the 2+1D case, we find that the transverse phonons of the solid acquire a mass, indicating that shear stresses can no longer propagate through the liquid at length scales larger than the {\em shear penetration depth}, in close analogy with the way that magnetic forces cannot propagate in an electromagnetic superconductor. In addition, the quantum nematic is also a regular superfluid.
It is the same mechanism as in 2+1D: the glide constraint encodes for the fact that dislocations ``do not carry volume'' and therefore the compressional stress is not affected by the dislocation condensate. The result
is that the longitudinal phonon of the solid turns into the second sound/phase mode of the superfluid. Last but not least, a new feature in 3+1D is the way that the rotational Goldstone bosons (or `torque photons' in stress
language) arise in the quantum nematic. The gross mechanism is the same as in 2+1D; using the `dynamical Ehrenfest constraint' formulation~\cite{BeekmanWuCvetkovicZaanen13,QLC2D} it becomes manifest that these modes are quite literally {\em confined} in the solid, while deconfining and becoming massless in the quantum liquid crystal with a rigidity that is residing in the dislocation condensate itself. The novelty is that in 3+1D we find according to expectations three such modes, that separate in two degenerate `transverse' modes and a `longitudinal' one, characterized by a parametrically different velocity.
As we discovered in 2+1D, the topological melting view offers a most elegant way of also dealing with the quantum smectic type of order. This just exploits the freedom to choose preferential directions for the
Burgers vectors in the dislocation condensate. In the nematic all Burgers directions contribute equally, while in the 2+1D smectic dislocations proliferate with their Burgers vectors oriented in one particular spatial
direction, only restoring translations in that particular dimension. We found that the long-wavelength physics of such quantum smectics is surprisingly rich. Intuitively one expects that a smectic is a system that
is one direction behaving like a liquid, remembering its solid nature in the other direction. However, we found that matters are quite a bit more interesting with the solid and liquid features being `intertwined'
in the literal sense of the word. We show in Sec.~\ref{sec:smectic} that much of the same pattern occurs in 3+1D . This landscape is now enriched by the fact that the dislocations can proliferate with Burgers vectors in
one or two directions, defining the {\em columnar} and {\em smectic} quantum phases. There is room for even more richness to occur. Dealing with the quantum smectic (``stacks of liquid planes''), when the momentum
of the propagating modes lie precisely in the liquid-like plane we find that the response is indistinguishable from a 2D quantum nematic, except for small, dimension-dependent differences in the velocities of the massless modes.
When the momentum lies in a solid--liquid plane it instead behaves like 2D quantum smectic. Precisely along the solid direction a longitudinal phonon is recovered which is at first sight surprising since
the shear modulus is contributing despite the fact that the transverse directions are liquid-like. Last but not least, we find one rotational Goldstone mode associated to the plane where translational symmetry
is restored, in accordance with recent predictions~\cite{WatanabeMurayama13}.
In Sec.~\ref{sec:columnar} we deal with the 3D quantum columnar phase with its two solid directions (``arrays of liquid lines''). We find that the longitudinal phonon and one transverse phonon remain massless, while a second transverse phonon picks up a Higgs mass.
There is also a massive mode due to the fluctuations of the dislocation condensate itself, although these two massive modes are coupled for almost all directions of momentum.
In the special cases that momentum lies exactly in the plane orthogonal to the liquid-like direction, or in a plane with one solid-like and the liquid-like direction, one obtains response
similar to the 2D solid and 2D smectic respectively. Since there is no plane with vanishing shear rigidity, rotational Goldstone modes are absent.
As we showed in QLC2D it is straightforward to extend the theory from neutral substances to electrically charged ones, which is the subject of Sec.~\ref{sec:Charged quantum liquid crystals}. We now depart from a charged `Wigner crystal' keeping track of the coupling to electromagnetic fields
when the duality transformation is carried out. There is now the technical difference that the stress gauge fields have a two-form and the EM gauge fields a one-form nature; the effect is that not all stress fields couple to the electromagnetic fields. As a novelty we find that the `longitudinal' rotational Goldstone mode is a purely neutral entity. Different from its transverse partners, it stays electromagnetically quiet even in the finite-momentum regime where all collective modes turn into electromagnetic observables in 2+1D.
Notwithstanding, the highlights of the 2+1D case all carry over to 3+1D. Most importantly, we show that the quantum nematics are characterized by a genuine electromagnetic Meissner effect proving directly that these are literal superconductors, while smectic and columnar phases have strongly-anisotropic superconductivity.
In Sec.~\ref{sec:Conclusions} we shall discuss the relevance of this work for real-world materials, and highlight roads for future research.
Finally a brief explanation of of our conventions regarding units and terminology. We work almost always in Euclidean time $\tau = \mathrm{i} t$, and the quantum partition function at zero temperature is expressed as an Euclidean path integral $Z = \int
\exp (- \mathcal{S} ) = \int \exp ( - \int \mathrm{d} \tau \mathrm{d}^3 x \, \mathcal{L})$. We employ relativistic notation in which the temporal component $\mathfrak{t} = c \tau$ has units of length, where $c$ is an appropriate velocity, usually the shear velocity $c_\mathrm{T}$. Greek indices $\mu, \nu,\ldots$ run over space and time while Roman indices $m,n,\ldots$ run over space only. Like in QLC2D, we will almost always work in one of two Fourier--Matsubara coordinate systems, where the axes are parallel or orthogonal to momentum. In the first system $(\mathfrak{t},\mathrm{L},\mathrm{R},\mathrm{S})$, the temporal coordinate $\mathfrak{t}$ is unchanged, but the three spatial coordinates are divided into one longitudinal $\mathrm{L}$, and two transverse directions $\mathrm{R},\mathrm{S}$ with respect to the spatial momentum $\mathbf{q}$. The directions $\mathrm{R},\mathrm{S}$ are orthogonal but otherwise arbitrary. The second system $(0,1,\mathrm{R},\mathrm{S})$, has one direction, $0$, parallel to spacetime momentum $p_\mu = (\frac{1}{c} \omega_n,\mathbf{q})$, where $\omega_n$ is a Matsubara frequency. The second direction, $1$ is orthogonal to $p_\mu$, but within the $(\mathfrak{t}\mathrm{L})$-plane, while the transverse direction $\mathrm{R},\mathrm{S}$ are as before. The explicit coordinate transformations are given in Appendix~\ref{sec:Fourier space coordinate systems}, where we make, without loss of generality, one particular choice of axes.
We set $\hbar \equiv 1$ everywhere.
\subsection{Incorporating the glide constraint}\label{subsec:nematic glide constraint}
The first thing we need to do is to implement the glide constraint by integrating out the Lagrange multiplier field $\lambda$. This is done in the usual way: complete the square, shift the integrand and perform the path integral to leave an overall constant factor~\cite{QLC2D}. Eq.~\eqref{eq:dislocation Higgs term nematic} is rewritten as,
\begin{align}
\mathcal{L}_\mathrm{Higgs} &=
\frac{1}{4}\sum_{a=x,y,z} \frac{\Omega^2}{ c_\mathrm{T}^2 \mu} \Big[
2 \lvert \tilde{b}^a_{1\mathrm{R}} \rvert^2 + 2\lvert \tilde{b}^a_{1\mathrm{S}} \rvert^2 + 2\lvert \tilde{b}^a_{\mathrm{R}\mathrm{S}} \rvert^2 \nonumber\\
&\phantom{mm}
+ \lambda^\dagger \lambda \epsilon_{\tau \mu \nu a} \epsilon_{\tau \kappa\lambda a} (\delta_{\mu\kappa} - \frac{\tilde{p}_\mu \tilde{p}_\kappa}{\tilde{p}^2} ) (\delta_{\nu\lambda} - \frac{\tilde{p}_\nu \tilde{p}_\lambda}{\tilde{p}^2} ) \nonumber\\
&\phantom{mm}
+ \lambda^\dagger \epsilon_{\tau \mu \nu a} (\delta_{\mu\kappa} - \frac{\tilde{p}_\mu \tilde{p}_\kappa}{\tilde{p}^2} ) (\delta_{\nu\lambda} - \frac{\tilde{p}_\nu \tilde{p}_\lambda}{\tilde{p}^2} ) \tilde{b}^a_{\kappa\lambda} \nonumber\\
&\phantom{mm}
+ \tilde{b}^{a\dagger}_{\mu\nu} (\delta_{\mu\kappa} - \frac{\tilde{p}_\mu \tilde{p}_\kappa}{\tilde{p}^2} ) (\delta_{\nu\lambda} - \frac{\tilde{p}_\nu \tilde{p}_\lambda}{\tilde{p}^2} ) \epsilon_{\tau \kappa\lambda a} \lambda\Big].
\label{eq:dislocation Higgs term nematic 2}
\end{align}
As usual, the factors of 2 in the first line arise from summing over antisymmetric components. For the last two lines, we recognize that the projectors $(\delta_{\mu\kappa} - \tilde{p}_\mu\tilde{p}_\kappa/\tilde{p}^2)$ etc.
enforce the Lorenz gauge fix on the dual stress gauge fields. The only remaining terms remaining are a quadratic term $\lambda^\dagger\lambda$ and $\lambda^\dagger \epsilon_{\tau mn a} b^a_{mn}$ and its Hermitian conjugate. For the second line, we calculate
\begin{align}
&\epsilon_{\tau \mu \nu a} \epsilon_{\tau \kappa\lambda a} (\delta_{\mu\kappa} - \frac{\tilde{p}_\mu \tilde{p}_\kappa}{\tilde{p}^2} ) (\delta_{\nu\lambda} - \frac{\tilde{p}_\nu \tilde{p}_\lambda}{\tilde{p}^2} ) \nonumber\\
&\phantom{mm}= \epsilon_{\tau \mu \nu a} \epsilon_{\tau \mu\nu a} - 2\epsilon_{\tau \mu \nu a}\epsilon_{\tau \kappa \nu a} \frac{\tilde{p}_\mu \tilde{p}_\kappa}{\tilde{p}^2} \nonumber\\
&\phantom{mm}= 2 - 2 \frac{q^2 - q_a^2}{\tilde{p}^2} = 2\frac{\frac{1}{c_\mathrm{d}^2}\omega_n^2 + q_a^2}{\tilde{p}^2}, \qquad \text{no sum } a.\label{eq:glide constraint multiplier field prefactor}
\end{align}
Here we used $\tilde{p}^2 = \frac{1}{c_\mathrm{d}^2}\omega_n^2 + q^2$. For the nematic, we sum over $a= x,y,z$ to find
\begin{equation}
\sum_{a = x,y,z} 2\frac{\frac{1}{c_\mathrm{d}^2}\omega_n^2 + q_a^2}{\tilde{p}^2} = 2\frac{3 \frac{1}{c_\mathrm{d}^2} \omega_n^2 + q^2}{\tilde{p}^2}.
\end{equation}
All together Eq.~\eqref{eq:dislocation Higgs term nematic 2} becomes
\begin{align}
\mathcal{L}_\mathrm{Higgs} &=
\frac{\Omega^2}{4 c_\mathrm{T}^2 \mu} \Big[ \sum_{a=x,y,z}\Big(
2\lvert \tilde{b}^a_{1\mathrm{R}} \rvert^2 + 2\lvert \tilde{b}^a_{1\mathrm{S}} \rvert^2 + 2\lvert \tilde{b}^a_{\mathrm{R}\mathrm{S}} \rvert^2 \Big) \nonumber\\
&\phantom{mmmmm}
+ \lambda^\dagger \lambda \ 2\frac{3 \frac{1}{c_\mathrm{d}^2} \omega_n^2 + q^2}{\tilde{p}^2}
\nonumber\\
&\phantom{mmmmm}
+ 2\lambda^\dagger (\tilde{b}^x_{yz} +\tilde{b}^y_{zx} +\tilde{b}^z_{xy}) + \mathrm{h.c.}\Big] \nonumber\\
&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu} \Big[ \sum_{a=x,y,z}\Big(
\lvert \tilde{b}^a_{1\mathrm{R}} \rvert^2 + \lvert \tilde{b}^a_{1\mathrm{S}} \rvert^2 + \lvert \tilde{b}^a_{\mathrm{R}\mathrm{S}} \rvert^2 \Big) \nonumber\\
&\phantom{mmnmm}- \frac{\tilde{p}^2}{3 \frac{1}{c_\mathrm{d}^2} \omega_n^2 + q^2} \lvert \tilde{b}^x_{yz} +\tilde{b}^y_{zx} +\tilde{b}^z_{xy} \rvert^2 \Big].\label{eq:dislocation Higgs term nematic 3}
\end{align}
The last equality arises after integrating out the Lagrange multiplier field $\lambda$. We will soon verify that the compression mode remains massless in the dislocation condensate
precisely because of the extra term arising from the glide constraint on the last line of Eq. \eqref{eq:dislocation Higgs term nematic 3}. In the case of the gauge-fixed nematic we can use the relation
\begin{align}
\tilde{b}^x_{yz} +\tilde{b}^y_{zx} +\tilde{b}^z_{xy}
&= \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}} + \mathrm{i} \frac{\omega_n}{c_\mathrm{d} \tilde{p}} \tilde{b}^\mathrm{R}_{1 \mathrm{S}}- \mathrm{i} \frac{\omega_n}{c_\mathrm{d} \tilde{p}} \tilde{b}^\mathrm{S}_{1 \mathrm{R}} \nonumber\\
&= \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}} + \mathrm{i} \sqrt{2} \frac{\omega_n}{c_\mathrm{d} \tilde{p}} \tilde{b}_{1 -}
\end{align}
Furthermore, for any field $A^a$ we have $\sum_{a = x,y,z} \lvert A^a \rvert^2 = \sum_{E = \mathrm{L},\mathrm{R},\mathrm{S}} \lvert A^E \rvert^2$. For the nematic, the Higgs term in the Lagrangian splits up
into the same five sectors as the elastic solid Lagrangian Eq.~\eqref{eq:solid stress gauge field Lagrangian tilde}, where only the $\mathrm{L} 1$-sector is modified by the glide constraint:
\begin{widetext}
\begin{align}\label{eq:nematic Higgs term gauge fields}
\mathcal{L}_\mathrm{Higgs} &=
\frac{\Omega^2}{2 c_\mathrm{T}^2 \mu} \bigg[ ( \lvert \tilde{b}^\mathrm{L}_{1\mathrm{S}} \rvert^2 + \lvert \tilde{b}^\mathrm{R}_{\mathrm{R}\mathrm{S}}\rvert^2 ) + ( \lvert \tilde{b}^\mathrm{L}_{1\mathrm{R}} \rvert^2 + \lvert \tilde{b}^\mathrm{S}_{\mathrm{R}\mathrm{S}}\rvert^2 ) + ( \lvert \tilde{b}^\mathrm{R}_{1\mathrm{R}} \rvert^2 + \lvert \tilde{b}^\mathrm{S}_{1\mathrm{S}}\rvert^2 )
+\lvert b_{1+} \rvert^2 \nonumber\\
& \phantom{mmmmm}
+ \begin{pmatrix} \tilde{b}^{\dagger}_{1 -} \\ \tilde{b}^{\mathrm{L}\dagger}_{\mathrm{R} \mathrm{S} } \end{pmatrix}^\mathrm{T}
\Big[ \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} -
\frac{1}{3 \frac{1}{c_\mathrm{d}^2} \omega_n^2 + q^2}
\begin{pmatrix}
2\frac{1}{c_\mathrm{d}^2}\omega_n^2 & -\mathrm{i} \sqrt{2} \tfrac{1}{c_\mathrm{d}} \omega_n \tilde{p} \\
\mathrm{i} \sqrt{2}\tfrac{1}{c_\mathrm{d}} \omega_n \tilde{p} & \tilde{p}^2
+ \lvert b_{1+} \rvert^2 \end{pmatrix} \Big]
\begin{pmatrix} \tilde{b}^{\dagger}_{1 -} \\ \tilde{b}^{\mathrm{L}\dagger}_{\mathrm{R} \mathrm{S}} \end{pmatrix}\bigg].
\end{align}
\end{widetext}
\subsection{Collective modes of the quantum nematic}\label{subsec:Collective modes in the quantum nematic}
To obtain the spectrum of modes in the nematic phase, we should add Eq.~\eqref{eq:dislocation Higgs term nematic 3} to Eq.~\eqref{eq:solid stress gauge field Lagrangian tilde} and calculate the propagators Eqs.~\eqref{eq:longitudinal propagator gauge field tilde}, \eqref{eq:transverse propagator gauge field tilde}. After a Wick rotation to real frequency $\omega$, we obtain the first main result of this paper, the stress propagators of the isotropic 3D quantum nematic:
\begin{widetext}
\begin{align}
G_\mathrm{L} &= \frac{1}{\mu} \frac{ - c_\mathrm{T}^2 q^2 (\omega^2 - \tfrac{1}{3}c_\mathrm{d}^2 q^2 - \Omega^2)}{(\omega^2 - c_\mathrm{L}^2 q^2)(\omega^2 - \tfrac{1}{3}c_\mathrm{d}^2 q^2) - \Omega^2 (\omega^2 - c_\kappa^2 q^2)},\label{eq:nematic longitudinal propagator}\\
G_{\mathrm{T} 1} = G_{\mathrm{T} 2} &= \frac{1}{\mu} \frac{ -c_\mathrm{T}^2 q^2 (\omega^2 -c_\mathrm{d}^2 q^2) - \Omega^2 ( \omega^2 - 2 c_\mathrm{T}^2 q^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2 - \Omega^2 )}{(\omega^2 - c_\mathrm{T}^2 q^2)(\omega^2 - c_\mathrm{d}^2 q^2) - \Omega^2 (\omega^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2)},\label{eq:nematic GT1}\\
G_{\mathrm{T} 3} &= \frac{1}{\mu} \frac{-\Omega^2}{\omega^2 - c_\mathrm{d}^2 q^2}.\label{eq:nematic GT3}
\end{align}
Recall that $c_\mathrm{L}$ and $c_\mathrm{T}$ are the velocities of the longitudinal and transverse phonons, respectively, while $c_\mathrm{d}$ is the velocity we assigned to the dislocation condensate. We defined here $c_\kappa^2 = \frac{\kappa}{\rho} = \frac{2}{D} \frac{ 1+ \nu}{1 - (D-1)\nu} c_\mathrm{T}^2$ as the {\em compression velocity} depending only on the compression modulus $\kappa$ and not the shear modulus $\mu$. This sets the (second) sound velocity in the quantum liquid.
It is useful to compare these to the corresponding expressions in two dimensions from QLC2D, where we use the appropriate definitions of $c_{\mathrm{L},\mathrm{2D}}$ and $c_{\kappa,\mathrm{2D}} = \sqrt{\kappa / \rho}$ via Eqs.~\eqref{eq:longitudinal velocity definition}, \eqref{eq:Poisson ratio definition},
\begin{align}
G^\mathrm{2D}_\mathrm{L} &= \frac{1}{\mu} \frac{-c_\mathrm{T}^2 q^2 ( \omega^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2 - \Omega^2)}{(\omega^2 - c_{\mathrm{L},\mathrm{2D}}^2 q^2)(\omega^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2) - \Omega^2 (\omega^2 - c_{\kappa,\mathrm{2D}}^2 q^2)}\label{eq:2D nematic GL},\\
G^\mathrm{2D}_\mathrm{T} &= \frac{1}{\mu}\frac{-c_\mathrm{T}^2 q^2 (\omega_n^2 - c_\mathrm{d}^2 q^2) - \Omega^2 ( \omega^2 - 2 c_\mathrm{T}^2 q^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2 - \Omega^2 )}{(\omega^2 - c_\mathrm{T}^2 q^2)(\omega^2 - c_\mathrm{d}^2 q^2) - \Omega^2( \omega^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2)}.\label{eq:2D nematic GT}
\end{align}
\end{widetext}
We see that the longitudinal propagator is the same apart from the definitions of $c_\mathrm{L}$ and $c_\kappa$ and a dimensional factor $1/D$ in front of $c_\mathrm{d}^2$. Furthermore, the transverse propagators associated with the `remnant' transverse phonons $G_{\mathrm{T} 1}, G_{\mathrm{T} 2}$ are seen to be completely independent of dimensionality. The only novelty is the $G_{\mathrm{T} 3}$ propagator which vanishes in the solid
but now describes a massless mode propagating with the condensate velocity in the quantum nematic.
\begin{figure*}
\begin{center}
\includegraphics[width=7.7cm]{3DnemT.png}
\hfill
\includegraphics[width=7.7cm]{3DnemL.png}\\
{\centering
\includegraphics[scale=.6]{legends_nematic.pdf}
}\\
\caption{Spectral functions of the quantum nematic in units of the inverse shear modulus $1/\mu$, with Poisson ratio $\nu = 0.2$. For a clear picture we have arbitrarily set $c_\mathrm{d} = 2 c_\mathrm{T}$. The inset is a zoom up near the origin. The width of the poles is artificial and denotes the relative pole strengths: these ideal poles are actually infinitely sharp. a) Transverse spectral function of $G_\mathrm{T} = G_{\mathrm{T} 1} + G_{\mathrm{T} 2} + G_{\mathrm{T} 3}$ from Eqs.~\eqref{eq:nematic GT1}, \eqref{eq:nematic GT3}. At low energies we find the massless rotational Goldstone modes with velocities $\frac{1}{\sqrt{2}} c_\mathrm{d}$ and $c_\mathrm{d}$ respectively. There is also a massive pole due the gapped shear phonons with Higgs gap $\Omega$, which have vanishing pole strength as $q\to 0$. At high energies we retrieve the transverse phonons with velocity $c_\mathrm{T}$ while the condensate modes with velocity $c_\mathrm{d}$ have vanishing pole strength as $q \to \infty$. b) Longitudinal spectral function of $G_\mathrm{L}$ from Eq.~\eqref{eq:nematic longitudinal propagator}. The massless pole has the pure compression velocity $c_\kappa$ at low energy while it extrapolates to the longitudinal phonon at high momenta. There is also a gapped condensate mode with gap $\Omega$ that extrapolates to a linearly dispersing mode with velocity $\frac{1}{\sqrt{3}} c_\mathrm{d}$ at high momentum, with vanishing pole strength both as $q \to 0$ and as $q \to \infty$.}\label{fig:nematic spectral functions}
\end{center}
\end{figure*}
The longitudinal ($\mathrm{L}$) and transverse ($\mathrm{T} 1, \mathrm{T} 2$) propagators were analyzed already in detail in the 2+1D case\cite{QLC2D} but let us repeat this exercise here for completeness.
The spectral functions are plotted in Fig.~\ref{fig:nematic spectral functions}. Let us first focus on the longitudinal response Eq.~\eqref{eq:nematic longitudinal propagator}. This propagator reveals
one massless and one massive pole with gap $\Omega$. At small energies and to lowest order in momentum their dispersion relations are
\begin{align}
\omega^{\mathrm{L}}_1 &= c_\kappa q + \ldots ,\nonumber \\
\omega^{\mathrm{L}}_2 &= \Omega + \frac{c_\mathrm{d}^2 + 4 c_\mathrm{T}^2}{6 \Omega} q^2 + \ldots . \label{eq:nematic longitudinal dispersions}
\end{align}
The massless $\omega^{\mathrm{L}}_1$ pole reveals the second sound mode, characterized by the purely compressional velocity $c_\kappa = \sqrt{\kappa/\rho}$ instead of the longitudinal phonon velocity $c_\mathrm{L} = \sqrt{(\kappa + 2 \mu)/\rho}$.
This was a highlight of QCL2D: the dislocation condensate has destroyed the shear rigidity of the solid at long wavelength, while it does not affect the compressional property of the solid since the dislocations ``do not carry volume'' as expressed by the glide constraint. In more detail, the structure of the longitudinal propagator reveals that the longitudinal phonon acquires a mode coupling with a condensate mode that effectively removes its shear rigidity, turning it into a pure sound mode. The second mode $\omega^{\mathrm{L}}_2$ is the massive counterpart; at $q=0$ the mode coupling disappears (the shear component of the longitudinal phonon arises only for finite spatial gradients) and then $\omega^\mathrm{L}_2$ is characterized by the Higgs mass of the stress superconductor $\Omega$, the only scale in the problem. It propagates with a combination of the transverse and condensate velocities, revealing its mixed origin.
At very high energies, the two poles disperse linearly with velocities $c_\mathrm{L}$ and $\sqrt{1/3}c_\mathrm{d}$ respectively. High momentum means small length scales, and we expect to retrieve here the signature of the underlying crystal lattice. This is indeed
the case: at large momenta the mode coupling with the condensate mode switches off and one recovers the longitudinal phonon, while the condensate mode turns `pure', propagating with just $c_\mathrm{d}$. The reader
should notice that the sound mode is completely universal at long wavelength: this can be continued adiabatically all the way to the gaseous limit described by the Bogoliubov theory of the weakly interacting Bose gas. However,
the ``massive shear photon'' is special for the maximal-correlation limit. As a finite-energy mode, it requires the ``locally solid'' correlations in the liquid to propagate; in approaching the gaseous limit it will get increasingly damped
to disappear completely in the weakly-interacting limit. One may want to view the roton found in $^4$He as a remnant of this mode in the regime where the crystalline correlation length becomes of the order of the interatomic distance.
\begin{table}
\begin{tabular}{ccc}
\toprule
sector & massless modes & massive modes \\
\hline
$\mathrm{L}$ & $c_\kappa q $ & $\Omega + \frac{\frac{1}{2}c_\mathrm{d}^2 + 2 c_\mathrm{T}^2}{3 \Omega} q^2 +\ldots$ \\
$\mathrm{T} 1$ & $\frac{1}{\sqrt{2}} c_\mathrm{d} q$ & $\Omega + \frac{\frac{1}{2}c_\mathrm{d}^2 + c_\mathrm{T}^2}{2 \Omega} q^2 +\ldots$ \\
$\mathrm{T} 2$ & $\frac{1}{\sqrt{2}} c_\mathrm{d} q$ & $\Omega + \frac{\frac{1}{2}c_\mathrm{d}^2 + c_\mathrm{T}^2}{2 \Omega} q^2 +\ldots$ \\
$\mathrm{T} 3$ & $c_\mathrm{d} q$ & - \\
\hline
total & 4 & 3 \\
\botrule
\end{tabular}
\caption{Collective modes in the isotropic quantum nematic. Indicated are the dispersion relations to lowest orders in momentum. The longitudinal phonon is protected by the glide constraint from obtaining a Higgs gap, and turns into a purely compressional mode. The gapped mode could be seen as the shear parts of the longitudinal phonon `being eaten' by the dislocation condensate. In the transverse sectors we see the two transverse phonons picking up a Higgs gap. Furthermore there are three rotational Goldstone modes emerging (deconfining), two with velocity $\frac{1}{\sqrt{2}}c_\mathrm{d}$ and one with velocity $c_\mathrm{d}$.}\label{table:nematic spectrum}
\end{table}
Let us now turn to the transverse sector,
and first consider the $\mathrm{T} 1, \mathrm{T} 2$ sectors, which were already identified as identical to the transverse sector in 2+1D. As in the longitudinal sector, these propagators describe two modes with the following dispersions at small momenta:
\begin{align}
\omega^{\mathrm{T} 1,2}_1 &= \frac{1}{\sqrt{2}} c_\mathrm{d} q + \ldots ,\label{eq:transverse rotational Goldstone dispersions}\\
\omega^{\mathrm{T} 1,2}_2 &= \Omega + \frac{c_\mathrm{d}^2 + 2 c_\mathrm{T}^2}{4 \Omega} q^2 + \ldots .
\end{align}
The massive modes $\omega^{\mathrm{T} 1,2}_2$ are the transverse phonons which have acquired a mass through the Anderson--Higgs mechanism. This comprises one of the main predictions of the dual gauge field theory of these maximally-correlated quantum liquid crystals: transverse phonons do not disappear but should be detectable as massive, propagating modes in the spectrum~\cite{ZaanenNussinovMukhin04,CvetkovicNussinovMukhinZaanen08,QLC2D}. Just as for the longitudinal sector, these modes should get increasingly damped when the solid correlations weaken, to completely disappear in the weakly-coupled, gaseous limit.
The massless poles $\omega^{\mathrm{T} 1,2}_1$ are universal: exactly the same mode is found in 2+1D where we identified it as the {\em rotational Goldstone mode} or {\em torque mode} in the stress formalism, since it propagates torque stresses within the quantum nematic~\cite{BeekmanWuCvetkovicZaanen13,QLC2D}. It is an exclusive feature of the zero-temperature superconducting/superfluid nematic. It does not exist in the `high-temperature' classic nematic fluid, the reason being a the `anomaly' of the Goldstone physics in liquid crystals~\cite{DeGennesProst95,ChaikinLubensky00} mentioned in Sec.~\ref{sec:Symmetry principles of quantum liquid crystals}. The trouble is rooted in the fact that the classical nematic fluid is also a regular hydrodynamical fluid. This supports circulation and it turns out that the rotational Goldstone bosons have a finite coupling to this hydrodynamical circulation even in the long-wavelength limit. The effect is that the rotational boson just gets overdamped by the coupling to this circulation. This situation changes drastically in the superfluid/superconductor. Circulation is now massive because of the quantization of the vorticity and at low energies the rotational Goldstone modes cannot be damped, but constitute the propagating modes.
As will be highlighted in the next subsection \ref{subsec:Torque stress in the quantum nematic}, the recovery of these rotational Goldstone bosons is a highlight of the duality construction. They are `confined' in the solid as we already discussed in Sec.~\ref{subsec:Torque stress gauge fields}, while they become deconfined (becoming massless and propagating) in the nematic. As elucidated in QLC2D the associated {\em torque rigidity} originates in the dislocation condensate itself:
is is observed that the velocity of the modes $\omega^{\mathrm{T} 1,2}_1$ is set by the condensate velocity $c_\mathrm{d}$~\cite{BeekmanWuCvetkovicZaanen13,QLC2D}. Furthermore, the pole strength is proportional to $\Omega^2$, and therefore vanishes when the dislocation condensate is absent.
The only difference between 2+1D and 3+1D is in the number of rotational Goldstone modes. In two space dimensions there is only one rotational plane where the rotational symmetry is broken ($O(2)$), and accordingly there is one Goldstone boson. In three dimensions there are three rotational planes ($O(3)$) and considering an `isotropic nematic' (or, say,
a cubic $O_\mathrm{h}$ nematic) the rotational symmetry is broken in all three planes, causing the existence of three rotational Goldstone bosons. The reader should notice that the most common uniaxial nematics are in this regard quite pathological. Their point group is $D_{\infty \mathrm{h}}$, characterized by breaking symmetry in two rotational planes, and therefore only two rotational Goldstones are present.
Where is the third rotational Goldstone boson? It is found in the $G_{\mathrm{T} 3}$ propagator. As we stressed in Eq.~\eqref{eq:GT3 solid}, this propagator is vanished identically in the solid but in the nematic it turns into the propagator of a single massless mode,
\begin{equation}
\omega^{\mathrm{T} 3} = c_\mathrm{d} q. \label{eq:longitudinal rotational Goldstone dispersions}
\end{equation}
Compared to the $\omega^{\mathrm{T} 1,2}$-modes its velocity larger by a factor $\sqrt{2}$, indicating that this mode is symmetry-wise not equivalent to the other two. As we will see in a moment, this is due to the fact that the
`polarizations' of these rotational Goldstone modes count in the same way as for phonons: the ${\mathrm{T} 1,2}_1$ modes are `transverse' while the ${\mathrm{T} 3}$ modes turns out to be `longitudinally polarized'. The formalism discussed in the next subsection will yield further insights in these torque stresses.
For completeness, we ask what happens at large momenta? The two modes described by each of the $G_{\mathrm{T} 1}$, $G_{\mathrm{T} 2}$ disperse linearly with velocities $c_\mathrm{T}$ resp. $c_\mathrm{d}$. On short length scales, we retrieve the transverse phonons of the solid, and the decoupled modes of the dislocation condensate with vanishing spectral weight as $q \to \infty$, just as in the longitudinal sector. The single mode in $G_{\mathrm{T} 3}$ has the velocity $c_\mathrm{d}$ at all momenta, with spectral weight $\propto 1/q$.
\subsection{Torque stress in the quantum nematic}\label{subsec:Torque stress in the quantum nematic}
We can obtain a better understanding of the rotational Goldstone modes by considering the torque stress and torque stress gauge fields of Sec.~\ref{subsec:Torque stress gauge fields}. Here we are interested in the physics at frequencies
small compared to the Higgs mass, and we send $\Omega \to \infty$, keeping only the Higgs term and ignoring the `phonon part'.
We work with the torque stress fields
\begin{align}
\tilde{\tau}^c_\mu &= \epsilon_{cba} \epsilon_{b\mu \kappa\lambda}\tfrac{1}{2} \tilde{b}^a_{\kappa\lambda},\label{eq:torque stress definition condensate unitary gauge}\\
&= \epsilon_{\mu\nu\kappa\lambda} \tilde{\partial}_\nu \tfrac{1}{2} \tilde{h}^c_{\kappa\lambda}.
\end{align}
where all fields are rescaled with respect to the dislocation velocity $c_\mathrm{d}$, as we have done throughout this section.
This definition of the torque stress is not gauge invariant, which in the duality is attributed to the presence of shear rigidity. However, in the dislocation condensate shear rigidity is lost and torque stress becomes a
physical quantity. In the presence of the dislocation condensate phase degrees of freedom $\tilde{\phi}^a_\kappa$, the torque stress is rather defined as
\begin{equation}\label{eq:torque stress definition condensate}
\tilde{\tau}^c_\mu = \epsilon_{cba} \epsilon_{b\mu \kappa\lambda} (\tilde{b}^a_{\kappa\lambda} + \tilde{\partial}_\kappa \tilde{\phi}^a_\lambda -\tilde{\partial}_\lambda \tilde{\phi}^a_\kappa).
\end{equation}
being gauge invariant under the transformations Eq.~\eqref{eq:stress gauge transformation dislocation condensate}. Let us take the unitary gauge fix $\tilde{\phi}^a_\lambda = 0$, bringing us back to Eq.~\eqref{eq:torque stress definition condensate unitary gauge}.
The Higgs term becomes Eq.~\eqref{eq:dislocation condensate minimal coupling step 1}, supplemented by the glide constraint, while the Higgs mass is rescaled by a factor of 2 like in Eq.~\eqref{eq:dislocation Higgs term nematic}:
\begin{align}\label{eq:nematic Higgs term unitary gauge fix}
\mathcal{L}_\mathrm{Higgs} &= \frac{\Omega^2}{4 c_\mathrm{T}^2 \mu } \Big[ \lvert \tilde{b}^a_{\mu\nu} \rvert^2 - \frac{\tilde{p}^2}{3 \frac{\omega_n^2}{c_\mathrm{d}^2} + q^2} \lvert \tilde{b}^x_{yz} + \tilde{b}^y_{zx} + \tilde{b}^z_{xy} \Big]\nonumber\\
&= \frac{\Omega^2}{4 c_\mathrm{T}^2 \mu } \Big[ (2 \lvert \tilde{b}^E_{\mathfrak{t} G} \rvert^2 + \lvert \tilde{b}^E_{F G} \rvert^2 \nonumber\\
&\phantom{mmmmm} - \frac{\tilde{p}^2}{3 \frac{\omega_n^2}{c_\mathrm{d}^2} + q^2} \lvert \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}} - \tilde{b}^\mathrm{R}_{\mathrm{S}\mathrm{L}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}} \rvert^2 \Big].
\end{align}
Here the capital indices sum over Fourier components $E,F,G = \{ \mathrm{L},\mathrm{R},\mathrm{S}\}$ and we have used the identity $\tilde{b}^x_{yz} + \tilde{b}^y_{zx} + \tilde{b}^z_{xy} = \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}} - \tilde{b}^\mathrm{R}_{\mathrm{S}\mathrm{L}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}}$. Some components of $\tilde{b}^a_{\kappa\lambda}$, which were completely absent in the solid Lagrangian, now appear. These could be said to originate in the dislocation condensate itself, transferred to the dual gauge fields via the Anderson--Higgs mechanism.
In the unitary gauge fix, the explicit correspondence between $\tilde{\tau}^c_\mu$ and $\tilde{b}^a_{\kappa\lambda}$ is just Eq.~\eqref{eq:torque stress in stress gauge field components} with all fields replaced by their tilde-equivalents. We can now express the Higgs term explicitly in the torque stress:
\begin{widetext}
\begin{align}
\mathcal{L}_\mathrm{Higgs} &= \mathcal{L}_{\mathrm{T1, Higgs}} + \mathcal{L}_{\mathrm{T2, Higgs}} + \mathcal{L}_{\mathrm{T3, Higgs}} + \mathcal{L}_{\mathrm{L1, Higgs}} + \mathcal{L}_{\mathrm{L2, Higgs}} + \mathcal{L}_{\mathrm{X, Higgs}}, \\
\mathcal{L}_{\mathrm{T1, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[ \frac{1}{2} \lvert \tilde{\tau}^\mathrm{S}_\mathfrak{t}\rvert^2 + \lvert \tilde{\tau}^\mathrm{S}_\mathrm{L}\rvert^2 + \frac{1}{2} \lvert \tilde{b}^\mathrm{L}_{\mathrm{L}\mathrm{S}} + \tilde{b}^\mathrm{R}_{\mathrm{R}\mathrm{S}}\rvert^2\Big],\\
\mathcal{L}_{\mathrm{T2, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[ \frac{1}{2} \lvert \tilde{\tau}^\mathrm{R}_\mathfrak{t}\rvert^2 + \lvert \tilde{\tau}^\mathrm{R}_\mathrm{L}\rvert^2 + \frac{1}{2} \lvert \tilde{b}^\mathrm{L}_{\mathrm{L}\mathrm{R}} + \tilde{b}^\mathrm{S}_{\mathrm{R}\mathrm{S}}\rvert^2\Big],\\
\mathcal{L}_{\mathrm{T3, Higgs}} &= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[ \frac{1}{2} \lvert \tilde{\tau}^\mathrm{L}_\mathfrak{t}\rvert^2
+ \frac{1}{2} \lvert \tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{R}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{S}}\rvert^2
+ \frac{3}{4} \lvert \tilde{\tau}^\mathrm{L}_\mathrm{L} \rvert^2
+ \frac{3}{4} \lvert \tilde{\tau}^\mathrm{R}_\mathrm{R} \rvert^2
+ \frac{3}{4} \lvert \tilde{\tau}^\mathrm{S}_\mathrm{S} \rvert^2 \nonumber\\
&\phantom{mmmmmmm}
+ \frac{1}{4} (\tilde{\tau}^{\mathrm{L}\dagger}_\mathrm{L} \tilde{\tau}^\mathrm{R}_\mathrm{R} +\tilde{\tau}^{\mathrm{R}\dagger}_\mathrm{R} \tilde{\tau}^\mathrm{L}_\mathrm{L})
+ \frac{1}{4} (\tilde{\tau}^{\mathrm{R}\dagger}_\mathrm{R} \tilde{\tau}^\mathrm{S}_\mathrm{S} +\tilde{\tau}^{\mathrm{S}\dagger}_\mathrm{S} \tilde{\tau}^\mathrm{R}_\mathrm{R})
- \frac{1}{4} (\tilde{\tau}^{\mathrm{L}\dagger}_\mathrm{L} \tilde{\tau}^\mathrm{S}_\mathrm{S} +\tilde{\tau}^{\mathrm{S}\dagger}_\mathrm{S} \tilde{\tau}^\mathrm{L}_\mathrm{L})
\Big],\\
\mathcal{L}_{\mathrm{L1, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[
\frac{1}{2} \lvert \tilde{\tau}^\mathrm{S}_\mathrm{R} - \tilde{\tau}^\mathrm{R}_\mathrm{S}\rvert^2
+ \lvert \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}}\rvert^2
+ \frac{1}{2}\lvert\tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{S}} - \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}} \rvert^2
- \frac{\tilde{p}^2}{3 \frac{\omega_n^2}{c_\mathrm{d}^2} + q^2} \lvert \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}} - \tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{S}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}} \rvert^2
\Big],\\
\mathcal{L}_{\mathrm{L2, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[
\frac{1}{2} \lvert \tilde{\tau}^\mathrm{S}_\mathrm{R} + \tilde{\tau}^\mathrm{R}_\mathrm{S}\rvert^2 +
\frac{1}{2}\lvert\tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{S}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}} \rvert^2
\Big],\\
\mathcal{L}_{\mathrm{X, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[\lvert \tilde{\tau}^\mathrm{L}_\mathrm{R}\rvert^2 + \lvert \tilde{\tau}^\mathrm{L}_\mathrm{S}\rvert^2 \Big].
\end{align}
\end{widetext}
Here the naming of different sectors follows that of Eq.~\eqref{eq:solid stress gauge field Lagrangian}, although this classification is slightly ambiguous. In particular the components $\tilde{b}^E_{\mathfrak{t}\mathrm{L}}$ do not have a counterpart in linear elasticity. This also leads to the introduction of a new sector $\mathcal{L}_{\mathrm{X}}$ which is decoupled from all others. We will comment on the interpretation of these degrees of freedom below. The next step is to substitute the torque stress gauge field Eq.~\eqref{eq:torque stress gauge field definition}, in the Lorenz gauge fix Eq.~\eqref{eq:torque stress gauge field components}. This leads to
\begin{widetext}
\begin{align}
\mathcal{L}_{\mathrm{T1, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[ (\frac{\omega_n^2}{c_\mathrm{d}^2} + \frac{1}{2} q^2) \lvert \tilde{h}^\mathrm{S}_{\mathrm{R}\mathrm{S}} \rvert^2 + \frac{1}{2} \lvert \tilde{b}^\mathrm{L}_{\mathrm{L}\mathrm{S}} + \tilde{b}^\mathrm{R}_{\mathrm{R}\mathrm{S}}\rvert^2\Big],\\
\mathcal{L}_{\mathrm{T2, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[ (\frac{\omega_n^2}{c_\mathrm{d}^2} + \frac{1}{2} q^2) \lvert \tilde{h}^\mathrm{S}_{\mathrm{R}\mathrm{S}} \rvert^2 + \frac{1}{2} \lvert \tilde{b}^\mathrm{L}_{\mathrm{L}\mathrm{R}} + \tilde{b}^\mathrm{S}_{\mathrm{R}\mathrm{S}}\rvert^2\Big],\\
\mathcal{L}_{\mathrm{T3, Higgs}} &= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[ \frac{1}{4}
\begin{pmatrix}
\tilde{h}^{\mathrm{L}\dagger}_{\mathrm{R}\mathrm{S}} & \tilde{h}^{\mathrm{R}\dagger}_{1 \mathrm{S}} & \tilde{h}^{\mathrm{S}\dagger}_{1 \mathrm{R}}
\end{pmatrix}
\begin{pmatrix}
3 \frac{\omega_n^2}{c_\mathrm{d}^2} + 2 q^2 & - \mathrm{i} \frac{\omega_n}{c_\mathrm{d}} \tilde{p} & \mathrm{i} \frac{\omega_n}{c_\mathrm{d}} \tilde{p} \\
\mathrm{i} \frac{\omega_n}{c_\mathrm{d}} \tilde{p} & 3 \tilde{p}^2 & \tilde{p}^2 \\
- \mathrm{i} \frac{\omega_n}{c_\mathrm{d}} \tilde{p} & \tilde{p}^2 & 3 \tilde{p}^2
\end{pmatrix}
\begin{pmatrix}
\tilde{h}^{\mathrm{L}}_{\mathrm{R}\mathrm{S}} \\ \tilde{h}^{\mathrm{R}}_{1 \mathrm{S}} \\ \tilde{h}^{\mathrm{S}}_{ 1 \mathrm{R}}
\end{pmatrix}
+ \frac{1}{2} \lvert \tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{R}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{S}}\rvert^2
\Big],\\
\mathcal{L}_{\mathrm{L1, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[
\frac{1}{2} \tilde{p}^2 \lvert \tilde{h}^\mathrm{S}_{1\mathrm{S}} - \tilde{h}^\mathrm{R}_{1\mathrm{R}} \rvert^2
+ \lvert \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}}\rvert^2
+ \frac{1}{2}\lvert\tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{S}} - \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}} \rvert^2
- \frac{\tilde{p}^2}{3 \frac{\omega_n^2}{c_\mathrm{d}^2} + q^2} \lvert \tilde{b}^\mathrm{L}_{\mathrm{R}\mathrm{S}} - \tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{S}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}} \rvert^2
\Big],\\
\mathcal{L}_{\mathrm{L2, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[
\frac{1}{2} \tilde{p}^2 \lvert \tilde{h}^\mathrm{S}_{1\mathrm{S}} + \tilde{h}^\mathrm{R}_{1\mathrm{R}} \rvert^2
+ \frac{1}{2}\lvert\tilde{b}^\mathrm{R}_{\mathrm{L}\mathrm{S}} + \tilde{b}^\mathrm{S}_{\mathrm{L}\mathrm{R}} \rvert^2
\Big],\\
\mathcal{L}_{\mathrm{X, Higgs}}&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu } \Big[ \tilde{p}^2 \lvert \tilde{h}^\mathrm{L}_{1\mathrm{S}}\rvert^2 + \tilde{p}^2 \lvert \tilde{h}^\mathrm{L}_{1\mathrm{R}} \rvert^2 \Big].
\end{align}
\end{widetext}
We have arrived at a point where the interpretation of the nature of the torque stress carried by the quantum nematic becomes clear. From Sec.~\ref{subsec:Torque stress gauge fields} we know that $h^c_{\mathrm{R}\mathrm{S}}$ represent the propagating rotational (Goldstone) modes while $h^c_{1 \mathrm{R}}$, $h^c_{1 \mathrm{S}}$ represent static forces. In $\mathcal{L}_{\mathrm{T1, Higgs}}$, $\mathcal{L}_{\mathrm{T2, Higgs}}$ we find two rotational Goldstone modes $h^\mathrm{S}_{\mathrm{R}\mathrm{S}}$, $h^\mathrm{R}_{\mathrm{R}\mathrm{S}}$ with velocity $\frac{1}{\sqrt{2}} c_\mathrm{d}$. Inverting the matrix in $\mathcal{L}_{\mathrm{T3, Higgs}}$, we find a third rotational Goldstone mode $h^\mathrm{L}_{\mathrm{R}\mathrm{S}}$ with velocity $c_\mathrm{d}$. This confirms our findings
summarized in Eqs.~\eqref{eq:transverse rotational Goldstone dispersions}, \eqref{eq:longitudinal rotational Goldstone dispersions}. Similarly, we can see that the static forces are also deconfined, and mediate long-range interactions between disclination sources (as usual, one should
mobilize the Coulomb gauge to find out that for instance $\langle \tilde{h}^{\mathrm{R}\dagger}_{\mathfrak{t}\mathrm{R}} \tilde{h}^{\mathrm{R}}_{\mathfrak{t}\mathrm{R}}\rangle \propto 1/q^2$).
Let us reconsider the discussion regarding the number of degrees of freedom that we started in Sec.~\ref{subsec:Interpretation of stress components}. In the 3+1D solid, we start with 12 stress components $\sigma^a_\mu$. There are three conservation laws $\partial_\mu \sigma^a_\mu =0$ and three Ehrenfest constraints $\sigma^a_m = \sigma^m_a$, such that we are left with six physical stress components: three phonons and three Coulomb forces. Going to
the dual stress gauge fields $b^a_{\kappa\lambda}$, there are at first 18 independent components due to antisymmetry under $\kappa \leftrightarrow \lambda$. In the solid, by accounting for the gauge freedom and Ehrenfest constraints, these still encode for the same six degrees of freedom in Eqs.~\eqref{eq:stress gauge field propagator electric shear}--\eqref{eq:stress gauge field propagator magnetic shear} in the limit $\ell \to 0$ Eq.~\eqref{eq:rotational force suppression}. In the Higgs phase however, {\em all} components of $b^a_{\kappa\lambda}$ obtain a physical meaning. The Anderson-Higgs mechanism transfers the dislocation phase degrees of freedom to the dual stress gauge field in Eq.~\eqref{eq:dislocation condensate minimal coupling step 1} when taking the unitary gauge fix as in Eq.~\eqref{eq:nematic Higgs term unitary gauge fix}. As long as one is solely interested in correlations in linear stress $\sigma^a_\mu$, only the nine gauge-invariant components of $b^a_{\kappa\lambda}$, which are explicitly employed in Eqs.~\eqref{eq:solid stress gauge field Lagrangian},\eqref{eq:nematic Higgs term gauge fields}, are accessible. These contain the spectrum as enumerated in Table~\ref{table:nematic spectrum}. However, as we have seen just now, by looking at correlations in torque stress $\tau^c_\mu$, other degrees of freedom become activated, which originate in the dislocation condensate. This is the best evidence we have for the applicability of the `stringy' form of the minimal couping Eq.~\eqref{eq:dislocation condensate minimal coupling step 1}.
We can also elucidate the difference between the rotational Goldstone modes in $\mathcal{L}_{\mathrm{T} 1}$, $\mathcal{L}_{\mathrm{T} 2}$ on the one hand and $\mathcal{L}_{\mathrm{T} 3}$ on the other hand. The Goldstone mode $h^\mathrm{L}_{\mathrm{R}\mathrm{S}}$ couples to disclinations $\Theta^\mathrm{L}_{\mathrm{R}\mathrm{S}}$ such that the Frank vector is parallel to the momentum. Apparently the longitudinal Goldstone mode, propagating in the direction of the Frank vector, has a different, typically higher, velocity than those propagating perpendicular to the Frank vector, i.e. within the rotational plane. This Goldstone mode is excited by probing the medium with a torque in the plane of the surface and measuring that torque on the opposite side of the medium, see Fig.~\ref{fig:rotational Goldstone modes}.
Let us clarify the origin of the $\sqrt{2}$ difference between the longitudinal and transverse velocity. The torque stress gauge fields $h^c_{\kappa\lambda}$ are dual to rotational fields $\omega^c$. We have summarized the theory of rotational elasticity in Sec.~\ref{subsec:Rotational elasticity}, where we noted there are longitudinal and transverse velocities in Eq.~\ref{eq:rotational velocities}. We see that for the relation $c_\mathrm{L}^\mathrm{rot} = \sqrt{2} c_\mathrm{T}^\mathrm{rot}$ is satisfied when the rotational Poisson ratio in Eq.~\eqref{eq:rotational Poisson ratio} is $\nu^\mathrm{rot} = 0$. This makes perfect sense: a non-zero Poisson ratio would mean that external `longitudinal' torques would also excite transverse rotational modes. In the nematic liquid crystal we expect no such couplings: all the modes should be strictly independent to lowest order.
\subsection{Elasticity as a quantum field theory}\label{subsec:Field-theoretic elasticity}
In QLC2D we provided an exposition of the quantum-field-theoretic formulation of linear elasticity. Let us here summarize the highlights. The principal quantities are displacement fields $u^a(\mathbf{x})$, referring to the deviation in direction $a$ from the equilibrium position of the constituent particle at position $\mathbf{x}$ in the coarse-grained continuum limit. The long-wavelength finite-energy configurations are enumerated in terms of the gradients of the displacement field $\partial_m u^a$. Departing from equilibrium the potential energy density of solids takes the familiar form known from elasticity theory~\cite{Kleinert89b}
\begin{align}
e^{(1)}_\mathrm{solid}(\mathbf{x}) &= \tfrac{1}{2} \partial_m u^a C_{mnab} \partial_n u^b,\label{eq:first order elastic energy}\\
e^{(2)}_\mathrm{solid}(\mathbf{x}) &= \tfrac{1}{2} \partial_m \partial_r u^a C^{(2)}_{mrnsab} \partial_n \partial_s u^b.\label{eq:second order elastic energy}
\end{align}
Here $C_{mnab}$ is called the elastic tensor and its independent non-zero components are called elastic constants, while $C^{(2)}_{mrnsab}$ represents the second-order contributions in the gradient expansion.
The elastic tensor is subjected to a number of symmetries and constraints. Importantly, antisymmetric combinations
\begin{equation}\label{eq:rotation field definition}
\omega^{ab} = \frac{1}{2} (\partial_a u^b - \partial_b u^a),
\end{equation}
represent local rotations that must vanish to first order since these cannot change the energy of the crystal. Accordingly, $C_{mnab}$ must be symmetric in $m,a$ and in $n,b$ and Eq.~\eqref{eq:first order elastic energy}
contains only the symmetric combinations called {\em strains}:
\begin{equation}
u^{ab} = \frac{1}{2} ( \partial_a u^b + \partial_b u^a).
\end{equation}
The crystalline symmetry in terms of its {\em space group} further reduces the number of independent elastic constants.
We extend this well-known theory of elasticity to the quantum regime by taking into account the quantum kinetic energy~\cite{ZaanenNussinovMukhin04}. We shall employ the Euclidean coherent-state path-integral formalism in an expansion of fluctuations around the maximally-correlated crystalline state,
defined by the partition function
\begin{align}
Z_\mathrm{solid} &= \int \mathcal{D} u^a\; \mathrm{e}^{- \mathcal{S}_\mathrm{solid} },\label{eq:solid partition function}\\
\mathcal{S}_\mathrm{solid} &= \int \mathrm{d} \tau \mathrm{d}^D x\; \mathcal{L}_\mathrm{solid} ,\label{eq:solid action}\\
\mathcal{L}_\mathrm{solid} &= \mathcal{L}_\mathrm{kin} + \mathcal{L}_\mathrm{pot},\\
\mathcal{L}_\mathrm{kin} &= \frac{1}{2\rho} (\partial_\tau u^a)^2,\label{eq:elasticity kinetic term}\\
\mathcal{L}_\mathrm{pot} &= e^{(1)}_\mathrm{solid}(x) + e^{(2)}_\mathrm{solid}(x).\label{eq:elasticity potential term}
\end{align}
Here the argument $x$ of the displacement fields $u^a(x)$ is understood to contain both space and time $x = (\tau, \mathbf{x})$, and the sign of the potential energy is consistent with our convention for imaginary time~\cite{ZaanenNussinovMukhin04,QLC2D}.
Although the formalism is valid for general elastic tensors, we shall treat explicitly only the case of the {\em isotropic solid}. Even though the crystal breaks rotational symmetry, the long-distance physics may still be effectively isotropic, as is the case for for instance the triangular lattice in 2D. In 3D, solids consisting of many crystalline and glasses (``amorphous solids'') are effectively isotropic~\cite{ChaikinLubensky00,Kleinert89b}. Isotropic solids are described by only two elastic constants: the {\em bulk} or {\em compression modulus} $\kappa$ and the {\em shear modulus} $\mu$. In contrast, in liquids or gases there is only a compression modulus, while for instance crystals with cubic symmetry are characterized
by three independent elastic constants.
The potential energy for the isotropic solid in $D$ space dimensions is defined in terms of the elastic moduli
\begin{equation}\label{eq:isotropic solid elastic constants}
C_{mnab} = D \kappa P^{(0)}_{mnab} + 2\mu P^{(2)}_{mnab}.
\end{equation}
where the projectors of `angular momentum' $s = 0,1,2$ on the space of (1,1)-tensors under $SO(D)$-rotations~\cite{Kleinert89b}:
\begin{align}
P^{(0)}_{mnab} &= \frac{1}{D} \delta_{ma} \delta_{nb},\label{eq:P0}\\
P^{(1)}_{mnab} &= \frac{1}{2} ( \delta_{mn}\delta_{ab} - \delta_{mb} \delta_{na}),\label{eq:P1}\\
P^{(2)}_{mnab} &= \frac{1}{2} ( \delta_{mn}\delta_{ab} + \delta_{mb} \delta_{na}) - \frac{1}{D} \delta_{ma} \delta_{nb}.\label{eq:P2}
\end{align}
These projectors satisfy $P^{(s)}_{mnab} P^{(s')}_{nkbc} = \delta_{ss'} P^{(s)}_{mkac}$ and
\begin{equation}
P^{(0)}_{mnab} + P^{(1)}_{mnab} + P^{(2)}_{mnab} = \delta_{mn} \delta_{ab}.
\end{equation}
The absence of a term proportional to $P^{(1)}$ in Eq.~\eqref{eq:isotropic solid elastic constants} signifies that local rotations Eq.~\eqref{eq:rotation field definition} cannot change the energy of the crystal. The strain component that is singled out by $P^{(0)}$ is called {\em compression strain} while the components in the $P^{(2)}$-subspace are called {\em shear strain}. In $D$ dimensions there are $\tfrac{1}{2} D^2 + \tfrac{1}{2}D - 1$ shear components, in particular
there are 2 shears in $D=2$ and 5 in $D=3$.
The relation between the compression and shear modulus can be expressed using the Poisson ratio $\nu$ via
\begin{align}\label{eq:Poisson ratio definition}
\kappa &= \mu \frac{2}{D} \frac{1+\nu}{1 - (D-1)\nu},&
\nu &= \frac{ D \kappa - 2 \mu}{D(D-1) \kappa + 2\mu}.
\end{align}
The Poisson ratio takes values in $-1 \le \nu \le 1/(D-1)$, and is usually positive. Another quantity used frequently is the Lam\'e constant $\lambda = \kappa - \frac{2}{D} \mu$. Combining the
kinetic and potential terms Eqs.~\eqref{eq:elasticity kinetic term},\eqref{eq:elasticity potential term} we define
\begin{align}\label{eq:relativistic solid Lagrangian}
\mathcal{L}_\mathrm{solid}
&= \frac{1}{2} \partial_\mu u^a C_{\mu\nu ab} \partial_\nu u^b,\nonumber\\
C_{\mu\nu ab}
&= \frac{1}{\mu} \delta_{\mu \mathfrak{t}} \delta_{\nu \mathfrak{t}} \delta_{ab} + C_{mnab}.
\end{align}
Throughout this paper we will use the `relativistic' time $\mathfrak{t} = c_\mathrm{T} \tau = \sqrt{\mu/\rho}\ \tau$ with the unit of length, while $c_\mathrm{T}$ is the shear velocity such that $\partial_\mu = (\frac{1}{c_\mathrm{T}} \partial_\tau, \partial_m)$. Since there cannot be a displacement in the time direction $u^\tau \equiv 0$, the strains $\partial_\mu u^a$ are characterized by a relativistic `spacetime' index $\mu$ and a purely spatial `lattice' index $a$.
The second-order term Eq.~\eqref{eq:second order elastic energy} reduces greatly due to the symmetry of the isotropic solid~\cite{Kleinert89b}:
\begin{align}
e_2 (\mathbf{x}) &= \frac{1}{2} 2\mu \big \lbrack \tfrac {1 - (D-2) \nu}{1 - (D-1) \nu}
\ell'^2 \partial_m \partial_j u^j \partial_m \partial_k u^k \nonumber\\
&\phantom{mmm} + \ell^2 \partial_m
\omega^{ab} \partial_m \omega^{ab} \big \rbrack. \label{eq:isotropic solid second gradient energy}
\end{align}
Here $\ell$ {is the {\em length scale of rotational stiffness}: at length scales smaller than $\ell$, contributions} due to local rotations become important. Similarly, $\ell'$ is the length scale below which second-order compressional contributions become important, but these do not change anything qualitatively and will be ignored in the remainder of this work.
The dynamical properties of the solid can be found by applying infinitesimal external stresses and measuring the responses. In other words, we are interested in the Green's function (propagator) $\langle u^a \;u^b \rangle$.
For the isotropic solid, these have the simple form
\begin{equation}
\langle u^a \; u^b \rangle = \frac{1}{\rho} \left[ \frac{P^\mathrm{L}_{ab}}{\omega_n^2 + c_\mathrm{L}^2 q^2(1 + \ell^{\prime 2} q^2) } + \frac{P^\mathrm{T}_{ab}}{\omega_n^2 + c_\mathrm{T}^2 q^2(1 + \ell^2 q^2)} \right].\label{eq:displacement propagator}
\end{equation}
using the longitudinal and transverse projectors $P^\mathrm{L}_{ab} = q_a q_b /q^2$, $P^\mathrm{T}_{ab} = \delta_{ab} - P^\mathrm{L}_{ab}$. In addition, the longitudinal and transverse velocity are, respectively:
\begin{align}
c_\mathrm{L} &= \sqrt{ \frac{ \kappa + 2\frac{D-1}{D}\mu }{\rho}} = \sqrt{\frac{2\mu}{\rho} \frac{1- (D-2)\nu}{1 - (D-1)\nu}} ,\label{eq:longitudinal velocity definition}\\
c_\mathrm{T} &= \sqrt{\frac{\mu}{\rho}}.
\end{align}
From Eq.~\eqref{eq:displacement propagator} we see that there is one longitudinal acoustic phonon with velocity $c_\mathrm{L}$ and $D-1$ transverse acoustic phonons with velocity $c_\mathrm{T}$. These correspond of course to
the Goldstone modes due to spontaneous breaking of $D$ translational symmetries.
After the dislocation-unbinding phase transition, the displacement fields $u^a$ are no longer well defined, and these propagators lose their meaning. We can however still consider the strain propagators $\langle \partial_m u^a \; \partial_n u^b\rangle$, that have a well defined meaning both in the ordered and disordered phases~\cite{ZaanenNussinovMukhin04,CvetkovicZaanen06a}. We are particularly interested in the longitudinal ($\mathrm{L}$) and transverse ($\mathrm{T}$) propagators.
In the solid these correspond to,
\begin{align}
G_\mathrm{L} &= \langle \partial_a u^a \; \partial_b u^b \rangle = \frac{1}{\mu} \frac{c_\mathrm{T}^2 q^2}{\omega_n^2 + c_\mathrm{L}^2 q^2(1 + \ell^{\prime 2} q^2)}, \label{eq:longitudinal propagator} \\
G_\mathrm{T} &= 2 \langle \omega^{ab} \; \omega^{ab} \rangle = \frac{1}{\mu} \frac{ (D-1) c_\mathrm{T}^2 q^2}{\omega_n^2 + c_\mathrm{T}^2 q^2(1 + \ell^2 q^2)}.\label{eq:transverse propagator}
\end{align}
Here the factor $D-1$ in the transverse propagator arises from summing the contributions of the $D-1$ transverse phonons.
\subsection{Stress--strain duality}\label{subsec:Stress--strain duality}
Following QLC2D, the first step in the dualization procedure is to define the canonical four-momenta conjugate to the displacement field $u^a$ via
\begin{equation}\label{eq:stress tensor definition}
\sigma^a_\mu = - \mathrm{i} \frac{\delta \mathcal{S}}{\delta (\partial_\mu u^a)} = -\mathrm{i} C_{\mu\nu ab} \partial_\nu u^b.
\end{equation}
where we used Eq.~\eqref{eq:relativistic solid Lagrangian} while $-\mathrm{i}$ follows from the standard conventions in the Euclidean formalism~\cite{Kleinert89b,ZaanenNussinovMukhin04,QLC2D}. The quantity $\sigma^a_\mu$ is called the {\em stress tensor}. In static elasticity the stress tensor has only spatial components $\sigma^a_m$ while it is symmetric under $a \leftrightarrow m$ since only symmetric strains $u^{ab}$ are allowed. Similar to the strain fields, in the
imaginary-time extension of the quantum theory the upper (Latin) labels are purely spatial while the lower (Greek) indices are referring to spacetime
since there are no displacements in the time direction, $u^\tau \equiv 0$. The absence of antisymmetric stress components is known as {\em Ehrenfest constraints}~\cite{QLC2D}:
\begin{equation}\label{eq:Ehrenfest constraints}
\epsilon_{cma} \sigma_m^a = 0 \quad \forall \; c.
\end{equation}
Let us now focus on the dual Lagrangian, where the principal variables are the stresses $\mathcal{L}_\mathrm{dual} = \mathcal{L}_\mathrm{dual}[\sigma^a_\mu]$. This can be derived equivalently
by a Legendre transformation or by a Hubbard--Stratonovich transformation of the original Lagrangian. In both cases we need to `invert' the elastic tensor $C_{\mu\nu ab}$. However, due to the absence of antisymmetric strains, the elastic tensor has zeros
amongst its eigenvalues and it cannot be inverted directly. However, the Lagrangian surely contains only physical fields and the dualization operation can be carried out `component-by-component'. It is most useful to bring the original
Lagrangian into a block-diagonal form, to then invert the respective non-zero blocks. For the isotropic solid with elastic tensor Eq.~\eqref{eq:isotropic solid elastic constants}, these correspond to the $P^{(0)}$- and $P^{(2)}$-parts
as well as the kinetic energy. In QLC2D we already derived the dual stress action in arbitrary spatial dimension $D$,
\begin{align}
\mathcal{Z}_\mathrm{solid} &= \int \mathcal{D} \sigma^a_\mu \mathcal{D} u^b \ \mathrm{e}^{-\mathcal{S}_\mathrm{dual}}, \label{eq:dual partition sum} \\
\mathcal{S}_\mathrm{dual} &= \int_0^\beta \mathrm{d} \tau \int \mathrm{d}^D {\mathbf{x}} \ \mathcal{L}_\mathrm{dual} + \mathrm{i} \sigma^a_\mu \partial_\mu u^a,\label{eq:dual action} \\
\mathcal{L}_\mathrm{dual} &= \frac{1}{2\mu} \lvert\sigma^a_\mathfrak{t}\rvert^2 + \frac{1}{2} \sigma^a_\mu \big( \frac{1}{D \kappa} P^{(0)}_{\mu\nu ab} + \frac{1}{2\mu}P^{(2)}_{\mu\nu ab} \big) \sigma^b_\nu \nonumber\\
&= \frac{1}{2\mu} \lvert\sigma^a_\mathfrak{t}\rvert^2 + \frac{1}{8\mu} \big[ \sigma^a_m \sigma^a_m + \sigma^a_m \sigma^m_a - \frac{2\nu}{1 +\nu} \sigma^a_a \sigma^b_b \big]. \label{eq:dual solid Lagrangian}
\end{align}
with the Poisson ratio $\nu$ given by Eq.~\eqref{eq:Poisson ratio definition}.
In appendix ~\ref{sec:Fourier space coordinate systems} we introduce a quite convenient Fourier space coordinate system. In this system, the direction $\mathrm{L}$ is parallel to the momentum $\mathbf{q}$ while $\mathrm{R},\mathrm{S}$ are two transverse directions perpendicular to $\mathrm{L}$ and to each other. All fields in the action are real-valued, and we demand that $\sigma^\dagger(p) = \sigma(-p)$ in momentum space~\cite{Kleinert89a,ZaanenNussinovMukhin04,QLC2D}.
The dual Lagrangian is then block diagonal, containing five sectors:
\begin{align}\label{eq:solid stress Lagrangian components}
\mathcal{L}_\mathrm{dual} &= \mathcal{L}_{\mathrm{T} 1} + \mathcal{L}_{\mathrm{T} 2} +\mathcal{L}_{\mathrm{T} 3} + \mathcal{L}_{\mathrm{L} 1} + \mathcal{L}_{\mathrm{L} 2},\\
\mathcal{L}_{\mathrm{T} 1} &= \frac{1}{8\mu}
\begin{pmatrix} \sigma^{\mathrm{R} \dagger}_\mathfrak{t} \\ \sigma^{\mathrm{R} \dagger}_\mathrm{L} \\ \sigma^{\mathrm{L} \dagger}_\mathrm{R} \end{pmatrix}^\mathrm{T}
\begin{pmatrix}
4 & 0 & 0\\
0 & 1 & 1 \\
0 & 1 & 1
\end{pmatrix}
\begin{pmatrix} \sigma^\mathrm{R}_\mathfrak{t} \\ \sigma^{\mathrm{R} }_\mathrm{L} \\ \sigma^{\mathrm{L}}_\mathrm{R} \end{pmatrix},\nonumber\\
\mathcal{L}_{\mathrm{T} 2} &= \frac{1}{8\mu}
\begin{pmatrix} \sigma^{\mathrm{S} \dagger}_\mathfrak{t} \\ \sigma^{\mathrm{S} \dagger}_\mathrm{L} \\ \sigma^{\mathrm{L}\dagger}_\mathrm{S} \end{pmatrix}^\mathrm{T}
\begin{pmatrix}
4 & 0 & 0\\
0 & 1 & 1 \\
0 & 1 & 1
\end{pmatrix}
\begin{pmatrix} \sigma^\mathrm{S}_\mathfrak{t} \\ \sigma^{\mathrm{S} }_\mathrm{L} \\ \sigma^{\mathrm{L}}_\mathrm{S} \end{pmatrix},\nonumber\\
\mathcal{L}_{\mathrm{T} 3} &= \frac{1}{8\mu}
\begin{pmatrix} \sigma^{\mathrm{R} \dagger}_\mathrm{S} \\ \sigma^{\mathrm{S} \dagger}_\mathrm{R} \end{pmatrix}^\mathrm{T}
\begin{pmatrix}
1 & 1 \\
1 & 1
\end{pmatrix}
\begin{pmatrix} \sigma^{\mathrm{R} }_\mathrm{S} \\ \sigma^{\mathrm{S}}_\mathrm{R} \end{pmatrix},\nonumber\\
\mathcal{L}_{\mathrm{L} 1} &= \frac{1}{8\mu} \frac{2}{1+\nu}
\begin{pmatrix} \sigma^{\mathrm{L} \dagger}_\mathfrak{t} \\ \sigma^{\mathrm{L} \dagger}_\mathrm{L} \\ \sigma_- \end{pmatrix}^\mathrm{T}
\begin{pmatrix}
\scriptstyle 2(1+\nu) & 0 & 0 & 0\\
0 & 1 & \sqrt{2}\nu \\
0 & \sqrt{2}\nu & 1-\nu
\end{pmatrix}
\begin{pmatrix} \sigma^\mathrm{L}_\mathfrak{t} \\ \sigma^{\mathrm{L} }_\mathrm{L} \\ \sigma_- \end{pmatrix},\nonumber\\
\mathcal{L}_{\mathrm{L} 2} &= \frac{1}{4\mu} \lvert \sigma_+ \rvert^2.
\end{align}
We will soon find out that $\mathcal{L}_{\mathrm{L} 1}$ contains the longitudinal phonon, while $\mathcal{L}_{\mathrm{T} 1}$ and $\mathcal{L}_{\mathrm{T} 2}$ contain the transverse phonons in the $\mathrm{R}$- resp. $\mathrm{S}$-transverse directions. Here we have defined
\begin{align}
\sigma_+ &= \frac{1}{\sqrt{2}}(\sigma^\mathrm{R}_\mathrm{R} + \sigma^\mathrm{S}_\mathrm{S}), &
\sigma_- &= \frac{1}{\sqrt{2}}(\sigma^\mathrm{R}_\mathrm{R} - \sigma^\mathrm{S}_\mathrm{S}).
\label{eq:sigma plus minus definition}
\end{align}
In fact, we could have defined similar symmetry and antisymmetric combinations for the transverse sectors, but we refrain from doing so because we need second-order contributions as we shall explain just below. For more context about this division into sectors, see Sec.~\ref{subsec:Interpretation of stress components}.
To express the transverse propagator Eq.~\eqref{eq:transverse propagator} in the dual stress fields, we need the contributions from second-gradient elasticity since
the first three matrices in Eq.~\eqref{eq:solid stress Lagrangian components} are not invertible (see Sec.~\ref{subsec:Dual propagator relations}).
From the second-gradient contribution Eq.~\eqref{eq:isotropic solid second gradient energy}, we will only use the rotational part. Expressed in $\omega^c = \tfrac{1}{2} \epsilon^{cab} \omega^{ab}$ these become,
\begin{equation}\label{eq:second-gradient elasticity}
\mathcal{L}^{(2)} = \frac{1}{2} 4 \mu \ell^2 (\partial_c\omega^c)^2.
\end{equation}
The canonical momentum conjugate to the rotation field $\omega^c = \tfrac{1}{2} \epsilon^{cab} \omega^{ab}$ is the {\em torque stress} $\tau^c_m$:
\begin{equation}\label{eq:torque stress definition}
\tau^c_m = -\mathrm{i} \frac{\delta \mathcal{S} }{\delta (\partial_m \omega^c) } = -\mathrm{i} 4 \mu \ell^2 \partial_m\omega^c.
\end{equation}
There is no separate temporal component, since the rotations $\omega^c$ are descendant from displacements $u^a$ via Eq.~\eqref{eq:rotation field definition} and do not have their own dynamics.
The dual second-gradient Lagrangian is then~\cite{QLC2D}
\begin{equation}\label{eq:dual torque stress Lagrangian}
\mathcal{L}^{(2)}_\mathrm{dual} = \frac{1}{8 \mu \ell^2} (\tau^c_m)^2.
\end{equation}
In the presence of torque stress, the Ehrenfest constraints are softened, and read~\cite{QLC2D}
\begin{equation}\label{eq:Ehrenfest constraint torque stress}
\epsilon_{cma} \sigma_m^a = \partial_m \tau_m^c.
\end{equation}
substituting his equation in Eq.~\eqref{eq:dual torque stress Lagrangian} yields,
\begin{align}\label{eq:second gradient Lagrangian in stress tensor}
\mathcal{L}^{(2)}_\mathrm{dual} &= \frac{1}{8 \mu \ell^2} \frac{1}{q^2} \epsilon_{cma} \sigma_m^{a\dagger} \epsilon_{cnb} \sigma^b_n\\
&=\frac{1}{8\mu \ell^2 q^2} \Big[
\begin{pmatrix} \sigma^{\mathrm{R} \dagger}_\mathrm{L} \\ \sigma^{\mathrm{L} \dagger}_\mathrm{R} \end{pmatrix}^\mathrm{T}
\begin{pmatrix}
1 & -1 \\
-1 & 1
\end{pmatrix}
\begin{pmatrix} \sigma^{\mathrm{R} }_\mathrm{L} \\ \sigma^{\mathrm{L}}_\mathrm{R} \end{pmatrix}\nonumber\\
&\phantom{mmmm} +
\begin{pmatrix} \sigma^{\mathrm{S} \dagger}_\mathrm{L} \\ \sigma^{\mathrm{L}\dagger}_\mathrm{S} \end{pmatrix}^\mathrm{T}
\begin{pmatrix}
1 & -1 \\
-1 & 1
\end{pmatrix}
\begin{pmatrix} \sigma^{\mathrm{S} }_\mathrm{L} \\ \sigma^{\mathrm{L}}_\mathrm{S} \end{pmatrix}\nonumber\\
&\phantom{mmmm} +
\begin{pmatrix} \sigma^{\mathrm{R} \dagger}_\mathrm{S} \\ \sigma^{\mathrm{S} \dagger}_\mathrm{R} \end{pmatrix}^\mathrm{T}
\begin{pmatrix}
1 & -1 \\
-1 & 1
\end{pmatrix}
\begin{pmatrix} \sigma^{\mathrm{R} }_\mathrm{S} \\ \sigma^{\mathrm{S}}_\mathrm{R} \end{pmatrix}
\Big].
\end{align}
To find the propagators of Eqs.~\eqref{eq:longitudinal propagator}, \eqref{eq:transverse propagator} on the dual (stress) side, one needs to introduce stress gauge fields, which we do in Sec.~\ref{subsec:Stress gauge fields} below.
\begin{figure*}
\null\hfill
\subfloat[longitudinal normal $\omega^\mathrm{L}$]{\includegraphics[scale=.3]{rot_Goldstone_long}\label{subfig:longitudinal rotational Goldstone}}
\hfill
\subfloat[transverse normal $\omega^\mathrm{R}$, $\omega^\mathrm{S}$]{\includegraphics[scale=.3]{rot_Goldstone_trans}\label{subfig:transverse rotational Goldstone}}
\hfill\null
\caption{Rotational modes geometry. Depicted are rotational modes between two torque sources indicated as black dots. The momentum is in the direction of the dotted green line connecting these dots, in the $z$- (left) and $x$-direction (right). The blue dashed semicircle line denotes the rotational plane, normal to the index $c$ of $\omega^c$, in this case the $z$-direction, which is longitudinal (left) and transverse (right). The rotational field $\omega^c$ dualizes to the two-form torque stress gauge field $h^c_{\kappa\lambda}$, see Sec.~\ref{subsec:Torque stress gauge fields}. The red cross denotes the transverse polarization of this gauge field $(\kappa\lambda) = (\mathrm{R} \mathrm{S})$, perpendicular to momentum. These transverse components of the dual torque stress gauge field represent the rotational (Goldstone) mode.
}\label{fig:rotational Goldstone modes}
\end{figure*}
\subsection{Rotational elasticity}\label{subsec:Rotational elasticity}
In a medium that is completely translationally symmetric, but which has rotational rigidity, one can also write down the general form of the elastic energy of the long-distance, low-energy excitations. The displacement fields $u^a$ are now ill-defined, but the rotation fields $\omega^c = \tfrac{1}{2} \epsilon^{cab} \omega^{ab}$ are good quantities, and are the fundamental Goldstone fields of this medium. They can be derived using differential geometry as the small fluctuations of the contortion tensor~\cite{Kleinert89b,BohmerDownesVassiliev11,BohmerTamanini15}
We are of course thinking of a nematic liquid crystal where all translational symmetry is restored, but this theory would hold for any translation-invariant medium with spontaneously broken rotational symmetry. For an isotropic medium, like our `isotropic nematic phase', this problem has been studied previously, see for instance Refs.~\onlinecite{BohmerDownesVassiliev11,BohmerTamanini15} and references therein. For systems with broken rotational symmetry, both discrete and continuous, the enumeration of elastic moduli was derived in Ref.~\onlinecite{StallingaVertogen94}. Here we focus on the isotropic case, which as before should be thought of as $O_\mathrm{h}$-symmetry in the limit of vanishing cubic anisotropy.
In close analogy to Sec.~\ref{subsec:Field-theoretic elasticity}, the general form of the rotational-elastic Lagrangian is~\cite{BohmerTamanini15}
\begin{align}\label{eq:general rotational Lagrangian}
\mathcal{L}_\mathrm{rot} &= \tfrac{1}{2} \rho_\mathrm{rot} (\partial_\tau \omega^c)^2 \nonumber\\
&\phantom{mm} + \tfrac{1}{2} \partial_m \omega^a ( \kappa_0 P^{(0)}_{mnab} + \kappa_1 P^{(1)}_{mnab} + \kappa_2 P^{(2)}_{mnab} ) \partial_n \omega^b.
\end{align}
Here $\rho_\mathrm{rot}$ is the `density' of the rotationally rigid medium, rather to be thought of as moment of inertia. The constants $\kappa_{0,1,2}$ define the rotationally elastic properties. Note that there is now no reason why the antisymmetric sector $P^{(1)}$ should be absent. Using relations Eqs.~\ref{eq:P0}--\ref{eq:P2} it can be shown that
\begin{equation}\label{eq:rotational sector dependence}
\int \partial_m \omega^a P^{(2)}_{mnab} \partial_n \omega^b = \int \partial_m \omega^a (2P^{(0)}_{mnab} + P^{(1)}_{mnab}) \partial_n \omega^b.
\end{equation}
Here we performed partial integrations and assumed the surface term vanishes. Then the rotational Lagrangian can be rewritten as
\begin{align}
\mathcal{L}_\mathrm{rot}
&= \tfrac{1}{2} \rho_\mathrm{rot} (\partial_\tau \omega^c)^2 + \tfrac{1}{2} \partial_m \omega^a \Big( (\kappa_0 + 2 \kappa_2) P^{(0)}_{mnab}
\nonumber\\
&\phantom{mmmmmmmmmmmmm} + (\kappa_1 + \kappa_2) P^{(1)}_{mnab} \Big) \partial_n \omega^b \nonumber\\
&= \tfrac{1}{2} \rho_\mathrm{rot} (\partial_\tau \omega^c)^2 + \tfrac{\kappa_0 + 2 \kappa_2}{6} (\partial_c \omega^c)^2 + \tfrac{\kappa_1 + \kappa_2}{4} (\epsilon_{abc} \partial_b \omega^c)^2\nonumber\\
&= \tfrac{1}{2} ( \rho_\mathrm{rot} \omega_n^2 + \tfrac{\kappa_0 + 2 \kappa_2}{3} q^2) (\omega^\mathrm{L})^2 \nonumber\\
&\phantom{mm} + \tfrac{1}{2} ( \rho_\mathrm{rot} \omega_n^2 + \tfrac{\kappa_1 + \kappa_2}{2} q^2) \big( (\omega^\mathrm{R})^2 + (\omega^\mathrm{S})^2\big).
\end{align}
We see that there are again longitudinal and transverse velocities, given by
\begin{align}\label{eq:rotational velocities}
c_\mathrm{L}^\mathrm{rot} &= \sqrt{\frac{\kappa_0 + 2 \kappa_2}{3\rho_\mathrm{rot}}}, &
c_\mathrm{T}^\mathrm{rot} &= \sqrt{\frac{\kappa_1 + \kappa_2}{2\rho_\mathrm{rot}}}.
\end{align}
Note however that the interpretation of these propagating modes is slightly subtle. The vector $\omega^c$ describes rotational deformations in the plane {\em perpendicular} to $c$. So the field $\omega^\mathrm{L}$, where $\mathrm{L}$ is parallel to spatial momentum, describes rotations in the plane perpendicular to the propagating direction. This is counterintuitive when thinking of phonons or photon polarizations, and the reader should take caution when using these terms in the rotational context. For clarity, we have illustrated this in Fig.~\ref{fig:rotational Goldstone modes}.
The longitudinal and transverse velocities are related via the Poisson ratio $\nu$ according to Eq.~\eqref{eq:longitudinal velocity definition}. Therefore one can define a `rotational Poisson ratio' as~\cite{BohmerTamanini15}:
\begin{equation}\label{eq:rotational Poisson ratio}
\nu_\mathrm{rot} = \frac{\kappa_0 - 3 \kappa_1 - \kappa_2}{2\kappa_0 - 3\kappa_1 + \kappa_2}.
\end{equation}
The interpretation of this quantity is as follows: if one perturbs the system by an external rotational torque such as $\tau^x_x$, there can be rotational strain in both the parallel direction ($\partial_x \omega^x$) as well as the perpendicular directions ($\partial_y \omega^y$ and $\partial_z \omega^z$). The negative ratio between the longitudinal and the transverse response is the rotational Poisson ratio. In an ordinary solid, the Poisson ratio is usually positive, meaning that a longitudinal elongation is accompanied by a transverse contraction. Translating this to the rotational context, a positive rotational Poisson ratio means a positive response parallel to the external torque would be accompanied by negative rotational strain in the orthogonal rotational planes. We will come back to this when discussing the rotational Goldstone modes of the quantum nematic in Sec.~\ref{subsec:Torque stress in the quantum nematic}.
\begin{figure*}
\subfloat[edge dislocation]{\includegraphics[height=2.7cm]{edgedislocation.png}\label{subfig:edge dislocation}}
\hfill
\subfloat[screw dislocation]{\includegraphics[height=2.7cm]{screwdislocation.png}\label{subfig:screw dislocation}}
\hfill
\subfloat[wedge disclination]{\includegraphics[height=2.7cm]{disclination.png}\label{subfig:wedge disclination}}
\caption{Elementary topological defects in solids. The defects are always linelike, indicated by the red line. The topological charge, Burgers vector for dislocations and Frank vector for disclinations, is indicated by an arrow. \protect\subref{subfig:edge dislocation} Edge dislocation, the Burgers vector (light green arrow) is perpendicular to the dislocation line (red). \protect\subref{subfig:screw dislocation} Screw dislocation, the Burgers vector is parallel to the dislocation line. \protect\subref{subfig:wedge disclination} Wedge disclination, with Frank vector parallel to the disclination line. }\label{fig:topological defects}
\end{figure*}
\subsection{Dislocations and disclinations}\label{subsec:Dislocations and disclinations}
The very idea of topological melting/weak-strong duality is that the (quantum) phase transition from the solid to the (quantum) liquid corresponds to a proliferation of free topological defects.
In general topological defects depend on the topological structure of order parameter space, which is the coset $G/H$ where $G$ is the symmetry group of the Lagrangian, and $H \subset G$ is the subgroup of unbroken symmetries. Topological defects can come in any dimensionality that is lower than the dimension of space. For instance, in three space dimensions, there are two-dimensional defects (like domain walls), one-dimensional line defects and zero-dimensional point defects. These $\bar{D}$-dimensional defects are classified by the $(D-1 -\bar{D})$th homotopy group of the space $G/H$ denoted by $\pi_{D-1-\bar{D}}(G/H)$.
When dealing with the breaking of the Euclidean group $G = E = \mathbb{R}^D \rtimes O(D)$ to a discrete subgroup (space group) $H = \mathbb{Z}^D \rtimes \bar{P}$, one finds $\pi_2\big((\mathbb{R}^D \rtimes O(D))/(\mathbb{Z}^D \rtimes \bar{P})\big) \simeq \pi_1(\mathbb{Z}^D \rtimes \bar{P}) \simeq \pi_0(1) \simeq 1$~\cite{Mermin79} (since $O(D)$ is not simply connected, the actual homotopy group sequence is slightly different, but in any case as long as $H$ is discrete $\pi_0 \simeq 1$). In other words, topological monopoles do not occur. The $\pi_1$-defects are the dislocations and disclinations~\cite{Mermin79,Kleinert89b,ChaikinLubensky00,SinghDunmur02,QLC2D}. Dislocations are the defects associated with the translational symmetry breaking. These can be pictured as inserting or removing a half-plane of material (this mental cutting and gluing procedure is called {\em Volterra process}). Traversing a contour around the dislocation core will result in a deficient lattice vector, which is called the {\em Burgers vector} $B^a$. It is the topological charge of the dislocation, since it does not depend on the details of the contour. In three dimensions, the dislocation is a line defect. When the Burgers vector is orthogonal to this line, it is called an {\em edge dislocation}, whereas a defect line parallel to the Burgers vector is called a {\em screw dislocation}, see Fig.~\ref{fig:topological defects}. Later, we consider closed dislocation loops of constant Burgers vector, which must be of edge-type somewhere along the loop.
The defects associated with rotational symmetry breaking are called {\em disclinations} and the Volterra process consists of inserting or removing wedges of material, which lead to deficient rotations, see Fig.~\ref{subfig:wedge disclination}. The magnitude is characterized by a deficit angle $\Omega$ and the topological charge is a tensor normal to the plane of rotation with $D-2$ indices called the {\em Frank tensor}. In 3D this becomes the Frank vector $\Omega^c$. If the Frank vector is parallel to the disclination line, it is called a {\em wedge disclination}, and otherwise a {\em bend} or {\em twist disclination}~\cite{Kleinert89b,ChaikinLubensky00,SinghDunmur02}. In 3D, the definitions of the Burgers and Frank vectors are:
\begin{align}
B^a
&= \oint_{\partial \mathcal{S}} \mathrm{d} x^m \; \partial_m u^a,\label{eq:Burgers vector definition}\\
\Omega^c
&= \oint_{\partial \mathcal{S}} \mathrm{d} x^m \; \partial_m \tfrac{1}{2}\epsilon^{cab} \omega^{ab}. \label{eq:Frank vector definition}
\end{align}
Here ${\partial \mathcal{S}}$ is an arbitrary closed contour encircling the core of the topological defect, such that the surface area $\mathcal{S}$ enclosed by $\partial \mathcal{S}$ is pierced by the defect line. An arbitrary defect can
have both translational and rotational character, but this can always be decomposed into multiple elementary defects that are pure dislocations and disclinations. Furthermore, dislocations and disclinations are not independent, see Eq.~\eqref{eq:dislocation disclination interdepence} below. One of the consequences is that a disclination--anti-disclination pair is not topologically trivial but equivalent to a dislocation line~\cite{Kleinert89b,QLC2D}.
In fact, upon pulling apart such a disclination-antidisclination pair in the solid the energy increases with the square of the distance: disclinations are `quadratically confined' in the solid, whereas these dipole pairs attain finite energy in their separation and `deconfine' in the (quantum) nematic. The precise (de)confinement mechanism is a highlight of the weak--strong duality in the 2+1D case~\cite{BeekmanWuCvetkovicZaanen13,QLC2D}, and in Secs.~\ref{subsec:Torque stress gauge fields}, \ref{subsec:Torque stress in the quantum nematic} we will see that it
also applies to 3+1D. Notice that although the gross physical meaning of (de)confinement is similar to that found in non-Abelian Yang--Mills theory, it appears to be due to a different mechanism.
Turning to the 3+1D quantum theory, the spatial dislocation/disclination loops on a time slice turn into worldsheets in spacetime. In our quantum field theoretical setting,
these topological defects take the form of non-critical (Nielsen--Olesen) bosonic strings because we depart from solids formed from bosonic constituents. These
in turn interact with each other via the excitations of the ordered background, the phonons, and these long-range interactions can be represented by effective gauge fields as we will discuss in great detail in Sec.~\ref{subsec:Stress gauge fields}.
On the time slice, the density of these strings can be enumerated in terms of defect density fields, defined by
\begin{align}
J^a_n (x) &= \epsilon_{nkl} \partial_k \partial_l u_a(x) = \delta_n (L,x) B^a,\label{eq:dislocation density definition}\\
\Theta^c_n (x) &= \epsilon_{nkl} \partial_k \partial_l \tfrac{1}{2} \epsilon^{cab} \omega^{ab}(x) = \delta_m (L,x) \Omega^c.
\end{align}
for the dislocations ($J$) and disclinations ($\Theta$). The displacement fields $u^a$ and $\omega^{ab}$ are singular at the core of the defects and become multivalued fields with non-commuting partial derivatives~\cite{Kleinert89b,Kleinert08}. Eqs.~\eqref{eq:Burgers vector definition}, \eqref{eq:Frank vector definition} can be retrieved by integrating these densities over the surface $\mathcal{S}$ and using Stokes' theorem. On the right-hand side, we use the definition of the delta function on the defect line $L$ parametrized by $s$, given by~\cite{Kleinert89b}:
\begin{equation}\label{eq:line delta function}
\delta_n (L,x) = \int_L \mathrm{d} s\ \partial_{s} x^L_n(s) \delta^{(D)}\big(x - x_k^L(s) \big).
\end{equation}
There is a dynamical constraint acting on the motion of edge dislocations. In a crystal they only move in the direction of their Burgers vector, and this is called {\em glide motion}~\cite{Friedel64,CvetkovicNussinovZaanen06}. The reason is that this motion is only a rearrangement of constituent particles, whereas motion orthogonal to the Burgers vector ({\em climb motion}) entails addition or removal of interstitial particles. In real crystals these are energetically very costly
and accordingly their density is small at not too high temperatures. In quantum crystals at zero temperature these occur only as virtual fluctuations involving a finite energy scale with the effect that the glide constraint becomes absolute in the deep IR. In fact, our limit of ``maximal crystalline correlations'' can be viewed as being equivalent to the demand that such constituent particles are infinite-energy excitations. The precise formulation of this {\em glide constraint} will be given in Eq.~\eqref{eq:glide constraint} after we have discussed in more detail the nature of the dislocation worldsheets.
\subsection{Condensation of particle-like topological defects}
\subsubsection{Vortex condensation in 2+1D}\label{subsubsec:Vortex condensation}
Before turning to the dislocation worldsheets in 3+1D, let us shortly review the results of Abelian-Higgs duality in 2+1D. The Berezinskii--Kosterlitz--Thouless phase transition~\cite{Berezinskii70,KosterlitzThouless72,KosterlitzThouless73} takes place in 2+0D at finite temperature where true long-range order is absent due to the Mermin--Wagner--Hohenberg theorem~\cite{MerminWagner66,Hohenberg67}. Thus starting from a state with algebraic long-range order at low temperature (e.g. a two-dimensional superfluid), increasing the temperature introduces vortex--antivortex pairs as thermal excitations. At the critical temperature, such pairs {\em unbind} and vortices and antivortices can appear freely, with the effect that the superfluid order has been destroyed. This is the disordered phase.
In the zero-temperature quantum setting, we must introduce the imaginary time axis and consider the problem in 2+1D spacetime. The quantum phase transition is now tuned by the quantum fluctuations, induced in the superfluid by the charging energy of the constituent bosons.
Departing from the superfluid phase, with its zero temperate broken global $U(1)$-symmetry the disordering quantum fluctuations are completely enumerated in terms of closed loops formed from vortex--antivortex worldlines.
Assuming that the distance between the vortices is large compared to the lattice cut-off (the low vortex-fugacity limit) the physics is completely captured in terms of the vortices as a system of relativistic bosonic particles, interacting via
long-range Coulomb-like forces, with a mass that is vanishing at the quantum phase transition. In the superfluid, defects only appear as small, closed spacetime loops of vortex worldlines, representing the creation and subsequent annihilation of vortex--antivortex pairs. Increasing quantum fluctuations increases the occurrence and the size of such loops, and at the quantum critical point the loops `blow out', becoming as large as the system size and forming a `tangle of free (anti)vortices'. The disordered phase therefore corresponds to a Bose condensate, minimally coupled to the gauge fields mediated by the locally superfluid medium found at distances smaller than the separation between the vortex
particles~\cite{ZaanenBeekman12}. This is known as the {\em vortex--boson} or {\em Abelian-Higgs duality}, in the condensed matter context introduced by Fisher and Lee~\cite{FisherLee89}, where it flourished in studies of the boson-Hubbard model~\cite{FisherEtAl89}. It has been further elaborated on in for instance Refs.~\onlinecite{Kleinert89a,KiometzisKleinertSchakel95,HerbutTessanovic96,CvetkovicZaanen06a} and is supported by very strong numerical evidence~\cite{NguyenSudbo99,HoveSudbo00,HoveMoSudbo00,SmisethSmorgravSudbo04,SmisethEtAl05,SmorgravEtAl05}.
In the ordered (superfluid) phase, the low-energy excitation is simply its Goldstone mode (superfluid second sound), which has the action of a free real scalar field--in the strong-correlation limit amplitude fluctuations are suppressed. This is dualized into a 2+1D vector field $b_\mu$ with vortex sources $J_\mu$ and action~\cite{FisherEtAl89,Kleinert89a,ZaanenNussinovMukhin04,CvetkovicZaanen06a,QLC2D},
\begin{equation}
\mathcal{L}_\mathrm{superfluid} = \frac{g}{2} (\epsilon_{\mu\nu\lambda} \partial_\nu b_\lambda)^2 + \mathrm{i} b_\lambda J_\lambda.
\end{equation}
Here $g$ is the coupling constant. This is of the form Eq.~\eqref{eq:general dual Langrangian Coulomb phase}. In the vortex condensate individual vortex sources $J_\mu$ have lost their identity, being replaced by
the collective condensate field $\Phi(x)$, the amplitude of which obtains a vacuum expectation value in the disordered phase (the vortex `superfluid density'). In 2+1D this `second-quantization' procedure can be derived rigorously
employing a lattice formulation~\cite{Kleinert89a,KiometzisKleinertSchakel95,Cvetkovic06}. Here one obtains the partition function of a grand canonical ensemble of meandering vortex (world)lines as bosonic particles, which are charged under the dual gauge field. In the disordered phase the Lagrangian is given by~\cite{FisherLee89,LeeFisher91,KiometzisKleinertSchakel95,Cvetkovic06}
\begin{equation}\label{eq:vortex-boson Higgs action}
\mathcal{L}_\mathrm{diso.} = \frac{g}{2} (\epsilon_{\mu\nu\lambda} \partial_\nu b_\lambda)^2 + \frac{1}{2} \lvert (\partial_\mu - \mathrm{i} b_\mu) \Phi \rvert^2 + \frac{\alpha}{2} \lvert \Phi\rvert^2 + \frac{\beta}{4} \lvert \Phi \rvert^4.
\end{equation}
Here $\alpha$ and $\beta$ are Ginzburg--Landau parameters. While $\beta >0$ always, when $\alpha$ becomes negative the quantum phase transition takes place. This is precisely the Ginzburg--Landau--Wilson action of a
relativistic superconductor, and the dual gauge fields $b_\mu$ become massive due to the Anderson--Higgs mechanism, implying that the interactions they mediate fall off exponentially with distance: the dual Meissner effect.
We refer to the second term in Eq.~\eqref{eq:vortex-boson Higgs action} as the {\em Higgs term} which is added to the {\em Coulomb term}.
Summarizing, there are three important requirements for condensation of topological defects to take place:
\begin{enumerate}
\item This construction only applies to {\em bosonic} forms of matter: when the constituents are bosons, motion of the topological defects does not lead to any fermion interchanges. If the defects are furthermore Abelian (i.e. the belong to an Abelian homotopy group), they behave just as bosonic particles themselves, forming eventually the dual Bose condensate.
This is self evident for superfluids but it is unknown how to formulate such a dualization for e.g. a crystal formed from fermions.
\item The strong-correlation or low defect-fugacity limit is hard wired in the construction. Only in this limit is the physics of the disordered state captured entirely in terms of the dual vortex fields.
\item Equivalently, non-topological defects such as interstitials in crystals are completely ignored which is another way of imposing the strong correlation limit. The effect will be that a variety of
propagating massive excitations are found that will get `(over)damped by the interstitials' upon moving away from the limit of maximal correlation.
\end{enumerate}
\subsubsection{Dislocation condensation in 2+1D}\label{subsubsec:Dislocation condensation}
Let us now turn to elasticity. The idea that melting of a solid takes place due to proliferation of dislocations is quite old. Starting with the Kosterlitz--Thouless papers~\cite{KosterlitzThouless72,KosterlitzThouless73}, the classical theory of dislocation-mediated melting in 2D was established in the late 1970s by Nelson, Halperin and Young~\cite{HalperinNelson78,NelsonHalperin79,Young79} so that we now speak of the KTNHY-transition of a 2D crystal to a 2D liquid crystal. This famously includes the prediction of the hexatic phase, which we refer to as $C_6$-nematic~\cite{LiuEtAl15,LiuEtAl16b,NissinenEtAl16,LiuEtAl16,QLC2D}. It was also realized that melting dislocations but not disclinations leads to the liquid crystal, while proliferation of dislocations and disclinations at the same time is in fact the ordinary first-order solid--liquid transition~\cite{Kleinert83}. Furthermore, dislocations can proliferate with Burgers vectors preferentially in one direction which `turns liquid' while the other direction `remains solid'~\cite{OstlundHalperin81}. This is the smectic liquid crystal with partial translational symmetry restoration.
It was the insight of Ref.~\onlinecite{ZaanenNussinovMukhin04} that dislocation-mediated quantum melting in 2+1D at zero temperature can be achieved by a generalization of Abelian-Higgs duality. Here the topological charge of the defects is not an integer in $\mathbb{Z}$ but a Burgers vector in $\mathbb{Z}^D$. Departing from a bosonic crystal, the dislocations braid as bosons as well. Accordingly, the dislocations form a conventional Bose condensate and in essence
the melting of the solid in a quantum liquid crystal is governed as well by the Abelian-Higgs duality we just discussed.
So far, a rigorous derivation of the Higgs term for dislocation-mediated quantum melting has not been achieved and no numerical simulations have been performed. However, we can rest on general principles to obtain these terms~\cite{ZaanenNussinovMukhin04,Cvetkovic06,QLC2D}. The condensation of bosons can only be the standard Ginzburg--Landau complex scalar field affair with the $\lvert\Phi\rvert^4$-potential as in Eq.~\eqref{eq:vortex-boson Higgs action}. For dislocations with Burgers vector $B^a$, $a = x,y$, there are two copies of $U(1)$-fields $\Phi^a$. There is {\em a priori} no $SU(2)$ symmetry between these fields, so the symmetry group is simply $U(1) \times U(1)$. The reason is that Burgers vectors can only point in the lattice directions, so we cannot make an arbitrary rotation in Burgers space. Once a condensate with Burgers vector in, say, the $x$-direction is established, all lattice points in the $x$-direction become equivalent. This also implies that any remaining dislocations must point orthogonal to the already-condensed direction, regardless of the original space group of the crystal. We already mentioned this in Sec.~\ref{sec:Symmetry principles of quantum liquid crystals} and in Fig.~\ref{fig:hexagonal melting}.
The reader should
be aware that the construction of the condensate was for the first time fully understood only in QLC2D --- the earlier papers were not quite correct on this issue.
The bottom line is that the Higgs term of the dislocation condensate is identical to that of a two-component Bose--Einstein condensate~\cite{KasamatsuEtAl05}, having the general form~\cite{QLC2D}
\begin{align}
\mathcal{L}_\mathrm{Higgs} &= \frac{1}{2} \sum_{a = x,y} \lvert (\partial_\mu - \mathrm{i} b^a_\mu) \Phi^a\rvert^2 +
\tfrac{\alpha_x}{2} | \Phi^x |^2 + \tfrac{\alpha_y}{2} | \Phi^y |^2
\nonumber\\ &\phantom{mmm}
+ \tfrac{\beta_x}{4}| \Phi^x |^4 + \tfrac{\beta_y}{4} | \Phi^y |^4
+ \tfrac{\gamma}{2} | \Phi^x |^2 | \Phi^y |^2. \label{eq:two-condensate potential}
\end{align}
The first term is the minimal coupling of the dual stress gauge fields to the dislocation condensates, replacing the coupling term $\mathrm{i} b^a_\mu J^a_\mu$ and is a direct generalization of Eq.~\eqref{eq:vortex-boson Higgs action}. The other terms are the Higgs potential. The parameters $\alpha_{x}$, $\alpha_y$, $\beta_x$, $\beta_y$, $\gamma$ depend on the parent crystal structure and other material details. The last term is the coupling between the two condensates: for large positive $\gamma$ only one of $\Phi^x$, $\Phi^y$ will obtain a vacuum expectation value and only dislocations with Burgers vector in that direction condense; this is the smectic phase characterized by a restored translational symmetry in only one direction. On the other hand, for small $\gamma$ the condensates are `locked together' and condense at the same time: this is the nematic phase restoring translational symmetry in both spatial directions~\cite{QLC2D}. The remnant rotational order in the nematic phase depends on the details of the parameters in Eq.~\eqref{eq:two-condensate potential}, and ultimately on the parent crystal symmetry~\cite{LiuEtAl15}.
The Higgs action Eq.~\eqref{eq:two-condensate potential} together with the original stress action of the solid phase is sufficient to completely capture the low-energy spectrum of the quantum smectic and nematic phases, which was worked out in detail for the 2+1D isotropic solid in QLC2D. We can already anticipate the contents of the next subsection: in 3D there will be three dislocation condensate fields $\Phi^a$ that could condense one-by-one to restore translational symmetry in one, two or three directions respectively. The order and nature of this condensation depend on the Ginzburg--Landau phenomenological parameters, ultimately descendant from the crystal structure and other details of the microscopic physics.
\subsection{Condensation of line-like topological defects}
We gave now arrived at the key section of this work: the construction of the Higgs terms for dislocation fields in 3+1 dimensions. Unfortunately, we will not derive this term in a systematic way. Ideally, one would procure the partition function of a grand canonical ensemble of dislocation worldsheets. This is however complicated by two factors. First, even in 2+1D the Higgs term for dislocation worldlines with vectorial topological charges has not been derived rigorously. But more importantly, the worldsheet nature of linelike defects is a serious obstacle. The theory describing the quantum mechanics of a single line-like object is string theory. However, the quantum field theory of an ensemble of such strings is {\em string field theory}~\cite{Witten86}, which is far from completely settled. But again, resting on general principles it is possible to argue what the form of the Higgs term associated with the spacetime `string foam' formed out of the proliferated dislocations must be.
\subsubsection{Vortex worldsheet condensation}\label{subsubsec:Vortex worldsheet condensation}
The condensation of extended objects has naturally received substantial attention in the literature. Especially on the lattice, it has been long ago established as a physical phenomenon~\cite{Kogut79,Savit80}. In field theory condensation of two-form fields was considered in Ref.~\onlinecite{Rey89} and in the context of vortex--boson duality in Refs.~\onlinecite{MotrunichSenthil05,Franz07}, and very recently in Ref.~\onlinecite{YeGu15}. The main issue here is how to couple the two-form field $b_{\mu\nu}$ to the phase degree of freedom of the condensate field $\Phi$. These works solve the problem by assigning a `vectorial phase' to the condensate field. The condensate field is assumed to be of the form
\begin{equation}\label{eq:vortex condensate vectorial phase}
\Phi(x) = \lvert \Phi(x) \rvert \mathrm{e}^{\mathrm{i} \int \mathrm{d} X^\mu(\sigma) \phi_\mu[X(\sigma)]}.
\end{equation}
Here $\sigma = \sigma_{1,2}$ are coordinates on the worldsheet, $X(\sigma)$ is a map from the worldsheet to real space (called {\em target space} in this context), and the integral is over the entire worldsheet. We also see immediately that shifting $\phi_{\mu}\to \phi_{\mu}+\partial_{\mu} \zeta$ with a total derivative is a redundancy (a constant phase shift) This is to be contrasted with the usual complex scalar field $\Phi = \lvert \Phi \rvert \mathrm{e}^{\mathrm{i} \phi}$. One obtains the seemingly straightforward generalization of the minimal coupling term in the London limit where amplitude fluctuations are suppressed:
\begin{equation}\label{eq:string minimal coupling}
\mathcal{L}_\mathrm{min.coup.} = \frac{1}{4} \lvert \Phi \rvert^2 (\partial_\mu \phi_\nu - \partial_\nu \phi_\mu - b_{\mu\nu} )^2.
\end{equation}
The extra factor of $\frac{1}{2}$ is to compensate over the sum over antisymmetric indices.
In Ref.~\onlinecite{BeekmanSadriZaanen11} we argued that this form cannot be correct at least for the case of vortex--boson duality in the superfluid--Mott insulator transition due to a too large number of propagating modes. The counting as follows. It is known that a massive two-form field $b_{\mu\nu}$ in 3+1 dimensions has three propagating degrees of freedom. This can be seen as follows: from Sec.~\ref{subsec:Two-form gauge fields} we know that a free real scalar field is dual to a free two-form gauge field in 3+1D, and hence the latter carries a single propagating degree of freedom. The dynamics of vortex condensate of the form Eq.~\eqref{eq:vortex condensate vectorial phase} is carried by $\phi_\mu$, and a free vector gauge field in 3+1D carries two propagating degrees of freedom. When the vortex condensate is coupled to the two-form gauge field, its two degrees of freedom are transferred to (``eaten by'') the gauge field through the Anderson--Higgs mechanism, making a total of three propagating modes.
However the (boson) Mott insulator has two propagating modes (the `doublon' and `holon' excitations) in any spatial dimension, and for this reason it cannot be described by such a `stringy' condensate. We put forward the possibility that the condensation of vortices with a finite size and a core energy could be different from that of coreless, critical string. We argued therefore that the vortex condensate cannot carry more than one degree of freedom, and must be described by an ordinary scalar phase field $\phi$. We proposed two alternative ways to coupling this phase field to the two-form gauge field in Refs.~\onlinecite{BeekmanSadriZaanen11,BeekmanZaanen12}.
In the current context of dislocation condensation, it turns out that as long as one is interested in correlations of linear stress $\sigma^a_\mu$ only, for instance the longitudinal and transverse propagators in Eqs.~\eqref{eq:longitudinal propagator stress propagator relation}, \eqref{eq:transverse propagator stress propagator relation} any Higgs term that gives a mass to the dual gauge fields $b^a_{\kappa\lambda}$ leads to correct results. The reason is that these quantities are by definition independent of the longitudinal components of $b^a_{\kappa\lambda}$ and as such cannot be influenced by the dislocation condensate phase degrees of freedom at all. However, we are also particularly interested in the fate of torque stress $\tau^c_\mu$, which we announced to become deconfined in the liquid crystals. To properly take this into account, it seems that the dislocation worldsheet condensate must follow the `stringy' form Eq.~\eqref{eq:string minimal coupling}. Therefore, below we shall consider the flavored generalization of Eqs.~\eqref{eq:vortex condensate vectorial phase}, \eqref{eq:string minimal coupling}.
\subsubsection{Dislocation worldsheet condensation}\label{subsubsec:Dislocation worldsheet condensation}
As we just argued, we shall not attempt to derive the full Higgs action of a condensate of dislocation worldsheets, but instead assume that this is governed by the minimal coupling form Eq. (\ref{eq:string minimal coupling}). In all other
regards, there is no difference of principle compared to the construction in the 2+1D case~\cite{QLC2D}. Let us nevertheless spell out once again the assumptions.
\paragraph{Assumptions.}
As in 2+1D, the quantum liquid crystals in 3+1D can be characterized topologically as quantum liquids with the property that disclinations are massive topological excitations. There is surely the practical question whether forms of matter exist
where disclinations are sufficiently costly compared to dislocations so that the simultaneous proliferation of dislocations and disclinations can be avoided; the latter case results in the standard first-order quantum phase transition from the crystal to the isotropic superfluid as realized for instance in $^4$He. This is however a matter of microscopic details which is beyond the field-theoretical scope: here we just assert that in principle such quantum liquid crystals can be formed in 3+1D, and study their low-energy properties.
A next concern is the precise symmetry of the resulting quantum liquid-crystalline states. The quantum nematics are in this regard most straightforward. These break the $O(3)$ to some (usually discrete) point group. As we discussed in Sec.~\ref{sec:Symmetry principles of quantum liquid crystals}, the order parameter theory already in the classical case becomes a quite complicated affair when dealing with anything else than the simple uniaxial nematics found in LCD screens and so forth~\cite{LiuEtAl15,LiuEtAl16b,NissinenEtAl16,LiuEtAl16}.
Here we will just avoid these difficulties by asserting that our quantum nematics are `isotropic', which means ``indistinguishable from the full spontaneous symmetry breaking of $O(3)$ itself'' when it comes to observable properties like the velocities of the rotational Goldstone bosons, or more generally the elasticity theory of the nematic~\cite{StallingaVertogen94}. This is the same `emergent' isotropy as which one encounters in the isotropic solid, see Sec.~\ref{subsec:Field-theoretic elasticity}. In principle this can be improved on by departing from the elasticity theory associated with the real 3D space groups, but this is just much more menial labor while it will
not affect the basic construction. We leave this as an open problem for follow-up work.
With these assumptions, we can afford to be rather ignorant regarding the details of the point group symmetries of the crystal and the quantum liquid crystals, given that the main rules for the construction of the dislocation condensate are
insensitive to these details. As we learned in 2+1D~\cite{QLC2D}, the first two rules are as follows: (i) For every space dimension a separate dislocation condensate is needed, to restore the translation symmetry in that particular direction. In three space dimensions we are dealing with three separate condensates. Depending on the sign and the strength of the interactions between these condensates one obtains a columnar, a smectic
or a nematic phase corresponding to dislocation condensates switching on in one, two or three spatial directions, see Fig.~\ref{fig:hexagonal melting}.
(ii) A finite `Burgers vector magnetization' is coincident with a finite disclination density and this should be forbidden in the quantum liquid crystal phases. Accordingly, the dislocation condensate in every spatial direction is characterized by the requirement that the Burgers vectors are constrained to be locally anti-parallel.
In fact, by far the most important assumption with regard to the limitations on the physics we can address is the low-fugacity assumption. We have emphasized it already a number of times: it is assumed that all the degrees of
freedom are collective, either corresponding to the phonons or the dislocations. In order to reach this limit, the distance between the dislocations in the condensate has to be very large compared to the lattice constant. The
quantum liquid is described as a system that is locally like a solid carrying propagating phonons. All the `disordering action' is due to the dislocations: the system turns into a liquid at distances larger than the average distance
between the dislocations. This has the physical consequence that all the collective degrees of freedom are {\em propagating}, including the finite energy `massive shear photons'. This should be considered as just an extreme limit that will never precisely be reached in any physical system. The degrees of freedom that are ignored are the interstitials, in essence loose constituent particles that would be the fundamental degrees of freedom were one departing
from the weakly-interacting `kinetic gas' limit which is the traditional starting point constructing theories of liquid crystals. These interstitials can be viewed in turn as bound dislocation--antidislocation pairs that could be included perturbatively departing from our weak--strong duality limit. This has not been studied systematically yet, but their qualitative effects are obvious: these would damp our massive propagating modes, to the extent that these modes can be entirely overdamped reaching the gaseous limit. In fact, these interstitials are much more of an issue in 3+1D as compared to the 2+1D case. In 2+1D both dislocations and interstitials are particle-like excitations and at least
in the Euclidean continuum the core energies of interstitials are just much higher than those of dislocations for ubiquitous microscopic reasons. However, in 3+1D interstitials continue to be particle-like, while the dislocations turn into strings. Accordingly, interstitials fluctuate much more easily and their gain in quantum kinetic energy may well overwhelm their high core-energies. The state of matter formed by a proliferation of interstitials while dislocations stay massive is the {\em supersolid} and it is well understood that for these reasons supersolids are in principle ubiquitous in 3+1D. It depends yet again on the microscopic numbers but it appears for the same reasons to be much more problematic to reach the limit of maximal crystalline correlations in 3+1D as compared to the situation in lower dimensions.
As a corollary to this ``no interstitial'' rule, the (anti-)dislocations can proliferate preserving the total number of particles: ``dislocations do not carry volume'', with the ramification that (iii) the glide constraint becomes absolute
(Sec.~\ref{subsec:Dislocations and disclinations}), and the ``no climb'' condition has to be imposed on the condensate. Just as in 2+1D~\cite{QLC2D}, this will have the consequence that the dislocations do not couple to compressional stress. Accordingly, sound stays massless in the (nematic) quantum liquid and it can be proven that this system is also a conventional superfluid.
In summary, the construction of the dual stress superconductor in 3+1D follows in every regard the template we studied throughly in 2+1D, except for the fundamental description of the string condensate. We will assume that the condensate in each Burgers direction is governed by the minimal coupling form Eq.~\eqref{eq:string minimal coupling}.
\paragraph{The Higgs potential.}
Given these assumptions, we have to construct a Higgs potential of $D=3$ complex scalar fields $\Phi_i$ ($i=x,y,z)$ associated with the separate dislocation condensates with Burgers directions in three spatial directions. As argued above and in Fig.~\ref{fig:hexagonal melting}, these directions must be orthogonal. Up to fourth order in the fields, the most general potential involving only density--density couplings is
\begin{align}
\mathcal{L}_\mathrm{\lvert \Phi \rvert} &= \tfrac{1}{2} \alpha_x | \Phi^x |^2 + \tfrac{1}{2} \alpha_y | \Phi^y |^2 + \tfrac{1}{2} \alpha_z | \Phi^z |^2
\nonumber\\ &\phantom{mm}
+ \tfrac{1}{4} \beta_x | \Phi^x |^4 + \tfrac{1}{4} \beta_y | \Phi^y |^4 + \tfrac{1}{4} \beta_z | \Phi^z |^4 \nonumber\\&\phantom{mm}
+ \tfrac{1}{2}\gamma_{xy} | \Phi^x |^2 | \Phi^y |^2
+ \tfrac{1}{4}\gamma_{xz} | \Phi^x |^2 | \Phi^z |^2 \nonumber\\
&\phantom{mm}
+ \tfrac{1}{4}\gamma_{yz} | \Phi^y |^2 | \Phi^z |^2.
\label{eq:three-condensate potential}
\end{align}
This potential is identical to the one describing a three-component Bose--Einstein condensate, which was studied in Ref.~\onlinecite{RobertsUeda06}. There it was established that the condensate consists of either one,
two or three species of bosons, pending the values of the couplings $\gamma_{xy}$, $\gamma_{xz}$, $\gamma_{yz}$. In the present context, it is possible to have dislocation condensates which restore translational symmetry
in one, two or three directions.
In analogy to the classical liquid crystals we will call these the {\em columnar}, {\em smectic} and {\em nematic} phases respectively, regardless the nature of the remnant rotational symmetry breaking, see Fig.~\ref{fig:phases} and Fig.~\ref{fig:hexagonal melting}.
\paragraph{Minimal coupling.}
Having established the form of the Higgs potential, we now need to couple the condensate fields $\Phi^a$ to the dual stress gauge fields $b^a_{\mu\nu}$. Following Sec.~\ref{subsubsec:Vortex worldsheet condensation}, the coupling is
\begin{align}
\mathcal{L}_\mathrm{min.coup.} &= \frac{1}{4} \sum_a \lvert \Phi^a \rvert^2 ( \partial_\mu \phi^a_\nu - \partial_\nu \phi^a_\mu - b^a_{\mu\nu} )^2,\label{eq:dislocation condensate minimal coupling step 1}
\end{align}
where $\phi^a_\mu$ are the vectorial phase fields of condensate $\Phi^a$ as in Eq.~\eqref{eq:string minimal coupling}.
The next issue to address is the velocity of the dislocation condensate. In the vortex--boson duality governing superfluids, the circulation around vortices is obviously ordinary superflow, so the vortex itself also moves with the superfluid phase velocity. From a different viewpoint, at the quantum critical point with emergent relativity there can be only one velocity scale. Nevertheless, it is useful to keep the vortex velocity scale formally distinct from the superfluid phase velocity in the calculations, to determine the nature of the massive modes in the Higgs phase~\cite{CvetkovicZaanen06a,QLC2D}.
For dislocations, the situation is slightly different. First of all there is the glide constraint restriction certain dislocation motion. Furthermore, it is known that screw and edge dislocations may move at different velocities, see Sec.~\ref{subsec:Dislocation worldsheets}.
Below, we will first treat an idealized case where all dislocations move with a single velocity $c_\mathrm{d}$, as it contains the important physics such as the emergent rotational Goldstone modes while the equations remain tractable. Afterwards, we will discuss the more realistic case with two velocity scales $c_\mathrm{s}$ and $c_\mathrm{e}$. In the first case, the minimal coupling Eq.~\eqref{eq:dislocation condensate minimal coupling step 1} becomes,
\begin{align}
\mathcal{L}_\mathrm{min.coup.} &= \frac{1}{4} \sum_a \lvert \Phi^a \rvert \Big( \frac{1}{c_\mathrm{d}^2} 2(\partial_\tau \phi^a_n - \partial_n \phi^a_\tau - b^a_{\tau n} )^2 \nonumber\\
&\phantom{mmmmm} + (\partial_m \phi^a_n - \partial_n \phi^a_m - b^a_{m n} )^2\Big).
\label{eq:dislocation condensate minimal coupling step 2}
\end{align}
Here summation over the Latin indices $m,n$ is implied and the factor of 2 in the first term accounts for the antisymmetry in $\tau, n$. It is easiest to perform the calculations in the dislocation-Lorenz gauge fix
\begin{equation}
\frac{1}{c_\mathrm{d}^2} \partial_\tau b^a_{\tau \nu} + \partial_m b^a_{m \nu} = 0,\label{eq:dislocation Lorenz gauge fix}
\end{equation}
since we can be sure that the condensate phase degrees of freedom are decoupled. It is then convenient to rewrite the whole expression with temporal components rescaled with $c_\mathrm{d}$ instead of $c_\mathrm{T}$.
For this purpose we define fields and derivatives indicated with a tilde, as follows~\cite{QLC2D}:
\begin{align}
\tilde{b}^a_{\kappa\lambda} &= ( \frac{1}{c_\mathrm{d}} b^a_{\tau l} , b^a_{kl} ) ,& \tilde{\phi}^a_\lambda &= (\frac{1}{c_\mathrm{d}} \phi_\tau,\phi_l), \nonumber\\
\tilde{\partial}_\mu &= ( \frac{1}{c_\mathrm{d}} \partial_\tau , \partial_m ), &
\tilde{p}_\mu &= (\frac{1}{c_\mathrm{d}} \omega_n , q_m).\label{eq:tilde fields definition}
\end{align}
The relations between the original and the tilde fields are
\begin{align}
\tilde{b}^a_{\mathfrak{t} n} &= \frac{c_\mathrm{T}}{c_\mathrm{d}} b^a_{\mathfrak{t} n} = \frac{1}{c_\mathrm{d}} b^a_{\tau n}, &
\tilde{p}_\mathfrak{t} = \frac{c_\mathrm{T}}{c_\mathrm{d}} p_\mathfrak{t} = \frac{1}{c_\mathrm{d}} \omega_n.
\end{align}
Clearly this affects the $0,1,\mathfrak{t}$-components, but the $\mathrm{L},\mathrm{R},\mathrm{S}$-components are unchanged: $A_{\mathrm{L},\mathrm{R},\mathrm{S}} = \tilde{A}_{\mathrm{L},\mathrm{R},\mathrm{S}}$ for any field $A_\mu$. With these redefinitions, the minimal coupling term can be written as
\begin{align}
\mathcal{L}_\mathrm{min.coup.} &= \frac{1}{4}\sum_a \lvert \Phi^a \rvert ( \tilde{\partial}_\mu \tilde{\phi}^a_\nu - \tilde{\partial}_\nu \tilde{\phi}^a_\mu - \tilde{b}^a_{\mu\nu} )^2,
\label{eq:dislocation condensate minimal coupling step 3}
\end{align}
\paragraph{Glide constraint.}
Finally, we need to implement the glide constraint Eq.~\eqref{eq:glide constraint} which dictates that edge dislocations can only move in the direction parallel to their Burgers vector, and which is rooted in particle number conservation. In three dimensions, it is possible that the dislocation line moves perpendicular to its Burgers vector at some part along the line, while it moves in the opposite direction somewhere else along its line, effectively `borrowing' an interstitial particle from itself~\cite{CvetkovicNussinovZaanen06}. This is sometimes called {\em restricted climb motion}. The glide constraint Eq.~\eqref{eq:glide constraint} then holds for the line as a whole, so that we have the integral identity $\int \mathrm{d}^D x\; \epsilon_{\mathfrak{t} mna} J^a_{mn}(x) = 0$. The glide constraint can be enforced in the path integral using a Lagrange multiplier field $\lambda(x)$:
\begin{equation}
Z_\mathrm{glide} = \int \mathcal{D} \lambda \mathrm{e}^{\int \lambda \epsilon_{\mathfrak{t} mna} J^a_{mn}}.
\end{equation}
Since the dual stress gauge field is minimally coupled to dislocation sources, this amounts to the shift~\cite{QLC2D},
\begin{equation}\label{eq:glide constraint substitution}
\tilde{b}^a_{\mu\nu} \to \tilde{b}^a_{\mu\nu} + \lambda \epsilon_{\mathfrak{t} \mu\nu a}.
\end{equation}
in the minimal coupling term.
Notice that since the Levi-Civita symbol already contains an entry $\mathfrak{t}$, we can replace $m \to \mu$, $n \to \nu$ with impunity. Also, since the glide constraint only concerns components with spatial indices, we can replace tilde fields as $\tilde{b}^a_{mn} = b^a_{mn}$. With regard to the dislocation condensate, the factors $b^a_{\mu\nu}$ can be simply substituted by Eq.~\eqref{eq:glide constraint substitution}. This is exactly the same procedure as followed in the 2+1D case~\cite{QLC2D,Cvetkovic06}. Later we will see that this constraint has the effect that the compression mode always stays massless although the shear modes acquire a Higgs mass.
This of course makes sense: the quantum liquid crystal is at the same time a superfluid, characterized by a massless sound mode.
\paragraph{Higgs term for a single velocity.}
We have now collected all the pieces and we can write down the Higgs term of the dislocation condensate coupled to two-form dual stress gauge fields:
\begin{widetext}
\begin{align}
\mathcal{L}_\mathrm{Higgs} &= \frac{1}{4}\sum_{a} \lvert \Phi^a \rvert^2
(\tilde{\partial}_\mu \tilde{\phi}^a_\nu - \tilde{\partial}_\nu \tilde{\phi}^a_\mu - \tilde{b}^a_{\mu\nu} - \lambda \epsilon_{\mathfrak{t} \mu \nu a} ) ^2 \\
&=
\frac{1}{4} \sum_a \frac{(\Omega^{a})^2}{c_\mathrm{T}^2 \mu} \Big[
\lvert \tilde{\partial}_\mu\tilde{\phi}^a_\nu - \tilde{\partial}_\nu\tilde{\phi}^a_\mu \rvert^2 +
(\tilde{b}^{a\dagger}_{\mu\nu} + \lambda^\dagger \epsilon_{\tau \mu \nu a}) (\delta_{\mu\kappa} - \frac{\tilde{p}_\mu \tilde{p}_\kappa}{\tilde{p}^2} )(\delta_{\nu\lambda} - \frac{\tilde{p}_\nu \tilde{p}_\lambda}{\tilde{p}^2} ) (\tilde{b}^a_{\kappa\lambda} + \lambda \epsilon_{\tau \kappa\lambda a})\Big].
\label{eq:dislocation Higgs term}
\end{align}
\end{widetext}
Here we have defined the Higgs mass $\Omega^a = c_\mathrm{T} \sqrt{\mu} \lvert \Phi^a \rvert$ with units of energy, and imposed the Lorenz gauge fix in the second line, as indicated by the projectors. The dislocation condensate phase modes $\tilde{\phi}^a_\mu$ are clearly decoupled in this gauge fix, and only the gauge-invariant parts of the stress gauge fields appear. The sum over $a$ is the sum over the Burgers directions of the dislocation condensate, which could be in one, two or three spatial directions depending on each of the $\Omega^a$. As argued in QLC2D and in Secs.~\ref{sec:Symmetry principles of quantum liquid crystals}, \ref{subsubsec:Dislocation condensation} these directions are strictly orthogonal to each other, see Fig.~\ref{fig:hexagonal melting}, while at least one of them lies along a crystal axis.
\paragraph{Lagrangian and propagators of the solid in the dislocation gauge fix.}
Eq.~\eqref{eq:dislocation Higgs term} is the Higgs term we shall use in all the calculations below. For future use, let us rewrite the Lagrangian of the elastic medium Eq.~\eqref{eq:solid stress gauge field Lagrangian} in terms of the tilde fields:
\begin{widetext}
\begin{align}
\mathcal{L}_\mathrm{dual} &= \mathcal{L}_{\mathrm{T} 1} + \mathcal{L}_{\mathrm{T} 2} + \mathcal{L}_{\mathrm{T} 3} + \mathcal{L}_\mathrm{L} + \mathrm{i} \tilde{b}^{a\dagger}_{\mu\nu} \tilde{J}^a_{\mu\nu},\label{eq:solid stress gauge field Lagrangian tilde}\\
\mathcal{L}_{\mathrm{T} 1} &= \frac{1}{8\mu}
\begin{pmatrix}\tilde{b}^{\mathrm{L}\dagger}_{1 \mathrm{S}} & \tilde{b}^{\mathrm{R}\dagger}_{\mathrm{R}\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
\frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2(1 + \frac{1}{\ell^2 q^2}) & \mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n \frac{c_\mathrm{d}}{c_\mathrm{T}}\tilde{p}(1 - \frac{1}{\ell^2 q^2}) \\
-\mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n \frac{c_\mathrm{d}}{c_\mathrm{T}}\tilde{p}(1 - \frac{1}{\ell^2 q^2}) & \frac{1}{c_\mathrm{T}^2}\omega_n^2(1 + \frac{1}{\ell^2 q^2}) + 4 q^2
\end{pmatrix}
\begin{pmatrix}\tilde{b}^{\mathrm{L}}_{1 \mathrm{S}} \\ \tilde{b}^{\mathrm{R}}_{\mathrm{R}\mathrm{S}} \end{pmatrix}, \\
\mathcal{L}_{\mathrm{T} 2} &= \frac{1}{8\mu}
\begin{pmatrix}\tilde{b}^{\mathrm{L}\dagger}_{1 \mathrm{R}} & \tilde{b}^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
\frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2(1 + \frac{1}{\ell^2 q^2}) & \mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n \frac{c_\mathrm{d}}{c_\mathrm{T}}\tilde{p}(1 - \frac{1}{\ell^2 q^2}) \\
-\mathrm{i} \frac{1}{c_\mathrm{T}}\omega_n \frac{c_\mathrm{d}}{c_\mathrm{T}}\tilde{p}(1 - \frac{1}{\ell^2 q^2}) & \frac{1}{c_\mathrm{T}^2}\omega_n^2(1 + \frac{1}{\ell^2 q^2}) + 4 q^2
\end{pmatrix}
\begin{pmatrix}\tilde{b}^{\mathrm{L}}_{1 \mathrm{R}} \\ \tilde{b}^{\mathrm{S}}_{\mathrm{R}\mathrm{S}} \end{pmatrix}, \\
\mathcal{L}_{\mathrm{T} 3} &= \frac{1}{8\mu}
\begin{pmatrix}\tilde{b}^{\mathrm{R}\dagger}_{1 \mathrm{R}} & \tilde{b}^{\mathrm{S}\dagger}_{1\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
\frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2(1 + \frac{1}{\ell^2 q^2}) & \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2(1 - \frac{1}{\ell^2 q^2}) \\
\frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2(1 - \frac{1}{\ell^2 q^2}) & \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2(1 + \frac{1}{\ell^2 q^2})
\end{pmatrix}
\begin{pmatrix}\tilde{b}^{\mathrm{R}}_{1 \mathrm{R}} \\ \tilde{b}^{\mathrm{S}}_{1 \mathrm{S}} \end{pmatrix}, \label{eq:L3 tilde}\\
\mathcal{L}_{\mathrm{L} 1} &= \frac{1}{8\mu} \frac{2}{1+\nu}
\begin{pmatrix}\tilde{b}^\dagger_{1 -} & \tilde{b}^{\mathrm{L} \dagger}_{\mathrm{R}\mathrm{S}} \end{pmatrix}
\begin{pmatrix}
(1-\nu) \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2 & \mathrm{i}\sqrt{2} \nu \frac{1}{c_\mathrm{T}}\omega_n \frac{c_\mathrm{d}}{c_\mathrm{T}}\tilde{p} \\
-\mathrm{i} \sqrt{2}\nu \frac{1}{c_\mathrm{T}}\omega_n \frac{c_\mathrm{d}}{c_\mathrm{T}}\tilde{p} & \frac{1}{c_\mathrm{T}^2}\omega_n^2 + 2 (1 + \nu) q^2
\end{pmatrix}
\begin{pmatrix}\tilde{b}_{1 -}\\ \tilde{b}^{\mathrm{L} }_{\mathrm{R}\mathrm{S}} \end{pmatrix}. \label{eq:solid Lagrangian gauge field L1 tilde}\\
\mathcal{L}_{\mathrm{L} 2} &= \frac{1}{4\mu} \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2 \lvert \tilde{b}_{1+} \rvert^2\label{eq:solid Lagrangian gauge field L2 tilde}
\end{align}
Using the relation $A_1 = \frac{c_\mathrm{d}}{c_\mathrm{T}}\frac{\tilde{p}}{p} \tilde{A}_1$ for any field $A_\mu$, the stress propagators Eqs.~\eqref{eq:longitudinal propagator stress propagator relation}, \eqref{eq:transverse propagator gauge field}
become in this gauge fix,
\begin{align}
G_\mathrm{L} &=\frac{1}{\kappa} - \frac{1}{(3\kappa)^2} \Big[
\frac{\omega_n^2}{c_\mathrm{T}^2} \langle \tilde{b}^{\mathrm{L}\dagger}_{\mathrm{R} \mathrm{S}}\; \tilde{b}^\mathrm{L}_{\mathrm{R} \mathrm{S}} \rangle
+ 2 \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2} \tilde{p}^2 \langle \tilde{b}^\dagger_{ 1 -}\; \tilde{b}_{1 -} \rangle
+ \mathrm{i} \sqrt{2} \frac{\omega_n}{c_\mathrm{T}} \frac{c_\mathrm{d}}{c_\mathrm{T}} \tilde{p}
\Big( \langle \tilde{b}^{\mathrm{L}\dagger}_{\mathrm{R} \mathrm{S}}\; \tilde{b}_{1 -} \rangle
- \langle \tilde{b}^\dagger_{1 -}\; \tilde{b}^{\mathrm{L}}_{\mathrm{R} \mathrm{S}} \rangle
\Big)
\Big],
\label{eq:longitudinal propagator gauge field tilde}\\
G_{\mathrm{T} 1} &= \frac{1}{\mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \Big[
\frac{c_\mathrm{d}^2}{c_\mathrm{T}^2} \tilde{p}^2 \langle \tilde{b}^{\mathrm{L}\dagger}_{1 \mathrm{S}} \; \tilde{b}^{\mathrm{L}}_{1 \mathrm{S}} \rangle + \frac{\omega_n^2}{c_\mathrm{T}^2} \langle \tilde{b}^{\mathrm{R}\dagger}_{\mathrm{R}\mathrm{S}}\; \tilde{b}^\mathrm{R}_{\mathrm{R} \mathrm{S}}\rangle + \mathrm{i} \frac{\omega_n}{c_\mathrm{T}} \frac{c_\mathrm{d}}{c_\mathrm{T}} \tilde{p} \Big( \langle \tilde{b}^{\mathrm{R}\dagger}_{\mathrm{R}\mathrm{S}} \; \tilde{b}^\mathrm{L}_{1 \mathrm{S}}\rangle - \langle \tilde{b}^{\mathrm{L}\dagger}_{1\mathrm{S}}\; \tilde{b}^\mathrm{R}_{\mathrm{R} \mathrm{S}}\rangle \Big)
\Big],\nonumber\\
G_{\mathrm{T} 2} &= \frac{1}{\mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \Big[
\frac{c_\mathrm{d}^2}{c_\mathrm{T}^2}\tilde{p}^2 \langle \tilde{b}^{\mathrm{L}\dagger}_{1 \mathrm{R}}\; \tilde{b}^{\mathrm{L}}_{1 \mathrm{R}} \rangle + \frac{\omega_n^2}{c_\mathrm{T}^2} \langle \tilde{b}^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}}\; \tilde{b}^\mathrm{S}_{\mathrm{R}\mathrm{S}}\rangle + \mathrm{i} \frac{\omega_n}{c_\mathrm{T}} \frac{c_\mathrm{d}}{c_\mathrm{T}} \tilde{p} \Big( \langle \tilde{b}^{\mathrm{S}\dagger}_{\mathrm{R}\mathrm{S}} \; \tilde{b}^\mathrm{L}_{1 \mathrm{R}}\rangle - \langle \tilde{b}^{\mathrm{L}\dagger}_{1\mathrm{R}}\; \tilde{b}^\mathrm{S}_{\mathrm{R}\mathrm{S}}\rangle \Big)
\Big],\nonumber\\
G_{\mathrm{T} 3} &= \frac{1}{\mu \ell^2 q^2} - \frac{1}{(2 \mu \ell^2 q^2)^2} \Big[\frac{c_\mathrm{d}^2}{c_\mathrm{T}^2} \tilde{p}^2 \langle \tilde{b}^{\mathrm{R}\dagger}_{1 \mathrm{R}} \; \tilde{b}^{\mathrm{R}}_{1 \mathrm{R}} \rangle + \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2} \tilde{p}^2 \langle \tilde{b}^{\mathrm{S}\dagger}_{1 \mathrm{S}} \; \tilde{b}^\mathrm{S}_{1\mathrm{S} }\rangle + \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2} \tilde{p}^2 \langle \tilde{b}^{\mathrm{S}\dagger}_{1\mathrm{S}} \; \tilde{b}^\mathrm{R}_{1 \mathrm{R}}\rangle + \frac{c_\mathrm{d}^2}{c_\mathrm{T}^2} \tilde{p}^2 \langle \tilde{b}^{\mathrm{R}\dagger}_{1\mathrm{R}}\; \tilde{b}^\mathrm{S}_{1\mathrm{S} }\rangle \Big].\label{eq:transverse propagator gauge field tilde}
\end{align}
\end{widetext}
\paragraph{Dislocation condensates with two velocities.}
Eqs.~\eqref{eq:dislocation Higgs term} and \eqref{eq:solid stress gauge field Lagrangian tilde} suffice to calculate the long wavelength properties of the columnar, smectic and nematic phases in the idealized case where there is only one dislocation velocity scale $c_\mathrm{d}$. For completeness we now consider the more general case where screw dislocations move with velocity $c_\mathrm{s}$ while edge dislocations move with velocity $c_\mathrm{e}$, see Sec.~\ref{subsec:Dislocation worldsheets}. As we mentioned there, the dislocation line density $J^a_{\tau n}$
represents screw dislocations when $a = n$ and edge dislocations when $a \neq n$. Since these dislocations couple minimally to dual stress gauge field components $b^a_{\tau n}$, it is clear that in the generalization of Eq.~\eqref{eq:dislocation condensate minimal coupling step 2}, the terms $\lvert b^a_{\tau n} \rvert^2$ with $a = n$ should come with the screw dislocation velocity $c_\mathrm{s}$ while those with $a \neq n$ should come with the edge dislocation velocity $c_\mathrm{e}$. Explicitly, the Higgs term is (with implicit summation over $m,n$)
\begin{widetext}
\begin{align}
\mathcal{L}_\mathrm{min.coup.} &= \frac{1}{4} \sum_a \lvert \Phi^a \rvert^2 \Big( \frac{1}{c_\mathrm{s}^2} 2(\partial_\tau \phi^a_a - \partial_a \phi^a_\tau - b^a_{\tau a} )^2
+ \sum_{b \neq a}\frac{1}{c_\mathrm{e}^2} 2(\partial_\tau \phi^a_b - \partial_b \phi^a_\tau - b^a_{\tau b} )^2
+ (\partial_m \phi^a_n - \partial_n \phi^a_m - b^a_{m n} )^2\Big).
\label{eq:dislocation condensate minimal coupling step 4}
\end{align}
\end{widetext}
\paragraph{The Lorenz gauge fix with two velocities.}
It is possible to choose the Lorenz gauge fix $\frac{1}{c^2} \partial_\tau b^a_{\tau \nu} + \partial_m b^a_{m\nu} = 0$ with a separate velocity for each combination of $a,\nu$. That is, the gauge conditions are
\begin{align}
\frac{1}{c_\mathrm{s}^2} \partial_\tau b^a_{\tau a} + \partial_m b^a_{m a} &= 0\qquad {\text{(no sum } a \text{)}},\label{eq:screw gauge fix}\\
\frac{1}{c_\mathrm{e}^2} \partial_\tau b^a_{\tau b} + \partial_m b^a_{m b} &= 0\qquad a \neq b.\label{eq:edge gauge fix}
\end{align}
This can be seen as follows. The gauge transformation is Eq.~\eqref{eq:stress gauge transformation}, which in the presence of the condensate fields is
\begin{align}\label{eq:stress gauge transformation dislocation condensate}
b^a_{\mu\nu} (x) &\to b^a_{\mu\nu}(x) + \partial_\mu \varepsilon^a_\nu(x) - \partial_\nu \varepsilon^a_\mu(x).\nonumber\\
\phi^a_\nu(x) &\to \phi^a_\nu(x) + \varepsilon^a_\nu(x) + \partial_\nu \zeta^a(x).
\end{align}
where $\zeta^a$ is any flavored scalar field independent of $\varepsilon^a_\nu$, while the
Lagrangian Eq.~\eqref{eq:dislocation condensate minimal coupling step 4} is invariant under the addition of the gradient. However, two sets of transformations of $\varepsilon^a_\nu$, $\zeta^a$ that differ as
follows with respect to a flavored scalar field $\eta^a$
\begin{align}
\varepsilon^a_\nu &\to \varepsilon^a_\nu + \partial_\nu \eta^a, &
\zeta^a &\to \zeta^a - \eta^a
\end{align}
will lead to the exact same gauge transformations Eq.~\eqref{eq:stress gauge transformation dislocation condensate}. In others words, there is a redundancy in the gauge transformations themselves, a ``gauge-in-the-gauge''. We can use this freedom to choose $\eta^a$ in such a way that
\begin{equation}
\frac{1}{c_\mathrm{s}^2} \partial_\tau \varepsilon^a_{\tau} + \partial_m \epsilon^a_m = 0. \label{eq:gauge-in-the-gauge fix}
\end{equation}
Consider the transformation of the following quantity (no sum over $a$),
\begin{align}
&\frac{1}{c_\mathrm{s}^2} \partial_\tau b^a_{\tau a} + \partial_m b^a_{m a} \to \nonumber\\
& \frac{1}{c_\mathrm{s}^2} \partial_\tau b^a_{\tau a} + \partial_m b^a_{m a} + (\frac{1}{c_\mathrm{s}^2} \partial_\tau^2 + \partial_m^2)\varepsilon^a_a + \partial_a (\frac{1}{c_\mathrm{s}^2} \partial_\tau \varepsilon^a_\tau + \partial_m \varepsilon^a_m).
\end{align}
The last term vanished by the gauge-in-the-gauge fix, and we see that we can choose $\varepsilon^a_a$ in such a way that Eq.~\eqref{eq:screw gauge fix} holds. Afterwards,
taking the divergence with respect to the velocity $c_\mathrm{e}$ and performing a gauge transformation leads for the edge components $b^a_{\mu b}$ ($a \neq b$) to,
\begin{align}
&\frac{1}{c_\mathrm{e}^2} \partial_\tau b^a_{\tau b} + \partial_m b^a_{m b} \to \nonumber\\
& \frac{1}{c_\mathrm{e}^2} \partial_\tau b^a_{\tau b} + \partial_m b^a_{m b} + (\frac{1}{c_\mathrm{e}^2} \partial_\tau^2 + \partial_m^2) \varepsilon^a_b + (1 - \frac{c_\mathrm{s}^2}{c_\mathrm{e}^2}) \partial_b \partial_m \varepsilon^a_m.
\end{align}
Here we used Eq.~\eqref{eq:gauge-in-the-gauge fix}. Now we can choose $\varepsilon^a_b$ ($a \neq b$) in such a way that this whole expression vanishes.
After imposing these gauge fixes Eqs.~\eqref{eq:screw gauge fix}, \eqref{eq:edge gauge fix}, the condensate phase degrees of freedom have been decoupled and do not explicitly contribute to the stress propagators. The recipe to perform the calculations is to add the solid Lagrangian Eq.~\eqref{eq:solid stress gauge field Lagrangian} in the same gauge fix to the Higgs term Eq.~\eqref{eq:dislocation condensate minimal coupling step 4} and use this to calculate the various stress propagators.
\subsection{Higgs term and glide constraint}
Again, our starting point is Eq.~\eqref{eq:dislocation Higgs term} where the sum over $a$ is now over two orthogonal directions which span the plane in which translational symmetry is restored. Since we depart from isotropic elasticity, all planes of dislocation condensation are equivalent. But even in the general case, as long as the condensate amplitudes $\Omega^a$ are the same for all Burgers directions, the form of the Higgs term is independent of the original crystal space group, due to the fact that the directions of translational symmetry restoration must be orthogonal as explained in Sec.~\ref{sec:Symmetry principles of quantum liquid crystals}. Without loss of generality, we can choose the Burgers vectors of condensed dislocations to lie in the $xz$-plane. Below we always assume $\Omega^x = \Omega^z \equiv \Omega$, which can be achieved for judicious choices of Ginzburg--Landau parameters in Eq.~\eqref{eq:three-condensate potential}.
As before, we choose the dislocation Lorenz gauge fix Eq.~\eqref{eq:dislocation Lorenz gauge fix} and disregard the condensate phase degrees of freedom as they do not couple to stress.
First we need to take care of the glide constraint in Eq.~\eqref{eq:dislocation Higgs term}. This follows a derivation similar to that of the nematic in Sec.~\eqref{subsec:nematic glide constraint}. For the prefactor of $\lambda^\dagger \lambda$, we have the result of Eq.~\eqref{eq:glide constraint multiplier field prefactor}, which now must be summed over the two directions. Integrating out the Lagrange multiplier field $
\lambda$ will lead to
\begin{align}
\mathcal{L}^{(xz)}_\mathrm{glide} = - 2 \frac{\Omega^2}{4 c_\mathrm{T}^2 \mu} \frac{ \tilde{p}^2}{2\frac{1}{c_\mathrm{d}^2} \omega_n^2 + q_x^2 + q_z^2} \left\lvert \tilde{b}^x_{yz} + \tilde{b}^z_{xy} \right\rvert^2.
\end{align}
If one instead chooses the condensation in the $xy$- or $yz$-plane, similar terms are obtained. The factor of 2 in front comes from summing over antisymmetric indices.
The Higgs term Eq.~\eqref{eq:dislocation Higgs term} for the smectic is then given by:
\begin{align}
\mathcal{L}^{(xz)}_\mathrm{Higgs}
&= \frac{\Omega^2}{2 c_\mathrm{T}^2 \mu} \Big[ \sum_{a=x,z}\Big(
\lvert \tilde{b}^a_{1\mathrm{R}} \rvert^2 + \lvert \tilde{b}^a_{1\mathrm{S}} \rvert^2 + \lvert \tilde{b}^a_{\mathrm{R}\mathrm{S}} \rvert^2 \Big) \nonumber\\
&\phantom{mmnmm}- \frac{\tilde{p}^2}{2 \frac{1}{c_\mathrm{d}^2} \omega_n^2 + q_x^2+q_z^2} \lvert \tilde{b}^x_{yz} +\tilde{b}^z_{xy} \rvert^2 \Big].
\label{eq:dislocation Higgs term smectic xz}
\end{align}
with analogous expressions for the $xy$- and $yz$-plane condensates with the coordinates $x,y,z$ permuted. In contrast to the nematic, the glide constraint term does not have a nice, short expression in Fourier space coordinates. We will leave the term as it is, while in the calculations we convert it to Fourier space coordinates, leading in general to a $9 \times 9$ matrix in the basis $\tilde{b}^E_{1 \mathrm{R}}$, $\tilde{b}^E_{1 \mathrm{S}}$, $\tilde{b}^E_{\mathrm{R}\mathrm{S}}$, $E=\mathrm{L},\mathrm{R},\mathrm{S}$. Without loss of generality, we can focus on just one of these three choices, the other two can be obtained through a simple coordinate transformation, in our isotropic (or cubic) case.
\begin{figure}[t]
\includegraphics[width=8cm]{smectic_velocity.pdf}
\caption{Velocities of the massless modes in the smectic as a function of interrogation angle $\eta$ from Eq.~\eqref{eq:smectic L T1 massless dispersion}. Here we have chosen a representative value of the Poisson ration $\nu = 0.3$. The dotted lines represent $c_\mathrm{L}$, $c_\kappa$ and $c_\mathrm{T}$, while the dashed lines represent $\frac{1}{\sqrt{1-2\nu}}c_\mathrm{T}$ and $\sqrt{\frac{2+2\nu}{3+2\nu}} c_\mathrm{T}$ as explained in the text. The longitudinal mode (blue) reaches the maximum value $c_\mathrm{L}$ at $\eta = \pi/2$, but not at $\eta =0$, unlike the 2D case. Its minimum value is $c_\kappa$. The transverse mode (yellow) never reaches the full transverse photon velocity $c_\mathrm{T}$. Its pole strength vanishes both as $\eta \to 0$ and as $\eta \to \pi/2$.}\label{fig:smectic velocity}
\end{figure}
\begin{figure*}
\includegraphics[width=7.7cm]{3DsmecxzT2.png}
\hfill
\includegraphics[width=7.7cm]{3DsmecxzL.png}\\
{\centering
\includegraphics[scale=.6]{legends_smecxz.pdf}
}\\
\caption{Spectral function of the quantum smectic in units of the inverse shear modulus $1/\mu$, with Poisson ratio $\nu = 0.2$, at a representative angle $\eta = 3 \pi/16$. For a clear picture we have arbitrarily set $c_\mathrm{d} = 3 c_\mathrm{T}$. The inset is a zoom up near the origin. The width of the poles is artificial and denotes the relative pole strengths: these ideal poles are actually infinitely sharp. Now the poles of the sectors $\mathrm{L}, {\mathrm{T} 1}$ are mixed, and of $\mathrm{T} 2, \mathrm{T} 3$ as well. We plot only $S_{\mathrm{T} 2}$ and $S_{\mathrm{L}}$, as $S_{\mathrm{T} 3}$ resp. $S_{\mathrm{T} 1}$ have the same poles although with different pole strengths. (a) In the purely transverse response we find the rotational Goldstone mode although with modified velocity. There are two gapped modes with gaps $\Omega$ resp. $\frac{1}{\sqrt{2}} \Omega$. (b) In the $\mathrm{L}$--$\mathrm{T} 1$ sector there are two gapless modes with velocities Eq.~\eqref{eq:smectic L T1 massless dispersion}, which extrapolate to the longitudinal and transverse phonon with velocities $c_\mathrm{L}$ and $c_\mathrm{T}$ at high energies. Furthermore there are two gapped modes with gaps $\Omega$ and $\frac{1}{\sqrt{2}}\Omega$ that extrapolate to the `transverse' condensate mode with velocity $c_\mathrm{d}$ and the `longitudinal' condensate mode with velocity $\frac{1}{\sqrt{2}}c_\mathrm{d}$ respectively at high energies. }\label{fig:smectic spectral functions}
\includegraphics[width=7.7cm]{3DsmecxyT.png}
\hfill
\includegraphics[width=7.7cm]{3DsmecxyL.png}\\
{\centering
\includegraphics[scale=.6]{legends_smecxy.pdf}
}\\
\caption{Spectral function of the quantum smectic with momentum in the liquid plane ($\eta = 0$), in units of the inverse shear modulus $1/\mu$, with Poisson ratio $\nu = 0.2$. For a clear picture we have arbitrarily set $c_\mathrm{d} = 3 c_\mathrm{T}$. The inset is a zoom up near the origin. The width of the poles is artificial and denotes the relative pole strengths: these ideal poles are actually infinitely sharp. The response is independent of the in-plane interrogation angle $\eta$. (a) The transverse response shows the rotational Goldstone mode with velocity $\frac{1}{\sqrt{2}}c_\mathrm{d}$ as well as the quadratically dispersing undulation mode. There are three gapped poles, one with gap $\Omega$ and two with gap $\frac{1}{\sqrt{2}}\Omega$ (the splitting between these latter two modes at intermediate velocities is just barely visible). (b) In the longitudinal response we find a massless mode with velocity $\frac{1}{\sqrt{1-2\nu}} c_\mathrm{T} = \sqrt{c_\mathrm{L}^2 -c_\mathrm{T}^2}$ at low energies extrapolating to the longitudinal phonon at high energies, and a gapped mode with gap $\Omega$.}\label{fig:eta0 smectic spectral functions}
\end{figure*}
\subsection{Collective modes in the quantum smectics}
\begin{table*}
\hfill
\begin{tabular}{clclc}
\multicolumn{3}{c}{general $\eta$}\\
\toprule
sector & & massless & &massive \\
\hline
$\mathrm{L}$ & \rdelim\}{2}{1mm}[] & \multirow{2}{*}{Eq.~\eqref{eq:smectic L T1 massless dispersion}} & \rdelim\}{2}{1mm}[] & $\Omega$ \\
$\mathrm{T} 1$ & & & & $\frac{1}{\sqrt{2}} \Omega$ \\
$\mathrm{T} 2$ &\rdelim\}{2}{1mm}[] & \multirow{2}{*}{$\frac{1}{\sqrt{2}} \sqrt{1 + \sin^2 \eta} \;c_\mathrm{d} q$} &\rdelim\}{2}{1mm}[] & $\Omega$ \\
$\mathrm{T} 3$ & & & & $\frac{1}{\sqrt{2}} \Omega$ \\
\hline
total & & 3 & & 4 \\
\botrule
\end{tabular}
\hfill
\begin{tabular}{cccc}
\multicolumn{3}{c}{$\eta = 0$}\\
\toprule
sector & massless & & massive \\
\hline
$\mathrm{L}$ & $\frac{1}{\sqrt{1-2\nu}} c_\mathrm{T} q $ & & $\Omega + \frac{\frac{1}{2}c_\mathrm{d}^2 + c_\mathrm{T}^2}{2 \Omega} q^2$ \\
$\mathrm{T} 1$ & $\frac{1}{\sqrt{2}\Omega} c_\mathrm{d} c_\mathrm{T} q^2$ & & $\frac{1}{\sqrt{2}}\Omega + \frac{c_\mathrm{d}^2 + c_\mathrm{T}^2}{\sqrt{2} \Omega} q^2 $ \\
$\mathrm{T} 2$ & $\frac{1}{\sqrt{2}} c_\mathrm{d} q$ & &$\Omega + \frac{\frac{1}{2} c_\mathrm{d}^2 + c_\mathrm{T}^2}{2 \Omega} q^2 $ \\
$\mathrm{T} 3$ & - & & $\frac{1}{\sqrt{2}}\Omega + \frac{c_\mathrm{d}^2}{\sqrt{2} \Omega} q^2$ \\
\hline
total & 3 & & 4 \\
\botrule
\end{tabular}
\hfill
\begin{tabular}{cccc}
\multicolumn{3}{c}{$\eta = \pi/2$}\\
\toprule
sector & massless & & massive \\
\hline
$\mathrm{L}$ & $c_\mathrm{L} q $ & & - \\
$\mathrm{T} 1$ & - & & $\sqrt{\frac{1}{2} \Omega^2 + c_\mathrm{T}^2 q^2} $ \\
$\mathrm{T} 2$ & - & & $\sqrt{\frac{1}{2} \Omega^2 + c_\mathrm{T}^2 q^2} $ \\
$\mathrm{T} 3$ & $c_\mathrm{d} q$ & & - \\
\hline
total & 2 & & 2 \\
\botrule
\end{tabular}
\hfill
\null\\
\caption{Collective modes in the smectic phases. Indicated are the dispersion relations to lowest orders in momentum.
The leftmost table shows the case for general angle $0 < \eta < \pi/2$. The propagators $G_\mathrm{L}$ and $G_{\mathrm{T} 1}$ share their poles: two highly $\eta$-dependent massless poles with velocities depicted in Fig.~\ref{fig:smectic velocity}, and two massive ones. The transverse propagator $G_{\mathrm{T} 2}$ and $G_{\mathrm{T} 3}$ contain the rotational Goldstone mode the velocity interpolates between $\frac{1}{\sqrt{2}} c_\mathrm{d}$ and $c_\mathrm{d}$. The middle table shows the results for $\eta = 0$, momentum in the liquid plane. As in the 2D smectic, the $T1$-transverse pole is the undulation mode with quadratic dispersion. The rotational mode is now exclusive found in $G_{\mathrm{T} 2}$. The rightmost table show the case for momentum in the solid direction, $\eta = \pi/2$. The number of modes is greatly reduced. We find results similar to the 2D smectic: the compression mode obtains the full longitudinal-phonon velocity, while the shear phonons are gapped. The third propagator $G_{\mathrm{T} 3}$ probes torque correlations in the liquid plane and picks up the rotational Goldstone mode.}\label{table:smectic spectrum}
\end{table*}
To obtain the spectrum of modes in the smectic phase, we add any coordinate of permutation of Eq.~\eqref{eq:dislocation Higgs term smectic xz} to Eq.~\eqref{eq:solid stress gauge field Lagrangian tilde} and calculate the propagators Eqs.~\eqref{eq:longitudinal propagator gauge field tilde}, \eqref{eq:transverse propagator gauge field tilde}. This is most easily performed on a computer. The result is highly dependent on the direction of momentum just as it was in 2+1D. Indeed, one would assume that the parent crystal anisotropy remains influential when not all translational symmetry is restored. However, in the present case of an isotropic crystal, matters do simplify quite a bit, as we can suffice with the polar angle $\eta$ while setting the azimuthal angle $\zeta = 0$ , see Eq.~\eqref{eq:eta zeta angles} and Fig.~\ref{fig:smectic eta angle}. Then the momentum $\mathbf{q} = ( q \cos \eta, q \sin \eta ,0)$ is in the $xy$-plane, and the polar angle $\eta$ tunes between momentum completely in the liquid plane $\eta = 0$, and parallel to the solid direction $\eta = \pi/2$. With these choices, the transverse $\mathrm{S}$-direction is parallel to the $z$-axis, while $\mathrm{L}$ and $\mathrm{R}$ lie in the $xy$-plane. As before, we perform analytic continuation to real time $\omega_n \to \mathrm{i} \omega - \delta$, where we drop the infinitesimal factors $\mathrm{i} \delta$ at the last step of the calculation for ease of notation.
The general form of the propagators is too complicated to write down explicitly. However, we can inspect the poles of the propagators to lowest order in momentum. As we have seen before in QLC2D, the longitudinal and transverse sectors mix at general angle $\eta$, and the characteristics of the responses are hybrid. In the 3D case, for the $xz$-condensate, the pair $G_\mathrm{L}$--$G_{\mathrm{T} 1}$ share the same poles, as do the pair $G_{\mathrm{T} 2}$--$G_{\mathrm{T} 3}$. The pole strengths all differ, however. In Fig.~\ref{fig:smectic spectral functions} we have plotted the spectral functions of $G_\mathrm{L}$ and $G_{\mathrm{T} 2}$ respectively. In the $\mathrm{L}$--$\mathrm{T} 1$-sector we find four poles. Two of the poles are massive with gaps $\Omega$ and $\Omega/\sqrt{2}$ respectively, to be interpreted as coming from the dislocation condensate in two Burgers directions (recall there were three massive modes in the nematic, see Table~\ref{table:nematic spectrum}). The two other poles are massless with dispersions
\begin{widetext}
\begin{align}
(\omega^{\mathrm{L}, \mathrm{T} 1}_{1,2})^2 = \frac{(3-2\nu) - (1-2\nu)\cos 2 \eta \pm \sqrt{ 1 - 4\nu + 20\nu^2 - 2 (3 - 8\nu + 4\nu^2) \cos 2 \eta + 3(3 - 4\nu - 4\nu^2) \cos^2 2 \eta} }{4(1-2\nu)}c_\mathrm{T}^2 q^2 + \ldots
\label{eq:smectic L T1 massless dispersion}
\end{align}
\end{widetext}
Overall this result is very comparable to 2+1D: there are two massless modes, one extrapolating to the longitudinal phonon and one to the transverse phonon, showing up in both propagators. There is obviously a complicated angle-dependence here. We have plotted the velocity as a function of $\eta$ for a representative value $\nu = 0.3$ in Fig.~\ref{fig:smectic velocity}. We can identify the pole with the plus sign as having longitudinal character while the minus sign has transverse character. Recall that in 2D, the longitudinal velocity varies from $c_\mathrm{L}$ at $\eta = 0$ to $c_\kappa$ at $\eta = \pi/4$ back to $c_\mathrm{L}$ at $\eta =\pi/2$. Conversely, the 2D transverse pole has vanishing pole strength and vanishing velocity at $\eta = 0,\pi/2$, while obtaining the full transverse velocity $c_\mathrm{T}$ at $\eta = \pi/4$~\cite{QLC2D}. Here in 3D the behavior is somewhat different. For the longitudinal mode, the maximum velocity is $c_\mathrm{L}$ at $\eta =\pi/2$ and the minimum velocity is $c_\kappa$ at $\eta = \tfrac{1}{2} \arccos \tfrac{1}{3}$. At $\eta =0$ the velocity is smaller than $c_\mathrm{L}$, namely $\frac{1}{\sqrt{1-2\nu}}c_\mathrm{T} = \sqrt{c_\mathrm{L}^2 - c_\mathrm{T}^2}$. The transverse pole has again vanishing pole strength and vanishing velocity at $\eta = 0,\pi/2$. However, the maximum velocity is smaller than $c_\mathrm{T}$, namely $\sqrt{\frac{2+2\nu}{3+2\nu}} c_\mathrm{T}$ attained at $\eta = \tfrac{1}{2} \arccos \frac{1-2\nu}{3+2\nu}$. Note that $c_\mathrm{L},c_\kappa \to \infty$ as $\nu \to 0.5$.
Compared to the 2D smectics we have two more propagators, namely $G_{\mathrm{T} 2}$ and $G_{\mathrm{T} 3}$.
Again we find is that these two propagators share their poles although the poles strengths are different for each case. For $G_{\mathrm{T} 2}$ and $G_{\mathrm{T} 3}$ we find three modes with low-energy dispersions:
\begin{align}
\omega^{\mathrm{T} 2, \mathrm{T} 3}_1 &= \frac{1}{\sqrt{2}} \sqrt{1 + \sin^2 \eta} \;c_\mathrm{d} q + \ldots ,\nonumber \\
(\omega^{\mathrm{T} 2, \mathrm{T} 3}_2)^2 &= \Omega^2 + \big( \frac{1}{2} c_\mathrm{d}^2( 1 + \sin^2 \eta) + c_\mathrm{T}^2 \cos^2 \eta \big) q^2 + \ldots,\nonumber\\
(\omega^{\mathrm{T} 2, \mathrm{T} 3}_3)^2 &= \frac{1}{2} \Omega^2 + ( c_\mathrm{d}^2 \cos^2\eta + c_\mathrm{T}^2 \sin^2 \eta ) q^2 + \ldots.\label{eq:smectic T2 T3 poles}
\end{align}
The first mode is a rotational Goldstone mode! We will say much more about this below in Sec.~\ref{subsec:Rotational Goldstone mode smectic}. Its velocity depends on the interrogation angle, interpolating between the `longitudinal torque' and `transverse torque' characteristics explained in Sec.~\ref{subsec:Torque stress in the quantum nematic}. This mode is somehow `divided' between $G_{\mathrm{T} 2}$ and $G_{\mathrm{T} 3}$. Next to that there are two gapped modes, mixing the condensate phase mode with the gapped shear phonons.
Eqs.~\eqref{eq:smectic L T1 massless dispersion}, \eqref{eq:smectic T2 T3 poles} are not strictly valid for the special angles $\eta = 0$ (momentum parallel to the `liquid' $x$-direction) and $\eta=\pi/2$ (momentum parallel to the `solid' $y$-direction), although the limiting behavior is correct. In fact, the form of the propagators change and the number of poles is reduced, to such an extent that we can write down the results explicitly.
For $\eta = 0$, momentum in the liquid plane, the propagators read
\begin{widetext}
\begin{align}
G^{(\eta = 0)}_\mathrm{L} &= \frac{1}{\mu} \frac{-c_\mathrm{T}^2 q^2 ( \omega^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2 - \Omega^2)}{(\omega^2 - c_\mathrm{L}^2 q^2)(\omega^2 - \tfrac{1}{2} c_\mathrm{d}^2 q^2) - \Omega^2 (\omega^2 -\frac{1}{1-2\nu} c_\mathrm{T}^2 q^2)},
\label{eq:GL eta0 smectic}\\
G^{(\eta = 0)}_{\mathrm{T} 1} &= \frac{1}{\mu} \frac{ \frac{1}{2}\omega^2 \Omega^2 - c_\mathrm{T}^2 q^2 ( \omega^2 - c_\mathrm{d}^2 q^2 -2\Omega^2)}{(\omega^2 - c_\mathrm{T}^2 q^2)(\omega^2 - c_\mathrm{d}^2 q^2) - \omega^2 \frac{1}{2}\Omega^2},\label{eq:GT1 eta0 smectic}\\
G^{(\eta = 0)}_{\mathrm{T} 2} &= \frac{1}{\mu} \frac{ - c_\mathrm{T}^2 q^2 ( \omega^2 - c_\mathrm{d}^2 q^2) - \Omega^2 ( \omega^2- 2c_\mathrm{T}^2 q^2 - \frac{1}{2}c_\mathrm{d}^2 q^2 - \Omega^2)}{(\omega^2 - c_\mathrm{T}^2 q^2)(\omega^2 - c_\mathrm{d}^2 q^2) - \Omega^2( \omega^2 - \frac{1}{2}c_\mathrm{d}^2 q^2)},\label{eq:GT2 eta0 smectic}\\
G^{(\eta = 0)}_{\mathrm{T} 3} &= \frac{1}{\mu} \frac{ -\frac{1}{2}\Omega^2}{\omega^2 - c_\mathrm{d}^2 q^2 - \frac{1}{2}\Omega^2}.\label{eq:GT3 eta0 smectic}
\end{align}
\end{widetext}
It is instructive to compare this to the propagators of the 2D nematic in Eqs.~\eqref{eq:2D nematic GL}, \eqref{eq:2D nematic GT}. Indeed the longitudinal propagator is identical except for the form of the longitudinal velocity, due to the different dimensionality in Eq.~\eqref{eq:longitudinal velocity definition}. Also, the velocity $\frac{1}{\sqrt{1-2\nu}} c_\mathrm{T} = \sqrt{c_\mathrm{L}^2 -c_\mathrm{T}^2}$ is different from $c_\kappa$ in 3D. (However, in 2D there is the relation $c_\kappa^\mathrm{2D} = \sqrt{(c_\mathrm{L}^\mathrm{2D})^2 - c_\mathrm{T}^2}$.) Notice as well that Eq.~\eqref{eq:GL eta0 smectic} is different from the 3D nematic Eq.~\eqref{eq:nematic longitudinal propagator} in particular considering the velocity of the massive condensate mode: it contains a prefactor of $\frac{1}{2}$ instead of $\frac{1}{3}$. The longitudinal propagator contains
two modes, with dispersion relations
\begin{align}
\omega^{\mathrm{L}}_1(\eta = 0) &= \frac{1}{\sqrt{1 - 2\nu}} c_\mathrm{T} q + \ldots ,\nonumber \\
\omega^{\mathrm{L}}_2 (\eta = 0)&= \Omega + \frac{\frac{1}{2}c_\mathrm{d}^2 + c_\mathrm{T}^2}{2 \Omega} q^2 + \ldots . \label{eq:eta0 smectic longitudinal dispersions}
\end{align}
The first, massless mode is a mixture of the 2D compression mode and the 3D longitudinal phonon, while the second, massive mode is identical to the 2D nematic condensate mode.
The transverse propagator $G_{\mathrm{T} 2}$ Eq.~\eqref{eq:GT2 eta0 smectic} is identical to that of the 2D nematic Eq.~\eqref{eq:2D nematic GT}. This makes sense: from Eq.~\eqref{eq:transverse propagator from torque correlators} we know that $G_{\mathrm{T} 2}$ measures the shear correlations in the $\mathrm{L}\mathrm{S}$-plane which for $\eta,\zeta = 0$ is exactly the liquid $xz$-plane. This propagator does not know anything about the third dimension, and it only sees a 2D dislocation condensate: a 2D nematic. This propagator also contains one massless and one massive mode, with dispersion relations,
\begin{align}
\omega^{\mathrm{T} 2}_1(\eta = 0) &= \frac{1}{\sqrt{2}} c_\mathrm{d} q + \ldots ,\nonumber \\
\omega^{\mathrm{T} 2}_2(\eta = 0) &= \Omega + \frac{\tfrac{1}{2} c_\mathrm{d}^2 + c_\mathrm{T}^2}{2 \Omega} q^2 + \ldots . \label{eq:eta0 smectic T2 dispersions}
\end{align}
These are the rotational Goldstone mode and a massive condensate--shear mode.
Next, look at $G_{\mathrm{T} 1}$ in Eq.~\eqref{eq:GT1 eta0 smectic}, and compare this with the transverse propagator of the 2D smectic at interrogation angle $\eta_{\rm 2D}=0$ in the limit $\ell \to 0$, which we take from QLC2D:
\begin{equation}
G^\mathrm{2D smec}_{\mathrm{T}} (\eta_{\rm 2D} = 0) = \frac{1}{\mu}\frac{ -\tfrac{1}{2}\omega^2 \Omega^2 - c_\mathrm{T}^2 q^2 (\omega^2 - c_\mathrm{d}^2 q^2 + 2 \Omega^2)}{(\omega^2 - c_\mathrm{T}^2 q^2) (\omega^2 - c_\mathrm{d}^2 q^2 ) - \tfrac{1}{2} \omega^2 \Omega^2 }.
\end{equation}
Here we have rescaled the Higgs mass by a factor of $\sqrt{2}$ due to differing definitions in QLC2D. This equation is indeed identical to Eq.~\eqref{eq:GT1 eta0 smectic}. Again, this makes perfect sense. The 2D smectic propagator at interrogation angle $\eta_{\rm 2D}=0$ has momentum along the liquid direction while the transverse direction is in the solid direction. Now $G_{\mathrm{T} 2}$ is the propagator of shear correlations orthogonal to the $\mathrm{R}$ direction, that is in the $\mathrm{L}\mathrm{S}$-plane. For $\zeta = 0$, $\mathrm{S}$ is along the solid $z$-axis, while the longitudinal direction is within the liquid plane. So this is a perfect match for the 2D smectic propagator. We recall the dispersion relations from QLC2D:
\begin{align}
\omega^{\mathrm{T} 1}_1 (\eta = 0) &= \frac{c_\mathrm{d} c_\mathrm{T} }{\frac{1}{\sqrt{2}}\Omega} q^2 + \ldots ,\nonumber \\
\omega^{\mathrm{T} 1}_2(\eta = 0) &= \frac{1}{\sqrt{2}}\Omega + \frac{c_\mathrm{d}^2 + c_\mathrm{T}^2}{\sqrt{2}\Omega} q^2 + \ldots . \label{eq:smectic xy T2 dispersions}
\end{align}
The first mode is massless with quadratic dispersion. This is the {\em undulation mode}~\cite{CvetkovicZaanen06b,QLC2D} known from classical smectic liquid crystals~\cite{DeGennesProst95,SinghDunmur02,ChaikinLubensky00}: the smectic can still support a reactive response to shear stress, but (at low energies) it is less strong than that of a true solid. This is reproduced here in the 3D smectic.
Finally we have the propagator $G_{\mathrm{T} 3}$ of Eq.~\eqref{eq:GT3 eta0 smectic}, representing shear correlations in the purely transverse plane $\mathrm{R}\mathrm{S}$ (the $yz$-plane for $\eta=0$), which has no counterpart in 2D liquid crystals. It has one massive mode with the exact dispersion
\begin{align}
(\omega^{\mathrm{T} 3})^2(\eta = 0) &= \tfrac{1}{2} \Omega^2 + c_\mathrm{d}^2 q^2.\label{eq:eta0 smectic T3 dispersion}
\end{align}
Is it clearly a massive condensate mode, not influenced by any phonon degrees of freedom, i.e. it does not depend on $c_\mathrm{T}$. Resoundingly, even though the dislocation condensate seems not to couple to stress in our gauge choice, this mode does show up in the physical transverse stress propagator.
In short at $\eta = 0$, for the in-plane response we recover the 2D nematic in the $\mathrm{L}$--$\mathrm{T} 1$ sector apart from some slightly different velocities, while smectic behavior including the undulation mode shows up for momenta transverse to the liquid plane.
The final special case is $\eta = \pi/2$, for which the momentum lies in the solid $y$-direction. At this angle, all four propagators contain only a single mode, with the following dispersion relations:
\begin{align}
\omega^{\mathrm{L}}(\eta = \pi/2) &= c_\mathrm{L} q,\label{eq:smectic longitudinal phonon}\\
(\omega^{\mathrm{T} 1})^2 (\eta = \pi/2)&= \frac{1}{2} \Omega^2 + c_\mathrm{T}^2 q^2,\label{eq:smectic pi/2 gapped shear}\\
(\omega^{\mathrm{T} 2})^2(\eta = \pi/2) &= \frac{1}{2} \Omega^2 + c_\mathrm{T}^2 q^2,\\
\omega^{\mathrm{T} 3} (\eta = \pi/2)&= c_\mathrm{d} q.
\end{align}
For the longitudinal direction coinciding with the solid direction, we retrieve the longitudinal phonon in Eq. \eqref{eq:smectic longitudinal phonon}! Even though the two transverse directions do not support shear rigidity, the velocity of the compressional mode at this angle is identical to that of a crystal. just as it was in 2+1D. Meanwhile, the two transverse phonons in the liquid plane have picked up the Higgs gap $\Omega/\sqrt{2}$, while not being mixed with the condensate phase degree of freedom as indicated by the absence of $c_\mathrm{d}$. The final mode is the rotational Goldstone mode which exists in the transverse, liquid plane. We had already seen in Eq.~\eqref{eq:longitudinal rotational Goldstone dispersions} of the nematic that this Goldstone mode has a different velocity than those appearing in $G_{\mathrm{T} 1}$ and $G_{\mathrm{T} 2}$.
In conclusion, we have derived the spectra of the 3D `isotropic' smectic liquid crystal and uncovered the collective modes that are expected on grounds of the symmetries.
We find for the 3D smectic a similar pattern as in the 2D smectic: depending on the interrogation angle, i.e. the angle of the momentum relative to the smectic plane, we retrieve some solid-like (massless phonons) and some liquid-like (gapped shear, compression mode) features. However, the modes do not in general exhibit a decomposition of the form `solid $\times$ liquid' , but instead show a mixture of the two characteristics through mode coupling. Assuredly, at the special angle $\eta =0$, the undulation mode of classical smectics is reproduced. From QLC2D, we know that $\eta_{\rm 2D} = \pi/4$ (or $[11]$-direction) for the 2D-condensate is a special angle displaying a perfect `solid $\times$ liquid' response, but we have not been able to detect similar `magic angle' in 3D, cf. Fig.~\ref{fig:smectic velocity}, implying that it could be accidental for 2D as there is only one transverse direction.
Naturally, a full analysis incorporating the space group symmetries of the 3D crystal should further affect the mode spectrum in terms of $\eta$ and even $\zeta$. We leave this for future considerations.
\subsection{Rotational Goldstone mode}\label{subsec:Rotational Goldstone mode smectic}
We have seen the emergence of a rotational Goldstone mode of rotations within the liquid plane. We can also see this from the torque stress in the deep Higgs limit $\Omega \to \infty$, similar to Sec.~\ref{subsec:Torque stress in the quantum nematic}.
There is a complementary insight into this phenomenon from symmetry principles. It is of course well established that spontaneous breaking of continuous global symmetries gives rise to massless Goldstone modes, which can be viewed as the low-energy fluctuations along the flat directions of order parameter space. However, it is equally well known that in solids, which break both translational and rotational symmetry, only the translational Goldstone modes, the phonons, are ever seen. This apparent contradiction to the Goldstone theorem has been featured even in modern textbooks~\cite{Sethna06}. In recent years, a good understanding of possibilities for having a reduced number of Goldstone modes has developed. First of all, it is possible that two broken symmetry generators excite the same Goldstone mode. This occurs when the vacuum expectation value of the commutator of (the densities of) these symmetry generators is non-vanishing~\cite{WatanabeBrauner11,WatanabeMurayama12,Hidaka13}. This is for instance the case in Heisenberg ferromagnets, where there is only one spin wave although there are two broken generators.
The current issue is however a different one. Recall that the Goldstone modes are infinitesimally small deviations from the preferred order parameter value (the Goldstone theorem is a statement about the limit of energy going to zero). Now, an infinitesimal rotation $\omega^{ab} = \partial_a u^b - \partial_b u^a$ can not locally be distinguished from an infinitesimal translation $u^a$~\cite{LowManohar02}. This simple observation can be extended into a quite general statement according to Ref.~\onlinecite{WatanabeMurayama13}. Namely, when the densities of broken symmetry generators are linearly dependent on each other, they do not excite independent Goldstone modes. In our case, the density of the generator of rotations around the $c$-axis $R^c(x)$ depends upon the generators of translations in the $a$-direction $T^a(x)$ in the Lie algebra of the Poincar\'e group via~\cite{WatanabeMurayama13,Kleinert08}:
\begin{equation}\label{eq:rotations depend on translations}
R^c (x) = \epsilon^{cba} x^b T^a(x).
\end{equation}
Roughly speaking, if one tries to excite a rotational Goldstone mode in a solid, one will instead excite (transverse) phonons.
In Ref.~\onlinecite{BeekmanWuCvetkovicZaanen13} we showed that in two dimensions indeed the rotational Goldstone mode emerges when translational symmetry is restored (the 2D nematic phase). This is completely consistent with the discussion above: as long as translational symmetry is broken, the relation Eq.~\eqref{eq:rotations depend on translations} prevents the rotational Goldstone mode to emerge as an independent excitation. But now in 3D an interesting scenario crops up. Clearly, if all translational symmetry is restored, three rotational Goldstone modes emerge as we have seen in Sec.~\ref{sec:nematic}. However, if translational symmetry is restored in two directions, according to Eq.~\eqref{eq:rotations depend on translations} one rotational degree of freedom is independent of any other broken symmetries. In this section we have confirmed that indeed one rotational Goldstone mode emerges, precisely in the plane where translational symmetry is restored.
\subsection{Generalizing nematic order: `isotropic' versus `cubic' nematics}\label{subsec:Generalizing nematic order}
Another issues is the form of the order parameter theory associated with liquid crystals in general. The reader should be familiar with the textbook cartoon, revolving around the kinetics of ``rod-like molecules''. In the isotropic fluid
these rods are both translationally and rotationally disordered with the rods pointing in all space directions. In the nematic phase these rods line up while they continue to be translationally disordered. Upon further lowering temperature
these rods may form liquid layers, that stack in a periodic array in the direction perpendicular to the layer: the smectic. At the lowest temperatures full crystalline order may set in. This cartoon is quite representative for much of
the classical liquid crystals; for deep reasons of chemistry, stiff, rod-like molecules are abundant and nearly all existent liquid crystals are of this `uniaxial kind'. However, viewed from a general symmetry breaking perspective these
uniaxial nematics are highly special and even pathological to a degree. Group theory teaches that the symmetry group describing the isotropy of Euclidean space $O(3)$ encompasses {\em all} three-dimensional point groups as its subgroups.
The uniaxial nematics are associated with the $D_{\infty \mathrm{h}}$ point group that is special in the regard that it only breaks the rotational isotropy in two of the three rotational planes of the $O(3)$ group. One ramification is
that it is characterized by only {\em two} rotational Goldstone modes. More generic 3D point groups break the isotropy in all three independent rotational planes and the Goldstone modes count in a way similar to the phonons
of the crystal: there are two `transverse' and one `longitudinal' acoustic modes associated with the rotational symmetry breaking, see Sec.~\ref{subsec:Rotational elasticity} and \ref{subsec:Torque stress in the quantum nematic}.
In the present duality setting we depart from the maximally symmetry breaking state: a crystal breaking both translations and rotations, characterized by one of the 230 space groups. By
proliferating the topological defects we {\em restore} the symmetry step-by-step. The principle governing the existing vestigial liquid-crystalline phases is that a priori, the topological defects associated with the restoration
of translational symmetry (the dislocations) can be sharply distinguished from those that govern the restoration of the isotropy of space -- the disclinations. Given the right microscopic circumstances, the disclinations can `stay massive' (not proliferating in the vacuum), while the dislocations have proliferated and condensed forming our dual `stress superconductor' with restored translational invariance and a liquid nature of the state of matter. Since these liquid crystals are `descendants' of the crystal, they are characterized by the `leftover' point group symmetry of the crystal. Point groups that are not compatible with the crystalline breaking of translations (encapsulated by the space groups) involving e.g. 5-fold rotations are therefore excluded.
It is now merely a matter of technical convenience to begin with the {\em most} symmetric space groups. In fact, to avoid as much as possible the details coming from crystalline anisotropies that
just obscure the essence we will look at from the simplest possible solid: the one described by the theory of isotropic elasticity in three space dimensions. This is similar in spirit to the famous
KTNHY theory of topological melting in 2D, which considers the special case of a {\em triangular} lattice, which is unique in the regard that its long-wavelength theory is precisely isotropic elasticity in two dimensions.
Upon proliferating the dislocations a nematic-type liquid crystal is formed that was named the ``hexatic'' since it is characterized by the six-fold rotational symmetry ($C_6$ point group) descending from the crystal. For the long-wavelength properties the precise form of the remnant discrete rotational symmetry is insignificant, the only thing that matters is that there is rotational rigidity. This is the reason we group all these states under the umbrella ``nematics'' (see also below).
In 3D there is no
space group that is described precisely by isotropic elasticity, characterized by merely a bulk (compression) and a shear modulus. This of course has influence on the descendant liquid crystals. The `rotational elasticity' theory of generalized nematics (characterized by any 3D point group) has been systematically enumerated~\cite{StallingaVertogen94} and it follows that even the most symmetric point groups such as the $O_\mathrm{h}$-group describing `cube-like' nematics (instead of the `rod-like' uniaxial ones) are characterized by three independent moduli. As we will see, departing from the isotropic solid there is only room for a single rotational modulus. Accordingly, the reader
should appreciate our `isotropic nematics' as being like a `cubic nematic' where we have switched off the moduli encoding for the cubic anisotropies by hand.
In fact, inspired by the considerations in the previous paragraph some of the authors felt a need to understand better the order parameter theory of such `generalized' (beyond uniaxial) nematics~\cite{LiuEtAl16b,LiuEtAl16,NissinenEtAl16}. They found out
that a systematic classification is just missing in the soft-matter literature, actually for a good reason. As it turns out, one is dealing with quite complex tensor order parameters involving tensors up to rank 6 for the most
symmetric point groups! It was subsequently found that discrete, non-Abelian gauge theory can be mobilized to compute both the explicit order parameters as well as the generic statistical physics associated with this
symmetry breaking in a relatively straightforward way. With regard to the latter, it was found that in case of the most symmetric point groups one runs into thermal fluctuation effects of an unprecedented magnitude~\cite{LiuEtAl16b}.
In the present context we just ignore these complications. We are primarily interested in the infinitesimal fluctuations around the ordered states and these are not sensitive to the intricacies of the `big-tensor' order parameters. In fact, all one needs to know is that our isotropic nematic is breaking rotations much like a cubic nematic, with the ramification that it should be characterized by two transverse and one longitudinal rotational Goldstone boson, see Sec.~\ref{subsec:Rotational elasticity} and \ref{subsec:Torque stress in the quantum nematic}.
\subsection{Quantum smectics: neither crystals nor superfluids}
In the vestigial order hierarchy the next state one meets is the smectic type (translational order in $D-1$ dimensions), sandwiched in between the crystal and the nematic type states. Yet again the textbook version is, from the viewpoint of general symmetry principles, of a very special kind. It is entirely focused on the `rod-like' $D_{\infty \mathrm{h}}$-molecules that now first arrange in liquid two-dimensional layers, which in turn stack in an array periodic perpendicular to these layers, breaking translations in this direction. Even more so than for the nematics a truly general `effective field theory description' departing from tight symmetry principles is lacking. This deficit becomes on the foreground especially when dealing with the zero-temperature {\em quantum} smectic states of matter. The `liquid nature' becomes now associated with superfluidity, and there should be a well-defined sector of long-wavelength Goldstone-type excitations. Are these like phonons resp. superfluid phase modes (second sound) depending on whether one looks along the `solid' resp. `liquid' directions? We shall see that these characteristics do shimmer through, but this is only a small part of the story. We found in the 2+1D case a remarkably complex assortment of collective modes reflecting the truly intertwined nature of superfluid and elastic responses~\cite{CvetkovicZaanen06b,QLC2D}. In part, this is already understood in the soft-matter literature in the form of the {\em undulation mode}: the transverse mode propagating in the liquid direction
acquires a {\em quadratic dispersion} since the lowest-order interactions between the liquid layers are associated with their curvature~\cite{DeGennesProst95,ChaikinLubensky00}. These are impeccably reproduced in our smectics seen as dual stress superconductors
of a particular kind. Yet again, in 3+1D there is even more to explore than in 2+1D; much of the sections on {\em quantum smectic} (\ref{sec:smectic}) and {\em columnar} (\ref{sec:columnar}) order are dedicated to charting this rich landscape.
Although a Landau-style `direct' order parameter theory is lacking for ``generalized (quantum) smectics'' (i.e. going beyond $D_{\infty \mathrm{h}}$), the topological principles beyond our weak--strong
duality are sufficiently powerful to formulate such a theory in the dual language of stress superconductivity. Like for the nematics, the main limitation is that we have formulated this theory departing from isotropic elasticity. The effects of the anisotropies associated with the real 3D space groups are presently unexplored and may be taken up as an open challenge.
It was realized in the classic literature on thermal topological melting that smectic-type order is actually a natural part of this agenda~\cite{OstlundHalperin81}. It appears that is was first addressed in the quantum context
independently in the early work by us~\cite{ZaanenNussinovMukhin04}, and by Bais \& Mathy who studied the possible liquid crystal phases with the fanciful Hopf symmetry breaking formalism~\cite{BaisMathy06,MathyBais07}. This works as follows:
as before, we depart from the crystal with a particular point group embedded in its space group. The dislocations are characterized by their topological charge: the Burgers vector.
These are associated with the deficient translations in the crystal lattice and accordingly they point only in lattice directions and are equivalent under the point-group transformations.
In a cubic lattice, for instance, Burgers vectors point in orthogonal spatial $x$- ,$y$- and $z$-directions, while in a hexagonal crystal these point in the $z$ direction or in are six equivalent
directions in the $xy$-plane associated with the sixfold axis, see Fig.~\ref{subfig:hexagonal melting crystal}.
\begin{figure*}
\hfill
\subfloat[hexagonal crystal]{\includegraphics[height=5cm]{hexagonal_crystal.pdf}\label{subfig:hexagonal melting crystal}}
\hfill
\subfloat[columnar]{\includegraphics[height=5cm]{hexagonal_columnar_vertical.pdf}\label{subfig:hexagonal melting columnar vertical}}
\hfill
\subfloat[smectic]{\includegraphics[height=5cm]{hexagonal_smectic_vertical.pdf}\label{subfig:hexagonal melting smectic vertical}}
\hfill\null
\\
\hfill
\subfloat[columnar]{\includegraphics[height=5cm]{hexagonal_columnar_horizontal.pdf}\label{subfig:hexagonal melting columnar horizontal}}
\hfill
\subfloat[smectic]{\includegraphics[height=5cm]{hexagonal_smectic_horizontal.pdf}\label{subfig:hexagonal melting smectic horizontal}}
\hfill
\subfloat[nematic]{\includegraphics[height=5cm]{hexagonal_nematic.pdf}\label{subfig:hexagonal melting nematic}}
\hfill\null
\caption{Sequential dislocation-mediated melting of a hexagonal crystal with $D_{6\mathrm{h}}$ point group. Grey lines are bonds in the original hexagonal crystal and are guides to the eye only in the other phases. Green dashed lines indicate the elementary Burgers vectors of dislocations. Red color indicates lines/planes/volumes with translational symmetry due to condensation of dislocations with Burgers vectors in blue dotted lines. The black arrows are the `rotational cross' indicating the broken rotational symmetry that is the same throughout all the phases. \protect\subref{subfig:hexagonal melting crystal} All symmetry is broken, and the elementary Burgers vectors can point in six in-plane and two out-of-plane directions. \protect\subref{subfig:hexagonal melting columnar vertical} Dislocations with Burgers vectors in the vertical direction (blue) condense and restore translational symmetry, resulting in a 2D array of liquid lines. Since the remaining translational order is orthogonal to the liquid directions, the remaining Burgers vectors (green) match the original in-plane Burgers vectors of the crystal. \protect\subref{subfig:hexagonal melting smectic vertical} If we furthermore melt along an in-plane direction, the result is a periodic stack of liquid planes: the quantum smectic. Note that the Burgers vectors (green) in this smectic do no longer point along the original crystal axes. Because points that are separated by vectors along the liquid directions (vectors in blue) are equivalent, the remaining Burgers vectors must be orthogonal to the liquid planes.
\protect\subref{subfig:hexagonal melting columnar horizontal} Alternatively, translational symmetry restoration can take place in the 6-fold plane. In-plane Burgers vectors in this columnar phase are not parallel to those of the parent crystal. \protect\subref{subfig:hexagonal melting smectic vertical} Melting all in-plane translational order leads to stacks of liquid planes with $C_6$ in-plane rotational order, i.e. a stack of hexatics.
\protect\subref{subfig:hexagonal melting nematic} Proliferation of all dislocations restores translational symmetry completely. Rotational symmetry remains broken because disclinations are forbidden. The rotational order is remembered from the original crystal point group (black cross).
}\label{fig:hexagonal melting}
\end{figure*}
The master principle governing both the smectic- and nematic-type vestigial phases is that the dislocations are allowed to proliferate while keeping the disclinations ``out of the vacuum''. The point group symmetry of the crystal is maintained while translational symmetry is restored. But we just learned that there is quite a variety of Burgers vectors; how should these be arranged
in the dislocation condensate? This is governed by precise topological rules. The first rule is that {\em the Burgers vectors of dislocations have to be `locally antiparallel'.} A disclination is topologically
identical to a macroscopic number of dislocations with parallel Burgers vectors~\cite{Kleinert89b,ZaanenNussinovMukhin04,QLC2D}. These are not allowed in the vacuum and therefore we have to insist that on the microscopic scale a dislocation
with Burgers vector pointing in the $\vec{B}$-direction of the lattice is always accompanied one pointing in precisely the opposite $-\vec{B}$-direction. The second rule is that {\em the translational symmetry gets restored precisely in the direction of the Burgers vectors}. In other words, points that differ by a (not necessarily integer) multiple of the Burgers vector become equivalent.
In the generalized nematic, translation symmetry is restored in {\em all} spatial directions and this implies
that all Burgers vector directions are populated equally in the dislocation condensate of the dual stress superconductor. One notices that this dislocation condensate remembers the point group of the
crystal through the requirement that it is formed out of ``Burgers vectors pointing in the allowed directions''. In fact, as we will discuss in more detail in section~\ref{subsec:Torque stress in the quantum nematic} the rotational elasticity of the nematic is carried by the dislocation condensate itself.
However, this ``equal Burgers vector population'' need not to be the case: it is perfectly compatible with the topological rules to populate only the pair-antiparallel Burgers vectors in e.g. one particular direction. Accordingly, translational symmetry is restored in that one of the three space dimensions and this
is the topological description of the {\em columnar} state, Fig~\ref{subfig:columnar}. In a next step, the condensate can pick Burgers vectors such that the translational symmetry is restored in two orthogonal
space directions, leaving the third axis unaffected: this is the {\em smectic} state in three dimensions, Fig.~\ref{subfig:smectic}. One notices a peculiar tension between the point group of the crystal and the way that the `liquid directions' emerge. Translational symmetry can only be independently restored in the three orthogonal ($x,y,z$) spatial directions since points that differ by a vector in a liquid direction are equivalent.
Accordingly, the `liquid' can occur either in one direction (the columnar phase), one plane (the smectic) or in all three directions (the nematic). In a cubic crystal this is straightforward; the three cubic axes are coincident with the three orthogonal `translational' directions and by proliferating dislocations in either one or two directions one obtains immediately the smectic and columnar phases shown in the cartoons Fig.~\ref{fig:phases}.
However, dealing with e.g. a hexagonal crystal this gets more confusing, see Fig.~\ref{fig:hexagonal melting}. The first melting transition to a columnar phase takes pairs of antiparallel Burgers vectors along one of the crystal axes. For instance, one can choose the direction perpendicular to the 6-fold plane, Fig.~\ref{subfig:hexagonal melting columnar vertical}. The result is a regular triangular array of liquid lines. The dislocations in this columnar phase are still along the original crystal axes. If dislocation condensation takes place with Burgers vectors in a second direction, a smectic is obtained, Fig.~\ref{subfig:hexagonal melting smectic vertical}. This is a periodic stack of liquid planes. Note that the periodicity is no longer along an axis of the parent crystal, but obviously perpendicular to the planes. Accordingly, the dislocations in this smectic have Burgers vectors in this perpendicular direction, not commensurate with the Burgers vectors of the parent crystal. Here we see the two important consequences of the rules mentioned above:
1) dislocation melting always takes place restoring translations symmetry in orthogonal directions, even though the elementary Burgers vectors of the parent crystal may not be orthogonal.
2) the remnant rotational order is {\em independent} of the translational symmetry restoration and is completely inherited from the parent crystal.
This can be clearly seen in e.g. Fig.~\ref{subfig:hexagonal melting smectic vertical}.
Conversely, we could first melting along an in-plane direction as in Fig.~\ref{subfig:hexagonal melting columnar horizontal}. Now we have `three kinds of physics': liquid-like in one in-plane direction, solid-like in the orthogonal in-plane direction, and solid-like in the out-of-plane direction, which was already inequivalent due to the original crystal anisotropy. Again the Burgers vectors have to be perpendicular to the liquid lines, not necessarily parallel to the original crystal axes. If the next melting step is again in-plane, we end up with a periodic stack of liquid layers, see Fig.~\ref{subfig:hexagonal melting smectic horizontal}. Each layer is like a 2D hexatic phase with $C_6$-symmetry in the plane. We will verify this explicitly in Sec.~\ref{sec:smectic}. The overall structure is a particular 3D smectic.
In all cases, once all translational symmetry has been restored due to melting of dislocations with Burgers vectors in all direction, a generalized nematic is obtained, see Fig.~\ref{subfig:hexagonal melting nematic}. The rotational order is the point group of the parent crystal, $D_{6\mathrm{h}}$ in this case.
It takes some special care to precisely formulate the equations describing this `Burgers vector population' affair in the construction of the dual dislocation condensates. For the 2+1D quantum liquid crystals
this was for the first time put in correct form in QLC2D --- although the Higgs terms in the effective dual actions were correct in earlier work, the derivation was flawed. As it turns out, this procedure straightforwardly generalizes to the 3+1D case, which is the topic of Sec.~\ref{sec:Dislocation-mediated quantum melting}. In the sections dealing with the smectic (\ref{sec:smectic}) and columnar (\ref{sec:columnar}) phases we will expose the remarkably rich landscape of `intertwined' liquid--solid responses of these systems. Once again, given our specialization to the strictly isotropic case this description is far from complete and we leave it to future work to find out a complete inventory of the long-wavelength physics that follows from this peculiar interplay of partial translational and full rotational symmetry breaking.
|
\section{Introduction}
\label{sec:intro}
Memento TimeMaps serve as an index for the mementos for an original resource (\mbox{URI-R}) contained in an archive \cite{rfc7089}. Web archives return TimeMaps with a list of identifiers (URI-Ms) for the HTTP transactions observed at archival time. TimeMaps have generally been used as a count of the number of representations of a \mbox{URI-R} present in an archive. But, TimeMaps may include \mbox{URI-Ms} for archived representations (HTTP 2XX), archived redirections (HTTP 3XX), and archived errors (HTTP 4XX or 5XX) \cite{rfc7231}. Only the URI-Ms that result in an HTTP 2XX when dereferenced match the general notion of a ``capture'' of the contents of a webpage. But, the status that results when a \mbox{URI-M} is dereferenced (to the extent of returning an archived entity body) is not present in a TimeMap. Further, TimeMaps do not explicitly return a ``count'' value to indicate the number of mementos listed in the TimeMap that produce a non-redirecting (non-3XX) HTTP status code when dereferenced. This can cause problems when using the number of \mbox{URI-Ms} in a TimeMap as a proxy for the number of captures of a Web page.
Various tools \cite{memgator, jordan-jcdl2015, kelly-dl2014-mink} and access points into the Web archives return a different count of the number of captures for a \mbox{URI-R} depending on the heuristic implemented and the source of the archival listings. For example, the Internet Archive's web interface when queried with the URI-R example.com states, ``Saved 11,771 times between January 20, 2002 and May 20, 2016''. The file returned from Internet Archive's CDX endpoint returns 69,162 entries. The TimeMap from Internet Archive for example.com contains 40,641 URI-Ms with a rel value of ``memento''. The heuristic of determining how many captures are represented by URI-Ms in a TimeMap cannot be completed without dereferencing.
Researchers can use the inline metadata about the \mbox{URI-M}, without the need to dereference the \mbox{URI-M} in a TimeMap, including its temporal ordering, datetime (through the datetime HTTP Link attribute \cite{rfc7089}), etc. to infer characteristics about a dereferenced memento. However, dereferencing some URI-Ms in a TimeMap produces an HTTP redirect \cite{rfc7231} that instructs the client to access a \mbox{URI-M} with a different datetime, to obtain the requested content. For example, a TimeMap for \url{http://vimeo.com} from Internet Archive contained 199,262 URI-Ms with an associated ``rel'' value of ``memento''. However, when a user accesses over 57\% of these \mbox{URI-Ms}, an HTTP Redirect is returned pointing to another memento whose \mbox{URI-M} is in the TimeMap that returns a HTTP Status OK. A different extreme of memento count results when a user accesses the TimeMap from \url{http://odu.edu}, whose percentage of redirects is around 9.7\% of the URI-Ms listed.
Redirection in a Web archive can be attributed to a variety of canonicalization rules including a scheme change (e.g., http to https), an obsolete subdomain (e.g., www2 to www), a slash added to a URI (\hyperref[]{http://foo.com/\textasciitilde joe} to \hyperref[]{http://foo.com/\textasciitilde joe/}), among others. Preserving and replaying these redirects allows an archive to accurately reproduce the HTTP transactions that would have occurred when the URI being accessed resided on the live Web.
When a \mbox{URI-M} in a TimeMap is dereferenced, it may redirect to another \mbox{URI-M} listed in the TimeMap. Because of this, the heuristic of counting URI-Ms with relation values of ``memento'' is an inaccurate means of determining the number of unique representations inferred from a TimeMap. We further emphasize the distinction per the Memento specification that the identifiers for mementos (URI-Ms) in a TimeMap are identifiers for archived HTTP transactions (e.g., transmission of HTTP 2XX, 3XX, 4XX, etc.) rather than identifiers for representations.
Based on the number of URI-Ms in a TimeMap not necessarily resolving to unique mementos when archival redirects are followed, we examined the mementos from contemporarily large TimeMaps to evaluate the patterns and schemes used in Memento canonicalization. Through this, we identify the difference between the number of mementos available as reported by the TimeMap through naive ``rel'' counting heuristics to the temporally unique mementos identified once these mementos are dereferenced.
\section{Background}
\label{sec:bg}
This section includes background information and an overview of the state-of-the-art of archival technologies relevant to this work including Memento aggregation, URI canonicalization, archival indexing formats, and \mbox{URI-R} opacity.
\subsection{Memento Aggregation}
\begin{figure}
\begin{lstlisting}
<http://example.com>; rel="original",
<http://web.archive.org/web/20020120142510/http://example.com/>; rel="first ^memento^"; datetime="Sun, 20 Jan 2002 14:25:10 GMT",
<http://web.archive.org/web/20020804094019/http://www.example.com/>; rel="^memento^"; datetime="Sun, 04 Aug 2002 09:40:19 GMT",
<http://web.archive.org/web/20160728014649/http://www.example.com/>; rel="^memento^"; datetime="Thu, 28 Jul 2016 01:46:49 GMT",
<http://web.archive.org/web/20160728114745/http://www.example.com>; rel="^memento^"; datetime="Thu, 28 Jul 2016 11:47:45 GMT",
<http://web.archive.org/web/20160728123024/http://example.com/>; rel="last ^memento^"; datetime="Thu, 28 Jul 2016 12:30:24 GMT",
<http://localhost:1208/timemap/link/http://example.com>; anchor="http://example.com"; rel="timemap"; type="application/link-format",
<http://localhost:1208/timegate/http://example.com>; anchor="http://example.com"; rel="timegate"
\end{lstlisting}
\caption{A partial Link formatted TimeMap from a local instance of MemGator. Highlighted rel values constitute inclusion in the sum described in Equations~\ref{eq:naive} and \ref{eq:naive2}.}
\label{fig:tm_link}
\end{figure}
The Memento Framework \cite{rfc7089} allows navigation of Web archives in the dimension of time using content from Web archives and resource versioning systems. A Memento TimeMap (Figure~\ref{fig:tm_link}) is a structured list of identifiers (URI-Ms) for archived captures (mementos) returned from an archive when queried with a \mbox{URI-R} as the parameter. A TimeMap may also contain references to other TimeMaps and TimeGates.
A Memento aggregator is a software implementation of Memento that takes a URI as the parameter, queries multiple supported archives, combines and temporally orders the returned mementos, and returns this list as a TimeMap to the user. We used the MemGator \cite{memgator} implementation of a Memento aggregator in collecting data for our analysis. MemGator provides its own heuristic for determining and reporting the number of mementos present in an aggregated TimeMap using the non-standard \texttt{X-Memento-Count} HTTP header. Despite this, we used the contents of the aggregated TimeMap returned from MemGator instead of this header as the basis for further investigating the number of mementos present.
\subsection{URI Canonicalization}
\label{sec:canonicalization}
URI canonicalization associates differently formatted URIs \cite{rfc6596}. For example, \hyperref[]{http://example.com} might be associated with:
\begin{itemize}
\item \hyperref[]{http://www.example.com}
\item \hyperref[]{https://www.example.com}
\item \hyperref[]{http://example.com/}
\item \hyperref[]{http://example.com/index.html}
\item \href{http://example.com/\string#articles}{http://example.com/\#articles}
\end{itemize}
Canonicalization allows after-the-fact clustering of URIs that likely reference the same resource. As URI schemes from a Web site change over time, canonicalization is critical for retaining a cohesive, comprehensive listing of the mementos available for a Web page. Internet Archive's Wayback CDX Server API\footnote{\hyperref[]{https://github.com/internetarchive/wayback/tree/master/wayback-cdx-server}} and endpoint\footnote{Example access at \hyperref[]{http://web.archive.org/cdx/search/cdx?url=example.com}} is one of multiple endpoints\hreffootnote{https://archive.org/help/wayback_api.php}{https://archive.org/help/wayback_api.php} that provides access to the indexes of the archive's holdings. A partial example (that corresponds with the TimeMap shown in Figure~\ref{fig:tm_link}) of the data returned from Internet Archive's CDX server is shown in Figure~\ref{fig:cdx}. As an alternative to their Memento endpoint \cite{wsdl-ia-memento}, the CDX endpoint provides the HTTP status code of the capture as well as Sort-friendly URL Reordering Transformed (SURT) URIs. Part of the SURT-generation process involves canonicalizing the \mbox{URI-R}. A canonicalized URI is present in the first space-delimited field (Figure~\ref{fig:cdx}) where the ``www'' subdomain is not present despite being part of the URI in the query parameter. The non-canonicalized \mbox{URI-R} attributed to the record in the CDX is available as the third field in the CDX record. Figure~\ref{fig:cdx} shows the \mbox{URI-R} variations including no subdomain, the ``www'' subdomain, with and without a trailing slash, and the explicit inclusion of the port number as all canonicalizing to the same URI in the CDX.
\begin{figure}
\begin{lstlisting}
com,example)/ 20020120142510 http://example.com:80/ text/html 200 HT2DYGA5UKZCPBSFVCV3JOBXGW2G5UUA 1792
com,example)/ 20020804094019 http://www.example.com:80/ text/html 200 UY3I2DT2AMWAY6DECFCFYMT5ZOTFHUCH 457
com,example)/ 20160728014649 http://www.example.com/ unk 302 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ 339
com,example)/ 20160728114745 http://www.example.com unk 302 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ 340
com,example)/ 20160728123024 http://example.com/ text/html 200 ASIFPQKKLDWATFDIO1OJJ3NSK34KLLMN 577
\end{lstlisting}
\caption{A CDX response returned from Internet Archive's CDX Server. The space-delimited fields are representative of the canonicalized (SURTed) URI, datetime, original URI, MIME type of original document (where applicable), HTTP response code, digest of WARC response record, and length of response record.}
\label{fig:cdx}
\end{figure}
\subsection{Archival Indexing}
A CDX record with a 3XX HTTP status code does not contain the ultimate \mbox{URI-M} that the user will experience when the URI-M is dereferenced. Further, the CDX in Figure~\ref{fig:cdx} is only representative of IA's holdings. The corresponding service of aggregating CDX records in an entity like Memento's concept of combining of TimeMaps from different archives through a Memento aggregator does not exist in standard practice for CDX files. Archives that provide a Memento endpoint are not required and frequently do not expose a CDX endpoint like Internet Archive. This prevents simply referencing all aggregated archives' CDX files for a URI to determine the non-redirecting count of mementos. In this work, we utilize the aggregated holdings of multiple Web archives as well as the CDXJ format \cite{salam-cdxj}, an extension of CDX. MemGator's CDXJ generation is derived from the archives' Memento endpoints, specifically their Link formatted TimeMaps, and transformed into the CDXJ format that allows quicker, more reliable parsing of the datetime that the included URI-Ms represent.
\subsection{Opacity of \mbox{URI-Ms}}
\label{sec:opacity}
\begin{sloppypar}
It is tempting to extract \mbox{URI-R} and Memento-Dateime values directly from \mbox{URI-Ms}. For example, it is likely that \url{http://web.archive.org/web/20140417054441/http://google.com/} is a memento for \url{http://google.com} appeared at April 17, 2014 at 05:44:41 GMT. However, we cannot be sure until we dereference the URI-M and check its response headers for the values in \texttt{rel="original"} and Memento-Datetime. While it is unlikely that the IA will deceive us, the \mbox{URI-M} may redirect to another URI-M with a different Memento-Datetime, or in the case of an archived HTTP redirection, the URI-M might end up at an altogether different \mbox{URI-R} \cite{alsum-redirects}. Furthermore, some archives issue \mbox{URI-Ms} without semantics --- for example these URI-Ms are all mementos for \url{google.com} but neither this nor the Memento-Datetime can be ascertained without dereferencing: \url{webcitation.org/query?id=1398456230796350}, \url{archive.is/sz8b9}, and \url{perma.cc/H3YY-BQN5}. For these reasons, we treat \mbox{URI-Ms} as fully opaque \cite{webArchitecture-opacity} and dereference all URI-Ms to extract values for \mbox{URI-R} and Memento-Datetime.
\end{sloppypar}
\section{Related Work}
Bar-Yossef et al. \cite{bar-yossef} introduced the term, ``Soft 404s'' to identify Web pages that report a status code other than HTTP 404 despite the page not existing. Meneses et al. \cite{soft404s} described the process of identifying ``Soft 404s'' based on a signature of the page's contents. In this work we describe ``soft 3XXs'' where content is returned from an archive with a status code of 200 yet the contents of the capture consist of an archived HTTP 3XX redirect. With archives that implement Memento, the Accept-Datetime header instructs the archive to return the originally archived status code (Section~\ref{sec:redirects}).
AlSum et al. \cite{alsum-redirects} analyzed memento redirection patterns relating to HTTP redirects to supply the user with the correct memento when a redirect is encountered in the archives. They introduced the notion of ``URI stability'' to give a quantitative measure of the presence of HTTP 3XX status codes that result when URI-Ms in TimeMaps are dereferenced.
Rosenthal has discussed Memento aggregator merits and downsides. He cautioned against using TimeMap magnitude for determining the number of URI-Ms available, stating, ``Even if we assume that an archive is correct in announcing that it contains a valid copy of the resource at a particular URL at a particular time, that does not imply that it is willing to satisfy a browser's request for that copy.'' \cite{rosenthal-mementoAggregation}.
\begin{table}
\centering
\begin{tabular}{ l | r }
Archive & Memento count\\\hline
Internet Archive & 636,246 \\
Archive-It & 62,828 \\
Webcitation & 7,551 \\
Stanford Web Archive & 4,734 \\
UK National Archives Web Archive & 1,510 \\
Archive.is & 1,301 \\
PRONI Web Archive & 173 \\
UK Parliament Web Archive & 127 \\
\hline
Total & 714,470\\
\end{tabular}
\caption{Distribution of mementos for google.com for a collection of archives defined by a locally deployed Memento aggregator.}
\label{tab:google}
\end{table}
\section{Data Collection}
\label{sec:dataCollection}
To analyze the degree to which archival identifiers result in redirects, we needed to acquire the HTTP response headers for all URI-Ms accumulated from multiple Web archives for a URI-R. The concept of a Memento aggregator allows us to accomplish this task, albeit parsing the standardized Link formatted resulting TimeMaps is potentially prone to error.
We deployed a local instance of MemGator\footnote{\hyperref[]{http://github.com/oduwsdl/memgator}} version 1.0-RC4 configured to query the archives listed in Table~\ref{tab:google}. The MemGator instance was initialized with 25 minutes as the value for the ``restimeout'' (response timeout for each archive) and ``hdrtimeout'' (header timeout for each archive) parameters. Declaring these timeout values ensured that the server portion of data collection was not prematurely returned because of network latency with communication to the considered archives. Our client script that queried the local MemGator instance was also setup to access this instance with equally large timeout values.
We leveraged MemGator's CDXJ \cite{salam-cdxj} interface (example output in Figure~\ref{fig:tm_cdxj}) for simple datetime extraction, structured JSON-formatted metadata of each memento's attributes, and more human readable output compared to the conventional Link (Figure~\ref{fig:tm_link}) or JSON formatted TimeMaps. Collection was run on a late 2013 MacBook Pro running OS X version 10.11.4 with a 2.4 GHz Intel i5 processor, 8 GB of RAM, and a 250 GB SSD disk. Data was collected mid-May, 2016. We performed an initial analysis of the mementos contained within the TimeMap without dereferencing any mementos. The client script to query the local MemGator was created in Python 2.7.10 using the ``requests'' library\footnote{\hyperref[]{http://python-requests.org}} and the built-in JSON parser.
\begin{figure}
\begin{lstlisting}[basicstyle=\footnotesize]
@meta {"original_uri": "http://example.com"}
@meta {"timegate_uri": "http://localhost:1208/timegate/http://example.com"}
@meta {"timemap_uri": {
"link_format": "http://localhost:1208/timemap/link/http://example.com",
"json_format": "http://localhost:1208/timemap/json/http://example.com",
"cdxj_format": "http://localhost:1208/timemap/cdxj/http://example.com"
}
...
20090418233448 {"uri": "http://web.archive.org/web/20090418233448/http://www.example.com/", "rel": "memento", "datetime": "Sat, 18 Apr 2009 23:34:48 GMT"}
20090421223547 {"uri": "http://wayback.vefsafn.is/wayback/20090421223547/http://www.example.com/", "rel": "memento", "datetime": "Tue, 21 Apr 2009 22:35:47 GMT"}
20090421231335 {"uri": "http://webarchive.loc.gov/all/20090421231335/http://www.example.com/", "rel": "memento", "datetime": "Tue, 21 Apr 2009 23:13:35 GMT"}
...
\end{lstlisting}
\caption{A partial CDXJ formatted TimeMap returned from a local instance of MemGator containing URI-Ms from multiple archives.}
\label{fig:tm_cdxj}
\end{figure}
\section{Analysis based on TimeMaps}
\label{sec:initialAnalysis}
We obtained a TimeMap for google.com from our locally deployed Memento aggregator (MemGator instance) containing 714,470 URI-Ms from 8 different Memento-compliant archives. Table~\ref{tab:google} shows the distribution of the mementos using a simple URI-based association algorithm to attribute URI-Ms to an archive. 89.1\% of the URI-Ms returned were from Internet Archive.
\subsection{Variation in Scheme}
Two schemes \cite{rfc7230} are used for the URI-Rs contained within the URI-Ms returned: HTTP and HTTPS. Table~\ref{tab:scheme} shows the breakdown of the URI-Rs contained in the URI-Ms based on the TimeMap, inclusive of the inferred \mbox{URI-Rs} whose scheme could not be determine solely from the \mbox{URI-M}. As discussed in Section~\ref{sec:opacity}, the more accurate method to attribute a \mbox{URI-R} to a \mbox{URI-M} is to obtain the memento's ``original'' Link header values but this section focuses on analyzing the contents of the TimeMap without requesting the \mbox{URI-Ms}.
From the URI-Rs that could be extracted, 86.2\% used the HTTP scheme.
Table \ref{tab:google2} shows the \mbox{URI-R}-based memento count for each scheme using a substring-based grouping approach similar to that used for Table~\ref{tab:google}. The TimeMap contained canonicalized variants of the URI-Rs embedded as substrings within the URI-Ms with our query for the CDXJ-formatted TimeMap. Our query supplied the \mbox{URI-R} variant containing the HTTP scheme, www sub-domain, and no trailing characters combination (i.e., \hyperref[]{http://www.google.com}) via the query to MemGator\footnote{\hyperref[]{http://localhost:1208/timemap/cdxj/http://www.google.com}}.
\begin{table}
\centering
\begin{tabular}{ r r }
Scheme & URI-R count \\\hline
http & 609,274 \\
https & 97,645 \\
unknown & 7,551 \\ \hline
& 714,470
\end{tabular}
\caption{Scheme distribution among the URI-Rs within the mementos in the TimeMap for google.com.}
\label{tab:scheme}
\end{table}
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{allArchives.png}
\caption{The average time between consecutive mementos has decreased with time. This plot shows a year-based bucketing of the difference in time between adjacent mementos from different archives for google.com. }
\label{fig:avgTime}
\end{figure}
\subsection{Grouping by Year}
We separated the URI-Ms as reported by the TimeMap into year-based buckets using the ``datetime'' attribute for each \mbox{URI-M} (and not the embedded 14-digit date stamp per Section~\ref{sec:opacity}) as well as by-archive for google.com. We calculated the average time between URI-Ms within a year-based bucket to show that the velocity of capturing google.com is generally increasing in time (Figure~\ref{fig:avgTime}). The quantity of captures from Internet Archive for google.com in 2015 was significantly lower than the trend would indicate. Table~\ref{tab:tcOverTM} indicates this dramatic drop in google.com captures from both the IA CDX endpoint and from the Memento endpoint. Also because data collection occurred in May 2016, the partial year data points for 2016 are on-par with the trend of years prior to 2015.
\begin{table}
\centering
\begin{tabular}{ r l r }
Scheme & Format & \mbox{URI-R} count \\ \hline
http & http://www.google.com & 541,160 \\
& http://google.com & 67,811 \\
& http://\textit{other}.google.com & 303
\\ \hline
https & https://www.google.com & 96,853 \\
& https://google.com & 792 \\
& https://\textit{other}.google.com & 0
\\ \hline
\end{tabular}
\caption{Count of URI-Rs contained within URI-Ms for google.com.}
\label{tab:google2}
\end{table}
\begin{table}
\centering
\begin{tabular}{ r r r r }
year & M$_{TM}$ & M$_{RC}$ & $DI$\\\hline
1998 & 4 & 4 & $\infty$ \\
1999 & 19 & 19 & $\infty$ \\
2000 & 132 & 87 & 1.933 \\
2001 & 1,185 & 579 & 0.955 \\
2002 & 176 & 137 & 3.513 \\
2003 & 75 & 55 & 2.750 \\
2004 & 197 & 143 & 2.648 \\
2005 & 1,236 & 414 & 0.504 \\
2006 & 735 & 483 & 1.917 \\
2007 & 1,055 & 842 & 3.953 \\
2008 & 1,376 & 894 & 1.855 \\
2009 & 6,074 & 4,335 & 2.493 \\
2010 & 9,326 & 6,530 & 2.335 \\
2011 & 20,634 & 9,279 & 0.817 \\
2012 & 102,533 & 16,240 & 0.188 \\
2013 & 228,405 & 25,203 & 0.124 \\
2014 & 164,865 & 22,738 & 0.160 \\
2015 & 17,978 & 11,286 & 1.686 \\
2016 & 139,520 & 5,805 & 0.043 \\
\hline
\end{tabular}
\caption{Google over time, bucketed by year, based on IA mementos extracted from the MemGator CDXJ TimeMap. M$_{TM}$ is the memento count based solely on the data in the TimeMap, M$_{RC}$ is the count based on exclusion of redirects when dereferenced, and $DI$ is the ratio of non-redirecting mementos to redirecting mementos, per Section~\ref{sec:canonicalizationPatterns}.}.
\label{tab:tcOverTM}
\end{table}
\subsection{``TimeMap'' from CDX Server}
We also obtained the CDX for google.com from Internet Archive (IA). We compared the HTTP response codes we received when dereferencing the IA URI-Ms from the CDXJ TimeMap (Section~\ref{sec:dataCollection}) with the response codes explicitly provided in the CDX listing IA returned. Per Section~\ref{sec:canonicalization}, a CDX endpoint is not available as a user-accessible endpoint from most Web archives. We used the available endpoint at IA as a sanity check for correctness of the data obtained when the URI-Ms in a TimeMap are dereferenced. The intention of analyzing the TimeMap and not simply deferring to the CDX Server, despite the majority of mementos in the aggregated TimeMap being from IA, is to extrapolate the dereferencing strategy to other Memento-compliant Web archives.
\section{Analysis based on Mementos}
\label{sec:mementoBasedAnalysis}
In Section~\ref{sec:initialAnalysis} we analyzed the archival presence of a URI based solely on the TimeMap supplied by a local aggregator when querying the aggregator with one canonicalized variant of the URI-R. In this section, we dereference the URI-Ms in the TimeMap for further analysis. Equations~\ref{eq:naive} and \ref{eq:naive2} set the basis for counting mementos in a TimeMap by merely counting the entries where the ``rel'' attribute in a TimeMap contains a value of ``memento''. For example, Figure~\ref{fig:tm_link} shows a Link formatted TimeMap where each highlighted entry containing a ``memento'' rel value (with the potential additional inclusion of other values like ``last'' and ``first'') increments the count according to Equation~\ref{eq:naive2}.
\begin{equation}\label{eq:naive}
M = \text{URI-M} \subset \text{TimeMap}, \text{if}\ ``memento" \subset \text{valuesOf(}rel\text{)}\\
\end{equation}
\begin{equation}\label{eq:naive2}
|TM|_{rel} = \sum_{m=1}^{len(M)} 1
\end{equation}
\subsection{Redirects in Mementos}
\label{sec:redirects}
The number of non-redirecting (non-3XX) mementos in a TimeMap cannot be counted with the TimeMap data alone. When URI-Ms are dereferenced, they need not contain an entity body but may consist only of an archived HTTP response, which might not be a 200. This occurs in cases where the live Web site returned an HTTP 302 redirect, among other circumstances. This redirect was captured, retained, and is replayed by the archives. When replayed, the datetime originally requested for a URI-R, as is often the case, will be different than the datetime of the memento ultimately served to the user. This ``archived 302'' is different from a 3XX code returned from an archive that is not representative of an archival capture; for example, when a datetime for a URI is requested where no capture for the URI is contained within the archive's holdings.
\begin{sloppypar}
Assuming the TimeMap in Figure~\ref{fig:tm_link} is wholly inclusive of all of the mementos contained by Internet Archive for example.com, requesting the \mbox{URI-M} \hyperref[]{http://web.archive.org/web/20160728114743/http://www.example.com} (two seconds before a listed \mbox{URI-M} neglecting datetime semantics per Section~\ref{sec:bg}) will result in a 302 from the archive pointing to the nearest capture. This behavior is a function of the archive, is not mandatory behavior to exhibit, and is functionality independent of the Memento protocol. Were there a capture at the former datetime where the archival crawler experienced a 302 from the live Web at the time, the TimeMap would contain the \mbox{URI-M} \hyperref[]{http://web.archive.org/web/20160728114743/http://www.example.com} with a rel value of ``memento'' indistinguishable from the memento entry at \hyperref[]{http://web.archive.org/web/20160728114743/http://www.example.com} regardless of the status code that occurs from the archive when each \mbox{URI-M} is dereferenced.
\end{sloppypar}
\subsection{Direct and Indirect Mementos}
\label{sec:userAgent}
Users interacting with an archive via a Web browser will not directly experience intermediary HTTP transactions (the user agent automatically redirects the user to a non-3XX status), we introduce the term URI-M$_{D}$ (for ``direct'') to indicate a \mbox{URI-M} in a TimeMap that does not require any intermediary transaction to resolve. Thus, a URI-M$_{D}$ is a case of a \mbox{URI-M} where the \mbox{URI-M} originally requested by the user is the identifier for the ultimate memento served. URI-Ms in a TimeMap that exhibit the behavior where a HTTP 3XX class code is replayed and the datetime differs from that requested by the user are indicated with URI-M$_{I}$ (for ``indirect'').
In Equation~\ref{eq:rm3XX} we filter $|TM|_{rel}$ from Equation~\ref{eq:naive2} to exclude mementos that resolve to HTTP 3XX status codes. $|TM|_D$ represents the count of mementos that result in non-3XX statuses based on the URI-Ms in a TimeMap. Section 2.2.4 of the Memento RFC \cite{rfc7089} states that a link with a datetime attribute must match the value of the Memento-Datetime header when the link is dereferenced.
\begin{equation}\label{eq:rm3XX}
|TM|_D = \sum_{m=1}^{len(M)}
\begin{cases}
0\ & 300 \geq httpStatus(m) < 400,\\
1\ & \text{otherwise.}
\end{cases}
\end{equation}
\begin{equation}\label{eq:just3XX}
|TM|_I = |TM|_{rel} - |TM|_D
\end{equation}
We quantify the ratio of mementos with non-redirecting HTTP status codes (Equation~\ref{eq:rm3XX}) to those with redirects (Equation~\ref{eq:just3XX}) in Equation~\ref{eq:rat} as $DI$.
\begin{equation}\label{eq:rat}
DI =
\begin{cases}
\frac{|TM|_D}{|TM|_I}\ & |TM|_I > 0,\\
\infty & \text{otherwise.}
\end{cases}
\end{equation}
\label{sec:scenario}
As an example, Figure~\ref{fig:balls} contains 11 URI-Ms that result in non-redirecting archived HTTP status codes when dereferenced, inclusive of eight 200 codes, two 4XX codes, and one archived 504. TimeMap A represents a domain for an organization (for example) that is acquired by another organization, whose domain is represented by TimeMap B. At the point of acquisition, TimeMap A redirects to TimeMap B, as represented by the HTTP 301. At two points prior to acquisition, an archival crawler attempted to capture the URI-R for TimeMap A but received a server-side redirect, which is reflected in the preserved HTTP 302 responses. As the acquisition proceeded, the URI-R may have been deleted (the 404 in TimeMap A) and the server misconfigured in the transition (HTTP 504). An intermittent HTTP 401 (Unauthorized) error is also experienced in the URI-R for TimeMap B as the servers are reconfigured to accept the additional traffic from the acquisition.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{mementoBalls4.png}
\caption{
Dereferencing URI-Ms in a TimeMap may produce a variety of HTTP status code, some of which are redirects both to other URI-Ms within the TimeMap and URI-Ms not included in the initial TimeMap. Counting the number of mementos without dereferencing URI-Ms is therefore problematic.}
\label{fig:balls}
\end{figure}
Three \mbox{URI-Ms} resulted in 3XX redirects when dereferenced. Using Equation~\ref{eq:rat}, $DI = 11/3 \approx 3.7$. Sparsely archived URIs will often contain a list of URI-Ms where all result in an HTTP 200 status code when dereferenced, which would result in $DI$ being undefined. It is far less likely that all URI-Ms in a TimeMap return in an HTTP redirect when dereferenced.
During the data acquisition process (Section~\ref{sec:dataCollection}) we experienced intermittent HTTP 5XX status codes \cite{rfc7231} in responses from Internet Archive, namely HTTP 503 (Service Unavailable) and 504 (Gateway Timeout). In much of the same way that sending an Accept-Datetime header causes a ``soft'' HTTP status code to ``harden'', we repeated the request via curl\footnote{\url{https://curl.haxx.se/}} with the inclusion of an Accept-Datetime HTTP header \cite{rfc7089}. This additional step caused no change in the subsequently returned results compared to the originally results response. Repeating the collection procedure for select URI-Ms in the future remedied this issue, allowing us to attribute the error to the archive and not the archive returning a capture of an archived 5XX. If the response instead indicated that the returned 5XX status codes was representative of the state of the \mbox{URI-R} at the respective time (through providing a Memento-Datetime response header) and not an intermittent result attributable to the archive, the \mbox{URI-M} would signify an increment in the Equation~\ref{eq:rm3XX} summation.
\subsection{Canonicalization Patterns}
\label{sec:canonicalizationPatterns}
In observing the mementos for \hyperref[]{http://www.google.com}, we encountered 3 canonicalization patterns for URI-Ms that surface those that are URI-M$_{D}$ versus those that are URI-M$_{I}$. We define $M_{TM}$ to be the memento count for a TimeMap when using only the data contained in the TimeMap without dereferencing mementos (Section~\ref{sec:initialAnalysis}). We define the {\em representative count} ($M_{RC}$) of the number of mementos present in a TimeMap to be the number of URI-M$_{D}$ where $M_{RC} \leq M_{TM}$. These canonicalization patterns observed are Inter-scheme, Slash-added, and subdomain redirect patterns, described in this section.
\def\createlinenumber#1#2{
\edef\thelstnumber{%
\unexpanded{%
\ifnum#1=\value{lstnumber}\relax
#2%
\fi}%
\ifx\thelstnumber\relax\else
\expandafter\unexpanded\expandafter{\thelstnumber}%
\fi
}
}
\let\thelstnumber\relax
\createlinenumber{3}{$M_1$}
\createlinenumber{5}{$M_2$}
\createlinenumber{6}{$M_3$}
\createlinenumber{8}{$M_4$}
\createlinenumber{9}{$M_5$}
\createlinenumber{11}{$M_6$}
\createlinenumber{12}{$M_7$}
\begin{figure*}[t]
\begin{lstlisting}[language=bash, basicstyle=\footnotesize, numbers=left,stepnumber=1, showstringspaces=false,tabsize=1,breaklines=true,breakatwhitespace=false]
<http://google.com>; rel="original",
...
<http://web.archive.org/web/20011124163711/http://www2.google.com/>; rel="memento"; datetime="Sat, 24 Nov 2001 16:37:11 GMT",
...
<http://web.archive.org/web/20130101000813/http://www.google.com/>; rel="memento"; datetime="Tue, 01 Jan 2013 00:08:13 GMT",
<http://web.archive.org/web/20130101003310/https://www.google.com/>; rel="memento"; datetime="Tue, 01 Jan 2013 00:33:10 GMT",
...
<http://web.archive.org/web/20140425221431/http://www.google.com>; rel="memento"; datetime="Fri, 25 Apr 2014 22:14:31 GMT",
<http://web.archive.org/web/20140425221433/https://www.google.com/>; rel="memento"; datetime="Fri, 25 Apr 2014 22:14:33 GMT",
...
<http://web.archive.org/web/20160519223823/http://www.google.com/>; rel="memento"; datetime="Thu, 19 May 2016 22:38:23 GMT",
<http://web.archive.org/web/20160520165954/http://google.com/>; rel="memento"; datetime="Fri, 20 May 2016 16:59:54 GMT",
...
\end{lstlisting}
\caption{A partial TimeMap in Link format for google.com with annotations highlighting various URI-Ms, discussed in Section~\ref{sec:canonicalizationPatterns}.}
\label{fig:tm_link_Ms}
\end{figure*}
\subsubsection{Inter-scheme \mbox{URI-M} Redirect}
\label{sec:interScheme}
As adoption of the secure HTTPS scheme over HTTP becomes more prevalent on the live Web \cite{freedompress, httpsBreaking}, the trend becomes apparent in the archives through canonicalizing the HTTP and HTTPS site to be one in the same. For example, observe two mementos from 2013 from the TimeMap for \hyperref[]{google.com} (Figure~\ref{fig:tm_link_Ms}). The status code returned for $M_2$ is 200 with no HTTP location header present (an example of a URI-M$_{D}$). However, the status code returned for $M_3$ is an HTTP 302 with an HTTP location response header of /web/20130101000813/http://www.google.com/, i.e., $M_3$ redirects to $M_2$ when dereferenced. Thus, $M_3$ is a URI-M$_{I}$. Were the naive but often applied Equation~\ref{eq:naive2} used for determining how many mementos are represented by the URI-Ms $M_2$ and $M_3$, both would be included while dereferencing each \mbox{URI-M} would result in a count of only a single memento. This highlights an important distinction and discrepancy between the number of identifiers (URI-Ms) and the number of representations (mementos).
\subsubsection{Slash-added \mbox{URI-M} Redirect}
\label{sec:slashAdded}
Queries for the URI \hyperref[]{http://www.google.com} are sometimes redirected to the same URI with an appended slash. For example, the two mementos $M_4$ and $M_5$ both exist in a TimeMap (Figure~\ref{fig:tm_link_Ms}).
When dereferenced, $M_4$ returns a 302 with a location header pointing to $M_5$, captured two seconds later based solely on the embedded datetime. Both URI-Ms are reported by the TimeMap while only the latter contains an entity body when dereferenced.
\begin{table}
\centering
\begin{tabular}{ r | r }
Time Gap Bucket &URI-M count\\\hline
0 seconds &24,541\\
1 second &34,577\\
2 seconds &26,153\\
3 seconds &62,526\\
4 seconds &46,738\\
5 seconds &14,215\\
6 seconds &12,213\\
7 seconds &9,748\\
8 seconds &7,431\\
9 seconds &6,610\\
$>$ 9 seconds, $\leq$ 1 minute &101,868\\
$>$ 1 minute, $\leq$ 1 hour &247,192\\
$>$ 1 hour, $\leq$ 1 day &37,399\\
$>$ 1 day &5,034\\ \hline
\end{tabular}
\caption{A range of time differences exists between adjacent captures of a \mbox{URI-R}. This table represents the instances of these differences between adjacent URI-Ms from the TimeMap for google.com reported by Internet Archive.}
\label{tab:google4}
\end{table}
\subsubsection{Subdomain \mbox{URI-M} Redirect}
\label{sec:subdomainSwitch}
It is also useful to observe canonicalization that does not result in a redirect. Google has used a variety of subdomains of the sort containing the literal ``www'' followed by a digit over the years, as with $M_1$ (Figure~\ref{fig:tm_link_Ms}). Accessing this memento (dereferencing $M_1$) results in an HTTP 200 status code. The other, much more common subdomain of www, as with $M_6$ returns an HTTP 302 redirected to $M_7$, both present in the TimeMap. Table~\ref{tab:datatatata} shows the magnitude of redirects for google.com based on the URI-R$_{orig}$ scheme, URI-R$_{orig}$ subdomain, URI-R$_{dest}$ scheme, and URI-R$_{orig}$ subdomain. The breakdown in Table~\ref{tab:datatatata} was also generated for comparison to the \mbox{URI-Rs} vimeo.com and wikipedia.org in Tables~\ref{tab:datatatata_vimeo} and ~\ref{tab:datatatata_wikipedia}, respectively.
\subsubsection{Analysis}
\label{sec:analysis}
Juxtaposing the proportion exhibited for each of the four permutations of scheme transitions (HTTP-to-HTTP, HTTP-to-HTTPS, etc.) from vimeo.com and wikipedia.org as compared to google.com, the inter-scheme transition seems more common with the former pair while the bulk of the results for google.com reside in redirects that retain both the HTTP scheme in the \mbox{URI-R} but also the www subdomain. Focusing specifically on the HTTP-to-HTTPS inter-scheme transition, Figure~\ref{fig:interScheme} serves as an interesting cross-section of Table~\ref{tab:datatatata} broken down by time. Disregarding the anomalous captures from Internet Archive in 2015, the overall trend of inter-scheme redirects is leaning toward more HTTP-to-HTTPS than HTTPS-to-HTTP as the secure scheme is adopted by more sites on the live Web. Disregarding 2015, the HTTP-to-HTTPS redirects for vimeo.com appear to be monotonically increasing with normalization for the partial year results for 2016 (collection was performed in May of that year). Figure~\ref{fig:interScheme} also shows a steeper quantity of captures containing these redirects for wikipedia.org with few captures of redirects exhibiting this inter-scheme permutation prior to 2014. The rapid increase in each site may temporally correspond with the adoption of the HTTPS scheme by the live Web site, thereby forwarding all traffic accessing the HTTP version of the site using an HTTP 3XX response.
An additional nuance to account for the large quantity of redirects from HTTP \mbox{URI-Ms} to HTTP \mbox{URI-Ms} for google.com can be observed by the large quantity of ``revisit'' entries in IA's CDX results for google.com. A revisit entry occurs when an archival crawler is returned content that is identical to a previous capture, often attributed using a hashing scheme on the live page's content. If an archive reports revisit records as an HTTP redirect based on the CDX listing, and this redirect is propagated to the archive's Memento endpoint thus producing a unique URI-M, the $DI$'s value for the \mbox{URI-R} decreases. Requesting the \mbox{URI-M} using the Accept-Datetime HTTP header then observing the Memento-Datetime response header's presence often reveals this nuance, but by relying on the TimeData data without requesting each \mbox{URI-M} , the $DI$ for the \mbox{URI-R} is unknown.
\begin{table}
\centering
\begin{tabular}{ c | r }
Year &Memento count\\\hline
2001 & 68 \\
2005 & 391 \\
2006 & 8 \\
2007 & 81 \\
2008 & 62 \\
2009 & 153 \\
2010 & 124 \\
2011 & 616 \\
2012 & 5,564 \\
2013 & 25,914 \\
2014 & 40,819 \\
2015 & 1,367 \\
2016 & 10,104 \\
\end{tabular}
\caption{Google memento analysis group by memento year where the time between two mementos is less than or equal to 2 seconds.}
\label{tab:lessThan2}
\end{table}
\subsection{Inter-Memento Temporality}
We measured the time between each pair of consecutive mementos, shown in Table \ref{tab:google4}. We found that 38.4\% had a time gap less than 9 seconds, indicating (in some cases) that a redirect would occur when the \mbox{URI-M} is dereferenced. Using a yearly bucketing scheme (Table~\ref{tab:allThePlots}), we plotted the difference in time between adjacent mementos from IA based on the scheme in the URI-M$_{I}$ and the ultimate URI-M that results when the URI-M$_{I}$ is dereferenced. For each log-log plot, shown with more details in the Appendix, a point's quadrant positioning is indicative of the quantity of mementos with a seconds-level granularity of time. For example, a point in the top-left quadrant of a plot indicates that there are many temporally consecutive memento pairs with a very small time difference between them. Top right would indicate many pairs with a large time difference between them; bottom right: few memento pairs with a large time difference; bottom left: few memento pairs with a small time difference. Many more points in the left half of a plot than the right indicates much less time between captures, i.e., the capture frequency was higher that year. More points being in the right half of the plot indicates that more time passes between consecutive captures. The trend for google.com excluding 2015 shows fewer pairs with a small time difference (more points in the bottom right) as time goes on for all redirect patterns other than HTTP-to-HTTP.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{find3XXForwards_line.png}
\caption{Inter-scheme redirects for google.com, vimeo.com, and wikipedia.org from the mementos in the TimeMap from MemGator for IA that result in a 3XX. }
\label{fig:interScheme}
\end{figure}
\subsection{Temporal Closeness as an Indicator of Redirection}
In only brief examination of the TimeMap, some temporally consecutive URI-Ms appeared in ``pairs'' where a second URI-M exists from the same archive within seconds of the previous. Table~\ref{tab:lessThan2} lists the by-year breakdown filtering to only include the URI-Ms where the time between the two is less than two seconds. The trend generally increases with time. This plot can also be cross-referenced with Table~\ref{tab:datatatata}, which shows the overall inter-scheme redirect breakdown totals independent of time with the additional subdomain granularity.
Table~\ref{tab:lessThan2} also shows a peak in 2014 at 40,819 pairs where the redirect is less than or equal to two seconds apart from the URI-R$_{orig}$ to the URI-R$_{dest}$. Given the quantity of inter-scheme redirects in Figure~\ref{fig:interScheme} for 2014 totaling around 30,000 as the sum and Table~\ref{tab:datatatata} showing a significantly larger number of same scheme redirects (e.g., 490,836 just for HTTP-to-HTTP both with the www subdomain), many redirects over the archived history of google.com can be attributed to something other than a scheme switch. The large number of aforementioned identical scheme and subdomain redirects indicates patterned responses like slash-added (Section~\ref{sec:slashAdded}) rather than scheme (Section~\ref{sec:interScheme}) or subdomain switch (Section~\ref{sec:subdomainSwitch}).
\begin{table}
\centering
\begin{subtable}{\textwidth}
\centering
\begin{tabular}{| r | r | r | r | r | r | r | r | r | r |}
\hline
\diagbox{URI-R$_{orig}$ scheme}{URI-R$_{dest}$ scheme} &
& \multicolumn{3}{|c|}{http} & \multicolumn{3}{|c|}{https} \\ \cline{3-8}
\multirow{3}{*}{http} & & none & www & other & none & www & other \\ \hline
& none & 1,279 & 68,837 & 55 & \ooorange{12} & \ooorange{20,825} & \ooorange{27} \\
& www & 8,934 & 490,836 & 204 & \ooorange{32} & \ooorange{77,610} & \ooorange{16} \\
& other & 0 & 224 & 22 & \ooorange{0} & \ooorange{26} & \ooorange{2} \\ \hline
\multirow{3}{*}{https} & none & 14 & 731 & 0 & 0 & 296 & 1 \\
& www & 1,117 & 72,874 & 27 & 15 & 18,525 & 2,101 \\
& other & 0 & 0 & 0 & 0 & 0 & 0 \\ \hline
\end{tabular}
\caption{Scheme and subdomain for redirects when dereferencing URI-Ms for google.com.}
\label{tab:datatatata}
\end{subtable}
\begin{subtable}{\textwidth}
\centering
\begin{tabular}{| r | r | r | r | r | r | r | r | r | r |}
\hline
\diagbox{URI-R$_{orig}$ scheme}{URI-R$_{dest}$ scheme} &
& \multicolumn{3}{|c|}{http} & \multicolumn{3}{|c|}{https} \\ \cline{3-8}
\multirow{3}{*}{http} & & none & www & other & none & www & other \\ \hline
& none & 1,642 & 104 & 0 & \greeeeen{82,637} & \greeeeen{0} & \greeeeen{0} \\
& www & 1,273 & 50 & 0 & \greeeeen{6,355} & \greeeeen{0} & \greeeeen{0} \\
& other & 0 & 0 & 0 & \greeeeen{0} & \greeeeen{0} & \greeeeen{0} \\ \hline
\multirow{3}{*}{https} & none & 315 & 6 & 0 & 35,293 & 1 & 0 \\
& www & 10 & 0 & 0 & 1,149 & 0 & 0 \\
& other & 0 & 0 & 0 & 0 & 0 & 0 \\ \hline
\end{tabular}
\caption{Scheme and subdomain for redirects when dereferencing URI-Ms for vimeo.com.}
\label{tab:datatatata_vimeo}
\end{subtable}
\begin{subtable}{\textwidth}
\centering
\begin{tabular}{| r | r | r | r | r | r | r | r | r | r |}
\hline
\diagbox{URI-R$_{orig}$ scheme}{URI-R$_{dest}$ scheme} &
& \multicolumn{3}{|c|}{http} & \multicolumn{3}{|c|}{https} \\ \cline{3-8}
\multirow{3}{*}{http} & & none & www & other & none & www & other \\ \hline
& none & 91 & 10,575 & 0 & \brown{0} & \brown{4,140} & \brown{0} \\
& www & 110 & 5,099 & 0 & \brown{1} & \brown{44,104} & \brown{0} \\
& other & 0 & 0 & 0 & 0 & \brown{0} & \brown{0} \\ \hline
\multirow{3}{*}{https} & none & 1 & 46 & 0 & 1 & 804 & 0 \\
& www & 14 & 1,014 & 0 & 0 & 5,602 & 0 \\
& other & 0 & 0 & 0 & 0 & 0 & 0 \\ \hline
\end{tabular}
\caption{Scheme and subdomain for redirects when dereferencing URI-Ms for wikipedia.org.}
\label{tab:datatatata_wikipedia}
\end{subtable}
\caption{When URI-Ms for three select domains (google.com, vimeo.com, and wikipedia.org), are dereferenced and produce an HTTP redirect, the originally accessed URI-R$_{orig}$ can result in a URI-R$_{dest}$ with a different scheme and subdomain. Cell colors correspond to lines in Figure~\ref{fig:interScheme}. The scheme and subdomain is ``Unknown'' for URI-Ms (like those from webcitation) that obfuscate \mbox{URI-R} with which the \mbox{URI-M} is associated. See Section~\ref{sec:bg}.}
\label{tab:dataTables}
\end{table}
\begin{table}
\centering
\begin{tabular}{ l | r | r || r | r }
host & \% 3XX & \% 200 & M$_{TM}$ & $DI$ \\\hline
google & 84.89 & 15.11& 695,525 & 0.178\\
yahoo & 88.16 & 11.83 & 418,896 & 0.134\\
sourceforge & 73.34 & 26.63 & 31,408 & 0.363\\
instagram & 67.32 & 32.65 & 55,228 & 0.485\\
vimeo & 57.04 & 42.94 & 199,262 & 0.752\\
cnn & 49.97 & 50.01 & 87,148 & 1.001\\
wikipedia & 44.62 & 55.19 & 25,973 & 1.240\\
whitehouse & 44.57 & 55.24 & 26,006 & 1.243\\
\end{tabular}
\caption{Dereferencing 7 other large Web sites' TimeMaps from Internet Archive produces the above distribution of status codes for each site.}
\label{tab:sites}
\end{table}
\subsection{Beyond Google}
We then evaluated the applicability of the observations for google.com with other archived Web sites. We dereferenced the TimeMaps of 7 additional large Web sites (Table~\ref{tab:sites}) with a variety of adoption trends of HTTPS and ephemerality as well as 13 home pages of various universities and colleges (Table~\ref{tab:edu}). From this further analysis, we observed how prevalent the trend is as exhibited by google.com with a hypothesis that the relatively static, fundamentally unchanging Google homepage is a reason for the relatively low $DI$.
\begin{figure}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_google.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_yahoo.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_vimeo.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_cnn.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_instagram.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_sourceforge.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_whitehouse.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_stanford.png}
\includegraphics[width=0.32\linewidth]{percentRedirectPlots/year_percentRedirects_generatePlot_ufl.png}
\caption{Nine URI-Rs from Tables~\ref{tab:sites} and~\ref{tab:edu} exhibit different degrees of redirection over time.}
\label{fig:redirectPlots}
\end{figure}
For the select academic institutions in Table~\ref{tab:edu}, $DI$ is inversely proportion to M$_{TM}$, albeit not strictly as evidenced by ``gatech'' and ``odu''. This pattern does not generally hold in comparison to the large sites in Table~\ref{tab:sites} though the selection of sites for each may contain some inadvertent bias. Figure~\ref{fig:redirectPlots} shows nine plots representing the percentage of redirects over time as determined when all URI-Ms with a rel value in the respective TimeMaps from IA are dereferenced.
\begin{table}
\centering
\begin{tabular}{ l | r | r || r | r }
host & \% 3XX & \% 200 & M$_{TM}$ & $DI$ \\\hline
stanford & 62.14 & 37.84 & 19,309 & 0.609\\
princeton & 60.10 & 39.88 & 9,355 & 0.663\\
columbia & 48.01 & 51.88 & 9,882 & 1.082\\
harvard & 33.91 & 65.96 & 7,699 & 1.948\\
caltech & 33.13 & 66.86 & 5,474 & 2.017\\
mit & 26.57 & 73.24 & 6,379 & 2.763\\
gatech & 26.03 & 73.94 & 3,907 & 2.841\\
ufl & 24.76 & 75.23 & 4,927 & 3.038\\
vt & 23.07 & 76.92 & 4,061 & 3.334\\
lsu & 15.06 & 84.93 & 2,974 & 5.638\\
nsu & 13.82 & 86.00 & 1,208 & 6.233\\
odu & 9.727 & 90.27 & 1,727 & 9.279\\
tcc & 5.429 & 94.57 & 884 & 17.41\\
\end{tabular}
\caption{Dereferencing the TimeMaps from 13 academic institutions' Web sites from Internet Archive produces the above distribution of status codes for each site.}
\label{tab:edu}
\end{table}
\begin{table}
\centering
\def0{0}
\begin{tabular}{ c | c c c c }
\multirow{2}{*}{year} & \multicolumn{4}{c}{scheme$_{orig}\rightarrow $scheme$_{dest}$} \\ \cline{2-5}
& http$\rightarrow$http & \ooorange{http$\rightarrow$https} & https$\rightarrow$http & https$\rightarrow$https \\ \hline
2010 & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTP_2010_small}.png} & & & \\
2011 & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTP_2011_small}.png} & & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTP_2011_small}.png} & \\
2012 & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTP_2012_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTPS_2012_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTP_2012_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTPS_2012_small}.png} \\
2013 & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTP_2013_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTPS_2013_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTP_2013_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTPS_2013_small}.png} \\
2014 & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTP_2014_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTPS_2014_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTP_2014_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTPS_2014_small}.png} \\
2015 & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTP_2015_small}.png} &
&
\includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTP_2015_small}.png} & \\
2016 & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTP_2013_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPtoHTTPS_2016_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTP_2016_small}.png} & \includegraphics[width=0.16\linewidth]{schemeForwardPlots/{archive.org_HTTPStoHTTPS_2016_small}.png} \\
\end{tabular}
\caption{Count of time delta instances for google.com from IA TimeMaps with 3XX redirects. Highlighted columns correspond to Figures~\ref{tab:datatatata} and \ref{fig:interScheme}.The plots in this Figure are available with more detail in the Appendix.
}
\label{tab:allThePlots}
\end{table}
\section{Conclusions}
In this work we identified the problem of attempting to count the number of mementos in a TimeMap based solely on the contents of the TimeMap. We progressively built a method for counting the number of archived captures of Web pages that contain content when dereferenced from a TimeMap. Through observing google.com, a \mbox{URI-R} with a contemporarily large apparent number of mementos, we dereferenced all URI-Ms in an aggregated TimeMap for the URI-R to show that a 84.9\% of the URI-Ms are redirects to other \mbox{URI-Ms} in the TimeMap.
We establish the nomenclature of M$_{RC}$, a means of communicating the number of URI-Ms in a TimeMap that contain a capture with an entity body, as compared to the more naive M$_{TM}$ as calculated using solely the contents of the TimeMap for a \mbox{URI-R}. We analyzed the TimeMaps for the \mbox{URI-Rs} of seven other contemporary large web sites and the TimeMaps from 13 academic institutions. We introduced the $DI$ metric to evaluate the ratio of non-redirecting \mbox{URI-Ms} in a TimeMap to the ratio of redirecting \mbox{URI-Ms} when all \mbox{URI-Ms} in a TimeMap are dereferenced. Five of the eight large Web sites' \mbox{URI-Rs} and two of the thirteen academic institutions' \mbox{URI-Rs} contained more redirecting than non-redirecting mementos when dereferenced ($DI < 1.0)$.
From the \mbox{URI-Ms} for google.com that redirected, we split the results between those that changed schemes and those that maintained the same URI-R scheme after the redirect. We split the results on an annual basis to show the effect that the introduction of HTTPS has had on \mbox{URI-R} canonicalization over time. We found that despite an anomalous set of captures in the year 2015, the number of redirects per year on the live Web from HTTP \mbox{URI-Rs} to HTTPS \mbox{URI-Rs} as preserved by the archive has superseded the number of redirects of HTTPS \mbox{URI-Rs} to HTTP \mbox{URI-Rs}. Though the quantity of holdings by Internet Archive for redirects from HTTPS to HTTP is not yet larger than the total of other permutations of HTTP(S) to HTTP(S) redirects (Table~\ref{tab:datatatata}), the rapid growth of redirects to the secure scheme for captures confirms and quantifies the increased adoption of HTTPS on the live Web.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
In \cite{Goppa1,Goppa2} Goppa used algebraic curves to construct linear error correcting codes, the so called algebraic geometric codes (AG codes).
The construction of an AG code with alphabet a finite field $\mathbb F_q$ requires that the underlying curve is $\mathbb F_q$-rational and involves two $\mathbb F_q$-rational divisors $D$ and $G$ on the curve.
In general, to construct a ``good" AG code over $\mathbb F_q$ we need a curve $\mathcal X$ with low genus $g$ with respect to its number of $\mathbb F_q$-rational points. In fact, from the Goppa bounds on the parameters of the code it follows that the relative Singleton defect is upper bounded by the ratio $g/N$, where $N$ can be as large as the number of $\mathbb F_q$-rational points of $\mathcal X$ not in the support of $G$.
Maximal curves over $\mathbb F_q$ attain the Hasse-Weil upper bound for the number of $\mathbb F_q$-rational points with respect to their genus and for this reason they have been used in a number of works. Examples of such curves are the Hermitian curve, the GK curve \cite{GK2009}, the GGS curve \cite{GGS}, the Suzuki curve \cite{DL1976}, the Klein quartic when $\sqrt{q}\equiv6\pmod7$ \cite{MEAGHER2008}, together with their quotient curves. Maximal curves often have large automorphism groups which in many cases can be inherited by the code: this can bring good performances in encoding \cite{Joyner2005} and decoding \cite{HLS1995}.
Good bounds on the parameters of one-point codes, that is AG codes arising from divisors $G$ of type $nP$ for a point $P$ of the curve, have been obtained by investigating the Weierstrass semigroup at $P$. These results have been later generalized to codes and semigroups at two or more points; see e.g. \cite{MATTHEWS2001,HOMMA1996,HK2001,CT2005,CK2009,LC2006,Kim1994}.
AG codes from the Hermitian curve have been widely investigated; see \cite{HK2006,HK2005,Tiersma1987,DK2011,HK2006_2,YK1991,Stichtenoth1988} and the references therein. Other constructions based on the Suzuki curve and the curve with equation $y^q + y = x^{q^r+1}$ can be found in \cite{Matthews2004} and \cite{ST2014}.
More recently, AG Codes from the GK curve have been constructed in \cite{FG2010,CT2016,BMZ}.
In the present work we investigate one-point AG codes from the $\mathbb F_{q^{2n}}$-maximal GGS curve, $n\geq5$ odd.
The GGS curve has more short orbits under its automorphism group than other maximal curves, see \cite{GOS}, and hence more possible structures for the Weierstrass semigroups at one point.
On the one hand this makes the investigation more complicated; on the other hand it gives more chances of finding one-point AG codes with good parameters.
One achievement of this work is the determination of the Weierstrass semigroup at any $\mathbb F_{q^2}$-rational point.
We show that the one-point codes at the infinite point $P_\infty$ inherit a large automorphism group from the GGS curve; for many of such codes, the full automorphism group is obtained.
Moreover, for $q=2$, we compute explicitly the Feng-Rao designed minimum distance, which improves the Goppa designed minimum distance.
As an application, we provide families of codes with $q=2$ whose relative Singleton defect goes to zero as $n$ goes to infinity.
We were not able to produce analogous results for an $\mathbb F_{q^2}$-rational affine point $P_0$, because of the more complicated structure of the Weierstrass semigroup.
In a comparison between one-point codes from $P_\infty$ and one-point codes from $P_0$, it turns out that the best codes come sometimes from $P_\infty$, other times from $P_0$; we give evidence of this fact with tables for the case $q=2$, $n=5$.
Note that in general, many of our codes are better than the comparable one-point Hermitian codes on the same alphabet.
In fact, let $C_{1}$ be a code from a one-point divisor $G_{1}$ on the $\mathbb F_{q^{2n}}$-maximal GGS curve with genus $g_1$, with alphabet $\mathbb F_{q^{2n}}$, length $N_{2}$, designed dimension $k_{1}^*=\deg G_{1}-g_{1}+1$, and designed minimum distance $d_{1}^*=\deg G_{1}-(2g_{1}-2)$. In the same way, let $C_{2}$ be a code from a one-point divisor $G_{2}$ on the $\mathbb F_{q^{2n}}$-maximal Hermitian curve with genus $g_2$, with the same alphabet $\mathbb F_{q^{2n}}$ and length $N_2=N_{1}$ as $C_{1}$, designed dimension $k_{2}^*=\deg G_{2}-g_{2}+1$, and designed minimum distance $d_{2}^*=\deg G_{2}-(2g_{2}-2)$.
In order to compare $C_1$ and $C_2$, we can choose $G_1$ and $G_2$ such that $k_1^*=k_2^*$. Then the difference $d_1^*-d_2^*$, like the difference $\delta_2^*-\delta_1^*$ between the designed Singleton defects, is equal to $g_2-g_1=\frac{1}{2}(q^{2n}-q^{n+2}+q^3-q^2)\gg0$.
Finally, we apply our results on AG codes to construct families of quantum codes and convolutional codes.
\section{Preliminaries}\label{Sec:Preliminaries}
\subsection{Curves and codes}\label{Sec:Preliminaries_Curves}
Let $\mathcal{X}$ be a projective, geometrically irreducible, nonsingular algebraic curve of genus $g$ defined over the finite field $\mathbb{F}_q$ of size $q$.
The symbols $\mathcal{X}(\mathbb{F}_q)$ and $\mathbb{F}_q(\mathcal{X})$ denote the set of $\mathbb{F}_q$-rational points and the field of $\mathbb{F}_q$-rational functions, respectively.
A divisor $D$ on $\mathcal{X}$ is a formal sum $n_1P_1+\cdots+n_rP_r$, where $P_i \in \mathcal{X}(\mathbb{F}_q)$, $n_i \in \mathbb{Z}$, $P_i\neq P_j$ if $i\neq j$.
The divisor $D$ is $\mathbb F_q$-rational if it coincides with its image $n_1P_1^q+\cdots+n_rP_r^q$ under the Frobenius map over $\mathbb F_q$.
For a function $f \in \mathbb{F}_q(\mathcal{X})$, $div(f)$ and $(f)_{\infty}$ indicate the divisor of $f$ and its pole divisor.
Also, the Weierstrass semigroup at $P$ will be indicated by $H(P)$.
The Riemann-Roch space associated with an $\mathbb F_q$-rational divisor $D$ is
$$\mathcal{L}(D) := \{ f \in \mathcal{X}(\mathbb{F}_q)\setminus\{0\} \ : \ div(f)+D \geq 0\}\cup\{0\}$$
and its dimension over $\mathbb{F}_q$ is denoted by $\ell(D)$.
Let $P_1,\ldots,P_N\in \mathcal{X}(\mathbb{F}_q)$ be pairwise distinct points and consider the divisor $D=P_1+\cdots+P_N$ and another $\mathbb F_q$-rational divisor $G$ whose support is disjoint from the support of $D$. The AG code $C(D,G)$ is the image of the linear map $\eta : \mathcal{L}(G) \to \mathbb{F}_q^N$ given by $\eta(f) = (f(P_1),f(P_2) ,\ldots,f(P_N))$. The code has length $N$ and if $N>\deg(G)$ then $\eta$ is an embedding and the dimension $k$ of $C(D,G)$ is equal to $\ell(G)$. The minimum distance $d$ satisfies $d\geq d^*=N-\deg(G)$, where $d^*$ is called the designed minimum distance of $C(D,G)$; if in addition $\deg(G)>2g-2$, then by the Riemann-Roch Theorem $k=\deg(G)-g+1$; see \cite[Th. 2.65]{HLP}. The dual code $C^{\bot} (D,G)$ is an $AG$ code with dimension $k^{\bot}=N-k$ and minimum distance $d^{\bot}\geq \deg{G}-2g+2$. If $G=\alpha P$, $\alpha \in \mathbb{N}$, $P \in \mathcal{X}(\mathbb{F}_q)$, the AG codes ${C} (D,G)$ and ${C}^{\bot} (D,G)$ are referred to as one-point AG codes. Let $H(P)$ be the Weierstrass semigroup associated with $P$, that is
$$H(P) := \{n \in \mathbb{N}_0 \ | \ \exists f \in \mathbb{F}_q(\mathcal{X}), (f)_{\infty}=nP\}= \{\rho_1=0<\rho_2<\rho_3<\cdots\}.$$
Denote by $f_{\ell}\in \mathbb{F}_q(\mathcal{X})$, $\ell\geq 1$, a rational function such that $(f_{\ell})_{\infty}=\rho_{\ell}P$. For $\ell \geq0$, define the \emph{Feng-Rao function}
$$\nu_\ell := | \{(i,j) \in \mathbb{N}_0^2 \ : \ \rho_i+\rho_j = \rho_{\ell+1}\}|.$$
Consider ${C}_{\ell}(P)= {C}^{\bot}(P_1+P_2+\cdots+P_N,\rho_{\ell}P)$, $P,P_1,\ldots,P_N$ pairwise distint points in $\mathcal{X}(\mathbb{F}_q)$. The number
$$d_{ORD} ({C}_{\ell}(P)) := \min\{\nu_{m} \ : \ m \geq \ell\}$$
is a lower bound for the minimum distance $d({C}_{\ell}(P))$ of the code ${C}_{\ell}(P)$, called the \emph{order bound} or the \emph{Feng-Rao designed minimum distance} of ${C}_{\ell}(P)$; see \cite[Theorem 4.13]{HLP}. Also, by \cite[Theorem 5.24]{HLP}, $d_{ORD} ({C}_{\ell}(P))\geq \ell+1-g$ and equality holds if $\ell \geq 2c-g-1$, where $c=\max \{m \in \mathbb{Z} \ : \ m-1 \notin H(P)\}.$
A numerical semigroup is called telescopic if it is generated by a sequence $(a_1,\ldots,a_k)$ such that
\begin{itemize}
\item $\gcd(a_1, \ldots , a_k)=1$;
\item for each $i=2,\ldots,k$, $a_i/d_i \in \langle a_1/d_{i-1},\ldots, a_{i-1}/d_{i-1}\rangle$, where $d_i=\gcd(a_1,\ldots,a_i)$;
\end{itemize}
see \cite{KP}.
The semigroup $H(P)$ is called symmetric if $2g-1\notin H(P)$. The property of being symmetric for $H(P)$ gives rise to useful simplifications of the computation of $d_{ORD}(C_\ell(P))$, when $\rho_\ell >2g$. The following result is due to Campillo and Farr\'an; see \cite[Theorem 4.6]{CF}.
\begin{proposition} \label{campillo} Let $\mathcal X$ be an algebraic curve of genus $g$ and let $P \in \mathcal X(\mathbb{F}_q)$. If $H(P)$ is a symmetric Weierstrass semigroup then one has $$d_{ORD}(C_\ell(P))=\nu_{\ell},$$ for all $\rho_{\ell+1}=2g-1+e$ with $e \in H(P) \setminus \{0\}$.
\end{proposition}
\subsection{The automorphism group of an AG code $C(D,G)$}
In the following we use the same notation as in \cite{GK2008,JK2006}.
Let $\mathcal{M}_{N,q}\leq{\rm GL}(N,q)$ be the subgroup of matrices having exactly one non-zero element in each row and column.
For $\gamma\in{\rm Aut}(\mathbb {F}_q)$ and $M=(m_{i,j})_{i,j}\in{\rm GL}(N,q)$, let $M^\gamma$ be the matrix $(\gamma(m_{i,j}))_{i,j}$.
Let $\mathcal{W}_{N,q}$ be the semidirect product $\mathcal M_{N,q}\rtimes{\rm Aut}(\mathbb {F}_q)$ with multiplication $M_1\gamma_1\cdot M_2\gamma_2:= M_1M_2^\gamma\cdot\gamma_1\gamma_2$.
The \emph{automorphism group} ${\rm Aut}({C}(D,G))$ of ${C}(D,G)$ is the subgroup of $\mathcal{W}_{N,q}$ preserving ${C}(D,G)$, that is,
$$ M\gamma(x_1,\ldots,x_N):=((x_1,\ldots,x_N)\cdot M)^\gamma \in {C}(D,G) \;\;\textrm{for any}\;\; (x_1,\ldots,x_N)\in {C}(D,G). $$
Let ${\rm Aut}_{\mathbb {F}_q}(\mathcal X)$ denote the $\mathbb {F}_q$-automorphism group of $\mathcal X$. Also, let
$$ {\rm Aut}_{\mathbb {F}_q,D,G}(\mathcal X)=\{ \sigma\in{\rm Aut}_{\mathbb {F}_q}(\mathcal X)\,\mid\, \sigma(D)=D,\,\sigma(G)\approx_D G \}, $$
where $G'\approx_D G$ if and only if there exists $u\in\mathbb {F}_q(\mathcal X)$ such that $G'-G=(u)$ and $u(P_i)=1$ for $i=1,\ldots,N$, and
$$ {\rm Aut}_{\mathbb {F}_q,D,G}^+(\mathcal X):=\{ \sigma\in{\rm Aut}_{\mathbb {F}_q}(\mathcal X)\,\mid\, \sigma(D)=D,\,\sigma(|G|)=|G| \}, $$
where $|G|=\{G+(f)\mid f\in\overline{\mathbb F}_q(\mathcal X)\}$ is the linear series associated with $G$.
Note that ${\rm Aut}_{\mathbb {F}_q,D,G}(\mathcal X)\subseteq {\rm Aut}_{\mathbb {F}_q,D,G}^+(\mathcal X)$.
\begin{remark}\label{Coincidono}
Suppose that ${\rm supp}(D)\cup{\rm supp}(G)=\mathcal X(\mathbb {F}_q)$ and each point in ${\rm supp}(G)$ has the same weight in $G$. Then
$$ {\rm Aut}_{\mathbb {F}_q,D,G}(\mathcal X) = {\rm Aut}_{\mathbb {F}_q,D,G}^+(\mathcal X) = \{\sigma\in{\rm Aut}_{\mathbb {F}_q}(\mathcal X)\,\mid\,\sigma({\rm supp}(G))={\rm supp}(G) \}. $$
\end{remark}
In \cite{GK2008} the following result was proved.
\begin{theorem}{\rm(\!\!\cite[Theorem 3.4]{GK2008})}\label{Aut}
Suppose that the following conditions hold:
\begin{itemize}
\item $G$ is effective;
\item $\ell(G-P)=\ell(G)-1$ and $\ell(G-P-Q)=\ell(G)-2$ for any $P,Q\in\mathcal X$;
\item $\mathcal X$ has a plane model $\Pi(\mathcal X)$ with coordinate functions $x,y\in\mathcal L(G)$;
\item $\mathcal X$ is defined over $\mathbb F_p$;
\item the support of $D$ is preserved by the Frobenius morphism $(x,y)\mapsto(x^p,y^p)$;
\item $N>\deg(G)\cdot\deg(\Pi(\mathcal X))$.
\end{itemize}
Then
$$ {\rm Aut}({C}(D,G))\cong ({\rm Aut}_{\mathbb {F}_q,D,G}^+(\mathcal X)\rtimes{\rm Aut}(\mathbb {F}_q))\rtimes \mathbb{F}_q^*. $$
\end{theorem}
If any non-trivial element of ${\rm Aut}_{\mathbb {F}_q}(\mathcal X)$ fixes at most $N-1$ $\mathbb {F}_q$-rational points of $\mathcal X$ then ${\rm Aut}({C}(D,G))$ contains a subgroup isomorphic to $ ({\rm Aut}_{\mathbb {F}_q,D,G}(\mathcal X)\rtimes{\rm Aut}(\mathbb {F}_q))\rtimes \mathbb{F}_q^*$; see \cite[Proposition 2.3]{BMZ}.
\subsection{The GGS curve}
Let $q$ be a prime power and consider an odd integer $n$.
The GGS curve $GGS(q,n)$ is defined by the equations
\begin{equation}\label{GGS_equation}
\left\{
\begin{array}{l}
X^q + X = Y^{q+1}\\
Y^{q^2}-Y= Z^m\\
\end{array}
\right.
,
\end{equation}
where $m= (q^n+1)/(q+1)$; see \cite{GGS}.
The genus of $GGS(q,n)$ is $\frac{1}{2}(q-1)(q^{n+1}+q^n-q^2)$, and $GGS(q,n)$ is $\mathbb F_{q^{2n}}$-maximal.
Let $P_0=(0,0,0)$, $P_{(a,b,c)}=(a,b,c)$, and let $P_{\infty}$ be the unique ideal point of $GGS(q,n)$.
Note that $GGS(q,n)$ is singular, being $P_\infty$ its unique singular point. Yet, there is only one place of $GGS(q,n)$ centered at $P_\infty$; therefore, we can actually construct AG codes from $GGS(q,n)$ as described in Section \ref{Sec:Preliminaries_Curves} (see \cite[Appendix B]{Sti} and \cite[Chapter 8]{HKT} for an introduction to the concept of place of a curve).
The divisors of the functions $x,y,z$ satisfying $x^q + x = y^{q+1}$, $y^{q^2}-y= z^m$ are
$$
(x)=m(q+1)P_0-m(q+1)P_{\infty},
$$
$$
(y)=m\sum_{\alpha^q+\alpha=0} P_{(\alpha,0,0)}-mqP_{\infty},
$$
$$
(z)=\sum_{\scriptsize\begin{array}{l} \alpha^q+\alpha=\beta\\ \beta \in \mathbb{F}_{q^2}\\ \end{array}} P_{(\alpha,\beta,0)}-q^3P_{\infty}.
$$
Throughout the paper we indicate by $\overline D$ and $\tilde D$ the divisors
\begin{equation}\label{Dbarra}
\overline D=\sum_{P\in GGS(q,n)(\mathbb{F}_{q^{2n}}),\ P\ne P_{\infty}}P,\qquad \tilde D=\sum_{P\in GGS(q,n)(\mathbb{F}_{q^{2n}}),\ P\ne P_{0}}P.
\end{equation}
\subsection{Structure of the paper}
The paper is organized as follows. In Section \ref{Sec:dord2} the value of $d_{ORD}(C_{\ell}(P_\infty))$ for $q=2$ and $n\geq 5$ is obtained, where $C_\ell(P_\infty)=C^{\perp}(\overline{D},\ell P_{\infty})$; this is applied in Section \ref{Sec:Application1} to two families of codes with $q=2$ whose relative Singleton defect goes to zero as $n$ goes to infinity.
In Section \ref{Sec:P_0} we determine the Weierstrass semigroup at $P_0$, and hence at any $\mathbb F_{q^2}$-rational affine point of $GGS(q,n)$.
The tables in Section \ref{Sec:tabelle} describe the parameters of $C_\ell{P_\infty}$ and $C_\ell(P_0)$ in the particular case $q^{2n}=2^{10}$.
Sections \ref{Sec:Application2} and \ref{Sec:Application3} provide families of quantum codes and convolutional codes constructed from $C_\ell(P_\infty)$ and $C_\ell(P_0)$.
Finally, we compute in Section \ref{Sec:Auto} the automorphism group of the AG code $C(\overline{D},\ell P_{\infty})$ for $q^n+1 \leq \ell \leq q^{n+2}-q^3$.
\section{The computation of $d_{ORD}(C_{\ell}(P_\infty))$ for $q=2$}\label{Sec:d_ORD}\label{Sec:dord2}
In this section we deal with the codes $C_{\ell}(P_\infty)={C}^{\bot}(\overline D,\rho_{\ell}P_\infty)$, where $\overline D$ is as in \eqref{Dbarra}. Our purpose is to exhibit the exact value of $d_{ORD}(C_{\ell}(P_\infty))$ for the case $q=2$. First of all we determine the values of $\nu_\ell$ (Subsection \ref{Sec:d_ORD_sub1}); in Subsections \ref{Sec:d_ORD_sub2}, \ref{Sec:d_ORD_sub3}, \ref{Sec:d_ORD_sub4} we compute $d_{ORD}(C_{\ell}(P_\infty))$.
\subsection{The Feng-Rao function $\nu_\ell$ for $q=2$}\label{Sec:d_ORD_sub1}
Assume that $q=2$ and $n \geq 5$ is odd. Let $m=\frac{2^n+1}{3}$. Then, from \cite[Corollary 3.5]{GOS},
$$H(P_\infty)= \bigg\{i(2^n+1)+2j \frac{2^n+1}{3}+8k \mid i \in \{0,1\}, \ j \in \{0,1,2,3\}, \ k \geq 0 \bigg\}.$$
\begin{remark} \label{scrittura} Let $\rho_\ell=i(2^n+1)+2j \frac{2^n+1}{3}+8k \in H(P_\infty)$. Then $\rho_\ell$ is uniquely determined by the triple $(i,j,k)$.
\end{remark}
\begin{proof}
Assume that $i(2^n+1)+2j \frac{2^n+1}{3}+8k=i'(2^n+1)+2j' \frac{2^n+1}{3}+8k'$. Then $i \equiv i' \pmod2$ and since $i,i' <2$ we have that $i=i'$. Thus, $2j\frac{2^n+1}{3}+8k=2j' \frac{2^n+1}{3}+8k'$. Since this implies that $j \equiv j' \pmod4$ and $j,j'<4$, we have that $j=j'$ and $k=k'$ and the claim follows.
\end{proof}
According to Remark \ref{scrittura}, the notation $(i,j,k)$ is used to indicate the non-gap at $P_\infty$ associated with the choices of the parameters $i,j,k$. In order to compute $d_{ORD}(C_{\ell}(P_\infty))$ the following definition is required. Let $\rho_\ell \in H(P_\infty)$ be fixed. Assume that $\rho_{\ell+1}=(i,j,k)$. Then,
$$\nu_\ell= \big| \{(i_r,j_r,k_r), \ r=1,2 \mid (i,j,k)=(i_1,j_1,k_1)+(i_2,j_2,k_2) \}\big|.$$
In the following lemmas we determine the value of $\nu_\ell$.
\begin{lemma} \label{i1} Let $\rho_\ell \in H(P_\infty)$ be fixed. Assume that $\rho_{\ell+1}=(1,j,k)$ for some $j=0,1,2,3$ and $k \geq 0$. Then,
$$\nu_\ell= \begin{cases} 2(j+1)(k+1), \ \ \textrm{if} \ \ k<m, \\ 2(j+1)(k+1)+2(3-j)(k-m+1), \ \ \textrm{otherwise}. \end{cases}$$
\end{lemma}
\begin{proof}
Let $i_1,i_2,j_1,j_2,k_1$, and $k_2 \in \mathbb{N}$ be such that
$$(2^n+1)+2jm+8k=(i_1+i_2)(2^n+1)+2(j_1+j_2)m+8(k_1+k_2)=3(i_1+i_2)m+2(j_1+j_2)m+8(k_1+k_2).$$
Then $i_1+i_2 \equiv 1 \pmod2$ and since $i_1+i_2 \leq 2$ we have that $i_1+i_2=1$. This implies that
$$3m+2jm+8k=3m+2(j_1+j_2)m+8(k_1+k_2),$$
and hence
\begin{equation}
\label{eq1}
jm+4k=(j_1+j_2)m+4(k_1+k_2).
\end{equation}
Assume that $j=0$. Then from (\ref{eq1}), $(j_1+j_2)m \equiv 0 \pmod 4$ and so, $j_1+j_2=4h$ for some integer $h$. Since $0 \leq j_1+j_2 \leq 6$ we have that $h=0$ or $h=1$. In the first case $k_1+k_2=k$, in the second case $k_1+k_2=k-m$. Since $k_1+k_2 \geq 0$, if $k < m$ the second case cannot occur. Thus, if $k < m$, since we have $2$ possible choices for $i_1$ and $(k+1)$ choices for $k_1$ (while $i_2$ and $k_2$ are determined according to the choices of $i_1$ and $k_1$, respectively), then $\nu_\ell=2(k+1)$.
Also, if $k \geq m$ we have that
$$\nu_\ell=2(k+1)+2\cdot|\{(j_1,j_2) : 0 \leq j_1,j_2 \leq 3, \ j_1+j_2=4\}|\cdot(k-m+1)=2(k+1)+6(k-m+1)$$
and the claim follows by direct checking. \\
Assume that $j=1$. Then from (\ref{eq1}), $(j_1+j_2)m \equiv m \pmod 4$ and so $j_1+j_2=1+4h$ for some integer $h$. Since $0 \leq j_1+j_2 \leq 6$ we have that $h=0$ or $h=1$. In the first case $k_1+k_2=k$, in the second case $k_1+k_2=k-m$. Since $k_1+k_2 \geq 0$ if $k < m$ the second case cannot occur. Thus, if $k < m$, since we have $2$ possible choices for $i_1$, $2$ possible choices for $j_1$ and $(k+1)$ choices for $k_1$, then $\nu_\ell=4(k+1)$.
Also, if $k \geq m$ we have that,
$$\nu_\ell=4(k+1)+2\cdot|\{(j_1,j_2) : 0 \leq j_1,j_2 \leq 3, \ j_1+j_2=5\}|\cdot(k-m+1)=4(k+1)+4(k-m+1),$$
and the claim follows by direct checking. \\
Assume that $j=2$. Then from (\ref{eq1}), $(j_1+j_2)m \equiv 2m \pmod4$ and so $j_1+j_2=2+4h$, for some integer $h$. Since $0 \leq j_1+j_2 \leq 6$ we have that $h=0$ or $h=1$. In the first case $k_1+k_2=k$, in the second case $k_1+k_2=k-m$. Since $k_1+k_2 \geq 0$, if $k < m$ the second case cannot occur. Thus, if $k < m$, since we have $2$ possible choices for $i_1$, $3$ possible choices for $j_1$ and $(k+1)$ choices for $k_1$, then $\nu_\ell=6(k+1)$.
Also, if $k \geq m$ we have that
$$\nu_\ell=6(k+1)+2(k-m+1)\cdot\big |\{(j_1,j_2) : 0 \leq j_1,j_2 \leq 3, \ j_1+j_2=6\}\big | =6(k+1)+2(k-m+1),$$
and the claim follows by direct checking.
Assume that $j=3$. Then from (\ref{eq1}), $(j_1+j_2)m \equiv 3m \pmod 4$ and so $j_1+j_2=3+4h$, for some integer $h$. Since $0 \leq j_1+j_2 \leq 6$ we have that $h=0$. Since this implies that $k_1+k_2=k$, we have that $\nu_\ell=8(k+1)$.
\end{proof}
Using a similar approach we can prove the following.
\begin{lemma} \label{i0}
Let $\rho_\ell \in H(P_\infty)$ be fixed. Assume that $\rho_{\ell+1}=(0,j,k)$ for some $j=0,1,2,3$ and $k \geq 0$. Then,
$$\nu_\ell= \begin{cases} (j+1)(k+1) + \lfloor \frac{j}{3} \rfloor (k+1), \ \ \textrm{if} \ \ k<m, \\ (j+1)(k+1) + \lfloor \frac{j}{3} \rfloor (k+1) + (5-2 \max\{0,j-2\})(k-m+1), \ \ \textrm{if} \ \ m \leq k <2m, \\ (j+1)(k+1) + \lfloor \frac{j}{3} \rfloor (k+1) + \\ + (5-2 \max\{0,j-2\})(k-m+1) + \max \{ 0,2-j\} (k-2m+1), \ \ \textrm{otherwise}. \end{cases}$$
\end{lemma}
\subsection{Computation of $d_{ORD}(C_{\ell}(P_\infty))$ for $ \rho_{\ell+1}=(1,j,k)$ and $\rho_{\ell} \leq 2g$}\label{Sec:d_ORD_sub2}
Let $\rho_\ell \in H(P_\infty)$. Assume that $\rho_{\ell+1}=(1,j,k)$ for some $j=0,1,2,3$ and $k \geq 0$.
Recall that $C_\ell(P_\infty)$ is the dual code of the AG code $C(\overline {D},\rho_{\ell}P_{\infty})$, where $\overline D$ is as in \eqref{Dbarra}.
\begin{lemma} \label{dord1}
If $\rho_{\ell+1}=(1,0,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 2, \ \ \textrm{if} \ \ k=0, \\ 3, \ \ \textrm{if} \ \ k \leq \lfloor \frac{m}{8} \rfloor, \\ 4, \ \ \textrm{if} \ \ \frac{m}{8} < k \leq \lfloor \frac{m}{4} \rfloor, \\ 5, \ \ \textrm{if} \ \ \frac{m}{4} < k \leq \lfloor \frac{3m}{8} \rfloor, \\ 6, \ \ \textrm{if} \ \ \frac{3m}{8} < k \leq \lfloor \frac{m}{2} \rfloor, \\ 8, \ \ \textrm{if} \ \ \frac{m}{2} < k \leq \lfloor \frac{3m}{4} \rfloor, \\ 8 \big(\lceil k - \frac{3m}{4}\rceil +1 \big), \ \ \textrm{if} \ \ \frac{3m}{4} \leq k \leq m-2, \\ \nu_\ell=2m, \ \ \textrm{if} \ \ k=m-1. \end{cases}$$
\end{lemma}
\begin{proof}
For $\rho_s \in H(P_\infty)$ the following system of inequalities is considered:
\begin{equation}
\label{sist}
\begin{cases} \rho_{s+1} \geq \rho_{\ell+1}, \\ \nu_s \leq \nu_\ell. \end{cases}
\end{equation}
In order to compute $d_{ORD}(C_\ell(P_\infty))$ we take the minimum value of $\nu_s$ such that System \eqref{sist} is satisfied. Also, a case-by-case analysis with respect to $a\in\{0,1\}$ is required. Assume that $\rho_{s+1}=(a,b,c)$ for some $a\in\{0,1\}$, $b\in\{0,1,2,3\}$ and $c \geq 0$. From Lemma \ref{i1}, System \eqref{sist} reads,
\begin{equation}
\label{sist1}
\begin{cases} 3am+2bm+8c \geq 3m+8k, \\ \nu_s \leq 2(k+1). \end{cases}
\end{equation}
{\bf Case 1: $a=1$ and $c<m$.} From Lemma \ref{i1}, System \eqref{sist1} reads
$$\begin{cases} 2bm+8c \geq 8k, \\ 2(b+1)(c+1) \leq 2(k+1). \end{cases}$$
\begin{itemize}
\item If $b=0$ then $c=k$ and so the unique solution is $\nu_\ell$ itself.
\item If $b=1$ then $c \geq \lceil k-\frac{m}{4} \rceil$ and $c \leq \lfloor \frac{k-1}{2} \rfloor$. Such a $c$ exists if and only if $ \lceil k-\frac{m}{4} \rceil \leq \lfloor \frac{k-1}{2} \rfloor$.
Assume that $k$ is odd. Then $k-\lfloor \frac{m}{4} \rfloor =\lceil k-\frac{m}{4} \rceil \leq \lfloor \frac{k-1}{2} \rfloor =\frac{k-1}{2}$ if and only if $k \leq 2\lfloor \frac{m}{4} \rfloor -1$.
Similarly if $k$ is even then $c$ exists if and only if $ k-\lfloor \frac{m}{4} \rfloor \leq \frac{k-2}{2}$, that is $k \leq 2\lfloor \frac{m}{4} \rfloor -2$. For these cases the minimum is obtained taking $c=\max\{0,\lceil k-\frac{m}{4} \rceil$\} and hence $\nu_s=4(\max\{0, \lceil k-\frac{m}{4} \rceil \} +1)$.
\item If $b=2$ then $c \geq \lceil k-\frac{m}{2} \rceil$ and $c \leq \lfloor \frac{k-2}{3} \rfloor$. As before, such a $c$ exists if and only if $\lceil k-\frac{m}{2} \rceil \leq \lfloor \frac{k-2}{3} \rfloor$.
This is equivalent to $k \leq \frac{3}{2}(\lfloor \frac{m}{2} \rfloor -1)$ if $k \equiv 0 \pmod3$, to $k \leq \frac{3}{2}\lfloor \frac{m}{2} \rfloor -2$ if $k \equiv 1 \pmod3$, to $k \leq \frac{3}{2}\lfloor \frac{m}{2} \rfloor -1$ if $k \equiv 2 \pmod3$.
For these cases the minimum is obtained taking $c=\max\{0, \lceil k-\frac{m}{2} \rceil\}$ and hence $\nu_s=6(\max\{0, \lceil k-\frac{m}{2} \rceil\} +1)$.
\item If $b=3$ then $c \geq \lceil k-\frac{3m}{4} \rceil$ and $c \leq \lfloor \frac{k-3}{4} \rfloor$.
As before, such a $c$ exists if and only if $\lceil k-\frac{3m}{4} \rceil \leq \lfloor \frac{k-3}{4} \rfloor$.
By direct checking, this is equivalent to $k\leq m-2$.
Here the minimum is obtained taking $c=\max\{0, \lceil k-\frac{3m}{4} \rceil\}$ and hence $\nu_s=8(\max\{0, \lceil k-\frac{3m}{4} \rceil\} +1)$.
\end{itemize}
When $k > \frac{3m}{4}$ and $k \leq m-2$ the minimum value above is obtained as $\nu_s=8(\lceil k-\frac{3m}{4} \rceil +1)$. We observe that if $k=m-1$ then $\nu_\ell=2(k+1)=2m$ and $8(\max\{0, \lceil k-\frac{3m}{4} \rceil\} +1)=8(\lceil m-1-\frac{3m}{4} \rceil\} +1)>2m$. This implies that if $k=m-1$ then the minimum value is $\nu_\ell=2m$ itself.
Thus, combining the previous results we obtain
\begin{equation}
\label{min1}
\min\{\nu_s \mid a=1 \ \ \textrm{and} \ \ c<m\}=\begin{cases} 2, \ \ \textrm{if} \ \ k=0, \\ 4 \ \ \textrm{if} \ \ 1 \leq k \leq \lfloor \frac{m}{4} \rfloor, \\ 6, \ \ \textrm{if} \ \ \frac{m}{4} < k \leq \lfloor \frac{m}{2} \rfloor, \\ 8, \ \ \textrm{if} \ \ \frac{m}{2} < k \leq \lfloor \frac{3m}{4} \rfloor, \\ 8(\lceil k-\frac{3m}{4} \rceil +1), \ \ \textrm{if} \ \ \frac{3m}{4} <k \leq m-2, \\ 2m=\nu_\ell, \ \ \textrm{if} \ \ k=m-1. \end{cases}
\end{equation}
{\bf Case 2: $a=1$ and $c\geq m$.} From Lemma \ref{i1} System \eqref{sist1} reads,
$$\begin{cases} 2bm+8c \geq 8k, \\ 2(b+1)(c+1) +2(3-b)(c-m+1) \leq 2(k+1). \end{cases}$$
Since $2(b+1)(c+1) +2(3-b)(c-m+1) \geq 2(c+1)$ and $c>k$ this case cannot occur.
\ \\
{\bf Case 3: $a=0$ and $c<m$.} From Lemma \ref{i0} \eqref{sist1} reads,
$$\begin{cases} 2bm+8c \geq 3m+8k, \\ (b+1)(c+1) + \lfloor \frac{b}{3} \rfloor (c+1) \leq 2(k+1). \end{cases}$$
\begin{itemize}
\item If $b=0$ then $c \geq \lceil k+\frac{3m}{8} \rceil$ and $c \leq 2k+1$.
Such a $c$ exists if and only if
$k \geq\lfloor \frac{3m}{8} \rfloor$.
For these cases, the minimum is obtained taking $c= \lceil k+\frac{3m}{8} \rceil$ and hence $\nu_s=(\lceil k+\frac{3m}{8} \rceil\ +1)$.
\item The case $b=1$ cannot occur. In fact we have $c \geq \lceil k+\frac{m}{8} \rceil$ and $2(c+1) \leq 2(k+1)$, a contradiction.
\item If $b=2$ then $c \geq \lceil k-\frac{m}{8} \rceil$ and $c \leq \lfloor \frac{2k-1}{3} \rfloor$.
Such a $c$ exists if and only if $k+\lceil -\frac{m}{8} \rceil \leq \lfloor \frac{2k+1}{3} \rfloor$.
This is equivalent to $k \leq 3 \lfloor \frac{m}{8} \rfloor -1$ if $2k\equiv1\pmod3$, to $k \leq 3 \lfloor \frac{m}{8} \rfloor +1$ if $2k\equiv2\pmod3$, to $k \leq 3 \lfloor \frac{m}{8} \rfloor -3$ if $2k\equiv0\pmod3$.
For these cases, the minimum is obtained taking $c=\max\{0, \lceil k-\frac{m}{8} \rceil\}$ and hence $\nu_s=3(\max\{0, \lceil k-\frac{m}{8} \rceil\} +1)$.
\item If $b=3$ then $c \geq \lceil k-\frac{3m}{8} \rceil$ and $c \leq \lfloor \frac{2k-3}{5} \rfloor$.
Such a $c$ exists if and only if $k+\lceil -\frac{3m}{8} \rceil \leq \lfloor \frac{2k-3}{5} \rfloor$.
This is equivalent to $k \leq \frac{5}{3} \lfloor \frac{3m}{8} \rfloor -\frac{5}{3}$ if $2k\equiv0\pmod5$, to $k \leq \frac{5}{3} \lfloor \frac{3m}{8} \rfloor -2$ if $2k\equiv1\pmod5$, to $k \leq \frac{5}{3} \lfloor \frac{3m}{8} \rfloor -\frac{7}{3}$ if $2k\equiv2\pmod5$, to $k \leq \frac{5}{3} \lfloor \frac{3m}{8} \rfloor -1$ if $2k\equiv3\pmod5$, to $k \leq \frac{5}{3} \lfloor \frac{3m}{8} \rfloor -\frac{4}{3}$ if $2k\equiv4\pmod5$.
In these cases, the minimum is obtained taking $c=\max\{0, \lceil k-\frac{3m}{8} \rceil\}$ and hence $\nu_s=5(\max\{0, \lceil k-\frac{3m}{8} \rceil\} +1)$.
\end{itemize}
Thus, we obtain
\begin{equation}
\label{min3}
\min\{\nu_s \mid a=0 \ \ \textrm{and} \ \ c<m\}= \begin{cases} 3, \ \ \textrm{if} \ \ k \leq \lfloor \frac{m}{8} \rfloor, \\ 5, \ \ \textrm{if} \ \ \frac{m}{8} < k \leq \lfloor \frac{3m}{8} \rfloor, \\ 5(\lceil k-\frac{3m}{8} \rceil +1), \ \textrm{if} \ \ \lceil \frac{3m}{8} \rceil \leq k \leq \lfloor \frac{5}{3} \lfloor \frac{3m}{8} \rfloor-\frac{7}{3} \rfloor, \\ (\lceil k+\frac{3m}{8} \rceil\ +1), \ \textrm{otherwise}.\end{cases}
\end{equation}
{\bf Case 4: $ a=0$ and $m \leq c < 2m$.} From Lemma \ref{i0} System \eqref{sist1} reads,
$$\begin{cases} 2bm+8c \geq 3m+8k, \\ (b+1)(c+1) + \lfloor \frac{b}{3} \rfloor (c+1) +(5-2\max\{0,b-2\})(c-m+1) \leq 2(k+1). \end{cases}$$
Since $(b+1)(c+1) + \lfloor \frac{b}{3} \rfloor (c+1) +(5-2\max\{0,b-2\})(c-m+1) \geq (b+1)(c+1)$ and $c>k$, cases $b=1,2,3$ cannot occur. Thus $b=0$ and
$$\begin{cases} 8c \geq 3m+8k, \\ (c+1) +5(c-m+1) \leq 2(k+1). \end{cases}$$
Hence $c \geq \lceil k+\frac{3m}{8}\rceil$ and $c \leq \lfloor \frac{2k+5m-4}{6} \rfloor$. Since $c \geq m$ then $k \geq \frac{m+4}{2}$. The minimum value is obtained (when it is possible) for $c = \lceil k+\frac{3m}{8}\rceil$. By direct checking the minimum $\nu_\ell$ is bigger than the one obtained in \eqref{min1}, and hence we can discard this case.
\ \\
{\bf Case 5: $a=0$ and $c\geq 2m$.} From Lemma \ref{i0} System \eqref{sist1} reads,
$$\begin{cases} 2bm+8c \geq 3m+8k, \\ (b+1)(c+1) + \lfloor \frac{b}{3} \rfloor (c+1) +(5-2\max\{0,b-2\})(c-m+1)+\max\{0,2-b\}(c-2m+1) \leq 2(k+1). \end{cases}$$
Since $(b+1)(c+1) + \lfloor \frac{b}{3} \rfloor (c+1) +(5-2\max\{0,b-2\})(c-m+1)+\max\{0,2-b\}(c-2m+1) \geq (b+1)(c+1) \geq 2m+1$ and $2(k+1) \leq 2m$ this case cannot occur.
Taking the minimum of the values in \eqref{min1} and \eqref{min3} the claim follows.
\end{proof}
Using the same arguments the following results are obtained.
\begin{lemma} \label{dord3}
If $\rho_{\ell+1}=(1,3,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\nu_\ell.$$
\end{lemma}
\begin{lemma} \label{dord2}
If $\rho_{\ell+1}=(1,1,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 4, \ \ \textrm{if} \ \ k=0, \\ 5, \ \ \textrm{if} \ \ k \leq \lfloor \frac{m}{8} \rfloor, \\ 6 \ \ \textrm{if} \ \ \frac{m}{8} < k \leq \lfloor \frac{m}{4} \rfloor, \\ 8, \ \ \textrm{if} \ \ \frac{m}{4} < k \leq \lfloor \frac{m}{2} \rfloor, \\ 8(\lceil k -\frac{m}{2} \rceil+1), \ \ \textrm{if} \ \ \lceil \frac{m}{2} \rceil \leq k\leq \lfloor \frac{3m}{4} \rfloor-2 ,
\\ 2(\lceil \frac{m}{4} +k \rceil+1)+6(\lceil \frac{m}{4} +k \rceil-m+1), \ \ \textrm{if} \ \ \lfloor \frac{3m}{4} \rfloor-1 \leq k \leq m-2, \\ 4m, \ \ \textrm{if} \ \ k=m-1. \end{cases}$$
\end{lemma}
\begin{lemma} \label{dord4}
If $\rho_{\ell+1}=(1,2,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 6, \ if \ k=0, \\ 8, \ \ \textrm{if} \ \ k \leq \lfloor \frac{m}{4} \rfloor, \\ 8(\lceil k -\frac{m}{4} \rceil+1), \ \ \textrm{if} \ \ \lceil \frac{m}{4} \rceil \leq k \leq \lfloor \frac{m}{2} \rfloor -2,
\\ 2(\lceil k +\frac{m}{2} \rceil+1)+6(\lceil k +\frac{m}{2} \rceil-m+1), \ \ \textrm{if} \ \ \lfloor \frac{m}{2} \rfloor -1 \leq k \leq \lfloor \frac{3m}{4} \rfloor -2,
\\ 4(\lceil k +\frac{m}{4} \rceil+1)+4(\lceil k +\frac{m}{4} \rceil-m+1), \ \ \textrm{if} \ \ \lfloor \frac{3m}{4} \rfloor -1 \leq k \leq m-2,
\\ \nu_\ell=6m, \ \ \textrm{if} \ \ k= m-1. \end{cases}$$
\end{lemma}
\subsection{Computation of $ d_{ORD}(C_{\ell}(P_\infty))$ for $ \rho_{\ell+1}=(0,j,k)$ and $\rho_{\ell} \leq 2g$}\label{Sec:d_ORD_sub3}
Using the same arguments as above we obtain the following results in the case $\rho_{\ell} \leq 2g$.
\begin{lemma} \label{dord5}
If $\rho_{\ell+1}=(0,0,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 2, \ \ \textrm{if} \ \ k \leq \lfloor \frac{3m}{8} \rfloor, \\ 3, \ \ \textrm{if} \ \ \lceil \frac{3m}{8} \rceil \leq k \leq \lfloor \frac{m}{2} \rfloor, \\ 4, \ \ \textrm{if} \ \ \lceil \frac{m}{2} \rceil \leq k \leq \lfloor \frac{5m}{8} \rfloor, \\ 5, \ \ \textrm{if} \ \ \lceil \frac{5m}{8} \rceil \leq k \leq \lfloor \frac{3m}{4} \rfloor, \\ 6, \ \ \textrm{if} \ \ \lceil \frac{3m}{4} \rceil \leq k \leq \lfloor \frac{7m}{8} \rfloor, \\ 8, \ \ \textrm{if} \ \ \lceil \frac{7m}{8} \rceil \leq k \leq m-1. \end{cases}$$
\end{lemma}
\begin{lemma} \label{dord6}
If $\rho_{\ell+1}=(0,1,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 2, \ \ \textrm{if} \ \ k \leq \lfloor \frac{m}{8} \rfloor, \\ 3, \ \ \textrm{if} \ \ \lceil \frac{m}{8} \rceil \leq k \leq \lfloor \frac{m}{4} \rfloor, \\ 4, \ \ \textrm{if} \ \ \lceil \frac{m}{4} \rceil \leq k \leq \lfloor \frac{3m}{8} \rfloor, \\ 5, \ \ \textrm{if} \ \ \lceil \frac{3m}{8} \rceil \leq k \leq \lfloor \frac{m}{2} \rfloor, \\ 6, \ \ \textrm{if} \ \ \lceil \frac{m}{2} \rceil \leq k \leq \lfloor \frac{5m}{8} \rfloor, \\ 8(\max\{0,\lceil k-\frac{7m}{8} \rceil\}+1), \ \ \textrm{if} \ \ \lceil \frac{5m}{8} \rceil \leq k \leq m-1. \end{cases}$$
\end{lemma}
\begin{lemma}
If $\rho_{\ell+1}=(0,3,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 6, \ \ \textrm{if} \ \ k \leq \lfloor \frac{m}{8} \rfloor, \\ 8(\max\{0,\lceil \frac{3m}{8} \rceil\}+1), \ \ \textrm{if} \ \ \lceil \frac{m}{8} \rceil \leq k \leq m-2, \\ \nu_\ell=5(k+1), \ \ \textrm{if} \ \ k=m-1. \end{cases}$$
\end{lemma}
\begin{lemma} \label{dord7}
If $\rho_{\ell+1}=(0,2,k)$ for some $k<m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 4, \ \ \textrm{if} \ \ k \leq \lfloor \frac{m}{8} \rfloor, \\ 5, \ \ \textrm{if} \ \ \lceil \frac{m}{8} \rceil \leq k \leq \lfloor \frac{m}{4} \rfloor, \\ 6, \ \ \textrm{if} \ \ \lceil \frac{m}{4} \rceil \leq k \leq \lfloor \frac{3m}{8} \rfloor, \\ 8(\max\{0, \lceil k-\frac{5m}{8}\rceil \}+1), \ \ \textrm{if} \ \ \lceil \frac{3m}{8} \rceil \leq k \leq \lfloor \frac{7m}{8} \rfloor-2, \\ 2(\lceil k+\frac{m}{8} \rceil+1), \ \ \textrm{if} \ \ \lfloor \frac{7m}{8} \rfloor -1 \leq k \leq m-3, \\ 3(k+1)=\nu_\ell, \ \ \textrm{if} \ \ k \in \{m-2, m-1\}. \end{cases}$$
\end{lemma}
\begin{lemma} \label{dord8}
If $\rho_{\ell+1}=(0,0,k)$ for some $m \leq k <2m$ then
$$d_{ORD}(C_\ell(P_\infty))=\begin{cases} 8(\lceil k-\frac{9m}{8}\rceil+1), \ \ \textrm{if} \ \ m \leq k < \lfloor\frac{11m}{8} -1 \rfloor, \\
2(\lceil k-\frac{3m}{8} \rceil+1)+\max\{0, 6(\lceil k-\frac{3m}{8}\rceil -m+1)\}, \ \ \textrm{if} \ \ \lfloor\frac{11m}{8} -1 \rfloor \leq k <2m. \end{cases}$$
\end{lemma}
\subsection{Computation of $d_{ORD}(C_{\ell}(P_\infty))$ for $\rho_{\ell}>2g$} \label{Sec:d_ORD_sub4}
\begin{proposition}\label{rimarco}
The Weierstrass semigroup $H(P_\infty)=\langle q^3, m q, q^n+1 \rangle$ is telescopic.
\end{proposition}
\begin{proof}
Let $a_1=q^3$, $a_2=mq$, $a_3=q^n+1$, $d_0=0$, $d_1=q^3$, $d_2=\gcd(q^3,mq)=q$, $d_3=\gcd(q^3,mq,q^n+1)=1$. Then $a_i/d_i\in\langle a_1/d_{i-1},\ldots,a_{i-1}/d_{i-1} \rangle$ for $i=2,3$; that is, $H(P_\infty)$ is telescopic.
\end{proof}
Proposition \ref{rimarco} implies that $H(P_\infty)$ is symmetric, from \cite[Lemma 6.5]{KP}.
This also follows from the fact that the divisor $(2g-2)P_\infty$ is canonical; see \cite[Lemma 3.8]{GOS} and \cite[Remark 4.4]{KP}.
In the following, Proposition \ref{campillo} is used to reduce the direct computation of $d_{ORD}(C_\ell(P_\infty))$ with $\rho_{\ell}>2g$, only to those cases for which $\rho_{\ell+1}\ne2g-1+e$ for any $e \in H(P_\infty) \setminus \{0\}$. Since the cases in which $\rho_{\ell+1}=(0,0,k)$ for $k<2m$ or $\rho_{\ell+1}=(i,j,k)$ for $k<m$ have been already studied, they can be excluded.
\begin{proposition} \label{farran} Let $\rho_{\ell} \in H(P_\infty)$ with $\rho_{\ell}>2g$ and $\rho_{\ell+1}=(i,j,k)$ and $k \geq m$. If $\rho_{\ell+1} \ne (0,0,k)$ for any $k \in [m,2m)$, then $\rho_{\ell+1}-2g+1 \not\in H(P_\infty)$ if and only if $\rho_{\ell+1}=(0,1,k)$ for some $k \in [m,2m)$.\end{proposition}
\begin{proof}
Write $k=m+s$ for some $s \geq 0$. We prove the claim using a case-by-case analysis with respect to the values of $i$ and $j$. We recall that $2g-1=(2^{n+1}+2^{n}-4)-1=9m-8$.
\ \\ \\
{\bf Case 1}: $i=1$. Clearly, $\rho_{\ell+1}=3m+2jb+8m+8s$.
\begin{itemize}
\item If $j=0$, then $\rho_{\ell+1}=3m+8m+8s=(9m-8)+(2m+8(s+1))=2g-1+e$. Writing $e=(0,1,s+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur.
\item If $j=1$, then $\rho_{\ell+1}=3m+2m+8m+8s=(9m-8)+(4m+8(s+1))=2g-1+e$. Writing $e=(0,2,s+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur.
\item If $j=2$, then $\rho_{\ell+1}=3m+4m+8m+8s=(9m-8)+(6m+8(s+1))=2g-1+e$. Writing $e=(0,3,s+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur.
\item If $j=3$, then $\rho_{\ell+1}=3m+6m+8k=(9m-8)+(8(k+1))=2g-1+e$. Writing $e=(0,0,k+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur.
\end{itemize}
\ \\
{\bf Case 2}: $i=0$. Clearly, $\rho_{\ell+1}=2jb+8k=2jb+8m+8s$.
\begin{itemize}
\item If $j=0$, then in particular we can write $k=2m+t$ for $t \geq 0$, since $k \geq 2m$. Thus, $\rho_{\ell+1}=16m+8t=(9m-8)+(7m+8(t+1))=2g-1+e$. Writing $e=(1,2,t+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur.
\item If $j=1$, then $\rho_{\ell+1}=2m+8k$. We first assume that $k \geq 2m$ and so that $k=2m+t$ for some $t \geq 0$. In this case $\rho_{\ell+1}=2m+16m+8t=(9m-8)+(9m+8(t+1))=2g-1+e$. Writing $e=(1,3,t+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur. Thus, $k \in [m,2m)$. In this case, $\rho_{\ell+1}=2m+8m+8s=(9m-8)+(m+8(s+1))=2g-1+e$. By direct computation $e \not\in H(P_\infty)$ and the claim follows.
\item If $j=2$, then $\rho_{\ell+1}=4m+8m+8s=(9m-8)+(3m+8(s+1))=2g-1+e$. Writing $e=(1,0,s+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur.
\item If $j=3$, then $\rho_{\ell+1}=6m+8m+8s=(9m-8)+(5m+8(s+1))=2g-1+e$. Writing $e=(1,1,s+1)$ we have that $e \in H(P_\infty)$, so this case cannot occur.
\end{itemize}
\end{proof}
Since from Proposition \ref{rimarco} the Weierstrass semigroup $H(P_\infty)$ is symmetric, its conductor is $c=2g$; equivalently, its largest gap is $2g-1$. The following theorem shows that the exact value of $d_{ORD}(C_{\ell}(P_\infty))$ is known for $\rho_{\ell+1} \geq 4g$; see \cite[Proposition 4.2 (iii)]{CF}.
\begin{theorem} \label{fengrao}
Let $H(P)$ be a Weierstrass semigroup. Then $d_{ORD}(C_{\ell}(P)) \geq \ell+1-g$ and equality holds if $\rho_{\ell+1} \geq 4g$.
\end{theorem}
According to the results obtained in the previous sections, Remark \ref{farran}, and Theorem \ref{fengrao}, to complete the computation of $d_{ORD}(C_\ell(P_\infty))$ for every $\rho_{\ell} \in H(P_\infty)$, only the case $\rho_{\ell} \in [2g,4g-1)$ with $\rho_{\ell+1}=(0,1,k)$ and $k \in [m,2m)$ has to be considered.
\begin{proposition} \label{perquant} Let $\rho_\ell \in H(P_\infty)$ be such that $\rho_{\ell}>2g$ and $\rho_{\ell+1}=(0,1,k)<4g$ for $k \in [m,2m)$. Then $$d_{ORD}(C_{\ell}(P_\infty))=\begin{cases} \nu_{\ell+5}=8k-7m+13, \ \ \textrm{if} \ \ k < \frac{9m-11}{8}, \\ \nu_{\ell+3}=8k-7m+11, \ \ \textrm{if} \ \ \frac{9m-11}{8} \leq k < \frac{11m-9}{8}, \\ \nu_{\ell+1}=8k-7m+9, \ \ \textrm{if} \ \ k \geq \frac{11m-9}{8}.\end{cases} $$ \end{proposition}
\begin{proof} Arguing as in the previous propositions one can prove that the value of $d_{ORD}(C_{\ell}(P_\infty))$ is obtained by $\nu_{\ell+5}$, $\nu_{\ell+3}$, and $\nu_{\ell+1}$, if $k < \frac{9m-11}{8}$, $\frac{9m-11}{8} \leq k < \frac{11m-9}{8}$, and $k \leq \frac{11m-9}{8}$ respectively. Since $\rho_{\ell+1} \geq 2g$, we have that $\rho_{\ell+t}=\rho_{\ell+1}+(t-1)$ for every $t \geq 1$.
Assume that $k<\frac{9m-11}{8}$. By direct checking $\rho_{\ell+6}=\rho_{\ell+1}+5=(1,3,\tilde k)$, where $\tilde k =k-\frac{7m-5}{8}$. Hence from Lemma \ref{i1}, $d_{ORD}(C_{\ell}(P_\infty))=8(k-\frac{7m-5}{8}+1)=8k-7m+13$, as $\tilde k < \frac{9m-11}{8}- \frac{7m-5}{8}<m$.
Assume that $\frac{9m-11}{8} \leq k < \frac{11m-9}{8}$. By direct checking $\rho_{\ell+4}=\rho_{\ell+1}+3=(1,0,\tilde k)$ where $\tilde k =k-\frac{m-3}{8}$. Hence $\tilde k \geq m-1$ and from Lemma \ref{i1}, $d_{ORD}(C_{\ell}(P_\infty))=2m=8k-7m+11$ if $\tilde k =m-1$, while $d_{ORD}(C_{\ell}(P_\infty))=2(\tilde k+1)+6(\tilde k -m+1)=8k-7m+11$ if $\tilde k \geq m$.
Assume that $k \geq \frac{11m-9}{8}$. By direct checking $\rho_{\ell+2}=\rho_{\ell+1}+1=(1,1,\tilde k)$ where $\tilde k =k-\frac{3m-1}{8}$. Hence $\tilde k \geq m-1$ and from Lemma \ref{i1}, $d_{ORD}(C_{\ell}(P_\infty))=4m=8k-7m+9$ if $\tilde k =m-1$, while $d_{ORD}(C_{\ell}(P_\infty))=4(\tilde k+1)+4(\tilde k -m+1)=8k-7m+9$ if $\tilde k \geq m$.\end{proof}
For $q\ne2$, we cannot determine $d_{ORD}(C_{\ell}(P_\infty)$ for all $\ell$. Yet, this is possible for certain $\ell$, as shown in the following propositions.
\begin{proposition}
If $\rho_{\ell+1}\leq (q-1)(q^n+1)$,
then
$$ d_{ORD}(C_{\ell}(P_\infty)) = j+1, $$
where $j\leq q-1$ satisfies $(j-1)(q^n+1) < \rho_{\ell+1} \leq j(q^n+1)$.
\end{proposition}
\begin{proof}
Since $H(P_\infty)$ is telescopic from Proposition \ref{rimarco}, we can apply \cite[Theorem 6.11]{KP}. The claim then follows because $q^n+1=\max\{\frac{q^3}{1},\frac{mq}{1},\frac{q^n+1}{1}\}$.
\end{proof}
\begin{proposition}
If $\frac{3}{2}(q-1)(q^{n+1}+\frac{1}{3}q^n-q^2-\frac{2}{3})-2<\ell\leq \frac{3}{2}(q-1)(q^{n+1}+q^n-q^2)-2$,
then
$$ d_{ORD}(C_{\ell}(P_\infty)) = \min\{\rho_t\mid \rho_t\geq \ell+1-g\}. $$
\end{proposition}
\begin{proof}
This is the claim of \cite[Theorem 6.10]{KP}.
\end{proof}
\subsection{Application for $q=2$: families of AG codes with relative Singleton defect going to zero}\label{Sec:Application1}
In this section, we assume that $q=2$ and provide two families of codes of type $C_{\ell}(P_\infty)$ in the cases $\rho_\ell=9m$ and $\rho_\ell=9m+8$, with relative Singleton defect going to zero as $n$ goes to infinity.
We denote by $\delta$ and $\Delta$ the Singleton defect and the relative Singleton defect of $C_\ell(P_\infty)$, respectively.
\
\begin{lemma}
Fix $n\geq5$ odd. Then $9m-1, \ 9m, \ 9m+1 \in H(P_\infty)$.
\end{lemma}
\begin{proof}
A direct computation shows that $9m-1=(0,3,\frac{2^n}{8})$, $9m=(1,3,0)$, and $9m+1=(0,2,\frac{5\cdot2^{n-3}+1 }{3})$, thus the claim follows.
\end{proof}
We now assume that $\rho_\ell=9m$. Since $\rho_{\ell+1}=9m+1=(0,2,\frac{5\cdot2^{n-3}+1 }{3})$ the following result follows from Lemma 1.3.
\begin{corollary} Assume that $\rho_\ell=(1,3,0)$. Then $\nu_\ell=3 \cdot \big(\frac{5 \cdot 2^{n-3}+1}{3}+1 \big) \geq 24$.
\end{corollary}
\begin{proposition} \label{9m} The code $C_\ell(P_\infty)$ is an $[N,k,d]_{2^{2n}}$-linear code with
\begin{itemize}
\item $N=(3m-1)^2+(3m-1)(9m-7)$,
\item $k=N-\frac{9m+9}{2}$,
\item $d \geq d_{ORD}(C_\ell(P_\infty)) = 16$,
\item $\delta \leq N-k+1- d_{ORD}(C_\ell(P_\infty))=\frac{9m-21}{2}$,
\item $\Delta =\frac{\delta}{N} \leq \frac{9m-21}{2(3m-1)(12m-8)}$; hence, $\Delta$ goes to zero as $n$ goes to infinity.
\end{itemize}
\end{proposition}
\begin{proof}
Since $GGS(q,n)$ is $\mathbb{F}_{2^{2n}}$-maximal, we have
$$N=(2^{2n}+1+2g(GGS(q,n))2^n)-1=2^{2n}+2^n(9m-7)=(3m-1)^2+(3m-1)(9m-7);$$
the last equality follows from $m=(2^n+1)/3$.
Since $C_\ell(P_\infty)=C^{\perp}(\overline D,\rho_\ell P_\infty)$, $k=N-\tilde{k}$ where $\tilde{k}$ is the dimension of $C(\overline D,\rho_\ell P_\infty)$. As $\deg(\rho_\ell P_\infty)>2g(GGS(q,n))-2$, from the Riemann-Roch Theorem follows
$$k=N-\tilde{k}=N-(\deg(\rho_\ell P_\infty)+1-g(GGS(q,n)))=N-\left(9m+1-\frac{9m-7}{2}\right)=r-\frac{9m+9}{2}.$$
By Lemma \ref{dord7}, $d_{ORD}(C_\ell(P_\infty))\geq 16$. To prove the claim is sufficient to show that there exists $\rho_s \geq \rho_\ell$ such that $\nu_s=16$. To this end we take $\rho_{s+1}=(1,3,1)=9m+8>9m+1$. From Lemma 1.1, $\nu_s=2(b+1)(c+1)=2(3+1)(1+1)=16$ and the claim follows. Now the claim on $\delta$ and $\Delta$ follows by direct computation.
\end{proof}
We now assume that $\rho_\ell=9m+8$, so that $\rho_{\ell+1}=9m+9=(0,2,\frac{5\cdot2^{n-3}+1}{3}+1)=(0,2,\frac{5m+9}{8})$.
Arguing as in the proof of Proposition \ref{9m} and using Lemma \ref{dord7}, the following result is obtained.
\begin{proposition} The code $C_\ell(P_\infty)$ is an $[r,k,d]_{2^{2n}}$-linear code with
\begin{itemize}
\item $r=(3m-1)^2+(3m-1)(9m-7)$,
\item $k=r-\frac{9m+25}{2}$,
\item $d \geq d_{ORD}(C_\ell(P_\infty)) = 2 \big(\lceil \frac{6m+9}{8} \rceil +1 \big)$,
\item $\delta \leq r-k+1- d_{ORD}(C_\ell(P_\infty))=\frac{9m+25}{2} -2 \big(\lceil \frac{6m+9}{8} \rceil +1 \big) <\frac{6m+21}{2}$,
\item $\Delta =\frac{\delta}{r
< \frac{6m+21}{2(3m-1)(12m-8)}$; hence, $\Delta$ goes to zero as $n$ goes to infinity.
\end{itemize}
\end{proposition}
\section{Weierstrass semigroup at $P_0$}\label{Sec:P_0}
In this section we describe the Weierstrass semigroup at $P_0$, and hence at any $\mathbb F_{q^2}$-rational affine point by Lemma \ref{Orbite}.
Consider the functions
\begin{equation}\label{functions}
\frac{y^r z^t}{x^s}, \qquad s\in [0,q^2-1], r\in [0,s], t\in \left[0,\left\lfloor\frac{sm(q+1)-rqm}{q^3} \right \rfloor \right].
\end{equation}
All these functions belong to $H(P_0)$. In fact,
$$\left(\frac{y^r z^t}{x^s}\right)=(mr+t-m(q+1)s)P_0+(m(q+1)s-mqr-tq^3)P_{\infty}$$
and by assumption
$$m(q+1)s-mqr-tq^3\geq 0.$$
\begin{proposition}
Let $t \in \left[0,\min\left(\left\lfloor\frac{sm(q+1)-rqm}{q^3} \right \rfloor ,m-1\right)\right]$ and
$$s\in [0,q], r\in [0,s]$$
or
$$s\in [q+1,q^2-1], r\in [0,q].$$
Then all the integers $mr+t-m(q+1)s$ are distinct.
\end{proposition}
\proof
Suppose $mr+t-m(q+1)s=m\overline r+\overline t-m(q+1)\overline s$. Then $t\equiv \overline t \pmod{m}$, which implies $t=\overline t$. Now, from $mr-m(q+1)s=m\overline r-m(q+1)\overline s$, $r\equiv \overline r \pmod{q+1}$, which yields $r=\overline r$ and $s=\overline s$.
\endproof
\begin{proposition}
Consider the following sets
$$\begin{array}{l}
\mathcal{L}_1 := \Big\{-t-rm+m(q+1)s \mid s\in [0,q], r\in [0,s], \\
\hspace{8 cm} t\in \left[0,((s-r)q+s)\frac{m-q^2+q-1}{q^3}+s-r \right]\Big\};\\
\\
\mathcal{L}_2 := \Big\{-t-rm+m(q+1)s \mid s\in [q+1,q^2-q], r\in [0,q],\\
\hspace{8 cm} t\in \left[0,((s-r)q+s)\frac{m-q^2+q-1}{q^3}+s-r \right]\Big\};\\
\\
\mathcal{L}_3 := \Big\{-t-rm+m(q+1)s \mid s\in [q^2-q+1,q^2-2],\\
\hspace{8.5 cm} r\in [0,q+s-q^2-1], t\in \left[0,m-1 \right]\Big\};\\
\\
\mathcal{L}_4 := \Big\{-t-rm+m(q+1)s \mid s\in [q^2-q+1,q^2-2], r\in [q+s-q^2,q],\\
\hspace{8 cm} t\in \left[0,((s-r)q+s)\frac{m-q^2+q-1}{q^3}+s-r \right]\Big\};\\
\\
\mathcal{L}_5 := \left\{-t+m(q+1)(q^2-1) \mid t\in \left[q^3,m-1 \right]\right\};\\
\\
\mathcal{L}_6 := \left\{-t-rm+m(q+1)(q^2-1) \mid r\in [1,q-2], t\in \left[0,m-1 \right]\right\};\\
\\
\mathcal{L}_7 := \Big\{-t-rm+m(q+1)(q^2-1) \mid r\in [q-1,q], \\
\hspace{6.5 cm} t\in \left[0,((q^2-1-r)q+q^2-1)\frac{m-q^2+q-1}{q^3}+q^2-1-r \right]\Big\}.\\
\end{array}
$$
Then each $\mathcal{L}_i$ is contained in $\mathcal{L}((2g-1)P_0)$.
\end{proposition}
\proof
By direct computations.
\endproof
Finally, we can give the description of the Weierstrass semigroup $H(P_0)$.
\begin{proposition}\label{SemigruppoP0}
$$\bigcup _{i=1}^7\mathcal{L}_i =H(P_0) \cap \{0,\ldots, 2g-1\}.$$
\end{proposition}
\proof
By direct computations, since
$$|\mathcal{L}_1|=\left(\frac{q^4 + 5q^3 + 8q^2 + 4q}{6}\right)\left(\frac{m-q^2+q-1}{q^3}\right)+\frac{(q+1)(q+2)(q+3)}{6},$$
$$|\mathcal{L}_2|=\left(\frac{q^6 - q^5 - q^4 - 3q^2 - 2q}{2}\right)\left(\frac{m-q^2+q-1}{q^3}\right)+\frac{q^5 - 2q^4 + 2q^3 - q^2 - 6q}{2},$$
$$|\mathcal{L}_4|=\left(\frac{3q^5 + 2q^4 - 20q^3 + q^2 + 8q + 12}{6}\right)\left(\frac{m-q^2+q-1}{q^3}\right)+\frac{3q^4 - q^3 - 18q^2 + 22q - 12}{6},$$
$$|\mathcal{L}_3|=\frac{m(q-2)(q-1)}{2}, \qquad |\mathcal{L}_5|=m-q^3, \qquad |\mathcal{L}_6|=\sum_{r=1}^{q-2} m=(q-2)m,$$
$$|\mathcal{L}_7|=\frac{m-q^2+q-1}{q^3}(2q^3-q-2)+2q^2-2q+1$$
and $\mathcal{L}_i \cap \mathcal{L}_j=\emptyset $ if $i \neq j$.
\endproof
Let $C_\ell(P_0)=C^{\perp}(\tilde D,\tilde\rho_\ell P_0)$, where $\tilde D$ is as in \eqref{Dbarra} and $\tilde\rho_\ell$ is the $\ell$-th positive non-gap at $P_0$.
In this case it has not been possible to determine $d_{ORD}(C_\ell(P_0))$ for any $n$ since we do not have a basis of the Weierstrass semigroup at $P_0$.
Nevertheless, Tables \ref{tabella1}, \ref{TabMeglio}, and \ref{tabquantum1} give evidence that for some specific values of $\ell$ the AG codes and AG quantum codes from $C_\ell(P_0)$ are better than $C_\ell(P_\infty)$, since the designed relative Singleton defect of $C_\ell(P_0)$ is smaller than the one of $C_\ell(P_\infty)$.
\section{AG codes on the GGS curve for $q=2$ and $n=5$}\label{Sec:tabelle}
In this section a more detailed description of the results obtained in the previous sections is given for the particular case $q=2$, $n=5$. Recall that in this case
$$ H(P_\infty)=\{0,8,16,22,24,30,32,33,38,40,41,44,46,48,49,52\}\cup \{54,\ldots,57\} $$
$$ \cup\,\{60\}\cup\{62,\ldots, 66\} \cup\{68\}\cup\{70,\ldots, 74\}\cup\{76,\ldots,82\}\cup\{84,\ldots,90\}\cup\{92,\ldots\}. $$
For the point $P_0$ (and hence for any $\mathbb F_{q^2}$-rational point), we have from Proposition \ref{SemigruppoP0}
$$ H(P_0)=\{0,21,22\}\cup\{29,\ldots,33\}\cup\{42,43,44\}\cup\{50,\ldots,55\}\cup\{58,\ldots,66\}\cup\{71,\ldots,77\}\cup\{79,\ldots\}. $$
Table \ref{tabella1} contains the parameters of the codes $C_{\ell_\infty}(P_\infty)$ and $C_{\ell_0}(P_0)$; in particular, their common length $N=3968$ and dimension $k$, their Feng-Rao designed minimum distance $d_{ORD}^{\infty}$ and $d_{ORD}^{0}$, their designed Singleton defects $\delta_\infty=N+1-k-d_{ORD}^{\infty}$ and $\delta_0=N+1-k-d_{ORD}^{0}$, and their designed relative Singleton defects $\Delta_{\infty}=\frac{\delta_{\infty}}{N}$ and $\Delta_{0}=\frac{\delta_{0}}{N}$.
\begin{center}
\tabcolsep = 0.5 mm
{\scriptsize
\begin{longtable}{|c|c||c|c|c|c||c|c|c|c|}
\caption[Codes $C_{\ell_\infty}(P_\infty)$ and $C_{\ell_0}(P_0)$, $q^n=2^5$]{Codes $C_{\ell_\infty}(P_\infty)$ and $C_{\ell_0}(P_0)$, $q^n=2^5$} \label{tabella1} \\
\hline
& & & & & & & & & \\[-2 mm]
$N$ & $k$ & $\rho_{\ell_\infty}$ & $d_{ORD}^{\infty}$ & $\delta_\infty \leq$ & $\Delta_\infty \leq$ & $\rho_{\ell_0}$ & $d_{ORD}^{0}$ & $\delta_0 \leq$ & $\Delta_0 \leq$\\
& & & & & & & & & \\[-2 mm]
\endfirsthead
\multicolumn{10}{c}%
{{ \tablename\ \thetable{} : continued from previous page}} \\
\hline
& & & & & & & & & \\[-2 mm]
$n$ & $k$ & $\rho_{\ell_\infty}$ & $d_{ORD}^{\infty}$ & $\delta_\infty \leq$ & $\Delta_\infty \leq$ & $\rho_{\ell_0}$ & $d_{ORD}^{0}$ & $\delta_o \leq$ & $\Delta_0 \leq$\\
& & & & & & & & & \\[-2 mm]
\endhead
\hline
\endfoot
\hline
\endlastfoot
\hline 3968 & 3966 & 8 & 2 & 1 & 0,0003 & 21 & 2 & 1 & 0,0003\\
\hline 3968 & 3965 & 16 & 2 & 2 & 0,0006 & 22 & 2 & 2 & 0,0006\\
\hline 3968 & 3964 & 22 & 2 & 3 & 0,0008 & 29 & 2 & 3 & 0,0008\\
\hline 3968 & 3963 & 24 & 2 & 4 & 0,0011 & 30 & 2 & 4 & 0,0011\\
\hline 3968 & 3962 & 30 & 2 & 5 & 0,0013 & 31 & 2 & 5 & 0,0013\\
\hline 3968 & 3961 & 32 & 2 & 6 & 0,0016 & 32 & 2 & 6 & 0,016\\
\hline 3968 & 3960 & 33 & 3 & 6 & 0,0016 & 33 & 3 & 6 & 0,016\\
\hline 3968 & 3959 & 38 & 3 & 7 & 0,0018 & 42 & 3 & 7 & 0,018\\
\hline 3968 & 3958 & 40 & 3 & 8 & 0,0021 & 43 & 3 & 8 & 0,021\\
\hline 3968 & 3957 & 41 & 3 & 9 & 0,0023 & 44 & 3 & 9 & 0,023\\
\hline 3968 & 3956 & 44 & 4 & 9 & 0,0023 & 50 & 3 & 10 & 0,026\\
\hline 3968 & 3955 & 46 & 4 & 10 & 0,0026 & 51 & 3 & 11 & 0,028\\
\hline 3968 & 3954 & 48 & 4 & 11 & 0,0028 & 52 & 3 & 12 & 0,031\\
\hline 3968 & 3953 & 49 & 4 & 12 & 0,0031 & 53 & 3 & 13 & 0,033\\
\hline 3968 & 3952 & 52 & 4 & 13 & 0,0033 & 54 & 3 & 14 & 0,036\\
\hline 3968 & 3951 & 54 & 4 & 14 & 0,0036 & 55 & 3 & 15 & 0,038\\
\hline 3968 & 3950 & 55 & 5 & 14 & 0,0036 & 58 & 4 & 15 & 0,038\\
\hline 3968 & 3949 & 56 & 5 & 15 & 0,0038 & 59 & 5 & 15 & 0,038\\
\hline 3968 & 3948 & 57 & 5 & 16 & 0,0041 & 60 & 5 & 16 & 0,041\\
\hline 3968 & 3947 & 60 & 5 & 17 & 0,0043 & 61 & 5 & 17 & 0,043\\
\hline 3968 & 3946 & 62 & 5 & 18 & 0,0046 & 62 & 5 & 18 & 0,046\\
\hline 3968 & 3945 & 63 & 5 & 19 & 0,0048 & 63 & 5 & 19 & 0,048\\
\hline 3968 & 3944 & 64 & 5 & 20 & 0,0051 & 64 & 5 & 20 & 0,051\\
\hline 3968 & 3943 & 65 & 5 & 21 & 0,0053 & 65 & 5 & 21 & 0,053\\
\hline 3968 & 3942 & 66 & 6 & 21 & 0,0053 & 66 & 6 & 21 & 0,053\\
\hline 3968 & 3941 & 68 & 6 & 22 & 0,0056 & 71 & 6 & 22 & 0,056\\
\hline 3968 & 3940 & 70 & 6 & 23 & 0,0058 & 72 & 6 & 23 & 0,058\\
\hline 3968 & 3939 & 71 & 6 & 24 & 0,0061 & 73 & 6 & 24 & 0,061\\
\hline 3968 & 3938 & 72 & 6 & 25 & 0,0064 & 74 & 6 & 25 & 0,064\\
\hline 3968 & 3937 & 73 & 6 & 26 & 0,0066 & 75 & 6 & 26 & 0,066\\
\hline 3968 & 3936 & 74 & 6 & 27 & 0,0069 & 76 & 6 & 27 & 0,069\\
\hline 3968 & 3935 & 76 & 6 & 28 & 0,0071 & 77 & 6 & 28 & 0,0071\\
\hline 3968 & 3934 & 77 & 8 & 27 & 0,0069 & 79 & 8 & 27 & 0,069\\
\hline 3968 & 3933 & 78 & 8 & 28 & 0,0071 & 80 & 8 & 28 & 0,0071\\
\hline 3968 & 3932 & 79 & 8 & 29 & 0,0074 & 81 & 8 & 29 & 0,0074\\
\hline 3968 & 3931 & 80 & 8 & 30 & 0,0076 & 82 & 8 & 30 & 0,0076\\
\hline 3968 & 3930 & 81 & 8 & 31 & 0,0079 & 83 & 8 & 31 & 0,0079\\
\hline 3968 & 3929 & 82 & 8 & 32 & 0,0081 & 84 & 8 & 32 & 0,0081\\
\hline 3968 & 3928 & 84 & 8 & 33 & 0,0084 & 85 & 8 & 33 & 0,0084\\
\hline 3968 & 3927 & 85 & 8 & 34 & 0,0086 & 86 & 8 & 34 & 0,0086\\
\hline 3968 & 3926 & 86 & 8 & 35 & 0,0089 & 87 & 8 & 35 & 0,0089\\
\hline 3968 & 3925 & 87 & 8 & 36 & 0,0091 & 88 & 8 & 36 & 0,0091\\
\hline 3968 & 3924 & 88 & 8 & 37 & 0,0094 & 89 & 8 & 37 & 0,0094\\
\hline 3968 & 3923 & 89 & 8 & 38 & 0,0096 & 90 & 8 & 38 & 0,0096\\
\hline 3968 & 3922 & 90 & 8 & 39 & 0,0099 & 91 & 8 & 39 & 0,0099\\
\hline 3968 & 3921 & 92 & 8 & 40 & 0,0101 & 92 & 8 & 40 & 0,0101\\
\hline 3968 & 3920 & 93 & 8 & 41 & 0,0104 & 93 & 8 & 41 & 0,0104\\
\hline 3968 & 3919 & 94 & 8 & 42 & 0,0106 & 94 & 8 & 42 & 0,0106\\
\hline 3968 & 3918 & 95 & 8 & 43 & 0,0109 & 95 & 8 & 43 & 0,0109\\
\hline 3968 & 3917 & 96 & 8 & 44 & 0,0111 & 96 & 8 & 44 & 0,0111\\
\hline 3968 & 3916 & 97 & 8 & 45 & 0,0114 & 97 & 8 & 45 & 0,0114\\
\hline 3968 & 3915 & 98 & 8 & 46 & 0,0116 & 98 & 8 & 46 & 0,0116\\
\hline 3968 & 3914 & 99 & 16 & 39 & 0,0099 & 99 & 9 & 46 & 0,0116\\
\hline 3968 & 3913 & 100 & 16 & 40 & 0,0101 & 100 & 16 & 40 & 0,0101\\
\hline 3968 & 3912 & 101 & 16 & 41 & 0,0104 & 101 & 21 & 36 & 0,0091\\
\hline 3968 & 3911 & 102 & 16 & 42 & 0,0106 & 102 & 22 & 36 & 0,0091\\
\hline 3968 & 3910 & 103 & 16 & 43 & 0,0109 & 103 & 22 & 37 & 0,0094\\
\hline 3968 & 3909 & 104 & 16 & 44 & 0,0111 & 104 & 22 & 38 & 0,0096\\
\hline 3968 & 3908 & 105 & 16 & 45 & 0,0114 & 105 & 22 & 39 & 0,0099\\
\hline 3968 & 3907 & 106 & 16 & 46 & 0,0116 & 106 & 22 & 40 & 0,0101\\
\hline 3968 & 3906 & 107 & 22 & 41 & 0,0104 & 107 & 22 & 41 & 0,0104\\
\hline 3968 & 3905 & 108 & 22 & 42 & 0,0106 & 108 & 22 & 42 & 0,0106\\
\hline 3968 & 3904 & 109 & 22 & 43 & 0,0109 & 109 & 22 & 43 & 0,0109\\
\hline 3968 & 3903 & 110 & 22 & 44 & 0,0111 & 110 & 22 & 44 & 0,0111\\
\hline 3968 & 3902 & 111 & 22 & 45 & 0,0114 & 111 & 26 &41 & 0,0104\\
\hline 3968 & 3901 & 112 & 22 & 46 & 0,0116 & 112 & 29 & 39 & 0,0099\\
\hline 3968 & 3900 & 113 & 24 & 45 & 0,0114 & 113 & 29 & 40 & 0,0101\\
\hline 3968 & 3899 & 114 & 24 & 46 & 0,0116 & 114 & 29 & 41 & 0,0104\\
\hline 3968 & 3898 & 115 & 30 & 41 & 0,0104 & 115 & 29 & 42 & 0,0106\\
\hline 3968 & 3897 & 116 & 30 & 42 & 0,0106 & 116 & 29 & 43 & 0,0109\\
\hline 3968 & 3896 & 117 & 30 & 43 & 0,0109 & 117 & 29 & 44 & 0,0111\\
\hline 3968 & 3895 & 118 & 30 & 44 & 0,0111 & 118 & 29 & 45 & 0,0114\\
\hline 3968 & 3894 & 119 & 30 & 45 & 0,0114 & 119 & 29 & 46 & 0,0116\\
\hline 3968 & 3893 & 120 & 30 & 46 & 0,0116 & 120 & 30 & 46 & 0,0116\\
\hline 3968 & 3892 & 121 & 32 & 45 & 0,0114 & 121 & 31 & 46 & 0,0116\\
\hline 3968 & 3891 & 122 & 32 & 46 & 0,0116 & 122 & 36 & 42 & 0,0106\\
\hline 3968 & 3890 & 123 & 33 & 46 & 0,0116 & 123 & 37 & 42 & 0,0106\\
\hline 3968 & 3889 & 124 & 38 & 42 & 0,0106 & 124 & 37 & 43 & 0,0109\\
\hline 3968 & 3888 & 125 & 38 & 43 & 0,0109 & 125 & 37 & 44 & 0,0111\\
\hline 3968 & 3887 & 126 & 38 & 44 & 0,0111 & 126 & 37 & 45 & 0,0114\\
\hline 3968 & 3886 & 127 & 38 & 45 & 0,0114 & 127 & 37 & 46 & 0,0116\\
\hline 3968 & 3885 & 128 & 38 & 46 & 0,0116 & 128 & 38 & 46 & 0,0116\\
\hline 3968 & 3884 & 129 & 40 & 45 & 0,0114 & 129 & 39 & 46 & 0,0116\\
\hline 3968 & 3883 & 130 & 40 & 46 & 0,0116 & 130 & 40 & 46 & 0,0116\\
\hline 3968 & 3882 & 131 & 41 & 46 & 0,0116 & 131 & 41 & 46 & 0,0116\\
\hline 3968 & 3881 & 132 & 44 & 44 & 0,0111 & 132 & 42 & 46 & 0,0116\\
\hline 3968 & 3880 & 133 & 44 & 45 & 0,0114 & 133 & 46 & 43 & 0,0109\\
\hline 3968 & 3879 & 134 & 44 & 46 & 0,0116 & 134 & 48 & 42 & 0,0106\\
\hline 3968 & 3878 & 135 & 46 & 45 & 0,0114 & 135 & 48 & 43 & 0,0109\\
\hline 3968 & 3877 & 136 & 46 & 46 & 0,0116 & 136 & 50 & 42 & 0,0106\\
\hline 3968 & 3876 & 137 & 48 & 45 & 0,0114 & 137 & 50 & 43 & 0,0109\\
\hline 3968 & 3875 & 138 & 48 & 46 & 0,0116 & 138 & 50 & 44 & 0,0111\\
\hline 3968 & 3874 & 139 & 49 & 46 & 0,0116 & 139 & 50 & 45 & 0,0114\\
\hline 3968 & 3873 & 140 & 52 & 44 & 0,0111 & 140 & 50 & 46 & 0,0116\\
\hline 3968 & 3872 & 141 & 52 & 45 & 0,0114 & 141 & 51 & 46 & 0,0116\\
\hline 3968 & 3871 & 142 & 52 & 46 & 0,0116 & 142 & 52 & 46 & 0,0116\\
\hline 3968 & 3870 & 143 & 54 & 45 & 0,0114 & 143 & 53 & 46 & 0,0116\\
\hline 3968 & 3869 & 144 & 54 & 46 & 0,0116 & 144 & 56 & 44 & 0,0111\\
\hline 3968 & 3868 & 145 & 55 & 46 & 0,0116 & 145 & 57 & 44 & 0,0111\\
\hline 3968 & 3867 & 146 & 56 & 46 & 0,0116 & 146 & 58 & 44 & 0,0111\\
\hline 3968 & 3866 & 147 & 57 & 46 & 0,0116 & 147 & 58 & 45 & 0,0114\\
\hline 3968 & 3865 & 148 & 60 & 44 & 0,0111 & 148 & 58 & 46 & 0,0116\\
\hline 3968 & 3864 & 149 & 60 & 45 & 0,0114 & 149 & 59 & 46 & 0,0116\\
\hline 3968 & 3863 & 150 & 60 & 46 & 0,0116 & 150 & 60 & 46 & 0,0116\\
\hline 3968 & 3862 & 151 & 62 & 45 & 0,0114 & 151 & 61 & 46 & 0,0116\\
\hline 3968 & 3861 & 152 & 62 & 46 & 0,0116 & 152 & 62 & 46 & 0,0116\\
\hline 3968 & 3860 & 153 & 63 & 46 & 0,0116 & 153 & 63 & 46 & 0,0116\\
\hline 3968 & 3859 & 154 & 64 & 46 & 0,0116 & 154 & 64 & 46 & 0,0116\\
\hline 3968 & 3858 & 155 & 65 & 46 & 0,0116 & 155 & 66 & 45 & 0,0114\\
\hline 3968 & 3857 & 156 & 66 & 46 & 0,0116 & 156 & 66 & 46 & 0,0116\\
\hline 3968 & 3856 & 157 & 68 & 45 & 0,0114 & 157 & 67 & 46 & 0,0116\\
\hline 3968 & 3855 & 158 & 68 & 46 & 0,0116 & 158 & 68 & 46 & 0,0116\\
\hline 3968 & 3854 & 159 & 70 & 45 & 0,0114 & 159 & 69 & 46 & 0,0116\\
\hline 3968 & 3853 & 160 & 70 & 46 & 0,0116 & 160 & 70 & 46 & 0,0116\\
\hline 3968 & 3852 & 161 & 71 & 46 & 0,0116 & 161 & 71 & 46 & 0,0116\\
\hline 3968 & 3851 & 162 & 72 & 46 & 0,0116 & 162 & 72 & 46 & 0,0116\\
\hline 3968 & 3850 & 163 & 73 & 46 & 0,0116 & 163 & 73 & 46 & 0,0116\\
\hline 3968 & 3849 & 164 & 74 & 46 & 0,0116 & 164 & 74 & 46 & 0,0116\\
\hline 3968 & 3848 & 165 & 76 & 45 & 0,0114 & 165 & 75 & 46 & 0,0116\\
\hline 3968 & 3847 & 166 & 76 & 46 & 0,0116 & 166 & 76 & 46 & 0,0116\\
\hline 3968 & 3846 & 167 & 77 & 46 & 0,0116 & 167 & 77 & 46 & 0,0116\\
\hline 3968 & 3845 & 168 & 78 & 46 & 0,0116 & 168 & 78 & 46 & 0,0116\\
\hline 3968 & 3844 & 169 & 79 & 46 & 0,0116 & 169 & 79 & 46 & 0,0116\\
\hline 3968 & 3843 & 170 & 80 & 46 & 0,0116 & 170 & 80 & 46 & 0,0116\\
\hline 3968 & 3842 & 171 & 81 & 46 & 0,0116 & 171 & 81 & 46 & 0,0116\\
\hline 3968 & 3841 & 172 & 82 & 46 & 0,0116 & 172 & 82 & 46 & 0,0116\\
\hline 3968 & 3840 & 173 & 84 & 45 & 0,0114 & 173 & 83 & 46 & 0,0116\\
\hline 3968 & 3839 & 174 & 84 & 46 & 0,0116 & 174 & 84 & 46 & 0,0116\\
\hline 3968 & 3838 & 175 & 85 & 46 & 0,0116 & 175 & 85 & 46 & 0,0116\\
\hline 3968 & 3837 & 176 & 86 & 46 & 0,0116 & 176 & 86 & 46 & 0,0116\\
\hline 3968 & 3836 & 177 & 87 & 46 & 0,0116 & 177 & 87 & 46 & 0,0116\\
\hline 3968 & 3835 & 178 & 88 & 46 & 0,0116 & 178 & 88 & 46 & 0,0116\\
\hline 3968 & 3834 & 179 & 89 & 46 & 0,0116 & 179 & 89 & 46 & 0,0116\\
\hline 3968 & 3833 & 180 & 90 & 46 & 0,0116 & 180 & 90 & 46 & 0,0116\\
\hline 3968 & 3832 & 181 & 92 & 45 & 0,0114 & 181 & 91 & 46 & 0,0116\\
\hline 3968 & 3831 & 182 & 92 & 46 & 0,0116 & 182 & 92 & 46 & 0,0116\\
\hline 3968 & $3968-\ell_{\infty}$ & $\rho_{\ell_\infty} \geq 183$ & $\ell_{\infty}-45$ & 46 & 0,0116 & $\rho_{\ell_0} \geq 183$ & $\ell_{0}-45$ & 46 & 0,0116 \\
\end{longtable}
}
\end{center}
Table \ref{TabMeglio} provides some examples in which codes of type $C_{\ell_0}(P_0)$ have better parameters than codes of type $C_{\ell_\infty}(P_\infty)$.
In particular, the length $n$ of the two codes is $3968$, the dimension $k_0$ and the Feng-Rao designed minimum distance $d_{ORD}^0$ of $C_{\ell_0}(P_0)$ are greater than or equal to the corresponding parameters $k_\infty$ and $d_{ORD}^\infty$ of $C_{\ell_\infty}(P_\infty)$, and the designed Singleton defect $\delta_0=n+1-k_0-d_{ORD}^0$ of $C_{\ell_0}(P_0)$ is strictly smaller than the designed Singleton defect $\delta_\infty=n+1-k_\infty-d_{ORD}^{\infty}$ of $C_{\ell_\infty}(P_\infty)$.
\begin{table}[htbp]
\begin{scriptsize}
\caption{Designed Singleton defect of $C_{\ell_0}(P_0)$ and $C_{\ell_\infty}(P_\infty)$, $q^n=2^5$}\label{TabMeglio}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline $\ell_0$ & 3 & 4 & 5 & 6 & 8 & 9 & 10 & 19 & 20 & 21 & 22 & 23 & 24 & 26 & 27\\
\hline $\ell_\infty$ & 4 & 5 & 6 & 7 & 9 & 10 & 11 & 20 & 21 & 22 & 23 & 24 & 25 & 27 & 28\\
\hline $\delta_\infty-\delta_0$ & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1\\
\hline\hline $\ell_0$ & 28 & 29 & 30 & 31 & 32 & 34 & 35 & 36 & 37 & 38 & 39 & 40 & 41 & 42 & 43\\
\hline $\ell_\infty$ & 29 & 30 & 31 & 32 & 33 & 35 & 36 & 37 & 38 & 39 & 40 & 41 & 42 & 43 & 44\\
\hline $\delta_\infty-\delta_0$ & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1\\
\hline\hline $\ell_0$ & 44 & 45 & 46 & 47 & 48 & 49 & 50 & 51 & 52 & 55 & 56 & 56 & 57 & 57 & 58\\
\hline $\ell_\infty$ & 45 & 46 & 47 & 48 & 49 & 50 & 51 & 52 & 53 & 56 & 56 & 57 & 57 & 58 & 58\\
\hline $\delta_\infty-\delta_0$ & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 5 & 6 & 6 & 7 & 6\\
\hline\hline $\ell_0$ & 58 & 59 & 59 & 60 & 60 & 61 & 61 & 62 & 63 & 64 & 65 & 66 & 66 & 67 & 67\\
\hline $\ell_\infty$ & 59 & 59 & 60 & 60 & 61 & 61 & 62 & 63 & 64 & 65 & 66 & 66 & 67 & 67 & 68\\
\hline $\delta_\infty-\delta_0$ & 7 & 6 & 7 & 6 & 7 & 6 & 1 & 1 & 1 & 1 & 1 & 4 & 6 & 7 & 6\\
\hline\hline $\ell_0$ & 68 & 68 & 69 & 77 & 77 & 78 & 88 & 88 & 89 & 89 & 90 & 90 & 91 & 91 &\\
\hline $\ell_\infty$ & 68 & 69 & 69 & 77 & 78 & 78 & 88 & 89 & 89 & 90 & 90 & 91 & 91 & 92 &\\
\hline $\delta_\infty-\delta_0$ & 5 & 4 & 5 & 4 & 4 & 4 & 2 & 3 & 4 & 3 & 2 & 3 & 4 & 3 &\\
\hline\hline $\ell_0$ & 92 & 92 & 93 & 93 & 94 & 99 & 99 & 100 & 100 & 101 & 101 & 102 & 110 & 110 & \\
\hline $\ell_\infty$ & 92 & 93 & 93 & 94 & 94 & 99 & 100 & 100 & 101 & 101 & 102 & 102 & 110 & 111 & \\
\hline $\delta_\infty-\delta_0$ & 2 & 1 & 2 & 2 & 1 & 2 & 2 & 2 & 2 & 2 & 2 & 1 & 1 & 1 & \\
\hline
\end{tabular}
\end{center}
\end{scriptsize}
\end{table}
\section{Quantum codes from one-point AG codes on the GGS curves}\label{Sec:Application2}
In this section we use families of one-point AG codes from the GGS curve to construct quantum codes. The main ingredient is the so called {\it CSS contruction} which enables to construct quantum codes from classical linear codes; see \cite[Lemma 2.5]{LGP}.
We denote by $q$ a prime power.
A $q$-ary quantum code $Q$ of length $N$ and dimension $k$ is defined to be a $q^k$-dimensional Hilbert subspace of a $q^N$-dimensional Hilbert space $\mathbb H=(\mathbb C^q)^{\otimes n}=\mathbb C^q\otimes\cdots\otimes\mathbb C^q$. If $Q$ has minimum distance $D$, then $Q$ can correct up to $\lfloor\frac{D-1}{2}\rfloor$ quantum errors.
The notation $[[N,k,D]]_q$ is used to denote such a quantum code $Q$.
For a $[[N,k,D]]_q$-quantum code the quantum Singleton bound holds, that is, the minimum distance satisfies $D\leq 1+(N-k)/2$.
The quantum Singleton defect is $\delta^Q:=N-k-2D+2\geq0$, and the relative quantum Singleton defect is $\Delta^Q:=\delta^Q/N$.
If $\delta^Q=0$, then the code is said to be quantum MDS.
For a detailed introduction on quantum codes see \cite{LGP} and the references therein.
\begin{lemma} \label{ccs} {\rm (CSS construction)}
Let $C_1$ and $C_2$ denote two linear codes with parameters $[N,k_i,d_i]_q$, $i=1,2$, and assume that $C_1 \subset C_2$. Then there exists an $[[N,k_2-k_1,D]]_q$ code with $D=\min\{wt(c) \mid c \in (C_2 \setminus C_1) \cup (C_1^\perp \setminus C_2^\perp)\}$, where $wt(c)$ is the weight of $c$.
\end{lemma}
We consider the following \textit{general t-point construction} due to La Guardia and Pereira; see \cite[Theorem 3.1]{LGP}. It is a direct application of Lemma \ref{ccs} to AG codes.
\begin{lemma} \label{lem1} {\rm (General t-point construction)}
Let $\mathcal X$ be a nonsingular curve over $\mathbb F_q$ with genus $g$ and $N+t$ distinct $\mathbb F_q$-rational points, for some $N,t>0$.
Assume that $a_i,b_i$, $i=1,\ldots,t$, are positive integers such that $a_i \leq b_i$ for all $i$ and $2g-2 < \sum_{i=1}^{t} a_i < \sum_{i=1}^t b_i < N$. Then there exists a quantum code with parameters $[[N,k,D]]_{q}$ with $k=\sum_{i=1}^{t} b_i - \sum_{i=1}^{t} a_i$ and $D \geq \min \big\{ N - \sum_{i=1}^{t} b_i, \sum_{i=1}^{t} a_i - (2g-2)\big\}$.
\end{lemma}
Let $n\geq 5$ be an odd integer.
We apply Lemma \ref{lem1} to one-point codes on the GGS curve.
\begin{proposition} \label{qua}
Let $a,b\in\mathbb N$ be such that
$$(q-1)(q^{n+1}+q^n-q^2)-2<a<b<q^{2n+2}-q^{n+3}+q^{n+2}.$$
Then there exists a quantum code with parameters $[[N,b-a,D]]_{q^{2n}}$, where
$$N=q^{2n+2}-q^{n+3}+q^{n+2},$$
$$D \geq \min \left \{ q^{2n+2}-q^{n+3}+q^{n+2}-b, a- (q-1)(q^{n+1}+q^n-q^2)+2 \right \}.$$
\end{proposition}
\begin{proof}
Let $GGS(q,n)$ be the GGS curve with equations \eqref{GGS_equation}, genus $g$, and infinite point $P_\infty$.
Consider the divisors $\overline D$ as in \eqref{Dbarra}, $G_1=a P_\infty$, and $G_2=b P_\infty$.
Note that ${\rm supp} (G_1) \cap {\rm supp}(\overline D)={\rm supp}(G_2) \cap {\rm supp}(\overline D)=\emptyset$.
From Lemma \ref{lem1}, there exists a quantum code with parameters $[[N,b-a,D]]_{q^{2n}}$, where $D \geq \min \big \{N-b, a-(2g-2) \big \}=\min \big \{ q^{2n+2}-q^{n+3}+q^{n+2}-b, a- (q-1)(q^{n+1}+q^n-q^2)-2 \big \}$.
\end{proof}
Another application of the CSS construction can be obtained looking at the dual codes of the one-point codes from the GGS curve. Let $P \in GGS(q,n)$. Fix $a=\rho_\ell \in H(P)$ and $b=\rho_{\ell+s} \in H(P)$ with $C_2=C_\ell(P)=C_\ell$ and $C_1=C_{\ell+s}(P)=C_{\ell+s}$, where $s \geq 1$. Clearly $C_1 \subset C_2$, as $C_{\ell} \subsetneq C_{\ell+s}$ for every $s \geq 1$. The dimensions of $C_2$ and $C_1$ are $k_2=N-h_\ell$ and $k_1=N-h_{\ell+s}=N-h_\ell-s$ respectively, where $h_i$ denotes the number of non-gaps at $P$ which are smaller than or equal to $i$. Thus, $k_2-k_1=s$. According to the CSS construction, these choices induce an $[[N, s,D]]_{q^{2n}}$ quantum code, where $N=q^{2n+2}-q^{n+3}+q^{n+2}$ and $D=\min\{wt(c) \mid c \in (C_2 \setminus C_1) \cup (C_1^\perp \setminus C_2^\perp)\}=\min\{wt(c) \mid c \in (C_{\ell} \setminus C_{\ell+s}) \cup (C(D,G_1) \setminus C(D,G_2)) \},$ with $G_2=\rho_\ell P$ and $G_1=\rho_{\ell+s} P$.
In particular,
\begin{equation} \label{Dquant}
D \geq \min \{d_{ORD}(C_\ell), d_1\},
\end{equation}
where $d_1$ denotes the minimum distance of the code $C(D,G_1)$. Following this construction and using an improvement of Inequality \eqref{Dquant}, the next theorem is obtained.
\begin{theorem} \label{quant1}
Let $g=(q-1)(q^{n+1}+q^n-q^2)/2$ and $N=q^{2n+2}-q^{n+3}+q^{n+2}$.
For every $\ell\in\left[3g-1,N-g\right]$ and $s\in\left[1,N-2\ell\right]$, there exists a quantum code with parameters $[[N,s,D]]_{q^{2n}}$, where $D \geq \ell+1-g$.
\end{theorem}
\begin{proof}
Since $\ell\geq3g-1$, we have $\rho_{\ell+s}=g-1+\ell+s$, and hence $d_1\geq N-\deg(G_1)=N-\rho_{\ell+s}=N-\ell-s-g+1$.
From Theorem \ref{fengrao}, $d_{ORD}(C_\ell)=\ell+1-g$. Thus, $D \geq \min \{d_{ORD}(C_\ell), d_1\}=\ell+1-g$. The claim follows.
\end{proof}
For fixed $q$, we can construct as a direct consequence of Theorem \ref{quant1} families of quantum codes depending on $n$ such that their relative quantum Singleton defect goes to zero as $n$ goes to infinity.
An example is the following.
\begin{corollary} \label{quant2}
Let $g=(q-1)(q^{n+1}+q^n-q^2)/2$ and $N=q^{2n+2}-q^{n+3}+q^{n+2}$.
For every $\ell\in[3g-1,N-g]$, fix $s=N-2\ell$.
Then there exists a quantum code with parameters $[[N,s,D]]_{q^{2n}}$ with $D\geq\ell+1-g$, whose relative quantum Singleton defect $\Delta_n^Q=(N-s-2D+2)/N$ satisfies
$$\Delta_n^Q\leq\frac{2g}{N}=\frac{(q-1)(q^{n+1}+q^n-q^2)}{q^{2n+2}-q^{n+3}+q^{n+2}}.$$
Hence, $\lim_{n \rightarrow \infty}{\Delta_n^Q} = 0$.
\end{corollary}
Using the computation of $d_{ORD}(C_{\ell}(P_\infty))$ in Section \ref{Sec:d_ORD}, we produce infinite families of quantum codes in which the lower bound in \eqref{Dquant} is explicitely determined. We look at those cases for which \eqref{Dquant} reads $D \geq d_{ORD}(C_{\ell}(P_\infty))>\ell+1-g$ and this bound is better than the one stated in Theorem \ref{quant1}. According to Proposition \ref{farran}, we choose $\rho_\ell \in H(P_\infty)$ such that $\rho_{\ell+1}=(0,1,k)$ for some $k \in [m,2m)$.
\begin{proposition}
Let $q=2^n$ for $n \geq 5$ odd, $g=(q-1)(q^{n+1}+q^n-q^2)/2$, and $N=q^{2n+2}-q^{n+3}+q^{n+2}$.
Let $\ell\in[g,3g-1]$ be such that $\rho_{\ell+1}\in H(P_\infty)$ is of type $(0,1,k)$ for some $k \in [m,2m)$. Let $s\in[1,N-2\ell-5]$.
Then there exists a quantum code with parameters $[[N,s,D]]_{q^{2n}}$ where
$$D \geq \ell+1-g+\begin{cases} 5, & \textrm{if} \ k<m \ \textrm{or} \ m \leq k <\frac{9m-11}{8}, \\ 3, & \textrm{if} \ \frac{9m-11}{8} \leq k < \frac{11m-9}{8}, \\ 1, & \textrm{if} \ \frac{11m-9}{8} \leq k. \end{cases}$$
\end{proposition}
\begin{proof} Arguing as in the proof of Theorem \ref{quant1}, we have that $d_1 \geq N-\ell-s-g+1$. Thus, from Proposition \ref{perquant} and Lemma \ref{dord6}, Inequality \eqref{Dquant} reads $$D \geq d_{ORD}(C_{\ell}(P_\infty))=\begin{cases} 8k-7m+13, & \textrm{if} \ k<m \ \textrm{or} \ m \leq k <\frac{9m-11}{8}, \\ 8k-7m+11, & \textrm{if} \ \frac{9m-11}{8} \leq k < \frac{11m-9}{8}, \\ 8k-7m+9, & \textrm{if} \ \frac{11m-9}{8} \leq k. \end{cases}$$ Since $\ell+1-g=\rho_{\ell+1}-2g+1=2m+8k-(9m-7)+1=8k-7m+8$, the claim follows. \end{proof}
\section{Convolutional codes from one-point AG codes on the GGS curves}\label{Sec:Application3}
In this section we use a result due to De Assis, La Guardia, and Pereira \cite{ALGR} which allows to construct unit-memory convolutional codes with certain parameters $(N,k,\gamma;m,d_f)_q$ starting from AG codes.
Consider the polynomial ring $R=\mathbb {F}_q[X]$. A convolutional code $C$ is an $R$-submodule of rank $k$ of the module $R^N$.
Let $G(X)=(g_{ij}(X))\in\mathbb {F}_q[X]^{k\times N}$ be a generator matrix of $C$ over $\mathbb {F}_q[X]$, $\gamma_i=\max\{\deg g_{ij}(X)\mid1\leq j\leq N\}$, $\gamma=\sum_{i=1}^k \gamma_i$, $m=\max\{\gamma_i\mid1\leq i\leq k\}$, and $d_f$ be the minimum weight of a word $c\in C$.
Then we say that $C$ has length $N$, dimension $k$, degree $\gamma$, memory $m$, and free distance.
If $m=1$, $C$ is said to be a unit-memory convolutional code.
In this case we use for $C$ the notation $(N,k,\gamma;m,d_f)_q$.
For a detailed introduction on convolutional codes see \cite{ALGR,RS1999} and the references therein.
\begin{lemma} \label{lem21}{\rm (\!\!\cite[Theorem 3]{ALGR})}
Let $\mathcal X$ be a nonsingular curve over $\mathbb F_q$ with genus $g$. Consider an AG code $C^{\bot}(D,G)$ with $2g-2<\deg(G)<N$. Then there exists a unit-memory convolutional code with parameters $(N,k-\ell,\ell;1,d_f \geq d)_q$, where $\ell \leq k/2$, $k=\deg(G)+1-g$ and $d \geq N-\deg(G)$.
\end{lemma}
We apply Lemma \ref{lem21} to one-point AG codes from the GGS curve.
\begin{proposition}\label{finiamo}
Consider the $\mathbb F_{q^{2n}}$-maximal GGS curve $GGS(q,n)$ and let $\rho_\ell \in H(P_\infty)$ be such that $(q-1)(q^{n+1}+q^n-q^2)-2<\rho_\ell<N$, where $N=q^{2n+2}-q^{n+3}+q^{n+2}$. Then there exists a unit-memory convolutional code with parameters $(N, k- s,s;1, d_f \geq d_{ORD}(C_\ell(P_\infty)))$, where $k=\rho_{\ell}+1-\frac{(q-1)(q^{n+1}+q^n-q^2)}{2}$ and $s \leq k/2$.
\end{proposition}
\begin{proof}
The result follows from Lemma \ref{lem21}. The inequality $d_f \geq d_{ORD}(C_\ell(P_\infty))$ follows from $d_f \geq d$ and Theorem \ref{fengrao} applied to the dual code $C_\ell(P_\infty)$.
\end{proof}
In particular, Theorem \ref{fengrao} yields the following corollary.
\begin{corollary}
Consider the $\mathbb F_{q^{2n}}$-maximal GGS curve $GGS(q,n)$ and let $\rho_\ell \in H(P_\infty)$ be such that $(q-1)(q^{n+1}+q^n-q^2)-2<\rho_\ell<N$, where $N=q^{2n+2}-q^{n+3}+q^{n+2}$ and $\ell \geq 3\frac{(q-1)(q^{n+1}+q^n-q^2)}{2}$. Then there exists a unit-memory convolutional code with parameters $(N, k- s,s;1, d_f)$, where $k=\rho_{\ell}+1-\frac{(q-1)(q^{n+1}+q^n-q^2)}{2}$, $s \leq k/2$, and $d_f \geq \ell+1-\frac{(q-1)(q^{n+1}+q^n-q^2)}{2}$.
\end{corollary}
\section{The Automorphism group of $C(\overline D,\ell P_{\infty})$}\label{Sec:Auto}
In this section we investigate the automorphism group of the code $C(\overline D,\ell P_{\infty})$, where $\overline D$ is as in \eqref{Dbarra}.
\begin{lemma}\label{Orbite}
The automorphism group $\mathrm{Aut}(GGS(q,n))$ has exactly two short orbits on $GGS(q,n)$; one consists of $P_\infty$, the other consists of the $q^3$ $\mathbb{F}_{q^2}$-rational points other than $P_\infty$.
\end{lemma}
\begin{proof}
From \cite{GMP,GOS}, $\mathrm{Aut}(GGS(q,n))=Q\rtimes\Sigma$, where $Q=\{Q_{a,b}\mid a,b\in\mathbb{F}_{q^2},a^q+a=b^{q+1}\}$ and $\Sigma=\langle g_{\zeta} \rangle$, with
\begin{equation}\label{AutDef} Q_{a,b}= \begin{pmatrix} 1 & b^q & 0 & a \\ 0 & 1 & 0 & b \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix},\quad
g_{\zeta} = \begin{pmatrix} \zeta^{q^n+1}&0&0&0 \\ 0&\zeta^{\frac{q^n+1}{q+1}}&0&0 \\ 0&0&\zeta&0 \\ 0&0&0&1 \end{pmatrix}, \end{equation}
$\zeta$ a primitive $(q^n+1)(q-1)$-th root of unity.
Therefore, $\mathrm{Aut}(GGS(q,n))$ fixes $P_\infty$. Also, $\mathrm{Aut}(GGS(q,n))$ acts transitively on the $q^3$ affine points of $GGS(q,n)$ having zero $Z$-coordinate, which coincide with the $\mathbb F_{q^2}$-rational points of $GGS(q,n)$ other than $P_\infty$.
Suppose $\mathrm{Aut}(GGS(q,n))$ has another short orbit $\mathcal O$.
Since $GGS(q,n)$ has zero $p$-rank and $\mathrm{Aut}(GGS(q,n))$ fixes $P_\infty$, $\mathcal O$ is tame. Hence, by Schur-Zassenhaus Theorem \cite[Theorem 9.19]{Rose}, the stabilizer of a point $P\in\mathcal O$ is contained up to conjugation in $\Sigma$.
This is a contradiction, as $\Sigma$ acts semiregularly out of the plane $Z=0$.
\end{proof}
Note from \eqref{AutDef} that $\mathrm{Aut}(GGS(q,n))$ is defined over $\mathbb F_{q^{2n}}$.
Let $\pi_a$ be the plane $Z=a$. The points of $\pi_0\cap GGS(q,n)$ are exactly the $q^3+1$ $\mathbb{F}_{q^2}$-rational points of $GGS(q,n)$, while all coordinates of any point of $GGS(q,n)\setminus\pi_0$ are not in $\mathbb{F}_{q^2}$.
The group $\Sigma$ fixes all points in $\pi_0\cap GGS(q,n)$ and acts semiregularly on the planes $\pi_a$,
while the group Q acts transitively on $\pi_0\cap GGS(q,n)$ and fixes $GGS(q,n)\cap\pi_a$ for all $a$.
Also, $Q$ acts faithfully on the Hermitian curve $\mathcal{H}_q:Y^{q+1}=X^q+X$ by $(X,Y,T)\mapsto\bar Q\cdot(X,Y,T)$, where $\bar Q$ is obtained from $Q$ deleting the third row and column.
\begin{proposition}
The automorphism group of ${C}(\overline D,\ell P_{\infty})$ contains a subgroup isomorphic to
$$ (\mathrm{Aut}(GGS(q,n))\rtimes\mathrm{Aut}(\mathbb{F}_{q^{2n}}))\rtimes\mathbb{F}_{q^{2n}}^{*}. $$
\end{proposition}
\begin{proof}
The set $S_{\sigma}$ of points of $GGS(q,n)$ fixed by a non-trivial automorphism $\sigma$ of $\mathrm{Aut}_{\mathbb{F}_{q^{2n}}}(GGS(q,n))=\mathrm{Aut}(GGS(q,n))$ has size $N_{\sigma}\leq q^3+1$.
In fact, if $\sigma\notin Q$, then $S_{\sigma}\subseteq\pi_0$.
If $\sigma\in Q$, then from $\sigma(P_\infty)=P_\infty$ we have that the induced automorphism $\bar\sigma\in\mathrm{Aut}(\mathcal{H}_q)$ fixes only $\mathbb{F}_{q^2}$-rational points of $\mathcal H_q$; hence, $\sigma$ fixes only $\mathbb{F}_{q^2}$-rational points of $GGS(q,n)$, that is, $S_{\sigma}\subseteq\pi_0$.
Since $|GGS(q,n)\cap\pi_0|=q^3+1$, $N_{\sigma}\leq q^3+1$.
Now the claim follows from \cite[Proposition 2.3]{BMZ}.
\end{proof}
\begin{proposition}
If $q^n+1 \leq \ell \leq q^{n+2}-q^3$ and $\{\ell,\ell-1\}\subset H(P_\infty)$, then
$$ \mathrm{Aut}({C}(\overline D,\ell P_{\infty})) \cong (\mathrm{Aut}(GGS(q,n))\rtimes\mathrm{Aut}(\mathbb{F}_{q^{2n}}))\rtimes\mathbb{F}_{q^{2n}}^{*}. $$
\end{proposition}
\begin{proof}
We apply \cite[Theorem 3.4]{GK2008}.
\begin{itemize}
\item The divisor $G=\ell P_\infty$ is effective.
\item A plane model of degree $q^n+1$ for $GGS(q,n)$ is
\begin{equation}\label{modello} \Pi(GGS(q,n)):\quad Z^{q^n+1} = X^{q^3}+X - (X^q+X)^{q^2-q+1}. \end{equation}
In fact, $Z^{m(q+1)}=Y^{q+1}h(X)^{q+1}=X^{q^3}+X - (X^q+X)^{q^2-q+1}$; also, Equation \eqref{modello} is irreducible since it defines a Kummer extension $\mathbb K(x,z)/\mathbb K(x)$ totally ramified over the pole of $x$.
Therefore, $\mathbb K(GGS(q,n))=\mathbb K(x,z)$, and $x,z\in\mathcal L(G)$ from the assumption $\ell\geq q^n+1$.
\item The support of $D$ is preserved by the Frobenius morphism $\varphi:(x,z)\mapsto(x^p,z^p)$, since $\varphi(P_\infty)=P_\infty$ and ${\rm supp}(D)=GGS(q,n)(\mathbb F_{q^{2n}})\setminus\{P_\infty\}$.
\item Let $N$ be the length of ${C}(\overline D,\ell P_{\infty})$. Then the condition $N>\deg(G)\cdot\deg(\Pi(GGS(q,n)))$ reads
$$ q^{2n+2}-q^{n+3}+q^{n+2} > \ell (q^n+1), $$
which is implied by the assumption $\ell\leq q^{n+2}-q^3$.
\item
\begin{itemize}
\item If $P=P_\infty$, then $\mathcal L(G)\ne\mathcal L(G-P)$ since $\ell\in H(P_\infty)$.
\item If $P\ne P_\infty$, then $1\in\mathcal L(G)\setminus\mathcal L(G-P)$.
\item If $P=Q=P_\infty$, then $\mathcal L(G-P)\ne\mathcal L(G-P-Q)$ since $\ell-1\in H(P_\infty)$.
\item If $P=P_\infty$ and $Q\ne P_\infty$, then $1\in\mathcal L(G-P)\setminus\mathcal L(G-P-Q)$.
\item If $P\ne P_\infty$ and $Q=P_\infty$, then $f-\mu\in\mathcal L(G-P)\setminus\mathcal L(G-P-Q)$, where $f\in\mathcal L(G)$ has pole divisor $\ell P_\infty$ and $\mu=f(P)$.
\item If $P,Q\ne P_\infty$ and $P\ne Q$, choose $f=z-z(P)$ or $f=x-x(P)$ according to $z(P)\ne z(Q)$ or $x(P)\ne x(Q)$; then $f\in\mathcal L(G-P)\setminus\mathcal L(G-P-Q)$.
\item If $P=Q\ne P_\infty$, then $z-z(P)\in\mathcal L(G-P)\setminus\mathcal L(G-P-Q)$.
\end{itemize}
\end{itemize}
Thus we can apply \cite[Theorem 3.4]{GK2008} to prove the claim.
\end{proof}
|
\section{Introduction}\enlargethispage{8pt}
Borides, carbides, and nitrides were among the early compound
superconductors discovered in the first half of the previous century\cite{Meissner1930}
(see, e.g., Ref.\ \onlinecite{Hott2016} for a recent review).
However, it was only in 2001, with the
discovery of superconductivity in MgB$_2$ at 39\,K,\cite{Akimitsu2001}
that researchers intensified the search for superconductivity in
other diborides. Based on a large number of studies, MgB$_2$ was identified as a two-band
two-gap superconductor. Its peculiar Fermi surface exhibits two-dimensional
hole-like cylinders from the $p_{x,y}$ bands, a hole-like tubular network
due to the bonding $p_z$ bands, and an electron-like tubular network
due to the antibonding $p_z$ bands.\cite{Kortus2001} Due to this electronic configuration and to a distinct anisotropy of the electron-phonon
interaction strength, the electronic excitation spectrum of MgB$_2$ adopts two gaps in the superconducting phase: a
large gap of 7.2\,meV in the $p_{x,y}$ ($\sigma$) bands, and a small gap
of 2.8\,meV in the bonding and antibonding $p_z$ ($\pi$) bands. At the same
time, an upper critical field anisotropy has been observed, with
${\mu_0H_{c2}}^{c} \sim 30$\,T and ${\mu_0H_{c2}}^{ab} \sim 3$\,T at zero
temperature.\cite{Sologubenko2002}
Besides alkaline-earth metals (such as Mg), diborides of other elements have
been proposed to be checked for superconductivity.
Because of the presence of partially filled 3$d$, 4$d$, and 5$d$ orbitals,
considered as promising for superconductivity, these new attempts involved
icosagens (Al) and various $d$-type transition metals (T). The latter (TB$_2$),
which are claimed to combine average coupling constants with comparable phonon
frequencies to MgB$_2$\cite{Heid2003} (due to the presence of light boron atoms),
were the natural candidates in this search. Unfortunately, these renewed efforts
proved unsuccessful and to date there are no reports of superconductivity for the
majority of TB$_2$ materials.
ZrB$_2$\ and HfB$_2$\ are two such non-superconducting refractory materials
with melting points of $\sim 3000$\,K, behaving essentially as Pauli
paramagnets down to low temperatures. The electrons in the 4$d$ shell of zirconium (Zr) and those in
the 5$d$ shell of hafnium (Hf) are less localized than those of the 3$d$ row.
In a recent study,\cite{Renosto2013} it was found that by replacing small amounts of Zr or Hf with V,
the resulting compounds Zr$_{1-x}$V$_{x}$B$_2$ and Hf$_{1-x}$V$_{x}$B$_2$
are superconductors. Maximum superconducting temperatures
$T_c = 8.33$\,K and 7.31\,K were reached in Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$, respectively, at the
upper solubility limit of V ($x \approx 0.04$). X-ray powder diffraction (XRD) patterns\cite{Renosto2013}
indicate that an increase in V doping does not change the in-plane lattice parameter $a$, while it reduces
the inter-layer distance $c$. At the same time, it has been shown that the structural and electronic properties of these compounds are
influenced by the presence of B vacancies.\cite{Dahlqvist2015}
In Ref.\ \onlinecite{Renosto2013}, the properties of superconducting Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ were investigated by means of macroscopic
techniques.\cite{Renosto2013} In this work we aimed at combining macro- and
microscopic techniques (including magnetometry, NMR, resistivity, and preliminary
$\mu^+$SR experiments) on Zr-based diborides and extend our study to include
the Hf-based compound. Since we succeeded in synthesizing
samples with less magnetic impurities (below 10 ppm
with respect to the previous ones,\cite{Renosto2013}
whenever relevant, a comparison between the two batches is included.
Our extensive data sets
allowed us to unravel clear analogies
and differences between the T-diborides Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\
and the well-known MgB$_2$. In Sec.~\ref{ssec:NMR} we show that spin-lattice relaxation processes
in T-diborides are two orders of magnitude slower that in MgB$_2$, indicating significantly different
electronic structures.
Combined magnetometry (Sec.~\ref{ssec:SQUID}) and resistivity (Sec.~\ref{ssec:res})
measurements were performed to evaluate the upper and lower critical fields, respectively. These compounds
prove to be extreme type-II superconductors, as reflected by the high values ($\sim$ 100) of the Ginzburg-Landau $\kappa = \nicefrac{\lambda}{\xi}$ parameter. The London penetration depth $\lambda$ was evaluated through $\mu^+$SR (Sec.~\ref{ssec:musr}) experiments and the coherence length $\xi$ via upper critical field measurements (Sec.~\ref{ssec:res}). Thanks to the efficient complementarity of these four techniques, we argue that, besides the qualitatively different Fermi surfaces of MgB$_2$ and transition-metal diborides, in both cases we are dealing with $s$-wave superconductors. While the rather high $T_c$ of MgB$_2$ is understood as a consequence of a favorable electronic structure and electron-phonon interaction, the drastic effect of V-doping in the T-diborides is still
rather surprising.
\begin{figure}[tbh]
\centering
\includegraphics[width=0.33\textwidth]{fig01}
\caption{\label{fig:ZrB2_structure}AlB$_2$ structure of the transition-metal diborides highlighting the graphene-like
B layers (magenta atoms) and the hexagonal metal layers (blue atoms).}
\end{figure}
\vspace{-5pt}
\section{Electronic structures of Z\lowercase{r}B$_2$ and H\lowercase{f}B$_2$\label{sec:fermisurf}}\enlargethispage{8pt}
Both pure and V-doped compounds crystallize in the layered AlB$_2$ structure with a $P6/mmm$ hexagonal
space group, where the Zr (or Hf) atoms and the B atoms occupy, respectively,
the $1a$ (0,0,0) and $2d$ (1/3, 2/3, 1/2) positions.
As in MgB$_2$, their crystal lattices are characterized by hexagonal metal
layers alternating with graphite-like B layers (see Fig.~\ref{fig:ZrB2_structure}).
Previous results of band-structure calculations\cite{Shein2008,Grechnev2009} and Fermi surface (FS) representations\cite{Shein2008} are available in the literature for both ZrB$_2$\ and HfB$_2$. Since, however, the printed version of the published FS of ZrB$_2$\ suffers from low quality, we chose to present the result of our own calculation in Fig.~\ref{fig:FS_ZB}, intended to serve for a comparison with the published FS of Mg$B_2$,\cite{Mazin2003} well known for its amazingly high critical temperature $T_c$ of the order of 40\,K for the onset of superconductivity. Our density-functional calculations use an LDA approach\cite{Perdew1992} based on the DMol$^3$ band-structure modeling program.\cite{Delley1990,Delley2000} The self-consistent field (SCF) was accomplished using the standard DNP\cite{Delley1990}
variational basis set and a $\Gamma$-centered $12\times 12\times 12$ $k$-mesh in the reciprocal cell. Pseudized scalar relativistic corrections\cite{Delley1998} were applied. The calculations are based on the experimental geometries published in Ref.~\onlinecite{Bsenko1974}.
The FS plot was generated using the XCrySDen program\cite{Kokalj1999} based on the DMol$^3$ output for a $99\times 99\times 99$ $k$-mesh.
The ZrB$_2$\ FS consists of four short cylindrical hole-type pockets around the $A$ point and four tripod-shaped electron-type pockets, each consisting of a triangular ring around the $K$ point and three elliptic extensions near the $\mathrm{\Gamma}AH$ plane. Since the FS of HfB$_2$\ exhibits essentially the same features, we refrained from presenting it in a separate figure. These features are really quite different from those of the FS of MgB$_2$. Most obvious is the reduction of the two almost two-dimensional hole sheets centered near the $A$ point for MgB$_2$ to one much smaller and 3-D type pocket for Zr- and Hf-diboride. Likewise the 3-D parts are quite different in shape for MgB$_2$. It is therefore not surprising that the two borides investigated here are less favorably conditioned for superconductivity than MgB$_2$ and indeed for both pure compounds, no superconductivity was detected above 1\,K. The band-structure calculations for the T-borides considered here indicate that in both cases the density of electronic states at the Fermi energy $N(E_\mathrm{F})$ is located in a region where $N(E)$ exhibits a pseudo-gap, i.e., is much reduced. It is thus remarkable that a very small (4--5\%) V-for-Zr or V-for-Hf substitution results in onsets of superconductivity up to approximately 8\,K. More details on the crystal structure and defect-induced phase instabilities of these types of compounds
are discussed in Ref.~\onlinecite{Renosto2013}.
\begin{figure}[tbh]
\centering
\includegraphics[width=0.35\textwidth]{fig02}
\caption{\label{fig:FS_ZB} Fermi surface of ZrB$_2$.}
\end{figure}
\vspace{-5pt}
\section{Experimental details\label{sec:details}}\enlargethispage{8pt}
Polycrystals of ZrB$_2$, HfB$_2$, Zr$_{0.96}$\-V$_{0.04}$\-B$_2$, and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\ were synthesized via boron carbide
reduction and structurally characterized as described in Ref.\ \onlinecite{Renosto2013}.
For our magnetic measurements, the samples in form of fine powders, with masses between
50 and 100\,mg, were sealed in Teflon (PTFE) tubes. The NMR investigations including line-shape and spin-lattice relaxation measurements were performed in an applied magnetic field of 3.505\,T, since higher fields would reduce the possibility to resolve
the quadrupolar effects. In our case, the $^{11}$B nucleus (spin
$I=\nicefrac{3}{2}$) proved to be the most suitable one, since it allows a direct comparison
between the four samples and it is four times more abundant than $^{10}$B.
The NMR signals were monitored by means of standard spin-echo sequences, consisting in
${\pi}/{2}$ and $\pi$ pulses of 2 and 4\,$\mu$s, respectively, with recycle delays
ranging from 10 to 100\,s, depending on the temperature ranging between 4 and 295\,K.
The NMR line-shapes were obtained via the fast Fourier Transform (FFT) of the echo
signal which, due to the high sensitivity of $^{11}$B, could be acquired using relatively few scans (from 4 to 2048).
The spin-lattice relaxation times $T_1$ were measured with the inversion recovery method,
using a $\pi$-${\pi}/{2}$-$\pi$ pulse sequence. The magnetometry measurements were performed by using a commercial
Magnetic Property Measurement System (MPMS XL) from Quantum Design,
equipped with a 7-T magnet and covering the temperature range from 2 to 400\,K. For the resistivity
measurements, the samples were densely packed in
cylindrical pellets with a diameter of 1.4\,mm and a thickness of 5 mm,
produced by applying high external pressures. The electrical contacts with a four-probe
configuration were made by means of an electrically-conducting silver epoxy.
Preliminary $\mu^+$SR measurements were made at the GPS spectrometer
of the S$\mu$S facility of Paul Scherrer Institute (PSI) in Villigen, Switzerland. The available
sample mass (300\,mg) was sufficient to stop the 4-MeV muons without additional
degraders and with a minimal background signal. To avoid pinning effects,
well known for distorting the vortex lattice (VL) in the superconducting phase of
MgB$_2$,\cite{Niedermayer2002} the transverse-field (TF) muon-spin rotation measurements
were made at the highest field available (0.6\,T).
\vspace{-5pt}
\section{Experimental results and discussion\label{sec:results}}\enlargethispage{8pt}
\subsection{Nuclear Magnetic Resonance\label{ssec:NMR}}
The ${}^{11}$B NMR lines of all the samples (both pure and V-doped)
were measured from 5 to 295\,K; typical data are shown in Fig.~\ref{fig:Diborides_NMR_lines_comparison}.
The reference ${}^{11}$B NMR frequency in an applied magnetic field of 3.505\,T was
evaluated to be $\nu_0 = 47.8844$\,MHz.
In our case, the NMR lines exhibit peaks which are very close to the reference, with an absolute positive shift
of only about 6\,kHz, corresponding to a Knight shift of 120\,ppm.
In the covered temperature range between 5 and 295\,K the ${}^{11}$B NMR lines practically
coincide (see Fig.~\ref{fig:Diborides_NMR_lines_comparison}), implying temperature-independent
Knight shifts for all the measured samples, compatible with the $\chi(T)$ plateaux
observed in the magnetometry data, measured under zero-field cooled
conditions (see Fig.~\ref{fig:chi_vs_T_10Oe_allsamples}).
The trend of the Knight shift in the superconducting phase could not be resolved because of the appreciable
width of the resonance signal.
As the line position, also the full width at half maximum (FWHM) is practically constant upon varying the
temperature. For the Zr-based samples its value is 13\,kHz (with 1\,kHz of additional broadening below $T_c$),
while for those containing Hf the width is 14\,kHz (+1\,kHz at low temperatures). The typical quadrupole splitting
$\nu_Q$ in ZrB$_2$\ and HfB$_2$\ is approximately the same, i.e. 54\,kHz. This relatively small value implies
a rather small electric-field gradient (EFG), especially if compared with MgB$_2$, for which
$\nu_Q \simeq 860$ kHz,\cite{Papavassiliou2001} which is a signature of a different electronic charge distribution,
confirmed by the different orbitals involved in the bonds, i.e. only $s$ and $p$ orbitals for MgB$_2$ and also $d$ orbitals for T-borides.
\begin{figure}[tbh]
\centering
\includegraphics[width=0.45\textwidth]{fig03}
\caption{\label{fig:Diborides_NMR_lines_comparison} ${}^{11}$B NMR lines in the four TB$_2$ compounds, respectively - from bottom to top, ZrB$_2$, Zr$_{0.96}$\-V$_{0.04}$\-B$_2$, HfB$_2$, and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$. The measurements were performed at 3.505\,T and for each sample the lines measured at 5 and at 295\,K are superposed, showing no relevant change. The dashed black line superposed on the ZrB$_2$\ spectra represents a simulation of a S = $\nicefrac{3}{2}$ powder pattern with small quadrupole splitting (reflected by the two shoulders around the central transition).}
\end{figure}
For typical powder spectra of a $I = \nicefrac{3}{2}$ nucleus with a small quadrupole splitting,
an analytical expression for the line-shape can be derived by considering the quadrupole term as a first-order
perturbation in the main Zeeman Hamiltonian.\cite{Goc1987, Cohen1954} As shown in Fig.~\ref{fig:Diborides_NMR_lines_comparison},
the experimental spectrum and the simulated line-shape (dashed line) for ZrB$_2$\ agree quite well.
The simulation of the powder spectrum, employing a Matlab code, performs the integration according to Euler's method over all
the possible orientations of the NMR line-shape factor\cite{Goc1987} $F$, i.e., the integration of the transition frequencies, derived from the quadrupolar theory and weighted using a Gaussian broadening function.\cite{Alderman1986, Hodgkinson2000}
From a quantitative analysis, the electric-field gradient (EFG) tensor can be evaluated using
\begin{figure}[tbh]
\centering
\includegraphics[width=0.45\textwidth]{fig04}
\caption{\label{fig:chi_vs_T_10Oe_allsamples} Magnetic susceptibility data [$\chi(T)$ in SI units]
at 1\,mT for ZrB$_2$, HfB$_2$, Zr$_{0.96}$\-V$_{0.04}$\-B$_2$, and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$. The undoped samples (and the doped ones above $T_c$) exhibit
diamagnetic behavior (plateaux). Small ZrB$_{12}$ impurities in ZrB$_2$\ induce a superconducting transition at $\sim 6$\,K. Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\ become superconductors at $T_c = 8.33$\,K and 7.31\,K, respectively.}
\end{figure}
\begin{equation} \label{eq:EFG}
eq = \frac{2I(2I-1)\,h\nu_Q}{3eQ},
\end{equation}
with $eq$ the largest EFG component (parallel to the applied magnetic field),
$I$ the nuclear spin, and $Q$ the quadrupole moment of the nucleus.
By considering the hexagonal symmetry of the AlB$_2$ structure,
we can assume that the in-plane anisotropy parameter $\eta = (V_{xx}-V_{yy})/V_{zz} = 0$ and, therefore,
$V_{xx} = V_{yy}$. Since the $V_{i,j}$ tensor is traceless, by evaluating $eq = V_{zz}$ from the simulated $\nu_Q$
data (see Eq.~\ref{eq:EFG}), we get $V_{zz} = 1.1\times10^{20}$\,Vm$^{-2}$ and $V_{xx} = V_{yy} = -5.5\times10^{19}$\,Vm$^{-2}$.
\begin{figure}[tbh]
\centering
\includegraphics[width=0.45\textwidth]{fig05}
\caption{\label{fig:Diborides_T1_compar} ${}^{11}$B NMR $T_1^{-1}(T)$ data
at 3.505\,T for all the investigated samples, compared with MgB$_2$ and AlB$_2$ data
(from Ref.\ \onlinecite{Baek2002}). From a standard fit,
assuming only one spin-lattice relaxation time,\cite{Mcdowell1995}
we obtain a linear trend, with Korringa constants as reported in Table~\ref{tab:summary}.
Materials with transition metal cations, such as Zr and Hf, exhibit slow spin-lattice relaxation processes,
two orders of magnitude slower than in MgB$_2$, whereas AlB$_2$ is an intermediate case.}
\end{figure}
As summarized in Table~\ref{tab:summary}, the Knight shift values are of the same order of magnitude as in MgB$_2$,\cite{Kotegawa2001,Papavassiliou2001,Baek2002,Pavarini2003}
but we note a difference of two orders of magnitude (!) in
the spin-lattice relaxation rates. The measured $T_1T$ value for MgB$_2$ is $1.8\times 10 ^2$\,sK which,
by considering its 70\,ppm Knight shift, implies an experimental Korringa constant $S_0 \equiv T_1TK^2 = 8.85\times 10^{-7}$\,sK,
approximately 3 times smaller than the theoretical value $S_{\mathrm{th}} = \hbar(\gamma_e/\gamma_{\mathrm{B}})^2 /(4\pi k_{\mathrm{B}}) = 2.56 \times 10^{-6}$\,sK. Upon V-doping, the very slow relaxation processes in T-diborides are non significantly
altered but, nevertheless, this apparently insignificant doping induces superconductivity at relatively high critical temperatures. Furthermore, due to strong covalent bonds, the largest contribution to the electronic density of states (DOS) at the Fermi level $D(E_F)$ is due to itinerant electrons in the boron layers. With respect to AlB$_2$ and even more so to MgB$_2$, $D(E_F)$ of our compounds is drastically reduced, as clearly confirmed by the spin-lattice relaxation data, in turn in good agreement with our theoretical calculations on the electronic structure of ZrB$_2$\ and HfB$_2$.
\begin{table}
\centering
\caption{\label{tab:summary} Key NMR parameters for the investigated samples compared with those of AlB$_2$ and MgB$_2$.}
\begin{ruledtabular}
\begin{tabular}{ l c c c }
Material & Shift (ppm) & $T_1T$ (10$^4$ s\,K) & $S_0$ (10$^{-4}$ s\,K) \\ \hline
ZrB$_2$ & 120 & 1.81 $\pm$ 0.07 & 2.60 $\pm$ 0.18 \\
Zr$_{0.96}$V$_{0.04}$B$_2$ & 120 & 1.89 $\pm$ 0.07 & 2.72 $\pm$ 0.11 \\
HfB$_2$ & 140 & 1.67 $\pm$ 0.05 & 3.27 $\pm$ 0.12 \\
Hf$_{0.97}$V$_{0.03}$B$_2$ & 140 & 1.56 $\pm$ 0.07 & 3.06 $\pm$ 0.09 \\
AlB$_2$ & -10 & 0.14 $\pm$ 0.04 & 0.009 $\pm$ 0.001 \\
MgB$_2$ & 70 & 0.018 $\pm$ 0.006 & 0.009 $\pm$ 0.001 \\
\end{tabular}
\end{ruledtabular}
\end{table}
The raw data, i.e., ${}^{11}$B NMR $T_1$ inverse saturation recovery curves, for the temperatures above $T_c$ were fitted by assuming the standard magnetization recovery formula for a single spin-lattice relaxation time.\cite{Mcdowell1995} However, due to the expected high anisotropy of $\mu_0H_{c2}$, and by analogy with MgB$_2$,\cite{Sologubenko2002} in powder samples we expect grains with different orientations, i.e.,
where the applied magnetic field lies in the $ab$ plane or is parallel to the $c$ axis. \textit{A priori} the orthogonal and parallel magnetic susceptibilities and the upper critical field depend on the direction of the field.
The existence of nonequivalent grains, due to the anisotropy of the upper critical field, is also confirmed by magnetometry (Fig.~\ref{fig:chi_vs_T_10Oe_allsamples}) and resistivity (Fig.~\ref{fig:Diborides_R_vs_T_all}) measurements, which show a superconducting width transition $\Delta T \sim 2.5$\,K, hence suggesting the persistence of inhomogenous domains. In a first approximation and following a procedure employed in the MgB$_2$ case,\cite{Baek2002, Pavarini2003}, we can fit the data, by assuming two relaxation processes (see Eq.~\ref{eq:IRfit}), related to the normal phase and to the superconducting phase, respectively:
\begin{equation} \label{eq:IRfit}
I(\tau_d) \propto \alpha \exp(\tau_d/T_{1s})^\beta +
(1 - \alpha) \exp(\tau_d/T_{1n})^\beta.
\end{equation}
Here $\tau_d$ is the time delay in the NMR pulse sequence, $T_{1s}$ and $T_{1n}$ the spin-lattice relaxation times
of the superconducting and normal grains, respectively, $\alpha$ the superconducting volume fraction and $\beta$
the stretching parameter (close to 1 in this case). We assume that $\alpha$ is a temperature-dependent
fit parameter, ranging from $\sim 1$ (in case of maximum superconducting fraction,
as evaluated from magnetometry data), down to 0. In the transition region the two plateaux are connected with a sigmoidal function.
The $T_1^{-1}(T)$ spin-lattice relaxation data above $T_c$ follow the linear behavior in $T$ of a simple metal (see Fig.~\ref{fig:Diborides_T1_compar}), where nuclear relaxation occurs mostly via interactions with the conduction electrons.
On the other hand, below $T_c$, the superconducting grains may exhibit two trends: a power-law with an integer exponent, typical of anisotropic superconductors, or an exponential trend, as expected for $s$-wave superconductors. If the sample is not perfectly homogeneous,
a second relaxation component could persist as a linear trend associated to normal grains, as shown in Fig.~\ref{fig:diborides_gap}.
The hypothesis of the two relaxation times is justified by the good fit results that we obtain for the evaluation of spin-lattice relaxation rates $R_1 = T_1^{-1}$.
A quantitative analysis is however hampered by a contribution to the relaxation from the flux vortex lines (i.e., their thermal motion).\cite{Jung2001, Rigamonti1998} Since we performed a field cooling (FC) measurement, the formation of a flux line lattice (FLL) with a regular arrangement of vortices, most likely with hexagonal symmetry, is expected. The geometric parameter of this lattice is the intervortex spacing $d(\mu_0H) = (2\Phi_0/3^{1/2}\mu_0H)^{1/2}$, which implies $d(3.505 T) = 26$ nm in our case. This value is approximately 6 times the diameter of the vortices $\xi$, as evaluated in Sec.~\ref{ssec:res}, implying that the measured
T$_1$ values consist of the sum of a slow contribution from outside the vortex cores and a faster contribution from the normal region within the vortices.\cite{Rigamonti1998}
We speculate that the considerable anisotropy of the upper critical field of MgB$_2$,\cite{Sologubenko2002} is also a characteristic of our materials. A rigorous confirmation would be obtained by relevant experiments on single crystals as, e.g., thermal conductivity measurements. In general, $s$-wave superconductors exhibit an exponential decrease of $T_1^{-1}(T)$ well below $T_c$, from which the gap value $\Delta$ can be extracted.
The appearance of a Hebel-Slichter coherence peak is usually interpreted as confirming the Cooper-pairing with spherical symmetry. In our case, the absence of a coherence peak does not rule out an $s$-wave parity, since the size of the peak can be significantly reduced by the pair-breaking mechanism in the presence of high fields.\cite{Masuda1969}
Due to the above-mentioned complexity of the relaxation processes and the quality of our data, it is impossible to extract the gap value $\Delta$ from the exponential decrease of the spin-lattice relaxation time. In any case a first evidence for $s$-wave superconductivity is the increasing deviation from a power-law behavior with an exponent 3 towards lower temperatures (see Fig.~\ref{fig:diborides_gap}).
To justify the similarities between the phonon-mediated $s$-wave superconductivity mechanism in MgB$_2$ and our V-doped samples we note that vanadium, given its $3d^3$ orbital, has one more electron, if compared to Zr ($4d^2$) and Hf ($5d^2$).
An analogous electron doping is confirmed in the case of MgB$_2$, where the $s$-states of Mg are pushed up by the B $p_z$ orbitals and, therefore, fully donate their electrons to the boron-derived conduction band.\cite{Kortus2001} This doping mechanism occurs also in the opposite direction (reduction of the $T_c$ value) in MgB$_2$. In this case, both the substitution of Mg with Li (hole doping) and of boron with carbon or Al (electron doping) reduce the $T_c$ of the material.\cite{Karpinski2008} In this case, it is claimed that the electrons fill the $\sigma$ band and holes occupy the $\pi$ band, therefore making charge compensation impossible. Furthermore, a recent paper\cite{Gil2013} supports the hypothesis of two gaps in Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ from critical current density $J_c$ measurements in different fields. In fact, $J_c$ can suitably be fitted by the sum of two contributions $J_1$ and $J_2$, respectively, related to the first and the second gap, following an exponential trend as a function of the applied magnetic field.
\begin{figure}[tbh]
\centering
\includegraphics[width=0.45\textwidth]{fig06}
\caption{\label{fig:diborides_gap} ${}^{11}$B NMR $1/T_1(T)$ data at 3.505\,T for Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ (main plot) and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\ (inset), representing
the two relaxation times of Eq.~\ref{eq:IRfit}, which result from fitting the inversion recovery curves.
Both superconducting samples exhibit a conventional metallic trend (red points) and a superconducting dropdown (magenta points); The blue lines are power laws with critical exponent 3, that would be a signature of $d$-wave superconductivity. In both the samples we argue that the faster relaxation rates below 5.5\,K with respect to the power-law support the hypothesis of $s$-wave pairing.}
\end{figure}
\subsection{SQUID Magnetometry\label{ssec:SQUID}}
\begin{figure}[tbh]
\centering
\includegraphics[width=0.45\textwidth]{fig07}
\caption{\label{fig:Hc1_trend} The temperature dependence of $\mu_0H_{c1}$ exhibits a negative curvature. By using
parabolic fits (dashed lines) the approximate $H_{c1}(0)$ values are obtained: 2.5 mT for Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and 3.6 mT for Hf$_{0.97}$\-V$_{0.03}$\-B$_2$. Inset: the $M$ vs.\ $H$ plots (zero field cooling) in Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ at 3, 5, 9, and 12\,K}
\end{figure}
SQUID magnetometry measurements were made on all the samples (ZrB$_2$, HfB$_2$, Zr$_{0.96}$\-V$_{0.04}$\-B$_2$, and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$).
The high sensitivity (10$^{-10}$\,Am$^2$) of the Reciprocating Sample Option (RSO) of the MPMS magnetometer allowed us to detect small impurities. In particular, in the ZrB$_2$\ and Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ samples a small mass fraction of about 0.5\,\% exhibits a superconducting transition at 5.5 K, the typical $T_c$ of ZrB$_{12}$ impurities. Smaller impurity contributions are also visible from the $\chi(T)$ plots at fields between 0.1 and 7\,T, exhibiting a steady increase below 20\,K.
Therefore, an accurate measurements of $T_c(H = 0)$ was achieved by applying
a small magnetic field of 1\,mT. The obtained values of $T_c$ are 8.33\,K for Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and 7.31\,K for Hf$_{0.97}$\-V$_{0.03}$\-B$_2$.
As shown in the inset of Fig.~\ref{fig:Hc1_trend}, the typical type-II SC cycles can be observed below $T_c$ ;
above $T_c$, a clear diamagnetic trend is confirmed with a typical
$\chi_m = -7.5\times 10^{-10}$\,m$^3$mol$^{-1}$,
a value approximately 5 times smaller than bismuth.
The low-field region (from 0.2 to 6\,mT), exhibits an initial linear trend in $M(H)$. It is possible to extract an approximate value of $\mu_0H_{c1}$, defining it as the field at which the deviation from the linear trend (called \textit{Meissner line}) exceeds the sensitivity of the instrument. Performing this analysis, for each of the $M$ vs.\ $H$ curves, we obtain $\mu_0H_{c1}(T)$, as shown in Fig.~\ref{fig:Hc1_trend} for both Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$. According to the two-band Ginzburg-Landau theory applied to MgB$_2$, the lower critical field exhibits a change in concavity (from negative to positive, upon cooling) at $T/T_c \sim$ 0.5.\cite{Askerzade2002} Since this change is scarcely distinct and the model depends upon the interband mixing of the two order parameters and of their gradients, it is difficult to interpret the $\mu_0H_{c1}(T)$ trend which, as reported in the literature for MgB$_2$,\cite{Sharoni2001, Li2001} can be fitted even with a line. On the other hand, the previously reported Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ data,\cite{Renosto2013}
show a pronounced upturn, which is not present in our case. The reason for this discrepancy is unclear at the moment.
\subsection{Resistivity\label{ssec:res}}
Systematic resistivity measurements were performed in zero field on all the samples
(ZrB$_2$, HfB$_2$, Zr$_{0.96}$\-V$_{0.04}$\-B$_2$, and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$), as shown in Fig.~\ref{fig:Diborides_R_vs_T_all},
and on the superconducting samples Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ (see the inset of Fig.~\ref{fig:Diborides_R_vs_T_all})
and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\ in magnetic fields up to 7\,T.
\begin{figure}[tbh]
\centering
\includegraphics[width=0.45\textwidth]{fig08}
\caption{\label{fig:Diborides_R_vs_T_all}$R$ vs. $T$ data for all the samples (ZrB$_2$, HfB$_2$, Zr$_{0.96}$\-V$_{0.04}$\-B$_2$, and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$) in zero field. The superconducting dropdown in Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\ are evident and confirm the SQUID magnetometry measurement. On the other hand, undoped samples exhibit a constant plateau, as
reflected also by the constance of Knight shift (see Sec.~\ref{ssec:NMR}). We attribute the small dropdown of ZrB$_2$\ to ZrB$_{12}$ impurities. The inset shows the resistivity data of Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ between 2 and 35\,K has been measured at different fields (from 0 to 7 T), confirming a reduced steepness towards higher fields ($\Delta T$ ranging from 2.5 to 3.5\,K towards higher fields) and the expected negative shift.}
\end{figure}
Each resistance measurements in zero field (ZF) was performed from 2 to 310\,K, while in field we focused our
attention on the superconducting transition region (range from 2 to 10\,K). Due to the small resistivity values, i.e., $\rho(T_c) \sim$ 0.7 $\mu \Omega$cm for Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and 0.8 $\mu \Omega$cm for Hf$_{0.97}$\-V$_{0.03}$\-B$_2$, we argue that we can analyze our data within the clean limit approximation.
Based on the data at different fields, the $\mu_0H_{c2}(T)$ values were evaluated. According to the theory of Werthamer, Helfand, and Hohenberg (WHH),\cite{Helfand1966} in the clean limit\cite{Khim2011} and for small spin-orbit couplings,\cite{ Werthamer1966} we have
\begin{equation} \label{eq:WHH}
\mu_0H_{c2}(0) = -0.73\,T_c\,\left.\frac{dH_{c2}}{dT}\right|_{T=T_c}.
\end{equation}
Within this approximation, $\mu_0H_{c2}(T_c)$ can be fitted with a parabolic curve:
\begin{equation} \label{eq:WHHfit}
\mu_0H_{c2}(T) = \mu_0H_{c2}(0)\,[1-(T/T_c)^2]
\end{equation}
\begin{figure}[tbh]
\centering
\includegraphics[width=0.45\textwidth]{fig09}
\caption{\label{fig:Diborides_Hc2_WHH_analyses} $\mu_0H_{c2}(T_c)$ plots for Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ (filled orange circles) and
Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\ (filled red triangles) are derived from resistivity measurements at different field from 0 to 7\,T. Via the WHH model, that implies
a parabolic fit (green and blue continuous lines), we can evaluate $\mu_0H_{c2}(0)$, i.e. 16.4\,T for Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and 10.8\,T for Hf$_{0.97}$\-V$_{0.03}$\-B$_2$. Since Hf is heavier than Zr, we argue that the lower value in the upper critical field is related to the lower average phonon frequency. The old data (Ref.\ \onlinecite{Renosto2013}, empty squares) and the new ones (circles) are in good agreement.}
\end{figure}
From the Ginzburg-Landau formula
$\xi(0) = [\phi_0/(2 \pi \mu_0 H_{c2}(0))]^{1/2}$,
our estimate of the coherence lengths $\xi(0)$ in both Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$\, are 4.5(1) and 5.5(1)\,nm, respectively.
By numerically solving
\begin{equation} \label{eq:Bc1Bc2}
\frac{\mu_0H_{c2}(0)}{\mu_0H_{c1}(0)} = \frac{2 \kappa^2}{\ln \kappa },
\end{equation}
we finally obtain the Ginzburg-Landau $\kappa$ parameter, with a value of 125 for Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and
80 for Hf$_{0.97}$\-V$_{0.03}$\-B$_2$, respectively indicating the strong type II nature of these superconductors. From
$\kappa = \lambda/\xi$, the London penetration depth $\lambda(0)$ in the two materials, is 570 and 445\,nm, respectively.
\subsection{Muon-spin rotation results in the SC phase\label{ssec:musr}}
Values of similar magnitude for the magnetic field penetration depth were obtained from preliminary muon-spin
rotation ($\mu$SR) experiments on Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ (see Fig.~\ref{fig:ZrVB2_TF_MuSR}).
Once implanted in matter, spin-polarized muons act as microscopic probes of magnetism,
which upon decay emit positrons preferentially along the muon-spin direction.
From the spatial anisotropy of the emitted positrons (i.e., the asymmetry signal)
one can reveal the distribution of the local magnetic fields. \cite{Blundell1999,Yaouanc2011}
In our case, by applying of 0.6\,T, a regular flux-line lattice (FLL)
develops in the superconducting phase below $T_c$. By uniformly sampling the FLL muons experience
an additional relaxation $\sigma_{\mathrm{sc}}$, which is related to the absolute magnetic
penetration depth $\lambda$ via:\cite{Brandt1988,Brandt2003}
\begin{equation}
\label{eq:lambda}
\frac{\sigma_{sc}^2}{\gamma^2_{\mu}}=0.00371 \cdot \frac{\varPhi_0^2}{\lambda^4}.
\end{equation}
Here $\varPhi_0 =2.068\times10^{-3}$\,T$\mu$m$^2$ is the magnetic flux quantum and
$\gamma_{\mu} = 2\pi \times 135.53$\,MHz/T, the muon gyromagnetic ratio.
\begin{figure}[bth]
\centering
\includegraphics[width=0.45\textwidth]{fig10}
\caption{\label{fig:ZrVB2_TF_MuSR} Transverse-field $\mu_SR$ data in the superconducting phase of Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ in a magnetic
field $\mu_0 H = 0.6$\,T.}
\end{figure}
Figure~\ref{fig:ZrVB2_TF_MuSR} shows the temperature dependence of $\sigma_{sc}$, proportional to the superfluid density ($\sigma_{sc} \propto n_s \propto \lambda^{-2}(T)$), together with a numerical fit with an average-field model $1/\lambda^2(T)=(1/\lambda^2(0))[1-(T/T_c)^n]$, which gives $1/\lambda^2(0)=1.43\pm0.2$\,$\mu$m$^{-2}$ and $n=2.0\pm0.1$. Subsequently, by using the relation $\lambda_{ab}(0)= \lambda_{\mathrm{eff}}(0)/1.31$, we estimate the in-plane magnetic penetration depth $\lambda_{ab}(0) = 638\pm11$\,nm.
This value is close to the one determined via macroscopic methods (see above), but it is very different from
$\lambda_{ab}(0) = 100$\,nm, also measured via $\mu$SR in MgB$_2$.\cite{Niedermayer2002}
This difference can be accounted for by considering the rather small electronic density of states in Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ compared to that
in MgB$_2$, compatible with the very different NMR relaxation-rate values reported in
Table~\ref{tab:summary}.
As a final note, we recall that the choice of the applied transverse field is crucial for the correct
determination of the field penetration depth. Indeed, detailed studies of the magnetic field dependence of
the muon-spin depolarization rate in MgB$_2$ (see, e.g., Figs.~1 and 2 in Ref.\onlinecite{Niedermayer2002})
have shown strong pinning effects in low applied fields (below 0.3\,T). These imply a considerably distorted
vortex lattice leading to a strong decay of the muon asymmetry and hence underestimated $\lambda_{ab}(0)$
values. However, since in fields exceeding 0.3\,T, only weak or no pinning effects were observed, we are confident
that by applying a transverse field of 0.6\,T, our results reflect the penetration depth.
\vspace{-5pt}
\section{Summary and conclusions\label{sec:Conclusions}}
From SQUID magnetometry, NMR, resistivity, and preliminary $\mu$SR experiment data on ZrB$_2$, HfB$_2$, Zr$_{0.96}$\-V$_{0.04}$\-B$_2$,
and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$, we argue that the latter two samples are $s$-wave superconductors, resulting from electron doping via
$d$ orbitals of vanadium. On the other hand, in MgB$_2$ the peculiar self-doping
exists thanks to the boron-like electrons at the Fermi level; at the same time also the $s$ states
of Mg donate their electrons to the boron-derived conduction bands (\textit{metallic} B).
It is also worth mentioning that B vacancies (a common defect occurring in diborides) enhance
the DOS even further (more than twice for doping of about 0.5\,\%).\cite{Dahlqvist2015} Besides the many
similarities with MgB$_2$, the lower $T_c$ values (7.31 and 8.33\,K instead of 39\,K) are accounted for
by the differences in the electronic structure and Fermi surface. The DOS per unit cell of MgB$_2$ at the Fermi level
(0.719\,states/eV) is dominated (about 60\%) by the $p$ orbitals of B atoms, while in the case of
the undoped transition-metal diborides (ZrB$_2$, for instance) 80\% of the DOS (0.130\,states/eV)\cite{Shein2008}
derives from the $d$ orbitals of the cation. Furthermore, our data support the hypothesis that we are dealing
with new two-band two-gap $s$-wave superconductors, since the coexistence of a normal and a superconducting
phase below $T_c$ suggests an anisotropy of the upper critical field, as well documented in the paradigmatic case of
MgB$_2$. More direct evidence for this assumption could be obtained by various experiments probing single
crystals of Zr$_{0.96}$\-V$_{0.04}$\-B$_2$\ and Hf$_{0.97}$\-V$_{0.03}$\-B$_2$.
\begin{acknowledgments}
This work was financially supported in part by the Schweizerische Nationalfonds zur F\"{o}rderung der Wissenschaftlichen
Forschung (SNF).
\end{acknowledgments}
|
\section{Introduction}\label{Introduction}
The late-time cosmic accelerated expansion
\cite{Perlmutter:1998np,Riess:1998cb} has posed important and
challenging problems to theoretical cosmology. Although the standard
model of cosmology has favored dark energy models
\cite{Copeland:2006wr} as fundamental candidates responsible for the
accelerated cosmic expansion, it is also viable that this expansion is
due to modifications of general relativity
\cite{Clifton:2011jh,Nojiri:2010wj}, which introduce new degrees of
freedom to the gravitational sector itself. Indeed, the phenomenology
of $f(R)$ gravity, where $R$ is the metric Ricci curvature scale and $f$ a
general function,
has been scrutinized motivated by the possibility to
account for the self-accelerated cosmic expansion without invoking
dark energy sources \cite{Sotiriou:2008rp,DeFelice:2010aj}. Besides,
this kind of modified gravity is capable of addressing the dynamics of
several self-gravitating systems alternatively to the presence of dark
matter \cite{Boehmer:2007kx,Bohmer:2007fh}.
In this approach of $f(R)$ gravity, and using its equivalent scalar-tensor
representation,
one can show that in order to satisfy local, i.e., solar system, observational constraints
a large mass of the scalar field is required,
which scales with the curvature through the chameleon mechanism \cite{Khoury:2003aq,Khoury:2003rn}.
In turn, this has undesirable effects at cosmological scales.
There are other drawbacks in these models,
see \cite{Sotiriou:2008rp,DeFelice:2010aj,Boehmer:2007kx,Bohmer:2007fh}.
A Palatini version of the $f(R)$ gravity
theory, where the connection rather than the metric represents the fundamental
gravitational field, has been proposed \cite{Olmo:2011uz}, and it has been
established that it has
interesting features but the manifest deficiencies and
downsides of the metric $f(R)$ gravity also appear
\cite{Olmo:2011uz}.
A hybrid combination of the two versions of the $f(R)$ gravity theories,
containing elements from both of
the two formalisms, i.e., a hybrid metric-Palatini theory,
consists of adding an $f({\cal R})$ term
constructed \`{a} la Palatini to the
Einstein-Hilbert Lagrangian \cite{Harko:2011nh}.
It turns out to be very successful in accounting for the observed
phenomenology and is able to avoid some of the shortcommings of the
original approaches \cite{Capozziello:2013uya}.
In the scalar-tensor representation of the hybrid metric-Palatini
theory there is a long-range light mass scalar field, which is able to
modify, in a way consistent with the observations, the cosmological
and galactic dynamics, but leaves the solar system unaffected
\cite{Capozziello:2012ny,Capozziello:2012qt}. This light scalar field
thus allows to evade the screening chameleon mechanism.
In addition, absence of instabilities in perturbations were
also verified in this model \cite{Koivisto:2013kwa}.
The theory was further developed into a generalized
hybrid metric-Palatini theory, in which a general function, $f\left(R,\mathcal R\right)$,
was postulated depending
on
both the metric and Palatini curvature scalars, $R$ and $\mathcal R$,
respectively
\cite{Tamanini:2013ltp}.
Of course, the positive
aspects of the initial hybrid metric-Palatini theory
are preserved, namely, the
consistency with
cosmological and galactic dynamics, and correct solar system
tests
\cite{Carloni:2015bua,Capozziello:2013yha}, see \cite{Capozziello:2015lza} for a review.
There are other developments.
Considering linear homogeneous perturbations, the stability regions of the Einstein static universe were analyzed, and it was shown that a large class of stable solutions exists \cite{Boehmer:2013oxa}.
In addition,
the full set of linearized evolution equations for the perturbed potentials, in the Newtonian and synchronous gauges, were derived. It was concluded that the main deviations from general relativity arise in the distant past, with an oscillatory signature in the ratio between the Newtonian potentials \cite{Lima:2014aza}. Furthermore, using specific models and a combination of cosmic microwave background, supernovae and baryonic acoustic oscillations background data, it was shown that the model's free parameters are in agreement with the observational constraints \cite{Lima:2015nma,Lima:2015nma,Lima:2016npg}.
It was also shown in this theory that the
initial value problem can always be well-formulated and be well-posed
depending on the adopted matter sources \cite{Capozziello:2013gza}.
The understanding and the building of solutions in complex theories like
the generalized hybrid
metric-Palatini gravity, and its scalar-tensor representation, is a difficult task.
Among other methods to find solutions,
the reconstruction technique method has been often valuable
in this search.
This method was first employed in
\cite{LM} in order to select the form for the inflation potential able
to resolve the open problems of the inflationary paradigm, e.g., the
graceful exit. Also other elegant attempts were made to
generalize this approach to nonminimally coupled scalar-tensor
theories with a single scalar field \cite{Ellis:1990wsa}.
In this work, we aim to find cosmological solutions in the generalized
hybrid metric-Palatini gravity proposed in \cite{Tamanini:2013ltp}
through the use of its scalar-tensor representation. We will then
devise a reconstruction technique algorithm for scalar-tensor theories
in a Friedmann-Lema\^itre-Robsertson-Walker (FLRW) universe and we
will show that the generalized theory through its scalar-tensor
representation provides a very rich structure in astrophysical and
cosmological applications. In particular, we will find that the
cosmology of the generalized hybrid metric-Palatini gravity can differ
in subtle ways from both general relativity and $f(R)$ gravity. In
principle these differences can be used in combination with
observational data to verify the viability of this class of theories.
This paper is organized in the following manner: In Sec.~\ref{sec:action}, we present the formalism of the generalized hybrid metric-Palatini gravity, and express the action in the scalar-tensor representation. In Sec.~\ref{secIII}, we consider the cosmological dynamics of FLRW spacetimes, using the scalar-tensor representation of the hybrid theory. In Sec.~\ref{secIV}, we obtain, from the
solutions found in the scalar-tensor representation, the specific forms for $f\left(R,\cal{R}\right)$. In Sec.~\ref{conclusion}, we set out our conclusions.
\section{Generalized hybrid gravity: Formalism}\label{sec:action}
\subsection{Action and field equations for the generalized hybrid gravity}
Consider the action of the generalized hybrid metric-Palatini modified theory of gravity, given by (the velocity of light is set to one)
\begin{equation}\label{genac}
S=\frac{1}{2\kappa^2}\int_\Omega\sqrt{-g}f\left(R,\cal{R}\right)d^4x+S_m(g_{ab}, \chi),
\end{equation}
where $\kappa\equiv 8\pi G$,
$G$ is the gravitational constant, $g$ is the determinant of the metric
$g_{ab}$, $f$ is a function of $R$ and $\mathcal{R}$, and $S_m$ is the matter action, in which matter is minimally coupled to the metric $g_{ab}$, and $\chi$ collectively denotes the matter fields. $R$ is the metric Ricci scalar and
$\mathcal{R}\equiv g_{ab}\mathcal{R}^{ab}$ is the Palatini scalar
curvature, with $\mathcal{R}^{ab}$
being
defined in terms of an independent connection $\hat\Gamma^c_{ab}$ as,
\begin{equation}
\mathcal{R}_{ab}=\partial_c\hat\Gamma^c_{ab}-\partial_b
\hat\Gamma^c_{ac}+\hat\Gamma^c_{cd}\hat\Gamma^d_{ab}-\hat\Gamma^c_{ad}
\hat\Gamma^d_{cb}\,.
\label{RicciPalatini}
\end{equation}
One can also use the additional independent connection as a building block to construct higher order curvature invariants, in the action
$
S = \frac{1}{2\kappa^2} \int {\rm d}^4x \sqrt{-g}\, f( R,\mathcal{R},{\cal{Q}}_H )
$,
where the term ${\cal{Q}}_H$, can take the following forms
${\cal{R}}^{\mu\nu} {\cal{R}}_{\mu\nu}$, $R^{\mu\nu}{\cal{R}}_{\mu\nu}$, ${\cal{R}}^{\mu\nu\alpha\beta}{\cal{R}}_{\mu\nu\alpha\beta}$, $R^{\mu\nu\alpha\beta}{\cal{R}}_{\mu\nu \alpha\beta}$, ${\cal{R}} R$, etc. However we will not consider these cases here. Relative to the presence of Ostrogradski instabilities, in the hybrid theory, like in $f(R)$-gravity, it is possible to avoid the problem by allowing a separation of the additional degrees of freedom into a harmless scalar degree of freedom \cite{Sotiriou:2008rp,DeFelice:2010aj}. In fact, it turns out that this feature is a similar exception in the larger space of metric-affine theories, where a generic theory is plagued by ghosts, superluminalities and other unphysical degrees of freedom \cite{Koivisto:2013kwa}.
Varying the action (\ref{genac}) with respect to the metric $g_{ab}$ and the independent connection, yields the following field equations
\begin{eqnarray}
\frac{\partial f}{\partial R}R_{ab}+\frac{\partial f}{\partial
\mathcal{R}}\mathcal{R}_{ab}-\frac{1}{2}g_{ab}f\left(R,\cal{R}\right)
\nonumber \\
-\left(\nabla_a\nabla_b-g_{ab}\Box\right)\frac{\partial f}{\partial R}=\kappa^2 T_{ab},
\end{eqnarray}
and
\begin{equation}
\hat\nabla_c\left(\sqrt{-g}\frac{\partial f}{\partial \cal{R}}g^{ab}\right)=0,
\label{indconnnection}
\end{equation}
respectively. The equation of motion (\ref{indconnnection}) implies that the independent connection is the Levi-Civita connection of a
new metric tensor $h_{ab}$ which is conformally related to $g_{ab}$ by the relation
\begin{equation}\label{hgf}
h_{ab}=g_{ab} \frac{\partial f}{\partial \cal{R}}\,.
\end{equation}
The independent connection can then be written in terms of the new metric as
\begin{equation}
\hat\Gamma^a_{bc}=\frac{1}{2}h^{ad}\left(\partial_b h_{dc}+
\partial_c h_{bd}-\partial_d h_{bc}\right).
\end{equation}
Thus, the new metric $h_{ab}$ is an auxiliary metric related to the independent connection, that was used to define the Palatini tensor, given by Eq. (\ref{RicciPalatini}). We emphasize that matter is coupled to the physical metric $g_{ab}$, so that only the Levi-Civita connection $\Gamma^a_{bc}(g)$ should be used in the geodesic equation applied to the metric-Palatini theory.
Note that since the matter action $S_m$ does not depend on the independent connection $\hat\Gamma$, the equation of motion for this connection is independent of the stress-energy tensor, whereas the same does not happen to the equation of motion of the metric $g_{ab}$.
\subsection{Scalar-tensor representation}
It is useful to express the action~\eqref{genac} in a scalar-tensor representation. This can be achieved by considering an action with two auxiliary fields, $\alpha$ and $\beta$, respectively, in the following form
\begin{eqnarray}
S=\frac{1}{2\kappa^2}\int_\Omega \sqrt{-g}\;\Big[f\left(\alpha,\beta\right)+\frac{\partial f}{\partial \alpha}\left(R-\alpha\right)
\nonumber \\
+\frac{\partial f}{\partial\beta}\left(\cal{R}-\beta\right)\Big]d^4x+S_m.
\label{gensca}
\end{eqnarray}
Putting $\alpha=R$ and $\beta=\mathcal{R}$ we recover action~\eqref{genac}.
To proceed, we define two scalar fields as
\begin{equation}\label{phidef}
\varphi=\frac{\partial f}{\partial\alpha}\,,
\end{equation}
\begin{equation}\label{psidef}
\psi=-\frac{\partial f}{\partial\beta}\,,
\end{equation}
where the negative sign in Eq.~\eqref{psidef}
is imposed to guarantee a positive kinetic energy for the scalar field.
Defining a potential $V$ as
\begin{equation}\label{potdef}
V\left(\varphi,\psi\right)=-f\left(\alpha,\beta\right)+\varphi\alpha-\psi\beta\,,
\end{equation}
the action equivalent to (\ref{gensca}) is of the form
\begin{equation}
S=\frac{1}{2\kappa^2}\int_\Omega \sqrt{-g}\left[\varphi R-\psi\mathcal{R}-V\left(\varphi,\psi\right)\right]d^4x+S_m\,,
\label{action3}
\end{equation}
where the variation of the matter action $S_m$ with respect to the metric $g_{ab}$
yields the stress-energy tensor $T_{ab}$.
It is important to stress that in this representation the scalar fields are not actually matter fields but only alternative representations of the curvatures $R$ and $\mathcal R$, and consequently they are allowed to assume values, such as negative ones, which would not be physically valid.Also, since we imposed the signs on the definitions of $\psi$ and $\varphi$ to guarantee that the kinetic energies are positive, then we should not worry about ghost instabilities. This range of values corresponds in the generalized hybrid gravity picture to some specific behavior of the derivatives of the function $f$ in the action. That said, it should also be clear that zeros for the scalar fields are problematic points in terms of the relation between the representations. In the following, the solutions we will obtain are only valid in the interval between the zeros of the scalar fields. In addition, it should be remarked that the scalar field representation will only be meaningful provided that $R$ and $\mathcal{R}$ can be expressed in terms of the scalar fields. This is not the case for all forms of the function $f$. In the next sections we will show some examples in which the consequences of these limitations become important.
Taking into account that $h_{ab}=-\psi\, g_{ab}$, see Eqs. (\ref{hgf}) and (\ref{psidef}), we have
\begin{equation}\label{confrt}
\mathcal{R}=R+\frac{3}{\psi^2}\partial^a \psi\partial_a \psi-\frac{3}{\psi}\Box\psi \,.
\end{equation}
Thus, we can replace $\cal{R}$ in the action (\ref{action3}),
to obtain
\begin{eqnarray}\label{genacts2}
S=\frac{1}{2\kappa^2}\int_\Omega \sqrt{-g}\Big[\left(\varphi-\psi\right) R-\frac{3}{2\psi}\partial^a\psi\partial_a\psi
\nonumber \\
-V\left(\varphi,\psi\right)\big]d^4x+S_m.
\end{eqnarray}
Varying the previous action, Eq.~(\ref{genacts2}), with respect to the metric $g_{ab}$ and the scalar fields $\varphi$ and $\psi$, and rearranging terms, we obtain the following equations of motion
\begin{eqnarray}
&&\left(\varphi-\psi\right) G_{ab}=\kappa^2T_{ab}+\left(\nabla_a\nabla_b-g_{ab}\Box\right)\left(\varphi-\psi\right)
\nonumber \\
&&\ +\frac{3}{2\psi}\partial_a\psi\partial_b\psi+\left(\frac{1}{2}V+\frac{3}{4\psi}\partial^c\psi\partial_c\psi\right)g_{ab}\,,
\label{genein2}
\end{eqnarray}
\begin{equation}\label{genphi}
\Box\varphi+\frac{1}{3}\left(2V-\psi V_\psi-\varphi V_\varphi\right)=\frac{\kappa^2T}{3} \,,
\end{equation}
\begin{equation}
\Box\psi-\frac{1}{2\psi}\partial^a\psi\partial_a\psi-\frac{\psi}{3}\left(V_\varphi+V_\psi\right)=0\,,
\label{genpsi}
\end{equation}
respectively, where $V_\varphi$ and $V_\psi$ are defined
by
\begin{equation}
V_\varphi\equiv \frac{\partial V}{\partial\varphi}\,, \qquad
V_\psi\equiv \frac{\partial V}{\partial\psi}\,,
\end{equation}
respectively.
The dynamical equation for $\varphi$, Eq.~(\ref{genphi}), can be obtained in the following manner. The variation of the action with respect to
$\varphi$ gives
\begin{equation}\label{rvphi}
R=V_{\varphi}.
\end{equation}
Then, taking the trace of Eq.~(\ref{genein2})
and inserting $R=V_{\varphi}$ gives Eq.~(\ref{genphi}).
The latter equation
shows an important difference between the two scalar fields, namely, $\varphi$ is coupled to matter whereas $\psi$, given through Eq.~(\ref{genpsi}), is not. This fact will have important consequences when we will look for cosmological solutions in which matter is present.
\section{Cosmological equations and solutions for the generalized hybrid gravity in the scalar-tensor representation}\label{secIII}
\subsection{Cosmological equations in the scalar-tensor representation}\label{seceqsstr}
In this section, we consider the FLRW spacetime with the spatial curvature parameter $k$, where $k$ can assume three values, $k=-1,0,1$. In spherical coordinates
$(t,r,\theta,\phi)$
the line element can be written as
\begin{equation}\label{FLRW}
ds^2=-dt^2+a^2\left(t\right)\left[\frac{dr^2}{1-kr^2}+r^2d\theta^2+r^2\sin^2\theta d\phi^2\right],
\end{equation}
where $a(t)$ is the scale factor.
We also assume that the cosmological scalar fields are time-dependent, $\varphi=\varphi\left(t\right)$, $\psi=\psi\left(t\right)$, and that the matter is described by a perfect fluid,
\begin{equation}\label{perffluid}
T^a_b={\rm diag}\left(-\rho,p,p,p\right)\,,
\end{equation}
so that the trace is given by
\begin{equation}\label{tabtrace}
T=T^a_a=-\rho+3p\,,
\end{equation}
with $\rho$ and $p$ being the energy density and the pressure of the fluid, respectively, and both are assumed to depend solely on $t$.
Then the modified cosmological equations for this case can be obtained
by computing directly the two independent components of
the gravitational field equation, Eq.~\eqref{genein2}, i.e., the Friedmann equation
\begin{eqnarray}\label{genH}
3\left(\frac{\dot a}{a}\right)^2+\frac{3k}{a^2}=\frac{1}{\varphi-\psi}\Bigg[\kappa^2\rho+\frac{V}{2}
\nonumber \\
+3\left(\frac{\dot\psi^2}{4\psi}-\left(\frac{\dot a}{a}\right)\left(\dot\varphi-\dot\psi\right)\right)\Bigg],
\end{eqnarray}
and the Raychaudhuri equation
\begin{eqnarray}
\label{gendH}
2\frac{d}{dt} {\left(\frac{\dot a}{a}\right)}-\frac{2k}{a^2}=\frac{1}{\varphi-\psi}\Bigg[-\kappa^2\left(\rho+p\right)-\frac{3\dot\psi^2}{2\psi}
\nonumber \\
+\left(\frac{\dot a}{a}\right)\left(\dot\varphi-\dot\psi\right)-\left(\ddot\varphi-\ddot\psi\right)\Bigg]\,,
\end{eqnarray}
respectively, where a dot means differentiation with respect
to $t$, $\;\dot{}\equiv d/dt$.
The evolution equations
for the scalar fields, Eqs.~\eqref{genphi} and
\eqref{genpsi}, take the following forms
\begin{equation}
\label{kgphi0}
\ddot\varphi+3\left(\frac{\dot a}{a}\right)\dot\varphi-\frac{1}{3}\left[2V-\psi V_\psi-\varphi V_\varphi\right]=-\frac{\kappa^2T}{3}\,,
\end{equation}
\begin{equation}\label{kgpsi0}
\ddot\psi+3\left(\frac{\dot a}{a}\right)\dot\psi-\frac{\dot\psi^2}{2\psi}+\frac{\psi}{3}\left(V_\varphi+V_\psi\right)=0\,,
\end{equation}
respectively.
Now, an equation for the potential can be obtained by multiplying Eq.~\eqref{gendH} by $3/2$, summing it to Eq.~\eqref{genH}, and using Eqs.~\eqref{kgphi0} and~\eqref{kgpsi0} to cancel the terms $\ddot\varphi$ and $\ddot\psi$, similarly to the method performed in \cite{Ellis:1990wsa}. The result is as follows,
\begin{equation}\label{modF1}
V_\varphi=6\left[
{\frac{d}{dt} \left(\frac{\dot a}{a}\right)
} +2
\left(\frac{\dot a}{a}\right)^2+\frac{k}{a^2}
\right]\,.
\end{equation}
Finally, the equations to describe the evolution of matter are an equation of state of the form
\begin{equation}\label{eos}
p=w\rho,
\end{equation}
where $w$ is a parameter without units,
and the conservation law for the stress-energy tensor $\nabla_a T^{ab}=0$, which becomes, using the equation of state~\eqref{eos},
\begin{equation}\label{econs}
\dot\rho=-3\,\frac{\dot a}{a}\,\rho\left(1+w\right).
\end{equation}
We have now a system of five independent equations, namely~\eqref{kgphi0}--\eqref{econs}, that
we have to solve for the seven independent variables $k,a, \rho, p, V, \varphi$, and $\psi$.
We will choose some specific geometry, i.e., we choose $k$.
We can then impose one extra constraint to the system in order for it to have a unique solution. In the following, we shall impose different constraints and obtain their respective solutions.
It is also useful to define
the Hubble function or parameter $H$ as usual
\begin{equation}\label{hubblep}
H=\frac{\dot a}{a}\,,
\end{equation}
which is a function of $t$. With Eq.~\eqref{hubblep}
we can trade $a(t)$ for $H(t)$ and vice versa.
\subsection{Cosmological solutions in the scalar-tensor representation}\label{secIVb}
\subsubsection{The de Sitter solution}\label{method1}
In this section, we try to find a solution of the form of a de Sitter expansion.
To do so
we
set the matter component to be vacuum, that is, $\rho=p=0$.
This choice means that we actually do not need Eqs.~\eqref{eos} and~\eqref{econs} to solve the system.
We also choose a flat universe, i.e., $k=0$
and a scale factor described by
\begin{equation}\label{scalefactordes1}
a\left(t\right)=Ae^{\sqrt\Lambda\, \left(t-t_0\right)}\,,
\end{equation}
where $A$ and $\Lambda$ are free constants, and
$\Lambda$ can be seen as some cosmological constant. This scale factor is plotted in Fig. \ref{fig:a1}.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{a1.pdf}
\caption{Plot of the scale factor $a(t)$ for the de Sitter case, given in Eq.~\eqref{scalefactordes1},
for the values $a_0=1$, $\Lambda=1$.}
\label{fig:a1}
\end{figure}
With these assumptions, Eq.~\eqref{modF1} becomes
\begin{equation}\label{solV1}
V_\varphi=12\Lambda.
\end{equation}
This equation can be directly integrated with respect to $\varphi$ to obtain a potential of the form
\begin{equation}
V\left(\varphi,\psi\right)=12\Lambda\varphi+b\left(\psi\right).
\end{equation}
where $b\left(\psi\right)$ is an arbitrary function on $\psi$ which arises from the fact that the potential is a function of both fields.
These results leave Eqs.~\eqref{kgphi0} and~\eqref{kgpsi0} as
\begin{equation}
\ddot\varphi+3\sqrt\Lambda\,\dot\varphi-\frac{1}{3}\left[12\Lambda\,\varphi+2b\left(\psi\right)-\psi b'\left(\psi\right)\right]=0\,,
\label{dotphi}
\end{equation}
\begin{equation}
\ddot\psi+3\sqrt\Lambda\,\dot\psi-\frac12\frac{\dot\psi^2}{\psi}+\frac13\,\psi\left[12\Lambda+b'\left(\psi\right)\right]=0,
\label{dotpsi}
\end{equation}
respectively.
In the particular case where
\begin{equation}
b\left(\psi\right)=-12\Lambda\,\psi\,,
\label{bchoice}
\end{equation}
the potential becomes
\begin{equation}
V\left(\varphi,\psi\right)=12\Lambda\left(\varphi-\psi\right)\,,
\label{potentialbchoice}
\end{equation}
and Eq.~\eqref{dotpsi} can be divided by $\dot\psi$ and directly integrated over $t$ to obtain the solution
\begin{equation}\label{solpsi}
\psi\left(t\right)=\psi_0 e^{-6 \sqrt\Lambda \, t} \left[e^{3\sqrt \Lambda \, \left(t-t_0\right)}-1\right]^2,
\end{equation}
where $t_0$ and $\psi_0$ are constants of integration.
In the same case of Eq.~\eqref{bchoice}, where $b\left(\psi\right)=-12\Lambda\,\psi$, Eq.~\eqref{dotphi}
becomes
$\ddot\varphi+3\sqrt\Lambda\,\dot\varphi-4\Lambda\,\left(\varphi-\psi\right)=0$, and
upon using Eq.~\eqref{solpsi} it
can be integrated to obtain the solution
\begin{eqnarray}\label{solvarphiades1}
\varphi\left(t\right)=&&\psi_0 e^{-6 \sqrt\Lambda \,t_0}-\frac{2}{7} \psi_0 e^{-6 \sqrt\Lambda t}\nonumber \\
&&-2 \psi_0 e^{-3 \sqrt\Lambda \left(t+t_0\right)}+\varphi_0 e^{-4 \sqrt\Lambda t}+\varphi_1 e^{\sqrt\Lambda t},
\end{eqnarray}
where $\varphi_0$ and $\varphi_1$ are constants of integration. The solutions for the scalar fields $\psi\left(t\right)$ and $\varphi\left(t\right)$ are plotted in Figs. \ref{fig:psi1} and \ref{fig:phi1}, respectively.
The solution is complete since $k$,
$a, \rho, p, V, \varphi$, and $\psi$ are known.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{psi1.pdf}
\caption{Plot of the scalar field $\psi(t)$ for the de Sitter case, given in Eq.~\eqref{solpsi},
for the values $\psi_0=\Lambda=1$, and $t_0=0$.}
\label{fig:psi1}
\end{figure}
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{phi1.pdf}
\caption{Plot of the scalar field $\phi(t)$ for the de Sitter case, given in Eq.~\eqref{solvarphiades1},
for the values $\phi_0=\phi_1=\Lambda=1$, and $t_0=0$.}
\label{fig:phi1}
\end{figure}
\subsubsection{Solution with a simplified potential equation I}
\label{method2}
In this section, we set again the matter component to be vacuum, i.e., $\rho=p=0$, and we choose again a flat universe, $k=0$. Inspecting Eq.~\eqref{modF1}, we can see that it simplifies if the first two terms within the brackets cancel each other, i.e.,
\begin{equation}
\dot H+2H^2=0,
\label{eqh1}
\end{equation}
where we have used the definition for the Hubble function $H$, Eq.~\eqref{hubblep}.
Equation~\eqref{eqh1}
has the following solution for $H$,
\begin{equation}
H=\frac{1}{2\left(t-t_0\right)},
\label{hs2}
\end{equation}
where $t_0$ is an integration constant. Now, using again Eq.~\eqref{hubblep},
then Eq.~\eqref{hs2} can be integrated to obtain
\begin{equation}
a\left(t\right)=a_0\sqrt{2\left(t-t_0\right)},
\label{a1pot}
\end{equation}
where $a_0$ is an integration constant. This solution is plotted in Fig.\ref{fig:a2}.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{a2.pdf}
\caption{Plot of the scale factor $a(t)$ for the case of the simplified potential I, given in Eq.~\eqref{a1pot},
for the values $a_0=1$, and $t_0=0$.}
\label{fig:a2}
\end{figure}
The interest in this solution resides in the fact that, for a universe populated by radiation, we expect that the behavior of the scale factor is proportional to $\sqrt{t}$, but in this case the same behavior can be obtained with $\rho=0$.
With this set of assumptions, the equation for the potential, Eq.~\eqref{modF1}, becomes
\begin{equation}\label{solV2}
V_\varphi=0,
\end{equation}
which can be directly integrated to obtain
\begin{equation}\label{solV13}
V\left(\varphi,\psi\right)=b\left(\psi\right),
\end{equation}
where $b\left(\psi\right)$ is an arbitrary function on $\psi$.
Using Eqs.~\eqref{a1pot} and~\eqref{solV13}, Eq.~\eqref{kgpsi0} for $\psi$ becomes
\begin{equation}\label{ddotpsieq2}
\ddot\psi+\frac{3\dot\psi}{2\left(t-t_0\right)}-\frac{\dot\psi^2}{2\psi}+\frac{\psi}{3}b'\left(\psi\right)=0.
\end{equation}
In the particular case where
\begin{equation}
b\left(\psi\right)=V_0,
\end{equation}
with $V_0$ a constant, the last term on the left-hand side of Eq.~\eqref{ddotpsieq2} vanishes and the equation can be integrated directly after dividing through by $\dot\psi$, yielding the solution
\begin{equation}\label{psisimppot1}
\psi\left(t\right)=\frac{\psi_1 \left(\psi_0 \sqrt{2\left(t-t_0\right)}-1\right)^2}{2\left(t-t_0\right)},
\end{equation}
where $\psi_0$ and $\psi_1$ are constants of integration. On the other hand, the equation for $\varphi$, namely Eq.~\eqref{kgphi0}, takes the form
\begin{equation}\label{varphiother1}
\ddot\varphi+\frac{3\dot\varphi}{2\left(t-t_0\right)}-\frac{1}{3}\left[2b\left(\psi\right)-\psi b'\left(\psi\right)\right]=0.
\end{equation}
Under the choice given in Eq.~\eqref{solV13} for $b\left(\psi\right)$, Eq.~\eqref{varphiother1}
decouples completely from the equation for $\psi$, yielding
\begin{equation}
\ddot\varphi+\frac{3\dot\varphi}{2\left(t-t_0\right)}-\frac{2}{3}V_0=0,
\end{equation}
which can be directly integrated to obtain the following solution,
\begin{equation}\label{phisimppot1}
\varphi\left(t\right)=\frac{2}{15}V_0t\left(t-2t_0\right)-\frac{\varphi_0}{\sqrt{2\left(t-t_0\right)}}+\varphi_1,
\end{equation}
where $\varphi_0$ and $\varphi_1$ are constants of integration. The solutions for the scalar field $\psi\left(t\right)$ and $\varphi\left(t\right)$ are plotted in Figs. \ref{fig:psi2} and \ref{fig:phi2}, respectively.
The solution is complete since
$k,a, \rho, p, V, \varphi$, and $\psi$ are known.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{psi2.pdf}
\caption{Plot of the scalar field $\psi(t)$ for the case of the simplified potential I, given in Eq.~\eqref{psisimppot1},
for the values $\psi_0=\psi_1=1$, and $t_0=0$.}
\label{fig:psi2}
\end{figure}
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{phi2.pdf}
\caption{Plot of the scalar field $\varphi(t)$ for the case of the simplified potential I, given in Eq.~\eqref{phisimppot1},
for the values $\varphi_0=\varphi_1=V_0=1$, and $t_0=0$.}
\label{fig:phi2}
\end{figure}
\subsubsection{Solution with a simplified potential equation II}
\label{method3}
In this section, we set again the matter component to be vacuum,
$\rho=p=0$, and we choose again a flat universe, $k=0$.
Taking into account Eq.~\eqref{modF1}, we look for solutions in which the first two
terms in the brackets are equal to a constant $\Omega^2/2$ say,
where the 2 enters for convenience. Then Eq.~\eqref{modF1}
is now
\begin{equation}
\dot H+2H^2+\frac{\Omega^2}{2}=0\,,
\label{assumption2}
\end{equation}
where we have used the definition for the Hubble function
$H$, Eq.~\eqref{hubblep}.
This condition yields the solution
\begin{equation}
H=-{\frac{\Omega}{2}}\tan\left[\Omega\left(t-t_0\right)\right]\,,
\label{hnew2}
\end{equation}
where $t_0$ is a constant of integration.
Now, using again Eq.~\eqref{hubblep},
then Eq.~\eqref{hnew2} can be integrated to obtain
the scale factor,
\begin{equation}\label{scale3}
a\left(t\right)=a_0\sqrt{\cos\left[\Omega\left(t-t_0\right)\right]},
\end{equation}
where $a_0$ is a positive constant of integration, $a_0>0$.
The behavior of $a(t)$ in Eq.~\eqref{scale3} is given in
Fig.~\ref{fig:am3}.
This solution is valid in between times
$t=-\frac{\pi}{2\Omega}+t_0$ and $t=\frac{\pi}{2\Omega}+t_0$,
or times that are translated from these ones
by $2\pi n$ with integer $n$.
In between times $t=\frac{\pi}{2\Omega}+t_0$ and $t=\frac{3\pi}{2\Omega}+t_0$,
or times that are translated from these ones
by $2\pi n$,
there are no physical solutions, since the scale factor in
Eq.~\eqref{scale3} gets imaginary values.
The solution thus
represents a universe which starts expanding at $t=-\frac{\pi}{2\Omega}+t_0$, attains a maximum value at $t=t_0$, and then collapses again at $t=\frac{\pi}{2\Omega}+t_0$.
The importance of this solution for $a\left(t\right)$ stands on the fact that in the later times of the universe expansion, where we might approximate the distribution of matter to be vacuum and the geometry to be flat, it is possible to obtain solutions for the scale factor with a dependence on time different than the usual $t^{1/2}$ power-law expected in standard general relativity.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{am3.pdf}
\caption{Plot of the scale factor $a(t)$ for the case
with a simplified potential equation II, given in Eq.~\eqref{scale3},
for the values $\Omega=1$, $a_0=1$, and $t_0=0$.}
\label{fig:am3}
\end{figure}
Assumption~\eqref{assumption2} leaves the equation for the potential, Eq.~\eqref{modF1}, in the form
\begin{equation}
V_\varphi=-3\Omega^2,
\end{equation}
which can be integrated to obtain
\begin{equation}
V\left(\varphi,\psi\right)=-3\Omega^2\varphi+b\left(\psi\right),
\end{equation}
where $b\left(\psi\right)$ is an arbitrary function of $\psi$. With these results, the equation for $\psi$, Eq.~\eqref{kgpsi0}, can be written as
\begin{equation}
\ddot\psi-\frac32\Omega \tan\left[\Omega\left(t-t_0\right)\right]\dot\psi-\frac{\dot\psi^2}{2\psi}+\frac{\psi}{3}\left[b'\left(\psi\right)-3\Omega^2\right]=0.
\label{psiassump3}
\end{equation}
In the particular case where
\begin{equation}
b\left(\psi\right)=3\Omega^2\psi\,,
\label{b2special}
\end{equation}
the potential becomes
\begin{equation}
V\left(\varphi,\psi\right)=-3\Omega^2\left(\varphi-\psi\right)\,,
\label{potiispec}
\end{equation}
the derivatives are $V_\varphi=-V_\psi$ and the last term in the
left-hand side of Eq.~\eqref{psiassump3}
vanishes. The equation can then be integrated analytically to obtain a solution for $\psi$ of the form
\begin{eqnarray}\label{solpsiassump3}
\psi\left(t\right)=&&\psi_1\sec\left[\Omega\left(t-t_0\right)\right]\left\{\sqrt{\cos\left[\Omega\left(t-t_0\right)\right]}\times\right.\nonumber \\
&&\times\left[\psi_0\frac{\Omega}{\sqrt2}-\sqrt{2}E\left({\frac{\Omega}{2}}\left(t-t_0\right)|2\right)\right]\nonumber \\
&&\left.+\sqrt{2}\sin\left[\Omega\left(t-t_0\right)\right]\right\}^2,
\end{eqnarray}
where $\psi_0$ and $\psi_1$ are constants of integration and $E\left(a|b\right)$ is the incomplete elliptic integral of the second kind defined as
$
E\left(a|b\right)=\int_0^a\sqrt{1-b\sin^2\theta}d\theta$.
Figure~\ref{fnumpsi} plots $\psi$ as a function of $t$.
As for the scale factor $a(t)$
the solution for $\psi(t)$ is valid in between times
$t=-\frac{\pi}{2\Omega}+t_0$ and $t=\frac{\pi}{2\Omega}+t_0$,
or times that are translated from these ones
by $2\pi n$ with integer $n$.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{psi.pdf}
\caption{Plot of the function $\psi(t)$ for the case
with a simplified potential equation II, given in Eq.~\eqref{solpsiassump3}, for $\Omega=1$,
$\psi_0=0$, $\psi_1=1$, and
$t_0=0$.}
\label{fnumpsi}
\end{figure}
Finally, the equation for $\varphi$, Eq.~\eqref{kgphi0}, becomes
\begin{equation}
\ddot\varphi-\frac32\Omega \tan\left[\Omega\left(t-t_0\right)\right]\dot\varphi+\frac{\psi}{3}b'\left(\psi\right)-\frac{2}{3}b\left(\psi\right)+\Omega^2\varphi=0.
\label{psieq3}
\end{equation}
Choosing the same function $b\left(\psi\right)$, see Eq.~\eqref{b2special}, Eq.~\eqref{psieq3}
simplifies to
\begin{equation}\label{numphieq2}
\ddot\varphi-\frac32\Omega \tan\left[\Omega\left(t-t_0\right)\right]\dot\varphi+\Omega^2\left(\varphi-\psi\right)=0.
\end{equation}
This equation does not have an analytical solution. However, it can be integrated numerically. In Fig.~\ref{fnumphi2} we plot the solution for $\varphi(t)$ for a specific choice of the parameters
and with $\varphi(0)\equiv\varphi_0=0$ and $\dot{\varphi}(0)\equiv\dot{\varphi_0}=0$.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{varphi.pdf}
\caption{Plot of the function $\varphi(t)$ for the case
with a simplified potential equation II, from numerical results for the integration of Eq.~\eqref{numphieq2} for $\Omega=1$,
$\psi_0=0$, $\psi_1=1$, $\varphi_0=0$, $\dot{\varphi_0}=0$, and
$t_0=0$.}
\label{fnumphi2}
\end{figure}
As for the scale factor $a(t)$ and for the scalar
$\psi(t)$ the solution for $\varphi(t)$ is valid in between times
$t=-\frac{\pi}{2\Omega}+t_0$ and $t=\frac{\pi}{2\Omega}+t_0$,
or times that are translated from these ones
by $2\pi n$ with integer $n$.
The solution is complete since
$k,a, \rho, p, V, \varphi$, and $\psi$ are known.
\subsubsection{Solution with simplified scalar field equations}
\label{method4}
In this section, we set again the matter component to be vacuum,
$\rho=p=0$, and choose a flat universe, $k=0$.
For simplicity, we introduce a potential of the following form
\begin{equation}\label{simpotential}
V=V_0\left(\varphi-\psi\right)^2,
\end{equation}
where $V_0$ is a constant.
Using this potential, the scalar field Eqs.~\eqref{kgphi0} and~\eqref{kgpsi0},
with the notation for the Hubble function $H$ given in
Eq.~\eqref{hubblep},
simplify to
\begin{equation}\label{simvarphi}
\ddot\varphi+3H\dot\varphi=0,
\end{equation}
and
\begin{equation}\label{simpsi}
\ddot\psi+3H\dot\psi-\frac{\dot\psi^2}{2\psi}=0,
\end{equation}
respectively. The first of these equations, Eq.~\eqref{simvarphi}, can be integrated directly after dividing through by $\dot\varphi$ to obtain the equation
\begin{equation}\label{soldphi1}
\dot\varphi=\frac{\varphi_0}{a^3},
\end{equation}
where $\varphi_0$ is an arbitrary constant of integration. To integrate the second equation, Eq.~\eqref{simpsi}, we divide it through by $\dot\psi$ which results in
\begin{equation}\label{simpsi2}
\frac{\dot\psi}{\sqrt{\psi}}=\frac{\psi_0}{a^3},
\end{equation}
where $\psi_0$ is a constant of integration. Equating Eq.~\eqref{simpsi}
to Eq.~\eqref{soldphi1}, one obtains a relation between the scalar fields given by
\begin{equation}\label{soldpsi1}
\frac{\dot\psi}{\psi_0\sqrt{\psi}}=\frac{\dot\varphi}{\varphi_0},
\end{equation}
which can be integrated to yield $\varphi=2\frac{\varphi_0}{\psi_0}\sqrt{\psi}+\varphi_1$,
which upon inverting gives,
\begin{equation}\label{relsca1}
\psi=\left(\frac{\psi_0}{2\varphi_0}\right)^2\left(\varphi-\varphi_1\right)^2,
\end{equation}
where $\varphi_1$ is a constant of integration.
Inserting the potential~\eqref{simpotential} in Eq.~\eqref{modF1}, we obtain
a relation between the Hubble function $H$, and so the scale factor, and the scalar fields as
\begin{equation}\label{fr1}
\dot H+2H^2=\frac{V_0}{3}\left(\varphi-\psi\right).
\end{equation}
Inserting Eq.~\eqref{relsca1} into Eq.~\eqref{fr1}
we obtain
\begin{equation}\label{fr2}
\dot H+2H^2=\frac{V_0}{3}\left[\varphi-\left(\frac{\psi_0}{2\varphi_0}\right)^2\left(\varphi-\varphi_1\right)^2\right].
\end{equation}
Now, after differentiating Eq.~\eqref{fr2}
with respect to time, using Eq.~\eqref{soldphi1} to cancel the $\dot\varphi$ term, and deriving again with respect to time, we can write Eq.~\eqref{fr2} as
\begin{equation}\label{eqndddH}
\dddot H+7\ddot H H+4\dot H^2+12 \dot H H^2=-\frac{V_0\psi_0^2}{6a^6}.
\end{equation}
This is a fourth degree nonlinear differential equation for $a$. It may not have an
analytical general solution, but one can try a particular solution through
an ansatz of the form $a=a_0t^\alpha$, for $a_0$ a constant and $\alpha$ a number.
This ansatz put into Eq.~\eqref{eqndddH} leads to
$\left(-12\alpha^3+18\alpha^2-6\alpha\right)\frac{1}{t^4}=-
\frac{V_0\psi_0^2}{6a_0^6}\frac{1}{t^{6\alpha}}$.
Equating the exponents gives $\alpha=2/3$.
So the scale factor $a(t)$ behaves as
\begin{equation}\label{scale31}
a(t)=a_0\,t^{2/3}\,
\end{equation}
This solution is plotted in Fig. \ref{fig:a4}.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{a4.pdf}
\caption{Plot of the scale factor $a(t)$ for the case
with simplified scalar field equations, given in Eq.~\eqref{scale31},
for the values $a_0=1$, and $t_0=0$.}
\label{fig:a4}
\end{figure}
Moreover, one finds that
$12\alpha^3-18\alpha^2+6\alpha=-\frac{4}{9}$, and so
equating the factors of the equation above yields
a constraint for the constants appearing in Eq.~\eqref{eqndddH}, namely,
\begin{equation}
\frac{V_0\psi_0^2}{6a_0^6}=-\frac{4}{9}\,.
\end{equation}
The solution for the scale factor, Eq.~\eqref{scale3},
allows us to integrate Eqs.~\eqref{soldphi1} and~\eqref{soldpsi1} directly over time, leading to the following solutions for the scalar fields
\begin{equation}\label{invphi4}
\varphi\left(t\right)=-\frac{\varphi_0}{a_0^3t}+\varphi_1 ,
\end{equation}
and
\begin{equation}\label{invpsi4}
\psi\left(t\right)=\left(-\frac{\psi_0}{2a_0^3t}+\frac{\psi_1}{2}\right)^2,
\end{equation}
respectively.These solutions are plotted in Figs. \ref{fig:phi4} and \ref{fig:psi4}, respectively.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{phi4.pdf}
\caption{Plot of the scalar field $\varphi(t)$ for the case
with simplified scalar field equations, given in Eq.~\eqref{invphi4},
for the values $\varphi_0=\varphi_1=1$, and $t_0=0$.}
\label{fig:phi4}
\end{figure}
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{psi4.pdf}
\caption{Plot of the scalar field $\psi(t)$ for the case
with simplified scalar field equations, given in Eq.~\eqref{invpsi4},
for the values $\psi_0=\psi_1=1$, and $t_0=0$.}
\label{fig:psi4}
\end{figure}
This solution is interesting because, for a universe dominated by
matter, i.e., $p=0$, one expects the scale factor to behave as
$t^\frac{2}{3}$, in this case we obtain this behavior considering
$\rho=0$.
The solution is complete since
$k,a, \rho, p, V, \varphi$, and $\psi$ are known.
\subsubsection{Solution with nonflat geometry}
\label{method5}
In this section, we
set again the matter component to be vacuum, $\rho=p=0$. Now, we consider a nonflat geometry, i.e.,
\begin{equation}\label{newk5}
k=1,-1\,.
\end{equation}
Looking at Eq.~\eqref{modF1}, we can try to find a simplified equation for the potential considering that the first three terms inside the brackets cancel each other. Using
Eq.~\eqref{hubblep} this ansatz can be written as
\begin{equation}\label{newhdot1}
\dot H+2H^2+\frac{k}{a^2}=0\,.
\end{equation}
Equation~\eqref{newhdot1} is in fact a nonlinear ordinary differential equation of the second order for $a\left(t\right)$. Integrating this equation, we find the solution
\begin{equation}\label{scale5}
a\left(t\right)=\sqrt{\frac{a_0^2-k^2\left(t-t_0\right)}{k}},
\end{equation}
where $k$ is one of the two values given in
Eq.~\eqref{newk5}, $a_0$ and $t_0$ are constants of integration, and we have
not considered the solution with negative sign for the scale factor $a(t)$. Notice that for $k=1$, there is an end value of $t$, given by $t_e=a_0^2+t_0$, for which the scale factor drops to zero, ending the solution,
as for $t>t_e$ it becomes a pure imaginary number. The same happens for $k=-1$ when $t<t_e$. Therefore, the solution for $k=1$ is only defined for $-\infty<t<t_e$ and the solution for $k=-1$ is only defined for $t_e<t<+\infty$, see Fig.~\ref{fig:a5}.
We know from observations that $k\sim 0$ is a very good approximation. However, the field equations \eqref{genein2} depend on $k$ in such a way that if $k$ varies even slightly from $0$, the structure of the equations changes dramatically. The relevance of the solution we have found, therefore, is to provide a glimpse of the role of spatial curvature in these theories.
\begin{figure}[h!]
\centering \includegraphics[scale=0.75]{am5.pdf}
\caption{Plot of the scale factor~\eqref{scale5} for $a_0=1$, $t_0=0$.}
\label{fig:a5}
\end{figure}
Taking into account Eq.~\eqref{newhdot1}, the equation for the potential, Eq.~\eqref{modF1}, reduces to $V_\varphi=0$. This can be directly integrated over $\varphi$ to obtain
\begin{equation}
V\left(\varphi,\psi\right)=b\left(\psi\right),
\end{equation}
where $b\left(\psi\right)$ is an arbitrary function of $\psi$.
With these results,
we further assume $b'\left(\psi\right)=0$, i.e.,
$b\left(\psi\right)=V_0$, so that
\begin{equation}
V\left(\varphi,\psi\right)=V_0,
\end{equation}
with $V_0$ is a constant,
Then, Eq.~\eqref{kgpsi0} becomes
\begin{equation}
\ddot\psi+\frac{2k^2\left(t-t_0\right)}{k^2\left(t-t_0\right)^2-a_0^2}\dot\psi-\frac{\dot\psi^2}{2\psi}=0.
\end{equation}
This equation can be solved analytically and we obtain
\begin{eqnarray}\label{psi5}
\psi\left(t\right)=&&\frac{\psi_1}{a_0^2-k^2\left(t-t_0\right)^2} \left[\left(a_0^4 \psi_0 k^2+1\right) \left(t-t_0\right)^2\right.\nonumber \\
&&-\left.a_0^6 \psi_0\right] e^{ 2 \tanh^{-1}\left[\frac{t-t_0}{a_0^2 \sqrt{\psi_0} \sqrt{a_0^2-k^2 \left(t-t_0\right)^2}}\right]},
\end{eqnarray}
where $\psi_0$ and $\psi_1$ are constants of integration.
On the other hand, Eq.~\eqref{kgphi0}, simplifies to
\begin{equation}\label{numphieq3}
\ddot \varphi-\frac{3k^2\left(t-t_0\right)}{a_0^2-k^2\left(t-t_0\right)^2}\dot\varphi
-\frac{2}{3}V_0=0.
\end{equation}
This equation is once again decoupled from $\psi$ and can be integrated directly to obtain
\begin{eqnarray}\label{varphi5}
\varphi\left(t\right)=&&\varphi_1+\frac{ t-t_0}{12a_0^2k \sqrt{a_0^2-k^2 \left(t-t_0\right)^2}}
\times\nonumber \\
&&\left[3 a_0^4 V_0 \tan ^{-1}\left(\frac{k\left(t-t_0\right)}{\sqrt{a_0^2-k^2 \left(t-t_0\right)^2}}\right)\right.\\
&&\left.+k \left(a_0^2 V_0 \left(t-t_0\right) \sqrt{a_0^2-k^2 \left(t-t_0\right)^2}+\varphi_0\right)\right],
\nonumber
\end{eqnarray}
where $\varphi_0$ and $\varphi_1$ are constants of integration. The solutions for $\psi\left(t\right)$ and $\varphi\left(t\right)$ are plotted in Figs. \ref{fig:psi5} and \ref{fig:phi5}, respectively.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{psi5.pdf}
\caption{Plot of the scalar field $\psi(t)$ for the case
with nonflat geometry, given in Eq.~\eqref{psi5},
for the values $\psi_0=\psi_1=a_0=1$, and $t_0=0$.}
\label{fig:psi5}
\end{figure}
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{phi5.pdf}
\caption{Plot of the scalar field $\varphi(t)$ for the case
with nonflat geometry, given in Eq.~\eqref{varphi5},
for the values $\varphi_0=\varphi_1=a_0=V_0=1$, and $t_0=0$.}
\label{fig:phi5}
\end{figure}
The solution is complete since
$k,a, \rho, p, V, \varphi$, and $\psi$ are known.
\subsubsection{Solution with perfect fluid matter}
\label{method6}
In this section we assume the presence of a perfect fluid, Eq.(\ref{perffluid}), with an equation
of state as given in Eq.~\eqref{eos}. Then
Eq.~\eqref{econs} provides solutions for $\rho$ and $p$ as
\begin{equation}\label{rho6}
\rho=\rho_0a^{-3\left(1+w\right)}\,,
\end{equation}
\begin{equation}\label{p6}
p=p_0a^{-3\left(1+w\right)},
\end{equation}
respectively,
where $\rho_0$ and $p_0$ are constants of integration that are related to each other by the equation of state~\eqref{eos}.
We also assume a flat universe, $k=0$.
Now, using Eqs.~\eqref{rho6} and~\eqref{p6} we have that the energy-momentum tensor
$T_{ab}$ has the trace $T$ given by
$T=\left(-\rho+3p\right)=-\rho_0\left(1-3w\right)a^{-3\left(w+1\right)}$.
Then Eq.~\eqref{kgphi0} becomes
\begin{equation}\label{newvarphi6}
\ddot\varphi+3H\dot\varphi-\frac{1}{3}\left[2V-\psi V_\psi-\varphi V_\varphi\right]=\frac{\kappa^2\rho_0}{3}\left(1-3w\right)a^{-3\left(w+1\right)}.
\end{equation}
In addition, Eq.~\eqref{kgpsi0} for $\psi$ and Eq.~\eqref{modF1} for $V_\varphi$ remain the same.
So, we just have to solve Eq.~\eqref{newvarphi6} for any of the particular cases discussed before,
the other functions remaining the same. We consider two cases, namely,
the de Sitter expansion of Sec.~\ref{method1} and
the simplified equation for the potential I of Sec.~\ref{method2}.
If we assume a de Sitter expansion, see Sec.~\ref{method1}, then the results
from Eq.~\eqref{scalefactordes1} to~\eqref{solpsi} hold,
in particular $a(t)$ behaves as in Fig.~\ref{fig:a1} and
$\varphi(t)$ as in Fig.~\ref{fig:phi1}.
However, instead of Eq.~\eqref{solvarphiades1} and the corresponding Fig.~\ref{fig:psi1},
we have to solve
Eq.~\eqref{newvarphi6}, which
can be integrated analytically to yield
\begin{eqnarray}\label{solvarphiades11}
\varphi\left(t\right)=&&-\frac{\kappa^2\rho_0\left(a_0e^{\sqrt{\Lambda} t}\right)^{-3w}e^{-3\sqrt{\Lambda} t}}{3a_0^3{\Lambda}\left(4+3w\right)}
\nonumber\\
&&+\psi_0e^{-6\sqrt{\Lambda} t_0}-\frac{2}{7}\psi_0e^{-6\sqrt{\Lambda} t}
\\
&&-2\psi_0e^{-3\sqrt{\Lambda}\left(t+t_0\right)}+\varphi_0e^{-4\sqrt{\Lambda} t}+\varphi_1e^{\sqrt{\Lambda} t}.\nonumber
\end{eqnarray}
This solution is plotted in Fig.~\ref{fig:phimat1}.
If we assume a simplified equation for the potential I, see Sec.~\ref{method2}, then the results
from Eq.~\eqref{eqh1} to~\eqref{psisimppot1} also hold,
in particular $a(t)$ behaves as in Fig.~\ref{fig:a2} and
$\varphi(t)$ as in Fig.~\ref{fig:phi2}.
Again, instead of Eq.~\eqref{varphiother1} and the corresponding Fig.~\ref{fig:psi2}, we have to solve
Eq.~\eqref{newvarphi6}, which
can be integrated analytically to yield
\begin{eqnarray}\label{x6}
\varphi\left(t\right)=\frac{2}{15} V_0 \left(t-t_0\right)^2-\frac{\varphi_0}{\sqrt{2\left( t-t_0\right)}}+\varphi_1
\nonumber \\
-\frac{2\kappa ^2 \rho_0 \left(t-t_0\right) }{3 a_0^2 (3 w-2)}\; \left(a_0 \sqrt{2\left( t-t_0\right)}\right)^{-\left(3w+1\right)}\,.
\end{eqnarray}
This solution is plotted in Fig.~\ref{fig:phimat2}.
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{phimat1.pdf}
\caption{Plot of the scalar field $\varphi(t)$ for the de Sitter case with perfect fluid matter, given in Eq.~\eqref{solvarphiades11},
for the values $\kappa=\rho_0=a_0=\Lambda=\psi_0=\varphi_0=\varphi_1=1$, and $t_0=w=0$.}
\label{fig:phimat1}
\end{figure}
\begin{figure}[h!]
\centering \includegraphics[scale=0.65]{phimat2.pdf}
\caption{Plot of the scalar field $\varphi(t)$ for the case
with a simplified potential equation with perfect fluid matter, given in Eq.~\eqref{x6},
for the values $\kappa=\rho_0=a_0=V_0=\varphi_0=\varphi_1=1$, and $t_0=w=0$.}
\label{fig:phimat2}
\end{figure}
The solutions are complete since
$k,a, \rho, p, V, \varphi$, and $\psi$ are known.
\section{Pulling the previous solutions into generalized hybrid gravity and the form of $f\left(R,\mathcal{R}\right)$}\label{secIV}
\subsection{The main equation for generalized hybrid gravity}
It is useful to obtain the form of the function $f\left(R,\mathcal{R}\right)$ in all the previous cases as the scalar-tensor form of the theory is only a derived form and used to provide an easier platform to perform the calculations. To obtain $f\left(R,\mathcal{R}\right)$, we are going to use the definition of the potential given by Eq.~\eqref{potdef}.
If we replace here the definitions of the scalar fields given in Eqs.~\eqref{phidef} and~\eqref{psidef}, we obtain a partial differential equation for $f$ as
\begin{equation}\label{difpot}
V\left(f_R,f_\mathcal{R}\right)=-f\left(R,\mathcal{R}\right)+f_R R+f_\mathcal{R} \mathcal{R},
\end{equation}
where the subscripts $R$ and $\mathcal{R}$ represent derivatives with respect to these variables, respectively.
As we have seen in the previous section the solutions for the scalar field can have zeros and assume a negative sign. While this does not compromise the physical meaning of the solution in the generalized hybrid gravity picture (in this picture the scalar fields are not real matter fields) it can lead to further constraints on the form of the action. One might require that the derivatives of $f$, which correspond to the scalar fields by Eqs.~(\ref{phidef}) and~(\ref{psidef}), are never zero. In this case the solution we have found will only be meaningful in the intervals in which the scalar field have a fixed sign.
Let us now analyze this equation for the various forms of the potential obtained so far.
\subsection{The previous solutions pulled into the generalized hybrid gravity}
\subsubsection{The de Sitter solution}\label{method1h}
The potential obtained in Sec.~\ref{method1} is given by $V=12{\Lambda}\left(\varphi-\psi\right)$ and therefore Eq.~\eqref{potdef} can be written as
\begin{equation}\label{eqdiff1}
f\left(R,\mathcal{R}\right)=f_R \left(R-12{\Lambda}\right)+f_\mathcal{R}\left(\mathcal{R}-12{\Lambda}\right).
\end{equation}
Defining two new variables as
$\bar{R}=R-12{\Lambda}$ and $\bar{\mathcal{R}}=\mathcal{R}-12{\Lambda}$,
the derivatives of the function $f$ remain the same for the new variables and the equation can be written in a simpler manner as
$
f\left(\bar{R},\bar{\mathcal{R}}\right)=f_{\bar{R}}\bar{R}+f_{\bar{\mathcal{R}}}\mathcal{R}
$,
which has a general solution of the form
$
f\left(\bar{R},\bar{\mathcal{R}}\right)=g\left({\bar{\mathcal{R}}}/{\bar{R}}\right)\bar{R}
$,
where $g$ is an arbitrary function. Coming back to the variables $R$ and $\mathcal{R}$,
we obtain
\begin{equation}\label{sollin}
f\left(R,\mathcal{R}\right)=g\left(\frac{\mathcal{R}-12{\Lambda}}{R-12{\Lambda}}\right)\left(R-12{\Lambda}\right)\,,
\end{equation}
where $g$ is still an arbitrary function.
As an example, let us consider a function $g$ of the form
\begin{equation}
g\left(\frac{\mathcal{R}-12{\Lambda}}{R-12{\Lambda}}\right)=\exp\left(\frac{R-12{\Lambda}}{\mathcal{R}-12{\Lambda}}\right).
\end{equation}
Inserting this function $g$ into Eq.~\eqref{sollin} to obtain $f$, inserting the result into Eq.~\eqref{difpot} and using the definitions of the scalar fields given by Eqs.~\eqref{phidef} and~\eqref{psidef}, we recover the result $V=12{\Lambda}\left(\varphi-\psi\right)$ for the potential and, consequently, the solutions for the scalar fields are the ones obtained in Sec.~\ref{method1}.
However, not all forms of the function $g$ can be used to obtain the correct solutions. Take for example a function $g$ of a power-law form, i.e.,
$
g\left(\frac{\mathcal{R}-12{\Lambda}}{R-12{\Lambda}}\right)=\left(\frac{\mathcal{R}-12{\Lambda}}{R-12{\Lambda}}\right)^\gamma
$, where the exponent $\gamma$ is a number.
This form of the function $g$ is compatible with Eq.~\eqref{eqdiff1} but it does not possess the solution we have reconstructed. This is due to the fact that the relationship between the scalar fields and the Ricci scalars given by $\varphi\left(R,\mathcal R\right)$ and $\psi\left(R,\mathcal R\right)$ cannot be inverted to obtain $R\left(\varphi,\psi\right)$ and $\mathcal R\left(\varphi,\psi\right)$. The equivalence between the metric and the scalar-tensor representations breaks down and one cannot immediately say that the theory will have the solutions reconstructed.
\subsubsection{Solution with a simplified potential equation I}
\label{method2h}
The potential obtained in Sec.~\ref{method2} is given by $V=V_0$, a
constant. We can then write Eq.~\eqref{potdef} as
\begin{equation}
f\left(R,\mathcal{R}\right)=f_R R+f_\mathcal{R} \mathcal{R}-V_0.
\end{equation}
This equation has a general solution of the form
\begin{equation}\label{solconst}
f\left(R,\mathcal{R}\right)=-V_0+g\left(\frac{\mathcal{R}}{R}\right)R,
\end{equation}
where $g$ is again an arbitrary function.
As an example, let us consider a function $g$ of the form
\begin{equation}
g\left(\frac{\mathcal R}{R}\right)=\exp\left(\frac{R}{\mathcal R}\right).
\end{equation}
This exponential form allows us to invert the relation between the
scalar fields and the Ricci tensors given by $\varphi\left(R,\mathcal
R\right)$ and $\psi\left(R,\mathcal R\right)$ to obtain
$R\left(\varphi,\psi\right)$ and $\mathcal
R\left(\varphi,\psi\right)$, and therefore we recover the constant
potential $V_0$ and the solutions for the scalar fields presented in
Sec.~\ref{method2}.
Interestingly, in this case
the
power-law form for the function $g$, i.e.,
$g\left(\frac{\mathcal R}{R}\right)=\left(\frac{\mathcal R}{R}\right)^\gamma$,
also allows us to invert the relation between the scalar fields and the Ricci tensors.
\subsubsection{Solution with a simplified potential equation II}
\label{method3h}
The simple potential obtained in Sec.~\ref{method3} is given by $V=-3\Omega^2\left(\varphi-\psi\right)$, see
Eq.~(\ref{potiispec}). This potential is of the same form of the one obtained in Sec.~\ref{method1} where the only difference is the factor $-3\Omega^2$ instead of $12\Lambda$. Therefore, the procedure to obtain the function $f$ is the same as in Sec.~\ref{method1h} and the solution is
\begin{equation}
f\left(R,\mathcal{R}\right)=g\left(\frac{\mathcal{R}+6c}{R+6c}\right)\left(R+6c\right).
\end{equation}
where $g$ is an arbitrary function.
\subsubsection{Solution with simplified scalar field equations}
\label{method4h}
The potential obtained in Sec.~\ref{method4} is given by $V=V_0\left(\varphi-\psi\right)^2$. With this potential, Eq.~\eqref{potdef} can be written as
\begin{equation}
f\left(R,\mathcal{R}\right)=f_R R+f_\mathcal{R} \mathcal{R}-V_0\left(f_R+f_\mathcal{R}\right)^2.
\end{equation}
This equation is a particular case of the Clairaut equation, with a general solution of the form
\begin{equation}
f\left(R,\mathcal{R}\right)=RC_1+\mathcal{R}C_2-V_0\left(C_1+C_2\right)^2,
\end{equation}
where $C_1$ and $C_2$ are arbitrary constants of integration.
\subsubsection{Solution with nonflat geometry}
\label{method5h}
The potential obtained in Sec.~\ref{method5} is given by $V=V_0$,
a constant potential. This is the same potential as in Sec.~\ref{method2} and therefore the procedure to find the function $f$ is the same and the solution is given by Eq.~\eqref{solconst}.
\subsubsection{Solution with perfect fluid matter}
\label{method6h}
In Sec.~\ref{method6} we presented two different solutions with perfect fluid matter. These solutions were generalizations of the solutions obtained in Secs.~\ref{method1} and~\ref{method2}, where the potentials were given by $V=12{\Lambda}\left(\varphi-\psi\right)$ and $V=V_0$, respectively. Therefore, the solutions for the function $f$ remain the same, given by Eqs.~\eqref{sollin} and~\eqref{solconst} respectively.
\section{Conclusions}\label{conclusion}
In this paper, we devised a number of reconstruction strategies to obtain new exact cosmological solutions in the context of the generalized hybrid metric-Palatini gravity. This theory can be recast in a scalar-tensor theory with two scalar fields whose solutions can be computed analytically for a convenient choice of the potential or the scale factor. Using the new methods we obtained a number of physically interesting solutions including power-law and exponential scale factors.
These solutions reveal some crucial differences not only with respect to GR but also to $f(R)$ gravity. For example it becomes evident that a given behavior of the scale factors in vacuum has always a counterpart in the presence of a perfect fluid, as shown in Sec.~\ref{method6}.
The impact in terms of the interpretation of the cosmological dark phenomenology via this class of theories is clear: a negligible matter distribution still gives rise to expansion laws which are compatible with the observations without requiring the total baryonic matter to be a relevant percentage of the total energy density. This is particularly important in the case of the de Sitter solution given in Sec.~\ref{method1}, because it implies that in these theories inflation could start even if matter is not negligible, and in the result of Sec.~\ref{method4}, which shows that in these theories matter is not central to obtain the classical $t^{2/3}$ Friedmann solution.
Another interesting aspect of the solutions we found concern the behavior of vacuum cosmology and the role of the spatial curvature. In the first case, see Sec.~\ref{method3}, we obtained that the generalized hybrid metric-Palatini theory can have a surprisingly complex behavior in the flat vacuum case. This has important consequences in terms, for example, of the far future of the cosmological models and the meaning of the cosmic no hair theorem in this framework.
Like in any higher order model, in the generalized hybrid metric-Palatini theory the spatial curvature has an important role. Even small deviations from perfect flatness are able to change dramatically the evolution of the cosmology. The result of Sec.~\ref{method5} gives us a glimpse of these differences and their magnitude, showing that vacuum closed models will collapse, whereas open ones expand forever both with a relatively simple expansion law.
It is also interesting to note that our results imply that generalized hybrid metric-Palatini gravity can generate a de Sitter evolution without the appearance of an explicit cosmological constant in the action. Instead, the presence of such a constant leads, throughout our method, to solutions which are radiationlike, see Sec.~\ref{method2}. This result gives us information on the properties of effective dark fluid related to the non Hilbert-Einstein terms present in this theory. It also implies, by the cosmological no-hair theorem, that these solutions should be unstable. The stability of the other solutions we have found cannot be obtained with the same ease. This is a limitation of the reconstruction in general. It allows to obtain some exact solution but it does not offer information on their stability which has to be investigated with different approaches.
Our results suggests that the hybrid theory space is a priori large and requires further investigation not only in terms of the exploration of the solution space, but also in terms of the stability of these solutions.
A final thought should be spent on the role of our results for the testability of this class of theory. It is clear that the solutions we have found can be used to perform a number of cosmological tests, e.g., the ones based on distances, for example. However, a more complete use of the accuracy of the current observations, such as cosmic microwave background anisotropies, requires a full analysis of the cosmological perturbations which is well outside of the scope of our contribution.
\begin{acknowledgments}
JLR acknowledges financial support from Funda\c{c}\~{a}o para a
Ci\^{e}ncia e Tecnologia (FCT) - Portugal for an FCT-IDPASC Grant
No.~PD/BD/114072/2015. SC is supported
by an Investigador FCT Research contract through project IF/00250/2013
and acknowledges financial support provided under the European Union's
H2020 ERC Consolidator Grant ``Matter and strong-field gravity: New
frontiers in Einstein's theory'' Grant Agreement No.~MaGRaTh646597,
and under the H2020-MSCA-RISE-2015 Grant No.~StronGrHEP-690904.
JPSL acknowledges financial support from
the FCT Project No.~UID/FIS/00099/2013,
the
FCT Grant
No.~SFRH/BSAB/128455/2017, the Grant
No.~88887.068694/2014-00 from
Aperfei\c coamento do Pessoal de N\'\i vel Superior (CAPES), Brazil,
and thanks Piotr Chru\'sciel and the
Gravitational Physics Group at the
Faculty of Physics, University of Vienna,
for hospitality.
FSNL acknowledges financial support of
FCT through an Investigador FCT Research contract, with reference
IF/00859/2012 and the Grant No.~PEst-OE/FIS/UI2751/2014.
\end{acknowledgments}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.